Long Green Times and Cycles at Congested Traffic Signals

Transcription

1 2009 D. GRANT MICKLE AWARD: Outstanding Paper in Operations and Maintenance Long Green Times and Cycles at Congested Traffic Signals Richard W. Denney, Jr., Eddie Curtis, and Larry Head Field data were collected and simulation experiments based on traffic at an intersection in Virginia were conducted to test the hypothesis that headways increase with long green times and to test the common assumption that throughput increases with longer cycles. The results showed that headways increased with long green times as a result of departing turning vehicles and that this effect could cause a significant increase in overall average approach headways. The results also showed that maximum throughput, defined as the point where additional offered load could not be served, did not increase with longer cycles. With values derived from the field data, increasing the cycle did not increase throughput. In simulation, increasing the cycle caused a reduction in throughput as a result of increasing the effect of departing turning traffic on the average headway. In ongoing research being conducted for FHWA, the authors interviewed a number of practitioners to identify their methods for responding to congested conditions by using intersection-based signal timing. The interviewees reported that stopline flow seemed to decrease after the first half minute or so of green even if the flow was still being fed by an emptying queue. This finding was reported both as a qualitative effect, where one interviewee suggested that the second 15-s interval following the onset of green showed the highest apparent flow, and as a quantitative effect, where an interviewee noted that the Sydney (Australia) Coordinated Adaptive Traffic System (SCATS) had to be detuned on approaches with unlimited queue because it tended to note a drop in density and incorrectly interpret it as a drop in demand. A review of the literature reveals only two studies that look into the stability of stopline flow directly and more that consider the relationship between capacity and green time. One was performed by Teply et al. for the Canadian Capacity Guide (1). Their data showed that flow peaked at around 2,000 vehicles/h in the interval between 15 and 30 s from the start of the green and then after 30 s diminished to a value closer to 1,800 vehicles/h. These results would seem to confirm the impression of the interviewees. (Additional work was done as part of the supporting work surrounding Project NCHRP 3-66, in which Smaglik et al. looked at flow with respect to the end of the green (2). However, they were primarily studying the flow of cars at the point where the signal controller gapped out, not at the flow of vehicles being fed by a standing queue.) R. W. Denney, Jr., 107 Carpenter Drive, Suite 230, Sterling, VA E. Curtis, FHWA Resource Center, 61 Forsyth Street SW, Suite 17T26, Atlanta, GA L. Head, System and Industrial Engineering Department, University of Arizona, Tucson, AZ Corresponding author: R. W. Denney, Jr., rwd@ iteris.com. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / Khosla, in a master s thesis at the University of Texas at Arlington (3), and Khosla and Williams (4) studied the matter at four intersections in the Dallas Fort Worth, Texas, metropolitan area. Khosla found that the flow increased with time into the green through about the first 20 s (the flows were reported in 10-s intervals) and then stabilized until the final one or two 10-s intervals, when it increased (presumably as an effect of rushing the clearance interval and considering that not all green periods were the same). At two of the four locations, the flow diminished starting with the interval at 40 s into the green (except the final interval). Interestingly, the two intersections that showed a reduction had auxiliary turn lanes for both left- and right-turners, though the data were not collected from the lane adjacent to the turn lanes. Two causal mechanisms have been hypothesized to explain the effect of diminishing stopline flow from an emptying queue, assuming that it exists. One is that turning vehicles may be trapped in a long queue in the through lanes. They will depart the through lanes for the turning lanes during the green for the through vehicles, thus lessening the flow on the through lanes. A more severe cause of this effect is an approach that widens to additional through lanes as it approaches the intersection. As the departing queue works its way back to the narrower section, the stopline can only be fed by the cars stored in that narrower section, which would starve the capacity of the wider section at the stopline by some amount. The second hypothetical mechanism might be that drivers are no longer able to respond to the displayed green light because their position in the queue puts them too far away to see the signal. If the drivers cannot see the signal, their perception reaction time may no longer overlap with that of drivers of vehicles in front of them as characterized originally by Greenshields (5). If they are reacting instead to the brake lights of the cars in front of them or to movement in an adjacent lane, some of the perception reaction time might be added back into their initial headway. The relevance of this issue is significant for saturated networks. One operating assumption common among practitioners is that intersection total throughput increases as cycle length increases, because the phase-change lost time, which is constant, diminishes as a percentage of the cycle. The common practice of increasing cycle length to increase capacity is based in part on the assumption that saturation flow remains constant once the initial lost time has been accommodated, and the effect of traffic departing from through lanes for turn lanes is ignored. If this assumption turns out to be false, or if the effect of departing turn traffic is significant, the common practice might not have the effect of increasing throughput for the congested movements, which is the initial objective in saturated conditions. A comprehensive study designed to model the change of saturation flow through long greens is beyond the scope of the study. However, the researchers evaluated one location with field studies and simulation to evaluate the effect of the two postulated causes and considered the impact of these effects on throughput as the basis for developing a throughput-based cycle length recommendation for practitioners. 1

2 2 Transportation Research Record 2128 STUDY SITE The intersection of Virginia Route 28 and Frying Pan Road provides a useful opportunity to evaluate any reduction in flow with long green times. The Virginia Department of Transportation has been systematically replacing at-grade signalized intersections along Route 28 with grade-separated interchanges, elevating Route 28 to expressway status. The intersection at Frying Pan Road is one of the few remaining at-grade intersections controlled by a traffic signal. The next signalized intersections both upstream and downstream from Frying Pan are about 5 mi away, and in both directions major facilities freely interchange with Route 28 before they reach those at-grade intersections. Thus, demand from upstream is little affected by traffic signals and there are no constraints on traffic downstream from the location. Demand at the intersection is very high; about 875 vehicles were actually served on the northbound approach in the peak 15 min of the study. This finding corresponds to an hourly flow of roughly 3,500 vehicles/h at the stopline, with significant queuing indicating demand not being served. Route 28 carries this traffic on three main lanes in the northbound direction, adding a right-turn lane about 500 ft upstream from the stopline. Frying Pan Road is a T-intersection with no northbound left turn. Signal operation at the intersection was fixed by observation. Green times for the northbound through movement were measured to be 180 s in a cycle of 270 s. FIELD DATA The researchers developed a software program that allowed a field observer to record the time each vehicle crossed the stopline in each lane by key press, in milliseconds from the start of the green. The observation method is subject to systematic errors based on the skill of the observer, with estimated errors of up to a half second in pressing the key. In order to filter out the noise induced by any flaws in the observation technique, vehicles were grouped in ranks of five and the headways between them were averaged. Data were collected over 11 consecutive cycles during one morning peak period in March Flow in each lane ranged from 58 to 101 vehicles during the green. Initially, the data were only evaluated for the first 60 vehicles crossing the stopline in each lane as representative of long green times sufficient for the objectives of the study. The headways between the first 60 vehicles in each lane were first evaluated to test for correlation to the cycle number. The data showed that the headways of vehicles from queue positions 40 to 60 were longer during the last three cycles of the study than during the first eight, suggesting that the queue had more limited demand at the end of the study period. Therefore, headways after the first 40 vehicles were ignored for the last three cycles. As a result, 11 cycles were used for the first 40 vehicles crossing the stopline, and 8 were used for the next 20. A total of 1,800 vehicle headways were included in the subsequent analysis. The 1,800 vehicles included do not reveal correlation by cycle numbers, as is illustrated in Figure 1, where those vehicles are Headway (ms) Cycle 1 Cycle 2 Cycle 3 Cycle 4 Cycle 5 Cycle 6 Cycle 7 Cycle 8 Cycle 9 Cycle 10 Cycle Position in Queue FIGURE 1 Headways by position in queue for different cycles (headways averaged across lanes in ranks of five vehicles) (ms = milliseconds).

3 Denney, Curtis, and Head 3 averaged in ranks of five. For all evaluations of headway, the first five ranks of vehicles were ignored to eliminate the effects of the start-up lost time. Another way to display these data is to show vehicle stopline crossing by time into green. Figure 2 shows time-into-green crossing by lane and cycle. The thick black diagonal line shows the Greenshields model with an assumption of 4 s lost time and a saturation flow of 1,900 vehicles per hour of green. One immediate observation is that the Greenshields relationship with currently accepted numbers proved to be an ideal not often exceeded by the field data. The results of primary interest are shown in Figure 3, where headways by lane (and across all lanes) are averaged in groups that include five ranks of vehicles, from the 6th to the 60th vehicle in the departing queue (40th vehicle for the last three cycles). Analysis Linear regressions were conducted for headways (in ranks of five) for each lane and for all lanes. The regressions provide an assessment of changing headway (the slope of the line) and correlation between queue position and headway. Correlation alone is insufficient to show that headways increase; statistical testing is also required on the coefficient of the position-in-queue parameter (which is the slope of the line) to determine that the slope is significant. Thus, a two-tailed test of the t-statistic was performed on the slope parameter. The linear regression for the headways in the right through lane was H = 22. 4( Q ) + 1, 781 ( 1) pos where H is the headway in milliseconds and Q pos is the position in the departing queue. With 11 data points in the regression, the calculated t-statistic was With 9 degrees of freedom, the value greater than which one rejects the hypothesis that the slope value is due to chance at a confidence of 95% is Thus, for the right lane (Lane 1), the slope is significantly nonzero. Similar calculations were made with the middle and left lanes, but the slopes of those lanes were not sufficiently different from zero to reject the hypothesis that they are caused by chance. The conclusion is that the right lane experienced increasing headways, but the center and left lanes did not. Total headways, averaged across all lanes, did show a significant slope based on the influence of the right lane in the average. The left lane was an ideal test for evaluating the hypothesis that driver reaction time starts later and adds to the headway beyond a certain distance from the stopline. Yet headways did not increase with position in queue in the left lane. Thus, the data do not support this hypothesis. 180, ,000 1,1 2,1 3,1 Time Into Green (ms) 140, , ,000 80,000 60,000 Line shows Greenshield equation using typical HCM numbers: Time into green = 4 + (3,600/1,900)n where n = vehicle position in queue Each point represents one vehicle 1,800 vehicles represented 1,2 2,2 3,2 1,3 2,3 3,3 1,4 2,4 3,4 1,5 2,5 3,5 1,6 2,6 3,6 1,7 2,7 3,7 1,8 2,8 3,8 40,000 1,9 2,9 3,9 20,000 1,10 2,10 3, Vehicle Position in Queue 1,11 2,11 3,11 FIGURE 2 Crossing times by lane and cycle, VA-28 at Frying Pan Road (first 60 vehicles in each lane, first 40 after Cycle 8).

4 4 Transportation Research Record y = x R 2 = Headway (ms) Position in Queue 1-40: 11 cycles represented Each point represents 55 vehicles in each lane Each point represents 165 vehicles in all lanes Position in Queue 41-60: 8 cycles represented Each point represents 40 vehicles in each lane Each point represents 120 vehicles in all lanes Slope is significant at 95% confidence (t test) for Lane 1 and all lanes y = x R 2 = y = x R 2 = y = x R 2 = Lane 1 Lane 2 Lane 3 All Lanes Linear (Lane 1) Linear (Lane 2) Linear (Lane 3) Linear (All Lanes) Reported correlation is not due to chance at 95% confidence (F test) for all regressions Position in Queue FIGURE 3 Headway by lane (averaged in ranks of five vehicles). The right lane was an ideal test for evaluating the effect of turning traffic leaving the through lane. The right-turn lane is about 500 ft long. Approximately 25 to 30 vehicles would be expected to fill a 500-ft queue, so one can assume that few right-turners will be in the first 25 or 30 queue positions because they would have already been able to move into the right-turn lane. Thus, one would expect to see headways mimic the center and through lanes through a position in queue of 25 or 30. The data do show a slight increase in headway through that much depth of queue and then jump significantly thereafter to a peak at a queue position of 30 to 35. After that peak, the headway drops back down around the 45th vehicle and then peaks again at the 55th vehicle. A possible explanation for this behavior is that the departing queue works its way back to beyond the opening of the right-turn lane, allowing right-turners to leave the right through lane to make their maneuver. They leave gaps. Drivers behind them may accelerate to fill those gaps but are unable to catch up to the gap by the time they reach the stopline. The opening creates a shock wave and flow increases again, resulting in shorter headways as the shock wave emerges. After the shock wave passes, headways increase again. The middle lane showed growing headways to a greater extent than the left lane, though this finding was not found to be statistically significant. If the effect is real, however, it could be explained by the situation in which some vehicles in the middle lane move into gaps created by departing right-turners in the right lane before they reach the stopline. Correlation was also tested. Some correlation was seen in all cases, but the correlation coefficients range from moderate to low. The F-statistic was used to evaluate whether these correlations could be attributed to chance, and in all cases, it exceeded the 95% critical value on the F-distribution, meaning that the reported correlation is probably not due to chance. Given the low correlation coefficients, however, it is not recommended that the linear regressions be used as a model to characterize headways. One question to be evaluated is the effect of the headways on various assumptions pertinent to calculating green times at saturated intersections. The Highway Capacity Manual (6) recommends a saturation flow value of 1,900 vehicles per hour of green. This value corresponds to a headway of just under 1.9 s per vehicle. Also, practitioners commonly estimate lost time at 4 s. Thus, the 60th car in a departing platoon should reach the stopline in 118 s, which is 4 s lost time plus 1.89 s headway multiplied by 60 vehicles. Figure 4 shows the average time required to serve cars at each rank in the departing queue. The numbers were calculated by averaging the time into green for all the cars in that rank and lane (or across all lanes). The solid diagonal line in Figure 5 illustrates the Greenshields model with modern parameters. The left lane corresponded very

5 Denney, Curtis, and Head 5 160, ,000 Stop Line Crossing Time (ms After Start of Green) 120, ,000 80,000 60,000 40,000 Solid Line: Green time = 4 + (3,600/1,900)n, where n = number of cars (Greenshields) Position in Queue 1-40: Each point represents 11 vehicles per lane, 33 vehicles all lanes Lane 1 Lane 2 Lane 3 All Lanes 20,000 Position in Queue 41-60: Each point represents 8 vehicles per lane, 24 vehicles all lanes Position in Queue FIGURE 4 Time into green by rank and lane (averaged across cycles). closely to the Greenshields model, and the middle lane was close, with an accumulated error of about 5 s by the 60th rank of vehicles. The right lane, however, fell behind the ideal by around 35 s. Nevertheless, the right lane closely corresponded to the ideal through about the 25th rank. This rank is the point at which headways spiked in response to the clearing back of the queue to the point where right-turners would again be in the through lane. Conclusions and Further Discussion The field data constitute a single condition and location and consequently cannot be used to construct models. But the data show that lanes adjacent to turning lanes will show significantly increased headways and reduced stopline flow once the queue has cleared to the upstream end of the turning lane. The field data showed only the effects of right-turners, which always receive the green at the same time as through vehicles. One would expect the effect to be more pronounced with left turns, with typically shorter left-turn bays and often separate signalization. Even the effect of a right lane, however, was sufficient to cause a significant increase in average headway across all lanes and an accumulated error in predicted required green time of about 15 s. Headways were not affected by turning traffic downstream of the turn-lane entrance. This finding suggests that maximum throughput is served when green times use the ability to feed the stopline with maximum flow. One can turn the analysis around by working backward through the foregoing illustrations to find the number of vehicles that can be served in a given amount of green. For example, Figure 5 shows how many vehicles can be served in, say, 118 s of green. The average across all lanes was 54, or 162 vehicles in total. To understand the effect of various green times, one might also ask how many vehicles can be served in, say, 48 s. From the data, the average across lanes is about 23, or 66 vehicles in total, per green period. If, hypothetically, the 118-s green were two-thirds of the cycle (as is the 180-s green in a 270-s cycle as observed in the field), the cycle would be 177 s. The hourly throughput would be cycles/h times 162 vehicles/cycle = 3,295 vehicles. However, if a 48-s green time was also two-thirds of a cycle, the cycle length would be 72 s, of which 50 could be operated in an hour. The total throughput for that hypothetical scenario would be 3,300 vehicles. The significance of 48 s is that this time is about the green time required to serve the vehicles that can queue up downstream of the turn-lane entrance. (A 72-s cycle is impractically short at this location because of the short times that would be available for the other movements if the same percentage of green is to be maintained.)

6 6 Transportation Research Record y = x R 2 = y = 1.084x R 2 = Headway (ms) cycles represented Each point represents 50 vehicles in each lane Each point represents 150 vehicles in all lanes y = x R 2 = Lane 1 Lane 2 Lane 3 All Lanes Linear (Lane 1) Linear (Lane 2) Linear (Lane 3) 1000 Slope is significant at 95% confidence (t test) for Lane 1 only Reported correlation is not due to chance at 95% confidence (F test) for Lane 1 and for all lanes regressed together Position in Queue FIGURE 5 Simulated headways by position in queue (averaged in ranks of five vehicles). Thus, by keeping the green time down to the point where only the queue to the upstream end of a 500-ft turn lane was served in each cycle, flow at the stopline is maintained close to ideal saturation and the overall throughput does not decrease. It is concluded, therefore, that the common belief that longer cycle lengths can be assumed to result in greater capacity cannot be supported by the behavior at this intersection. The actual signal timing at the intersection was 180 s of green in a cycle of 270 s. The measured flow during the first eight cycles, which were observed to have maintained queue-departure flow for the entire green time, was 1,986 vehicles in 2,160 s, for an hourly throughput of 3,310 vehicles. Again, increasing the cycle does not result in greater throughput. Thus, the reduction in lost time as a percentage of the cycle does not overcome the reduced flow caused by turning vehicles leaving the through lanes in very long green times. SIMULATION To analyze optional control approaches, VISSIM v.4.3 was used to provide microscopic simulation. Issues emerged during calibration, primarily that VISSIM allows lane-changers to stop at the final opportunity for the maneuver and block traffic indefinitely (VISSIM eliminates the vehicle after 60 s). This behavior is unrealistic. Even those drivers who do wait to the last minute usually give up on making the lane change, or benefit from a charitable driver in the lane into which they are moving, or force their way in more aggressively than VISSIM was prepared to simulate. To prevent these behaviors, the researchers set up the VISSIM network such that right-turners started in the right through lane at the entrance to the network so that they would already be in the right lane as they approached the entrance to the right-turn bay. The researchers also provided two parallel links covering the length of the right-turn lane. One link served the left two lanes and the other served the right through lane and the right-turn lane. This configuration made it possible for right-turners to move into the right-turn lane at any point along its length. At the right-turn island gore, the two parallel links are connected to a single link serving the three through lanes. Input volumes were chosen to ensure that the network was oversaturated, with an approach hourly volume of 6,000, 750 of which turned right. The 6,000 entering vehicles were divided equally among the three lanes. Approach links were made as long as the license of the copy used permitted, about 8,000 ft. The queue overflowed the approach links throughout the simulation, but this factor caused no visible effect at the intersection. The simulation was run for 4,000 s, and data were collected starting 1,080 s into the simulation (four cycles). The signal timing was the same as was observed in the field study. The simulation stopped

7 Denney, Curtis, and Head 7 during a red period, so data were collected from 10 consecutive departing queues. Figure 6 shows the layout of the VISSIM network, with links and connectors shown only by their centerlines. The northward direction is toward the top of the screen, and the northbound approach was the object of study. The Wiedemann 74 car-following model in VISSIM (7) was used, which includes two user-defined parameters bx_add and bx_mult. Several values were tried, and the resulting headway and time-intogreen, on the basis of a visual evaluation of the charts, with a bx_add of 2.6 and a bx_mult of 3.6 performed best. These values are similar to the default values of 2.0 and 3.0, respectively. VISSIM was configured to export raw data for data collection points at the stopline. The data were manually filtered to include only records that indicate a vehicle first entering a detector. These records were then evaluated in an Excel spreadsheet as with the data collected in the field. Figure 5 shows headways as recorded in the simulation. As with the field data, the headways were averaged across ranks of five queue positions. The simulated data show a similar shape to the field data, in that headways remain the same with respect to queue position for the middle and left lanes but lengthen as a result of departing rightturners in the right through lane. The simulation also shows a sud- den increase in headway in the right lane just upstream from the entrance to the right-turn lane, followed by a return to shorter headways, and then followed by another peak. No simulation values were found that matched the field-measured y-intercept of the linear regression for the right lane, but the slope and the range of values were similar. Also, the sudden increase in headways at the upstream end of the right-turn lane happened earlier in the departing queue by a few vehicles. The simulated headways in the right lane were just a bit short overall but were judged close enough to the field data to provide a reasonable means for comparing alternative strategies. Figure 7 shows the time into green, again with a solid diagonal line representing the Greenshields equation with modern values. COMPARISON OF ALTERNATIVE SIGNAL TIMING The main purpose of the field study at Frying Pan Road was to observe whether throughput does actually increase or decrease as a result of very long greens and cycles, and whether shorter or longer cycles affect the overall throughput in ways generally assumed by practitioners. FIGURE 6 VISSIM network, VA-28 at Frying Pan Road.

8 8 Transportation Research Record , , ,000 Stop Line Crossing Time (ms) 100,000 80,000 60,000 Lane 1 Lane 2 Lane 3 All Lanes 40,000 20, Position in Queue FIGURE 7 Simulated stopline crossing time (averaged across cycles). Evaluating the effect requires a means of characterizing the throughput of the intersection. The simulation was repeated to compare the offered load (demand) with the served load (throughput) on the northbound approach for a range of input volumes. All simulation parameters were kept the same except input volume. Right turns were kept at 12.5% of the approach volume. Offered load is the total input volume that VISSIM was asked to process that was routed onto the northbound through lanes (rightturners are excluded in order to provide a comparison with field data, which measured only through cars on a cycle-by-cycle basis). Because the residual queue spilled out of the simulation network, it was not possible to characterize demand upstream from the queue. A check of the similarity between VISSIM s nominal volume and actual volume found the two values to be within a few percentage points in all cases in which the residual queue did not spill out of the network. Thus, the offered load was assumed to be the hourly demand that was nominally programmed into the simulation. The served throughput was the total volume of traffic measured at data collection points in the through lanes at the stopline, as recorded from VISSIM. Figure 8 shows offered versus served loads at the simulated intersection and includes the northbound through movement. For an intersection that is able to serve all the demand presented to it, the illustration should show a one-to-one relationship between offered and served loads (shown by the solid diagonal line in Figures 8 and 9). When the intersection reaches the point at which additional loads cannot be served, the curve relating the two flattens to a horizontal line that represents maximum throughput. Figure 8 shows that throughput tracks demand in the absence of growing residual queues, as expected, until the maximum throughput is reached. At that point, additional demand cannot be accommodated and the served throughput reaches a ceiling. In the simulation, the maximum throughput was a little less than 3,500 vehicles/h, which corresponded reasonably well with the maximum observed throughput of about 3,300 vehicles/h reported in the field data. The difference can be explained by the slightly reduced headways as simulated versus those observed. This approach in identifying the maximum throughput by comparing offered and served loads provides a means of evaluating alternative strategies for minimizing congestion on the basis of the throughput objective. The simulations were repeated with two other sets of signal timing to correspond to the scenarios derived from the field results. In addition to the field case of a 180-s green within a 270-s cycle, simulations were conducted with a 48-s green in a 72-s cycle and a 118-s green in a 177-s cycle. It should be noted again that the 72-s cycle is impractically short and would result in absurdly short green times on the minor movements and be unattainable in practice. The 177-s

9 Denney, Curtis, and Head Served Load (Throughput)(vph) Offered Load (vph) FIGURE 8 Offered load versus throughput from simulation, through lanes only (180-s green, 270-s cycle). (Solid diagonal line shows ideal throughput.) cycle, in contrast, allowed reasonable phase times and resulted in no congestion in the other movements with the use of field volumes. Figure 9 shows the maximum throughput curves for the three timing scenarios. DISCUSSION OF RESULTS AND CONCLUSION Two influences on the throughput compete against one another. The first is the effect of phase-change lost time, which increases as a percentage of the cycle as the cycle length is reduced. As was mentioned at the outset, the consumption of the cycle by lost time is the usual motivation for using longer cycles in the presence of capacity problems and congestion, on the basis of the assumption that the longer cycle is less consumed by lost time and therefore more efficient. The competing influence is the effect of the right turn. As was seen in both the field data and the simulation, the headways and saturation flow of the right lane were similar to those of the middle and left lanes and also fairly constant until the queue emptied to a point adjacent to the right-turn bay entrance. Upstream from the right-turn lane entrance, the right through lane contains all the right-turners who are unable to reach the right-turn lane because of the residual queue. Once the queue clears back to the point where right-turners are present in the queue, their departure into the right-turn lane as the queue clears will leave gaps in the through traffic. Through cars attempt to fill the gap, partly from further back and partly from the middle lane. Both field data and simulation showed that headways peak, recover, but then peak again as the shock wave crosses the stopline. At the field site, the effect of the right turn was sufficiently dominant to have a significant effect on the overall average headway on the approach. Consideration of the field data for the number of vehicles served, on average, across all three lanes during the hypothetical green times of 48 s and 118 s and the actual green time of 180 s and then converting to hourly flows suggested that all three timing scenarios would provide about the same throughput. Given that the 48-s green time serves only cars queued downstream of the entrance to the right-turn lane, the headways were expected to be uniformly short and unaffected by right-turners. The simulation showed this result by indicating that the 48-s green time provided the highest maximum throughput, with the assumption that the green percentage of the cycle was held constant. The simulation also indicates that the larger the percentage of the green that can be consumed by flows unaffected by turning traffic, the higher the throughput. The medium cycle in the simulation provided higher throughput than the longest cycle but not as high as the

10 10 Transportation Research Record Served Load (Throughput)(vph) s Green, 270 s Cycle 118 s Green, 177 s Cycle 48 s Green, 72 s Cycle FIGURE Offered Load (vph) Simulated maximum throughput for three signal timings, through lanes. shortest cycle. This result is the opposite of the assumption made by most practitioners and indicates that in this simulated case at least, the effect of turning traffic on overall throughput overcomes the reduced cycle efficiency caused by lost time. The simulation was perhaps more optimistic than the field data, though the field data represented only the longest cycle. In any case, a safe conclusion is that at this site, the longer cycle is not increasing throughput. In conclusion, the field data did not support the hypothesis that lost time builds back into the traffic stream as queues empty through very long greens. However, turning traffic leaves gaps in the through lanes, and the data at this site show that those gaps do result in a significant decrease in throughput. The less green time is used to serve a queue containing turning vehicles, the greater the maximum throughput in the through lanes. ACKNOWLEDGMENT The work reported was part of an ongoing research project, Signal Timing in Saturated Conditions, sponsored by FHWA. The authors are grateful for the support provided by this research effort. REFERENCES 1. Teply, S., D. I. Allingham, D. B. Richardson, and B. E. Stephenson. Canadian Capacity Guide for Signalized Intersections, 2nd ed. Canadian Institute of Transportation Engineers, June Smaglik, E. J., D. M. Bullock, T. Urbanik II, and D. B. Bryant. Evaluation of Flow-Based Traffic Signal Control Using Advanced Detection Concepts. In Transportation Research Record: Journal of the Transportation Research Board, No. 1978, Transportation Research Board of the National Academies, Washington, D.C., 2006, pp Khosla, K. Determining Effects on Saturation Flow Rate During Longer Cycle Lengths, MS thesis. University of Texas, Arlington, Khosla, K., and J. C. Williams. Saturation Flow at Signalized Intersections During Longer Green Time. In Transportation Research Record: Journal of the Transportation Research Board, No. 1978, Transportation Research Board of the National Academies, Washington, D.C., 2006, pp Greenshields, B. D., J. R. Bibbins, W. S. Channing, and H. H. Miller. A Study of Traffic Capacity. Highway Research Board Proceedings, Vol. 14, Highway Capacity Manual. TRB, National Research Council, Washington, D.C., VISSIM 4.30 User Manual. Planung Transport Verkehr AG, Karlsruhe, Germany, The Traffic Signal Systems Committee sponsored publication of this paper.

11 Assessment of Traffic Signal Maintenance and Operations Needs at Virginia Department of Transportation Wenling Chen, Larrie Henley, and Jeff Price Budgeting for the maintenance and operations (M&O) of traffic signal assets can be done in many ways. Systematic, data-driven approaches are applied to various extents in transportation agencies. However, little has been documented regarding signal resource allocation practices. A traffic signal needs assessment framework recently developed by the Virginia Department of Transportation to support signal M&O funding allocation is presented. The framework offers the potential for incorporating several performance measures into the needs assessment process, including mean time between failures, mean time to repair, average signal component service life, average remaining service life, and share of preventive maintenance costs with respect to overall maintenance costs. Through adjustment of related parameters, the needs assessment framework allows users to set performance targets against key signal system performance measures and compare needs for meeting various performance scenarios. Budgeting and resource allocation for the maintenance and operations (M&O) of traffic signal assets can be done in many ways. Some agencies use simple guidelines such as unit cost per centerline mile or a percentage of present constructed value (1); others perform bottomup estimates based on anticipated labor and equipment needs. An FHWA 2004 signal asset management state-of-the-practice review indicates that half of the agencies reviewed (with system sizes ranging from 300 to 1,000 signals) apply systematic and data-driven approaches of various extents for signal budgeting (2). However, little detailed information is available documenting signal resource allocation practices. The traffic signal needs assessment methodology that the Virginia Department of Transportation (VDOT) recently developed to support signal M&O budget allocations is presented. Background on VDOT s system and its needs-based budgeting approach in general is presented, followed by elaboration of the elements of VDOT s signal needs assessment framework. This account is followed by a discussion of the potential that the framework provides for incorporating a number of key performance measures into a signal needs assessment process. Operations Planning Division, Virginia Department of Transportation, 1401 East Broad Street, Richmond, VA Corresponding author: W. Chen. Chen@vdot. virginia.gov. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / VDOT S SYSTEM AND NEEDS-BASED BUDGETING APPROACH Virginia has the third-largest state-maintained highway system in the country (after North Carolina and Texas), with approximately 57,000 centerline miles of roads and 19,293 structures (bridges and large culverts) (3). With regard to signalization, VDOT is responsible for more than 3,000 signalized intersections. Table 1 summarizes the various categories of assets in the Virginia state highway system. The row beginning with Signals is bolded as the item thematic to this paper. Similar to many other state DOTs or large transportation agencies, VDOT s signal M&O functions are served by separate units (4). Signal maintenance, including preventive maintenance (PM), repair, and replacement, is conducted by the field operations maintenance sections. Signal operations functions including signal timing, system monitoring, and system upgrades are carried out by field traffic engineering sections. In addition, specialized functions such as structural inspections required by professional engineering or rail crossing preemption inspections are performed by field structure and bridge units or by external entities such as the Virginia Department of Rail and Public Transportation. Maintaining and operating a large inventory of signals presents a great challenge to VDOT. In FY 2008, VDOT spent approximately $24.6 million on signal M&O. This figure is a nearly $8 million increase from the FY average. Decisions on resource allocation and ranking of M&O needs by priority are essential to the upkeep of system performance. In order to optimize the trade-offs across investments in system operations and preservation, it is critical to have methodologies and tools that support analysis and understanding of the costs, benefits, and performance outcomes associated with alternative investments. Beginning with FY 2006, VDOT began using a needs-based budgeting (NBB) approach to identify asset maintenance needs based on inventory and condition information of many assets (5). The NBB approach led to a 10% increase ($97.4 million) in the FY 2006 maintenance funding over the originally planned allocation (6). This NBB process has since been used to develop the annual budget request and to priority-rank the statewide allocation of available M&O funding. VDOT continues to develop data systems and processes as part of a vision to become a performance-based organization in which investments are linked to performance outcomes (3). The NBB approach has allowed VDOT to shift from allocating traffic signal funding on the basis of historical expenditures to using an objective, quantified method driven by maintenance needs and data. The assessment draws on the field-maintained signal inventory, generally accepted life-cycle assumptions, and first-hand knowledge of asset condition and serviceability (5). 11

12 12 Transportation Research Record 2128 TABLE 1 Inventory in Virginia State Highway System, 2007 Total VDOT Inventory Item Interstate Primary Secondary Frontage Maintained Lane mile 5,383 21,642 97, ,811 Bridges and culverts 3,010 5,012 11,271 19,293 Tunnels Toll facilities 3 3 Tolled lane mi Safety rest areas Welcome centers Ferries (vessels) Pipes a 8,000 58, , ,000 Ditches a (ft) 16,067,000 65,126, ,278, ,471,000 Unpaved shoulders a (ft) N/A 64,085, ,800, ,885,000 Pavement marking a (ft) 57,029, ,620, ,142, ,791,000 Guardrail a (ft) 9,353,000 11,739,000 6,655,000 27,747,000 Signs a 40, , , ,000 Signals 0 1,802 1,228 3,030 Cameras Dynamic message signs Traffic sensors 1,416 1,416 Count stations Roadway weather information systems Fog detectors HOV gates Highway advisory radio a Estimated inventory based on statistical sampling. SOURCE: VDOT, 2007 (3). FRAMEWORK FOR SIGNAL M&O NEEDS ASSESSMENT In essence, VDOT s signal M&O needs are determined on the basis of the size and age (or condition) of inventories, the deterioration pattern of signal equipment, and the quantity and unit cost of each type of maintenance activity. The framework for signal needs assessment (Figure 1) requires three types of input: Characteristics of the system or asset types, quantities, age, life expectancies, and cost of the physical components that make up the signal systems. Major categories of signal physical components include structure, hardware, software, detection devices, and the communications infrastructure that connects and controls the signals. Definition of M&O work work categories, definitions, frequencies, and resource requirements. Infrastructure performance criteria the objectives for meeting the needs, mainly referring to the performance targets set for the needs assessment. Currently, VDOT s signal M&O performance target is to maintain the system s existing condition. Characteristics of System Defining signal system assets establishes a baseline from which data, information, performance measures, and needs can be further identi- fied (7). The signal components and associated life expectancies, which are applied in VDOT s signal needs assessment process, are summarized in Table 2. The system components and life expectancies were compiled through consultation with field maintenance managers, who are in charge of signal maintenance work. Signal systems in VDOT s nine districts vary in complexity, and hence different districts have different compositions of the components listed. The component life expectancies provided in Table 2 represent expert knowledge of inventory performance and M&O practices within VDOT. They do not necessarily represent standards or practices applied in other agencies. For example, although manufacturers commonly rate light-emitting diodes (LEDs) for a 10-year expected life and provide a 5-year warranty (8), VDOT has found that early models failed much sooner. This outcome is especially true in areas of the state with warmer climates. Interviews with VDOT maintenance staff suggested that LEDs lasted about 5 years and then started to lose luminosity. In addition, LEDs close to 5 years of service were sometimes found with only a few effective diodes left. A 2005 LED performance survey by the California Energy Commission indicated that 23% of more than 300 participating jurisdictions reported an average of 4.5 years of LED life expectancy (9). Also, unlike agencies with large inventories of electromechanical controllers, VDOT has a different composition of controller types. Most of VDOT s signal controllers are computerized and are expected to become functionally obsolete within 7 years in service, which could be sooner than many other agencies would expect. Cabinets normally last much longer than controllers. However, the prac-

13 Chen, Henley, and Price 13 Inventory Data Requirements Asset (System) Define the System: - Asset Components (Subsystems) Types and Quantities - Age and Life Expectancy - Replacement Cost Needs Assessment Preventive Maintenance Needs Work Define M&O Works: - Categories of Work - Required Resources (Labor, Equipment, and Materials) - Resources Rates Determine Mix and Amount of Work Needed Apply Unit Cost to Work Quantities Repair Needs Replacement Needs Operations Needs Asset M&O Needs Work History Strategy Performance Impacts and Benefits Performance Criteria - Standards and Requirements - Performance Targets - Failure Patterns - Other Other Needs Stored Data Document Process Terminator FIGURE 1 Framework for signal M&O needs assessment. tice is usually to replace cabinets when controllers are replaced so as to make maintenance and problem diagnosis easier. Hence, VDOT s documented 10-year cabinet life can be shorter than elsewhere as well. Data on the characteristics of signal systems are normally obtained from signal inventories maintained by individual districts. The attributes generally available in signal inventories are unique asset identification, location, type, and installation year. Definition of M&O Work Four basic categories of M&O activities are captured in the signal needs assessment: PM, repair, replacement, and operating activities. There is no formula for determining the composition for maintenance actions since each system is unique (1). However, to ensure fair resource allocation, it is critical to clearly define each maintenance category and ensure that each is composed of a consistent list of activities across different districts in the state. Definitions of the four M&O categories and the activities they cover are presented next on the basis of field surveys, industry standards, and literature reviews. Data on work characteristics (frequencies, triggers, and resource requirements) are often obtained from field maintenance records or estimated by industry standards. PM Activities PM is defined as a set of checks and procedures to be performed at regularly scheduled intervals for the upkeep of traffic signal equipment; it includes inspection, record keeping, cleaning, and replacement based on the function and rated service life of the components (1). PM activities carried out by VDOT districts are as follows: 1. Inspect structures, foundations, mast arms, span (tether) wires, and poles; 2. Inspect control boxes, junction boxes, hand holes, and filters; 3. Inspect loops, signal heads, pedestrian signals, pole caps, hand-hole covers, and pedestrian buttons; 4. Inspect backplates, signal-based signs, and graffiti; 5. Inspect structure; 6. Check or test conflict monitors, load switches, auxiliary logic; 7. Check or clean cabinet (interior and exterior); 8. Change control box filters; 9. Check control box lamps; 10. Check and re-do caulk; 11. Check splices; 12. Replace conflict monitors that have up-to-date testing paperwork; 13. Check electromechanical control equipment (dial assemblies, cam assemblies, relays, flashers, disconnect switches, and terminal connections); 14. Check detectors (sensors, amplifiers, etc.); 15. Preemption test rail crossing (experts will carry out); 16. Test railroad pedestrian pushbutton devices; 17. Check if all paperwork matches system data; and 18. Replace batteries. Repair Activities Repair, or corrective maintenance, is defined as work required to restore a damaged or deteriorated asset to design, functionality, and capability (10). Repair actions are normally triggered by trouble calls regarding signal malfunction. Repair activities are as follows: 1. Repair equipment malfunction [lamp or lenses burnt out, local controller, master controller, detector sensor, amplifier, conflict monitor, flasher, time switch, load switch and relay, coordination unit, communication (interface, modem), and signal cables]; 2. Repair nonswitching signals, repair faulty pedestrian button, replace frayed cables;

14 14 Transportation Research Record 2128 TABLE 2 Components of Signal System Component Life (years) Component Life (years) Structure Mast arms 30 Pole (mast and strain) 30 Mast and pole foundations 30 Hardware Conflict monitor 7 Controller 7 Master controller 4 TS2 master controller 7 TS2 secondary controller 7 TS2 controller and cabinet 7 Priority control 2-channel 7 Priority control 4-channel 7 Cabinet 10 Modem 10 Signal heads 10 Flasher signal heads (1 section) 10 LED amber 5 LED red 5 LED green 5 Signal 12 in. section 10 Signal (1 section) 10 Signal (2 section) 10 Signal (3 section) 10 Signal (4 section) 10 Signal (5 section) 10 Signal (ped head) 10 Detection Loop detector unit 4 Camera 4 Camera processor (1) 4 Camera equipment (4) 4 Camera viewers 6 Camera monitor 6 TS2 loop amplifier (2-channel amp) 7 Priority control detector 7 Other TS2 BUS interface unit 7 TS2 power supply 7 Malfunction management unit 7 Uninterruptable power system 7 Opticom system (400) 4 Card rack (1) 4 Detector (4) 4 Card (1) 4 Two-way radio system (350) 4 Master antenna 4 Transceiver (1) 4 Luminaire 15 Central computer signal system Technician laptop 3 System desktop 3 System servers 4 System software 1 Support and maintenance 1 Clampmeter 1 Megger 1 Signs (avg. size 24 in. 36 in.) 10 Other box equipment 5 UPS backup battery pack 7 3. Clean graffiti, fix loose or lost bolts, fix missing covers; and 4. Repair knockdowns, damage that requires immediate safetyrelated repair (mast arm, strain pole, span wire or tether, pedestal, cabinet, signal heads). At VDOT, signal trouble calls are captured in various ways including 24-h signal repair hotlines; VA Traffic, a statewide incident input system; a transportation emergency operations center; five regional traffic operations centers; and 911 calls. Signal problems are reported by a wide range of sources including citizens; county, city, and state law enforcement; and VDOT s own operations staff who monitor traffic. Remote monitoring of trouble has been carried out in some areas of the state as crew time and workload permit. For example, one district has implemented the Management Information System for Transportation (MIST) to improve trouble detection through remote signal monitoring. With access to MIST, technicians can monitor signals and directly detect issues like detector malfunction. However, because of resource constraints, the current priority is on addressing major malfunctions and concerns raised by the public. Currently, the needs assessment does not account for the duration between when problems occur and when problems are reported. Replacement Activities Replacement maintenance is defined as replacement or complete restoration of assets that cannot be repaired. If the asset no longer functions, is obsolete, or does not conform to current federal or state mandates for design performance, it must be replaced or overhauled (10). The triggers for signal replacement include trouble calls, inspection outcomes, equipment obsolescence, standard changes, and system upgrades. At VDOT, a majority of replacement actions are triggered by known life-cycle dates, inspection results, or trouble calls on recurring malfunctions. Replacement maintenance activities are as follows: 1. Replace wiring with mast arms, 2. Upgrade detectors, 3. Replace lenses,

15 Chen, Henley, and Price Rebuild intersection to bring within life-cycle requirements, and 5. Replace foundations that have met life-cycle requirements. Operating Needs For signal systems, operations include all actions necessary for the proper functioning of a system (1). Operating actions include reviewing and updating timing plans, monitoring signal operations (with control software), updating system parameters, and dispatching crews to address operating concerns (11). Operating needs also include the following costs: 1. Electricity as metered, 2. Telephone, 3. Network communications, 4. Software licensing, and 5. Optimization. The framework for assessing needs requires four common parameters: number of years between the analysis year and the future needs year (t), number of signalized intersections in the inventory (A), estimated annual increase in the number of signalized intersections (B), and inflation rate (i). In the framework, each category is assessed individually. The common parameters are applied consistently across four maintenance categories. PM Needs Since PM is normally performed cyclically and the scope of the work is relatively consistent across different years, a unit cost based on historical expenditure is applied for assessing signal PM needs. As shown in Figure 2, the assessment draws on the per intersection PM cost for the analysis year and estimates the needs-year PM cost by multiplying this unit cost with the needs-year estimated number of intersections in addition to an inflation adjustment. Activities Not Captured in Framework Interviews with VDOT field staff indicated that additional activities performed by signal maintenance divisions are not captured under the needs assessment framework. However, such activities, mainly related to the installation of new signals, are normally funded through construction projects in VDOT and hence will not be included in the M&O needs assessment framework. These activities include the following: 1. Inspection of completed work for acceptance of new signals, 2. Marking of cables, 3. Preconstruction meetings and inspections (external entities, especially those related to fire and rescue, red-light running cameras, and video surveillance operations), and 4. Planning of reviews. Infrastructure Performance Criteria To determine the amount and type of work needed to maintain the current signal operating condition and the associated costs, the processes described in the remainder of this section are applied. Repair Needs Repair needs are determined by the estimated repair cost together with estimated lane closure cost. Repair cost refers to the associated labor and equipment costs, whereas lane closure cost refers to the cost associated with setting up the work zones on the lanes where signal malfunctions occur and repair operations take place. Since not all signal repair operations require lane closing, a percentage is applied to the lane closure cost representing the fraction of repair operations that require lane closure. As shown in Figure 3, the analysis (base) year per intersection repair cost and lane closure cost are determined by a number of parameters including the average number of repairs performed per intersection, average hours spent per repair, average labor and equipment cost required per repair, and average lane closure cost. Data on these factors are primarily obtained from signal maintenance records and VDOT s expenditure-tracking systems. The lane closure unit cost is determined on the basis of costs realized for various VDOT traffic signal projects. Currently, costs for primary and secondary road lane closure average $150/h, which covers labor and material for setting up and taking down signs, cones, barrels, and other measures needed for the lane closure operation. Base Year PM Cost per Intersection (P) Estimated Number of Intersections in t Years: (A + t x B) Estimated Inflation for Year t (1+i) t P X (A + t x B) X (1+i) t Estimated PM Needs in t Years (PMNt) Parameters Symbol Global # of years into the needs year t PM Specific # of intersections in the inventory (by system or by county) Annual increase of signalized intersections Annual inflation Base year PM cost per intersection A B i P FIGURE 2 PM needs assessment process.

16 16 Transportation Research Record 2128 Avg. Repair Cost/Hour (#Techs X TechRt + #Trucks X TruckRt) Avg. Repair Time (Hour): [(RT+TT)/60] Lane Closure Cost/Hour: (LcC) Avg. Repair Time (Hour): [(RT+TT)/60] [(RT+TT)/60] X (#Techs X TechRt + # Trucks X Truck Rt) [(RT+TT)/60] X LcC Avg. Cost Per Repair Avg. # of Repairs Per Year Per Intersection (#R) Avg. Lane Closure Cost/Repair Avg. # of Repairs Per Year Per Intersection (#R) Avg. % of Repairs That Close Lanes (LcR) Avg. Cost Per Repair X (#R) Avg. Lane Closure Cost/Repair X (#R) X (LcR) Global Parameters Symbol # of years into the needs year t # of intersections in the inventory (by system or by county) Annual increase of signalized intersections A B Annual inflation i Avg. Repair Cost Per Intersection Avg. Lane Closure Cost Per Intersection Total Repair & Lane Closure Costs Per Intersection Per Year (RPRN) Estimated Number of Intersections in t Years: (A + t X B) Estimated Inflation in t Years: (1+i) t RPR X(A + txb) X(1+i) t Estimated Repair Needs in t Years (RPRNt) Repair Specific Parameters Symbol Avg. # of Repair Trips Per Year Per Intersection #R Avg. Mins Spent Per Repair RT Avg. Round Trip Time in Mins TT Avg. # of Signal Techs Per Repair # Techs Avg. Tech Hrly Rate (incld. Benefits) TechRt Avg. # of Bucket Trucks Per Repair #Trucks Avg. Truck Hrly Rate TruckRt Lane Closure Cost/Hour LcC Avg. % of Repair Trips That Close Lanes LcR FIGURE 3 Repair needs assessment process. Specifically, repair and lane closure costs are estimated through the following calculations: base year per intersection = avg. intersection repairs per year * repair cost avg. repair time per repair trip * (labor cost per hour * number of technicians per repair + equipment cost per hour * number of equipments per repair) base year per intersection = avg. lane closure cost per hour * lane closure cost hours spent per repair operation * avg. intersection repairs per year * % of repairs requiring lane closure Combining these two costs results in the base year repair needs per intersection (RPRN), which are then multiplied by the needs year estimated number of intersections and adjusted by using the inflation factor to reach the estimated repair needs in year t (RPRNt). Replacement Needs Replacement needs are estimated at the signal component level and assessed with a life-cycle approach. Three main factors determine the level of replacement needs: The average replacement rate of each component is defined as 1 over its remaining service life expected (1/RLEn); component life expectancy (LEn) (Table 2) is defined as LEn less years in service (or age). With replacement rate defined as 1/RLE, a straightline deterioration pattern is assumed. Specifically, A replacement rate of 1 implies that the component s expected life will end during the needs year and is due to be replaced; A replacement rate that is zero or negative indicates that the component has served more than its expected life and a full replacement is needed; and A positive replacement rate of less than 1 implies that the component has two or more years of service life remaining; replacement rates in this range can be applied in the assessment framework as a contingency to account for replacement needs of components that fail ahead of their expected lives; longer remaining life means that a lower replacement rate is to be applied to the contingency; Component quantity per intersection (Qn/A) at each replacement rate is calculated as the quantity of each component that has the same number of years of service life remaining divided by the total number of signals in the inventory; and Component replacement cost at installation (RPCn) includes labor, material, and equipment costs required to replace the component. As shown in Figure 4, replacement needs are determined through the following formulas: base year per intersection = replacement cost of component n * replacement needs for Σ m 1 (per intersection quantity of comcomponent n (RPNn) ponent n at replacement rate m * replacement rate m) base year per intersection= Σ n 1RPNn replacement needs for all components (RPN)

17 Chen, Henley, and Price 17 Replacement Rate m of Component n: (1/RLEn,m) Component n Replacement Needs Per Intersection: (RPNn) (RLEn)=Asset Remaining Life in Years Per Intersection Quantity of Component n at Replacement Rate m: (Qn,m/A) Replacement Cost for n: RPCn m (Qn,m 1 A) (1 RLEn,m) RPCn Sum(RPN1+RPN2+...RPNn) Per Intersection Replacement Needs for All Components (1~n): (RPN) Estimated Number of Intersections in t Years: (A + t X B) RPN X (A+t x B) X (1+i) t Estimate d RPN in t Years (RPNt) Including material, and installation equipment and labor costs Estimated Inflation in t Years: (1+i) t Global Parameters Symbol Replacement Specific Parameters Symbol # of years into the needs year t #th signal component n # of intersections in the inventory (by system or by county) A Replacement cost of component n RPCn Annual increase of signalized intersections B Quantity of component n Qn Annual inflation i Remaining service life of component n RLEn FIGURE 4 Replacement needs assessment process. Replacement needs in year t (RPNt) are then estimated as the product of base year per intersection needs (RPN) and the estimated number of intersections in the year t inventory with adjustment for inflation. Replacement needs assessed through this method are sensitive to the LEn (or replacement rate) factor. Specifically, the assessment will likely generate more robust estimates for items with finite lives such as electronic components, which, when they fail, must be replaced or in limited cases can be repaired. However, structural components such as poles, cabinets, junction boxes, and mast arms are often able to provide service beyond their rated lives. Even though there is a higher risk of failure to extend the use of structural components beyond their rated lives, it is not an uncommon practice, especially when there are resource constraints. Conversely, it is not unusual for components to fail or be replaced before the end of their expected lives because of product deficiencies, replacement or upgrade of associated equipment, or external impacts such as lighting, power surges, or accidents. Hence, different inventory compositions and different patterns of external impacts result in different durability performance of agencies signal equipment. Ideally, life expectancy data should be adjusted to account for the foregoing factors in the needs assessment. A sensitivity analysis can be performed to determine different signal replacement needs based on various life expectancy assumptions. Operating Needs Operating needs (OPN) are determined by metered electric cost, telephone bills, network communications cost, annual software licensing fees, and optimization costs. As shown in Figure 5, the base year per intersection operating needs are estimated as the sum of all these costs divided by the number of intersections in the inventory. The operating needs in year t are then estimated as the product of the base-year per-intersection needs and the estimated number of intersections in year t and adjusted for inflation. Miscellaneous Needs Factors not included in any of the first four needs categories should be captured as miscellaneous needs to provide room for improving the accuracy of needs estimates. At VDOT, this category currently includes payments to localities for signal systems that localities are maintaining on behalf of VDOT. In addition, the framework is being reviewed annually by field staff to ensure that additional factors or updated parameter values, if any, are captured in the framework to better reflect the latest signal M&O practices. OPPORTUNITIES FOR PERFORMANCE-BASED ASSESSMENT The framework described has potential for incorporating a number of performance measures into the needs assessment process for signals. In general, signal system goals and objectives focus on two major areas: (a) performance of system equipment in terms of functional reliability and (b) level of service provided to the end users in terms of throughput and safety (11). In addition, from an asset management perspective, agency goals and objectives include maximizing asset service life and minimizing asset life-cycle costs. Key performance measures to evaluate signal systems against these goals and objectives include the following: Intersection vehicle hours of delay (level of service), Number of crashes at signalized intersections (level of service), Mean time to repair (MTTR) (level of service),

18 18 Transportation Research Record 2128 Electric Metered Cost: (EC X 12 X A) Average Phone Bill: (PhC X 12 X A) Network Communication Bills (Ntwk) Annual Software Licensing Fee (SftLcs) Sum all Total Annual Operating Cost Divide by A Operating Cost Per Intersection (OPC) Estimated Number of Intersections in t Years: (A + t*b) OPC X (A+t x B) X (1+i) t Estimated OPN in t Years (OPNt) Optimization Cost (Opt) Other Costs (Othr) Estimated Inflation in t Years: (1+i) t Global Parameters Symbol Operations Specific Parameters Symbol # of years into the needs year t Monthly electric metered cost per intersection EC # of intersections in the inventory (by system or by county) A Ave. monthly phone bill per intersection PhC Annual increase of signalized intersections B Private network communications bills Ntwk Annual inflation i Annual software licensing fee SftLcs Optimization cost Opt Other cost Othr FIGURE 5 Operating needs assessment process. Signal component mean time between failures (MTBF) (functional reliability), Signal component life expectancy LEn (life-cycle cost), and Life-cycle maintenance costs (life-cycle cost). As summarized in Table 3, through the highlighted parameters, the framework allows linking funding levels to performance outcomes for the foregoing measures except delays and crashes. It provides an analysis framework for users to compare various needs scenarios, each with different performance targets. The framework also provides a structure for users to better understand and articulate the performance consequences or costs of different maintenance investment strategies. MTTR and Needs Parameter Average Round-Trip Time MTTR represents the repair crew s average response time to address signal repair needs. MTTR can be estimated as half of the average round-trip time (ARTT) (for repairs). Assuming an equal productivity level, improved repair response time (or reduced ARTT) implies a greater number of technicians, a denser distribution of repair stations, and hence a higher level of resources required, and vice versa. Analyzing resources and funding requirements under various ARTT assumptions will help clarify the different levels of repair responsiveness (or repair level of service) that agencies should expect from various repair allocation scenarios. MTBF and Needs Parameter Average Repair Trips per Signal MTBF measures the reliability of signals based on the frequency of repairs needed. A longer MTBF indicates a more reliable signal system. On average, MTBF equals 1/average repair trips (ART). A per-signal annual ART of 4 suggests an MTBF of 0.25 year, or 3 months. Adopting more aggressive PM programs will result in reduced demand for responsive repair work. It is expected that an increase in PM work would lead to one or all of the following: lower ART (or higher MTBF), lower repair needs, and fewer overall life-cycle maintenance needs. Tracking historical values for the PM share of maintenance needs, the MTBFs, and overall persignal life-cycle maintenance needs will allow agencies to make informed decisions on maintenance strategies. Needs Parameter Life Expectancy n Life expectancy n (LEn) measures the durability performance of signal components. The average values of LEn are normally rated by manufacturers but can be heavily affected by agencies M&O practices. The higher the value of LEn, the lower the replacement needs. By adjusting LEn values, users can estimate the funding needs required under different asset life-span scenarios. Monitoring LEn over time will allow agencies to stay informed on overall inventory status and adjust M&O strategies as appropriate to bring assets toward manufacturer-rated (or healthy) LEn values. Needs Parameter Remaining Life Expectancy n Remaining life expectancy n (RLEn) equals life expectancy (LEn) minus component age (or years in service). Average RLEn measures the remaining usefulness of a component in the inventory. Monitoring values of RLEn in relationship to M&O needs allows agencies to plan financially for smooth system turnover. Over time it provides possibilities for agencies to research and identify the optimal inventory state that requires minimum M&O funding.

19 Chen, Henley, and Price 19 TABLE 3 Needs Assessment Parameters and Potential Performance Measures Needs Assessment Parameters Symbol Potential PMs Preventive maintenance Base year PM cost per intersection P Share of PM (versus reactive M) Repair Avg. no. of repair trips per year per intersection #R Mean time between failures (MTBF) Avg. minutes spent per repair RT Avg. round trip time in minutes TT Mean time to repairs (MTTR) Avg. no. of signal techs per repair # Techs Avg. tech hourly rate (including benefits) TechRt Avg. no. of bucket trucks per repair # Trucks Avg. truck hourly rate TruckRt Lane closure cost/h LcC Avg. % of repair trips that close lanes LcR Replacement #th signal component n Replacement cost of component n RPCn Quantity of component n Qn Component n life expectancy LEn Component life expectancy (LEn) Remaining service life (LEn years in service) RLEn Avg. remaining service life (RLEn) Operating needs Monthly electric metered cost per intersection EC Avg. monthly phone bill per intersection PhC Private network communications bills Ntwk Annual software licensing fee SftLcs Optimization cost Opt Other cost Othr Additional needs category Payment to localities CONCLUSIONS With an inventory of more than 3,000 signalized intersections, VDOT faces great challenges to maintaining the condition and service level of its signal assets. Decisions on priority-ranking M&O needs and resource allocation are essential to the upkeep of signal system performance. To optimize trade-offs across investments in system operations and preservation, it is critical to have methodologies and tools that allow understanding of costs, benefits, and performance outcomes associated with alternative investments. Budgeting for the M&O of traffic signal assets is performed in many ways at various transportation agencies. However, little documentation is available detailing signal resource allocation practices. A framework for traffic signal needs assessment is presented that VDOT recently developed to support signal M&O funding allocation. The framework provides the potential for incorporating a number of performance measures into the needs assessment process, including MTBF, MTTR, signal component LEn, average asset RLEn, and share of PM needs in the overall maintenance cost. Through adjusting values on the relevant parameters, the needs assessment framework allows setting performance targets for key signal system performance measures and comparing resources required to meet various performance scenarios. REFERENCES 1. Giblin, J. M. Traffic Control System Operations: Installation, Management, and Maintenance. Institute of Transportation Engineers, Washington, D.C., Cambridge Systematics, Inc. Technical Memorandum: Signal Systems Asset Management State-of-the-Practice Review. FHWA, U.S. Department of Transportation, April Biennial Report on the Condition and Performance of Surface Infrastructure in the Commonwealth of Virginia. In 2007 Acts of the Virginia General Assembly, VDOT, Richmond, Sept. 2007, Chaps. 335 and Walter, H. K. NCHRP Synthesis of Highway Practice 245: Traffic Signal Control Systems Maintenance Management Practices. TRB, National Research Council, Washington, D.C., Report to the General Assembly on Asset Management Methodology and Asset Management Systematic Mechanisms. In 2006 Virginia General Assembly Appropriations Act, VDOT, Richmond, Dec. 2006, Items 444 A.4 and B1. 6. Asset Management Program Budget and Expenditure Report. Maintenance Planning Leadership Group, VDOT, Sept. 14, Paral, J. M. Identification of Operations Assets. Publication FHWA- HOP FHWA, U.S. Department of Transportation, Sept NYS Traffic Signals: Increasing New York State Market Awareness and Demand for Energy-Efficient Light-Emitting Diode Traffic Signals. Lighting Research Center, Rensselaer Polytechnic Institute, Troy, N.Y., Bronson, M., J. Sugar, V. Hall, and R. L. Therkelsen. Light Emitting Diode (LED) Traffic Signal Survey Results. California Energy Commission, Sacramento, Jan Virginia Department of Transportation Activity Code Manual Pocket Edition. VDOT, Richmond, May Harrison, D. F., D. Krechmer, and J. Strasser. Elements of a Comprehensive Signals Asset Management System. Publication FHWA-HOP FHWA, U.S. Department of Transportation, Dec The opinions and conclusions are the responsibility of the authors and do not necessarily reflect the official views of the Virginia Department of Transportation. The Traffic Signal Systems Committee sponsored publication of this paper.

20 Performance Measures for Railroad-Preempted Intersections Thomas M. Brennan, Jr., Christopher M. Day, Darcy M. Bullock, and James R. Sturdevant The Manual on Uniform Traffic Control Devices provides guidance on when traffic signals should be interconnected with railroad crossings. Various other manuals and reports deal with how the traffic signal preemption logic should be configured. Maintenance practices typically require technicians to perform routine checks to ensure that preemption circuitry is functioning, and agencies often log the preemption events as part of a system-monitoring procedure. However, no procedure or software allows a controller to tabulate performance measures on how effective the preemption logic is at clearing conflicting phases before a train s arrival. This study introduces the concept of using an event-based data collection system to monitor the railroad preemption input, vehicle detectors, and phase indications to develop performance measures for evaluating the effectiveness of clearing movements that cross the tracks before train arrival. These techniques are applied to statistically compare the impact of steerable traffic signal indications at a location where the tracks are within 60 ft of the traffic signal stop bar. Based on a comparison of 700 preemptions before the signal heads were installed with 2,102 preemptions after signal heads were installed, no statistically significant difference was observed. The methods used to assess and design the appropriate minimum warning times, transfer times, and other interconnect design parameters for railroad highway grade crossings are outlined in manuals and reports published by various agencies (1 4). Although railroad highway grade crossing accident rates have steadily decreased from 1975, partly because of public education on railroad crossings, improved signaling systems, and abandonment of lines (5), there still remain a significant number of accidents each year that warrant evaluation of railroad signal preemption practices. Because of the infrequency of accident occurrences and the high number of potential contributing variables affecting accident frequency, analysis of accident data alone cannot satisfactorily characterize an intersection s preemption performance. Figure 1 shows a typical intersection where the railroad tracks are approximately 60 ft from the stop bar and the signal controller is interconnected with the railroad to receive notification that a train is T. M. Brennan, Jr., C. M. Day, and D. M. Bullock, Department of Civil Engineering, Purdue University, 550 Stadium Mall Drive, West Lafayette, IN J. R. Sturdevant, Indiana Department of Transportation, 100 North Senate Avenue, Room N925, Indianapolis, IN Corresponding author: D. M. Bullock, darcy@ purdue.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / approaching. When that notification is received by the signal, the current traffic signal phase times through a variable right-of-way transfer interval before going to a track-clearing green phase. The track-clearing green is designed to provide enough green time for queued traffic that is potentially in contact with the tracks to clear the railroad right-of-way before train arrival. Although there are detailed design procedures for calculating preemption parameters, there are no procedures or recommended practices for monitoring a traffic signal to confirm that the desired operation is being achieved. A methodology is proposed for providing performance measures that uses an event-based data logger (6 8) and vehicle detectors to confirm that the track-clearing green phase did in fact operate successfully. LITERATURE REVIEW In 1999 the Texas Transportation Institute instrumented an interconnected highway railroad intersection to monitor train arrivals (9, 10). One part of the research focused on the effects of variable railroad warning time (10). Variable warning times are one of the main contributors to the potentially hazardous preempt trap (11, 12), which can occur when the track-clearing green phase ends before the gates are fully horizontal and potentially allows a vehicle onto the tracks with no way to clear them without violating the traffic signal. Other modeling techniques were used to test the sensitivity of design parameters such as train speed and the impact on safety at a railroad crossing (13). INSTRUMENTED RAILROAD HIGHWAY GRADE CROSSING INTERSECTION The Indiana instrumented railroad highway grade crossing intersection (Figure 1) is configured to remotely log railroad interconnect data such as vehicle presence, traffic signal phase, start of active preempt, start of track-clearing green phase, end of track-clearing green, and end of active preempt. This intersection of US-36 SR-67 and Carroll Road is on the northeast side of Indianapolis just east of McCordsville. Figure 1 shows the detector layout (Figure 1a), signal phase layout (Figure 1b), and phase diagram (Figure 1c). As shown in Figure 1a, SA4 and SL7 represent the detectors SA-1, 2, 3 and SL-1, 2, 3 south of the railroad tracks, where the A in SA indicates the through lane closest to the centerline (one through lane southbound) and the L in SL represents the left-turn lane. Overlaps A and B (OLA and OLB) operate with a 3-s trailing green, yellow, and red with reference to Phases 4 and 7. Phases 4 and 7 20

21 Brennan, Day, Bullock, and Sturdevant 21 SA4 N SL7 SA 4,5,6,7 SL 4,5,6,7 Carroll Rd. SA 1,2,3 SL 1,2,3 EL 1,2,3,4 WL 1,2,3,4 camera view NL 1,2,3,4 NA 1,2,3,4 PTZ Camera US 36/SR 67 (a) N Approximate Visibility Region of Signal Heads Controlled by OLA and OLB φ4 φ7 Carroll Rd. φ7 Signal Head φ4 Signal Head OLA OLB Span Wire Poles (Typical) OLA & OLB Signal Head Span Wire φ1 φ6 φ6 φ3 OLD φ8 OLC (b) φ2 φ2 φ5 US 36/SR 67 OLB Signal Head OLA Signal Head Ø1 Ø2 OLD Ø3 Ø4 OLA = OLA TRAILING Ø4 = OLB TRAILING Ø7 Ø7 OLC = OLC TRAILING Ø8 Ø5 Ø6 OLB Ø8 = OLD TRAILING Ø3 FIGURE 1 Phasing and detection at instrumented railroad crossing: (a) detector layout and camera mount, (b) phase orientation and steerable signal head visibility zone, and (c) ring structure. (c)

22 22 Transportation Research Record 2128 Railroad Cabinet Detector Presence and Count Performance Measures Event Logger Phase Information Purdue Traffic Lab IP/VPN INTERNET RR Warning Preempt Circuit Digital Video for Ground Truthing FIGURE 2 Schematic data flow at instrumented railroad crossing. (IP internet protocol; VPN virtual private network.) control the signal heads on the north side of the span-wire installation (traffic signal heads hung from span wire, with single support poles located at four corners of intersection), and OLA and OLB control the heads on the south side of the span-wire installation. Overlaps C and D (OLC and OLD) operate with a similar 3-s trailing operation with reference to dummy Phases 8 and 3 to provide simultaneous phase termination of both compatible north- and southbound movements. Data are downloaded every 6 h from the control logger located at the site. Figure 2 conceptually identifies the flow of the data logged. EVALUATION OF STEERABLE SIGNAL HEADS The track is in close proximity to the intersection, and the Manual on Uniform Traffic Control Devices (MUTCD) offers the following guidance, under Section 8D.07 (2), for crossings with presignals: Consideration should be given to using visibility-limited signal faces (see Section 4A.02) at the intersection for the downstream signal faces that control the approach that is equipped with presignals. Therefore this intersection was a candidate for the application of steerable signal heads (14). These heads selectively limit the visibility of the OLA and OLB indication signals to vehicles in the region approximately bounded by the stop bar and the railroad tracks on the north side of the intersection (this trapezoidal region is marked in Figure 1b). A controller with event-based data logging was installed at the intersection to collect detector, phase, and railroad warning data (6). An IP-based pan-tilt zoom camera was installed with alarm recording capabilities to collect video data during preempt events. Data collection began on June 2, The following modifications were made to the intersection during the data collection period: July 1, Standard signal heads were removed and steerable signal heads installed at locations shown as OLA signal head and OLB signal head in Figure 1b but were not programmed to reduce the visibility of the OLA and OLB indications. July 16, Steerable signal heads (OLA signal head and OLB signal head, Figure 1b) were programmed to reduce the visibility of the OLA and OLB green and yellow indications to the trapezoidal region shown in Figure 1b. The red indications on these heads had no reduction in visibility zones programmed into the steerable head. August 28, Steerable signal heads were re-aimed to reduce the red indication visibility to the trapezoidal region shown in Figure 1b. The track-clearing green and amber indications remained as programmed on July 16. DATA COLLECTION AND PERFORMANCE MEASURES Performance measure plots were produced from the processed information obtained from the controller. Data collected on June 9, 2008, are used to demonstrate the types of performance measures that can be drawn from the data. Figure 3 shows the number of preempt events, the separation time between the start of these events, and the duration of active preempt occurrences. Preempt durations can imply the amount of delay a driver may encounter at this intersection during an active preempt and estimate an upper bound on the maximum queue length (15). Additional performance measures are defined in both graphical and equation form. Figure 4a shows the preempt events with reference to the preempt latency and duration of track-clearing green. Preempt latency is the time it takes for the traffic signal to start the track-clearing green once the signal is in active preempt. From Figure 4a at times 0:48:36 and 1:47:15, the preempt latency is

23 Brennan, Day, Bullock, and Sturdevant PREEMPT DURATION TIME FROM START OF ACTIVE PREEMPT (s) Mean Duration = s Standard Deviation = s Min Duration = 22.5 s Max Duration = s 0 0:00:00 1:30:00 3:00:00 4:30:00 6:00:00 7:30:00 9:00:00 10:30:00 12:00:00 13:30:00 15:00:00 16:30:00 18:00:00 19:30:00 21:00:00 22:30:00 0:00:00 START OF ACTIVE PREEMPT (TIME OF DAY) FIGURE 3 Railroad preemption events and durations on June 9, found to be 3 s and 8 s, respectively. Preempt latency is calculated as follows: P = t t l ctg, s p () 1 where P l = preempt latency, t ctg,s = start of track-clearing green, and t p = start of active preempt. For June 9, 2008, the preempt latency ranges from 3 to 21.3 s. The preempt latency represents the amount of time a vehicle could potentially be afoul of the tracks and waiting for a green light allowing it to proceed. Track-clearing green is the amount of time OLA and OLB are green during active preempt. From Figure 4a at times 0:48:36 and 1:47:15, the track-clearing green time is found to be 18.1 s for both events. Track-clearing green (g tc ) duration is defined as gtc = tctg, e tctg, s ( 2) where t ctg,e and t ctg,s are the end and start of the track-clearing phase, respectively. Under normal operating conditions g tc should not vary. However, variability may be introduced in cases where a gate down relay is used in the interconnect (11). The final component to be quantified is the margin between the last vehicle and the end of the track-clearing green. In Figure 4b the detector presence for SA4 or SL7 is superimposed on the active preempt duration, preempt latency, and duration of track-clearing green shown in Figure 4a. The track-clearing green margin time (m ctg ) is given by mctg = tctg, e t1 v () 3 where t ctg,e is the end of the track-clearing green and t lv is the time when the last vehicle is detected. This is the amount of time between detection of the last vehicle and the end of the track-clearing green. As shown in Figure 4b at times 16:52:50 and 22:05:14, the m ctg is observed to be 22 s and 13 s, respectively. At time 16:52:50 the track-clearing green margin was negative, meaning that either SA4 or SL7 did not clear before the end of the track-clearing green. At time 22:05:14, the track-clearing green margin was positive, meaning that a vehicle located on SA4 and SL7 had cleared during the track-clearing green. Tabulating negative track-clearing green margin time occurrences provides a measure of potential conflicts between an oncoming train and vehicles located near the track. Figure 5 shows data collected June 9, 2008, and it is refined to reflect only those times when an active preempt occurs with vehicle presence detected on either SA4 or SL7. Figure 6a and Figure 6b respectively show those times when SA4 or SL7 is occupied. From

24 24 Transportation Research Record Mean Preempt Latency = 7.52 s S.D. Preempt Latency = 5.16 Min Preempt Latency = 3 s Max Preempt Latency = 21.2 s Track Clear Green Duration = 18.1 s TRACK CLEAR PHASE IS GREEN, (ii) PREEMPT LATENCY, (i) PREEMPT DURATION, (iii) 0 0:48:36 1:47:15 4:22:30 7:02:07 10:22:10 13:25:58 15:39:15 16:52:50 16:56:48 17:03:29 17:11:53 17:16:03 17:17:37 17:22:13 18:18:33 20:18:23 21:29:08 21:33:18 21:46:18 21:49:55 21:56:16 22:05:14 22:15:03 TIME FROM START OF ACTIVE PREEMPT (s) 22:21:38 22:36:03 22:45:26 23:27:17 23:33:13 23:50:46 23:58:46 START OF ACTIVE PREEMPT (TIME OF DAY) (a) 100 TIME FROM START OF ACTIVE PREEMPT (s) TRACK CLEAR PHASE IS GREEN, (ii) PREEMPT LATENCY, (i) END OF ACTIVE PREEMPT, (iii) DETECTOR PRESENCE, (iv) 0 0:48:36 1:47:15 4:22:30 7:02:07 10:22:10 13:25:58 15:39:15 16:52:50 16:56:48 17:03:29 17:11:53 17:16:03 17:17:37 17:22:13 18:18:33 20:18:23 21:29:08 21:33:18 21:46:18 21:49:55 21:56:16 22:05:14 22:15:03 22:21:38 22:36:03 22:45:26 23:27:17 23:33:13 23:50:46 23:58:46 START OF ACTIVE PREEMPT (TIME OF DAY) (b) FIGURE 4 Track phase and detector status during train preemption, June 9, 2008: (a) time from start of preempt to start of track-clearing green (i ), to end of track-clearing green (ii), and to end of active preempt (iii) and (b) time from start of preempt to start of track-clearing green (i ), to end of track-clearing green (ii), and to end of active preempt (iii) with detector presence on SA4 or SL7 (iv).

25 Brennan, Day, Bullock, and Sturdevant TRACK CLEAR PHASE IS GREEN PREEMPT LATENCY END OF ACTIVE PREEMPT DETECTOR PRESENCE 0 10:08:30 12:17:52 13:54:14 15:39:15 16:45:35 TIME FROM START OF ACTIVE PREEMPT (s) 16:52:50 16:53:44 16:56:48 16:58:20 17:10:37 17:11:53 17:13:25 17:16:03 17:16:41 17:17:37 17:19:12 17:22:13 20:23:23 21:29:08 21:33:18 21:36:37 21:46:18 21:47:42 21:49:55 21:56:16 22:05:14 22:27:50 23:33:13 START OF ACTIVE PREEMPT (TIME OF DAY) FIGURE 5 Track-clearing phase timing only for preemptions with detector presence on SA4 or SL7, June 9, Figure 6a and Figure 6b the track-clearing green margin times for SA4 and SL7 were derived separately by applying Equation 3. To explain these concepts further, Figure 6a has callouts a through f, which are used to better describe the types of observations occurring at the site with reference to SA4: Callout a. Detector becomes active during transition to trackclearing green, and vehicles clear SA4 detectors during track-clearing green. In this situation an application of Equation 3 would result in a positive value for m ctg, which equals about 10 s. Callout b. Vehicle is present during preempt activation and departs before transition to track-clearing green is complete. Callout c. Preempt latency (P l ) is approximately 21 s. The first of two vehicles is detected on SA4 during the track-clearing green and clears the detector before the end of the track-clearing green. A second vehicle is detected on SA4 after the track-clearing green has ended, which would result in m ctg = 34 s when Equation 3 is applied a situation that might be addressed by using the gate-down relay to extend the track-clearing green interval until the gates are down (11). Callout d. After the track-clearing green has ended, a vehicle is detected toward the end of active preempt. In this situation the gates may be going up and a vehicle has crossed the tracks before the end of active preempt. Callout e. Either one or multiple vehicles queue over SA4 but are fully cleared before the end of the track-clearing green phase. Callout f. Presence on SA4 clears before the end of the trackclearing green, but toward the end of active preempt a vehicle is detected. As described in Callout d, a vehicle may have crossed the tracks as the gates were going up. Outlined in Figure 6b is Callout a, which is used to better describe the types of observations occurring at the site with reference to SL7. In Callout a at SL7, a vehicle presence is detected during the trackclearing green. One or more vehicles do not clear SL7 during the track-clearing green and do not clear until after the end of active preempt. This situation is more problematic than that for SA4 because the vehicles do not get an opportunity to clear SL7 once the rightof-way transfer is complete. This is another case that could likely be addressed by the use of a gate-down relay. PERFORMANCE MEASURES WITH SYNCHRONIZED VIDEO Video recordings were synchronized with the controller data to ground-truth the test bed site. The preempt event at 9:14:24 in Figure 7 shows a typical occurrence in which the vehicle clears the tracks during the track-clearing green. Figure 7 is synchronized with Figure 8, and the respective callouts in Figure 7 (a), (b), (d), (e), ( f ), and (g) correspond to specific frames in Figure 8. The callouts (c) and (h) are not shown in Figure 7 because the start of the railroad active warning system and the train arrival are not logged on to the controller at this time. With reference to Figure 8, (a) preempt call is received while a vehicle is present on SL7; (b) trackclearing green begins after a preempt latency of 8 s; (c) flashing railroad warning lights are observed in the video; (d) SL7 presence is dropped as the vehicle is observed clearing the area; (e) SA4 becomes occupied while the gates are observed going down; ( f ) SA4 clears while the gates are going down and before the end of the track-clearing green; (g) track-clearing green ends, the gates are

26 26 Transportation Research Record DETECTOR PRESENCE TRACK CLEAR PHASE IS GREEN PREEMPT LATENCY END OF ACTIVE PREEMPT 0 10:08:30 12:17:52 13:54:14 15:39:15 16:45:35 16:52:50 16:53:44 16:56:48 16:58:20 17:10:37 17:11:53 17:13:25 17:16:03 17:16:41 17:17:37 17:19:12 17:22:13 20:23:23 21:29:08 21:33:18 21:36:37 TIME FROM START OF ACTIVE PREEMPT (s) 21:46:18 21:47:42 21:49:55 21:56:16 22:05:14 22:27:50 23:33:13 a b c d e f START OF ACTIVE PREEMPT (TIME OF DAY) (a) TIME FROM START OF ACTIVE PREEMPT (s) DETECTOR PRESENCE TRACK CLEAR PHASE IS GREEN PREEMPT LATENCY END OF ACTIVE PREEMPT 0 10:08:30 12:17:52 13:54:14 15:39:15 16:45:35 16:52:50 16:53:44 16:56:48 16:58:20 17:10:37 17:11:53 17:13:25 17:16:03 17:16:41 17:17:37 17:19:12 17:22:13 20:23:23 21:29:08 21:33:18 21:36:37 21:46:18 21:47:42 21:49:55 21:56:16 22:05:14 22:27:50 23:33:13 a START OF ACTIVE PREEMPT (TIME OF DAY) (b) FIGURE 6 Track-clearing phase timing, June 9, 2008: (a) track-clearing phase timing only for preempts with detector presence on SA4 and (b) track-clearing phase timing only for preempts with detector presence on SL7.

27 Brennan, Day, Bullock, and Sturdevant TRACK CLEAR PHASE IS GREEN TIME FROM START OF ACTIVE PREEMPT (s) (d) (b) (a) (f) (g) (e) PREEMPT LATENCY END OF ACTIVE PREEMPT DETECTOR PRESENCE 0 0:15:49 6:00:52 9:14:24 10:31:18 16:53:15 20:21:57 20:30:53 20:49:03 23:15:39 23:57:09 START OF ACTIVE PREEMPT (TIME OF DAY) FIGURE 7 Detector presence for SA4 or SL7, July 23, observed to be down, and no vehicles are present on the tracks; and (h) train is observed arriving at the crossing. The preempt event at 12:08:58 in Figure 9 shows a situation in which the track-clearing green does not successfully clear the crossing. Figure 9 is synchronized with Figure 10, and the respective callouts for Figure 9 (a), (b), (e), ( f ), and (h) correspond to specific frames in Figure 10. The callouts for (c), (d), and (g) are not shown in Figure 9 because the start of the active warning system and the train arrival are not logged on to the controller at this time. In this situation the vehicle would have benefited from an extended green time on the track-clearing green interval until the gate was down. With reference to Figure 10, (a) preempt call is received while no vehicle is detected on SL7 or SA4; (b) track-clearing green begins after preempt latency of 3 s; (c) flashing railroad warning lights are observed in the video; (d) gates are observed going down; (e) SA4 becomes occupied and at the same time the gates are observed going down; ( f ) track-clearing green ends while SA4 is occupied; (g) gates are observed to be completely down with SA4 still occupied; and (h) train is observed in the right-hand side of the frame just as SA4 clears. The train does pass unhindered, but had the truck been a larger vehicle or been hauling a trailer the result could have been a collision. AGGREGATED PERFORMANCE MEASURES The number of times a vehicle is detected during an active preempt along with its relative position with reference to SA4 or SL7 are shown in Tables 1 through 4. The four tables are shown separately to distinguish between the data collected before and after the placement of steerable optically programmable heads on OLA and OLB. It was hypothesized that the new signal heads would reduce the frequency of vehicles detected at the end of the track-clearing green (Column 5, all tables). A binomial comparison of three different time periods of active data logging is shown in Table 5. CONCLUSION AND FUTURE RESEARCH The results of a binomial comparison of the proportion of vehicles observed failing to clear during the track-clearing green with the total railroad preempt events before steerable heads (Figure 1b) were installed, after steerable heads were installed but not programmed, after steerable heads were installed and programmed, and after programmed steerable heads were re-aimed are reflected in Table 5. Although a reduction in vehicles failing to clear during the trackclearing green was expected, a slight increase was observed. However, statistical tests indicated no statistically significant differences. Consequently, there is no statistical evidence to reject the null hypothesis that the steerable signal heads improved the vehicle clearance at SA4 or SA7 (post-track detectors) during an active preempt. Further research is under way to extend the data collection infrastructure shown in Figure 2 to include the gate-down relay and island circuit from the railroad. This extension will allow evaluation of the performance of the gate-down logic (11) and measure separation time without visual inspection of the video frames (Figure 10) to determine when the train arrives. ACKNOWLEDGMENTS This work was supported by the Joint Transportation Research Program administered by the Indiana Department of Transportation and Purdue University. The controller hardware used in this research was the ASC/3 controller manufactured by Econolite Control Products, Inc.

28 28 Transportation Research Record 2128 (a) (b) (c) (d) (e) (f) (g) (h) FIGURE 8 Track-clearing green sequence vehicle clears before track-clearing green ends, July 23, 2008: (a) preempt received, SL7 occupied (t 09:14:24); (b) start of track-clearing green (t 09:14:32); (c) flash starts, SL7 occupied (t 09:14:33); (d) SL7 clears (t 09:14:38); (e) SA4 occupied, gates going down (t 09:14:40); (f ) SA4 clears (t 09:14:42); (g) track-clearing green ends (t 09:14:51); and (h) train arrives, track is clear (t 09:14:57).

29 Brennan, Day, Bullock, and Sturdevant TIME FROM START OF ACTIVE PREEMPT (s) (h) (f) (b) (a) (e) DETECTOR PRESENCE TRACK CLEAR PHASE IS GREEN PREEMPT LATENCY END OF ACTIVE PREEMPT 0 0:00:51 2:31:45 5:29:08 6:29:54 8:34:27 10:18:42 11:58:58 12:08:58 13:51:33 15:05:25 16:03:43 18:51:34 19:33:03 FIGURE 9 Track-clearing green sequence for SA4, July 25, START OF ACTIVE PREEMPT (TIME OF DAY) (a) (b) (c) (d) FIGURE 10 Track-clearing green sequence vehicle clears after track-clearing green ends, July 25, 2008: (a) preempt received, no presence (t 12:08:58); (b) start of track-clearing green (t 12:09:01); (c) flash starts, SA4 clear (t 12:09:10); (d ) gates going down (t 12:09:14). (continued on next page)

30 30 Transportation Research Record 2128 (e) (f) (g) (h) FIGURE 10 (continued) Track-clearing green sequence vehicle clears after track-clearing green ends, July 25, 2008: (e) SA4 occupied, gates going down (t 12:09:18); (f ) track-clearing green ends (t 12:09:19); (g) gates down, vehicle not cleared (t 12:09:27); and (h) train arrives, SA4 is clear (t 12:09:38). TABLE 1 Track-Clearance Performance with Standard Signal Heads Detector Presence During Preempt Detector Presence During Preempt Detector Presence During Preempt (SA4 and SL7) (SA4) (SL7) (5) (6) (8) (9) (12) Between After Between After (11) After 28 Days Observed End of Track- End of Track- Between Track- Track- Clear Track- Clear End of Clear (3) (4) Clear Green (7) Clear Green (10) Track- Green Total Only Green and Only Green and Only Clear and (2) Preempt During and Before During and Before During Green Before Total Events Track- End of End of Track- End of End of Track- and End End of (1) Preempt with Clear Active Active Clear Active Active Clear of Active Active Date Events Detection Green Preempt Preempt Green Preempt Preempt Green Preempt Preempt 6/3/ /4/ /5/ /6/ /7/ /8/ /9/ /10/ /11/ /12/ /13/ /14/ /15/ /16/ /17/ (continued)

31 Brennan, Day, Bullock, and Sturdevant 31 TABLE 1 (continued) Track-Clearance Performance with Standard Signal Heads Detector Presence During Preempt Detector Presence During Preempt Detector Presence During Preempt (SA4 and SL7) (SA4) (SL7) (5) (6) (8) (9) (12) Between After Between After (11) After 28 Days Observed End of Track- End of Track- Between Track- Track- Clear Track- Clear End of Clear (3) (4) Clear Green (7) Clear Green (10) Track- Green Total Only Green and Only Green and Only Clear and (2) Preempt During and Before During and Before During Green Before Total Events Track- End of End of Track- End of End of Track- and End End of (1) Preempt with Clear Active Active Clear Active Active Clear of Active Active Date Events Detection Green Preempt Preempt Green Preempt Preempt Green Preempt Preempt 6/18/ /19/ /20/ /21/ /22/ /23/ /24/ /25/ /26/ /27/ /28/ /29/ /30/ Total TABLE 2 Track-Clearance Performance with Unprogrammed Signal Heads Detector Presence During Preempt Detector Presence During Preempt Detector Presence During Preempt (SA4 and SL7) (SA4) (SL7) (5) (6) (8) (9) (11) (12) Between After Between After Between After 14 Days Observed End of Track- End of Track- End of Track- Track- Clear Track- Clear Track- Clear (3) (4) Clear Green (7) Clear Green (10) Clear Green Total Only Green and Only Green and Only Green and (2) Preempt During and Before During and Before During and Before Total Events Track- End of End of Track- End of End of Track- End of End of (1) Preempt with Clear Active Active Clear Active Active Clear Active Active Date Events Detection Green Preempt Preempt Green Preempt Preempt Green Preempt Preempt 7/2/ /3/ /4/ /5/ /6/ /7/ /8/ /9/ /10/ /11/ /12/ /13/ /14/ /15/ Total

32 32 Transportation Research Record 2128 TABLE 3 Track-Clearance Performance with Programmed Steerable Signal Heads Detector Presence During Preempt Detector Presence During Preempt Detector Presence During Preempt (SA4 and SL7) (SA4) (SL7) (5) (6) (8) (9) (11) (12) Between After Between After Between After 41 Days Observed End of Track- End of Track- End of Track- Track- Clear Track- Clear Track- Clear (3) (4) Clear Green (7) Clear Green (10) Clear Green Total Only Green and Only Green and Only Green and (2) Preempt During and Before During and Before During and Before Total Events Track- End of End of Track- End of End of Track- End of End of (1) Preempt with Clear Active Active Clear Active Active Clear Active Active Date Events Detection Green Preempt Preempt Green Preempt Preempt Green Preempt Preempt 7/17/ /18/ /19/ /20/ /21/ /22/ /23/ /24/ /25/ /26/ /27/ /28/ /29/ /30/ /31/ /1/ /2/ /3/ /4/ /5/ /6/ /7/ /8/ /9/ /10/ /11/ /12/ /13/ /14/ /15/ /16/ /17/ /18/ /19/ /20/ /21/2008 8/22/ /23/ /24/ /25/ /26/ /27/ Total 1, NOTE: Data were not collected on August 21 because of system maintenance.

33 Brennan, Day, Bullock, and Sturdevant 33 TABLE 4 Track-Clearance Performance with Re-Aimed Steerable Signal Heads Detector Presence During Preempt Detector Presence During Preempt Detector Presence During Preempt (SA4 and SL7) (SA4) (SL7) (5) (6) (8) (9) (11) (12) Between After Between After Between After 33 Days Observed End of Track- End of Track- End of Track- Track- Clear Track- Clear Track- Clear (3) (4) Clear Green (7) Clear Green (10) Clear Green Total Only Green and Only Green and Only Green and (2) Preempt During and Before During and Before During and Before Total Events Track- End of End of Track- End of End of Track- End of End of (1) Preempt with Clear Active Active Clear Active Active Clear Active Active Date Events Detection Green Preempt Preempt Green Preempt Preempt Green Preempt Preempt 8/29/ /30/ /31/ /1/ /2/ /3/ /4/ /5/ /6/ /7/ /8/ /9/ /10/ /11/ /12/ /13/ /14/ /15/ /16/ /17/ /18/ /19/ /20/ /21/ /22/ /23/ /24/ /25/ /26/ /27/ /28/ /29/ /30/ Total

34 34 Transportation Research Record 2128 TABLE 5 Binomial Comparison of Aggregated Track-Clearance Performance Measures Detector Presence During Preempt (SA4 and SL7) Between End of Track-Clear Green and End of Active Preempt Dates Total Preempt Compared (2008) Description Events (4a) (4b) Z-Value, 95% Confidence Group (1) (2) (3) Count Percent (5) A B C D E 6/3 6/30 7/2 7/15 7/2 7/15 7/17 8/27 6/3 7/15 7/17 8/27 7/17 8/27 8/29 9/30 6/3 6/30 8/29 9/30 Standard signal heads Unprogrammed steerable signal heads Unprogrammed steerable signal heads Programmed steerable signal heads Standard and unprogrammed steerable signal heads Programmed steerable signal heads Programmed steerable signal heads Programmed steerable signal heads RE-AIMED 8/28 Standard signal heads Programmed steerable signal heads RE-AIMED 8/ , ,076 1, , no significant difference 0.64, no significant difference 1.35, no significant difference 0.06, no significant difference 1.29, no significant difference REFERENCES 1. NCHRP Synthesis of Highway Practice 271: Traffic Signal Operations Near Highway-Rail Grade Crossings. TRB, National Research Council, Washington, D.C., Manual on Uniform Traffic Control Devices. FHWA, U.S. Department of Transportation, Marshall, P., and W. Berg. Design Guidelines for Railroad Preemption at Signalized Intersections. ITE Journal, Vol. 67, No. 2, Alroth, W. A Proposed Recommended Practice: Preemption of Traffic Signals at or near Actively Protected Railroad Grade Crossings. ITE Journal, Vol. 65, No. 2, Mok, S., and I. Savage. Why Has Safety Improved at Rail-Highway Grade Crossing? Risk Analysis, Vol. 24, No. 4, 2005, pp Smaglik, E. J., A. Sharma, D. M. Bullock, J. R. Sturdevant, and G. Duncan. Event-Based Data Collection for Generating Actuated Controller Performance Measures. In Transportation Research Record: Journal of the Transportation Research Board, No. 2035, Transportation Research Board of the National Academies, Washington, D.C., 2007, pp Hubbard, S. M. L., D. M. Bullock, and C. M. Day. Opportunities to Leverage Existing Infrastructure to Integrate Real-Time Pedestrian Performance Measures into Traffic Signal System Infrastructure. Presented at 87th Annual Meeting of the Transportation Research Board, Washington, D.C., Day, C. M., E. J. Smaglik, D. M. Bullock, and J. R. Sturdevant. Quantitative Evaluation of Actuated Versus Nonactuated Coordinated Phases. In Transportation Research Record: Journal of the Transportation Research Board, No. 2080, Transportation Research Board of the National Academies, Washington, D.C., 2008, pp Translink Train Monitoring Project. Texas Transportation Institute, Texas A&M University, College Station Engelbrecht, R. The Effect of Variation in Railroad Warning Time on Traffic Signal Preemption. Presented at the Sixth International Symposium on Railroad-Highway Grade Crossing Research and Safety, Yohe, J. R., and T. Urbanik II. Advance Preempt with Gate-Down Confirmation: Solution for Preempt Trap. In Transportation Research Record: Journal of the Transportation Research Board, No. 2035, Transportation Research Board of the National Academies, Washington, D.C., 2007, pp Engelbrecht, R., S. Sunkari, T. Urbanik II, S. Venglar, and K. Balke. The Preempt Trap: How to Make Sure You Do Not Have One. Report Texas Transportation Institute, Texas A&M University, College Station, Cho, H., and L. R. Rilett. Modeling Signalized Intersections near Highway-Railroad Grade Crossings: Sensitity Analyses of Key Design Parameters. In Transportation Research Record: Journal of the Transportation Research Board, No. 1973, Transportation Research Board of the National Academies, Washington, D.C., 2006, pp Intelight Website. pdf. 15. Sharma, A., D. M. Bullock, and J. A. Bonneson. Input Output and Hybrid Techniques for Real-Time Prediction of Delay and Maximum Queue Length at Signalized Intersections. In Transportation Research Record: Journal of the Transportation Research Board, No. 2035, Transportation Research Board of the National Academies, Washington, D.C., 2007, pp The contents of this paper reflect the views of the authors, who are responsible for the facts and the accuracy of the data, and do not necessarily reflect the official views or policies of the National Academies, the American Association of State Highway and Transportation Officials, the Federal Highway Administration, or the Indiana Department of Transportation. These contents do not constitute a standard, specification, or regulation. The Traffic Signal Systems Committee sponsored publication of this paper.

35 Three-Dimensional Mapping of Inductive Loop Detector Sensitivity with Field Measurement Christopher M. Day, Thomas M. Brennan, Jr., Matthew L. Harding, Hiromal Premachandra, Allen Jacobs, Darcy M. Bullock, James V. Krogmeier, and James R. Sturdevant Inductance loops continue to be the most widely used sensing device for vehicle detection. Several different loop geometries are commonly used, but scant technical design literature quantifies their field of detection and sensitivity. Three-dimensional maps of loop response sensitivity based on field measurement with loops installed in asphalt are presented. Loop response was characterized for different metal objects at various heights from the pavement surface. Sensitivity maps were generated for 6-ft circular, 6-ft octagonal, 20-ft octagonal, and 20-ft quadrupole loops. From the field observations, it is concluded that the claims of increased vehicle sensitivity for quadrupole loops first published in the 1970s are inaccurate for most vehicles and quadrupole loops are in fact less sensitive than comparatively sized rectangular loops. Inductive loop detectors (ILDs) have been in use since the early 1960s and have remained the most common type of traffic sensor despite the emergence of competing technology. A great deal of work has been published on ILDs discussing their operation (1 3) and applications in performance measure calculation (4 6), freeway speed estimation (7), freeway travel time estimation (8, 9), and vehicle identification (10 12). There are few studies in the literature that describe how detection fields vary with geometry and sensitivity. Hamm and Woods (13) published a study in which the detection areas of loops were characterized at different sensitivity settings. Their focus was on factors such as pavement condition, depth of placement, and number of turns rather than on the geometry of the detection area. However, Hamm and Woods provide an example of how the detection zone of a particular loop geometry varies with different sensitivity settings. Kidarsa et al. (14) used a MATLAB simulation to characterize the response of ILDs to bicycles and to describe the shape of the detection zone in three dimensions. The methodology and data representation employed in the current study are similar to those used in C. M. Day, T. M. Brennan, Jr., M. L. Harding, H. Premachandra, and D. M. Bullock, School of Civil Engineering; and J. V. Krogmeier, School of Electrical and Computer Engineering, Purdue University, 550 Stadium Mall Drive, West Lafayette, IN A. Jacobs, Reno A&E, 4655 Aircenter Circle, Reno, NV J. R. Sturdevant, Indiana Department of Transportation, 100 North Senate Avenue, Room N925, Indianapolis, IN Corresponding author: D. M. Bullock, darcy@purdue.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / these previous studies, but the objective here is to document the loop response in three dimensions by using field data. THEORY OF LOOP RESPONSE TO METAL SURFACE The following analysis was adapted from the Traffic Detector Handbook (1). Loop sensitivity S is given by L L L S = = L L where L refers to a value of inductance, with L V and L NV referring to the loop inductance with (V) and without (NV) the presence of a vehicle above the loop, respectively. Loop sensitivity can also be stated in terms of the surface area of the metal vehicle body as well as the distance between the loop and the vehicle body. When a vehicle is brought into proximity with an inductive loop, it behaves like an air core transformer [a transformer in which the coils are wrapped around air (actually a hollow tube in an actual device) rather than ferrite], as shown in Figure 1. The surface of the vehicle undercarriage is modeled as a shorted loop. The sensitivity of such a system is given by 2 M S = LL Δ NV V L V where L L and L V are the self-inductance of the inductive loop detector and the vehicle body, respectively. The coupling or mutual inductance M between the loop and the vehicle body can be expressed as N A F M = μ 0 d L V L NV ( 2) () 3 where µ 0 = permeability of free space, N L = number of turns in loop, A V = surface area of vehicle parallel to loop surface area, F L = constant that adjusts for impact of nonuniform magnetic flux, and d = distance between vehicle and loop. () 1 35

36 36 Transportation Research Record 2128 Vehicle Body Induced Current l V A V L 1 M 21 Vehicle Loop Current Induced Current d Loop Detector Loop Current L 2 N L l L A L Magnetic Flux Loop Detector (a) (b) FIGURE 1 ILD vehicle interaction modeled as air core transformer: (a) electrical model and (b) physical model (d distance between vehicle and loop, A V surface area of vehicle parallel to loop service area, and N L number of turns in loop). L The self-inductance L L of the inductive loop detector is given by where A L is the surface area of the loop, L is the circumference of the loop, and the other terms are the same as in Equation 3. For the vehicle body, the self-inductance L V is given by L where F V is similar to the term in Equation 3, A V is the surface area of the vehicle as before, and V is the effective circumference of that area modeled as a shorted loop. A synthesis of the above three equations yields the following expression of loop sensitivity: S L V N A F = μ 2 0 l L L L AF V = μ 0 l V L V A llf AdF ( 4) = V L V L 6 2 L V () 5 This expression states the expected impact of numerous parameters on loop sensitivity. In particular, the inverse-square relationship between sensitivity and distance d is notable. The response of the loop should drop sharply as the distance of a metal object from the loop surface is increased. The response of the loop is also a function of the ratio of the surface areas of the loop (A L ) and of the vehicle body (A V ). When A L is larger than A V, the loop response is weaker because the region where currents can be induced is small. As A V increases, the sensitivity increases, within a practical limit. If an infinite plane of metal existed parallel to the loop surface, only a local region of it would experience the inductance and A V would be limited to that area. The reason for this response is that the influence of the loop decreases by an inverse square of the distance. One possible approximation for the maximum practical value of A V is A L. It is not the objective of this study to refine the model but rather to show that the loop response is expected to increase as a greater portion of the loop surface area is covered by the vehicle surface area (i.e., as A V /A L tends to unity). () Loop detectors do not directly measure ΔL/L but rather the more easily measured quantity Δf/f, the relative shift in the oscillating frequency of the loop. Depending on the loop amplifier, frequency shifts or period shifts may be measured. This quantity relates to ΔL/L by the following formula (14): ΔL 1 Δf S = = = L 2 f 1 2 where f V and f NV are the frequency with and without the presence of a vehicle above the loop. Figure 2 shows traces in ΔL/L values sampled over time as a vehicle passed over an octagonal ILD. The three traces correspond to various paths taken over the detector area by the vehicle. The strongest response is for the vehicle traveling along the centerline of the loop; this is where the maximum amount of detector surface area was covered by the vehicle. The peaks in the traces correspond to the points where the vehicle passed the center point of the loop, which would also correspond to the moment when the amount of loop area covered by the vehicle is maximized. This response is as predicted by Equation 6. METHODOLOGY ( fv fnv) f NV Figure 3a is a photograph of the test bed showing the four loop geometries examined in this study: L1, L2: 6-ft by 6-ft (1.8-m by 1.8-m) octagonal, L3, L4: 6-ft round, L5: 20-ft (6.1-m) octagonal, and L6: 20-ft quadrupole. ( 7) The loops were placed in recently set asphalt by conventional saw cuts. The pavement was less than a year old at the time of the study, with no water ponding or pavement damage known to have occurred. Figure 3b shows a wiring diagram of the test bed.

37 Day, Brennan, Harding, Premachandra, Jacobs, Bullock, Krogmeier, and Sturdevant Loop Sensitivity ΔL/L Vehicle traveling on loop centerline Vehicle traveling on loop edgeline 1 Vehicle traveling 2 ft outside of loop edgeline Number of Samples FIGURE 2 Example loop detector traces as vehicle passes over center of loop. L1 L2 L3 L4 L5 L6 (a) L1 L2 L3 L4 L5 L6 L1 L2 L3 L4 L5 L6 Lead-in terminals (b) FIGURE 3 Loop detector test bed: (a) photograph and (b) electrical diagram.

38 38 Transportation Research Record 2128 Two probes were used to measure ΔL/L values at various positions above the inductance loops, simulating the presence of a vehicle body at a fixed height above the pavement: Galvanized steel sheet 4 ft by 12 ft (1.2 m by 3.6 m) and Anodized aluminum sheet 4 ft by 12 ft. Different metal probes were used to examine the effects of metal type on the loop response. Different types of metals are known to elicit different responses from metal-detecting devices (15), particularly when ferrous and nonferrous metals are compared. The testing apparatus is illustrated in Figure 4. A wood frame was used to prevent the sheet from bending or warping. A gantry made 2" PVC 12"x12" Engineering Joist 2"x4" 8' 12' (a) 1.92' (b) FIGURE 4 Testing apparatus: (a) gantry used to control height during measurements and (b) test in progress with anodized aluminum sheet.

39 Day, Brennan, Harding, Premachandra, Jacobs, Bullock, Krogmeier, and Sturdevant Edge +12 Edge Line Edge Edge 60 Center Line Reference Point Edge 60 Center Horizontal Offset Edge 12 Edge Line Edge FIGURE Loop traces. of nonmetal components controlled the height of the probe. With the height adjusted, the probe was moved across transverse sections of the loop. Values of ΔL/L were recorded at 6-in. (15.2-cm) intervals along the transverse section lines starting from the center of the loop. This procedure was repeated at various heights starting from the pavement surface and moving upward in 6-in. increments. Several such traces were measured along different cross sections of the loop, as shown in Figure 5. For the 20-ft loops, traces were taken at the centerline, edge line, 60 in. (152.5 cm) inside the edge, 12 in. (30.5 cm) inside the edge, and 12 in. outside the edge. The 6-ft loops used the same traces, except of course for the 60-in. offset trace. Assuming symmetry along the longitudinal and center transverse axes of each loop, ΔL/L values were mapped for one quadrant and then mirrored across each axis. Test traces in the other quadrants validated the assumption of symmetry. In a typical field installation, loop amplifiers tune out persistent changes in sensitivity caused by introduction of a metal object (such as a parked vehicle) into the environment. The ΔL/L values were recorded with a Reno A&E loop amplifier, which was modified to not tune out the presence of the metal probe during measurement. The ΔL/L values were used to generate the plots of sensitivity presented in the next section. To give context to the ΔL/L values, the response of each loop type to different vehicles is presented in Table 1. The passenger car, although smaller than the pickup truck, was closer to the ground and drew a stronger response from the loops, except for the rectangular TABLE 1 Loop Detector Response for Different Vehicles and Loop Geometries ΔL/L Value Clearance 6-ft 6-ft 6-ft 20-ft 20-ft Vehicle Type (in.) Round Octagonal Quadrupole Rectangular Passenger car (Mitsubishi Gallant) Pickup truck (Chevrolet Silverado) Front-end loader >24 N/A N/A Two-axle dump truck >24 N/A N/A Bicycle (loop edge) On pavement Bicycle (loop center) On pavement

40 40 Transportation Research Record 2128 TABLE 2 Correspondence of Sensitivity Levels to L /L Values ΔL/L Reno A&E EDI Sensitivity (%) Sensitivity Level Level loop, for which the truck was slightly stronger. Table 2 gives the sensitivity levels of Reno A&E and Eberle Design Inc. detector cards. Both the car and the truck would have been detected at even the lowest sensitivity levels of both detector cards. Heavier vehicles such as the front end loader and dump truck had larger metal surface areas but were situated higher above the pavement. As a result, the reported ΔL/L-values were lower. Bicycles have the lowest sensitivity of all the vehicles tested. Bicycle wheels have a much smaller metal composition, but it is situated close to the pavement surface. They are virtually undetected at the center of the loop (except for the quadrupole loop, where a loop wire is present) and produce a weak response when situated over the loop edge. RESULTS OF FIELD MEASUREMENT Figures 6 and 7 show cross-sectional traces across the center of the loops, with traces shown for steel and aluminum sheets at 12, 24, and 36 in. from the pavement surface. Figure 6a shows traces for the quadrupole loop, Figure 6b for the rectangular loop, Figure 7a for the octagonal loop, and Figure 7b for the round loop. Figure 8 shows superimposed traces of the responses of all four loop geometries to the galvanized steel probe 12 in. above the pavement. Although variation in sensitivity was observed within the boundaries of the loops, there was virtually no difference in detector sensitivity outside the loop perimeters. As expected, a consistent decrease in loop response with increasing distance from the pavement was confirmed. Aluminum produced a slightly weaker response than steel, but the difference was only noticeable closer to the pavement. To document the low end of the sensitivity range, probe heights of 12 in. and above are shown. If probe heights of 6 in. are used, the peak ΔL/L values exceed or are similar to the values for the passenger car and pickup truck in Table 1. At heights greater than 12 in., there was little difference between aluminum and steel. The response in terms of ΔL/L was significantly lower than that of the passenger car and pickup truck (see Table 1), particularly 24 and 36 in. above the pavement. Across the center cross-sectional trace, the strongest response tended to be observed when the metal probe was positioned at the center of the loop, except for the quadrupole loop, where the response was strongest between the center and edge loop wires because at the center point, the amount of metal surface area directly above the loop was maximum. If a vertically aligned probe (such as a bicycle wheel) had been used, the peak would likely have been observed above the loop wires. Using a metal sheet approximates the footprint of a vehicle body, and this study confirms the strongest response when the vehicle is aligned directly above the loop. The octagonal and round loops responded similarly, with a notable variation being the greater difference between the response to aluminum and steel from the round loop (Figure 7). In general, the response of the octagonal loop was slightly stronger than that of the round loop (Figure 8). At 12 in. the rectangular loop was considerably less sensitive than the 6-ft loops, but as the height was increased, the difference in sensitivity became less prominent. At 36 in., the ΔL/L responses of rectangular, round, and octagonal loops were comparable. The quadrupole loop was found to be the least sensitive of all. The 20-ft loops exhibited a weaker response because the loop area A L was much larger than that of the 6-ft loops. If multiple 6-ft loops were used to form a wider detector area, the magnitude of their response would have been similar. Three-dimensional maps of loop sensitivity to the steel sheet 12 in. above the pavement are shown in Figures 9 and 10. Direction of vehicular travel noted in these graphs represents the direction in which the longer edge of the metal probe was aligned. The three-dimensional maps illustrate features of loop response consistent with what was observed in the cross sections. The response is strongest when the probe is centered above the open areas of the loops and decreases as it is moved away. As mentioned before, this is where the amount of loop surface area covered by metal is at its greatest. Among the different cross-sectional traces along the longitudinal axes of the 20-ft loops, there is not much difference in the sensitivity until the edge. As for the 6-ft loops, the three-dimensional images do not reveal radial symmetry. Instead, the shape of the curve is elongated along the direction of vehicular travel. This finding is a consequence of the shape of the metal probe. The probe covered almost the entire surface area of the loop when positioned at the north edge, whereas only about half of the loop was covered when positioned at the east edge. If a smaller, symmetric probe had been used radial symmetry would have been observed in these loops. However, the probe used in this study is intended to simulate a vehicle body. Figure 11 shows contour maps of the response of the four types of loops to the galvanized steel sheet 24 in. from the pavement. In these plots, the response is mapped in terms of detector sensitivity. The Reno A&E sensitivity levels were used because they are on the same order of magnitude as the variations among point measurements in these observations. Table 2 gives the ΔL/L values associated with each sensitivity level. In addition to the nine sensitivity levels, a line is shown to indicate where the loop response falls to zero, as calculated in the MATLAB model used to generate the contours. These plots indicate what regions 24 in. above the loop are detectable at various sensitivity levels with the test probe. The shapes of the plots are similar to the three-dimensional plots discussed previously, with the most responsive regions being at the center of the loop. The quadrupole loop is the least sensitive, being completely unable to detect the steel sheet at 24 in. at a sensitivity lower than 3. The sensing range drops off sharply with distance from the loop edges. The rectangular loop is more sensitive, and the detection ranges of the loop at each sensitivity level are larger than those of the quadrupole loop. The octagonal and round loops have similar contour plots. The widths of the detection ranges are greater than those of the rectangular loop.

41 Day, Brennan, Harding, Premachandra, Jacobs, Bullock, Krogmeier, and Sturdevant Steel at 12" 1.5 ΔL/L (%) Aluminum at 12" Similar response for steel and aluminum at 24" Aluminum at 36" (no response for steel) A A Centroid Horizontal Offset (in.) (a) (b) Steel at 12" Aluminum at 12" ΔL/L (%) Similar response for steel and aluminum at 24" Similar response for steel and aluminum at 36" B B Centroid Horizontal Offset (in.) (c) (d) FIGURE 6 Response of quadrupole and rectangular loops to different metals at various heights: (a) 20-ft quadrupole loop, (b) cross section A A, (c) 20-ft rectangular loop, and (d ) cross section B B.

42 42 Transportation Research Record Steel at 12" Aluminum at 12" ΔL/L (%) Similar response for steel and aluminum at 24" Similar response for steel and aluminum at 36" C (b) C Centroid Horizontal Offset (in.) (a) Steel at 12" ΔL/L (%) Aluminum at 12" Steel at 24" Aluminum at 24" Aluminum at 36" D (d) D Centroid Horizontal Offset (in.) (c) Steel at 36" FIGURE 7 Response of 6-ft octagonal and round loops to different metals at various heights: (a) 6-ft by 6-ft octagonal loop, (b) cross section C C, (c) 6-ft round loop, and (d) cross section D D.

43 Day, Brennan, Harding, Premachandra, Jacobs, Bullock, Krogmeier, and Sturdevant ft x 6 ft Octagonal 6 ft Round 20 ft Rectangular 20 ft Quadrupole ΔL/L (%) Centroid Horizontal Offset (in.) FIGURE 8 Comparison of loop response to galvanized steel 12 in. above pavement (6-ft loop traces represent sensitivity of single loop; multiple loops in series would have overall lower sensitivity). The 1976 patent (16) of the quadrupole loop claimed that the loop design would reduce false calls from adjacent lanes, provide better detection of small vehicles, and improve uniformity of loop sensitivity over the detection area. The patent provided a diagram of magnetic field strength across the cross section of the loop, as shown in Figure 12. The strength of the magnetic field was expected to be weakest above the center in the case of the rectangular loop (Figure 12a) and strongest above the center wires in the case of the quadrupole loop (Figure 12b). Although these expectations of magnetic field strength are correct, the results presented here demonstrate that such a setup does not necessarily translate into a more sensitive detector. These observations reveal trends in the loop sensitivity contradictory to the expectations displayed in Figure 12. Although the magnetic field directly measured at a point above the center wire of the quadrupole loop is certainly stronger than that at the same point above the rectangular pole loop does not couple as well with the metal probe. Further investigation is needed to explain why this is so, but it is hypothesized that it is because the weaker magnetic field above the loop perimeter leads to a decrease in induced current and therefore of sensitivity. In addition, these observations suggest that the likelihood of false calls originating from vehicles above adjacent lanes is not lower for a quadrupole loop than for a rectangular loop. The sensitivity plots outside the loop perimeter (Figure 8) are highly similar in shape. CONCLUSION Three-dimensional maps of ILD response of four common loop geometries to aluminum and steel objects situated at different heights above the pavement are presented. Although these results represent a specific set of conditions, a mapping of loop response in three dimensions has not previously been executed in terms of ΔL/L values. These data provide the following loop design insights: Vehicles with high ground clearance (such as trailers) present challenges for accurate detection because loop sensitivity decreases by the inverse square of the distance (Equation 6) between the vehicle undercarriage and the loop face. Although heavy vehicles are more massive, smaller vehicles are more easily detectible because they are closer to the ground. Generally, there was negligible difference in the loop response between the aluminum and steel probes. Even when the probes were 12 in. above the pavement, the response difference was barely visible on the sensitivity plots (Figures 6 and 7). Although sensitivity decreases with vertical distance from the pavement, the overall shape of the sensing range remains the same (Figures 6 and 7). Similar responses were found for 6-ft octagonal and 6-ft round loops (Figure 7). Quadrupole loops were less sensitive overall than rectangular loops of similar size (Figure 6). This finding held for both the test probe and several different types of large vehicles that were driven over the loops (Table 1). The finding runs contrary to the claim that quadrupole loops are more sensitive for vehicle detection. The quadrupole loop had lower sensitivity along its central axis, which was where sensitivity was expected to be greatest (Figure 6). The field sensitivity maps show no evidence that quadrupole loops are less susceptible than rectangular loops to false calls from vehicles in adjacent lanes (Figures 6 and 8). The outcomes of this study are applicable to the design of ILDs for detection of vehicles with large clearance areas. Accurate presence

44 44 Transportation Research Record 2128 ΔL/L (Percent) Direction of Vehicular Travel Horizontal Offset (Inches) (a) ΔL/L (Percent) Direction of Vehicular Travel Horizontal Offset (Inches) (b) FIGURE 9 Three-dimensional plots of inductance response to galvanized steel sheet elevated 12 in. from pavement: (a) 20-ft quadrupole loop and (b) 20-ft rectangular loop.

45 Day, Brennan, Harding, Premachandra, Jacobs, Bullock, Krogmeier, and Sturdevant 45 ΔL/L (Percent) Horizontal Offset (Inches) (a) Direction of Vehicular Travel ΔL/L (Percent) Horizontal Offset (Inches) (b) Direction of Vehicular Travel FIGURE 10 Three-dimensional plots of inductance response to galvanized steel sheet elevated 12 in. from pavement: (a) 6-ft by 6-ft octagonal loop and (b) 6-ft round loop.

46 46 Transportation Research Record Longitudinal Position (Inches) Longitudinal Position (Inches) Horizontal Position (Inches) (a) Horizontal Position (Inches) (b) Longitudinal Position (Inches) Horizontal Position (Inches) (c) Longitudinal Position (Inches) Horizontal Position (Inches) (d) FIGURE 11 Contour plots of inductive response to galvanized steel sheet at height of 24 in. from pavement surface: (a) 20-ft quadrupole loop, (b) 20-ft rectangular loop, (c) 6-ft by 6-ft octagonal loop, and (d) 6-ft round loop.

47 Day, Brennan, Harding, Premachandra, Jacobs, Bullock, Krogmeier, and Sturdevant 47 Vehicle Flow Vehicle Flow O (a) O (b) FIGURE 12 Plots of magnetic field strength for loop geometries reflected in 1976 quadrupole loop patent, after Koerner (16): (a) rectangular loop and (b) quadrupole loop. detection of heavy trucks with large spans between axles is one such application. Another would be the detection of wooden buggies, which have few metal components sitting relatively high above the pavement surface. The methods presented in this study may also be used to map the sensitivity of other ILD geometries for a wide range of applications ranging from the detection of aircraft during ground maneuvers to the detection of smaller metal objects passing through a given space. ACKNOWLEDGMENTS This work was supported by the Joint Transportation Research Program administered by the Indiana Department of Transportation and Purdue University. REFERENCES 1. Klein, L. A., D. R. P. Gibson, and M. K. Mills. Traffic Detector Handbook. Publ. FHWA-HRT FHWA, U.S. Department of Transportation, Mills, M. K. Self Inductance Formulas for Multi-Turn Rectangular Loops Used with Vehicle Detectors. In 33rd IEEE Vehicle Technology Conference Record, IEEE, New York, May 1983, pp Mills, M. K. Self Inductance Formulas for Quadrupole Loops Used with Vehicle Detectors. In 35th IEEE Vehicle Technology Conference Record, IEEE, New York, May 1985, pp Smaglik, E. J., D. M. Bullock, and A. Sharma. A Pilot Study on Real- Time Calculation of Arrival Type for Assessment of Arterial Performance. Journal of Transportation Engineering, ASCE, Vol. 133, No. 7, July 2007, pp Day, C. M., E. J. Smaglik, D. M. Bullock, and J. R. Sturdevant. Quantitative Evaluation of Fully Actuated Versus Nonactuated Coordinated Phases. In Transportation Research Record: Journal of the Transportation Research Board, No. 2080, Transportation Research Board of the National Academies, Washington, D.C., 2008, pp Bullock, D. M., C. M. Day, and J. R. Sturdevant. Signalized Intersection Performance Measures for Operations Decision Making. ITE Journal, Aug. 2008, pp Park, S., S. G. Ritchie, and C. Oh. An Innovative Single-Loop Speed Estimation Model with Advanced Loop Data. Presented at 86th Annual Meeting of the Transportation Research Board, Washington, D.C., Dailey, D. J. Travel-Time Estimation Using Cross-Correlation Techniques. Transportation Research, Vol. 27B, No. 2, 1993, pp Kwon, J., B. Coifman, and P. Bickel. Day-to-Day Travel-Time Trends and Travel-Time Prediction from Loop-Detector Data. In Transportation Research Record: Journal of the Transportation Research Board, No. 1717, TRB, National Research Council, Washington, D.C., 2000, pp Coifman, B. Vehicle Re-identification and Travel Time Measurement in Real-Time on Freeways Using Existing Loop Detector Infrastructure. In Transportation Research Record: Journal of the Transportation Research Board, No. 1643, Transportation Research Board of the National Academies, Washington, D.C., Sun, C., S. G. Ritchie, K. Tsai, and R. Jayakrishnan. Use of Vehicle Signature Analysis and Lexicographic Optimization for Vehicle Reidentification on Freeways. Transportation Research, Vol. 7C, 1999, pp Ndoye, M., V. Totten, B. Carter, D. M. Bullock, and J. V. Krogmeier. Vehicle Detector Signature Processing and Vehicle Reidentification for Travel Time Estimation. Presented at 87th Annual Meeting of the Transportation Research Board, Washington, D.C., Hamm, R. A., and D. L. Woods. Loop Detectors: Results of Controlled Field Studies. ITE Journal, Vol. 62, No. 11, 1992, pp Kidarsa, R., T. Pande, S. V. Vanjari, J. V. Krogmeier, and D. M. Bullock. Design Considerations for Detecting Bicycles with Inductive Loop Detectors. In Transportation Research Record: Journal of the Transportation Research Board, No. 1978, Transportation Research Board of the National Academies, Washington, D.C., 2006, pp Krznaric, S., and J. A. Gatto. Electromagnetic Properties of Ferrous and Nonferrous Metals. Presented at American Institute of Aeronautics and Astronautics 37th Structural Dynamics and Materials Conference, Salt Lake City, Utah, April 18 19, Koerner, S. J. Inductive Loop Structure for Detecting the Presence of Vehicles over a Roadway. U.S. Patent 3,984,764, filed Oct. 5, The contents of this paper reflect the views of the authors, who are responsible for the facts and the accuracy of the data, and do not necessarily reflect the official views or policies of the Indiana Department of Transportation. The contents do not constitute a standard, specification, or regulation. The Traffic Signal Systems Committee sponsored publication of this paper.

48 Cycle-Length Performance Measures Revisiting and Extending Fundamentals Christopher M. Day, Darcy M. Bullock, and James R. Sturdevant Webster s single-ring formulation for cycle length and the Highway Capacity Manual (HCM) intersection saturation metric are revisited, and the model for dual-ring operation is extended to provide a tool for evaluating the effectiveness and efficiency of a prescribed cycle length. The calculations from Webster s model and the HCM provide a framework for identifying periods of time when cycle length could be substantially shortened, periods of the day when an increase in cycle length would provide some modest improvements, and periods of the day when cycle length is adequate and capacity problems should be addressed by adjusting splits. The methodology is demonstrated for incorporating the number of split failures, overall intersection saturation, and Webster s optimal cycle length to evaluate the quality of split operation and the effectiveness of cycle length at a signalized intersection. Clearly defined performance measures are essential for the analysis of any system. A great deal of effort has been spent on the measurement of freeway performance, with real-time vehicle speeds and expected travel times becoming available to the motoring public in many cities. However, performance measures for signalized arterial highways are in their infancy. As the traffic signal industry enters this new territory, it is important to identify appropriate performance measures. There are several potential audiences for traffic signal performance measures, which represent the various stakeholders in arterial road networks: system users, traffic professionals (planners, engineers, and technicians), and elected officials. It is anticipated that system users and elected officials will be interested in performance measures similar to those developed for freeways, such as speed or travel time. However, the planning and operation of a signalized arterial are considerably different from those of a freeway. Considerable effort is needed to bridge the gap between individual signal operation and network performance and to identify performance measures necessary for identifying operational improvements for signalized arterials. The current state of the practice in data collection at signalized intersections is to manually count vehicles in 15-min intervals and apply the Highway Capacity Manual (HCM) (1) methodology to assess overall intersection performance with a letter grade for the level C. M. Day and D. M. Bullock, School of Civil Engineering, Purdue University, 550 Stadium Mall Drive, West Lafayette, IN J. R. Sturdevant, Indiana Department of Transportation, 100 North Senate Avenue, Room N925, Indianapolis, IN Corresponding author: D. M. Bullock, darcy@purdue.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / of service. Adaptive control systems (2) require real-time information, the fidelity of which has historically been constrained by memory and bandwidth limitations. Signal controllers themselves typically operate at a resolution of 0.1 s, but all of the information that the controller logic processor gathers is typically discarded. Somewhere within these limits, there is a level of aggregation that makes it possible to retrieve useful information at a much better resolution than that of the HCM method yet without taxing communication and computing resources. One example is ACS-Lite (3, 4), which aggregates information in 1-min bins. Rather than using arbitrary blocks of time, using intervals with intrinsic meaning to signal operations such as cycles and phases is proposed. Several papers have been published (5 8) showing examples of signal performance measures calculated on a phase-by-phase, cycleby-cycle basis. The benefit of using this resolution is that it provides an intermediate level of aggregation above the controller resolution of 0.1 s, yet it uses an interval of time that has intrinsic meaning to signal operation. This study examines measures that can be used to quantify overall intersection saturation and evaluate cycle length. By exploring performance measures that can be created on a cycleby-cycle basis, improvement of the quality of information that can be gathered from signalized arterials is sought. Finally, the ultimate objective would be to use this information to identify cycle changes that would improve operations. CYCLE-BASED PERFORMANCE MEASURES Opening Up the HCM Delay Equation A common method of selecting cycle length and splits is to measure traffic volumes in 15-min intervals with manual counts and use this information to derive the parameters with a software package such as HCS2000 (9) or a simulation software package. The HCM method focuses on the control delay, an estimation of which is given by d = d ( PF ) + d + d ( ) where d 1, d 2, and d 3 represent uniform, incremental, and initial queue delay, respectively, and PF is a progression adjustment factor. Of the three delay components, d 1 can be calculated by using easily measurable quantities such as green time, cycle length, and volume: d gi 05. C 1 C = gi ( 1 min ( 1, Xi )) C 2 ( 2) 48

49 Day, Bullock, and Sturdevant 49 where g i = effective green for movement i, C = cycle length, and X i = volume-to-capacity ratio of movement i, given by X i v C = () 3 s g i where v is the flow rate of the movement and s is the saturation flow rate, both in units of vehicles per hour per lane [(veh/h)/ln]. The flow rate v is calculated by (a) (c) (b) (d) 3 4 N 3, 600 v = n C where N = number of counted vehicles, n L = number of lanes, C = cycle length (s), and 3,600 = conversion term. The volume-to-capacity ratio is itself a useful quantity for determining the performance of a phase, especially when calculated on a cycle-by-cycle basis. The occurrence of split failures can be estimated by using the volume-to-capacity ratio (7). As X i increases, the probability that a split failure will occur also increases. In this study, split failures are identified as occurring when X i > 1, which is a somewhat arbitrary but reasonable threshold based on engineering judgment. In addition to calculation of the volume-to-capacity ratio, the individual movement volumes can be used to calculate two other quantities investigated here: intersection saturation and Webster s optimal cycle length. Degree of Intersection Saturation The degree of intersection saturation (X C ), also known as the critical volume-to-capacity ratio, is a measure that describes the degree to which the capacity of an intersection is utilized on the basis of the volumes of the critical movements. The HCM equation for this measure is X C L v C = () 5 s C L i ci ( 4) where L is the lost time, and in the summation term over critical phases ci, v represents the volume and s is the saturation flow rate. The terminology critical phase refers to that phase having the greater volume in situations where more than one phase runs concurrently; the sequence of critical phases in a given cycle is referred to as the critical path. With a project-scheduling analogy, the determination of the critical path through a dual-ring, eight-phase controller may be reduced to a comparison between which phase pair on either side of the barrier is on the critical path. For example, one such comparison would be between Phases 1 and 2 and Phases 5 and 6. The pair of phases with the highest combined volume is said to be the critical one. The reason for pairing {1, 2} for comparison against {5, 6} and {3, 4} against {7, 8} is that the boundary between phases within a pair is allowed to vary independently of that in the other pair, whereas both pairs simultaneously end at a barrier. Figure 1 shows the four possible critical paths through a dualring, eight-phase controller. With this method to compare phases, Equation 5 simplifies to X C = FIGURE 1 Possible critical paths through eightphase, dual-ring controlled intersection: (a) Critical Path 1234, (b) Critical Path 1278, (c) Critical Path 5634, and (d) Critical Path v v v v max, s s + max, s s C L C where, for example, v 12 = v 1 /s + v 2 /s. X C represents the degree to which the total intersection capacity is used by summing volumes moving through the critical path. The measure is not intended to determine whether capacity problems exist for individual phases. Individual phases in the critical path may have volumes approaching saturation. Consequently, the calculated X C -values will not necessarily reflect high intersection utilization unless the volumes are high for all the movements in the critical path. An example calculation of X C for eight cycles is presented in Tables 1 through 5. First, a comparison is conducted between volumes of phase pairs {1, 2} and {5, 6} on the left side of the barrier, as shown in Table 1 for eight midday cycles. Whichever phase pair has the greatest combined volume is said to be critical. For example, during the cycle beginning at 14:42:20, Phases 5 and 6 have a higher combined volume than Phases 1 and 2, and therefore they are TABLE 1 Sample Calculation of X C on Cycle-by-Cycle Basis: Volume-to-Saturation Ratios and Critical Paths for Left Side of Barrier Volume (veh/h/ln) Time v 1 v 2 v 5 v 6 v 1+2 v 5+6 CP L 14:42: {5,6} 14:44: {5,6} 14:45: {1,2} 14:47: {5,6} 14:49: {1,2} 14:51: {5,6} 14:52: {5,6} 14:54: {1,2} () 6

50 50 Transportation Research Record 2128 TABLE 2 Sample Calculation of X C on Cycle-by-Cycle Basis: Volume-to-Saturation Ratios and Critical Paths for Right Side of Barrier Volume (veh/h/ln) Time v 3 v 4 v 7 v 8 v 3+4 v 7+8 CP R 14:42: {3,4} 14:44: {3,4} 14:45: {3,4} 14:47: {3,4} 14:49: {3,4} 14:51: {3,4} 14:52: {3,4} 14:54: {3,4} TABLE 4 Sample Calculation of X C on Cycle-by-Cycle Basis: Lost Time for Right Side of Barrier Lost Time by Phase (s) Time CP R L R 14:42: {3,4} :44: {3,4} :45: {3,4} :47: {3,4} :49: {3,4} :51: {3,4} :52: {3,4} :54: {3,4} 7.1 considered to be critical. This comparison is repeated for phases on the right side of the barrier, as shown in Table 2. This calculation gives the critical path (CP) through the cycle as well as the needed volumes for the maximum terms in Equation 6. The total lost time L is a summation of the lost times associated with each occurring critical phase during a cycle. Lost time calculations are shown in Tables 3 and 4 for the left and right sides of the barrier, respectively. When a phase is skipped, its lost time is equal to zero. When the traffic volumes and lost time amounts have been found, the X C calculation of Equation 6 can proceed, as shown in Table 5. In this case, the intersection is clearly under saturation during these cycles and has significant reserve capacity. plan. Webster s equation was intended for a single ring, but the model can be extended to a dual-ring, eight-phase controller by applying the concept of the critical path used in the calculation of X C. Essentially, the optimal cycle length is considered as serving the critical path through the cycle as though it were a single ring, and it is assumed that this will be sufficient to serve the concurrent but less saturated phases. Table 5 shows the calculation of C W for the eight phases used in the previous example. The two inputs for the equation the lost time and the summation of volume-to-saturation ratios are the same. As for X C, only the lost time for served phases was included. For the cycles examined here, a shorter cycle length was recommended by Webster s equation in each case. Webster s Optimal Cycle Length Webster s (10) 1958 equation for optimal cycle length (C W ) is given as C W 15. L + 5 = φ 1 Y i = 1 i ( 7) where Y i in the summation term is the ratio of volume to saturation flow rate for phase i for all phases (1 to φ) in the signal operating TABLE 3 Sample Calculation of X C on Cycle-by-Cycle Basis: Lost Time for Left Side of Barrier Lost Time by Phase (s) Time CP L L 14:42: {5,6} :44: {5,6} :45: {1,2} 5 14:47: {5,6} 5 14:49: {1,2} :51: {5,6} :52: {5,6} :54: {1,2} 8.2 DATA COLLECTION Data used in this study were collected on March 14, 2007, at the intersection of State Routes 37 and 32/38 in Noblesville, Indiana. A plan view of the intersection showing detector locations is given in Figure 2, which also shows the ring diagram. This intersection operates under a standard dual-ring, eight-phase configuration with coordinated Phases 2 and 6 and one lagging left-turn phase (Phase 5). Detection was available on Phases 2 and 6, although it was not used to operate the intersection at that time since the coordinated phases were operating in recall. Vehicles were counted by using stop bar detectors for all phases except 2 and 6, where advance detectors were used. EXAMPLE CALCULATIONS WITH REAL DATA Intersection Saturation Figure 3 shows the volumes of all eight phases on March 24, The strong a.m. and p.m. peaking characteristics of coordinated Phases 2 and 6 represent flows entering and leaving Indianapolis at those times of day. Relatively weak peaking can also be seen on Phases 4 and 8. Of the left-turn phases, Phase 5 is the strongest, revealing relatively large demand for traffic northbound from Indianapolis seeking to enter Noblesville via SR-32.

51 Day, Bullock, and Sturdevant 51 TABLE 5 Sample Calculation of X C on Cycle-by-Cycle Basis: X C and C w Max Max v C Time (v 12,v 56 ) (v 34,v 78 ) i s ci L (s) C (s) C L X C C W 14:42: :44: :45: :47: :49: :51: :52: :54: Phase Sequence SR ft (123 m) SR 32/38 INDOT Cabinet Purdue Cabinet 405 ft (123 m) FIGURE 2 Intersection of SR 37 and 32/38 in Noblesville, Indiana.

52 52 Transportation Research Record 2128 Equivalent Hourly Flow Rate (veh/h) P1 P2 P3 P4 P6 P5 P7 P8 0 0:00 12:00 24:00 0:00 12:00 24:00 0:00 12:00 24:00 0:00 12:00 24:00 Time of Day FIGURE 3 Volumes for eight phases at Noblesville test intersection on March 14, 2007, with solid lines indicating 20-point moving averages. The frequency of split failures for all eight phases for these volumes is shown in Figure 4. Here, split failures have been counted every half-hour, with the determination of whether a split failure occurred based on whether the volume-to-capacity ratio X i is greater than one. Despite the heavy volumes of Phases 2 and 6, few split failures occurred for these phases because the coordinated phases tended to receive more green time from omissions and early terminations of preceding phases. Phases 1, 3, and 7 have relatively low volumes, and relatively few split failures occurred. Phases 4 and 5, however, reported many split failures, a combination of moderate volumes, and limited availability of green time for these phases. Figure 5 is a 24-h graph of X C at the Noblesville test intersection on March 14, Since X C is based on total critical volume, the trend is the result of an aggregation of the eight volume plots in Figure 3. There is very low use during the overnight (0:00 to 6:00 and 22:00 to 24:00) free period and moderate use during most of the day, with substantial peaks during the a.m. and p.m. peak hours as well as a less prominent peak around noon. The critical path is shown by the symbol marking each data point. For this weekday pattern, The north south phase group {5, 6} is typically the critical pair in the morning, whereas Phase Group {1, 2} is typically critical in the evening, and The east west phase group {3, 4} is dominant throughout most of the day; Phase Group {7, 8} rarely appears in the critical path. As X C approaches 1, more movements at the intersection are saturated, and it becomes more difficult to move green time between phases without sacrificing needed capacity. As X C approaches 1.0, perhaps only fractions of seconds exist to be transferred between phases. Lower X C -values indicate where there is more spare capacity to be redistributed. Splits may be adjusted without taking away used 12 P1 P2 P3 P4 Number of Split Failures P6 P5 P7 P8 0 0:00 12:00 24:00 0:00 12:00 24:00 0:00 12:00 24:00 0:00 12:00 24:00 Time of Day FIGURE 4 Number of split failures by phase, counted in half-hour bins.

53 Day, Bullock, and Sturdevant pt. Mov. Avg. Intersection Saturation :00 2:00 4:00 6:00 8:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 0:00 Time of Day FIGURE 5 X C plotted over 24 h. capacity from phases, and there should be more time that can be moved around. Combining X C with the counts of split failures allows the determination of when splits could be adjusted and what phases are in need of additional capacity. Figure 6 is a graph of X C plotted only for cycles in which split failures took place. Phases 4 and 5 appear to have the most frequent failures. During the a.m. and p.m. peak periods, when X C is close to 1, split failures are unlikely to be corrected by adjusting splits because there is little underutilized green time that can be moved. However, during the midday periods, where X C tends to be well below 1, there are likely many cycles in which the green time could have been better allocated to eliminate split failures. To expand on this idea, X C can be used to classify split failures as being correctable or not. An arbitrary threshold value of X T is selected to make this determination. A split failure is considered to be correctable when X C < X T and uncorrectable (by split adjustments) when X C > X T. Figure 7 shows the number of split failures during the midday (11:00 to 13:00) and p.m. peak (15:00 to 19:00) periods, with the portion of correctable split failures indicated for two different X T -values. These graphs indicate both the total number of split failures and what proportions of them are coincident with X C -values greater or less than X T. The value of X T that was selected implies the amount of effort that can be expended on improving the splits. Selecting X T = 0.75 means that one would consider making split adjustments to address persistent split failures when X C < However, if failures when X C > 0.75 were observed, making split adjustments would not be considered, since the amount of underutilized green time is quite small, approaching 0 as X C approaches 1. In those cases, additional measures may be required, such as increasing the cycle length or adding lanes. Choosing a higher value such as X T = 0.85 requires more engineering diligence in adjusting splits, since smaller increments of green time can be transferred from one phase to another. For example, during the peak p.m. period, Phase 4 experienced 60 split failures, as shown in Figure 7. If X T = 0.75 is selected, 34 of the 60 failures are reported to be correctable (i.e., they took place when X C < 0.75). If X T = 0.85 is chosen, that number increases to 51 of 60 as the number of split failures that could be corrected by split adjustments. Split failures reported as not correctable are not necessarily impossible to address. However, as X C approaches 1.0 there is little slack capacity in the intersection and hence it becomes difficult to correct split failures by only adjusting splits. The optimal value of X T is not easily defined since it is highly dependent on how much staff time is available for tuning signal systems. However, it is clear that choosing a low value would cause missed opportunities for improving splits, whereas choosing a high value would cause diminishing returns on the effort as smaller and smaller quantities of underutilized green time are redistributed. Optimal Cycle Length Figure 8 is a 24-h graph of Webster s optimal cycle length (C W ) superimposed over the actual cycle length at the Noblesville intersection on March 14, A comparison of this graph with the volumes in Figure 3 and the X C plot in Figure 5 reveals that C W tracks with overall intersection volume. During the free periods (before 6:00 and after 22:00), very short cycle lengths (20 to 30 s) are suggested by Webster s formulation. Actual cycle lengths during this time period were measured by finding the time between successive ends of green for Phase 2. Long cycle lengths are reported when long times passed in which the phases did not cycle. The minimum

54 54 Transportation Research Record X T = 0.85 Intersection Saturation X T = :00 2:00 4:00 6:00 8:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 0:00 Time of Day Phase 1 Phase 5 Phase 2 Phase 6 Phase 3 Phase 7 Phase 4 Phase 8 20 pt. Mov. Avg. FIGURE 6 X C plotted over 24 h with split failures by phase shown. points represent the shortest possible cycle length during the free periods, and they are about 5 to 7 s longer than C W. In the coordinated periods (6:00 to 22:00), the recommended cycle lengths are typically lower than the actual cycle lengths, except for some cycles when there were high volumes for many phases. In Figure 8, the circled point represents a cycle in which 1,568 (veh/h)/ln were served on the critical path, corresponding to Σ(v/c) ci = 0.83 and X C = The recommended cycle length is s. Table 6 shows descriptive statistics for C W for all of the coordinated time periods. The average C W for the 15:00 to 19:00 time period is 72.1 s, with a standard deviation of This value compares with the actual coordinated cycle length of 116 s. The high variance is a consequence of the considerable stochastic variation in volumes from cycle to cycle. There are many times of the day when a reduction in cycle length seems feasible, particularly 9:00 to 11:00 and 19:00 to 22:00. During the 9:00 to 11:00 time period, the average value of C W is 48 s, which is half of the actual cycle length of 96 s. During the 19:00 to 22:00 time period, the average value of C W is 39 s, which is 38% of the actual cycle length of 104 s. During these two time periods, the maximum value of C W did not exceed the actual cycle length. This finding was not true of the other time periods. The advantage of reducing cycle length is that it reduces delay for minor movements. Equation 2 reveals that delay is proportional to cycle length. If cycle length is reduced and the same splits are main- tained, the g/c ratios stay approximately the same and changes in cycle length directly translate into changes in delay. The disadvantage of reducing cycle length is that the signal operates less efficiently, with the larger proportion of time spent in clearance intervals than in green. Also, the green window for the coordinated movements is reduced in size, making it more sensitive to precise offset tuning. There is also an increased probability that arriving vehicles will be cut off. However, when demand for the coordinated phases is low, as it is in these example time periods, reductions in cycle length may provide significant improvements in delay for the crossing movements. At other times of day, increases in cycle length would be advisable if significant split failures and high intersection saturation occur. If the cycle length is changed, all of the other terms in Equation 5 should remain approximately the same. Supposing that one has a cycle length C = 100 s and a lost time L = 20 s, the C/(C L) term is equal to If C is increased by 20% to 120 s, then C/(C L) = The resulting change in X C is equal to a 4% decrease. This change might allow the relief of congestion by giving phases more green time. However, this relief comes at the cost of increasing delay (Equation 2). Also, it takes a relatively large amount of added cycle length to improve X C. Because of coordination, the cycle length is fixed to provide a predictable green window for Phases 2 and 6 and to meet the capacity needs of all intersections along the corridor. The comparison of C W with the actual cycle length reveals the cost of imposing coordina-

55 Day, Bullock, and Sturdevant 55 Frequency of Split Failures Phase (a) Frequency of Split Failures Phase (b) Frequency of Split Failures Phase (c) Frequency of Split Failures Phase (d) X C > X T, cycle length change needed X C > X T, split adjustment only FIGURE 7 Number of split failures with portion that could be remedied by split adjustments within same cycle length: (a) 11:00 to 13:00, X T 0.75; (b) 15:00 to 19:00, X T 0.75; (c) 11:00 to 13:00, X T 0.85; and (d ) 15:00 to 19:00, X T tion constraints. If the controller were fully actuated with short gap extension times, the actual cycle lengths would approach C W. However, the coordination constraint forces the cycle to hold to a fixed amount, typically longer, with most of that added time being assigned to the coordinated phases. ADDITIONAL CONSIDERATIONS Infrastructure Most intersections require vehicle presence detection to operate. In addition to vehicle presence data, this methodology relies on the availability of lane count detection and high-resolution controller event logging. Several vendors now offer cost-effective products for counting, but only a few agencies regularly deploy this technology (7, 11). With regard to high-resolution controller event logging, at least one vendor now offers a product that is capable of providing the necessary fidelity of data (5). For a newly constructed intersection, the incremental cost of providing enhanced detection and data logging capability is on the order of $2,000 when loop detectors are used and perhaps as low as $500 when video detection is used. When one considers that the construction cost for a new intersection is on the order of $120,000, these costs are quite modest, given the sub- stantial benefits that can accrue from more-informed management of the traffic signal system, particularly in growth corridors where traffic patterns are rapidly changing. Pedestrian Service Pedestrian phases add another dimension to the problem by adding more constraints to the way phases are timed. For example, concurrent pedestrian phases introduce an alternative lower bound (minimum) for phase split times. A second category of pedestrian phasing, exclusive pedestrian service, introduces an entirely new phase that results in a larger proportion of lost time in the cycle for serving vehicles. Recently some new work on real-time pedestrian performance measures has been initiated (12). Although pedestrian phases were not introduced in this study, they are an important area for further extension of this work. CONCLUSIONS The first plots of intersection saturation (X C ) and Webster s optimal cycle length (C W ) based on real-time data are presented here. How to combine X C with counts of split failures was demonstrated to identify

56 56 Transportation Research Record Optimal Actual :00 2:00 4:00 6:00 8:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 0:00 FIGURE 8 Actual cycle length versus Webster s optimal cycle length. times of day when opportunities exist for improving operations and to determine whether making split adjustments would be an appropriate response to these improvements. Further study would be needed to determine what threshold value (X T ) to use to determine whether split adjustments would be an effective response. The calculation of C W provides a basis of comparison for the selected cycle length, suggesting times of day when changes in cycle length may be appropriate. Examples of calculations based on real data were included. Additional research is needed to determine threshold values for deciding when to intervene; to evaluate the impact of cycle and split changes on corridor progression; and to identify what accommodations must be made in the methodology for pedestrian service. ACKNOWLEDGMENTS This work was supported by the National Cooperative Highway Research Program and the Joint Transportation Research Program administered by the Indiana Department of Transportation and Purdue University. TABLE 6 Comparison of Actual Cycle Length with Webster s Optimal Cycle Length for Coordinated Time Periods Webster s Optimal Cycle Length (s) Actual Cycle Standard Time Period Length (s) Minimum Maximum Median Average Deviation 6:00 9: :00 11: :00 13: :00 15: :00 19: :00 22:

57 Day, Bullock, and Sturdevant 57 REFERENCES 1. Highway Capacity Manual. TRB, National Research Council, Washington, D.C., Kessmann, R. Urban Traffic Control System Traffic Adaptive Network Signal Timing Program, Vol. 1: Overview Description. Publ. FHWA RD FHWA, U.S. Department of Transportation, Luyanda, F., D. Gettman, L. Head, S. Shelby, D. M. Bullock, and P. Mirchandani. ACS-Lite Algorithmic Architecture: Applying Adaptive Control System Technology to Closed-Loop Traffic Signal Control Systems. In Transportation Research Record: Journal of the Transportation Research Board, No. 1856, Transportation Research Board of the National Academies, Washington, D.C., 2003, pp Shelby, S. G., D. M. Bullock, and D. Gettman. Resonant Cycles in Traffic Signal Control. In Transportation Research Record: Journal of the Transportation Research Board, No. 1925, Transportation Research Board of the National Academies, Washington, D.C., 2005, pp Smaglik, E. J., A. Sharma, D. M. Bullock, J. R. Sturdevant, and G. Duncan. Event-Based Data Collection for Generating Actuated Controller Performance Measures. In Transportation Research Record: Journal of the Transportation Research Board, No. 2035, Transportation Research Board of the National Academies, Washington, D.C., 2007, pp Smaglik, E. J., D. M. Bullock, and A. Sharma. A Pilot Study on Real- Time Calculation of Arrival Type for Assessment of Arterial Performance. Journal of Transportation Engineering, ASCE, Vol. 133, No. 7, July 2007, pp Day, C. M., E. J. Smaglik, D. M. Bullock, and J. R. Sturdevant. Quantitative Evaluation of Fully Actuated Versus Nonactuated Coordinated Phases. In Transportation Research Record: Journal of the Transportation Research Board, No. 2080, Transportation Research Board of the National Academies, Washington, D.C., 2008, pp Bullock, D. M., C. M. Day, and J. R. Sturdevant. Signalized Intersection Performance Measures for Operations Decision Making. ITE Journal, Aug HCS2000: Highway Capacity Software. McTrans, Gainesville, Fla. 10. Webster, F. V. Traffic Signal Settings. TRRL Report 39. U.K. Transport and Road Research Laboratory, London, Using Existing Loops at Signalized Intersections for Traffic Counts: An ITE Informational Report. ITE Journal, Vol. 78, No. 2, Feb Hubbard, S. M. L., D. M. Bullock, and C. M. Day. Integration of Real-Time Pedestrian Performance Measures into Existing Infrastructure of Traffic Signal System. In Transportation Research Record: Journal of the Transportation Research Board, No. 2080, Transportation Research Board of the National Academies, Washington, D.C., 2008, pp The contents of this paper reflect the views of the authors, who are responsible for the facts and the accuracy of the data, and do not necessarily reflect the official views or policies of the National Academy of Sciences, the American Association of State Highway and Transportation Officials, the Federal Highway Administration, or the Indiana Department of Transportation. These contents do not constitute a standard, specification, or regulation. The Traffic Signal Systems Committee sponsored publication of this paper.

58 Active and Passive Bus Priority Strategies in Mixed Traffic Arterials Controlled by SCOOT Adaptive Signal System Assessment of Performance in Fortaleza, Brazil Francisco Moraes Oliveira-Neto, Carlos Felipe G. Loureiro, and Lee D. Han In recent years, bus priority techniques for signals controlled by traffic management centers have become a viable alternative to reduce passenger delays at signalized intersections, especially in mixed traffic corridors. However, before any bus signal priority strategy is deployed in such corridors, the impacts on the different users of the system should be evaluated. The main objective of this work was to assess the operational performance of passive and active bus priority techniques in fixed and real-time signal systems of one of the main arterial corridors in Fortaleza, Brazil. As a secondary objective, it also evaluated the operational benefits of a SCOOT adaptive signal control system, comparing it with welladjusted fixed-time plans optimized by TRANSYT, for periods of medium and high traffic volumes. In the evaluation of alternative scenarios, the following performance measures were considered: vehicle delay and number of stops simulated by SCOOT, as well as bus and automobile travel times observed in the field during the operation of each scenario. The results did not favor the adoption of passive and active priority schemes in the studied corridor; this led to the conclusion that SCOOT s real-time control, programmed for a good signal progression of the general traffic (buses and automobiles), is the best signal control strategy for an arterial corridor similar to the one under analysis. In recent years, several Brazilian metropolises established traffic management centers (TMCs) that, among their other functions, are able to implement transit signal priority (TSP) systems. Signal priority strategies, which seek to improve overall urban mobility by assigning signal timing plans that favor mass transit over passenger vehicles, have been tested in many cities in the United States, Canada, Japan, and Europe. In general, they are classified into two levels (1 3): passive priority and active priority. A passive priority system typically adjusts signal timing plans manually or with computational software such as TRANSYT to reduce the delay of transit vehicles. This type of technique is suitable for locales where bus volumes are high and the dwell times are predictable (4). An active priority system detects the approach or presence of transit vehicles and commences changes to the signal phases accordingly to accommodate these vehicles. It is believed F. M. Oliveira-Neto and L. D. Han, Department of Civil and Environmental Engineering, University of Tennessee, 223 Perkins Hall, Knoxville, TN C. F. G. Loureiro, Department of Transport Engineering, Federal University of Ceará, Campus do Pici, Bloco 703, CP , Fortaleza, CE , Brazil. Corresponding author: C. F. G. Loureiro, felipe@det.ufc.br. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / that the use of active priority strategies can significantly reduce bus delay and travel time, especially in corridors under TMC control. Past studies reported that the use of bus priority techniques in arterial corridors is challenging, particularly under mixed flow conditions, where multiple types of vehicles commingle. This practice has been criticized by local agencies as increasing passenger car delay and number of stops on major streets because of its disruptive nature on signal coordination as well as the excessive delay on cross streets due to capacity reduction (5). As such, many past studies (4 12) investigated the performance and impacts of various strategies (e.g., green extension, red truncation or early green, phase insertion, phase recovery) designed to provide bus priority in mixed traffic arterial corridors. In particular, to mitigate the adverse traffic impacts, most of these studies tested active conditional priority techniques, which rely either on schedule adherence or on traffic capacity constraints. Other efforts studied the inclusion of factors such as detector and bus stop locations in various TSP algorithms. For example, Liu et al. (13) proposed an analytical approach to assess the relationship between bus detector location and effectiveness of TSP in arterial corridors. Also, in the case of corridors with nearside bus stops, commonly found in the United States, Kim and Rilett (14) proposed a TSP algorithm to account for bus arrival uncertainties at signalized intersections caused by variation in dwell times. Most of the aforementioned studies rely mainly on analytical or simulation approaches instead of field experiments. Empirical analyses are usually limited with respect to duration and amount and type of data collected, as reported by Kimpel et al. (15), who conducted an extensive study based on a large amount of empirical data and analyzed several variables from the perspective of operators (e.g., mean and variance of running time) and passengers (e.g., on-time performance, bus headway) in 24 segments of urban arterial corridors in Portland, Oregon. Some studies tested TSP strategies in the field with both simulated and observed performance measures (16), because data such as vehicle delay, queue length, and number of stops are cumbersome and expensive to obtain in the field, whereas data on bus travel time (running time), automobile travel time, bus headways and frequency, schedule adherence (on-time performance), dwell times, and so on, can be more easily collected in the field. In Fortaleza, Brazil, a metropolis with a population of over 2.5 million, there is still no priority strategy for buses, which share the roadways with other vehicles. By the end of the 1990s, field studies showed that traffic in Fortaleza s main arterial corridors already experienced an average speed of around 20 km/h (12 mph) and an average bus speed of 15 km/h, slightly less than 10 mph (17). However, starting in 58

59 Oliveira-Neto, Loureiro, and Han , traffic control for the city s most congested area was upgraded from an outdated fixed-time control to a modern adaptive signal control system called SCOOT (Split, Cycle and Offset Optimization Technique) (18), which is housed in the Urban Traffic Control Center of Fortaleza (CTAFOR). Currently, this system controls 245 of a total of 545 signalized intersections in the city and enables the implementation of different signal control strategies in its main arterial corridors. In addition, the new system allows evaluation of alternative strategies by using performance measures (such as delay and number of stops) estimated by SCOOT and stored in a database called ASTRID (Automatic SCOOT Traffic Information Database) (19), aggregated in 15-min intervals. SCOOT has been used for assessing bus priority strategies by researchers worldwide. In England, where the SCOOT model was originally developed, many simulations and field trials were performed in London and Southampton (20). U.S. studies were conducted by using microsimulation to evaluate bus priority operations in arterial corridors in Washington, D.C., Northern Virginia (21), and Salt Lake City, Utah (22). In general, the scenarios evaluated in these studies considered the performance of TSP strategies under adaptive signal control as well as optimized fixed-time or actuated signal control. Taking advantage of the authors access to and familiarity with Fortaleza s CTAFOR system and its wealth of real-time data, this study assessed the operational performance of several scenarios with and without bus signal priority, during peak and off-peak traffic periods, in one of the city s major arterial corridors. Whereas previous studies mostly relied on simulations, this study innovatively experimented with passive and active signal priority under fixed and realtime controls in mixed traffic conditions as observed in large Brazilian cities. The main objective of this study is to demonstrate the analyses for assessing the performance of various signal priority strategies, some of them designed and embedded into the SCOOT system. ANALYSIS OF PASSIVE BUS PRIORITY STRATEGIES TRANSYT Model One of the TSP strategies analyzed was the passive signal priority model embedded in TRANSYT (Version 10), named BUS TRANSYT. The behaviors of buses and passenger cars in the arterial corridor were modeled separately with a technique called shared stop-line (23, 24); that is, the main link of each approach is accompanied by additional links that can represent the arrival of other vehicle types (e.g., buses or emergency vehicles). The model can handle up to five different vehicle classes, a handy function that can separate different flow origins (e.g., parking exits) and types of roads; it can also model buses that follow specific routes and serve certain bus stops. Because of the complex interaction between buses and other vehicle types in a traffic stream, buses are modeled in TRANSYT by using a combination of the shared stop-line method and a bus-specific dispersion model. This dispersion model considers the bus travel time in each link and implicitly considers the dwell times per bus stop. Given that the delays and the number of stops are computed for buses and cars separately, performance measures per person can be estimated by using the relative number of persons transported per vehicle. The optimizing routine will try to find the signal settings that minimize the overall person number of stops and delay. Therefore, in this study, the following parameters were set up for each shared link: bus frequencies, average bus cruising speeds, mean dwell time for each bus stop, and weights that express the relative number of persons transported by bus and car. SCOOT Real-Time Control In the SCOOT adaptive control system, certain parameters can be modified to improve the movement of buses in arterials. For example, there are two parameters split weighting multiplier (SPWM) and split weighting saturation (SPWS) that can be used to unbalance the link saturations at a signalized intersection and prioritize heaviertraffic links (25). Therefore, these two parameters, which may be applied on links other than bus links, allow the increase in green times of bus links to obtain better signal progression. SPWM establishes the level in which the saturation of a link should be increased and SPWS establishes the cut-off value of saturation, applied to the penalized link, to return to the equilibrium condition. The SCOOT model also has a parameter to mitigate congestion situations, the congestion importance factor, which extends the green length on congested links, reducing queues, and can be applied on bus links during peak hours. Another alternative for passive priority in SCOOT, which was one of the strategies analyzed in this study, consists in narrowing the allowable range for variation of the offsets on the basis of field observations of bus travel and dwell times. This strategy can be accomplished by changing a parameter called Link Bias (varying between 0 and 127, corresponding the lowest and highest restrictions), which limits the offset ranges and can be applied to the default offset values. SCOOT ACTIVE BUS PRIORITY MODEL The active bus priority feature was implemented in SCOOT only in 1995 (26). The priority logic consists in allocating extra green times, allowing a detected bus to pass through either the current or the next green indication. These extra times are inserted either in the beginning of each posterior green indication (which is named recall or early green ) or at the end of the current green (extensions). In general, the SCOOT model attempts to predict the arrival moment of a detected bus at an intersection, which in turn determines the type of strategy to be executed. There are three possible strategies: no priority is executed in the case of an arrival within the green indication, an extension is requested in the case of an arrival right after the green period, or otherwise a recall is requested, which once executed decreases the bus delay in the queue. In SCOOT, allowable values of green extension and recall can be set by using proper parameters. The priority requests are also conditioned by traffic capacity at intersections, defined in terms of imposed saturation conditions required before either extensions or recalls are executed (26). In other words, the priority requests are executed only if extra times due to extensions or recalls do not exceed the maximum allowable values, and there is spare capacity under normal operation of SCOOT without priority. It is also important to mention that after a bus leaves the intersection having received an extension or recall, there is a period of recovery to bring the signal to its normal operation. The SCOOT model allows four recovery methods: do nothing, DN; minimum stage, MS; degree of saturation, DS; and long stage, LS (26). DS was the recovery method used in this study. In DS recovery, an imposed saturation is set to execute the stages with duration less than what they would be under normal SCOOT operation without priority. Although it takes a little longer to return to normal signal

60 60 Transportation Research Record 2128 operation compared with MS and LS recoveries, DS recovery causes the least traffic disruption and is the most suitable method for arterials because it does not considerably affect offsets, as a DN recovery does. Moreover, it does not cause oversaturation and excessive delays, as could happen with MS and LS recovery methods. In the case of multiple priority requests for example, when other buses are detected while an extension, recall, or recovery is already in progress the model will evaluate each request independently and estimate the effect for each bus according to the preestablished constraints (i.e., spare capacity, allowable green extension, or recalls). In general, if importance factors for buses (e.g., buses with higher passenger volumes that run on major links) are defined, the model will act according to this predefinition; otherwise an extension takes precedence over a recall. The bus detection method can be based on transponders, automated vehicle location, or any other feasible technique. The system should be able to detect buses at predefined points upstream of each approach stop bar and after any existing bus stop, which means that SCOOT does not model the time spent at bus stops. CORRIDOR CHARACTERISTICS The 13 de Maio Avenue was selected as the case study for this research since it is one of the major two-way arterials linking the west and east regions of Fortaleza. Traffic volumes on this 1.5-mi-long arterial corridor are around 30,000 vehicles daily and up to 180 buses per hour, running in both directions. Its cross section presents four lanes, two in each direction, physically separated by a 30-in.-wide raised median. Along the corridor, there are 10 signalized intersections, closely spaced 240 m (800 ft) apart on average. Before operating under SCOOT control, all signals were controlled by electromechanical equipment, allowing only single plans without coordination. Currently, these signals are operating exclusively under real-time control. As for public transit operation, it is worth mentioning that buses take 62% of total person movements on the corridor, even though the traffic composition accounts for only 14% of buses on the arterial and 6% on the cross streets. It is important to point out that the high bus volumes observed on both main and cross streets can lead to frequencies as high as one bus per minute per intersection in off-peak periods and two buses per minute per intersection in peak periods. Regarding transit and traffic current performance levels, it was observed in field surveys that buses operating along the analyzed corridor usually spend in the range of 25% to 35% of their journey time stopped at signals, with no significant differences between peak and off-peak periods, whereas automobiles spend around 40% in off-peak periods and up to 58% in peak periods stopped at signalized intersections. The dwell time operations usually take a proportion of 15% (westbound) and 25% (eastbound) from the bus journey time. Regarding travel times along the corridor, the average values for automobiles were around 4 min in off-peak and 6 min in peak periods, whereas the average bus travel times were around 6 min in off-peak and 8 min in peak periods. For the purposes of this study, the 13 de Maio corridor was divided into two sections based on its cross-street traffic levels: Section 1 had light traffic on cross streets, composed mainly of private cars; and Section 2 had heavy traffic on cross streets, composed of both buses and private cars. Therefore, along its Section 2, the arterial is crossed by other important bus corridors, making it possible to analyze priority conflicts, with priority requests occurring on all approaches at each signalized intersection. EVALUATION METHOD The following set of evaluation scenarios were considered: Scenario 1. Before urban traffic control (pre-utc). Single-time plan without coordination, corresponding to the existing conventional operation before CTAFOR implementation; Scenario 2. Well-adjusted fixed-time (WAFT) plans. Fixed-time plans computed in TRANSYT by using traffic flow data collected in the field and adjusted by CTAFOR s traffic engineers according to real traffic conditions observed in the field; Scenario 3. Passive priority under fixed-time (PPFT) control. Fixed-time signal settings computed with the BUS TRANSYT model; Scenario 4. SCOOT control (SCOOT). Signal times calculated by the adaptive SCOOT control model, without bus priority strategy; Scenario 5. Passive priority under SCOOT control (PPSC). Signal times calculated by the adaptive SCOOT control model, with offset permissible range set based on the average bus speeds and dwell times; Scenario 6. Active priority under real-time control on main street (APRT1). Performance of the priority logic embedded into the SCOOT model for buses running on the main street; and Scenario 7. Active priority under real-time control on both main and cross streets (APRT2). Performance of the SCOOT bus priority logic for buses running on both main and cross streets; as mentioned before, this scenario permitted analysis of the effect of request conflicts between cross and main streets. The context described by the foregoing scenarios encompasses all possible alternatives to improve traffic signal operational performance and reduce congestion on an arterial corridor that is currently operating under a noncoordinated single-time signal plan. Moreover, the experiment was also designed to consider a situation representing arterials with heavy bus volumes, commonly observed in Brazilian urban networks. Therefore, the idea was to represent a decisionmaking process in which the traffic engineer would face the following sequence of choices: (a) keep the existing control technology and merely update the fixed-time plans; (b) update the fixed-time plans with a passive bus priority strategy; (c) implement an adaptive traffic control system; (d) implement an adaptive traffic control system with passive bus priority; or (e) implement an adaptive traffic control system with active bus priority. Performance Measures In the evaluation of the proposed scenarios, the following performance measures were considered: vehicle delay and number of stops simulated by SCOOT as well as bus and automobile travel times observed in the field during the operation of each scenario s control strategy. To assess the impact of the control strategies over the general traffic stream (in Scenarios 2 to 7), vehicle delays and numbers of stops were collected from the SCOOT database by using an interface called TRANSCOOT, which was built specifically to transfer data from the SCOOT database to a geographic information system platform (27). It is important to note that these modeled variables were originally stored per link and per interval of 15 min; however, in this study they were spatially aggregated, for each 15 min, over each direction of traffic by summation of the products of the corresponding link delay and volume, resulting in overall vehicle delay, in minutes per hour, and number of stops per direction and per 15 min. Field surveys were carried out mainly to evaluate the benefits of the control strategies for buses as well as the performance of general traf-

61 Oliveira-Neto, Loureiro, and Han 61 TABLE 1 Evaluation Scenarios and Their Corresponding Performance Measures Scenario Scenarios without priority Scenario 1 Pre-UTC fic in nonpriority scenarios. Thus, automobile travel time surveys took place only in scenarios without priority (Scenarios 1, 2, 4), whereas bus travel time surveys took place in all scenarios (1 through 7). In the pre-utc scenario, it was impossible to estimate vehicle delay and number of stops because the CTAFOR system had not been deployed yet. Therefore, this scenario was analyzed by using only the variables of bus and automobile travel times measured in the field. Table 1 shows the scenarios and their corresponding performance measures. Field Experiment Scenario 2 WAFT Scenario 4 SCOOT Scenarios with priority Scenario 3 PPFT Scenario 5 PPSC Scenario 6 APRT1 Scenario 7 APRT2 Performance Measures Bus travel time Auto travel time Bus travel time Auto travel time Vehicle delay Number of stops Bus travel time Vehicle delay Number of stops All scenarios were observed during weekdays (Tuesdays, Wednesdays, and Thursdays) in the morning off-peak (8:30 to 10:30 a.m.) and in the afternoon peak (5:15 to 6:45 p.m.). These time periods were selected to allow the assessment of TSP benefits on both medium and high demand traffic conditions. The SCOOT active priority model was tested only on Section 2 of the studied corridor, where observers placed at predefined points detected bus presence by using manual plugs wired to signal controllers. Buses operating on major routes serving the corridor s main and cross streets were detected, and the priority strategy was implemented regardless of the bus schedule status (ahead or late). Regarding travel time surveys, automobile data were collected with a probe car driving along the main street at approximately the average speed of its traffic stream. As for buses, their travel and dwell times were collected through onboard surveys. The use of script routines written in specific programming language (POCKETC) for palmtop computers reduced the size of the surveying crew and made it possible for one person to register both passage instants and loss times at traffic signals and bus stops. The sample size adopted for the travel time surveys was estimated by using data collected during the six weekdays of the pre-utc scenario. The resulting sample was used to estimate the number of days needed to test the subsequent scenarios. Therefore, a minimum sample size of 20 observations was determined assuming a 5% significance level and a 10% error. Hence, given the available resources (two observers to execute onboard surveys and one car with two observers to collect automobile travel times), three weekdays worth of data collection was adopted for each of the remaining scenarios. Analysis Methodology The comparison of the scenarios was based on the following research hypotheses: 1. Updating signal time plans will reduce both automobile and bus mean travel times without changing the variance of these variables; 2. Passive bus priority under fixed-time control in comparison with well-adjusted fixed-time control will benefit public transport vehicles in terms of mean travel time reduction without affecting the performance of the other types of vehicles traveling on both main and cross streets; 3. Real-time control compared with well-adjusted fixed-time and passive priority under fixed-time control strategies changes signal times accordingly to traffic fluctuations, thus decreasing both mean and variance of travel time and vehicle delay variables; 4. Passive bus priority under real-time control compared with SCOOT real-time control without priority strategy will benefit public transport vehicles in terms of mean travel time reduction without affecting the performance of the other types of vehicles traveling on both main and cross streets; and 5. Active bus priority under real-time control compared with SCOOT real-time control without priority strategy will reduce both mean and variance of bus travel time without significantly affecting general traffic running on main and cross streets. Tables 2 and 3 show the set of alternative hypotheses considered in the analysis methodology. Differences considered significant in bus and automobile mean travel times were the ones resulting in 5% significance levels and in travel time reductions of at least 5 s per signal. Significant differences in mean vehicle delays among scenarios were found at the 1% significance level, resulting in absolute differences in mean stopped delay (sd) of at least 5 s per stopped vehicle per traffic direction: sd 60 vd ns sveh () 1 = ( ) where vd is the mean vehicle delay in minutes and ns is the mean number of stops in number of vehicles. TABLE 2 Scenario Comparisons: Alternative Hypotheses for Updating Signal Plans 1. Updating Signal Plans Performance Measure Mean Travel Time per Direction Variance of Travel Time 2 versus 1 well-adjusted Bus: H a : µ btt2 < µ btt1 Bus: H a : σ 2 btt2 σ 2 btt1 fixed-time versus pre-utc Auto: H a : µ att2 < µ att1 Auto: H a : σ 2 att2 σ 2 att1 NOTE: btt = bus travel time; att = auto travel time; vd = vehicle delay per traffic direction; ns = number of stops per traffic direction; = cases not tested.

62 62 Transportation Research Record 2128 TABLE 3 Scenario Comparisons for Additional Performance Measures Main Links Main and Cross Links Performance Measure Mean Bus Travel Time Mean Auto Travel Time Vehicle Delay Number of Stops 2. Passive Priority Under Fixed-Time Control 3 versus 2 passive priority under fixed-time H a : µ btt3 < µ btt2 H a : µ vd3 µ vd2 H a : µ ns3 µ ns2 versus well-adjusted fixed-time H a : σ 2 btt3 σ 2 btt2 H a : σ 2 vd3 σ 2 vd2 H a : σ 2 ns3 σ 2 ns2 3. Real-Time Control 4 versus 2 SCOOT control versus well-adjusted H a : µ btt4 < µ tpo2 H a : µ att4 < µ att2 H a : µ vd4 < µ vd2 H a : µ ns4 µ ns2 fixed-time H a : σ 2 btt4 <σ 2 tpo2 H a : σ 2 att4 <σ 2 att2 H a : σ 2 vd4 <σ 2 vd2 H a : σ 2 ns4 σ 2 ns2 4 versus 3 SCOOT control versus passive priority H a : µ btt4 < µ btt3 H a : µ vd4 < µ vd3 H a : µ ns4 µ ns3 under fixed-time H a : σ 2 btt4 <σ 2 btt3 H a : σ 2 vd4 <σ 2 vd3 H a : σ 2 ns4 σ 2 ns3 4. Passive Priority Under Real-Time Control 5 versus 4 passive priority under real-time H a : µ btt5 < µ btt4 H a : µ vd5 µ vd4 H a : µ ns5 µ ns4 control versus SCOOT control H a : σ 2 btt5 σ 2 btt4 H a : σ 2 vd5 σ 2 vd4 H a : σ 2 ns5 σ 2 ns4 5. Active Priority Under Real-Time Control 6 versus 4 active bus priority on H a : µ btt2 < µ btt1 H a : µ vd2 µ vd1 H a : µ ns2 µ ns1 main street versus SCOOT control H a : σ 2 btt2 <σ 2 btt1 H a : σ 2 vd2 σ 2 vd1 H a : σ 2 ns2 σ 2 ns1 7 versus 6 active bus priority on both main and H a : µ btt3 < µ btt2 H a : µ vd3 µ vd2 H a : µ ns3 µ ns2 cross streets versus active priority on main street H a : σ 2 btt3 <σ 2 btt2 H a : σ 2 vd3 σ 2 vd2 H a : σ 2 ns3 σ 2 ns2 7 versus 4 active priority on both main and H a : µ btt3 <σ 2 btt1 H a : µ vd3 µ vd1 H a : µ ns3 µ ns1 cross streets versus SCOOT control H a : σ 2 btt3 <σ 2 btt1 H a : σ 2 vd3 σ 2 vd1 H a : σ 2 ns3 σ 2 ns1 NOTE: btt = bus travel time; att = auto travel time; vd = vehicle delay per traffic direction; ns = number of stops per traffic direction; = cases not tested. Regarding the variable number of stops, it was assumed at the 1% significance level, resulting in differences of at least 5% in the mean stop proportion: ns sp = (%) ( 2) n f where ns = mean number of stops per traffic direction, f = mean traffic flow per link, and n = number of links per traffic direction. When the difference in variances was tested, it was considered at a 5% significance level for bus and automobile travel times, whereas for variances of vehicle delay and number of stops those differences in which the null hypothesis was rejected were assumed as significant at a level of 2.5%. ANALYSIS RESULTS Initially control variables (traffic flows and dwell times) were compared among scenarios. Graphical analyses showed that temporal and spatial traffic flow profiles were similar over all scenarios. Statistical tests (ANOVA and tests of variance), with 5% significance levels, also showed that all scenarios presented similar behavior in terms of dwell time operation. Therefore, these results ensured that similar traffic conditions and bus operational characteristics were observed among the evaluated scenarios, which guaranteed that any changes in performance measure values resulted solely from the tested control strategies. As for the hypothesis tests on the mean, Table 4 shows the results for the analyses of Scenario Comparisons 2 versus 1, 3 versus 2, 4 versus 2, and 5 versus 4 considering passive priority strategies and SCOOT adaptive control, whereas Table 5 shows the results for the analyses of Scenario Comparisons 6 versus 4 and 7 versus 4, taking into account active priority strategies. The negative values in boldface type indicate evidence of significant differences between the two analyzed scenarios, confirming the research hypothesis in question. However, positive differences in boldface type indicate that the result contradicts the initial hypothesis. In general, the results showed the following hypotheses: Hypothesis 1 (updating signal time plans) was confirmed only for the morning off-peak period in Section 1; Hypothesis 2 (passive priority under fixed-time control) was not confirmed in any of the analyzed cases; Hypothesis 3 (real-time control) was confirmed in several cases, mainly relating to reduction in automobile travel time as well as number of stops on the main street and reduction in delay variance for the crossing traffic; Hypothesis 4 (passive priority under real-time control) was confirmed only for traffic moving east on the main street during the afternoon peak; and Hypothesis 5 (active priority under real-time control) was not confirmed in any of the analyzed cases. The fact that well-adjusted fixed-time plans (Scenario 2) did not perform significantly better than the pre-utc signal plans (Scenario 1) during peak hours, especially in Section 2, can be explained by the reduction in operational capacity on some of the main street approaches because of inclusion of actuated pedestrian phases as a result of the TMC s new policy on enhancing pedestrian safety. Regarding the results for the scenarios with passive priority strategies, the lack of confirmation of both Hypotheses 2 and 4 for almost all assessed cases showed the low effectiveness of this type of TSP strategy, either under fixed- or real-time control, on heavy mixed traffic bus corridors. Two main reasons could be associated with that

63 Oliveira-Neto, Loureiro, and Han 63 TABLE 4 Scenario Comparisons: Results Section 1 Section 2 Variable Difference EB WB NB SB EB WB NB SB 2 Versus 1 Well-Adjusted Fixed-Time Versus Pre-UTC (Testing Hypothesis 1: updating signal time plans) Morning off-peak period Δatt (s/auto/signal) Δbtt (s/bus/signal) Afternoon peak period Δatt (s/auto/signal) Δbtt (s/bus/signal) Versus 2 Passive Bus Priority Under Fixed-Time Versus Well-Adjusted Fixed-Time (Testing Hypothesis 2: passive priority under fixed-time control) Morning off-peak period Δbtt (s/bus/signal) Δsd (s/vehicle) Δsp (%) Afternoon peak Δbtt (s/bus/signal) Δsd (s/vehicle) Δsp (%) Versus 2 SCOOT Control Versus Well-Adjusted Fixed-Time (Testing Hypothesis 3: real-time control) Morning off-peak period Δatt (s/auto/signal) Δbtt (s/bus/signal) Δsd (s/vehicle) Δsp (%) Afternoon peak Δatt (s/auto/signal) Δbtt (s/bus/signal) Δsd (s/vehicle) Δsp (%) Versus 4 Passive Bus Priority Under Real Time Versus SCOOT Control (Testing Hypothesis 4: passive priority under real-time control) Morning off-peak period Δbtt (s/bus/signal) Δsd (s/vehicle) Δsp (%) Afternoon peak Δbtt (s/bus/signal) Δsd (s/vehicle) Δsp (%) NOTE: Δatt = differences in mean auto travel time per signal and per traffic direction; Δbtt = differences in mean bus travel time per signal and per traffic direction; Δsd = differences in mean stopped delay in seconds per vehicle and per traffic direction; Δsp = differences in mean stop proportion per traffic direction. result: first, a high level of interaction between buses and automobiles on the main street and, second, very heterogeneous behavior of bus dwell times, especially during peak hours. In contrast, the confirmation of Hypothesis 3, for almost all cases, highlighted the operational benefits of implementing an adaptive signal control in an urban network with characteristics similar to the ones observed in this case study. It is worth noting that for the cases of significant reductions in bus and automobile travel times, most gains were from 10% to 20% of the total travel times. As a matter of fact, in Scenario 4 (SCOOT control without TSP strategies) all offset values (default values of the parameter DEFO) were defined by using the values obtained from the well-adjusted fixed-time plans, having their variation bounded by the parameter Link Bias. In addition, all cross streets were penalized with the SPWM and SPWS parameters (with degrees of saturation varying from 80% to 100%). The aim was to provide under SCOOT control good traffic progression on the main street. This technical decision, also applied to most arterials controlled by CTAFOR, resulted from previous empirical studies that indicated that the best way to ensure good traffic progression in arterials under SCOOT adaptive control was to limit their signal offset variations and slightly increase the saturation of their cross streets. Thus, even though the responsive feature of real-time control was fairly compromised, significant reductions in vehicle stops and mean travel times on the main streets were observed, as was confirmed by the results of the hypothesis tests shown in Table 4. Moreover, in some cases, significant reductions in delay variances

64 64 Transportation Research Record 2128 TABLE 5 Analysis of SCOOT Active Bus Priority Logic: Hypothesis 5 Active Priority Under Real-Time Control Morning Off-Peak Afternoon Peak Variable Difference EB WB NB SB EB WB NB SB 64 Active Bus Priority on Main Street Versus SCOOT Control Δbtt (s/bus/signal) Δsd (s/vehicle) Δsp (%) Active Bus Priority on Main and Cross Streets Versus SCOOT Control Δbtt (s/bus/signal) Δsd (s/vehicle) Δsp (%) for the cross traffic were also observed. Therefore, while keeping good progression for the main street traffic, SCOOT control proved to handle traffic fluctuations on cross streets quite well. Last, in assessing the active bus priority logic embedded in the SCOOT model, as shown in Table 5, the results observed in this case study indicated that the implemented strategies (extension, recall, and recovery), contrary to expectations, did not significantly improve bus performance measures. Besides, as proved by the results of comparisons of Scenario 6 versus 4 and 7 versus 4, they also had negative impacts (increase in vehicle delay and number of stops) to the general traffic moving westbound (which is the heavier-traffic direction of the corridor), as a consequence of signal coordination disruption. The factor that might have mainly contributed to lower the performance of active priority strategies was that SCOOT bus priority logic applies strategies for each node individually, without taking into account adjacent nodes. That is, SCOOT may implement distinct strategies within the same cycle time in adjacent nodes; for instance, while an extension is being executed in one node, compensation might be taking place in another adjacent node. This situation can occur due to priority conflicts caused by multiple requests coming from distinct traffic directions (e.g., eastbound versus northbound requests) since, as expected, more than one bus detection every minute in each node was observed during the field experiments. In addition, as mentioned before, SCOOT control, represented by Scenario 4, was already adjusted to enhance the performance of traffic moving on the main street; thus, any coordination disruption due to execution of a priority strategy in Scenarios 6 or 7 caused additional vehicle delay and stops for the main street traffic. It is also worth mentioning that in Scenario 7, additional priority requests on cross streets increased vehicle delay and number of stops even more for the westbound traffic. In this case, besides loss of signal coordination, reduction in the main street s capacity was also observed caused by extra green times given to detected buses running on cross streets. CONCLUSIONS This study assessed the performance of passive and active bus priority strategies under both fixed- and real-time (adaptive) signal control on a high-volume mixed traffic two-way urban arterial in Fortaleza, Brazil. In addition, it also assessed the benefits of SCOOT adaptive control when compared with well-adjusted fixed-time plans on the same bus corridor. The first conclusion to be drawn from this research is that the results of the performance evaluations did not favor the adoption of the passive priority techniques tested in this case study. Some reasons may be given for such a negative outcome. First, in mixed traffic corridors with high bus volumes, as for the case analyzed, the tested passive priority techniques tended to increase the interaction between automobiles and buses along the main street. Thus, as signal offsets were adjusted to better fit bus travel times, most vehicles in the main traffic platoons had to stop more at intersections, increasing queues lengths. Moreover, since buses usually arrive at the downstream approach after the main traffic platoon, the additional queues increased bus delays and sometimes caused buses to wait in queue until the next green indication. Second, the high variability observed in dwell times, with coefficients of variation around 50%, made the operation with passive priority based on mean dwell times inefficient. Regarding the active bus priority strategies implemented only on the main street s approaches, this case study showed that the extra times given through extensions and recalls, as well as the recovery strategy, caused disruption of signal coordination, resulting in additional vehicle delay and number of stops for the main street traffic. The explanation for this low performance relates not only to the corridor s physical and operational conditions but also to SCOOT bus priority logic. The arterial s signalized nodes are closely spaced, requiring good traffic progression, which was provided by setting the SCOOT model to limit the variation of signal offsets and to slightly saturate the operation on cross streets. Therefore, any little changes in signal times caused undesirable disruption of signal coordination. In addition, SCOOT bus priority logic works individually on nodes, and since the studied corridor is a two-way arterial with high bus frequencies (more than one bus per node every cycle), there were many priority conflicts in which the SCOOT logic implemented different combinations of strategies for adjacent nodes. Such an effect was intensified when priority conflicts with cross streets were tested, increasing vehicle delays and number of stops for the main street traffic. In summary, the results of this empirical analysis led to the overall conclusion that among those tested, the most suitable type of signal control for a two-way arterial with mixed traffic and high bus volumes is SCOOT adaptive control without priority strategies, as long as it is precisely set up to prioritize the general traffic moving on the main street and to accommodate the randomness of the traffic on cross streets.

65 Oliveira-Neto, Loureiro, and Han 65 ACKNOWLEDGMENTS This research was partially supported by Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq), of the Ministry for Science and Technology of Brazil. The authors thank the Autarquia Municipal de Trânsito e Cidadania de Fortaleza for the opportunity to work in partnership with CTAFOR s staff and to have access to its database. The authors also thank the U.S. Department of Transportation for providing travel funds (under an Eisenhower Graduate Fellowship) for this paper to be presented at the 88th TRB Annual Meeting. REFERENCES 1. Wood, K. Urban Traffic Control, System Review. TRL Project Report 41. U.K. Transport Research Laboratory, Wokingham, Berkshire, England, Fox, K., F. Montgomery, S. Shepherd, C. Smith, S. Jones, and F. Biora. Integrated ATT Strategies for Urban Arterials: DRIVE II Project PRIMAVERA Bus Priority in SCOOT and SPOT Using TIRIS. Traffic Engineering and Control, Vol. 36, No. 6, 1995, pp Sunkari, R. S., P. S. Beasley, T. Urbanik II, and D. B. Fambro. Model to Evaluate the Impacts of Bus Priority on Signalized Intersections. In Transportation Research Record 1494, TRB, National Research Council, Washington, D.C., 1995, pp Skabardonis, A. Control Strategies for Transit Priority. In Transportation Research Record: Journal of the Transportation Research Board, No. 1727, TRB, National Research Council, Washington, D.C., 2000, pp Balke, K. N., C. L. Dudek, and T. Urbanik II. Development and Evaluation of Intelligent Bus Priority Concept. In Transportation Research Record: Journal of the Transportation Research Board, No. 1727, TRB, National Research Council, Washington, D.C., 2000, pp Agrawal, B. B., S. T. Waller, and A. Ziliaskopoulos. Modeling Approach for Transit Signal Preemption. In Transportation Research Record: Journal of the Transportation Research Board, No. 1791, Transportation Research Board of the National Academies, Washington, D.C., 2002, pp Dion, F., H. Rakha, and Y. Zhang. Evaluation of Potential Transit Signal Priority Benefits Along a Fixed-Time Signalized Arterial. Journal of Transportation Engineering, ASCE, Vol. 130, No. 3, Janos, M., and P. G. Furth. Bus Priority with Highly Interruptible Traffic Signal Control: Simulation of San Juan s avenida Ponce de Leon. In Transportation Research Record: Journal of the Transportation Research Board, No. 1811, Transportation Research Board of the National Academies, Washington, D.C., 2002, pp Chang, J., J. Collura, F. Dion, and H. Rakha. Evaluation of Service Reliability Impacts of Traffic Signal Priority Strategies for Bus Transit. In Transportation Research Record: Journal of the Transportation Research Board, No. 1841, Transportation Research Board of the National Academies, Washington, D.C., 2003, pp Wadjas, Y., and P. G. Furth. Transit Signal Priority Along Arterials Using Advanced Detection. In Transportation Research Record: Journal of the Transportation Research Board, No. 1856, Transportation Research Board of the National Academies, Washington, D.C., 2003, pp Liao, C.-F., and G. A. Davis. Simulation Study of Bus Signal Priority Strategy: Taking Advantage of Global Positioning System, Automated Vehicle Location System, and Wireless Communications. In Transportation Research Record: Journal of the Transportation Research Board, No. 2034, Transportation Research Board of the National Academies, Washington, D.C., 2007, pp Muthuswamy, S., W. R. McShane, and J. R. Daniel. Evaluation of Transit Signal Priority and Optimal Signal Timing Plans in Transit and Traffic Operations. In Transportation Research Record: Journal of the Transportation Research Board, No. 2034, Transportation Research Board of the National Academies, Washington, D.C., 2007, pp Liu, H., A. Skabardonis, W. Zhang, and M. Li. Optimal Detector Location for Bus Signal Priority. In Transportation Research Record: Journal of the Transportation Research Board, No. 1867, Transportation Research Board of the National Academies, Washington, D.C., 2004, pp Kim, W., and L. R. Rilett. Improved Transit Signal Priority System for Networks with Nearside Bus Stops. In Transportation Research Record: Journal of the Transportation Research Board, No. 1925, Transportation Research Board of the National Academies, Washington, D.C., 2005, pp Kimpel, T. J., J. Strathman, R. L. Bertini, and S. Callas. Analysis of Transit Signal Priority Using Archived TriMet Bus Dispatch System Data. In Transportation Research Record: Journal of the Transportation Research Board, No. 1925, Transportation Research Board of the National Academies, Washington, D.C., 2005, pp Zheng, J., G. Zhang, Y. Wang, and P. M. Briglia. Comprehensive Evaluation of Transit Signal Priority System Using Observed and Simulated Traffic Data. Presented at 87th Annual Meeting of the Transportation Research Board, Washington, D.C., Loureiro, C. F. G., and M. G. S. Oliveira. GIS Database/Interface for Urban Traffic Management in Brazil. Presented at the 69th ITE Annual Meeting, Las Vegas, Nev., SCOOT 0414: Executive Summary Description of SCOOT. SCOOT Traffic Handbook. U.K. Transport Research Laboratory, Wokingham, Berkshire, England, SCOOT 7730: ASTRID Operator Manual. SCOOT Traffic Handbook. U.K. Transport Research Laboratory, Wokingham, Berkshire, England, Hounsell, N., and G. Wall. Examples of New Intelligent Transportation Systems Applications in Europe to Improve Bus Services. Presented at 81st Annual Meeting of the Transportation Research Board, Washington, D.C., Zhang, Y. An Evaluation of Transit Signal Priority and SCOOT Adaptive Signal Control. MSc thesis. Virginia Polytechnic Institute and State University, Blacksburg, Feng, Y., H. J. Perrin, Jr., and P. T. Martin. Bus Priority of SCOOT Evaluated in VISSIM Simulation Environment. Presented at 82nd Annual Meeting of the Transportation Research Board, Washington, D.C., Robertson, D. I., and R. A. Vincent. Bus Priority in a Network of Fixed- Time Signals. TRRL Report LR666. U.K. Transport and Road Research Laboratory, Crowthorne, Berkshire, England, Pierce, J. R., and K. Wood. BUS TRANSYT: A User s Guide. TRRL Report SR266. U.K. Transport and Road Research Laboratory, Crowthorne, Berkshire, England, SCOOT 0474: Operational Guide to Customizing. SCOOT Traffic Handbook. Transport Research Laboratory, SCOOT 0484 Functional Description Bus Priority. SCOOT Traffic Handbook. U.K. Transport Research Laboratory, Wokingham, Berkshire, England, Loureiro, C. F. G., H. B. Meneses, F. M. de Oliveira Neto, and M. M. de Castro-Neto. Managing Congestion in Large Brazilian Urban Area Through Logical Interface Between SCOOT and GIS Platform. Presented at 88th Annual Meeting of the Transportation Research Board, Washington, D.C., The Traffic Signal Systems Committee sponsored publication of this paper.

66 Recommended Tolerances for Magnetometer Orientation and Field Calibration Procedure Joseph M. Ernst, Aaron Ault, James V. Krogmeier, and Darcy M. Bullock Magnetometers are increasingly being deployed for traffic detection in both fixed wired and wireless installations around the United States. Magnetometers respond to local magnetic field disturbances from the ambient magnetic field of the earth, and this response may be expected to depend on the relative orientation of the magnetometer axis and the direction of the earth s magnetic field lines at the point of deployment. As documented by U.S. Geological Survey (USGS) maps, the declination and inclination of the earth s magnetic field vary significantly over the continental United States, yet there is no traffic engineering design literature characterizing the effect of geographic magnetic field variation on magnetometer detection performance. This study briefly reviews the scientific theory, identifies the magnetometer installation parameters of most importance, and characterizes their impact on performance. In addition, recommendations are based on USGS data, which can be used to develop specifications for installation, testing, and acceptance of magnetometer-based detection technology. Finally, a simple procedure for field calibration of sensor orientation in a conduit to near-optimal rotation angle is presented. In recent years, magnetometers have become a viable alternative to inductive loops for vehicle detection and counting applications. Magnetometers require a less invasive installation procedure than do inductive loops because they are smaller. Microloop sensors and inductive loop sensors are somewhat similar in that their inductance changes in the presence of a vehicle. They differ in the physical phenomena that cause the inductance to change in response to a vehicle. When a conducting object appears close to an inductive loop, a mutual inductive coupling occurs due to eddy currents induced in the object. The net result is that the inductance of the inductive loop sensor is always reduced. In contrast, the inductance of a Microloop sensor can either increase or decrease in response to a perturbation in the earth s ambient magnetic field caused by the presence of a ferromagnetic object. A Microloop sensor would not respond to an all-aluminum vehicle, whereas an inductive loop sensor would. The two main methods used to install Microloops are cut and cover, similar to the installation of inductive loops, and directional J. M. Ernst, A. Ault, and J. V. Krogmeier, School of Electrical and Computer Engineering, Electrical Engineering Building, 465 Northwestern Avenue, and D. M. Bullock, School of Civil Engineering, 550 Stadium Mall Drive, Purdue University, West Lafayette, In Corresponding author: D. M. Bullock, darcy@purdue.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / boring, where the Microloop is pushed into a conduit extending under the pavement from the roadside. Directional boring can be done without disrupting traffic and is therefore the preferred method. The Microloop Operation Manual (1) and the Traffic Detector Handbook (2) recommend installation with the sensitivity axis in the vertical orientation. This orientation is difficult to ensure when the sensor is installed from the roadside through a conduit. The authors have noted variability in the sensitivity of Microloops deployed in different places, which has led to a careful investigation of the parameters that can affect performance. In this study the effects of suboptimal orientation are characterized and it is shown that the Microloop performs best when aligned with the earth s magnetic field as closely as the installation will allow. Because the orientation of a Microloop can be difficult to observe, a simple orientation tuning procedure is also proposed. RELEVANT WORK There is considerable interest and activity within the intelligent transportation systems community on the validation of data received from vehicle monitoring systems (3 5), where inductive loops, Microloops, microwave radar, and video detection are the most important technologies. As the quantity of sensors increases, the task of verifying the quality of the data becomes more cumbersome. If the data received are to be relied on, verifiable standard installation methods must be developed. As mentioned previously, the Microloop is based on a magnetoinductive effect whereby the magnitude of the projection of an external magnetic field on the sensor s axis of sensitivity is translated to a change in the inductance of the sensor. This change in inductance is measured by electronic circuitry to which the Microloop is attached. Another important magnetometer technology is the magnetoresistive type, in which the resistance changes in response to changes in field magnitude (5, 6). The commercial Microloop is a one-axis sensor, whereas commercial magnetoresistive sensors are available with three sensitivity axes (7). If all three axes of a three-axis sensor are properly processed, it would be expected that the overall performance of such a sensor would be essentially independent of its orientation with respect to that of the earth s magnetic field. A common installation method for Microloop detectors is directional boring. Detailed installation instructions can be found in chapter 5 of the Traffic Detector Handbook (2). The instructions dictate that the sensor be installed with a vertical orientation. The operating manual for the Microloop also stresses the importance of the probe s installation in a vertical position and the maintenance of the vertical position of the probe (1). The only recognition in the 66

67 Ernst, Ault, Krogmeier, and Bullock 67 literature of the importance of the direction of the earth s magnetic field to magnetometer performance is in the recommendation that they not be used near the equator (2). The direction of the earth s magnetic field varies significantly around the world and even within the continental United States. The U.S. Geological Survey (8) and the National Oceanic and Atmospheric Administration (9) have created models and maps of the magnetic field of the earth including inclination and declination angles (Figure 1). The electronic circuitry used to measure the change in inductance of either Microloops or inductive loops is based on the principle of a resonant oscillator, where the Microloop or inductive loop forms a part of the inductance of the oscillator s inductance capacitance tank. The oscillator produces a sinusoidal voltage with frequency dependent on the inductance. The oscillator frequency, which is always inversely proportional to the square root of the inductance, is measured and processed in order to detect vehicles or capture vehicle signatures. The instantaneous quantity reported by the detector is proportional to the difference between the measured oscillating frequency and an estimate of the nominal (i.e., no vehicle) oscillating frequency. Appropriately scaled, this difference is called the delta frequency. A signature is the signal formed by the delta frequencies as a function of time. EXPERIMENTAL METHOD In the following experiments, the sensitivity of the magnetometer is characterized by the peak-to-peak change in frequency (delta frequency) of the signal received from the Microloop. Measurements were taken in a controlled laboratory environment and in the field. As shown in Figure 2a, A C,TN and A C,MN are used to define the angle from true north and magnetic north to the axis of the conduit of a Microloop installation. Figure 2b shows an example road with W N E S Magnetic North True North Conduit Probe Probe A C,MN A C,TN (a) Sensor Conduit N R NS (a) N N W E S (b) (c) N W E Sensor Conduit R EW S E (d) E (e) FIGURE 1 Declination and inclination of magnetic field across the United States: (a) declination (degrees east from north) and (b) inclination (degrees down from horizontal). (b) FIGURE 2 Angle definitions for north south roads (A C,MN 90 degrees) and east west roads (A C,MN 0 degrees); north for map compass rose is magnetic north: (a) orientation angle definitions, (b) road with A C,MN 90 degrees, (c) cross-sectional view at N N, (d) road with A C,MN 0 degrees, and (e) crosssectional view at E E.

68 68 Transportation Research Record 2128 A C,MN = 90 degrees. Since the conduit is installed perpendicular to the road, the sensor has rotational freedom in the north south direction in this case. In Figure 2c, R NS is defined as the angle from the vertical. Equivalent diagrams are shown in Figures 2d and 2e for a road with A C,MN = 0 degrees. A compass was used in all experiments to locate magnetic north, since the local characteristics of sites may change the local magnetic field. Stationary Laboratory Test The stationary experimental setup is shown in Figure 3a. The Microloop was placed inside the conduit with a rotatable joint as shown in Figure 3b. This experiment was conducted with an orientation consistent with an installation in which A C,MN = 0 degrees, which corresponds to an eastbound-westbound road. Magnetic north in the laboratory was found by using a compass. R EW as defined in Figure 2e was varied from 0 degrees to 90 degrees in degree increments by rotating the conduit. The distance from the bottom of the bar magnet to the floor was also varied from 12 to 24 in. The bar magnet used was 3 in. long and 1 2 in. in diameter. Its magnetic field was measured to be approximately 20 millitesla at 14 in. For each data point, the delta-frequency output of the Microloop was first measured with no external influence and then measured with a bar magnet at the designated height. The difference between these two measurements was recorded and called the peak-to-peak delta frequency. The data points are plotted in Figure 4 and discussed in the experimental results and analysis section. Vehicle Tests Bar Magnet In the vehicle tests, a car was driven over a Microloop at various Microloop orientation angles. The site used for the vehicle tests is shown in Figure 5. A Microloop installed in a conduit can only rotate about the conduit s axis, which limits the achievable orientations. Therefore, it is not always possible to achieve optimal alignment with the earth s magnetic field at a given installation. Two representative vehicle tests were made: one for an eastbound westbound road where A C,MN = 0 degrees and one for a northbound southbound road where A C,MN = 90 degrees Microloop Rotation Vehicle Test 1 (for eastbound westbound road): A C,MN = 0 degrees In Vehicle Test 1, Microloop detector signatures were captured for a vehicle traveling over the loop where R EW was increased from 90 degrees to 90 degrees in degree increments. A signature is collected from a vehicle traveling over the sensor at each orientation. A subset of the signatures collected is shown in Figure 6. (a) Vehicle Test 2 (for northbound southbound road): A C,MN = 90 degrees Vehicle Test 2 was executed in the same fashion as Vehicle Test 1. The only change is that A C,MN = 90 degrees to correspond to a northbound-southbound road, which changes the Microloop s free axis of rotation to north and south. Therefore, R NS was varied from 90 degrees to 90 degrees in degree increments. Figure 7 shows a subset of the vehicle signatures collected. EXPERIMENTAL RESULTS AND ANALYSIS FIGURE 3 Evaluation of rotated angle on sensitivity with varying height (rotational angle R EW as defined in Figure 2): (a) experimental setup and (b) adjustable rotation at t-joint. (b) Since Microloops have sensitivity only along one axis, the measured magnetic field should be a projection of the earth s magnetic field onto this axis. Figure 4 shows the peak-to-peak change in frequency due to the presence of the bar magnet at various heights and degrees of rotation. Since data were collected at regular increments from 0 degrees to 90 degrees of Microloop rotation, quarter-wavelength sinusoidal curves were fit to the data points in Figure 4. These sinusoidal curves fit the data set at each height from 12 to 24 in. The results of the vehicle tests are summarized in Figure 8. The peak-to-peak changes in frequencies of the signatures are linearly scaled so that they best match a sinusoidal fit and are translated to

69 Ernst, Ault, Krogmeier, and Bullock 69 FIGURE 4 Summary of data from stationary test: Microloop sensitivity from point source and least-squares sinusoidal fit. Microloop (a) (b) (c) FIGURE 5 Field test site at Purdue University: (a) aerial view of field site, (b) field test site, and (c) close-up view of Microloop setup at field test site.

70 70 Transportation Research Record 2128 (a) (b) (c) (d) (e) (f) FIGURE 6 Microloop signatures at various rotational orientations for conduit installed with A C,MN 0 (approximate peak-to-peak change in frequency is shown in parentheses): (a) R EW 0 degrees (~1,400 Hz), (b) R EW 22.5 degrees (~1,300 Hz), (c) R EW 45 degrees (~1,200 Hz), (d) R EW 67.5 degrees (~500 Hz), (e) R EW 90 degrees (~100 Hz), and (f ) R EW 90 degrees (~100 Hz). have an average energy of one-half. An average sensitivity then has a value of one-half; zero sensitivity, a value of zero; and the maximum hypothetical sensitivity, a value of 1. These normalized values are then plotted against the angle of the Microloop from the vertical, R EW or R NS. Linear regression techniques were used to find the optimal phase offset of the best-fit sinusoid. Two sinusoids are plotted on each graph. One is the optimal-fit sinusoid based on the data collected, and the other assumes that vertical orientation is optimal by setting the phase offset of the sinusoid to 0 degrees. In Figure 8a the sinusoidal fit to the normalized peak-to-peak frequency shows that the empirical estimate of the optimal R EW is degrees for a sensor where A C,MN = 0 degrees. The sinusoid centered at R EW = 0 degrees represents the sinusoidal fit assuming that the vertical is optimum, which is the case where A C,MN = 0 degrees. In Figure 8b the sinusoidal fit to the normalized peak-topeak frequency shows that the empirical estimate of the optimal R NS is degrees for a sensor where A C,MN = 90 degrees. The sinusoid centered at R NS = 0 degrees represents the sinusoidal fit

71 Ernst, Ault, Krogmeier, and Bullock 71 (a) (b) (c) (d) (e) (f) FIGURE 7 Microloop signatures at various rotational orientations for conduit installed with A C,MN 90 (approximate peak-to-peak change in frequency is shown in parentheses): (a) R NS 0 degrees (~1,300 Hz), (b) R NS 22.5 degrees (~1,300 Hz), (c) R NS 45 degrees (~900 Hz), (d) R NS 67.5 degrees (~100 Hz), (e) R NS 90 degrees (~500 Hz), and (f ) R NS 90 degrees (~200 Hz). assuming that the vertical is optimum, which is not the case where A C,MN = 90 degrees. According to the National Geophysical Data Center (9), the angle of the earth s magnetic field is parallel to R NS = 21.6 degrees in Lafayette, Indiana. The optimal R EW is expected to be zero because when A C,MN = 0, there is no rotational freedom in the direction of the magnetic field. Vehicle Test 1 yielded a measured optimal R EW of degrees (Figure 8a). The error of degrees is not sufficient to reject the theoretical optimal value of 0 degrees. Vehicle Test 2 has a measured optimal R NS of degrees (Figure 8b). The error of degrees is not sufficient to reject the theoretical optimal value of 21.6 degrees Qualitatively, it is clear from Figure 8 and the raw signatures in Figures 6 and 7 that misalignment of the sensor up to 45 degrees from optimal yields sufficient sensitivity. As the optimal orientation moves farther from the vertical, installations that have assumed that the vertical is optimal will require much tighter tolerances. For example, in

72 72 Transportation Research Record 2128 Normalized delta frequency signal energy Vertical Best Fit Normalized delta frequency signal energy Best Fit Vertical Tilt (degrees) (a) Tilt (degrees) (b) FIGURE 8 Sinusoidal fit to sensitivity of Microloop for (a) R EW where A C,MN 0 degrees and (b) R NS where A C,MN 90 degrees. Lafayette, an installation with an offset from the vertical of 30 degrees will perform well if the offset is in the direction of the magnetic field because it will be only approximately 10 degrees from the optimal orientation. Unfortunately, if the sensor is 30 degrees from the vertical in the direction away from alignment with the magnetic field, that sensor will be about 50 degrees from optimal and will experience significant degradation in sensitivity. Although the difference is relatively small in Lafayette, southern installations will be more seriously affected because of the inclination shown in Figure 1b. For a Microloop to perform well, it is recommended that the degradation in sensitivity not exceed 3 db. Examples of rotational tolerances for both the vertical alignment and the magnetic field alignment can be found in Table 1. Orientations that are rotated between the minimum and maximum orientations will have sensitivity degradation of less than 3 db from optimal. Given that the optimal orientation is aligned with the magnetic field rather than the vertical, it is now possible to propose a procedure to reliably orient the Microloops in real installations. Figure 9 includes three signatures from each of the two vehicular tests (A C,MN = 0 and A C,MN = 90). The three signatures selected are the optimal orientation, optimal plus 45 degrees, and optimal minus 45 degrees. For each of the plots in Figure 9, the relevant negative deflections are designated with a circle. The negative deflections for the optimal cases in Figure 9b and e are balanced. The negative deflections of the other orientations of the sensors are unbalanced. The procedure described in the following section will equalize the negative deflections and therefore orient the Microloop to optimal. Figure 9a c has rotational freedom in the north south axis; therefore the angle from the vertical is R NS. In contrast, Figure 9d f has rotational freedom in the east west axis; therefore the angle from the vertical is R EW. The optimal angle can be found by equalizing the circled negative deflections. With an electromagnetic simulator (Maxwell 2D) the electromagnetic field was computed for a ferrous metal in a magnetic field (Figure 10a) normal to the earth s surface to represent an optimal sensor orientation and (Figure 10b) with a field rotated by 15 degrees to represent a suboptimal sensor orientation. The density of the lines designates the strength of the magnetic field. TABLE 1 Orientation Tolerances in Degrees for Microloop Installations to Remain Within 3 db of Optimal Detection Sensitivity Minimum and Maximum Orientation Angles from Vertical to Remain Within 3 db of Optimal Performance A C,TN = 0 degrees A C,TN = 45 degrees A C,TN = 90 degrees Maximum Orientation Tolerance for Arbitrary A C,TN 3 db Lower 3 db Upper 3 db Lower 3 db Upper 3 db Lower 3 db Upper Angle from Angle from City Bound Bound Bound Bound Bound Bound Vertical Optimal West Lafayette, Indiana Miami, Florida Seattle, Washington Los Angeles, California San Diego, California Baton Rouge, Louisiana Dallas, Texas New York, New York Kansas City, Missouri Phoenix, Arizona

73 Ernst, Ault, Krogmeier, and Bullock 73 (a) (b) (c) (d) (e) (f) FIGURE 9 Conduit with (a c) A C,MN 90 versus (d f ) A C,MN 0: (a) optimal R NS degrees, (b) optimal R NS degrees, (c) optimal R NS degrees, (d) optimal R EW degrees, (e) optimal R EW 0 degrees, and (f ) optimal R EW degrees. Figure 10a shows the symmetric displacement of the field if the field is vertical. This same symmetrical structure would be seen for the two-dimensional projection of the magnetic field when A C,MN = 0 degrees. Figure 10b shows the asymmetric displacement of the field, which would be seen for the two-dimensional projection of the magnetic field when A C,MN = 90 degrees. This displacement causes a field that is less strong on the right edge and stronger on the left edge. In Figure 10a it is clear that the magnetic field strength as seen by a Microloop at Point 1a will be the same as that at Point 3a, and both of these are less than the nominal field strength. Point 2a indicates the stronger magnetic field as the car is passing over the Microloop. In Figure 10b, both the angle and the density of the field lines at Points 1b and 3b are no longer the same because of the improper alignment of the sensor with respect to the magnetic field. This displacement causes the negative deflections at Points 1b and 3b to be asymmetric. Therefore, if the angle and density of the field lines are the same at Points 1 and 3, the Microloop is optimally aligned

74 74 Transportation Research Record 2128 Output of Microloop Car 0 Degrees 1a 2a 3a (a) Output of Microloop Car 15 Degrees 1b 2b 3b (b) FIGURE 10 Maxwell 2D numerically calculated fields around ferrous material: (a) vertical field representing properly oriented Microloop detector and (b) vertical field rotated by 15 degrees representing rotated Microloop detector. with the magnetic field, and if they are asymmetric, it is not optimally aligned. PROPOSED CALIBRATION PROCEDURE When Microloops are installed with the directional bore method, it can be difficult to align them properly. Even if the Microloops at the hand hold are aligned properly, the torque on the installation sled under the road could cause the Microloop orientation on the far side to be significantly different. For this reason, installations in which it is assumed that the vertical orientation is optimal have been unable to guarantee proper orientation. Given the new definition of optimal as aligned with the earth s magnetic field, the following procedure can now be proposed to facilitate the calibration of Microloop installations. With this method, the Microloop can be easily and properly oriented in the field without knowledge of the magnetic field at that location and without visual confirmation of the Microloop orientation. The installation procedure begins with the insertion of the Microloop sled into the conduit under the road. Signatures will be collected from several vehicles and analyzed. If the negative deflections are not deemed symmetrical, the Microloop orientation should be rotated 30 degrees in one direction. Several more signatures are collected and analyzed. If the difference between the negative deflections has decreased but the sign of the difference is the same (i.e., the deflection before rotation is greater than that after rotation, or vice versa), the Microloop should be rotated 15 more degrees in the same direction. If the sign of the difference is not the same, the optimal angle is between the first and

75 Ernst, Ault, Krogmeier, and Bullock 75 the second angles, so the Microloop should be rotated 15 degrees in the opposite direction. If the difference in deflections worsens, the Microloops should be rotated 60 degrees in the opposite direction (30 degrees backward from the original orientation). The procedure continues in this fashion, decreasing correction angles until the negative deflections are acceptably symmetric. The Microloop is now at near-optimal orientation for that installation. CONCLUSION The Microloop s less invasive installation procedure is attractive, but it also makes verifying the correct orientation of the sensor difficult. In recent years, Microloops have demonstrated variability in sensitivity between installations (3, 4). Current installation procedures assume that optimal orientation is always the vertical, but Figure 8b shows that this is not the case. The effect of variation in orientation on the sensitivity of the Microloop was investigated and it was shown that optimal Microloop orientation is with its axis of sensitivity parallel to the earth s magnetic field, which varies geographically. Orientations up to approximately 45 degrees from the optimal still yield reasonable signatures. On the basis of the electromagnetic theory outlined here, Table 1 gives several examples of the thresholds for rotation both from the vertical and the optimal that will lose less than 3dB of the optimal signal. Because of the nature of Microloop installation, rotational freedom exists only along one axis. For installations on an eastboundwestbound road where A C,MN = 0 degrees, the closest orientation to parallel with the magnetic field is R EW = 0 degrees. For installations on a northbound-southbound road where A C,MN = 90, the sensor can be set exactly parallel with the earth s magnetic field because of the definition of magnetic north. The residual effects of the rotational offset of the Microloop cause clear, asymmetric, negative deflections, as shown in Figure 9. By viewing the signatures from the Microloop when vehicles travel over the detection area, the optimal rotation can easily be found by turning the Microloop until the negative deflections are symmetrical. With this new definition of optimal orientation, a clear, simple, repeatable calibration procedure is provided that can be executed with minimal changes to existing construction procedures and existing infrastructure with no new equipment required. Example tolerances for both the vertical and optimal installation methods are provided for various U.S. cities. ACKNOWLEDGMENTS This work was supported by the Joint Transportation Research Program administered by the Indiana Department of Transportation and Purdue University. REFERENCES 1. Canoga Vehicle Detection System Model 701 Microloop Operation Manual. 3M Canada Company, London, Kell, J., I. Fullerton, and M. Mills. Traffic Detector Handbook, 2nd ed. FHWA, U.S. Department of Transportation, Wells, T. Health Monitoring Procedures for Freeway Traffic Sensors. MSCE thesis. Dec Achilledes, C., and D. M. Bullock. Performance Metrics for Freeway Sensors. Report FHWA/IN/JTRP-2004/37. Joint Transportation Research Program, Dec Turner, S. Defining and Measuring Traffic Data Quality. Presented at Traffic Data Quality Workshop, Dec jpodocs/repts_te/13767.html. 6. Lenz, J. E. A Review of Magnetic Sensors. Proc. IEEE, Vol. 78, No. 6, June 1990, pp Cheung, S.-Y., S. Coleri Ergen, and P. Varaiya. Traffic Surveillance with Wireless Magnetic Sensors. Proc., 12th ITS World Congress, San Francisco, Calif., Nov U.S. Geological Survey National Geophysical Data Center, National Oceanic and Atmospheric Administration. The contents of this paper reflect the views of the authors, who are responsible for the facts and the accuracy of the data, and do not necessarily reflect the official views or policies of the Federal Highway Administration or the Indiana Department of Transportation. The contents do not constitute a standard, specification, or regulation. The Traffic Signal Systems Committee sponsored publication of this paper.

76 Optimization of Coordinated Actuated Traffic Signal System Stochastic Optimization Method Based on Shuffled Frog-Leaping Algorithm Byungkyu (Brian) Park and Joyoung Lee Study has shown that about half of the traffic signalized intersections in the United States need to be reoptimized. Even though many traffic signal control systems have been upgraded from pretimed controllers to actuated and adaptive controllers, traffic signal optimization software has not kept up with such advances. For example, existing commercial traffic signal timing optimization programs such as SYNCHRO and TRANSYT-7F do not optimize advanced controller settings (e.g., minimum green time, extension time, detector settings) available in modern traffic controllers. This deficiency is due partly to existing programs, which are based on macroscopic simulation tools that do not explicitly consider individual vehicular movements. To overcome such a shortcoming, a stochastic optimization method (SOM) was proposed and successfully applied to a few case studies. This study presents some enhancements made in the SOM and case study results from an arterial network consisting of 16 signalized intersections. The proposed method employs a distributed computing environment for faster computation time and uses a shuffled frog-leaping algorithm (SFLA) for better optimization. The case study results showed that the proposed enhanced SOM method, called SFLASOM, improved total network travel times over field settings by 3.5% for midday times and 2.1% for p.m.-peak times. In addition, corridor travel times for both p.m.-peak and midday times were significantly improved. Of the approximately 300,000 signalized intersections in the United States (1), 56% urgently need reoptimization of their traffic signal timing plans (2). Further, traffic engineers scored themselves a grade of D on their traffic signal operations performance (3). Many factors have affected such a low score including insufficient staff, less-thandesired funding levels, and lack of adequate optimization software, among others. Unfortunately, none of the existing traffic signal optimization programs such as SYNCHRO 6 (4), TRANSYT-7F (5), and PASSER V (6) explicitly optimize advanced controller settings, including minimum green time, extension time, detector recall mode, and so on. This deficiency is because the fidelity of the traffic simulations used in these programs is not adequate to optimize these advanced settings. Department of Civil and Environmental Engineering, Thornton Hall B228, University of Virginia, 351 McCormick Road, P.O. Box , Charlottesville, VA Corresponding author: B. Park, bpark@virginia.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / Although basic parameters such as cycle length, offsets, and maximum green times can be easily obtained from existing tools, traffic engineers have to use default parameters for these advanced settings or choose parameters based on trial-and-error methods. In addition, the existing tools, which are based on macroscopic simulation models, are unable to account for individual vehicular driving behaviors. The two key components of traffic signal optimization are (a) the adequacy of the simulation tools used in the evaluation of timing plans and (b) the quality of the optimization method used. Thus, microscopic simulation models, once properly calibrated, can be very effective in modeling individual vehicular driving behaviors. A number of studies (7 11) have already demonstrated that the stochastic optimization method (SOM), which is based on a microscopic traffic simulation model and an optimizer, can be effective in optimizing advanced controller settings. The focus of this study is to explore the possibility of improving computation time by adopting a distributed computing environment (DCE) and obtaining a better solution with a new optimization technique. LITERATURE REVIEW Park et al. (12, 13) introduced the application of metaheuristic algorithms to traffic signal optimization methods. They applied a genetic algorithm (GA) as an optimizer to find the best timing plan for oversaturated traffic conditions, a problem never before solved by an analytical model. The study showed that the GA outperformed the existing commercial traffic signal optimization program, TRANSYT-7F. Park et al. (7) integrated a metaheuristic algorithm and a microscopic traffic simulation model to consider the variability in traffic due to drivers characteristics and other characteristics and to achieve better optimization results. They developed and applied a GA-based signal SOM (GA-SOM) to a coordinated pretimed signal corridor with CORSIM (14). The study clearly demonstrated that the GA-SOM outperformed TRANSYT-7F in every performance measure delays, queue times, and throughputs. Park and Schneeberger (8, 9) applied a GA-SOM to optimize the offsets of a coordinated-actuated signal corridor consisting of 11 intersections in Virginia. In this study, VISSIM (15) was employed as a microscopic-level evaluator. In addition, they expanded the SOM technique to the calibration of microscopic simulation models and emphasized the importance of appropriate calibration methods in traffic simulation work. 76

77 Park and Lee 77 Yun and Park (10) developed and applied a SOM for traffic signal controller setting optimization using CORSIM as an evaluator. A GA-based optimizer found optimal values for both traditional traffic signal settings (e.g., cycle length, offset, green splits, phase sequences) and the advanced controller settings mentioned in the previous section. The study showed that the proposed SOM improved the internal links average delay time by 40% over the base condition, whereas SYNCHRO improved average delay time by approximately 25% over the base condition. Stevanovic et al. (11) developed a state-of-the-practice tool for a signal SOM called VISSIM-based GA optimization of signal timings (VISGAOST). Although SOM programs developed earlier were not user friendly, VISGAOST included a graphical user interface that provided increased convenience for users. In addition, VISGAOST used a DCE to reduce simulation run times. Although evaluation studies performed on an actual network in Park City, Utah, demonstrated that VISGAOST and SYNCHRO optimized the field settings equally in three different traffic demand scenarios, SYNCHRO did not optimize the advanced controller settings. ENHANCEMENTS The previously developed SOM used metaheuristic algorithms such as the GA, tabu search algorithm, and simulated annealing. Among these, the GA outperformed other algorithms in finding an optimal solution (16). Thus, the GA was considered as an optimizer at the initial stage of this study. However, it was frequently observed that the GA optimizer was not able to obtain a solution better than the existing timing plan for the case study network consisting of 16 signalized intersections. In addition, the amount of time required to conduct the search was extremely lengthy. Even after more than 20 days worth of GA optimization runs, a better solution was not achieved. The remainder of this section discusses proposed enhancements made over the existing SOM program. Shuffled Frog-Leaping Algorithm A shuffled frog-leaping algorithm (SFLA) (17), an evolutionary algorithm, was developed recently and successfully applied to several optimization problems in a few engineering applications (18, 19). The SFLA is based on the evolution of memes carried by individuals and a global exchange of information within the population (17). The procedure itself is quite simple: it determines the next position by searching a direction given by the best and worst vectors in each memeplex. If a new solution is better than the worst one, the new solution replaces the worst one. If not, a randomly generated solution replaces the worst one. Figure 1 shows a conceptual framework of the SFLA. In each memeplex, the performances of individual frogs, which act as chromosomes in the GA, are examined, and the frog showing the best performance is selected. The best frog then carries out a local search by exchanging its information with the frog showing the worst performance within the same memeplex. The exchanging scheme is formulated as follows: ( ) X = X + α X X () 1 n b b w where the symbols are as defined in Figure 1. If the performance of the new frog is better than that of the worst frog, the new frog replaces the worst one. Otherwise, a brand new frog, which is randomly created, replaces the worst one. These local searches are simultaneously and independently performed in each memeplex. Once the local searches are complete, the global best frog over all memeplexes is selected and exchanges its information with the worst frogs in each memeplex: ( ) X = X + α X X ( 2) n g g w The same replacing procedure as in the local search is implemented, and this task is called the global search. The memeplexes are re-created after both local and global searches, and the same procedures are replicated until a stopping criterion is satisfied. Figure 2 (18, 19) explains the procedural framework of the SFLA. It was observed that the SFLA found a better solution with fewer evaluation numbers than the GA. Unlike the conventional binary GA, each frog (X), which forms a vector, in the SFLA consists of any arbitrary real numbers: X f0, f1, f2, f4,..., f n ( 3) = [ ] where f is an arbitrary real value. For the purpose of traffic signal timing plan optimization, the frogs in this study were designed to follow the fraction-based decoding scheme that was employed in previous Local Search 1 X n = X b + α (X b X w ) Xb & Xg Xn Replacing Search direction Global Search X n = X g + α (X g X w ) Xw Search direction Xb New Xb & Xg Xg : The global best frog Xw : The worst frog Xb : The best frog Xn : A new frog : Memeplex 1 : Memeplex 2 α : Random number(0~1) Xw Search direction Local Search 2 X n = X b + α (X b X w ) FIGURE 1 Conceptual framework of SFLA.

78 78 Transportation Research Record 2128 Local a Start Population size (p) Number of memeplexes (m) Iterations within each memeplex (it) m=m+1 m=0 it=0 it=it+1 Determine X b, X w, and X g Generate population (p) randomly Apply Equations (1) and (2) Evaluate the fitness of (p) Sort (p) in descending order Partition p into m memeplexes Yes Is new frog better than worst? No Apply Equations (1) and (2), replacing X b by X g Local search a Shuffle the memeplexes b Convergence criteria satisfied No Yes Is new frog better than worst? No Generate a new frog randomly Replace worst frog Yes Determine the best solution No m=no. of memeplexes Yes it=no. of iterations No End Yes b FIGURE 2 Procedural framework of SFLA (18, 19). studies (12, 13) for signal parameters (i.e., cycle length, offsets, maximum and minimum greens, and vehicle extension times). Each frog consists of n number of real values as in Equation 3. Distributed Computing Environment Given that the SOM employs a microscopic simulation model for the evaluation of traffic signal control settings during the optimization runs, the amount of time required for optimization is not trivial. For example, for an optimization that requires 200 independent evaluation runs with five replications each at 100 s per replication, the total simulation time required would be more than 27 h. If two computers were used at the same time, the run time would theoretically be reduced by half. This study employed a DCE to reduce the computation time and obtained the results within a practical time period. The conceptual framework of the master slave-type DCE is shown in Figure 3. The master of the DCE ensures communications between the master and the slaves, which are dedicated for simulation runs, and assigns a new simulation job to an idling slave. Evaluation results are transmitted back to the master for further processing. Master Slave Processor#1 Processor#2 FIGURE 3 Conceptual framework of DCE.

79 Park and Lee 79 TABLE 1 Time Saved by DCE Number of Slaves in DCE Elapsed time (min:s) 70:37 36:53 25:56 18:50 14:38 Time saving (%) 33:44 (47.7) 44:41 (63.3) 51:47 (73.3) 54:59 (77.8) Table 1 presents experimental results demonstrating the time savings of simulation evaluations under the DCE. The experiment was implemented with a single master and up to five slaves, which all used Intel Pentium 3.6-GHz processors and were connected via Internet Protocol addresses. The elapsed time to run 50 VISSIM replications was measured for each number of slaves in the DCE settings. The results show that the time savings increased as the number of slaves in the DCE increased. As expected, the rate of time savings gradually diminished, partly because of the communication time lags between the master server and the slave computers. Nonetheless, with a DCE, one can achieve significant computational time savings. Stand-Alone Application The previous SOM program developed by Yun and Park (10) was implemented within a MATLAB programming environment (20). Although MATLAB has several advantages (e.g., programming ease, high-performance numerical computing, reliable random number generator), it requires MATLAB to implement the SOM program. In other words, the program could not run without a computer installed with the MATLAB program. In order for the enhanced SOM program to run as a stand alone (i.e., without the MATLAB program), the core modules of the previous SOM program were recompiled by MATLAB.NET Builder (20) and transformed into external dynamic linked libraries (DLLs). These DLLs were fed into a single stand-alone program, SFLASOM, written by Microsoft.NET# (21). In addition, the SFLA and DCE were fed into SFLASOM. each of the midday and p.m.-peak periods, and five replications were made for each sample. The LHD is an experimental design approach that achieves maximum coverage of the vector space defined by the chosen calibration parameters and their levels by ensuring minimum correlations among factors considered. It has been used in a few previous studies (24, 25) and has been shown to reduce the number of combinations to a reasonable level while still adequately covering the entire parameter surface. Figure 4b shows the entire study corridor modeled in VISSIM. Figure 5 shows an x-y plot of simulated eastbound and westbound travel times from the 200 LHD samples, where each point represents the average of five replications. Depending on the parameter sets used in the experimental design, the ranges of simulated travel times vary significantly. The rectangular boxed areas indicate field-measured travel times for each direction of travel. The parameter sets within the rectangular box represent field conditions. Among the eight parameter sets within the box, the best parameter set was selected by comparing individual link travel times. The eight selected parameter sets were ordered according to how close they were to the mean of the field travel times (dashed lines) for each direction. The travel times of each link were also examined to compare with the observed link travel times. Figures 6 and 7 show the comparisons between the simulated and measured westbound link travel times for the best parameter sets. The link travel times appear to follow a bimodal distribution: those who went through during the green versus those who did not make the green and waited for the next one. It is apparent that the observed travel times (marked with down arrows) match up well with the histograms obtained from the best parameter set. IMPLEMENTATION Study Corridor and Scope A 6-mi-long section with 16 signalized intersections along Route 50 in Northern Virginia was selected. The study corridor located between the intersections of Pleasant Valley and Rugby Roads is being operated by coordinated actuated traffic signal control. The proposed method was implemented for both midday and p.m. peak periods. Figure 4a shows the aerial view of the study corridor (22). Data Collection and Model Calibration Traffic volume data were collected by video recordings at the entry and exit points of the study corridor. Turning movements and existing signal timing plans along the study corridor were provided by the Northern Region Operation (NRO) database. Seven probe vehicles equipped with Global Positioning System devices collected the travel time data for the entire corridor as well as for each link during the data collection periods. The study corridor was modeled in VISSIM. After fine tuning of the VISSIM simulation network, it was calibrated using the Latin hypercube design (LHD) approach (23) with 200 samples for Simulation Model Validation The best parameter set was evaluated with a different signal timing plan to validate its performance. The bars in Figure 8a and b show the frequencies of individual vehicles corridor travel times obtained from the VISSIM simulations for the new timing plan, and arrows indicate the observed field corridor travel times collected by the four probe vehicles. It can be seen that all collected travel times were within the range of travel times predicted by the simulation model. Taking the actual corridor travel time variations into consideration, the results clearly show that the calibrated VISSIM model adequately predicts the performance of the new field timing plan. Thus, the new SOM evaluation results obtained from VISSIM are considered to be reliable. Performance Comparisons Evaluation Scenarios Evaluations were performed for three scenarios SYNCHRO, field, and SOM for both midday and p.m.-peak periods. In the SYNCHRO scenario, the optimal timing plan was obtained from the SYNCHRO program. The field scenario represents the current timing plan used in

80 80 Transportation Research Record 2128 (a) (b) FIGURE 4 Study corridor: (a) aerial photograph (22) and (b) VISSIM network. the study corridor. It is noted that NRO staff optimizes its traffic signal timing plans by using SYNCHRO and then adjusts the timing plan on the basis of simulations and field observations. Thus, the field timing plan generally outperforms the SYNCHRO timing plan. Finally, the SOM represents a timing plan optimized by the proposed approach based on the SFLA with the following algorithm settings: Total number of individual frogs: 200 for midday and 250 for p.m. peak, Total number of memeplexes: 20 for midday and 25 for p.m. peak, Number of iterations within a memeplex: 1, and Number of replications per memeplex: 15. All scenarios were evaluated with VISSIM using emulated National Electrical Manufacturers Association signal controllers (15) embed- ded in VISSIM. The performance of each scenario was determined by two measures of effectiveness: vehicle hours traveled (VHT) and directional corridor travel time. VHT was selected because it considers the entire system performance, including those vehicles remaining on the network at the end of the simulation run. Directional corridor travel time was chosen because it considers progression along the corridor. These measures were obtained from the well-calibrated VISSIM simulations. To account for day-to-day variations in traffic demand, each scenario was replicated 50 times, and the distribution of measures of effectiveness was used in the comparisons. Table 2 summarizes the evaluation scenarios. The minimum green times, minimums of maximum, and vehicle extension times under the SOM were optimized. The minimums of maximum greens depend on the minimum greens. The optimized traffic signal timing plans were carefully investigated and no anomalies were found.

81 Park and Lee 81 Simulated Westbound Travel Time (Sec) Eastbound Field Travel Time Westbound Field Travel Time Simulated Eastbound Travel Time (Sec) 2210 FIGURE 5 Identification of best parameter set. System Performance Figure 9 shows the evaluation results of each scenario for both the midday and p.m.-peak periods. As illustrated, the traffic signal timing plans obtained from the SOM outperformed those obtained from SYNCHRO and the field. Under midday traffic conditions, the SOM timing plan improved the total network travel time by 7.5% and 3.5% over the SYNCHRO and field timing plans, respectively. In addition, under p.m.-peak traffic conditions, system performance using the SOM timing plan improved by 8.8% and 2.2% over the SYNCHRO and field timing plans, respectively. These performance benefits were all statistically significant at the 95% Frequency FIGURE 6 Midday westbound link travel times.

82 82 Transportation Research Record 2128 Frequency FIGURE 7 Westbound link travel times at p.m. peak Frequency Frequency WB Travel Time (Sec) (a) EB Travel Time (Sec) (b) FIGURE 8 Corridor travel time by simulation (bars) and by probe vehicles (arrows): (a) westbound and (b) eastbound.

83 Park and Lee 83 TABLE 2 Evaluation Scenario Descriptions Scenario Parameter SYNCHRO Field SOM Cycle length (s) Fixed (MD = 150, p.m. = 200) Fixed (MD = 150, p.m. = 200) Fixed (MD = 150, p.m. = 150) Offset Optimized Tweaked from SYNCHRO Optimized Max green Optimized Tweaked from SYNCHRO Optimized Min green Fixed Fixed (major street:15,20 s) (major street:15,20 s) Optimized (cross street:5 10 s) (cross street:5 10 s) Minimum of max green Fixed (10 s) Fixed (10 s) Optimized Vehicle extension time Fixed (4 7 s) Fixed (4 7 s) Optimized Amber and red clearance Fixed (5 7 s) Phase sequence NOTE: MD = midday; p.m. = p.m. peak. Fixed (current field setting) FieldSetting SY NCHRO Optimized Frequency VHT (a) Frequency FieldSetting(C=200) SYNCHRO(C=150) Optimized(C=150) 5 0 FIGURE VHT (b) Evaluation results of VHT for (a) midday and (b) p.m. peak.

84 84 Transportation Research Record 2128 TABLE 3 Summary of System Performance VHT Mean Standard Scenario (50 evaluations) Deviation Midday SOM Field SYNCHRO p.m. peak SOM 1, Field 1, SYNCHRO 1, confidence level. The benefits of the SOM are summarized in Table 3. The evaluation results for both midday and p.m.-peak conditions obtained from field settings were better than those from SYNCHRO. It is of interest to observe that the SOM outperformed the field timing plans, which indicates that the systematic approach used in the SOM was more effective than tweaks made in the field over SYNCHRO timing plans. Corridor Travel Time Westbound and eastbound travel times along the study corridor were also examined, comparing the field and SOM timing plans. The corridor travel times of SYNCHRO were not presented because the field settings outperformed the SYNCHRO settings as demonstrated in the system-level performance evaluations. Figure 10 shows the distributions of the corridor travel times for each direction during midday and p.m.-peak periods. Overall, the SOM timing plans outperformed the field timing plans. The average travel times under the SOM and field timing plans were significantly different at the 95% confidence level. Further, when the standard deviation results were compared, the SOM timing plans proved to be more robust than the field timing plans, as shown in Table 4. Under p.m.-peak traffic conditions, the SOM timing plan showed a much narrower distribution, indicating more reliable performance with respect to variations in day-to-day traffic demand. In addition, the evaluation results of the corridor travel times were consistent with those of network performance. CONCLUSIONS This study improved the previously developed SOM program by adopting an SFLA within a DCE and constructing a stand-alone program. The study showed that the DCE can significantly improve the computation time needed for the proposed SOM optimization. In addition, the improved performance of the SFLA was in part due to its nature of search, which combines local search as well as global search without requiring additional evaluations that would have been needed in the GA method. The performance of the enhanced SOM was investigated by using multiple simulation runs along a corridor with 16 signalized intersections. Compared with the current field settings, the SOM timing plan improved the total network travel time by 3.5% and 2.1% under midday and p.m.-peak traffic conditions, respectively. In addition, when the SOM timing plan is compared with the SYNCHRO timing plan, travel time savings were 7.5% and 8.8% under midday and p.m.-peak traffic conditions, respectively. In the performance of corridor travel time, the enhanced SOM outperformed the field settings, resulting in improvements of 16.9% and Frequency Field(WB) Optimized(WB) Frequency Field(EB) Optimized(EB) Westbound Travel Time (Sec) (a) Eastbound Travel Time (Sec) Frequency Field(WB) Optimized(WB) 880 Westbound Travel Time (Sec) 960 Frequency (b) Field(EB) Optimized(EB) 880 Eastbound Travel Time (Sec) 960 FIGURE 10 Evaluation of corridor travel time under traffic conditions at (a) midday and (b) p.m. peak.

85 Park and Lee 85 TABLE 4 2.3% for westbound and eastbound travel times, respectively, under midday traffic conditions. In addition, the SOM timing plan outperformed the field timing plan by 9.6% and 17.9% for westbound and eastbound travel times, respectively, under p.m.-peak traffic conditions. On the basis of these results, it can be concluded that the SOM approach is much more efficient than the current state-of-the-practice approach. Even though the SOM-based traffic signal timing plan optimization requires much more computation time and resources (e.g., this study required about 10 consecutive days with seven slave computers) than traditional methods, this deficiency could be overcome with future advances in computation technology. REFERENCES Comparison of Corridor Travel Times Standard Period Measure Scenario Mean (s) Deviation Midday Westbound Field travel time SOM Gain (%) (16.9%) Eastbound Field travel time SOM Gain (%) 17.1 (2.3%) p.m. peak Westbound Field travel time SOM Gain (%) 69.6 (9.6%) Eastbound Field travel time SOM Gain (%) (17.9%) Urban Mobility Report. Texas Transportation Institute, Texas A&M University System, College Station, Temporary Losses of Highway Capacity and Impacts on Performance: Phase 2. Report ORNL/TM-2004/209. Oak Ridge National Laboratory, Oak Ridge, Tenn., Nov The National Traffic Signal Report Card. Technical Report National Transportation Operations Coalition, Washington, D.C., Husch, D., and J. Albeck. Trafficware SYNCHRO 6 User Guide. Trafficware, Albany, Calif., Hale, D. Traffic Network Study Tool TRANSYT-7F, United States Version. McTrans Center, University of Florida, Gainesville, Jan PASSER V. Texas Transportation Institute, Texas A&M University System, College Station, Park, B., N. M. Rouphail, and J. Sacks. Assessment of Stochastic Signal Optimization Method Using Microsimulation. In Transportation Research Record: Journal of the Transportation Research Board, No. 1748, TRB, National Research Council, Washington, D.C., 2001, pp Park, B., and J. D. Schneeberger. Microscopic Simulation Model Calibration and Validation: Case Study of VISSIM Simulation Model for a Coordinated Actuated Signal System. In Transportation Research Record: Journal of the Transportation Research Board, No. 1856, Trans- portation Research Board of the National Academies, Washington, D.C., 2003, pp Park, B., and J. D. Schneeberger. Evaluation of Traffic Signal Timing Optimization Methods Using a Stochastic and Microscopic Simulation Program. Research Report UVACTS University of Virginia, Charlottesville, Jan Yun, I., and B. Park. Stochastic Optimization for Coordinated Actuated Traffic Signal Systems. Presented at 88th Annual Meeting of the Transportation Research Board, Washington, D.C., Stevanovic, A., P. T. Martin, and J. Stevanovic. VISSIM-Based Genetic Algorithm Optimization of Signal Timings. In Transportation Research Record: Journal of the Transportation Research Board, No. 2035, Transportation Research Board of the National Academies, Washington, D.C., 2007, pp Park, B., C. J. Messer, and T. Urbanik II. Traffic Signal Optimization Program for Oversaturated Conditions: Genetic Algorithm Approach. In Transportation Research Record: Journal of the Transportation Research Board, No. 1683, TRB, National Research Council, Washington, D.C., 1999, pp Park, B., C. J. Messer, and T. Urbanik II. Enhanced Genetic Algorithm for Signal-Timing Optimization of Oversaturated Intersections. In Transportation Research Record: Journal of the Transportation Research Board, No. 1727, TRB, National Research Council, Washington, D.C., 2000, pp CORSIM Reference Manual: Version 5.1. ITT Industries Inc., Systems Division and ATMS R&D and Systems Engineering Program Team, Colorado Springs, VISSIM 4.20 User Manual. Planung Transport Verkehr, Karlsruhe, Germany, Yun, I. Stochastic Optimization Method for Coordinated Actuated Signal Systems. PhD thesis. University of Virginia, Charlottesville, Eusuff, M. M., and E. Lansey. Optimization of Water Distribution Network Design Using the Shuffled Frog Leaping Algorithm. Journal of Water Resources Planning and Management, ASCE, Vol. 129, No. 3, 2003, pp Elbehairy, H., E. Elbeltagi, T. Hegazy, and K. Soudki. Comparison of Two Evolutionary Algorithms for Optimization of Bridge Deck Repairs. Computer-Aided Civil and Infrastructure Engineering, Vol. 21, No. 8, 2006, pp Elbeltagi, E., T. Hegazy, and D. Grierson. A Modified Shuffled Frog- Leaping Optimization Algorithm: Applications to Project Management. Structure and Infrastructure Engineering, Vol. 3, No. 1, March 2007, pp MATLAB and SIMULINK. MathWorks, Inc., Schmidt, M., and S. Robinson. Microsoft Visual C#.NET 2003: Developer s Cookbook. Sams Publishing, Google Maps McKay, M. D., R. J. Beckman, and W. J. Conover. A Comparison of Three Methods for Selecting Values of Input Variables in the Analysis of Output from a Computer Code. Technometrics, Vol. 21, No. 2, 1979, pp Park, B., J. Won, and I. Yun. Application of Microscopic Simulation Model Calibration and Validation Procedure: Case Study of Coordinated Actuated Signal System. In Transportation Research Record: Journal of the Transportation Research Board, No. 1978, Transportation Research Board of the National Academies, Washington, D.C., 2006, pp Park, B., and J. Won. Simulation Model Calibration and Validation: Phase II Development of Implementation Handbook and Short Course. Research Report FHWA/VTRC 07-CR5. Virginia Transportation Research Council, Charlottesville, Oct The Traffic Signal Systems Committee sponsored publication of this paper.

86 Bicyclist Intersection Crossing Times Quantitative Measurements for Selecting Signal Timing Steven E. Shladover, ZuWhan Kim, Meng Cao, Ashkan Sharafsaleh, and Jing-Quan Li In support of efforts to improve traffic signal timing to accommodate bicyclists needs, observations were made of the timing of bicyclists intersection crossing maneuvers. Video recordings were made of bicyclists crossings and the video images were processed to extract the bicyclists trajectories. These trajectories were synchronized with video images of the traffic signals so that the timing of the bicyclists maneuvers could be determined relative to the signal phases. The processed data yielded cumulative distributions of the crossing speeds of bicyclists who did not have to stop at the intersection and the start-up times and final crossing speeds of the bicyclists who had to cross from a standing start. A unique feature of these data is the timing information relative to the traffic signal, which is used to define recommended signal times to permit most bicyclists to cross wide arterial intersections safely. In October 2007, California Assembly Bill 1581 amended the California Vehicle Code, stating that bicyclists and motorcyclists are legitimate users of roadways in California and requiring trafficactuated signals to be installed and maintained so as to detect lawful bicycle or motorcycle traffic on the roadway. AB 1581 also requires the California Department of Transportation (Caltrans) to establish uniform standards, specifications and guidelines for the detection of bicycles and motorcycles by traffic-actuated signals and related signal timing. This legislation stimulated the need for Caltrans to determine how to specify the minimum green signal intervals throughout the state to give bicyclists sufficient time to cross wide arterials from a standing start. Of course, any increase in the green time allocated to minor cross streets has to be taken away from the time available for mainline traffic on the arterial, so a trade-off must be made between the needs of the mainline drivers and the needs of the crossing bicyclists. The current default minimum green interval is 4 s in California, but there are precedents for increasing this interval at wide intersections with heavy bicyclist traffic, especially when those bicyclists include school children. If bicyclists could be detected reliably at intersection approaches and clearly distinguished from motor vehicles, it would be possible to modify the signal timing so that the bicyclists would receive longer green S. E. Shladover, Z. W. Kim, A. Sharafsaleh, and J.-Q. Li, California PATH Program, University of California, Berkeley, 1357 South 46th Street, Building 452, Richmond, CA M. Cao, Department of Electrical Engineering and CE-CERT, University of California, Riverside, Corresponding author: S. E. Shladover, steve@path.berkeley.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / times than the motor vehicles. In this way, the longer green intervals would only be triggered when bicyclists are present, minimizing the negative effects on mainline green time. Unfortunately, although bicycle classification is possible with state-of-the-art image-processing technology, it is not yet sufficiently reliable or mature for practical deployment in the field, so bicyclists and motor vehicles will be receiving the same green signal time for the foreseeable future. This situation increases the urgency of determining the minimum length of green that is really needed by bicyclists, since this green interval will have to apply all the time, with or without bicyclists. The alternative available for bicyclists to actuate a green signal to cross a busy arterial is to push the pedestrian call button (when one is available). This alternative provides a significantly longer green time than bicyclists should need, since it has to be set on the basis of pedestrian walking speed, incurring a significant efficiency penalty on the mainline traffic. NEW INFORMATION NEEDED In order to make intelligent decisions about the length of green intervals to accommodate bicyclists, several kinds of information are needed: Time for bicyclists to start up from a stop after their signal turns green; Speed of bicyclists crossing a wide arterial once they have reached their cruising speed; Dependence of bicyclist crossing times on street width, road geometry details (grade, arterial crown), and bicyclist demographics (especially age); and Effects of shortened green time on mainline traffic conditions, with both coordinated and independent signals. Unfortunately, little such information is available in the literature, which meant that it was necessary to create much of that information for this project. RELEVANT PRIOR LITERATURE It is not easy to collect quantitative microscopic data about bicycling, and only limited resources have been available to support efforts to collect such data. This shortcoming has tended to limit the breadth and depth of the information available in the literature. The most directly relevant prior work is the 1995 review paper by Wachtel et al., which summarized a variety of preceding studies to point toward signal timing recommendations (1). The data cited directly in the paper are relatively limited and are described only in 86

87 Shladover, Kim, Cao, Sharafsaleh, and Li 87 terms of mean crossing times and speeds. In the absence of more detailed data it is difficult to know how to apply the findings more widely. The second relevant work is the 2005 paper by Rubins and Handy (2), which summarizes an extensive set of bicyclist crossing observations in Davis, California. These data, which represent total crossing times at intersections of varying widths, show significant scatter, making it challenging to use them as the basis for specific signal timing recommendations. Regressions showed the mean value of crossing time as a function of crossing width, and the authors derived estimates of fixed start-up time and crossing speed from these ensemble statistics. They focused on 2nd- and 15th-percentile statistics for formulating their recommendations, but the full distributions of crossing times were not reported, which makes it impossible to apply other percentile criteria based on their data. NEED FOR NEW DATA COLLECTION The decision about increasing minimum green intervals to accommodate bicyclists could have significant impacts on arterial traffic throughout California. Therefore, the state wanted to ensure that any such decision is based on current, accurate, and transparent data about the behavior of bicyclists. It was particularly important to understand the relationship between bicyclist crossing behavior and traffic signal phase, including when bicyclists start and complete their crossings relative to the green, yellow, and red signal onsets. Since none of the existing published studies on bicycling behavior included synchronization with traffic signals or comprehensive data in which all the dependencies could be explored, it was necessary to collect a new set of data. Through use of modern data collection methods, it was also possible to provide a level of detail and accuracy substantially beyond that of prior studies, which were based on manual tabulation of bicyclist crossing times using stopwatches. DATA COLLECTION AND PROCESSING APPROACH Data Collection System The quantification of bicyclist crossing behavior relied on digital video recording and offline video image processing of the bicyclist movements. A video camera was mounted on top of a mast at a height of 6 to 7 m, atop a trailer parked near the subject intersection. The location of the trailer was chosen to provide a clear view of the bicyclist crossing path, with the camera height selected to be high enough that passing vehicles could not occlude the view of the bicyclists. A second video camera was mounted at a height of 2 to 3 m, with a view of the traffic signal head governing the bicyclist crossing movement. An embedded computer (PC/104) equipped with an MPEG hardware compression board was used to convert the analog video signals into pairs of MPEG video clips. The video clips were synchronized to have less than 33 ms (one scan) error. Selection of Host Sites The initial criteria for selection of host sites for data collection were the following: High volume of bicyclist traffic, Diversity of bicyclist population across sites, Moderate to wide intersection, and Available location for mounting video camera in an adjoining building or on a parked trailer. The two intersections that were finally selected for the data collection reported here were El Camino Real and Park Boulevard in Palo Alto, which was recommended by the city of Palo Alto as the crossing with the highest bicyclist volume there, and Telegraph Avenue and Russell Street in Berkeley, where the Russell Street bicycle boulevard crosses the major arterial of Telegraph Avenue. The Palo Alto site is at one of the entrances to the Stanford University campus and therefore serves primarily Stanford students and employees, generally a population of young adults. The Berkeley site provides a more diverse population of bicyclists, including school-age children and more mature adults, in a more residential area. The characteristics of the two intersections are summarized in Table 1. Data Processing After the video data were recorded in the field, they were analyzed in the laboratory to extract the information of interest. The imageprocessing system automatically detects moving objects in the video and generates their trajectories (3). There are many challenges for fully automated detection, such as poor image resolution on the far side of the image and difficulty in discriminating bicyclists from pedestrians and other vehicles. To ensure error-free detection, an interactive image-processing tool [introduced by Kim (3)] was used. An operator manually confirmed, repositioned, or manually marked the initial position of each bicyclist. The tracking was then done automatically, but any visible tracking errors were fixed manually with the interactive tool. A screen shot of the tool, with the imageprocessing system s symbols identifying the bicyclists with a thick ellipse, is shown in Figure 1. The superimposed image of the traffic signal from the second video camera is visible in the upper left corner of each image. An image-processing tool was used to extract the signal phase from the image patch, and the result was shown to the user to ensure that the phase detection was correct. The error of the crossing time since the beginning of the green signal phase is less than 100 ms. The trajectories may contain some errors from the conversion from image to world coordinates (calibration errors), tracking errors due to poor image resolution, or both. Such an error can be up to 1 m within the intersection box and even larger behind the stop bar on the far side of the intersection, where the image resolution is poor. Therefore, determining the crossing starting time from the trajectory alone may result in errors, but this TABLE 1 Characteristics of Observed Intersections Palo Alto: Park Blvd. at El Camino Real Berkeley: Russell St. at Telegraph Ave. Width 125 ft (38 m), seven lanes 84 ft (25.6 m), four lanes Speed limit 40 mph ( 70 km/h) 25 mph (40 km/h) (cross traffic) Traffic Heavy Moderate Intersection Crowned Flat Visibility Limited Better Approach grades Flat 3.4%, +2.5% Bike traffic Evening commute All day Bicyclists Young adults Diverse

88 88 Transportation Research Record 2128 FIGURE 1 Video playback tool showing image-processing software tracking bicyclists across intersection (El Camino Real at Park, Palo Alto, California). factor is not a major concern because the timing relative to the green signal onset is of primary interest. The output of the video image-processing software, applied using the playback tool, is a plot of the location of the bicyclist crossing the intersection versus time. The traffic signal phase information is used to establish the reference time. The start of the green phase is defined as time zero for each crossing trajectory, and the times of the yellow and red onsets are marked by vertical yellow and red lines on the plots. An example of the plot generated for each bicyclist crossing is shown in Figure 2. Figure 2 shows a plot of a standing start by a bicyclist, with the trajectory indicated by the solid curve. The stop bar on the far side of the intersection where the bicyclist had to wait before starting to cross is at location 41 m and the curb line on the near side of the intersection, Crossing completed at opposite curb Y position along Park Ave Effective start-up (offset) time 5.5 s Speed = 5.9 m/s (13.6 mph) 3 sec yellow Stop bar (start of crossing) Time FIGURE 2 Example of extraction of needed values for standing start from bicyclist trajectory plot (videofile: 592, timestamp: standing start traffic lane).

89 Shladover, Kim, Cao, Sharafsaleh, and Li 89 representing completion of the crossing, is at 3 m. The vertical yellow line indicates the yellow onset after about 15.5 s of green, and the vertical red line indicates the red onset 3 s later. This bicyclist started moving about 3 s after the signal changed to green and reached full speed of 5.9 m/s (defined graphically by the slope of the tangent to the trajectory curve) about 7 s after the start of the green phase. In order to provide a single concise representation of standing starts that can be generalized to intersections of arbitrary width, the offset time was defined as the intersection of the full crossing speed tangent with the line representing the start of the crossing. In this example, that offset time is 5.5 s, and this complete crossing maneuver is described most concisely in terms of that offset time and the final crossing speed of 5.9 m/s (13.6 mph). ANALYSIS OF BICYCLIST CROSSING DATA The data collection system was used throughout the daylight hours for two days in Palo Alto and three days in Berkeley. In Palo Alto, 310 bicyclist crossings were observed, but only 255 of these produced usable data, with complete enough trajectories for analysis. Of the usable crossings, 180 were from standing starts and 75 were from rolling starts, the latter indicating that the signal was already green when the bicyclist arrived at the stop bar to begin the crossing. In Berkeley, 439 usable crossings were recorded in both directions of travel, with 279 standing starts and 160 rolling starts. The Berkeley site had more unusable crossings because of difficult lighting conditions in the late afternoon and a wider range of complicated bicyclist maneuvers that were not readily classifiable. The standing-start and rolling-start cases were analyzed separately because of their different implications for signal timing. The standing starts are important for determining the length of the minimum green interval and are therefore the primary focus of this analysis. The rolling starts could be considered as inputs to decisions about the length of the yellow and all-red intervals, but Caltrans researchers believe that those must be determined on the basis of other considerations. The key attribute of the rolling-start crossings is the crossing speed, the distribution of which is shown in Figure 3. The superimposed lines on the cumulative distribution plot show the 10th-, 20th-, and 50th-percentile values, indicating potential criteria for consideration in signal timing. The values of these key percentile values were approximately 6.5 km/h faster for the Palo Alto bicyclists (primarily young adult commuters) than for the more diverse Berkeley bicyclists. The Berkeley data may have been biased somewhat to the low side because more of the bicyclists crossed this intersection at a significant angle to the direction of traffic on Russell Street, whereas the analysis results show the component of their speed in the direction of traffic. There was a further contrast between the speeds of the eastbound and westbound bicyclists at the Berkeley intersection based on the grade on Russell Street. Westbound bicyclists approached the intersection on a 3.4% grade, whereas eastbound bicyclists approached on a +2.5% grade, leading to a difference of 3 km/h in their median speeds, 7.2 km/h in their 80th-percentile speeds and 8.8 km/h in their 90th-percentile speeds. The statistics of the standing-start bicyclist crossings are most important for selecting the length of the green interval. The distributions of the offset times and the final steady speed of crossing are shown in Figures 4 and 5, respectively. A scatter plot of these samples for the Palo Alto bicyclists was generated to visualize potential correlations between these parameters. Fortunately, there was virtually no correlation between them, so they can be treated as independent Russell WB Russell EB Park FIGURE 3 Cumulative distributions and histograms of speed of rolling-start bicycles in mph.

90 90 Transportation Research Record 2128 First in queue at Park Russell WB Russell EB Park FIGURE 4 Cumulative distributions and histograms of offset times for standing-start bicyclist crossings. Russell WB Russell EB Park Speed (km/h) FIGURE 5 Cumulative distributions and histograms of final crossing speed for standing-start bicyclist crossings.

91 Shladover, Kim, Cao, Sharafsaleh, and Li 91 variables, simplifying the development of recommendations for signal timing. Figure 4 shows a strong concentration of offset times in the range of 4 to 8 s for the Palo Alto bicyclists, which is considerably longer than the start-up times reported by Wachtel et al. (1) in the same city, whereas the Berkeley bicyclists were concentrated in the range of 2 to 5 s. The median value was about 6.5 s, and the eightieth and ninetieth percentiles were about 8.3 and 9.3 s, respectively, in Palo Alto, whereas the corresponding values were about 3 s less in Berkeley. These values are this large because they are counted from the green onset, accounting for the time the bicyclists need to recognize that the signal has changed, and they represent the intersection of the constant speed tangent curve with the starting location from Figure 2 rather than the actual start-up time. This second point is an important distinction, which is needed when the results from one intersection have to be applied to an intersection with a different width. The offset time is the component of the crossing time that is independent of intersection width. Because there was significant bicycle queueing at the Palo Alto intersection, its data were recomputed for the first bicyclists in the queue to produce the fourth plot in Figure 4, which is more appropriate for choosing the minimum crossing time. The 80thand 90th-percentile offset times were 7.5 and 8.4 s, respectively, for these bicyclists. The faster offset times in Berkeley were initially surprising, since their rolling-start crossing speeds were significantly slower. The video data indicated that the Berkeley bicyclists were more likely to start crossing before they had a green signal (which none of the Palo Alto bicyclists did). The Palo Alto bicyclists had to be more cautious about starting because they were crossing a wider street with heavier and much faster traffic and they had poorer visibility from their starting position. In addition, the Berkeley bicyclists had such a short minimum green interval (4 s) that they may have been tempted to start crossing early in order to be confident of completing their crossing safely. The offset times in Berkeley differed significantly for the two directions, with the westbound bicyclists having significantly lower mean and variance in their offset times. Closer inspection of the Berkeley intersection revealed that eastbound bicyclists normally had clearer visibility of the approaching cross traffic on the near side of Telegraph Avenue (southbound) because of a bus stop at the corner, encouraging them to start crossing when no cars were visible on Telegraph during its yellow phase, but that visibility could be blocked when a bus was stopped, producing some longer offset times as well. Figure 5 shows that the final crossing speeds of the Palo Alto bicyclists were clustered in the range of 15 to 30 km/h, with only a few outliers at significantly higher speeds, whereas the speeds of the Berkeley bicyclists were mainly clustered from 10 to 20 km/h. The median speed in Palo Alto was 21.3 km/h, with 10th- and 20thpercentile values of about 16.8 and 18.4 km/h, respectively, which were about 6.5 km/h faster than the corresponding values observed in Berkeley, an indication of the different demographic composition of the bicycling population at these two intersections. The final crossing speeds in Palo Alto may also have been increased by the pronounced crown of the El Camino road surface, putting the second half of their crossings on a negative grade. Wachtel et al. (1) described the 10th-percentile bicyclist speeds as 12.8 km/h for young adults and 9.6 km/h for children. The speeds reported here are direct measurements of the actual final cruising speed rather than an average speed that includes some of the start-up transient. The final crossing speeds for the standing-start bicyclists in Berkeley were similar for the two directions of travel, since the intersection is flat even though the approach blocks have significant grades. As a cross check on the independence of the offset time and final crossing speed measurements, the distributions of the total time that the bicyclists took to cross the 125-ft width of El Camino Real at Park Boulevard from a standing start were also plotted. These data are shown in Figure 6, indicating a median crossing time in excess Time (s) FIGURE 6 Distributions of total time for bicyclists to cross El Camino Real from standing start.

92 92 Transportation Research Record 2128 of 13 s, with the 80th and 90th percentiles at about 15 and 16.5 s, respectively. These times are all larger than the sum of the current minimum green plus yellow plus all-red intervals at this intersection ( = 11 s). That minimum setting only meets the needs of the fastest 20% of the bicyclists. With the values of offset time and final crossing speed, the crossing times can be estimated as follows: ( ) ( ) width ft crossing time = offset time + crossing speed mph * 068. With the measured values of offset time and crossing speed, the crossing time estimates for the median and 80th- and 90th-percentile bicyclists are Median = 12.9 s (compared with 13.3 s measured in data), Eightieth percentile (80th-percentile offset time and 20thpercentile crossing speed) = 15.7 s (compared with 15 s measured in data), and Ninetieth percentile (90th-percentile offset time and 10thpercentile crossing speed) = 17.4 s (compared with 16.5 s measured in data). Figure 7a shows the distribution of the observed duration of the green phase for Park Boulevard, with each sample representing one bicyclist crossing. This signal is traffic actuated, with the green duration ranging from 10 to 37 s. Clearly most of the bicycle traffic is occurring when there is also significant vehicular traffic triggering the detectors to extend the green phase beyond its minimum duration. The colors on the histogram indicate the signal phase when the bicyclist completed the crossing. Obviously, the shorter-duration green cycles caused a significant proportion of the bicyclists to complete their crossing in the yellow and red phases. This plot indicates the desirability of at least 13 s of green to enable most bicyclists to complete their crossing in the green at this intersection. The analogous distribution for Russell Street is shown in Figure 7b, where an even larger majority of the bicycle traffic was crossing when the green phase was more than adequate to meet their needs and only a few of the bicyclists had to complete their crossing in the red when the green phase was only 7 s long. SUMMARY OF BICYCLIST CROSSING DATA The video data reported here represent an unprecedented rich description of bicyclist intersection crossing behavior, particularly with respect to traffic signal phase. Key findings from these observations are the following: Even within a relatively homogeneous population of bicyclists, there was a wide range of speeds and start-up offset times. The bicyclist speeds and offset times were not correlated with each other, so their statistics could be analyzed independently. Start-up times were significantly longer than those reported in previous studies (1, 2), but the authors believe that their data are more accurate representations of reality because they are based on the individual bicyclist trajectories rather than being derived indirectly from ensemble statistics with large variability. A start-up time of at least 8 s was needed to represent the ninetieth percentile of bicyclists crossing a high-speed, high-density Completion on Red Yellow (a) (b) FIGURE 7 Distribution of duration of green phase in seconds when bicyclists were crossing (a) El Camino Real at Park Boulevard and (b) Telegraph at Russell Street (Berkeley, California).

93 Shladover, Kim, Cao, Sharafsaleh, and Li 93 arterial with limited visibility, whereas about 6 s was needed to represent a comparable percentile of bicyclists crossing a medium-density arterial with moderate-speed traffic and better visibility. About 90% of the primarily young adult commuter bicyclists in Palo Alto reached a steady cruising speed of at least 16.8 km/h during their crossing, while the comparable statistic for the more diverse bicyclists in Berkeley was 11.2 km/h. This steady cruising speed is the parameter that has to be used to extrapolate the data to other intersections with different street widths. Grades on the roadways approaching the intersection significantly influence the speed of bicyclists making rolling approaches, particularly in the upper speed percentiles. An average grade around 3% led to a difference of 3 km/h in the median speed and 8.8 km/h in the ninetieth-percentile speed of the bicyclists approaching from opposite directions. Higher crowns on the surface of the street being crossed may also increase the offset time and the final crossing speed. The contrast between the Palo Alto and Berkeley data indicates the sensitivity of bicyclist crossing time statistics to differences in the bicycling population and the physical and operational characteristics of the intersection being crossed. Considerations such as the speed and density of the crossing traffic, the crown of the road surface, and the ability of the bicyclists to see the cross traffic from their starting position can have a significant influence on the time needed to traverse an intersection. APPLICATIONS TO TRAFFIC SIGNAL TIMING If the green time available for minor cross streets is increased to accommodate bicyclists, there is a concern that the consequent reduction in mainline green time will exacerbate congestion. To assess the validity of this concern, increases in the minimum green time were tested in a VISSIM microsimulation of 10 km of the El Camino Real corridor in Palo Alto and Mountain View, California, where 25 consecutive actuated traffic signals are coordinated. The simulation represented peak-period traffic conditions in order to identify worst-case effects on this busy six-lane arterial. The minimum green interval for most of the cross streets is 7 s and a few are longer. The base case for the simulation represented the current coordinated signal timing plan, without pedestrian cycles. The test cases provided for increases of 2 s and 4 s in the minimum green for one cross street and for all cross streets, plus one case with the addition of 20 pedestrian cycles per hour (the number recorded during the period of heavy bicycle traffic) at one intersection. The test cases with the minimum green increase at one intersection produced network traffic delay increases of only 0.5 s (0.6% of travel time) and mean mainline queue length increases of only 2.2% and 4.4% at that intersection. The cases with the minimum green increase at all intersections produced negligible changes in network delay and increases of 3.5% and 9.2% in the longest mainline queue (at most one-fourth car length increase). By contrast, the test case with the pedestrian cycles at one intersection showed a much larger effect, with a network delay increase of 1.1 s (1.23%) and queue length increase of 50% (1.5 car lengths) for through traffic and 22% for left turns. The simulation cases with the increased minimum green times had such small effects on traffic because the peak-period traffic volumes on the cross streets were already sufficient to extend their green intervals well beyond the minimum. The only times when bicyclists are likely to encounter the minimum green intervals are off-peak periods with low vehicular cross traffic and limited mainline traffic congestion. This finding indicates that it should be possible to extend the minimum green with negligible impact on congestion. Recommendations for minimum signal timing were developed on the basis of the field data reported in the previous section, accounting for the different intersection widths. The bicyclist crossing times based on 80th- and 90th-percentile offset times and final crossing speeds were associated with the sum of the green, yellow, and all-red intervals: ( G + Y + AR) = offset time + For the young adult commuters crossing a wide arterial with fastmoving traffic in Palo Alto, the 80th- and 90th-percentile timing criteria are ( ) = + G+ Y+ AR W 80 ( ) = + G+ Y+ AR W 90 For the more diverse recreational bicyclists crossing a moderatewidth, moderate-speed arterial in Berkeley, the comparable criteria are ( ) = + G+ Y+ AR W 80 ( ) = + G+ Y+ AR W 90 These signal intervals are plotted, together with example 50thpercentile values, in Figure 8, where the locations of the diamonds superimposed on the plots indicate which intersection was used to generate the line of the same color. The slopes of the plots for the Berkeley data are larger because of the slower final speeds there, whereas the intercepts of the plots for the Palo Alto data are higher because of the larger offset times that those bicyclists adopted to contend with the hazards of crossing their heavily trafficked, wide, high-speed arterial. An example application of the criteria derived from the Palo Alto observations to the entire set of standing-start crossing data observed in Palo Alto is shown in Figure 9. All of the standing-start bicyclist crossing trajectories above the red lines would have been completed before the red onset if these criteria had been applied to the signal timing at that intersection. The outlying plots on the right-hand side of the figure represented bicyclists who were queued behind other bicyclists, waiting to cross. If the detection system can detect their presence, it can extend the green interval beyond the minimum so that they can also be accommodated. CONCLUSIONS [ width ( m) ] [ crossing speed ( m s) ] This study has shown how to accommodate the needs of bicyclists for adequate green time to cross wide arterials at signalized intersections. The detailed, accurate measurements of bicyclist crossing

94 94 Transportation Research Record 2128 FIGURE 8 Dependence of signal duration criteria based on 50th-, 80th-, and 90th-percentile bicyclist capabilities on street crossing width. times represent an unprecedented quantification of bicycling behavior and bicyclists needs, providing the foundation for defining signal timing. These measurements have been parameterized in terms of starting offset time and final crossing speed so that the results can be generalized to intersections of arbitrary width. Signal interval lengths are defined as a function of street width to accommodate 80% or 90% of the bicycling population observed at two different kinds of intersections. Comparable data collection and analysis should be done at additional intersections with even more diverse characteristics so that the individual effects on bicyclists crossing times and speeds of all the relevant intersection attributes can be identified. ACKNOWLEDGMENTS This work was sponsored by the Business, Transportation and Housing Agency, California Department of Transportation, under Partners for Advanced Transit and Highways (PATH) Task Order The authors thank Ahmad Rastegarpour, Kai Leung, and Jose Perez of Caltrans for their support of this work, as well as PATH colleagues Scott Johnston, Kathryn Choi, and Xiao-Yun Lu for their contributions to the data collection and the transportation staffs of the cities of Palo Alto and Berkeley for their cooperation in selecting intersections and providing the needed permits. Distance Curb line 50%ile criterion 80%ile criterion 90%ile criterion REFERENCES 1. Wachtel, A., J. Forester, and D. Pelz. Signal Clearance Timing for Bicyclists. ITE Journal, March 1995, pp Rubins, D. I., and S. Handy. Times of Bicycle Crossings: Case Study of Davis, California. In Transportation Research Record: Journal of the Transportation Research Board, No. 1939, Transportation Research Board of the National Academies, Washington, D.C., 2005, pp Kim, Z. Real-Time Object Tracking Based on Dynamic Feature Grouping with Background Subtraction. Proc., IEEE Conference on Computer Vision and Pattern Recognition, 2008, pp Time Stop bar on far side DISCUSSION Robert Shanteau California Association of Bicycling Organizations, 13 Primrose Circle, Seaside, CA 93955, rmshant@gmail.com. FIGURE 9 Application of signaling criteria based on 50th-, 80th-, and 90th-percentile bicyclist capabilities to all crossings from standing start at Park Boulevard in Palo Alto. This paper finds bicyclist crossing times for two signalized intersections, one each in Palo Alto, and Berkeley, California.

95 Shladover, Kim, Cao, Sharafsaleh, and Li 95 Unfortunately, the intersections are not typical of other signalized intersections because of a high crown in Palo Alto and a narrow street and early starts in Berkeley. I understand that the authors will be observing more intersections to obtain a more representative sample. This research is being used to help to conform bicycle signal timing to the requirements of California s AB 1581, which was signed in The paper develops a startup offset time and a final crossing speed. At the Palo Alto intersection, however, bicyclists queue in a bike lane, so those at the rear of the queue need to wait for the those at the front of the queue to clear before they can start. Unfortunately, the trajectories used to develop the startup crossing time are of all stopped bicyclists at the intersection instead of only those at the front of the queue, so the reported startup crossing time is too large. Any future research of bicyclist crossing times should address these issues. AUTHORS CLOSURE The Discussion comment was based on the data in the original version of the paper as presented at the 2009 Annual Meeting. The version of the paper published here includes a new version of Figure 4, showing separate curves for the offset times for all bicyclists and for the bicyclists at the head of the queue at the Palo Alto intersection. The formulas for estimating crossing times and the plot of crossing times as a function of street width in Figure 8 are based on the offset time distribution for the first bicyclist in the queue. The 90th-percentile value of offset time was reduced by 0.9 s and the 80th-percentile offset time was reduced by 0.8 s compared with the offset times for all of the bicyclists. The contents of this paper reflect the views of the authors, who are responsible for the facts and accuracy of the data. The contents do not necessarily reflect the official views or policies of the State of California. The Traffic Signal Systems Committee sponsored publication of this paper.

96 Guidelines for Multicriterion Decision-Based Left-Turn Signal Control Ozlem Ozmen, Zong Z. Tian, and Reed Gibby There are three main types of left-turn control at signalized intersections: permitted, protected, and protected permitted. Determining left-turn signal control types is one of the major elements in traffic signal design and operations, and it is a systematic approach involving complex decision rules. Existing guidelines for determining left-turn signal control have been primarily based on traffic volume, delay, geometry, crash experience, and other related factors. Such guidelines are generally presented in a flowchart format. One of the major shortcomings of the existing guidelines is that a single variable could dominate the decision process and thereby prevent a balanced consideration of all other important variables. A new guideline for determining left-turn control types mostly applicable to conventional left-turn displays is presented. The current format is designed for a four-leg intersection configuration; however, it can be easily adapted to the other formats. The guideline is based on the principles of multicriterion decision analysis and provides an indexbased recommendation: a numerical scale is used to compare each type of left-turn control with the others instead of an absolute left-turn control type. With the data collected from the field, it is shown that the proposed guideline can be calibrated and tailored to specific left-turn control policies. Therefore, the guideline can be easily adapted by any jurisdiction. Determining left-turn signal control is a systematic procedure involving complex decision rules. In general, there are three types of left-turn controls at signalized intersections: permitted only (PM), protected only (PT), and protected permitted (PP). The objectives of any recommended left-turn signal control should be to provide improved efficiency or improved safety, or both. Improved efficiency is reflected by an increase in capacity or a decrease in delay. Improved safety is reflected by a likely decrease in crashes. A number of factors are involved when left-turn signal control types are determined. These variables generally include traffic volume, delay, geometry, speed, accident history, and other related factors. Various guidelines have been developed for determining left-turn control types at signalized intersections. These guidelines are mostly presented in a flowchart format. Threshold values are generally set for each factor, and the decision process is usually dominated by a single factor if the threshold value is met for that factor. For example, most existing guidelines usually set a threshold of 1 for the number of O. Ozmen and Z. Z. Tian, Department of Civil and Environmental Engineering, University of Nevada, Reno, NV R. Gibby, Nevada Department of Transportation, Carson City, NV Corresponding author: O. Ozmen, ozlem@unr.nevada.edu, oozmen@yahoo.com. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / left-turn lanes; that is, protected left-turn control is recommended if the number of left-turn lanes exceeds 1. Such a decision process lacks a balanced consideration of other factors such as opposing traffic volume, which may be very low; therefore, the decision may not be rational at all locations or under all conditions. In many cases, an alternative left-turn control type might provide equal performance, especially when jurisdictions have clear policies and preferences toward a specific type of left-turn control. Because many factors cannot always be quantitatively evaluated, engineering judgment often plays a critical role in the selection of a left-turn control type. The primary objectives of this study are (a) to develop a new guideline for selecting left-turn signal control based on the principles of multicriterion decision-making analyses and (b) to test the feasibility and validity of the guideline for practical applications. BACKGROUND Left-Turn Control Guidelines Several research efforts by state and local agencies have developed warrants and guidelines for determining left-turn control types at signalized intersections. Most guidelines were developed on the basis of various factors and are primarily presented in a flowchart format. The major factors include traffic volume, delay, accident history, number of left and opposing through lanes, speed, sight distance, and qualitative measures by engineering judgment, described as follows: Traffic volume and the volume cross product are the most common factors. Cottrell recommends permitted left-turn control if left-turn volume is equal to or less than two vehicles per cycle (1). The cross product is the left-turn volume multiplied by the opposing through volume. Different threshold values were used in the studies depending on the number of opposing through lanes (2 4). Delay is found to be a common measure in several studies. Koupai and Kothari recommended PP left-turn control as a minimum requirement when the mean peak-hour delay per left-turning vehicle exceeds 35 s for rural areas and 50 s for urban areas (3). Cottrell recommended PP left-turn control when the mean peak-hour delay per left-turning vehicle exceeds 35 s and the total peak-hour left-turn delay exceeds 2.0 vehicle-h (1). Accident history is another critical factor. PT left-turn control is recommended by Asante if seven or more left-turn related accidents have occurred within three years with PP left-turn control (5). Koupai and Kothari recommend PT left-turn control if the number of leftturn accidents on two approaches is more than six per year or on one approach is more than four per year during any 12-month period within the last 3 years (3). Another study suggested two threshold values when PP left-turn control should be converted into PT con- 96

97 Ozmen, Tian, and Gibby 97 trol: five left-turn accidents per year and the critical accident rate of 32.6 accidents per 100 million of left-turn and opposing traffic volumes (1). The number of left-turn lanes is mostly consistent among the various guidelines. If there is more than one left-turn lane, PT left-turn phasing is recommended (1, 2, 5). The number of opposing through lanes is increased when the through volume increases, and this change may lead to increased accidents. PT left-turn phasing is recommended when there are three or more opposing through lanes (1, 2, 5). Speed, sight distance, and engineering judgment are other important factors for selection of left-turn control. PT left-turn control is recommended if the opposing speed is greater than 45 mph (2, 4, 5). Restricted sight distance may cause the increase of accidents at the intersections. Upchurch (4) recommends PT left-turn control when the opposing speed is 35 mph or less and the sight distance is 250 ft or less, or the opposing speed is 40 mph or more and the sight distance is 400 ft or less. Most guidelines and warrants recommend that engineering judgment be used in conjunction with the quantitative measures. Although most existing guidelines have covered these major factors, selection of a left-turn control type is usually dominated by a single factor when its recommended threshold is met. The decision on selecting left-turn control type must be evaluated by traffic engineers with both left-turn guidelines and engineering judgment. Even if some guidelines include engineering judgment as a factor, the existing practice of relying largely on engineering judgment is insufficient because of inconsistent application and exposure to liability challenges. As a result, the decision process lacks a comprehensive and balanced consideration of all relevant factors, which may yield undesirable solutions. Multicriterion Decision Analysis Multicriterion decision analysis is a technique to handle situations in which multiple criteria, usually conflicting criteria, and factors are involved (6). There are three stages in utilizing this technique (7): 1. Determine the relevant criteria and alternatives, 2. Assign numerical measures to the relative importance of the criteria and to the alternatives for these criteria, and 3. Process the numerical values to determine a ranking of each alternative. A wide range of multicriterion decision analysis methods can be used to achieve a final ranking or scoring of the decision alternatives. Common multicriterion decision analysis methods are the analytic hierarchy process (8), simple additive weighting method (9), TOPSIS method (10), and ELECTRE (10). Multicriterion decision analysis begins with establishing factors that can measure relevant goal accomplishments (11). Most of the applications of multicriterion decision analysis in the traffic engineering field are still in the development phase. Only a small number of studies were found, mostly related to assessment of transportation policies, highway asset management, and planning (12 16). Another researcher makes a case for the use of decision support systems in dealing with the complexities of urban traffic systems with a large number of conflicting goals (17). Multicriterion decision analysis methodology was applied as a decision support tool to evalu- ate the different alternatives for reducing congestion and improving the road system in the city of Campinas, Brazil (18). In the study, multicriterion decision analysis was built in three phases: structuring, evaluation, and recommendation. The problem formulation and objective identification were achieved in the first phase. During the evaluation phase, a value from 0 to 100 for each alternative for a given criterion was assigned (called the value function ). Then weights were determined to assess the relative importance of each criterion in the whole model. In the last phase, global evaluation was performed by multiplying the weights by the attractiveness of each criterion from the value function, as follows: V( a) = w v ( a) + w v ( a) wnvn( a) n V( a) = wv i i( a) i= 1 where V (a) = global value for alternative a, v 1(a), v 2(a),..., v n(a) = partial values of alternatives for Criteria 1, 2,..., n, w 1, w 2,..., w n = weights on Criteria 1, 2,..., n, and n = number of criteria within model. The alternative was chosen with the highest result (V (a) ). Another study was conducted to improve the efficiency of the public urban transportation system (19). In the study, a model of multicriterion decision making based on the analytical hierarchical model was presented with the aim of improving the efficiency of the public urban transportation system. Multicriterion decision analysis was applied to the development of an incident management program including the enhancement of safety, minimization of resource requirements, and the enhancement of traffic operations (20). Applications of decision support systems and multicriterion decision making in particular to transportation are described in an interim report for NCHRP Project (21). NCHRP Project 20-29(2) expanded the framework developed in NCHRP and developed a generic software package to facilitate the multimodal, multicriterion transportation investment analysis for both freight and passenger transportation (22). The simple additive weighting method is probably the best known and most used multicriterion decision analysis method (11). The main goal of the method is to obtain a weighted sum of the performance rating for each alternative over all factors. The main steps of the simple additive weighting method are as follows: Identification of the alternatives, criteria, and factors and generation of a decision matrix. The analysis begins by establishing criteria that can measure relevant goal accomplishments. The decision matrix (D) involves a set of n criteria C i (i = 1, 2,..., n) and a set of m alternatives A j ( j = 1, 2,..., m) (23). A decision matrix for m alternatives and n criteria is given as A A... A 1 2 C1 x11 x12... x1m C2 x21 x22... x2m D = Xij = Cn xn1 xn2... xnm m ( 2) where X ij represents the performance of alternative A j ( j = 1, 2,..., m) with respect to Criterion C i (i = 1, 2,..., n). () 1

98 98 Transportation Research Record 2128 Determination of criteria weighting. Typically, the relative importance of criteria is measured with a weighting vector W, written as W w1, w1,..., w n ( 3) Weights can be expressed at either an ordinal (qualitative) or cardinal (quantitative) measurement level. Normalization. Both decision matrix (X ij ) and weight vector (W) may contain qualitative and quantitative data and each criterion may have different measurement units. Therefore, normalization may be required to compare all alternatives in the decision matrix. Determining the overall performance. The overall performance value of each alternative (V j ) is obtained by V j where n wi = 1 0 rij 1 () 5 i= 1 = ( ) = n i= 1 wr i ij r ij is defined as the normalized performance rating of alternative A j on Criterion C i (23). The more-preferred alternative A j is obtained on the basis of the value of V j (the greater the value of V j, the more preferred the alternative is). PROPOSED LEFT-TURN GUIDELINE ( 4) The multicriterion decision-based guideline in this study was based on the simple additive weighting method. The computational engine was implemented in Microsoft Excel with Visual Basic programming and followed the following steps: Determine each alternative, that is, the three control types; Identify criteria in the decision process, that is, the factors; Generate subdecision matrices for each factor with respect to each alternative; Determine factor weights based on local data and agency preferences; Generate a decision matrix; and Rank the alternatives. The three left-turn signal control alternatives at signalized intersections are A1, PM left turn; A2, PP left turn; and A3, PT left turn. The nine evaluation factors that are mostly applicable for four-leg intersections with conventional left-turn displays are identified as C 1, subject left-turn volume; C 2, cross product (subject left-turn volume times the opposing through volume); C 3, number of subject left-turn lanes; C 4, number of opposing through lanes; C 5, opposing speed; C 6, sight distance; C 7, number of left-turn accidents on subject approach; C 8, Compound Factor 1: number of subject left-turn and opposing through lanes; and C 9, Compound Factor 2: opposing speed and number of opposing through lanes. The guideline includes the compounding factors for considering the combined effect of number of subject left-turn and opposing through lanes and opposing speed and number of opposing through lanes. The subdecision matrices (scores) are generated for each criterion based on literature review and local preferences. In addition, the criterion weights determined by using local data are w 1, weight for subject left-turn volume; w 2, weight for cross product; w 3, weight for subject left-turn lane; w 4, weight for opposing through lane; w 5, weight for opposing speed; w 6, weight for sight distance; w 7, weight for number of left-turn accidents; w 8, weight for Compound Factor 1; and w 9, weight for Compound Factor 2. The decision matrix is generated for three alternatives and nine factors by using Equation 2 as follows: D = C The performance of each alternative with respect to each factor (X ij ) is obtained from the subdecision matrices. The overall performance value of each alternative (V PM, V PP, and V PT ) is calculated for nine factors as follows: n V = w X PM i= 1 n V = w X PP i= 1 V = w X PT C C ( SubjectLT_volume) ( Cross_product) ( LT_lanes) ( OppTH_lanes C3 C ) C C C ( OppSpeed ) ( Sight_Distance) ( LT_accidents) ( LT_lanes & OppTH_lanes) ( Oppspeed ) C9 & OppTH_lanes n i= 1 where V PM = overall performance of PM left-turn control, V PP = overall performance of PP left-turn control, V PT = overall performance of protected left-turn control, w i = weight for each factor, X i1 = performance of PM left-turn control with respect to each factor, X i2 = performance of PP left-turn control with respect to each factor, and X i3 = performance of PT left-turn control with respect to each factor. CASE STUDY i i i i1 i2 i3 ( ) ( ) ( ) A PM A PP A PT X11 X12 X13 X21 X22 X23 X31 X32 X33 X41 X42 X43 X51 X52 X53 X61 X62 X63 X71 X72 X73 X81 X82 X83 X91 X92 X93 () 6 The proposed left-turn guideline procedure was applied in a case study based on the data collected in three major jurisdictions in Nevada, including the city of Reno, the city of Sparks, and most of the Las

99 Ozmen, Tian, and Gibby 99 Vegas metropolitan area, where signals are operated and maintained by Freeway Arterial Systems of Transportation (FAST). These agencies have clear differences in their left-turn control preferences and policies. For example, Reno has a rather conservative policy reflected by a tendency to use protected left-turn controls, whereas FAST has a rather aggressive policy reflected by a higher percentage of PP leftturn controls. This difference is attributed to the heavier traffic flow rates that need to be accommodated. Sparks is somewhere between. In order to assess the guideline with field data from these agencies, the factors and weights had to be tailored to reflect the differences in these policies. The guideline was implemented in an Excel spreadsheet using the Visual Basic programming language. The spreadsheet included a total of four worksheets and four macro modules. The factors were listed on the Inputs worksheet; the Calc1, Calc2, and Calc3 worksheets are the computational engines for the three jurisdictions. In these worksheets, the subdecision and decision matrices and weights were presented for each jurisdiction. The recommendation from the guideline for left-turn control was also included in the Inputs worksheet based on the results from the other worksheets. The overall computational procedure for implementing the guideline for multicriterion decision analysis within the left-turn control can be summarized by the following steps: Identification of all important factors (Inputs worksheet), Generation of subdecision matrices (Calc1, Calc2, and Calc3 worksheets), Assignment of weight factors for each alternative (Calc1, Calc2, and Calc3 worksheets), Generation of decision matrix (Calc1, Calc2, and Calc3 worksheets), and Determination of results of overall performance (Inputs worksheet). Identification and Selection of Factors The factors were obtained from the literature and a comprehensive nationwide agency survey. In the survey, specific questions were asked about the factors and their degrees of importance when left-turn control types are determined. The majority of the responses placed crash records at the top, followed by capacity and delay, size of intersection (e.g., number of lanes), and traffic progression. As a result, nine factors were identified, as discussed in the previous sections: subject left-turn volume, cross product, subject left-turn lane, opposing through lane, opposing speed, sight distance, and number of leftturn accidents, and the two compound factors: subject left-turn and opposing through lanes and opposing speed and opposing through lanes. These compound factors primarily reflected degrees of influence on the safety aspect of left-turn controls. Generation of Subdecision Matrices The subdecision matrices were generated to determine the performance of each alternative with respect to each individual factor by giving a score. The scores of each alternative for each individual factor were primarily obtained on an empirical basis. Besides the information obtained from the literature and the nationwide survey, the research panel composed of traffic engineers from the three jurisdictions also provided significant input. Each factor was scored a value between 0 and 10. The number reflects the level of preference for a control type. A score of 10 would suggest that a certain type of left-turn control is absolutely needed. For example, when the left-turn volume is 60 vph or less, scores of 10, 0, and 0 are assigned to PM, PP, and PT, respectively, indicating that permitted left-turn control is strongly favored. Subdecision matrices were generated for each jurisdiction separately. Sample subdecision matrices generated for the city of Reno are shown in Tables 1 and 2: a total of nine subdecision matrices for each factor are presented. Most of the scores in the subdecision matrices were very similar for each city; however, there were still some variations from city to city because of each city s preferred left-turn control type in similar conditions. The major variation is the scores of the compound factor of opposing through speed and number of opposing through lanes (OppSpeed&OppTHLane). In similar conditions, Reno has a tendency to use protected left-turn controls, whereas Las Vegas and Sparks have a tendency to use PP left-turn controls. For example, Sparks and Las Vegas would more likely prefer PP left-turn control when the speed limit is 35 mph and the number of opposing through lanes is 2 (scored as 0 for PM, 6 for PP, and 4 for PT for Sparks and 2 for PM, 7 for PP, and 1 for PT for Las Vegas). In contrast, Reno would prefer PT left-turn control at the same approach (scored as 0 for PM, 3 for PP, and 7 for PT). Assignment of Weights Once the factors were identified and the subdecision matrices for each factor were established, the next step was to estimate the relative importance of each factor (weights). The weights are major indicators of the left-turn control preferences and policies and have a significant impact on the final recommendation. At the testing stage of the guideline, weights were assigned by group experts and engineers. Then the weights were calibrated and derived for the three agencies based on the intersection data collected at these jurisdictions. They were determined on the basis of the best match between the existing left-turn control and the guideline-recommended left-turn control. Table 3 shows the factor weights for Reno, Sparks, and Las Vegas. The weights are dynamic rather than static; that is, the weights change as the variable (input) changes. For example, the weight for the opposing through lane may be assigned at a lower value of 0.1 when the number of opposing through lanes is less than 3 but may be assigned at a much higher value of 0.3 when the number of opposing through lanes is equal to or more than 3 for Reno. In most existing guidelines, the number of opposing through lanes becomes a dominant factor when greater than 2, generally resulting in a recommendation of PT left-turn control. Overall Performance Once the factors were identified, their subdecision matrices were established, and the weights were determined, the next step was to generate the decision matrix. The overall performance of each alternative (V PM, V PP, and V PT ) was obtained for the nine factors with Equation 6. Depending on whether the left-turn phase was on a coordinated system, an adjustment was made to shift 10% of the overall performance value of V PP to V PT, reflecting the case in which PT control benefits progression by using lead-lag phasing without the yellow trap normally existing under PP control.

100 100 Transportation Research Record 2128 TABLE 1 Subdecision Matrices C1 C5 and C7 C9 PM PP PT C1: SubjectLT_Volume > C2: Cross_Product (OppTH = 1) 50, , , , , , , , >160, C2: Cross_Product (OppTH 2) 100, , , , , , , , >320, C3: LT Lanes PM PP PT C4: OppTH_Lanes = = C5: Opp Speed mph C7: LT_accidents (acc/year) C8: LT_lanes & OppTH_Lanes 1/ / / / / / / / / PM PP PT C9: OppSpeed & OppTH Lanes 25/ / / / / / / / / / / / / / / / / / / / / / / / / / / / / ( ) ( ) V = V 01. V + V PP ad PP PP PT VPT ad = VPT VPP + V PT where V PP ad = adjusted overall performance of PP left-turn control, V PT ad = adjusted overall performance of PT left-turn control, V PP = overall performance of PP left-turn control, and V PT = overall performance of PT left-turn control. Then the overall performance values of each alternative were converted into a 100% scale that gives the index for each left-turn control (LT control index). It should be noted that the guideline recommends each individual left-turn control type and does not con- ( 7) sider the opposing through left-turn control type in the decision process. Therefore the left-turn control chosen for one direction may govern what options are available for the opposing movement. During the final decision process, traffic engineers should give important consideration to whether the intersection is located on an arterial street. The type of left-turn phasing is an important factor for maximizing the progression on an arterial street, and many agencies consider choosing lead-lag phasing to improve the progression. Using lead-lag PP left-turn phasing is further enhanced by allowing vehicles to turn left during the permitted interval and providing more green time for the coordinated movements. This technique is especially effective for coordinated arterial signals where the progressed platoons in each direction do not pass through the signal at exactly the same time (24). If such is the case, the final decision for selecting the

101 Ozmen, Tian, and Gibby 101 TABLE 2 Subdivision Matrix C6, Sight Distance TABLE 4 Factor Input Data for Illustrative Example Value left-turn control type may favor lead-lag PP left-turn phasing unless no safety issue exists. In addition, consistency of left-turn control type along a corridor should be considered during the final decisionmaking process unless there is an absolute reason to select another type of left-turn control. Illustrative Example Speed Sight Distance An example is presented next to demonstrate how the proposed guideline works. The input data are shown in Table 4, and Figure 1 shows the decision matrix. As discussed in the previous section, the factor weights were assigned dynamically according to the input values. Since none of the factors exceed their threshold values, the following weights were City preference (1 Reno, 2 Sparks, 3 Las Vegas) 2 Traffic volume: left-turn (vph) 147 Traffic volume: opposing through (vph) 747 Number of lanes: left-turn 1 Number of lanes: opposing through 2 Opposing speed (mph) 40 Sight distance (ft) 500 Number of LT accidents (acc./year) 0 Coordinated (1 yes, 0 no) 1 assigned for the nine factors: w 1 = 0.01, w 2 = 0.01, w 3 = 0.1, w 4 = 0.1, w 5 = 0.1, w 6 = 0.1, w 7 = 0.15, w 8 = 0.15, and w 9 = The overall performance of each left-turn control was calculated as follows: n VPM = wi Xi1 = i= = n VPP = wi Xi2 = i= = n VPT = wi Xi3 = i= = Since there are nine factors, the overall performance value of each alternative was obtained by multiplying the weighted sum values by 9. TABLE 3 Factor Weights Condition City w 1 w 2 w 3 w 4 w 5 w 6 w 7 w 8 w 9 If sight distance is restricted Reno Sparks Las Vegas If OppTH_lane = 1 and Reno left turn accident > 3 Sparks Las Vegas If OppTH_lane > 1 and Reno left turn accident > 3 Sparks Las Vegas If LT_lane > 1 Reno Sparks Las Vegas If OppTH_lane > 1 Reno If OppTH_lane > 2 Sparks If OppTH_lane > 3 Las Vegas If cross product > 160,000 Reno &OppTH_lane = 1 or Sparks cross product > 320,000 Las Vegas &OppTH_lane>1 If LT_volume 500 Reno If LT_volume 300 Sparks Otherwise Reno Sparks Las Vegas

102 102 Transportation Research Record 2128 Raw Scores Overall Performance Weights PM PP PT PM PP PT LT&TH Lanes PM PP PT City Preference 2 1/ Subject Left-turn, vph / Cross Product 109, / Number of Left-turn Lanes / Number of OppTH Lanes / Opposing Speed, mph / Sight Distance, ft / Number of LT Accidents, acc./yr / Compound Factors 2/ LT&OppTH Lanes Accidents/year PM PP PT OppSpeed&THLanes Score LT-Volume PM PP PT Speed, mph PM PP PT Decision Matrix Speed SD > Cross Product: 2 Lanes PM PP PT , , OppSpeed&TH PM PP PT , / , / , / , / , / Opp TH Lanes PM PP PT 320, / >320, / Lanes: LT PM PP PT 35/ / / / FIGURE 1 Decision matrix for illustrative example. Table 5 shows the guideline results for the given example. Since the left-turn phase was on a coordinated system, adjustment was made to shift 10% of the overall performance value of PP left-turn controls to the overall performance value of PT left-turn controls: V V PP ad PT ad TABLE 5 ( ) = ( ) = = = Guideline Recommendations for Illustrative Example PM PP PT Overall performance (V PM, V PP, V PT ) Coordination adjustment (V PM, V PP-ad, V PT-ad ) LT control index 22% 26% 52% Conclusion See index of each control Finally, these scores were converted into a 100% scale that would give the index for each type of control. In this example, PT was a favorable control with an index value of 52% compared with 22% for PM and 26% for PP. The guideline was evaluated by using data from 28 intersections in the three jurisdictions. Table 6 shows a summary of the major characteristics at these intersections. The existing left-turn control types at the study sites were compared with those recommended by the guideline, and the percent matches are shown in Figure 2. The proposed left-turn guideline provided the highest matches to the existing control type for the Reno sites (93%). The sites in Sparks had the lowest match at 68%. There are several reasons why the controls did not match: If there is a more restrictive control type (e.g., PT) for the opposing left turn, the more restrictive control will be selected for both approaches for consistent driver expectations. The number of sites TABLE 6 Major Characteristics of Intersections for Illustrative Example LT Volume Speed # of Approaches # of Approaches # of Approaches # of Left-Turn # of Opposing # of Left-Turn Jurisdiction (vph) (mph) with PM with PP with PT Lanes Through Lanes Accidents Reno Sparks Las Vegas

103 Ozmen, Tian, and Gibby % 93% 68% % Matching Sparks Las Vegas Reno Cities FIGURE 2 Percent matching left-turn controls based on existing left-turn controls versus guideline results. involved in such a case was 4% in Reno, 5% in Las Vegas, and 23% in Sparks. This finding indicated that the more restrictive left-turn controls should be selected for both approaches at these intersections. The crash data were available only at the sites in Sparks. The high number of left-turn crashes triggered a more restricted left-turn control, resulting in the lower percentage of matches in Sparks. It seems that the recommendations from the guideline are rational, and some intersections may need to be changed to more restrictive control types because of the high number of left-turn accidents. SUMMARY AND CONCLUSION A new left-turn control guideline based on multicriterion decisionmaking principles was presented. This guideline eliminates one of the major shortcomings of previous guidelines by providing a comprehensive and balanced consideration of all the factors involved. Instead of recommending a clear-cut absolute control type, the guideline produced control indices, closely resembling a typical decisionmaking process in practice. Such a guideline is of particular interest in cases where multiple factors are generally involved and engineering judgment plays a critical role in determining the left-turn control. The computational engine for the guideline was implemented in an Excel spreadsheet format with Visual Basic codes. With the data from 28 signalized intersections in three major Nevada jurisdictions (Las Vegas, Reno, and Sparks), the scores and weights used in the decision matrices were calibrated and derived to provide the best match between existing control types and those recommended by the guideline. The case study demonstrated how the scores and weights of specific factors can be established to tailor specific agency preferences and policies. As indicated in the case study, there are clear differences in left-turn control preferences in the three jurisdictions. For example, Reno adopts a more conservative policy by implementing more PT left-turn controls, whereas Las Vegas adopts a more aggressive policy by implementing a higher number of PP left-turn controls. The results from the case study generally showed good correlation between the existing control and the guideline recommendations. The sites in the Reno area showed the highest match at 93%, and the site in Sparks showed the lowest match at 68%. It seems, however, that the recommendations from the guideline for the sites in Sparks were rational. The sites that did not match generally involved a rather high number of crashes, where the guideline recommended PT controls for both approaches instead of the current PP controls. In conclusion, the proposed left-turn guideline showed excellent potential in assisting traffic engineers and jurisdictions to make rational decisions on left-turn control types. The methodology and model structure used in this study can be easily adapted by other jurisdictions developing or revising left-turn control guidelines. The next phase of this study should focus on validation of the guideline by using more intersections. ACKNOWLEDGMENT The work reported is part of a research project sponsored by the Nevada Department of Transportation. REFERENCES 1. Cottrell, B. H., Jr. Guidelines for Protected/Permissive Left-Turn Signal Phasing. In Transportation Research Record 1069, TRB, National Research Council, Washington, D.C., 1986, pp Agent, K. R. Guidelines for the Use of Protected/Permissive Left-Turn Phasing. ITE Journal, Vol. 57, No. 7, 1987, pp Koupai, P. A., and A. M. Kothari. Recommended Guidelines for Protected/Permissive Left-Turn Phasing. Compendium of Technical Papers, ITE 1999 Annual Meeting and Exhibit, Las Vegas, Nev., Upchurch, J. E. Guidelines for Selecting Type of Left-Turn Phasing. In Transportation Research Record 1069, TRB, National Research Council, Washington, D.C., 1986, pp Asante, S., S. A. Ardekani, and J. C. Williams. Selection Criteria for Left-Turn Phasing and Indication Sequence. In Transportation Research Record 1421, TRB, National Research Council, Washington, D.C., 1993, pp Mendoza, G. A., P. Macoun, R. Prabhu, D. Sukadri, H. Purnomo, and H. Hartanto. Guidelines for Applying Multicriteria Analysis to the Assessment of Criteria and Indicators. Center of International Forestry Research (CIFOR), Bogor, Indonesia, Triantaphyllou, E. Multicriterion Decision Making Methods: A Comparative Study. Springer-Verlag, Heidelberg, Germany, Saaty, T. L. Axiomatic Foundations of the Analytic Hierarchy Structures. Management Science, Vol. 32, 1986, pp Farmer, P. C. Testing the Robustness of Multiattribute Utility Theory in an Applied Setting. Decision Sciences, Vol. 18, 1987, pp

104 104 Transportation Research Record Hwang, C. L., and K. Yoon. Multiple Attribute Decision Making: Methods and Applications. Springer-Verlag, New York, Yoon, K., and C. L. Hwang. Multiple Attribute Decision Making: An Introduction. Sage University Paper Series: Quantitative Applications in the Social Sciences, No Sage Publications, Tsamboulas, D., and A. G. Kopsacheili. Methodological Framework for Strategic Assessment of Transportation Policies: Application for Athens 2004 Olympic Games. In Transportation Research Record: Journal of the Transportation Research Board, No. 1848, Transportation Research Board of the National Academies, Washington, D.C., 2003, pp Janarthanan, N., and J. Schneider. Multicriteria Evaluation of Alternative Transit System Designs. In Transportation Research Record 1064, TRB, National Research Council, Washington, D.C., 1986, pp Tsamboulas, D. A., G. S. Yiotis, and K. D. Panou. Use of Multicriteria Methods for Assessment of Transport Projects. Journal of Transportation Engineering, ASCE, Vol. 125, No. 5, 1999, pp Kapros, S., K. Panou, and D. A. Tsamboulas. Multicriteria Approach to the Evaluation of Intermodal Freight Villages. In Transportation Research Record: Journal of the Transportation Research Board, No. 1906, Transportation Research Board of the National Academies, Washington, D.C., 2005, pp Venter, C. J., T. Lamprecht, and W. Badenhorst. Simulating Land Use Development Through a Stochastic Allocation Procedure in Johannesburg, South Africa. In Transportation Research Record: Journal of the Transportation Research Board, No. 1977, Transportation Research Board of the National Academies, Washington, D.C., 2006, pp Bielli, M. A DSS Approach to Urban Traffic Management. European Journal of Operational Research, Vol. 61, 1992, pp Fioravanti, R. D., M. A. Amancio, and M. L. Galves. Alternatives to Reduce Congestion and Improve the Road System Using a Multi- Criteria Decision Analysis: A Case Study in the City of Campinas, Brazil. In Urban Transport XIII: Urban Transport and the Environment in the 21st Century, Wessex Institute of Technology, Southampton, United Kingdom, Kavran, Z., G. Stefancic, and A. Presecki. Multicriteria Analysis and Public Transport Management. In Urban Transport XIII: Urban Transport and the Environment in the 21st Century, Wessex Institute of Technology, Southampton, United Kingdom, Chowdhury, M. A. Toward an Optimal Incident Management Program Using a Multicriteria Decision Making Approach. Resource Papers for the ITE International Conference, Dana Point, Calif., March 1996, pp Development of a Multimodal Framework for Freight Transportation Investment: Consideration of Rail and Highway Trade-Offs. Interim Report, NCHRP Project Texas Transportation Institute, Texas A&M University System, College Station, Nov NCHRP Research Results Digest 258: Development of a Computer Model for Multimodal Multicriteria Transportation Investment Analysis. NCHRP Project (2). TRB, National Research Council, Washington, D.C., Sept Chung-Hsing, Y. A Problem-Based Selection of Multi-Attribute Decision Making Methods. International Transactions in Operations Research, Vol. 9, 2002, pp Brehmer, L. C., K. C. Kacir, D. A. Noyce, and M. P. Manser. NCHRP Report 493: Evaluation of Traffic Signal Displays for Protected/ Permissive Left-Turn Control. Transportation Research Board of the National Academies, Washington, D.C., The views expressed in this paper are those of the authors and do not reflect any policy or standard of the sponsor. The Traffic Signal Systems Committee sponsored publication of this paper.

105 Optimizing Traffic Control to Reduce Fuel Consumption and Vehicular Emissions Integrated Approach with VISSIM, CMEM, and VISGAOST Aleksandar Stevanovic, Jelka Stevanovic, Kai Zhang, and Stuart Batterman One way to reduce excessive fuel consumption and vehicular emissions on urban streets is to optimize signal timings. Historically, signal timing optimization tools were used to reduce traffic delay and stops. The concept of optimizing signal timings to reduce fuel consumption and emissions was addressed decades ago with tools that are now considered outdated. This study advocates a fresh approach to integrating existing state-of-the-art tools for reassessing and ultimately minimizing fuel consumption and emissions. VISSIM, CMEM, and VISGAOST were linked to optimize signal timings and minimize fuel consumption and CO 2 emissions. As a case study, a 14-intersection network in Park City, Utah, was used. Signal timings were optimized for seven optimization objective functions to find the lowest fuel consumption and CO 2 emissions. Findings show that a formula commonly used to estimate fuel consumption in traffic simulation tools inadequately estimates fuel consumption and cannot be used as a reliable objective function in signal timing optimizations. Some of the performance measures used as objective functions in the optimization process were proved to be ineffective. When CMEM-estimated fuel consumption is used as an objective function, estimated fuel savings are around 1.5%, a statistically significant decrease. Further research is needed to find an effective way to minimize fuel consumption and emissions by using the proposed approach. Both continuous transportation growth in the Western world and the recent economic boom in India, China, and many third-world countries have had a tremendous impact on the use of fossil fuels. The increase in fuel consumption affects the environment (the greenhouse effect), health (air pollutants), and the economy (increased fuel prices). Increased fuel consumption is mainly caused by two factors. First, millions of new drivers start using private cars as a main mode of transportation every year. Second, when these new travelers join existing traffic demand, traffic congestion increases because highway capacity does not increase commensurately with the new demand. The highest fuel consumption on urban arterials is associated with driving in congested traffic, characterized by higher speed fluctua- A. Stevanovic and J. Stevanovic, Department of Civil and Environmental Engineering, University of Utah, 122 South Central Campus Drive, Room 104, Salt Lake City, Utah Current affiliation for A. Stevanovic, Department of Civil Engineering, Florida Atlantic University, 777 Glades Road, Building 36, Room 231, Boca Raton, FL Current affiliation for J. Stevanovic: 2145 Northwest Third Court, Boca Raton, FL K. Zhang and S. Batterman, Environmental Health Sciences, School of Public Health, University of Michigan, 1420 Washington Heights, Room 6037, Ann Arbor, MI Corresponding author: A. Stevanovic, aleks.stevanovic@fau.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / tions and frequent stops at intersections. However, low traffic and continuous progression along streets do not guarantee the lowest fuel consumption and emissions. Excessive speeding, which may occur on roads with low traffic, may cause increased emissions for several pollutants. The best flow of traffic on arterial streets, in terms of fuel consumption and emissions, is the one with the fewest stops, shortest delays, and moderate speeds maintained throughout the commute (1). One of the ways to reduce excessive stop-and-go driving on urban streets is to optimize signal timings. Historically, signal timing optimization tools were developed to reduce delays and stops experienced by urban drivers. The concept of optimizing signal timings to reduce fuel consumption and emissions was first addressed by Robertson et al. (2). However, at that time traffic was simulated by macroscopic and analytical tools, and individual driving behavior was not considered. Similarly, the relationship between traffic activity, fuel consumption, and vehicular emissions, which was applied to all vehicles, was a simplistic and linear relationship (2). In recent years powerful tools for traffic modeling, fuel consumption, and emissions modeling have been developed. Microscopic simulation tools, such as VISSIM, have been used for more than a decade to model individual traffic behavior (3). Similarly, emissions models, such as the comprehensive modal emission model (CMEM), were developed to estimate second-by-second emissions of individual vehicles based on modes of a common driving cycle (4). These two types of microscopic models were coupled to estimate instantaneous emissions based on second-by-second activities of individually behaved vehicles (5 7). However, signal timing optimization models have been developed that now use microscopic traffic models to evaluate and improve the quality of signal timings (8, 9). Researchers have reported that these new signal optimization tools generate signal timings that reduce delays and stops when compared with the ones generated by macroscopic optimization tools (10). However, no research has been performed that integrates all these new microscopic tools in order to find the best signal timings that would minimize fuel consumption and emissions. The research reported here aims to fill that gap in existing practice by integrating a microscopic traffic simulator, a comprehensive microscopic emission estimation model, and a stochastic signal optimization tool to provide signal timings that minimize fuel consumption and vehicular emissions. BACKGROUND In previous decades, many researchers have evaluated the effects of traffic signal timings on the environment (11 18). Effects are evaluated through an investigation of the amount of fuel consumption 105

106 106 Transportation Research Record 2128 and vehicle emissions (both air pollutants and greenhouse gases) for various traffic conditions by using various methods and tools. Both field and theoretical tests have shown that optimized signal timings decrease fuel consumption and vehicle emissions compared with nonoptimized timings. Robertson et al. optimized signal timings by using TRANSYT 8 to minimize fuel consumption (2). They found that when signal timings are not optimized to reduce delays but to reduce total fuel consumption, the benefits of such signal timings may decrease fuel consumption by up to 3%. The fuel consumption was estimated from its linear relationship with traffic performance measures (delay, stops, and average speed) (2). The research set an industry standard in optimization of signal timings by defining a performance index (PI) as a linear combination of delay and stops that should be minimized to get minimal fuel consumption. Experiments showed that each stop should be associated with a penalty delay of 20 s if fuel consumption is going to be minimized. This PI became a standard objective function for optimizing signal timings, and the defined weights for delay and stops have not changed significantly since then. To estimate air pollutant concentrations, Park et al. coupled the VISSIM microsimulator model with MODEM, an emissions inventory database (19). Concentrations estimated by using a Gaussian dispersion model were comparable with those estimated from another macroscopic model but slightly different from levels measured in the field. Instead of using an emissions inventory database, Nam et al. coupled VISSIM with CMEM to estimate emissions from a single vehicle (5). The comparison with the field measurements found that CMEM is acceptable when capturing aggregated hydrocarbon (HC) and carbon monoxide (CO) trends but less accurate for carbon dioxide (CO 2 ) and nitrogen oxides (NO x ). An integrated VISSIM CMEM model was also used to show that the signal timings, optimized for progression in TRANSYT 9, significantly reduced pollutant emissions and fuel consumption on an arterial road (6). Oda et al. developed a simulator to estimate CO 2 emissions (20). They used a macroscopic traffic flow model to input traffic activities into the CO 2 simulator. The authors wanted to optimize traffic control settings to reduce CO 2 emissions. However, because of the huge computational burden needed to estimate CO 2 for all vehicles in the network, the authors simplified the experiments. Instead of minimizing CO 2 they minimized the number of stops, which they had shown was highly correlated with CO 2 (20). Another integrated VISSIM CMEM model was used to show that a scenario with optimal traffic control reduced various pollutant emissions (CO, HC, NO x ) from 3% to 15%. The research was done for a road network in Beijing by Chen and Yu (7). Qu et al. investigated impacts of reduced freeway speed limits on traffic emissions in Houston, Texas (21). The authors used TRANSIM to model traffic. The traffic activities were imported into three emission models: TRANSIMS (CMEM), MOBILE 5, and MOBILE 6. Emissions of three major pollutants [volatile organic compounds (VOC), NO x, and CO] were modeled in each of the three emissions models to investigate the effectiveness of freeway speed limit reductions as a way to decrease emissions. The results were mixed, showing that some models justify the reduction of speed limits while others do not. The study also showed TRANSIMS s inability to model changes in speed limits accurately because of its discrete approach in modeling vehicular speeds. Another attempt to determine signal timings that minimize fuel consumption and vehicular emissions was reported by Smith et al. (22), who briefly addressed SCOOT operations that minimize vehicle emissions. Traditionally, SCOOT has been used to minimize delays and stops in traffic by adjusting signal timings based on traffic demand measured in real time. The authors tested a new version of SCOOT that can minimize any of the five emission pollutants CO, CO 2, VOC, NO x, and PM 10 instead of the traditional PI. The pollutants were estimated on the basis of the SCOOT traffic model. The authors used a new SCOOT feature to minimize emissions by adjusting traffic control settings for the U.K. region of Leicester. The results showed that emissions for any of the pollutants could be reduced by up to 2% if an emission-related objective function is used during SCOOT optimizations. Unfortunately, these reductions were not statistically significant at the 95% confidence level. A major limitation of the approach was the fact that SCOOT s mesoscopic traffic model was not capable of modeling second-by-second modular operations (acceleration, cruising, idling) of individual vehicles. Rather, SCOOT bases its emission estimates on average emission rates for each vehicle class (four classes are available), and traffic flow and speed estimates are averaged over each link (13, 23). In summary, researchers have used various traffic simulation tools and various methods to estimate fuel consumption and vehicle emissions. Most applications have shown that optimized signal timings decrease fuel consumption or vehicle emissions or both but are based on macroscopic or mesoscopic models and unreliable objective functions. However, no research has addressed the optimization of signal timings based on evaluations of single-vehicle emissions and driving behavior. Further, without an objective function related to accurate fuel consumption and emission estimates, signal timings cannot be optimized to minimize these environmental impacts. Research presented here optimizes signal timings on the basis of CMEM emissions estimates for a population of vehicles whose individual driving behaviors were modeled in VISSIM. Optimization was used to minimize fuel consumption and CO 2 emissions. VISSIM CMEM VISGAOST CONCEPT VISSIM Model VISSIM is a microscopic, time-step and behavior-based model developed to simulate urban traffic and public transport operations. The program can analyze vehicle operations under different lane configurations, traffic composition, traffic signals, and public transport stops. This ability makes it a useful tool to evaluate traffic in alternative networks and to develop transportation engineering and planning measures of effectiveness (3). The accuracy of a traffic simulation model is mainly dependent on the quality of the vehicle modeling, such as the methodology of moving vehicles through the network. In contrast to less complex models that use constant speeds and deterministic car-following logic, VISSIM uses the psychophysical driver behavior model developed by Wiedemann (3). VISSIM has several ways of modeling traffic control. One of the most popular ways is the emulation of the industry standards in traffic control established by the National Electrical Manufacturers Association (NEMA). Recent experiments showed that signal timings generated by VISSIM s NEMA emulator do not differ practically from those generated by real-world controllers. CMEM Model CMEM is a physically based, power-demand model developed by the University of California at Riverside, the University of Michigan, and Lawrence Berkeley National Laboratory (4). After a variety of

107 Stevanovic, Stevanovic, Zhang, and Batterman 107 enhancements, the latest version (3.0) includes submodels for lightduty vehicles (LDVs) and heavy-duty diesel (HDD) vehicles. These submodels estimate vehicle tailpipe emissions (CO, HC, NO x, and CO 2 ) in different modes of vehicle operation, such as idling, cruising, acceleration, and deceleration. Scora and Barth suggested that temporal and vehicular aggregations were necessary in practice because CMEM was developed to predict emissions for vehicle categories (4). The temporal scale ranges from second-by-second, several seconds (mode) to driving cycle or scenario, and the vehicular scale ranges from a specific vehicle, vehicle technology category, to general vehicle mix or fleet. CMEM model inputs include traffic composition, vehicle and operation variables (e.g., speed, acceleration, and road grade), and modelcalibrated parameters (e.g., cold start coefficients and an engine friction factor) (4). Outputs are tailpipe emissions and fuel consumption. Emissions (in grams per second) are predicted as the product of fuel rate (FR, in grams per second), engine-out emission indices (grams of emission per grams of fuel), and time-dependent catalyst pass fraction (CPF), defined as the ratio of tailpipe to engineout emissions. CPF is mainly affected by the fuel-to-air ratio and engine-out emissions. LDV and HDD models have similar structures (4). Both are composed of six modules: engine power demand, engine speed, fuel-toair ratio for the LDV model or engine control unit for the HDD model, fuel rate, engine-out emissions, and CPF for the LDV model or aftertreatment pass fraction for the HDD model. Key parameters (e.g., vehicle mass, engine size, fuel type) depend on vehicle technology, fuel delivery system, emission control technology, vehicle age, and other factors. CMEM has been calibrated by using data from the National Cooperative Highway Research Program, which includes both engine-out and tailpipe emissions of CO, HC, NO x, and CO 2 for over 400 vehicles in 36 vehicle technology categories. VISGAOST Program VISGAOST is an optimization program for signal timings of traffic controllers based on their performance in VISSIM microscopic simulation. The program bases its optimization on the stochastic nature of genetic algorithms (GAs). The general structure of VISGAOST GA optimization is well documented (10). The basic version of VISGAOST is written in C++ and relies on VISSIM s input and output files (3). The key part of the program is a simple GA similar to other GAs used for signal timing optimization (24). The first version of VISGAOST enabled the optimization of all four basic signal settings: cycle, offset, split, and phase sequence. The program was tested and evaluated for the network in Park City, Utah, consisting of three groups of coordinated intersections and two actuated intersections. Results confirmed that VISGAOST can find timing plans that work better in VISSIM than the initial timing plans from the field (9). Further, the results showed that the GA-optimized plan was better than the timing plan generated by the traditional optimization tool SYNCHRO. VISGAOST application was extended to enable optimization of transit signal priority (TSP) settings. The two most common TSP settings green extension and early green were optimized for a corridor of seven signalized intersections in Albany, New York. Results showed that the optimized timing plan improved overall traffic performance and reduced person delay (10). The extended version of VISGAOST, presented in this paper, enables optimization of signal settings to minimize fuel consumption and vehicular emissions estimated by CMEM. The program has been modified to accommodate new linkage to CMEM and some new estimates from VISSIM. The steps below describe the basic operations in the VISGAOST optimization process. Step 0: Initializing G, total number of generations; T, total number of timing plans per generation;, convergence threshold; i, current number of population; i = 0. Generation of initial population p i of timing plans tp k k [1,..., T] Read field timing plan tp 1 from database, Generate tp k k [2,..., T]. Step 1: Evaluating population Evaluation of tp k p i k [1,..., T] Write tp k to database, Simulate tp k, Estimate emissions for tp k, Calculate fitness k. Step 2: Testing termination criteria Find b, fitness b for which fitness b = max( fitness 1,..., fitness T ); Find fitness a for which fitness a = average( fitness 1,..., fitness T ); Test rule. IF ((i = G) OR (( fitness b fitness a ) < )) Stop and RETURN tp b p i ELSE GO TO Step 3 Step 3: Generating new population i = i + 1 Generation of new population p i Select best-ranking timing plans from p i 1, Generate p i through GA operations. GO TO Step 1 VISSIM CMEM VISGAOST Integration Figure 1 shows the integration of VISSIM, CMEM, and VISGAOST to find signal timings that reduce fuel consumption and vehicular emissions. The optimization process starts with the VISGAOST generation of the initial population of signal timings, which is seeded by the existing set of signal timings from the field. Each generated signal timing plan is evaluated in VISSIM. As a result of the evaluation process, VISSIM outputs a vehicle record file with relevant second-by-second data for each vehicle in the network for the entire simulation period. The vehicle record file is processed by the VISSIM CMEM interface and sent to CMEM. CMEM estimates emissions and fuel consumed during the evaluation of that particular signal timing plan. The CMEM estimates are then summed for all vehicles in the network during the entire simulation period. VISGAOST receives the summed fuel consumption (or vehicular emissions) for each signal timing plan from the current population. A signal timing plan with the lowest fuel consumption (emissions) will be selected as the best one and saved to be compared with the best one from the next generation. Then the GA procedure within VISGAOST uses four basic GA operators to create a new population of signal timings. The whole process is repeated until one of two predefined termination criteria is met.

108 108 Transportation Research Record 2128 VISSIM VISSIM Input SignalGroups[8] =[1.0,2.0,3.0,4.0,5.0,6.0,7.0,8.0]; Split[1,8]=[[10.0,23.0,10.0, 23.0,10.0,23.0,10.0,23.0]]; Optimized Signal Settings LeadPhase[1,8]=[[1,0,0,1,1, 0,1,0]]; CycleLength[1]=[66.0]; Offset[1]=[30.0]; VISGAOST VISSIM Output Ve hi cl e Record t; Veh; Type; v; a; 1.0; 2; 1001; 23.18; 0.86; 1.0; 1; 1001; 25.75; 0.69; 1.0; 3; 1001; 24.55; 0.82; 2.0; 5; 1001; 23.80; 0.59; 2.0; 4; 1001; 24.60; 0.80; 2.0; 2; 1001; 23.76; 0.86; VISSIM Output Network Measures of Effectiveness Performance Simulation time: 600 to 4200 Parameter Value Total travel time[h] Total delay time[h] Number of stops Stopped delay[h] 84.2 Summed Estimations CO E7 CO HC NOx Fuel Dist VISSIM-CMEM-VISGAOST Interface OR LDV HDD Lookup Table Type CMEM Description MEM Code Percent ULEV PZEV Tier 1 < 50k, low ratio Tier 1 < 50k, high ratio LDGV 8 Tier 1 > 50k, low ratio Tier 1 > 50k, high ratio 0.20 CMEM CMEM Output Control File: veh- Activity File: veh- Distance Traveled 0.55 mi Fuel Use (grams/mile) CO2 = (grams) CO = (grams) HC = (grams) NOx = (grams) Control File: veh- Activity File: veh- Distance Control Traveled File: 0.55 veh- mi Fuel Activity Use File: veh- Control File: veh Di stance (grams/mil Traveled e) 0.15 mi Activity File: veh- CO2 Fuel = Use Distance Traveled (grams) 0.35 mi CO = (grams/mile) (grams) Control Fuel File: Useveh- HC CO2 = = (grams) (grams) Activity File: veh- (grams/mile) NOx CO = = (grams) (grams) Distance CO2 Traveled = (grams) mi HC CO = = (grams) Fuel Use 5 (grams) NOx HC = = (grams) (grams/mile) 91 (grams) CO2 NOx = = (grams) (grams) CO = (grams) HC = (grams) NOx = (grams) 3-way catalyst, FI, > 50k miles low way catalyst, FI, > 50k miles high HDDV , 4 stroke, Elect FIGURE 1 VISSIM CMEM VISGAOST integration. VISSIM CMEM VISGAOST Interface Connections between traffic microsimulation tools, such as VISSIM, and instantaneous emission models, such as CMEM, have been described elsewhere (5 7), and other studies provide more information about the VISGAOST and VISSIM interface (9, 10). Here the focus is on specific modifications of these two interfaces that enable functional communication among the three tools. The major limitations of several previous attempts to integrate traffic simulation and emissions estimation models were that they were not applicable to U.S. traffic conditions because the emissions were based on European vehicles (15, 25, 26). In other studies, heavy vehicles were not modeled directly because there was no HDD model in CMEM at that time (5, 27 29). Finally, in several studies vehicle emissions were overestimated because low-emitting vehicles were improperly represented in the old CMEM version (4). This study improves the emissions modeling approach by using a recent version of the CMEM software (3.0) and a representative sample of vehicles used in the United States. A program built in Java connects VISSIM with the LDV and HDD models in CMEM (Figure 1). The program s logic is similar to that described previously (6, 28). The program improves on the previous developments by modeling diesel trucks directly, calling either LDV or HDD core models for each individual vehicle (instead of the LDV batch model, which limits the number of records and vehicles that can be handled) (4), and using Java, a platform-independent language. For each vehicle VISSIM provides simulation time, a vehicle identifier, a vehicle type (LDV or truck), speed, and acceleration or deceleration on a second-by-second basis. The Java interface program imports the VISSIM output file to CMEM, which uses individual vehicle data to estimate instantaneous emissions for each vehicle. Each VISSIM vehicle type is assigned (by the Java program) to a CMEM vehicle category. The assignment of vehicle categories follows the mapping process described in Table 1, which maps the vehicle types from MOBILE 6.2 to the CMEM vehicle categories. It was assumed that the simulated vehicle fleet is composed of the light-duty gasoline vehicles (LDGVs) and heavy-duty diesel vehicles (HDDVs) defined in MOBILE 6.2.

109 Stevanovic, Stevanovic, Zhang, and Batterman 109 TABLE 1 Mapping of Vehicle Categories in MOBILE 6.2 and CMEM CMEM Type CMEM Description Code Percent LDGV ULEV PZEV Tier 1 < 50k, low ratio Tier 1 < 50k, high ratio Tier 1 > 50k, low ratio Tier 1 > 50k, high ratio way catalyst, FI, > 50k miles low way catalyst, FI, > 50k miles high HDDV , 4-stroke, electric NOTE: FI = fuel injected. It was also assumed that LDGVs can be represented in CMEM by two Tier 2 vehicle categories, ultra-low-emitting vehicles (ULEVs) and partial-zero-emitting vehicles (PZEVs); four Tier 1 vehicle categories; and two categories of old vehicles (Table 1). The LDGV category was matched to these eight CMEM categories according to the vehicle age distribution from MOBILE 6.2 (30) and the Tier 2 phase-in schedule (31). CMEM does not include HDDVs after 2002, and thus the CMEM category of HDDVs was chosen instead. Trucks manufactured before 1998 or after 2002 were not considered in the study. Many vehicle types can be defined in VISSIM; however, the initial experiments were constrained to two: passenger cars and heavy vehicles (trucks). Depending on the VISSIM vehicle type, the Java program utilizes either the CMEM LDV model or the CMEM HDD model. A CMEM model (LDV or HDD) computes fuel consumption and vehicular emissions for each vehicle in the simulation outputs. The Java program summarizes individual vehicles fuel consumption and emissions (CO, HC, NO x, and CO 2 ) to obtain the total values for the entire road network. CASE STUDY Study Network, Park City, Utah To optimize signal timings for minimal fuel consumption, the Park City road network, located in Utah near Salt Lake City, was chosen. The network consists of two suburban arterials, SR-224 and SR-248, and many crossroads. The network, shown in Figure 2, has 14 signalized intersections and average annual daily traffic of 32,000 and 20,000 on SR-224 and SR-248, respectively km FIGURE 2 VISSIM model of road network in Park City, Utah.

110 110 Transportation Research Record 2128 VISSIM Model of Park City Network Building, calibrating, and validating the VISSIM model required extensive field data collection and data reduction efforts. A team of 10 students was employed and trained to collect various traffic data during three weeks in August All data used in this study were collected between 4:00 and 6:00 p.m. on workdays under fair weather and dry pavement conditions. The following data were collected: turning-movement counts, saturation flow rates, stopped delay at the intersections, spot speed data, corridor vehicle classification counts, and corridor travel times. On the basis of the collected data the following parameters were adjusted to calibrate the VISSIM model: traffic inputs and routing decisions, the two car-following parameters in the Wiedemann 74 VISSIM model, control delays, desired speed decisions, and vehicle compositions. Validation of the model was done with the corridor travel times. All of the segment travel times from the field and VISSIM were close, but four of them (two in each direction) were still statistically different (two-tailed t-test was performed, with a = 0.05 and n = 15). A detailed description of the data, calibration process, and validation results was given elsewhere (9). Field Signal Timings The field signal timings were implemented under mixed actuatedcoordinated and actuated-uncoordinated control. Intersections of Bonanza Drive and SR-248 and SR-248 and Comstock Drive are actuated-uncoordinated, and all others are coordinated. The first three intersections in the Kimball Junction area were run on 128-s cycles. The other intersections all ran on 106-s cycles with exception of Deer Valley and Bonanza Drive, which used double cycling. The signal timings in the field were monitored regularly, but there were no recent major updates. Traffic engineers maintained the signal timings to achieve good progression between intersections, which was reflected in the initial signal timings in the optimization. VISGAOST Optimizations There were two major objectives for the VISGAOST optimization of signal timings. The first objective was to compare estimates of the fuel consumption from CMEM LDV and HDD models with those computed by VISSIM (node evaluation) based on a formula widely used by major traffic signal optimization tools (TRANSYT-7F and SYNCHRO) (8). The formula used by VISSIM, TRANSYT-7F, and SYNCHRO reads as follows: F = total travel k + total delay k + stops k where k 1 = Speed speed 2, k 2 = , k 3 = speed 2, F = fuel consumed (gal), speed = cruise speed (mph), total travel = vehicle miles traveled (veh mi), total delay = total signal delay (h), and stops = total stops (veh/h). The second objective was to show that in order to minimize fuel consumption or vehicular emissions, fuel consumption or particu- lar vehicular emissions should be used as an objective function when signal timings are optimized. In other words, when delays or stops are used in the objective function, minimal delays or stops are obtained but not necessarily the minimal fuel consumption or lowest vehicular emissions. Because of the time-consuming optimization process, these optimization experiments were limited to minimize fuel consumption and CO 2 emissions. CO 2 does not represent a criterion pollutant, but because of the threat of global warming, controlling this gas has become more important than ever. Other vehicular emissions (CO, HC, NO x ) can also be optimized by the proposed VISSIM CMEM VISGAOST approach. Seven optimization experiments were conducted. In total more than 100 control variables for all intersections in the Park City network were optimized to reduce total delay (for the entire network, in hours per hour), stops, throughput (total number of vehicles that completed their trips in the network), PI (PI = total delay + 10 * stops/h), CMEM fuel consumption, VISSIM fuel consumption, and CMEM CO 2 emissions. Multiple optimizations, with various objective functions, were performed to show the difference in the lowest fuel consumption and CO 2 achieved by each optimization method. Each optimization started with the same initial signal timings from the field. Each optimization was based on evaluations of traffic and emissions performances accumulated during 60 min of simulation time with an additional 10 min for warm-up. Simulation warm-up is necessary to achieve steady-state traffic conditions in the network. Each optimization had 12,000 evaluations of various signal timing plans; 20 signal timing plans were operated through GA procedures for each of 600 generations. Previous experiments showed that this combination of GA population and generations yields the best results (9). In addition, each signal timing plan was evaluated for five randomly seeded simulation runs to account for variability of traffic flows. The optimizations were performed on 20 dual-processor computers. Overall, it took around 20 days of continuous simulation run time to complete the optimizations. RESULTS AND DISCUSSION Evaluation Results Optimization of CMEM-estimated fuel consumption is shown in Figure 3, which demonstrates how the best fuel consumption and average fuel consumption vary over 600 generations. Spikes observable in Figure 3 reflect use of partial optimizations of signal timings (9). Similar trends were observed for six other optimization runs. Most of the final signal timings, which reduce fuel consumption and CO 2 emissions, seem to favor major-street operations and yield more delay to the side-street traffic. These signal timings exhibit higher cycle lengths and better progression on major streets. Once the optimizations were finished, each of the seven best signal timing plans was evaluated through 40 randomly seeded VISSIM simulations. VISSIM performance measures were recorded and average statistics were computed. The 40 VISSIM runs were also linked with CMEM. CMEM s estimates of fuel consumption and CO 2 were recorded and averaged. Mean values from these statistics are presented in Table 2 for all seven objective functions. Discussion of Results Almost all of the optimization experiments found signal timings that reduce CMEM fuel consumption when compared with the initial

111 Stevanovic, Stevanovic, Zhang, and Batterman Average Fuel Consumption Initial Fuel Consumption Best Fuel Consumption Fuel Consumption [gal] Number of Generations FIGURE 3 Optimization of CMEM fuel consumption. signal timings. However, not all of the objective functions worked effectively. Although delay, throughput, and PI could be used as objective functions without additional constraints, the other performance measures were not as effective in this role. The major problem is represented by the way that an objective function reports performance of the system if there is excessive delay in the network. In such a situation delay and PI significantly increase, whereas throughput significantly decreases. So if the GA procedure suggests a signal timing plan that causes a traffic jam, these three performance measures, when used as objective functions in the GA, will detect the problem and such a signal timing plan will be discarded. However, when other performance measures are used as objective functions, they may not necessarily recognize poor traffic conditions. For example, if traffic is jammed, vehicles move less and hence they stop less frequently (as recorded by VISSIM). So if the stops are minimized in the optimization, the traffic jam will be perceived as a favorable outcome. A similar situation occurs with emissions-related measures. If a vehicle is stopped and idling, it consumes less fuel than one that runs at 40 mph. So although its fuel-per-mile consumption is higher when the vehicle is idling, its fuel-per-second consumption and emissions are lower. For this reason sometimes both VISSIM s internal fuel calculation procedure and CMEM report low fuel consumption and emissions associated with poor traffic conditions (which are detected by the other performance measures). To illustrate the problem, such an example is provided in Table 2, where VISSIMreported fuel consumption is used as an objective function (note the minimal VISSIM-reported fuel consumption and the huge increases in number of stops and delay). Similar problems were observed when three other performance measures were used as objective functions in these experiments (stops, CMEM fuel use, and CMEM CO 2 ). The ultimate solution to this problem might be either selection of the reliable objective function or use of various metrics (e.g., stops and delay) as constraints in the optimization rather than in the objective function. To test this integration, the second method was used. The GA optimizations were constrained in such a way that poor solutions, which minimize one objective function TABLE 2 Measures of Effectiveness from 40-Run Tests Mean Values CMEM VISSIM CMEM CO 2 Throughput MOE Optimized Fuel Use [gal] Fuel Use [gal] [kg] Delay [h] Stops [veh] PI Initial (no optimization) , , , Delay [h] , , , Stops a , , , Throughput [veh] , , , PI , , , CMEM fuel use [gal] a , , , VISSIM fuel use [gal] 2, , , , , ,487.2 CMEM CO 2 [kg] a , , , a Constrained optimization results.

112 112 Transportation Research Record 2128 but negatively affect all others, are discarded. Table 2 shows the results of such an approach for the three objective functions whose optimizations were constrained (denoted by an asterisk). The results from Table 2 show that almost every optimization reduces CMEM-reported fuel consumption and CO 2 emissions. However, optimizations in which these two performance measures were minimized generated minimal results. At the same time, VISSIM s fuel consumption does not seem to be consistent even when proper performance measures are used as objective functions. Overall, VISSIM fuel consumption does not represent an accurate value when compared with CMEM fuel consumption. The fuel consumption reported by VISSIM should be lower than that from CMEM because VISSIM reports fuel consumption only on the links within node boundaries (user-defined areas around the intersections), whereas CMEM reports total fuel consumed on all links in the network. However, results from Table 2 show the opposite trend, because VISSIM does not calculate fuel consumption properly. There are multiple reasons for this inaccuracy in VISSIM s fuel consumption calculations: the formula used by VISSIM is based on aggregated measures (speed, stops, delay, etc.) and cannot provide a level of accuracy achieved by CMEM calculations; VISSIM does not report fuel consumption for those vehicles that are still within the node boundaries; VISSIM s formula might be based on outdated emissions characteristics of an average vehicular fleet. Further research is needed to investigate the inaccuracy of the fuel consumption reported by VISSIM. When CMEM fuel consumption is used as an objective function to optimize signal timings, the optimal signal timings generate significantly lower fuel consumption than if delay or PI is used as an objective function. The savings in fuel consumption when compared with the delay or the PI are around 1.5%. Although such savings may not be seen as important, it is interesting to see that after 12,000 evaluations the results do not show that there is a significant difference in fuel consumption between signal timings optimized for minimal delay and minimal PI. In the past, the difference in fuel consumption between signal timings that optimize delay and PI was estimated to be between 1% and 3% (2). The current findings confirm those from Smith et al., who directly minimized fuel consumption (within SCOOT) and reported similar benefits over PI optimization of around 2% (22). CONCLUSIONS The goal of this study was to present a new integration of existing traffic operation, emissions estimation, and signal optimization models. The study describes the integration of VISSIM, CMEM, and VISGAOST to optimize signal timings in such a way as to achieve minimal fuel consumption and vehicular emissions. The following conclusions were reached: 1. Number of stops, fuel consumption, and CO 2 emissions do not seem to be reliable objective functions in the optimization of signal timings. Instead, they should be combined with other traffic performance measures, or additional constraints need to be introduced in the optimization process. Further research is needed to investigate what the best objective function is to minimize fuel consumption and emissions. 2. The VISSIM formula for fuel consumption is heavily influenced by number of stops and does not seem to be a reliable objective function to minimize fuel consumption or emissions. VISSIM s method of estimating fuel consumption seems to significantly overestimate total fuel consumption when it is compared with CMEM fuel consumption. 3. If fuel consumption is used as an objective function in a constrained optimization of signal timings, the optimal signal timings will generate fuel consumption 1% to 1.5% lower than that obtained through a minimization of delay or PI. Although these results may seem insignificant, they have the same order of magnitude as the results obtained from the experiments that first investigated fuel minimization through signal timings (2). 4. Lengthy computation times make application of this research impractical for everyday optimization of signal timings. For this reason it is necessary to investigate which combination of delay, stops, and other potential performance measures would lead to minimal fuel consumption. 5. Future research should address additional optimization experiments with a variety of traffic networks and scenarios to develop a general strategy of how optimization of certain traffic metrics affects fuel consumption and vehicular emissions. Eventually, results from the simulation testing should be validated by field measurements. REFERENCES 1. Barth, M. J., and K. Boriboonsomsin. Real-World Carbon Dioxide Impacts of Traffic Congestion. In Transportation Research Record: Journal of the Transportation Research Board, No. 2058, Transportation Research Board of the National Academies, Washington, D.C., 2008, pp Robertson, D. I., C. F. Lucas, and R. T. Baker. Coordinating Traffic Signals to Reduce Fuel Consumption. TRL Report LR934. Transport Research Laboratory, Crowthorne, Berkshire, United Kingdom, VISSIM 4.30 User Manual. Planung Transport Verkehr AG, Karlsruhe, Germany, Scora, G., and M. Barth. Comprehensive Modal Emission Model (CMEM) Version 3.01 User s Guide. University of California, Riverside, Nam, E. K., C. A. Gierczak, and J. W. Butler. Comparison of Real-World and Modeled Emissions Under Conditions of Variable Driver Aggressiveness. Presented at 82nd Annual Meeting of the Transportation Research Board, Washington, D.C., Stathopoulos, F. G., and R. B. Noland. Induced Travel and Emissions from Traffic Flow Improvement Projects. In Transportation Research Record: Journal of the Transportation Research Board, No. 1842, Transportation Research Board of the National Academies, Washington, D.C., 2003, pp Chen, K., and L. Yu. Microscopic Traffic-Emission Simulation and Case Study for Evaluation of Traffic Control Strategies. Journal of Transportation Systems Engineering and Information Technology, Vol. 7, No. 1, 2007, pp Hale, D. Traffic Network Study Tool, TRANSYT-7F, U.S. Version T7F10. McTrans, University of Florida, Gainesville, Stevanovic, A., P. T. Martin, and J. Stevanovic. VISSIM-Based Genetic Algorithm Optimization of Signal Timings. In Transportation Research Record: Journal of the Transportation Research Board, No. 2035, Transportation Research Board of the National Academies, Washington, D.C., 2007, pp Stevanovic, J., A. Stevanovic, P. T. Martin, and T. Bauer. Stochastic Optimization of Traffic Control and Transit Priority Settings in VISSIM. Transportation Research, Vol. 16C, No. 3, 2008, pp Hallmark, S. L., and R. Guensler. Comparison of Speed and Acceleration Profiles from Field Data with NETSIM Output for Modal Air Quality Analysis of Signalized Intersections. In Transportation Research Record: Journal of the Transportation Research Board, No. 1664, TRB, National Research Council, Washington, D.C., 1999, pp Rouphail, N. M., C. H. Frey, J. D. Colyar, and A. Unal. Vehicle Emissions and Traffic Measures: Exploratory Analysis of Field Observations at Signalized Arterials. Presented at 80th Annual Meeting of the Transportation Research Board, Washington, D.C., Marsden, G., M. Bell, and S. Reynolds. Towards a Real-Time Microscopic Emission Model. Transportation Research, Vol. 6D, 2001, pp

113 Stevanovic, Stevanovic, Zhang, and Batterman Rakha, H. A., and Y. Ding. Impact of Vehicle Stops on Vehicle Energy and Emissions. Presented at 80th Annual Meeting of the Transportation Research Board, Washington, D.C., Unal, A., N. M. Rouphail, and H. C. Frey. Effect of Arterial Signalization and Level of Service on Measured Vehicle Emissions. In Transportation Research Record: Journal of the Transportation Research Board, No. 1842, Transportation Research Board of the National Academies, Washington, D.C., 2003, pp Midenet, S., F. Boillot, and J. C. Pierrelee. Signalized Intersection with Real-Time Adaptive Control: On-Field Assessment of CO 2 and Pollutant Emission Reduction. Transportation Research, Vol. 9D, 2004, pp Li, X., G. Li, S. Pang, X. Yang, and J. Tian. Signal Timing of Intersections Using Integrated Optimization of Traffic Quality, Emissions, and Fuel Consumption: A Note. Transportation Research, Vol. 9D, 2004, pp Coelho, M. C., T. L. Farias, and N. M. Rouphail. Impact of Speed Control Traffic Signals on Pollutant Emissions. Transportation Research, Vol. 10D, 2005, pp Park, J. Y., R. B. Noland, and J. W. Polak. Microscopic Model of Air Pollutant Concentrations: Comparison of Simulated Results with Measured and Macroscopic Estimates. In Transportation Research Record: Journal of the Transportation Research Board, No. 1750, TRB, National Research Council, Washington, D.C., 2001, pp Oda, T., M. Kuwahara, and S. Niikura. Traffic Signal Control for Reducing Vehicle Carbon Dioxide Emissions on an Urban Road Network. In Proceedings of the 11th World Congress on ITS, Nagoya, Japan, Qu, T., L. R. Rilett, and J. Zietsman. Estimating Impact of Freeway Speed Limits on Automobile Emissions. Presented at 82nd Annual Meeting of the Transportation Research Board, Washington, D.C., Smith, K., K. Wood, and A. Ash. Managing Emissions from Vehicles in Urban Systems. Presented at the AET European Transportation Conference, Homerton College, Cambridge, United Kingdom, SCOOT Traffic Handbook, Issue A 31-Dec Transport Research Laboratory Ltd., Crowthorne House, Wokingham, Berkshire, United Kingdom, Goldberg, D. E. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley Publishing Co. Inc., Reading, Mass., Fellendorf, M. Integrated Simulation of Traffic Demand, Traffic Flow, Traffic Emissions and Air Quality. Proc., 8th International Symposium on Transport and Air Pollution, Graz, Austria, ptv.de/download/traffic/library/1999%20istap%20symposium%20 VISSIM.pdf. 26. Niittymaki, J., and M. Granberg. Development of Integrated Air Pollution Modeling Systems for Urban Planning: DIANA Project. Presented at 80th Annual Meeting of the Transportation Research Board, Washington, D.C., Malcom, C., G. Scora, and M. Barth. Validating a Micro-Scale Transportation/Emissions Model with Tunnel Study Data. Proc., 11th CRC On-Road Vehicle Emissions Workshop, San Diego, Calif., Chevallier, E. Microscopic Modeling Framework for Estimating Emissions from Traffic Management Policies. MS thesis. University of London, 2005, pp Kim, B. Y., R. L. Wayson, and G. Fleming. Development of the Traffic Air Quality Simulation Model. In Transportation Research Record: Journal of the Transportation Research Board, No. 1987, Transportation Research Board of the National Academies, Washington, D.C., 2006, pp User s Guide to MOBILE6.1 and MOBILE6.2. U.S. Environmental Protection Agency, r03010.pdf. 31. U.S. EPA Finalizes Tier 2 Standards and Limits on Gasoline Sulfur. U.S. Environmental Protection Agency, default-file/mob21.pdf. The Traffic Signal Systems Committee sponsored publication of this paper.

114 Optimizing Signal Timings from the Field VISGAOST and VISSIM ASC/3 Software-in-the-Loop Simulation Aleksandar Stevanovic, Jelka Stevanovic, and Peter T. Martin Traditionally, when traffic signals are retimed, a significant difference is seen between signal timings recommended by optimization software and those implemented in field controllers. Those two sets of signal timings rarely match each other, and often a manual process is involved in transferring the data to and from the field controllers. A method is presented: signal timings are downloaded from field controllers, optimized by a software package, and then uploaded to field controllers. The method is VISGAOST, a stochastic optimization program, working with VISSIM ASC/3 software-in-the-loop simulation to optimize the signal timings obtained from the field. The method was applied to optimize signal timings for a five-intersection urban arterial segment in West Valley City, Utah. Traffic operations simulated by a high-fidelity VISSIM represented field observations reliably. After thousands of potential signal timings were evaluated, VISGAOST found a better set of signal timings than those used in the field. The final signal timings were tested for robustness under fluctuating traffic in microsimulation. The test results show that these optimized signal timings are more robust than those used in the field. Further applications of the method are needed to test the field performance of the signal timings optimized by VISGAOST. Signal timing optimization tools have come a long way from predominantly deterministic tools based on the analytical theory of traffic flow to the currently popular methods that optimize signal timings by utilizing the stochastic nature of traffic flow modeled through microscopic simulations (1, 2). The main purpose of developing signal optimization tools is to test various traffic-control strategies without disturbing traffic in the field. However, representing real trafficcontrol logic in a simulation environment is not a trivial task. Hence, traditionally only basic parameters such as cycle length, offset, and splits were optimized. Also, when the optimal signal timings were found, they were rarely implemented in the field controllers without further alterations. One reason for this inconsistency between optimized and implemented signal timings is the variety of traffic controllers, which, in spite of their compliance with industry standards, do not always offer identical traffic-control features. It is inefficient and potentially impossible to maintain signal optimization software that is compatible with all major traffic controllers from a variety of vendors. How- Department of Civil and Environmental Engineering, University of Utah, 122 South Central Campus Drive, Room 104, Salt Lake City, Utah Current affiliation for A. Stevanovic: Department of Civil Engineering, Florida Atlantic University, 777 Glades Road, Building 36, Room 231, Boca Raton, FL Current address for J. Stevanovic: 2145 Northwest Third Court, Boca Raton, FL Corresponding author: A. Stevanovic, aleks.stevanovic@fav.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / ever, most vendors of commercial optimization software do not want to tie their software to a single traffic controller. For these reasons, signal optimization software programs have been using trafficcontrol logic and settings based on industry standards such as those of the National Electrical Manufacturers Association (NEMA) and have few vendor-specific features. So far, the concept of using low-fidelity traffic models and emulation of standard traffic-control settings has worked to a certain degree. Signal timings generated by this approach are sufficient to be used as initial timings, which are then fine-tuned in the field. However, if signal timings are needed to deal with oversaturated traffic or to address a specific problem at an intersection (e.g., transit signal priority, preemption, special pedestrian timings), the process of adjusting signal settings can be labor intensive. In addition, when one is dealing with very specific traffic problems, the signal timings suggested by a traditional optimization program may not be better than the ones that are causing problems in the field. With oversaturation becoming a major problem in U.S. metropolitan areas, there is a need for development of new tools and methods to optimize signal timings that will be ready to work in the field without extensive tune-ups. The objective of this study is to present an enhanced version of a stochastic signal timing optimization tool that interfaces with a software-in-the-loop simulation (SILS). The SILS suite consists of the VISSIM microsimulator and the ASC/3 Econolite virtual signal controller. The stochastic optimization tool is an enhanced version of the previously presented VISSIM-based genetic algorithm (GA) optimization of signal timings (VISGAOST) (1). The interface between VISGAOST and VISSIM ASC/3 SILS enables optimization of ASC/3 controller signal timings, which are evaluated through VISSIM simulations. This concept for the first time allows optimization of practical signal timings from the field. Signal timings from ASC/3 field controllers databases are imported into VISSIM ASC/3 SILS. These signal timings are then modified through GA procedures in VISGAOST. The final ASC/3 databases contain signal timings that work best in multiple randomly seeded VISSIM simulations. The final ASC/3 databases (with optimized signal timings) are ready for acceptance testing before authorized traffic-control engineers upload them to the field controllers. The VISGAOST optimization of ASC/3 signal timings was tested in a case study of five signalized intersections along the 3500 South arterial in West Valley City, Utah. REVIEW OF PREVIOUS RESEARCH Historically, macroscopic optimization programs, which are primarily deterministic in structure, such as TRANSYT-7F (2), and SYNCHRO (3) were used to optimize traffic-signal timings. The major improvement in evaluating traffic-control strategies in a realistic traffic environment was introduced through the use of microscopic 114

115 Stevanovic, Stevanovic, and Martin 115 simulation tools as external evaluators (4). Because of the longer computational times needed to evaluate hundreds of various combinations of signal timings, these microscopic evaluation tools were coupled with stochastic optimization methods. The stochastic optimization methods had already proved to be efficient for optimizing traffic-control settings (5). The stochastic optimization tools do not evaluate each combination of signal timings as the conventional tools do. Instead, they introduce smart algorithms (e.g., GAs) to lead the signal timing combinations toward an acceptable solution through an evolutionary process (6). At the time of their initial applications these new signal optimization methods were still constrained to use unsophisticated emulation of real traffic-control logic (5). Later, these new optimization methods were coupled with microsimulation tools, which offered significantly close emulation of the standard industry features of current traffic-control logic (1, 7, 8). The consequence of these new developments was that for the first time some traffic controller settings (also emulated) that were traditionally neglected in the optimization process were adjusted to reduce user delay (9, 10). Although not applied in signal optimization programs, new ways of modeling traffic control in microsimulation were recognized even before sophisticated emulators became available (11). The concept of hardware-in-the-loop simulation (HILS) was first developed as a package by the University of Louisiana (12). There is a long list of HILS deployments and testing studies since the first deployment (12). Currently, there are at least six university centers across the country where HILS is used for testing, research, and education (13 18). The HILS concept contains four elements: a microscopic traffic simulation model (e.g., CORSIM, VISSIM, and SIMTRAFFIC), a traffic controller (e.g., McCain, Naztec, Siemens, Econolite), a controller interface device (CID), and interface software. The microsimulation model generates traffic, which places calls on virtual detectors in the model. The detector calls are transferred through a CID to the traffic controller, which processes the calls, applies its controller s logic, and returns signal states back to the microsimulation model. The process depends on physical communication between HILS elements and cannot be easily automated. As a result, repetitive evaluations of various traffic-control settings, which are required in the optimization process, are almost impossible with the current HILS developments. The major disadvantages of the HILS approach are that it cannot run faster or slower than real time, there is no synchronization between the controller s and the computer s clocks, and separate controller hardware is required for each intersection (12). These issues were addressed when SILS was developed. The SILS suite provides virtual controllers that run from the same code base as the hardware controllers from the field (12). At least two SILS applications were successfully developed in the past: Siemens s NextPhase, which is linked to CORSIM (19) and VISSIM (20), and Econolite s ASC/3, which connects to VISSIM (one of the few commercially available SILS). The SILS approach is integrated within a single personal computer platform, which provides the opportunity for automation of multiple SILS runs. However, other issues, which include access to the proprietary controller s database files, have restricted development of applications in which signal timings are optimized through SILS. To summarize, conventional signal timing optimization tools are traditionally based on macroscopic and predominantly deterministic models of traffic operations. The new stochastic (GA) optimization tools enable evaluation of signal timing plans in the much more fieldlike virtual reality of microscopic traffic models. However, the signal timings that should be optimized are still hosted in emulators of the controller logics with basic industry standards (NEMA). Transferability of the signal timings to the field controllers once the optimization is finished remains a major issue. Cycle lengths, offsets, splits, and other settings are manually transferred from signal optimization software outputs into the field controllers. Often the signal timings from the optimization software may not have operational counterparts in vendor-specific controllers from the field (or vice versa). Sometimes signal timings from the field cannot be optimized by the optimization software (e.g., minimum and maximum green). Other times, optimized signal timings cannot be directly applied in the field controllers without further manipulation (e.g., an optimized offset is in seconds but the field controller uses offsets in percentage of cycle length). An integration of VISGAOST and SILS is presented here that, when applied to Econolite ASC/3 controllers, addresses most of these issues. SILS VISGAOST CONCEPT VISSIM ASC/3 SILS Integration PTV America in partnership with Econolite Control Products has developed an ASC/3 software-in-the-loop controller embedded in VISSIM. The SILS integration provides multiple virtual ASC/3 controllers capable of running signal timings faster or slower than real time. These virtual ASC/3 controllers run from the same code base as the ASC/3 hardware controllers. This feature means that ASC/3 provides nearly identical behavior as the field version, ensuring that the performance between the two is also nearly identical. The ASC/3 controller in SILS is compliant with the National Transportation Communications for Intelligent Transportation Systems Protocol, utilizes the same database as the field controller, and has virtually all the features of a hardware controller except those necessary to operate in a cabinet or centralized traffic signal system (12). VISGAOST Program VISGAOST is written in C++ as a standalone GA optimization program (1). It relies on VISSIM s microsimulation runs and relevant VISSIM and ASC/3 virtual controller input and output files. The program consists of two distinctive-by-purpose modules: the GA optimizer and the ASC/3 software interface. The VISGAOST optimization framework is described elsewhere (10) but provided here are four basic steps that are executed during the VISGAOST optimization process: Step 0: Initializing G, total number of generations; T, total number of timing plans per generation;, convergence threshold; i, current number of population; i = 0. Generation of initial population p i of timing plans tp k, k [1,..., T] Read and decode tp 1 from ASC/3 database, Generate tp k, k [2,..., T]. Step 1: Evaluating population Evaluation of tp k p i, k [1,..., T] Encode and write tp k to ASC/3 database, Simulate and evaluate tp k, Calculate fitness k.

116 116 Transportation Research Record 2128 Step 2: Testing termination criteria Find b, fitness b for which fitness b = max( fitness 1,..., fitness T ), Find fitness a for which fitness a = average( fitness 1,..., fitness T ), Test rule. IF ((i = G) OR ( fitness b fitness a < )) Stop and RETURN tp b p i ELSE GO TO Step 3 Step 3: Generating new population i = i + 1 Generation of new population p i Select best-ranking timing plans from p i 1, Generate p i through GA operations. GO TO Step 1 The GA optimization process starts by populating the first set of signal timing plans. The process uses a signal timing plan from the field as a seed to create the rest of the population. The other signal timing plans in the population are created by using a Monte Carlo simulation process. In the next step, all signal timings are converted into binary strings as a preparation step for GA manipulation. These signal timings are evaluated through multiple VISSIM runs, when each signal timing plan is associated with its fitness value a measure of its quality in terms of the objective function used (delay, stops, etc.). Then the GA operators (selection, crossover, mutation, elitism) manipulate signal timings from the existing population to create a new population (of the next generation). In general, new signal timing plans are selected for breeding by mimicking the natural process of the survival of the fittest, where the individuals with higher fitness have a greater chance to survive. The pairs of new timing plans are created by copying, combining, and modifying the genetic material of the selected pairs of old timing plans. The procedure for generation of a new population of timing plans is repeated until the predefined end criterion is satisfied. The other VISGAOST module that interacts with the GA optimizer is the ASC/3 software interface. This module accesses the databases of the ASC/3 virtual controller, which contain signal timing plans. Once the database content is read, the ASC/3 software interface decrypts the data and reads specific signal timings that should be optimized. The signal timings are then converted to binary format to be manipulated by the GA optimizer. Once the GA optimizer modifies the signal timings, they are converted back to decimal format. Finally, the ASC/3 software interface encrypts the signal timings back into the ASC/3 database format. SILS VISGAOST Connection The integration of VISGAOST and VISSIM ASC/3 SILS is shown in Figure 1. Two major connections are crucial for this integration. First, the GA optimizer connects to SILS to access VISSIM evaluation files. These files contain measures of effectiveness (MOEs) that are used within the GA optimizer to calculate the fitness value for each signal timing plan. Second, the ASC/3 software interface connects to SILS to access the ASC/3 database to read and write the signal timings whenever this is necessary during the optimization process. Except for new operations related to the ASC/3 software interface, the VISGAOST optimization remains the same as that described in previous studies (10). The signal timings are still constrained within the fraction-based encoding structure, which is based on minimum ASC/3 Software VISSIM SILS ASC/3 Database File ª «¼ Q Rƒcq sbar? P Taq \ - Ç Æ Á  Р¹ º Ñ P! Q A \ Q " # C 2 B R S p R ƒ c ` q s b a W ' & % $ D Š 7 d G V U 6 F VISSIM Output File Network Performance! T dœ V é À VISGAOST ASC/3 Interface A & # GA - Optimizer Simulation time from 600 to 4200 Parameter Value Total travel time[h],all Veh Total delay time[h],all Veh Average speed[mph],all Veh 28.9 Number of stops,all Veh Distance traveled, All Veh Stopped delay[h],all Veh 84.2 FIGURE 1 VISSIM ASC/3 SILS VISGAOST connection. (ASC/3 Database File is encrypted.)

117 Stevanovic, Stevanovic, and Martin S A D B B D 5200 W 4800 W 4400 W 4155 W 4000 W 0.0 mile 0.25 mile 0.5 mile FIGURE 2 Study corridor along 3500 South Street, West Valley City, Utah. maximum cycle times and relevant phase minimums. Thus, available ranges for each signal timing parameter are much shorter than the ones allowed by the ASC/3 controller (e.g., 999 s for cycle length or 255 s for an offset or split). In this way, the domain of multidimensional search space is reduced in order to avoid the selection of illogical signal timings. Also, the time to find a (local) optimum is reduced. SILS CASE STUDY Study Corridor To test the VISGAOST VISSIM ASC/3 SILS integration, a 1.5-mi segment (from 4000 West to 5200 West) of the 3500 South urban arterial in West Valley City, Salt Lake County, Utah, was selected. The segment, shown in Figure 2, represents one of the major east west (E W) arterials in the county. The arterial connects the fastgrowing western portion of the county with the major north south (N S) highway and transit routes. From left to right, the controllers at the five signalized intersections run actuated-coordinated operations with cycle lengths of 60 s, 120 s (the three intersections in the middle), and 150 s, respectively. The intersection at 4000 West has the highest traffic flows with more than 3,400 veh/h. The least congested is 5200 West, with less than 1,900-veh/h flow. The level of service (denoted with circled capital letters) and basic geometry information are provided for each intersection in Figure 2. Each intersection has an ASC/3 Econolite controller, which was bought to support transit priority operations for a recently introduced bus rapid transit service. VISGAOST Optimizations of Traffic Signal Timings The main objective of the optimization experiments was to identify better signal timings than the current field timings. Previous research (1, 10) showed that VISGAOST can find better timing plans than conventional signal optimization tools. Currently, the emphasis is on the practicality of VISGAOST, which is achieved through manipulation of signal timings utilized in the field. To accomplish this, ASC/3 database files for the five signalized intersections were obtained from Model traffic counts [veh/h] R² = Field traffic counts [veh/h] (a) Calibration and Validation of Corridor The VISSIM model was built, calibrated, and validated on the basis of data from the field (network geometry, measures of traffic, and transit operations). In order to calibrate the VISSIM model, signal timings from the field, speed limits, 15-min turning-movement counts, and queue lengths at some intersections were used. To validate the model, travel times (floating car with Global Positioning System device) along the arterial were measured while passing times at each intersection were recorded. Figure 3 shows the results of the calibration and validation efforts. In Figure 3a the field turning-movement counts were compared with VISSIM measurements from data collection points at the relevant locations. Similarly, in Figure 3b field travel times for each segment were compared with the VISSIM travel time measurements for the same segment. The high level of correlation obtained for the two pairs of data sets shows that a reliable model of the current traffic conditions on this arterial segment was achieved. Model travel times per section [sec] R² = Field travel times per section [sec] (b) FIGURE 3 Validity of VISSIM model of 3500 South Street: (a) calibration results traffic counts and (b) validation results travel times.

118 118 Transportation Research Record Initial Performance Index Average Performance Index Best Performance Index Performance Index Cycle Length Cycle Length, Cycle Length & Offsets & Splits All Offsets Phase Sequence Number of Generations FIGURE 4 Optimization of PI. the Utah Department of Transportation (UDOT). The ASC/3 databases were uploaded and used as the initial signal timings in the VISSIM ASC/3 SILS experiments. The VISGAOST program optimized cycle length, offsets, splits, and phase sequences for the five intersections. However, in the field, there is a reason that these intersections do not share the same cycle length. One intersection (60 s) is double-cycling to reduce unnecessary delay, whereas the other has a longer cycle length (150 s) to correspond to a busy intersection, which was not part of this study. For these initial tests it was assumed that all five intersections were under the same cycle length. The optimization process consisted of a sequence of partial optimizations. The decision to optimize signal timings separately was based on the experience that such an approach produces an appropriate solution faster than if all signal timings are optimized simultaneously. As a fitness function, a linear function of the total delay and stops, the performance index (PI), was used. The PI was selected as an objective function because it represents a common objective function in conventional signal optimization tools (e.g., SYNCHRO and TRANSYT-7F). The method for deriving the PI from VISSIM s outputs is reported elsewhere (1). The entire optimization process had 8,000 evaluations of individual timing plans 20 timing plans for each of 400 generations. In reality this number should be multiplied by 3 because each evaluation was done for three various random seeds in VISSIM (a total of 24,000 VISSIM runs). Each VISSIM run was 20 min long 15 min of peak traffic conditions plus 5 min of simulation warm-up to reach steady-state traffic conditions. A grid of 20 computers running continuously for 7 days was used to complete the optimization experiments. Once the optimal signal timings were obtained, their robustness (and the robustness of the initial signal timings) was verified through 40 VISSIM runs with independent random seeds. Two-tailed tests for paired means were then used to test the null hypothesis (α =0.05) that means of the selected performance measures were the same for both initial and optimized signal timings. RESULTS AND DISCUSSION Evaluation Results Results from Figure 4 show the average and the best PIs along the sequence of partial GA optimizations. Each partial optimization used the best signal timings found in the previous partial optimization. The optimization results presented in Figure 4 show that VISGAOST was able to find better signal timings than the ones currently implemented in the field. Table 1 provides statistics mean and standard deviation (SD) for 11 performance measures collected from 40 simulation runs. The PI values from Figure 4 and Table 1 are not the same for the relevant timing plans. The difference between initial and final signal timings based on three simulation runs (Figure 4) was reduced when the same signal timings were evaluated through 40 simulation runs (Table 1). Nevertheless, the two-tailed tests performed for 40-run statistics confirmed that the optimized signal TABLE 1 Performance Measures from 40-Run Tests Initial Timings Optimized Timings Measure of Effectiveness Mean SD Mean SD Number of stops 2, , Throughput 1, , Vehicles in the network Total delay (h) Total stopped delay (h) Vehicle stopped delay (s) Vehicle delay (s) Speed (mph) Distance traveled (mi) 1, , Stops per vehicle Performance index

119 Stevanovic, Stevanovic, and Martin 119 timings are significantly better than the initial timings by any of the 11 performance measures. Discussion of Results The difference in performance measures between the initial and final signal timings observed in Table 1 is the largest one that the authors have ever achieved when signal timings were optimized with VISGAOST. Figure 4 shows that the biggest improvement in PI is achieved when cycle lengths are modified. This large difference may exist because one of the intersections (with a cycle length of 150 s) is heavily influenced by a large intersection (a multilane highway and 3500 South) that is not part of the model. The long cycle length of an outside intersection and a need for steady progression might be reasons why the three intersections in the middle have longer cycle lengths in the field than necessary, as related to the microsimulation model. VISGAOST has no information on these external circumstances so it reduces the cycle length to meet local demand and traffic flow. External traffic flows in the model were coded correctly, but the tidal nature of traffic platoons coming from the outside intersections was not simulated. Table 2 presents all of the initial and final signal timings to provide a closer look at their differences. The biggest difference is observed in cycle length, although there are some significant differences in splits and offsets. Only one intersection (4000 W) benefits from an altered phase sequence. By considering the signal timings from Table 2 one may speculate that better signal timings may be achieved (by altering some offsets and splits) than the ones obtained in the optimization process. Stochastic optimization with a GA is a process that goes through thousand of combinations and does not guarantee a global optimum, especially when the results (Figure 4) do not show convergence of the average solutions. However, unaltered results are provided in Table 2 to show both advantages and disadvantages of the GA optimization process. Figure 5 shows variations of four performance measures during the optimization process. The PI is the performance measure that is optimized, and the other three performance measures are recorded only to observe variations. The variance in total delay is the most similar to the PI. Number of stops also shows a similar pattern to the PI s pattern, with few exceptions. Variation of throughput is the most interesting, showing that the best throughput was achieved at some time during the optimization of cycle length and offsets. Later, when the phase sequence for Intersection 4000 W was altered, the throughput decreases while the delay also decreases slightly. Since the PI is insensitive to throughput and is very sensitive to the decrease in delay, the optimization selects these signal timings as the best (at that point). However, because of the small size of the study network and relatively low traffic volumes, this difference in throughput does not have a significant impact on traffic operations. Comparison of variations in different performance measures, as shown in Figure 5, provides an opportunity to review the optimization process and find a set of signal timings that may not be the best on the basis of the optimized performance measure but would satisfy some other criterion (e.g., the highest throughput). CONCLUSIONS When signal timings from the field are optimized in signal optimization programs, the process is performed manually. Traditionally, of all the signal timings found in field controllers, only some are manually entered into the signal optimization program. Conventional optimization tools optimize an even smaller portion of those entered signal timings. Once optimized, these signal timings are again returned manually to traffic controllers. If the optimization is done on the basis of signal timing evaluations in macroscopic traffic models, there is an excellent chance that these timings will require significant finetuning before being implemented in the field. All of these issues were addressed by presenting an integration of VISGAOST (a stochastic signal optimization tool) and VISSIM ASC/3 (a SILS suite). Field signal timings were downloaded from real ASC/3 Econolite controllers. The ASC/3 controller databases were then uploaded into virtual ASC/3 controllers of the VISSIM ASC/3 SILS setup. The signal timings were manipulated through thousands of VISGAOST s GA iterations. Each signal timing plan was evaluated multiple times through a high-fidelity VISSIM model of an arterial street in west Salt Lake City. Robustness of the final signal timings contained in the modified ASC/3 databases was verified afterward. After completion of these procedures the signal timings are ready for acceptance testing before being uploaded into real-world signal controllers by authorized traffic signal professionals. On the basis of the described integration process and observed findings, the conclusions are as follows: 1. The attempt to integrate the existing VISGAOST and the VISSIM ASC/3 SILS suite was successful. There were no observed problems with manipulations of the ASC/3 databases or illogical signal timing values. 2. The results of optimization of signal timings on a fiveintersection arterial segment in West Valley City show that VISGAOST can find better signal timings than those currently used in the field. When the signal timings were tested for robustness through 40 randomly seeded simulations, the results confirmed the superiority of the optimized signal timings for all investigated performance measures. TABLE 2 Initial Versus Optimized Signal Timings Phase Splits (%) Signal Timings Intersection Offset (s) Phase Sequence Initial cycle time = 60, 120, 4,000W and 150 s 4,155W ,400W ,800W ,200W Optimized cycle time = 77 s 4,000W ,155W ,400W ,800W ,200W

120 120 Transportation Research Record 2128 Number of Stops Throughout Performance Index Total Delay [h] Cycle Cycle & Offsets Cycle, Offsets & Phase Sequence Splits Number of Generations FIGURE 5 Change in various performance measures during PI optimization. 3. The fact that differences between optimized signal timings and initial signal timings are extensive can be attributed to two potential reasons: the modeled network is taken out of its broader context or the signal timings in the field are outdated, or both. 4. VISGAOST will not generate perfect signal timings for a short number of evaluations, but it can be used to adjust many signal timing parameters that cannot be adjusted by traditional signal optimization tools. This feature, along with the fact that VISGAOST manipulates the same databases that are used in the field, makes VISGAOST a practical tool despite its long optimization times. 5. Further research is needed to assess the quality of signal timings optimized through the described approach in the field. It is also necessary to identify how such sophisticated optimized signal timings compare with unaltered signal timings generated by conventional signal timing optimization tools (e.g., SYNCHRO and TRANSYT-7F). All REFERENCES 1. Stevanovic, A., P. T. Martin, and J. Stevanovic. VISSIM-Based Genetic Algorithm Optimization of Signal Timings. In Transportation Research Record: Journal of the Transportation Research Board, No. 2035, Transportation Research Board of the National Academies, Washington, D.C., 2007, pp Hale, D. Traffic Network Study Tool, TRANSYT-7F, United States Version T7F10. McTrans Center, University of Florida, Gainesville, Husch, D., and J. Albeck. Synchro 6, Traffic Signal Software User Guide. Trafficware Corporation, Albany, Calif., Park, B., C. J. Messer, and T. Urbanik II. Traffic Signal Optimization Program for Oversaturated Conditions: Genetic Algorithm Approach. In Transportation Research Record: Journal of the Transportation Research Board, No. 1683, TRB, National Research Council, Washington, D.C., 1999, pp Foy, M. D., R. F. Benekohal, and D. E. Goldberg. Signal Timing Determination Using Genetic Algorithms. In Transportation Research Record 1365, TRB, National Research Council, Washington, D.C., 1992, pp Goldberg, D. E. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley Publishing Company, Inc., Reading, Mass., Hadi, M. A., and C. E. Wallace. Hybrid Genetic Algorithm to Optimize Signal Phasing and Timing. In Transportation Research Record 1421, TRB, National Research Council, Washington, D.C., 1993, pp Park, B. B., and J. D. Schneeberger. Evaluation of Traffic Signal Timing Optimization Methods Using a Stochastic and Microscopic Simulation Program. Research Report UVACTS University of Virginia, Charlottesville, Jan Park, B. B., and I. Yun. Stochastic Optimization Method for Coordinated Actuated Signal Systems. Research Report UVACTS University of Virginia, Charlottesville, Jan Stevanovic, J., A. Stevanovic, P. T. Martin, and T. Bauer. Stochastic Optimization of Traffic Control and Transit Priority Settings in VISSIM. In Transportation Research, Vol. 16C, No. 3, June 2008, pp Urbanik, T. II, and S. P. Venglar. Advanced Technology Application: The SMART Diamond. In Compendium of Technical Papers, Institute of Transportation Engineers, 65th Annual Meeting, 1995, pp Urbanik, T. II, M. Kyte, and D. M. Bullock. Software-in-the-Loop Simulation of Traffic Signal Systems. SimSub Mid-Year Newsletter, July 2006, pp Koonce, P. J. V., T. Urbanik II, and D. M. Bullock. Evaluation of Diamond Interchange Signal Controller Settings by Using Hardware-in-the-Loop Simulation. In Transportation Research Record: Journal of the Transportation Research Board, No. 1683, TRB, National Research Council, Washington, D.C., 1999, pp Abbas, M., D. M. Bullock, and L. Head. Real-Time Offset Transitioning Algorithm for Coordinating Traffic Signals. In Transportation Research Record: Journal of the Transportation Research Board, No. 1748, TRB, National Research Council, Washington, D.C., 2001, pp Kyte, M., A. Abdel-Rahim, and M. Lines. Traffic Signal Operations Education Through Hands-On Experiences: Lessons Learned from a Workshop Prototype. In Transportation Research Record: Journal of the Transportation Research Board, No. 1848, Transportation Research Board of the National Academies, Washington, D.C., 2003, pp Bullock, D. M., B. Johnson, R. Wells, M. Kyte, and Z. Li. Hardware in the Loop Simulation. Transportation Research, Vol. 12C, No. 1, Feb. 2004, pp Smadi, A., and S. Birst. Use of Hardware-in-the-Loop Traffic Simulation in a Virtual Environment. In Applications of Advanced Technology in Transportation: Proceedings of the Ninth International Conference, Aug , 2006, Chicago, Ill., ASCE, New York, 2006, pp Yun, I., M. Best, and B. Park. Evaluation of Adaptive Maximum Feature in Actuated Traffic Controller: Hardware-in-the-Loop Simulation. In Transportation Research Record: Journal of the Transportation Research Board, No. 2035, Transportation Research Board of the National Academies, Washington, D.C., 2007, pp CORSIM User s Manual, Version FHWA, U.S. Department of Transportation, VISSIM 4.30 Manual. Planung Transport Verkehr (PTV) AG, Karlsruhe, Germany, The Traffic Signal Systems Committee sponsored publication of this paper.

121 Evaluation of Integrated Platoon-Priority and Advance Warning Flasher System at High-Speed Intersections Henry X. Liu and Sundeep Bhimireddy Most rural or suburban high-speed isolated intersections have higher traffic volumes on the major approach compared with the minor approach. Often these intersections are not close enough to one another to provide coordination and are not far enough apart to disperse vehicle platoons completely on the major approach. These major-approach vehicle platoons are forced to stop frequently because of conflicting calls placed by a few vehicles on the minor approach. As a result, these intersections operate poorly, especially during peak periods. In addition, advance warning flashers are used at these intersections to warn motorists of the end of the green. The conventional method uses the trailing overlap green, which holds the green for a fixed time after gap-out. This trailing overlap green replaces the existing dilemma-zone protection provided by loop detectors and also increases delay on the minor approach. Recently, platoon-priority signal control systems have been developed to progress platoons efficiently at these intersections. In addition, the Texas Transportation Institute has developed the advance warning endof-green system to provide advance warning at these intersections without the necessity of holding the green after gap-out. In this study, these systems are integrated and the performance of the system is evaluated in terms of delay, stops, and advance warning time. Cabinet-in-the-loop tests performed with a real scenario suggested potential benefits of 50% reduction in delay and stops on the major approach experiencing platoons. It was also found that the total intersection delay and stops were reduced by as much as 20%. Most rural or suburban highways have high-speed isolated intersections with higher traffic volumes on the major approach and comparatively less traffic on the minor approach. Often these intersections are not far enough apart to completely disperse the platoons formed during the queue discharge process on the major approach. Therefore, vehicle platoons from the upstream intersection are often forced to stop because of conflicting calls placed by a few vehicles on the minor approach. As a result, these intersections operate poorly, especially during peak periods. Various platoon dispersion models have been developed to design effective signal timing and provide coordination between a series of intersections (1 5). However, these passive systems do not perform well when there are side streets between Department of Civil Engineering, University of Minnesota, Twin Cities, 500 Pillsbury Drive Southeast, Minneapolis, MN Corresponding author: H. X. Liu, henryliu@umn.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / the interconnected intersections or there is unpredictability in traffic demand. Recently, active platoon-priority signal control systems that operate on the basis of real-time traffic conditions have been developed to deal with this problem (6 8). Furthermore, advance warning flashers (AWFs) are used on major approaches at rural high-speed signalized intersections to warn motorists of the end of the green. The conventional method uses a trailing overlap green (9). A fixed-length green hold, typically 7 to 8 s, is placed at the end of the phase. The advance warning beacons located upstream of the intersection start flashing along with the trailing overlap green. In addition, dilemma-zone detectors are used on high-speed approaches to prevent a gap-out when there is a vehicle in the dilemma zone. But when AWFs are used, there is a fixed-length hold over the phase after gap-out and no more green extensions are given afterward. This setup replaces the dilemma-zone protection (DZP) provided by the detectors, and the routinely placed fixed-time hold at the end of the phase increases delay on the minor approach. An advance warning end-of-green system (AWEGS) has been developed by Texas Transportation Institute (TTI) to provide advance warning without the necessity of holding the green (10). Several research studies have been done on the effectiveness of AWF signs. The accident studies conducted on isolated high-speed signalized intersections (11 13) indicated that the advance warning signs are effective in reducing accidents. Although the results of an accident study in Minnesota (14) were mixed, the study recommended advance warning signs on approaches with a posted speed limit of 55 mph or higher. Both platoon-priority and AWEGS have a great deal in common in terms of hardware requirements and operation. Both systems use advance detection to collect information regarding future vehicle arrivals at the intersection and analyze collected information in regard to respective system objectives and, when necessary, override normal signal controller operations to achieve their purpose. However, there has not been a research effort to integrate these systems. The purpose of this research study is to develop and evaluate the benefits of an integrated system that provides platoon priority, advance warning of the end of green, and also DZP at the end of green. The developed system is tested on a cabinet-in-the-loop system with a real scenario to evaluate its benefits. INTEGRATED SYSTEM Figure 1 illustrates the integrated system (IS) architecture. Similar architecture has been used in previous studies (7, 10). The IS uses advance detectors and a computer installed with IS software to detect future events such as platoon arrivals and phase detector actuation 121

122 122 Transportation Research Record 2128 AWF Advance Detectors Dilemma-Zone Detectors Detector Status Signal and AWF Status Industrial PC NIDAQ I/O Card Detector and Signal Status Cabinet Controller Over-ride Inputs Detector Inputs Controller Over-ride Inputs Signal Status Traffic Controller FIGURE 1 IS architecture. times at the intersection. The algorithm estimates the travel time of each vehicle detected by advance detectors to each dilemma-zone detector in its lane and to the stop bar. This estimation is done on the assumption that the vehicles travel at a constant speed between the advance detectors and the stop bar as measured by the advance detectors. The system also assumes that the vehicles do not overtake or change lanes. The system overrides normal signal controller operations, using a combination of transit signal priority (TSP) call and phase-hold, to provide platoon priority to the detected platoons for an identified priority phase. In addition, with the detector actuation time information and signal controller status information, the system predicts future phase gap-outs and provides advance warning of the end of the green to motorists on main-street (AWF) phases. Also, by using the phase-hold option, the system provides DZP to both cars and trucks separately. The IS software consists of eight tasks that are performed in a loop in real time as shown in Figure 2. Each of these task duties is briefly described in the order of their execution. Controller Status Data Acquirer The controller status data acquirer module acquires signal status information and phase detector information from the signal con- troller via the cabinet back panel. A digital input output card is used to facilitate communication between the cabinet back panel and the computer. The signal status information is deduced from Phase Green On and Ring Status Bit pin terminals on the cabinet back panel. Timer Module With the signal timing plan data and the controller status data, the module keeps track of the minimum green timer and the maximum green timer for each AWF phase separately and updates them in real time. The maximum green timer starts a countdown when a call is placed on one of the conflicting phases. To estimate the actual minimum green portion (g actual min ) of a phase, the module keeps track of the number of detector actuations received during the nongreen portion of the phase. Depending on the number of actuations (a n ), seconds per actuation parameter (s), minimum green parameter (g min ), and maximum initial green time parameter (g max min) set for the phasein timing plan, g actual min for the phase is calculated by the module as follows: { min { min n }} actual max g = max g, min g, a s min

123 Liu and Bhimireddy 123 FIGURE 2 AWF Watchdog Task Start Controller Status Data Acquirer Timer Module AWF Watchdog Task Hold Call Manager Advance Detection Data Acquirer AWF System Platoon-Priority System IS flowchart. TSP Call Manager Even though the AWF system (described later) is designed to accurately predict the green phase terminations, it has limitations. The system assumes that all the vehicles that pass over the advance detectors are through movements and includes them in dilemmazone detector actuation times, but some of these vehicle movements may be turning movements that may not pass over all the dilemma-zone detectors. Also, the system assumes that the vehicles travel at constant speeds when in fact the motorists may slow down or accelerate as they approach the intersection. These assumptions result in unexpected gap-outs. To accommodate these cases, the AWF watchdog task keeps track of the gap-out timer and identifies any impending unexpected gap-outs within the next 0.2 s for each AWF phase. Figure 3 is the flowchart for this task. The definitions of extension timing and minimum timing can be found in a National Electrical Manufacturers Association standards publication (15). The flowchart illustration uses the following notation: P i.gtimer: phase i gap-out timer and P e i.go: phase i unexpected gap-out predicted (Boolean, true or false). The system supports the simultaneous gap-out feature provided in controllers; the flashers are activated only if both phases are predicted to gap out. While determining gap-outs, the task also verifies if there are any phase-hold calls placed by the hold call manager. If the algorithm determines that the green phase is going to terminate in the next 0.2 s, the algorithm calls the DZP module. It places a hold on the phase for the amount of time required to clear any existing vehicles in their respective dilemma zones, defined as the area between where 90% of the motorists do not stop and 90% of the motorists do stop (16, 17). This area is typically between 2.5 and 5.5 s travel time to the stop bar. The exact area depends on vehicle speed and its type. The dilemma-zone times for cars were estimated from the dilemma-zone areas defined for different speeds in the Minnesota Department of Transportation (Mn/DOT) manual on detector configuration (9). A research study conducted by Zimmerman (18) concluded that trucks benefit by an additional 1.5 s of DZP over cars since they require longer stopping distances. These equivalent time-based dilemma zones are used to determine if a vehicle is in its dilemma zone or not. If the DZP module has not placed a phase hold and both the phases are predicted to have unexpected gap-outs, the watchdog task activates the AWF beacons immediately. Otherwise, the watchdog alarm task determines that the phase is not going to terminate and waits to evaluate the next loop cycle. Hold Call Manager The hold call manager manages the phase hold calls placed on the AWF phases and the platoon priority phase. If there is a holdphase request from the DZP module or the platoon-priority system after the previous execution time, the algorithm places a hold on the respective phase. If there is currently a hold on a phase, the algorithm will remove or continue it on the basis of the duration requested by the DZP module or platoon-priority system. The hold calls are placed on phase hold input pins via the cabinet back panel. Advance Detection Data Acquirer In order to predict future arrival times of vehicles at the stop bar and the dilemma-zone detectors, the algorithm needs information about the approaching vehicles in advance. The algorithm needs vehicle detection time at the advanced detector location, vehicle speed, and vehicle type to estimate arrival times. These data are acquired by this module from advance detectors.

124 124 Transportation Research Record 2128 Start Select AWF phase Extension timing? no Minimum timing? no Hold placed? yes P i.gtimer < 0.2 yes yes yes yes no yes DZP Hold placed? no P e i.go=f Pe i.go=t Next AWF phase? no P e i e.go=t P i.go=t no P e i e.go=f P i.go=f yes Activate flashers End FIGURE 3 AWF watchdog task flowchart (T true, F false). AWF System Figures 4 through 6 are the flowcharts for the AWF task. The following notation is used: P i.go: phase i gap-out predicted, P i max-timer: phase i maximum timer, t AWF : desired advance warning time, PG: passage gap set in controller, CT: current time, P i.cht: phase i critical hot time, P i.got: phase i gap-out predicted time, and P i d t : phase i earliest detector actuation time after P i.got. The AWF task estimates the future arrival times of vehicles at each dilemma-zone detector for each lane. It also predicts each vehicle s presence time on the dilemma-zone detectors on the basis of its speed and type. With this information and phase timing data (minimum green, maximum green, etc.), the algorithm predicts the phase gapout time. The algorithm estimates the phase detector on and off times based on vehicle arrival times and detector presence times. A gapout is determined if a detector-off period greater than the passage time set in the controller is predicted. If the algorithm does not find a critical detector-off period between the vehicles detected by the advance detectors, the algorithm has to wait until the critical hot time before determining a gap-out because the detector actuation times of the next arriving vehicle are unknown. The algorithm also waits until

125 Liu and Bhimireddy 125 Start Select AWF phase P i.go = T yes no P i.max-timer < t awf yes no Predict gap-out Simultaneous gap-out logic? yes yes P i.go = T no Simultaneous gap-out no yes Hold placed? P i.go = T P i.go = T yes DZP Hold placed? yes Activate flashers no no yes Next AWF phase? no End FIGURE 4 AWF system flowchart. the critical hot time even though a detector-off period greater than the passage time is found if there is at least one lane open (i.e., there are no vehicles between the advance detector and the first dilemma-zone detector on the lane). The critical hot time is the latest predicted detector-off time from the advance detector information plus the difference between passage time and the threshold travel time between the advance detectors and the first dilemma-zone detector. This threshold travel time is computed by using the 99th-percentile speed of the approach. Once the algorithm determines that the phase is going to terminate, it checks if there is a scheduled phase hold during the determined green termination time. If there is a hold, the algorithm waits until the hold end time to predict future gap-outs. Otherwise, the algorithm calls the DZP module, which, only if required, places a hold on the respective phase. If both the phases are predicted to gap out, the task activates the AWF beacons. The task also features simultaneous gap-out logic (Figure 6) and gap-outs are predicted accordingly, when activated. Platoon-Priority System The platoon-priority system module detects future platoon arrivals at the stop bar by using the advance detector information. The algorithm only keeps track of last n (minimum number of vehicles that can be deemed as a platoon, or minimum platoon size) vehicles that passed over the advance detectors. The algorithm is similar to that described

126 126 Transportation Research Record 2128 Start Find next detectoroff period Detector-off period > PG yes no no List empty? Lane open? yes yes no CT > P i.cht no yes P i.go = T P i.go = F yes Hold placed? no End FIGURE 5 Predict gap-out flowchart. by Chaudhary et al. (7) and has two stages for platoon-priority scheduling: Identification and Extension. Figure 7 is the flowchart for the platoon-priority system. During the platoon identification stage, the algorithm evaluates the last n vehicles on a rolling-horizon basis. If the difference between the arrival times of first and last vehicles of the last n vehicles detected at the stop bar is less than a predetermined arrival time threshold value (T), the algorithm recognizes the group of vehicles as a platoon and schedules an initial platoon-priority interval with start and end times. Once the initial platoon-priority interval is scheduled, the algorithm switches to the platoon extension stage. During this stage, all additional vehicles that pass over the detector are evaluated individually to determine whether they are a part of the previously detected platoon. This determination is made by comparing the headways between the subject vehicle and the last vehicle of the platoon with a predefined extension threshold value (e). If the headway is less than the threshold value, the algorithm extends the initially scheduled priority interval end time to accommodate the current vehicle. This process continues until a headway that does not meet the threshold criterion is found or the maximum priority green time is reached. The algorithm parameter minimum platoon size (n), arrival time threshold value (T), and extension threshold value (e)

127 Liu and Bhimireddy 127 Start Start P 1.go=T, P 2.go=T Platoon-Priority Constraints yes no P 1.got < P 2.got Lock algorithm? P 1 * = P 1 P 1 * = P 2 P 2 * = P 2 P 2 * = P 1 no Platoon detected? yes P 2.got < P 1.d t no yes no yes Lane open for P 1 *? no CT > P 2 *.CHT Platoon Identification Stage Platoon Extension Stage yes yes no End P 1.go=T, P 2.go=T P 1.go=F, P 2.go=F FIGURE 7 Platoon-priority system flowchart. FIGURE 6 End Simultaneous gap-out flowchart. are predetermined on the basis of platoon arrival characteristics at the intersection. Controller Manipulation When a platoon is detected on the priority-phase approach, the signal status could be in the green or nongreen portion of the cycle. If the signal status is not green, the system needs to safely terminate the phase that is currently being served, skip all other unserved phases, and switch to the priority phase by the time the first vehicle in the platoon arrives at the stop bar. If there is a queue already, it needs to be discharged before the platoon joins the queue. To achieve this status, the controller is overridden by using a combination of TSP calls and phase holds. The controller manipulation depends on the status of the priority phase during the platoon-priority start time. The signal can be in either one of these two states: green or nongreen. The controller manipulation for these two cases is described as follows: Green. If the signal status is already green, a phase hold is issued from the time a platoon is detected until the platoon-priority interval end time. This step ensures that the phase does not gap out until all the vehicles in the platoon pass through the intersection. This phase hold is subject to the maximum green timer of that particular phase and is not applied past the maximum timer expiration. This restriction is placed to avoid higher delays on conflicting approaches. Nongreen. In the nongreen case, the system issues a TSP call so that the phase being served is terminated and the priority phase is served by omitting all other unserved phases (TSP settings are set accordingly). As soon as the current serving phase enters its yellow clearance interval, the TSP call is dropped and a phase hold is applied until the priority interval end time. This process is called early green. If the priority phase receives an early green, the algorithm is locked at the end of priority phase until all the nonpriority phases that have calls are served once. This step restricts giving back-to-back early greens to the priority phase, which incurs unreasonable delays on nonpriority phases.

128 128 Transportation Research Record 2128 TSP Call Manager The behavior of the TSP call manager module depends on the status of the priority phase during the platoon-priority start time. If the priority-phase status is green during the platoon-priority start time, the module informs the hold call manager to hold the priority-phase green until the platoon-priority end time. If the priority-phase status is not green, the module sends a continuous pulsating signal to the controller activating the TSP sequence. As soon as the priority phase turns green, the TSP call is dropped and the hold call manager is informed to hold the green for the remainder of the platoon-priority interval. CABINET-IN-THE-LOOP ARCHITECTURE The IS was designed to work in conjunction with a traffic controller in real time. Software emulators lack advanced features such as phase holds. Therefore, the IS was tested by using cabinet-in-the-loop simulation with an actual traffic controller. Figure 8 illustrates the system architecture for the cabinet-in-the-loop simulation setup. It consists primarily of three components: personal computer, cabinet, and controller. In real time, the cabinet sends vehicle detector calls and any controller override input calls (TSP and phase hold inputs) to the controller and in turn receives the signal status from the controller. An Econolite ASC/2S traffic controller installed with TSP software was used for this research study. The computer consists of two subcomponents: VISSIM simulation software and IS software running on it. At each simulation time step, the vehicle detector calls generated by VISSIM are sent to the cabinet and the current signal status from the cabinet is acquired and updated in VISSIM accordingly. The IS software receives advance detector information such as vehicle type and speed from VISSIM at each time step. It also receives phase detector status and signal status from the cabinet at each time step and when required sends controller override input calls to the cabinet. A digital input output card is used to provide communication between the computer and cabinet. CASE STUDY The developed IS was tested for the intersection of Trunk Highway 55 and Argenta Trail in Inver Grove Heights, Minnesota. This intersection is an isolated high-speed approach intersection with posted speed limits of 65 mph. The intersection experiences a significant number of platoons on the eastbound approach of TH-55 from the upstream intersection of TH-55 and TH-149 during the p.m. peak period. The two-intersection corridor is shown in Figure 9. The p.m. peak-period through and turning-movement volumes are shown also for the test intersections. The main priority phase is the eastbound through phase on TH-55. The concurrent priority phase that is served along with the priority phase is the westbound through phase on TH-55. The priority phase and concurrent priority phase (through phases on TH-55) are AWF phases. Spot speed studies were conducted on eastbound TH-55 between the two intersections and on Argenta Trail to collect free-flow speed data. The speeds on TH-55 have a mean of 61 mph and a standard deviation of 4 mph. The 15th- and 85th-percentile speeds are 57 mph and 65 mph, respectively. The speeds on Argenta Trail have a mean of 40 mph with a standard deviation of 4 mph. Detector Status Signal Status Digital I/O Card COM Interface Detector Status and Controller Override Inputs VISSIM Advance Detector Information Cabinet Signal Status Controller IS Software Digital I/O Card Controller Override Inputs Detector and Signal Status FIGURE 8 Cabinet-in-the-loop architecture.

129 Liu and Bhimireddy 129 T.H 55 T.H 149 Argenta Trail FIGURE 9 Intersection layout for case study. A network representing the corridor of the intersections shown in Figure 4 was built in VISSIM. All of the detectors were placed true to the existing field configuration, and controller parameters were set to field settings. The signal control of the intersection of TH-55 and TH-149 was controlled by the ring-barrier signal controller (traffic controller emulator) provided in VISSIM, whereas the signal control at the intersection of TH-55 and Argenta Trail was controlled by the real ASC/2S traffic controller in the loop. Several simulation studies were conducted to identify optimal parameters for platoon identification. These results indicated a minimum platoon size of 6, arrival time threshold of 8 s, and extension threshold value of 3 s for the scenario. These values were used for the following simulation tests. To have a deeper understanding of the effects of the systems individually, simulation tests were carried out enabling only one system at a time initially. Later both systems were enabled at the same time and tested. These results are discussed in the following sections. Delay and Stops Platoon-Priority System To understand the behavior of the platoon-priority system and evaluate its benefits, simulation tests were carried out with only the platoon-priority system enabled. The simulation results are shown in Table 1. For the priority phase, the advance detector at 1,000 ft from the intersection produced the lowest delay and fewer stops. In TABLE 1 System Performance Advance Detector Locations Normal Trailing Overlap 1,000 ft 1,250 ft 1,500 ft Delay Stops Delay Stops Delay Stops Delay Stops Delay Stops Platoon-Priority System Priority phase Concurrent phase Nonpriority phases Total Advance Warning Flasher System Priority phase Concurrent phase Nonpriority phases Total Integrated System Priority phase Concurrent phase Nonpriority phases Total NOTE: Delay (s/veh); stops (%).

130 130 Transportation Research Record 2128 general, the platoon-priority system produced 45% to 50% lower delay and 55% to 60% fewer stops for the priority phase. However, for the concurrent priority phase, the lowest delay and fewer stops are found with advance detectors at 1,250 ft from the intersection. As expected, the normal signal timing plan produced better performance on the nonpriority phases. On the whole, advance detectors at 1,000 ft produced the lowest delay of s/veh and lowest percentage of stops of 50%. This result is a 14% decrease in delay and an 18% decrease in stops over the performance of the normal signal timing plan. AWF System Simulation results obtained from the tests done by enabling only the AWF system are shown in Table 1 against those for the normal signal timing plan and the trailing overlap green. As expected, the trailing overlap green system introduced unnecessary delays on nonpriority phases without any benefit on the priority and concurrent priority phases, which resulted in an overall increase in delay and stops. The AWF system provides advance warning by predicting the gap-out instead of holding the green past the gap-out and holds the green only when DZP is needed. As a result, the AWF system produced lower delays and fewer stops than the trailing overlap green system. However, the AWF system also produced slightly higher delays and more stops than the normal signal timing plan. This finding is expected since the AWF system sometimes holds the green to provide DZP for a few vehicles on the major approach, inducing delay for several vehicles waiting on the minor approach. The system was primarily designed to provide safety at high-speed intersections rather than efficient operation. IS Performance The IS performance results are shown in Table 1 against those for the trailing overlap green system. There is a slight increase in overall intersection delay over the platoon-priority system performance. This additional delay is induced by the AWF system, as explained previously. However, the AWF system had little or no effect on the number of stops. Although the three advance detector locations provided better performance, advance detector location at 1,250 ft provided the best performance. With the advance detectors at 1,250 ft, the delays were reduced by 51% and stops were reduced by 59% for the priority phase. Considering overall intersection performance, delay was reduced 19% and stops were reduced 21%. Advance Warning Time The histograms for advance warning time provided by the IS run for a 10-h simulation period for different advance detector locations are shown in Figure 10. The results obtained with enabling only the AWF system followed the same pattern as that for the IS. Therefore, because of space restrictions, only the results from the IS are presented. The length of the advance warning time provided depends on the distance of the advance detectors (the farther they are from the first dilemma-zone detector, the farther the algorithm can look into the Frequency Frequency Seconds Seconds (a) (b) Frequency Seconds (c) FIGURE 10 Advance warning time histograms for detector locations at (a) 1,000 ft, (b) 1,250 ft, and (c) 1,500 ft.

131 Liu and Bhimireddy 131 future and the longer the algorithm can provide advance warning times). However, there were cases when the advance warning time provided was less than 1 s. Ninety percent of these cases were unexpected gap-outs but were predicted by the watchdog alarm task. If there were any vehicles in the dilemma zone during those gap-outs, the algorithm would have extended the green. Advance warning was not necessary in those cases since there were no vehicles in the dilemma zone. In general, the advance warning time recommended by Mn/DOT on a typical high-speed approach with a posted speed limit of 65 mph is 7.5 s (9). Therefore, for the tested intersection, advance detection around 1,250 ft is suitable for providing the recommended advance warning time. CONCLUSION Vehicle-actuated signal control performs poorly at intersections experiencing platoons. Orthodox AWF signs, which use trailing overlaps, increase intersection delay and replace existing DZP provided by loop detectors. Platoon-priority systems and advance warning systems have been developed to deal with these problems separately in previous studies. This study discussed the integration of platoon-priority systems and advance warning systems to improve both efficiency and safety at high-speed rural isolated intersections. The developed system was tested with a real-world scenario by using cabinet-in-theloop architecture. The benefits were quantified in terms of delay, stops, and advance warning time provided. For the tested intersection, advance detection at 1,250 ft provided optimal performance. More than 50% reduction in delays and stops was found for the approach with platoon arrivals. Also, overall intersection delay and stops were reduced by 20%. With advance detection at 1,250 ft and an approach speed of 65 mph, the system was able to provide 6 to 7 s advance warning of the end of the green in the majority of cases. However, because of the high percentage of turning movements, there was a significant number of cases in which the advance warning time provided was less than a second. The proposed system is suitable for rural high-speed intersections that have fewer turning movements and frequently experience vehicle platoons on the major approach. REFERENCES 1. El-Reedy, T., and R. Ashworth. Platoon Dispersion Along a Major Road in Sheffield. Traffic Engineering and Control, Vol. 19, 1978, pp Robertson, D. I. TRANSYT: A Traffic Network Study Tool. Report LR253. U.K. Transport and Road Research Laboratory, Crowthorne, Berkshire, England, Manar, A., and K. G. Baass. Traffic Platoon Dispersion Modeling on Arterial Streets. In Transportation Research Record 1566, TRB, National Research Council, Washington, D.C., 1996, pp Michalopoulos, P. G., and V. Pisharody. Platoon Dynamics on Signal Controlled Arterials. Transportation Science, Vol. 14, No. 4, 1980, pp Yu, L., and R. F. Benekohal. Platoon Dispersion and Calibration Under Advanced Traffic Control Strategies. In Traffic Congestion and Traffic Safety in the 21st Century: Challenges, Innovations, and Opportunities, ASCE, Washington, D.C., 1997, pp Wasson, J., M. M. Abbas, D. M. Bullock, A. Rhodes, and C. Zhu. Reconciled Platoon Accommodations at Traffic Signals. Publication FHWA/IN/ JTRP-99/1. Indiana Department of Transportation, Indianapolis, Chaudhary, N. A., M. M. Abbas, H. Charara, and R. Parker. Platoon Identification and Accommodation System for Isolated Traffic Signals on Arterials. Publication FHWA/TX 05/ Texas Transportation Institute, Texas A&M University, College Station, Jiang, Y., S. Li, and D. E. Shamo. Development of Vehicle Platoon Distribution Models and Simulation of Platoon Movements on Indian Rural Corridors. Publication FHWA/IN/JTRP-2002/23. Indiana Department of Transportation, Indianapolis, Minnesota Manual on Uniform Traffic Control Devices. Minnesota Department of Transportation, St. Paul, Messer, C. J., S. R. Sunkari, H. A. Charara, and R. T. Parker. Development of Advance Warning System for End-of-Green Phase at High-Speed Traffic Signal. Publication FHWA/TX-04/ Texas Transportation Institute, Texas A&M University, College Station, Klugman, A., B. Boje, and M. Belrose. A Study of the Use and Operation of Advance Warning Flashers at Signalized Intersections. Publication MN/RC-93/01. Minnesota Department of Transportation, St. Paul, Pant, P. D., and Y. Xie. Comparative Study of Advance Warning Signs at High Speed Signalized Intersections. In Transportation Research Record 1495, TRB, National Research Council, Washington, D.C., 1995, pp Sayed, T., H. Vahidi, and F. Rodriguez. Advance Warning Flashers: Do They Improve Safety? In Transportation Research Record: Journal of the Transportation Research Board, No. 1692, TRB, National Research Council, Washington, D.C., 1999, pp Hughes, P. Guidelines for the Installation of Advance Warning Flashers. Technical Memorandum. Engineering Services Divisions, Minnesota Department of Transportation, St. Paul, NEMA Standards Publication TS v02.06: Traffic Controller Assemblies with NTCIP Requirements. National Electrical Manufacturers Association, Rosslyn, Va., Parsonson, P. S. Small-Area Detection at Intersection Approaches. Traffic Engineering, 1974, pp Zegeer, C. V. Effectiveness of Green-Extension Systems at High-Speed Intersections. Research Report 472. Bureau of Highways, Kentucky Department of Transportation, Frankfort, Zimmerman, K. H. Additional Dilemma Zone Protection for Trucks at High-Speed Signalized Intersections. In Transportation Research Record: Journal of the Transportation Research Board, No. 2009, Transportation Research Board of the National Academies, Washington, D.C., 2007, pp The Traffic Signal Systems Committee sponsored publication of this paper.

132 Prediction of Red Light Running Based on Statistics of Discrete Point Sensors Liping Zhang, Kun Zhou, Wei-bin Zhang, and James A. Misener A probabilistic model is proposed to predict red light running (RLR) for collision avoidance systems at arterial signalized intersections. This RLR predictor consists of an arrival time estimator and a statistical predictor of vehicle stop-and-go maneuvers with two discrete point sensors (capable of measuring speed). In addition, unlike most prediction models, which are designed to minimize mean errors, this model identifies two types of error: the false alarm and the missed report. The capability of distinguishing these two types of error is crucial to the effectiveness of RLR-related collision avoidance systems. Therefore, the Neyman Pearson criterion is employed: it keeps the false-alarm rate lower than a given threshold while at the same time minimizing the probability of missing error. To quantify the trade-off between these two types of error in the system design, a system operating characteristics (SOC) function is defined. The system parameters are determined by using an offline supervised parameter-setting procedure in which training data are collected from a field intersection in the San Francisco Bay Area in California with Autoscope video cameras. Effectiveness of the proposed model and its prediction algorithm are demonstrated by the collected field data. The theoretical system performance predicted by the SOC curve is matched with the evaluated performance by means of data collected from field intersections. For example, at a preset false-alarm rate of 3%, the correct prediction rate of RLR for three approaches of the field intersection ranges from 63% to 80%. Red light running (RLR) is increasingly a major national safety issue at signalized intersections. For example, in 2004, 2.5 million or 40% of all police-reported crashes in the United States occurred at or near intersections (1). Of these intersection-related crashes 8,619 or 22.5% were fatal and 848,000 or 46% resulted in injury. The RLR aspect of this problem is known and serious (2 4), since approximately 20% of all intersection crashes occur because of signal violations (4), resulting in an estimated monetization of $13 billion a year when fatalities, lost wages, medical costs, property damages, and insurance are considered (5). Not surprisingly, a significant effort has been undertaken to conceive engineering countermeasures to address this problem. These findings are summarized by Retting et al., and they run the gamut from optimization of signal timing through removing unwarranted signals (6). There has also been significant research and development in recent years in the application of technology to address the California Partners for Advanced Transit and Highways Program, University of California, Berkeley, 1357 South 46th Street, Richmond, CA Corresponding author: L. Zhang, lpzhang@path.berkeley.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / intersection collision problem. The central concept of a dynamic system is to actively inform or alert drivers of emerging, impending hazardous situations. In fact, the U.S. Department of Transportation and state transportation departments have formed several programs to investigate such collision avoidance systems. Efforts include the Intersection Decision Support Program (2, 6 8) and more recently the Cooperative Intersection Collision Avoidance System (CICAS) (3, 6). Academic investigators in these programs are from the University of California, Berkeley, Partners for Advanced Transit and Highways (PATH) Program; the University of Minnesota; and the Virginia Polytechnic and State University (9). Described here are the RLR modeling and prediction approaches within the CICAS Program signalized left-turn assistance and traffic signal adaptation conducted under the auspices of the University of California, Berkeley, effort. In essence, through applying the aforementioned central concept to RLR violations, a provocative and all-round solution, which potentially addresses the dilemma zone, willful violation, and intentional violation motivations, is to dynamically extend an all-red signal phase. All-red phases typically consume 1 2 to 1 s, whereas the solution considered here extends the all-red phase to 2 to 3 s. This approach therefore relies in large part on prediction of a signal violator to enact a dynamic or adaptive allred extension, and it is precisely this probabilistic prediction and implementation into an algorithm that is the focus here. To elaborate, the prediction of RLR is a crucial part of the collision avoidance system, which must be able to sense the dynamic characteristics of a vehicle approaching the intersection from a reasonable range. A dynamic status map of the intersection defined by Zennaro and Misener can then be built on the sensed components (10). From the status map, a violation needs to be predicted to enable the engineering measures such as issuing a warning or dynamically adapting the signal to protect vehicles affected by the violator (11). The problem of RLR prediction is addressed with a combination of predicting the vehicle stop-or-go motion plus estimating the arrival time. Bonneson et al. (12) identified the probability of stopping and the yellow interval duration as the two most significant driver-related factors contributing to the frequency of RLR. Therefore, to predict individual RLR at a signalized intersection with preset yellow interval durations, it is essential to be able to predict if the vehicle is stopping or not. Sheffi and Mahmassani introduced a statistical model in which the driver s decision to stop is normally distributed and a function of critical time (13). The Sheffi Mahmassani model is an important basis for the current research. The critical time itself, however, is not enough to be used as a predictor for individual RLR. More recent research on driver behavior during the intersection approach uses logit models (14, 15) or empirical models (16). Rakha et al. carried out a study in which stopping probabilities are characterized by using detailed driver information, including their perception 132

133 Zhang, Zhou, Zhang, and Misener 133 reaction time (PRT), age, and sex (16). The results as a whole show that the stopping probability is a function of multiple factors, including the distance to the intersection, speed, and headway (17, 14, 13). Among these factors, only the measures available from infrastructurebased sensors were chosen to supplement what is known and to develop an implementable RLR all-red timing and alert algorithm. The choices are the distance to the intersection and speed. On the basis of these measures, the Sheffi Mahmassani probabilistic model is extended to a multidimensional normal distribution in which measured vehicle-to-intersection factors can be incorporated. In short, the stopping of a vehicle is viewed as an evolving event for various distances to the intersection; therefore, the distance to the intersection, speeds of the approaching vehicles, and time into yellow when speed is known are important contributory factors to the model discussed here. When a prediction algorithm is applied to the collision avoidance system, there are usually two types of error: the missing-report error, in which an RLR is reported as a non-rlr (a false negative), and the false-alarm error, in which proceeding through the yellow signal or stopping before the stop bar is accepted as an RLR (false positive). Clearly, in any RLR countermeasure there could be different consequences for the two types of error, with the missing report giving a false sense of security and the false alarm being a nuisance; therefore, a means to minimize and control the balance is essential for any implementation. As an important note in implementing such a balance, the a priori probability of hypotheses can be different. For example, at most intersections, the occurrence of RLR could be significantly less than that of stopping. Thus the requirement for the false-alarm rate should be at or lower than a given threshold. This approach is in contrast to the traditional probabilistic analysis for RLR frequency estimation or dilemma-zone prediction, since these analyses do not distinguish the two types of error. The methodology considered here or the method to control the balance is addressed with the Neyman Pearson (NP) criterion, which governs the prediction of RLR occurrence based on empirical observations by maximizing the probability of prediction (minimizing the missed prediction rate) while keeping the false-alarm rate equal to or lower than a given level (18). By customizing and adjusting the preset level of the false-alarm rate, the system can therefore adapt to various scenarios and intersection requirements. The predictor variables were restricted to only those that could be easily obtained with automatic and real-time point detection methods. The reason for restricting the type of detectors is for deployability and cost considerations. In this study, the emulated speed loops of Autoscope cameras were used as discrete sensors. In summary, the study described expands previous work in several ways: In the application dimension, modeling and algorithms were explicitly developed for predicting RLR at arterial intersections for collision avoidance systems and In the probabilistic modeling aspect, the traditional onedimensional normal distribution model was extended to a multidimensional normal model. In addition, the two types of error of the predictor (false negative and false positive) were segregated by applying the NP criterion such that the system performance could be customized for different scenarios. From that a system operating characteristic (SOC) function was developed to quantify the trade-off in system design. The resulting model and predictor for RLR requires only point detectors, which gives it the potential for a wider range of applications. MODELING AND PREDICTING RLR RLR Model with Point Detector Data Three types of maneuvers for vehicles approaching an intersection are of interest: being the first to stop, going through the yellow, and going through the red (RLR). The latter two scenarios fall into the going group. These maneuvers are not driver-based taxonomy, and coupled with observation of only the vehicles, they obviate the need to define more microscopically driver-related information to include perception reaction time (PRT) and other personal information. Two of K point detectors (k 1, k 2 ) were chosen for modeling and prediction purposes. The group of detectors at various distances from the intersection is { 1 k1... k2 K} The detected speeds at these locations are { v k, v k } ( ) and the time instant (relative to the yellow onset) when these measurements are obtained is { t k, t k } ( ) The time stamps in Equation 2 are not independent, assuming that vehicles move at constant acceleration between the sensors. In this case, the data are highly dependent. Thus only the latest data point t(k 2 ) will be used. The predicted arrival time at the intersection from sensor k 2 is denoted tˆi(k 2 ), which is defined as ˆtI ( k2) = t( k2) + estimated time to intersection (3) In Figure 1, an example setup is plotted to show how these discrete observations were obtained. This diagram is for the eastbound part of the field intersection (Table 1). Six discrete sensors were configured to the collected data from approaching vehicle maneuvers and outputs (speeds and time stamps) from two of the sensors (Sensors 3 and 4) that were chosen to work as the predictor. (The selection of sensors for prediction is discussed later.) First, a probabilistic model of the prediction problem for the two types of scenarios is presented. The observations are T { v k, v k, t k } (4) where T is the transpose of a matrix. There are two groups of hypotheses: for the go-through maneuver (g) and the first-to-stop (s) maneuver. Among the go-through vehicles, the RLR and go-through-in-yellow vehicles differ only in their arrival time at the intersection, or by definition ti > Y t Y I ( ) ( ) ( ) ( ) ( ) ( ) ( ) RLR Go through yellow RLR is defined as arriving at an intersection (time of arrival at the intersection is denoted as t I ) after the red onset, and Y is the duration of the yellow interval. (5)

134 134 Transportation Research Record d(1) Distance to Intersection (ft) Estimated time to intersection d(2) d(3) d(4) = d(k 1 ) = d(k 2 ) 40 d(5) 15 0 d(6) d(7) Stop bar t(k 1 ) t(k 2 ) Time (sec) t 1 tˆ1 (k 2 ) t 1 : Actual arrival time at intersection FIGURE 1 Time space diagram of example setup. TABLE 1 Configuration of Field Intersection Page Mill Rd. at El Camino Real, Fully Actuated Signal Control Eastbound Westbound Southbound Lanes (recorded) Speed limit (mph) Average speed (mph) % speed < (mph) Yellow duration (s) All-red interval (s) Speed loops (exit) Speed loops (advance) Speed loops (stop) % RLR time-into-red < (s) 1.5s 1.5s 1.5s Description of the approach 0.15 mi from upstream No upstream intersection for 0.25 mi from upstream intersection more than 0.7 mi upstream intersection, main street Average speed of RLR 30 mph 35 mph 38 mph 85% of RLR speed < 40 mph 45 mph 45 mph 85% of v t inred of RLR < 60 ft 90 ft 90 ft Locations of speed loops for 63, 97 95, , 120 prediction (ft from stop bar) Locations of all speed loops at 0, 15, 40, 63, 97, 120, 190 0, 20, 40, 95, 139, 190 0, 30, 90, 120, 212 stop and advance area

135 Zhang, Zhou, Zhang, and Misener 135 The test statistics of the lower dimension are constructed here to compact the data while maintaining the ability to distinguish between hypotheses. The two-dimensional observation is defined as 1 v = v( k_ 1) v k_ 2 2 ( )+ ( ) v k v k ( ( )= ) ( ) t( k ) t( ) ak and the estimated arrival time tˆi(k 2 ) as X ˆ x v, x a k Since only RLR associated with going through is of interest here, the observation vector is conditionally modeled on the estimated arrival time. Because many random factors contribute, X can be reasonably hypothesized to be a bivariate normal random vector for either t I > Y or t I < Y. Thus to formulate the hypotheses, H H s g and = { = = ( )} Hg ti Y : X N ug, g (9) where 2 k 1 T (7) ( ) ( ) t > Y : X N u, I s s t > Y : X N u, I g g ( ) N = probability distribution function of the normal distribution; H s and H g = hypotheses for stopping and going through; u s, Σ s and u g, Σ g = mean and covariance matrices of X under the two hypotheses when t I > Y, respectively; and u g, Σ g = mean and covariance of X for vehicles going through when t I Y. Equation 8 defines the set of hypotheses for when predicted arrival time is later than a given threshold, and Equation 9 for the hypotheses when arrival time is earlier than the threshold. Since the arrival time to the intersection for H s is always greater than Y, there is no hypothesis H s t I Y. Prediction of RLR Based on NP Criterion To derive the prediction method for RLR, several definitions regarding RLR errors were generated. Definition 1. The selected vehicle measurements of interest are those from vehicles approaching the intersection during the yellow interval that meet the following condition: 0 t( k1 ) < Y (10) In other words, the vehicle arrives at the point detector k 1 during the yellow. Carefully selected d(k 1 ) can guarantee that most of the RLR will be covered in these measurements. For example, if d(k 1 ) is equivalent to 2 s to the intersection at the posted speed limit, the zone covered by Equation 10 ranges from 2 to Y + 2 s to the intersection; (6) (8) this range would cover most instances of RLR. The selection of d(k 1 ) and d(k 2 ) will be discussed later. Also excluded are all the samples when the above-mentioned vehicles are moving very slowly because of heavy traffic. In such scenarios, vehicles may experience stopping and going maneuvers, which are unlikely to be captured by only two point sensors. A lower speed limit of 10 mph is the average speed above which the case is of interest. During the field observation period for over one month, no RLR case was seen with a speed less than 10 mph. Definition 2. RLR is the event in which a vehicle arrives at the intersection later than the red onset and proceeds to go through the intersection: Definition 3. The predictor of vehicle maneuvers gives a binary decision between the hypotheses for stopping or going through: D X which distinguishes two hypotheses: H s t I > Y and H g. Definition 4. Predictor D(X) would present two types of error for t I > Y, the probability of which is a false alarm (FA): or a missing (M) report: M ( ) = { ( ) = } P s g ˆ P D X H H ( ) respectively. Definition 5. The prediction of RLR is a statistical decision based on the stop-versus-go prediction D(X) and the estimated arrival time. Because most RLR vehicles move at a constant speed or accelerate to go through at arterial intersections, a conservative estimator of tˆi(k 2 ) can then be defined such that P{ t ˆ I( k Y ti Y, ( ) 2 ) > go through} 0 15 In this case, for RLR prediction purposes, one can simply reject the scenario when tˆi(k 2 ) Y. Only those observations that satisfy the condition of tˆi(k 2 ) > Y could possibly be predicted as RLR. Thus, if the predictor of RLR is denoted as D RLR (X), it can be defined as follows: D RLR { I } RLR ˆ= H g, t > Y (11) ( )= { Hs Hg} ˆ or ( 12) P g s ˆ P D X H H, t Y ( ) FA ( )= ( )= g s I > 13 1 ( X)= 0 Definition 6. There are two types of errors of the RLR predictor. The missing report for predictor D RLR (X) is { } PM RLR= ˆ P DRLR( X)= 0 RLR ( 17) The false-alarm rate is { } ( )= ( )> if D X and tˆ k Y otherwise H g { } s g 14 PFA RLR = ˆ P DRLR ( X)= 1 not RLR ( 18) I 2 ( 16)

136 136 Transportation Research Record 2128 The rate of false alarms to all normal samples (except for RLR, which is rare) is the false-alarm rate, a low level of which could contribute many false reports, as can be seen from the previous example. That is one reason why the proposed method was developed, which is able to control the false-alarm rate for optimization. The definitions of missing and false-alarm rates in Equations 17 and 18 are simple. Their evaluations, however, are challenging. To derive an optimized detection algorithm, D RIR (X) is linked to D(X) to make the optimization a more tangible problem. The false-alarm rate and the missing rate for RLR prediction can be approximately expressed in terms of the probabilities of error of D(X): P P g s i P H s c ( 19) FA RLR FA ( ) ( )+ P P s g M RLR M( ) ( 20) P FA (g s) is defined in Equation 13 and P M (s g) in Equation 14, and the constant c can be determined by using historical data. The rationale of this approximation lies in the concern about the cases in which a stopping vehicle is falsely reported as RLR. Another reason is that there is relatively high confidence in the estimation of arrival time for vehicles going through the intersection since they tend to move at more constant speeds than vehicles that will stop. From the definitions of the RLR event and the two types of error related to its prediction, the (sub)optimal D(X) is formulated by using the NP criterion. Because of the highly related probabilities of their errors, D(X) is optimized instead of D RLR (X). P{X H s, t I > Y} and P{X H g, t I > Y} are defined as the probability density functions (PDFs) of X under hypotheses H s and H g, respectively (see Equation 8), and β is the false-alarm rate for D(X). To form the NP predictor, which minimizes the missing probability while keeping the false-alarm rate not greater than β, the following constrained optimization problem needs to solved: v> x 1 { v0, a0}= argmax P X, tˆ I k ( x1, x2) H ak ( g ( 2 )> Y dx 2)> x2 for any x, x, such that ( ) ak ( 2 )> x2 { } To keep the false-alarm rate of RLR less than a given threshold θ, θ c β = ( 22) P H s ( ) { } { } v> x 1 P X H s, tˆ I ( k2 )> Y dx is used where β is the false-alarm rate set in Equation 21. By using only the rectangular boundary as in Equation 21, the solution is no longer optimum; rather it is a suboptimal approximation to the NP solution. Also, no attempt is made to find an even simpler form of sufficient statistics of X because it is believed that the driver s decision-making procedure is affected by a host of factors that may be reflected in the observation data. It is therefore unlikely to find a single sufficient statistic in scalar form to contain all the information needed for the classification. Figure 2 shows a bivariate stop go model generated from 29 days of data collected at the field intersection (see Table 1, eastbound, for details). The left peak belongs to H s and right peak is H g. Several characteristics can be noted from this example: 1. The (first-to) stop maneuver and the going-through maneuver have distinct distributions in terms of their speeds and speed differences at two advance point sensors. This difference in their empirical statistics forms the basis for the prediction algorithm. 2. The distributions are also partially overlapped, which is exactly the reason for prediction errors. This overlap reflects the complex nature of driver behaviors, to which many factors may contribute, including the driver s PRT, the traffic, and others. The imperfection of the sensors may also contribute to the errors. The prediction β Probability Density acceleration (m/s 2 ) average speed (mph) v a acceleration (m/s 2 ) average speed (mph) FIGURE 2 Bivariate normal model of two hypotheses.

137 Zhang, Zhou, Zhang, and Misener 137 method has to focus on the minimization (and trade-off) of these errors according to some practical requirements, which is proposed here to be the NP criterion. From the field data, the difference in the speeds of the first-to-stop and going-through maneuvers was observed at an advance area (e.g., from 60 to 100 ft). This difference gets even more significant when vehicles were closer to intersection. For example, at 30 ft from the intersection, the stop-go maneuvers were well separated. However, as will be discussed later, advance sensors cannot be located that close because in that case some of RLR events will not be predicted before it is too late for the collision avoidance system to react. From field observations, the statistics of vehicle speed and acceleration (speed change between the two sensors) of the first-to-stop and going-through in the yellow or red were extracted as an example. The average speeds for first-to-stop and going-through vehicles ( ft from the stop bar) are 26 mph and 35 mph, respectively, and the acceleration rates are 2 m/s 2 and 0.2 m/s 2. There is a significant difference in the speed and acceleration for the two scenarios, especially the difference in the acceleration, which is just as was expected since the going-through vehicles tend to move at constant speed whereas stopping (especially the first-to-stop) vehicles would experience (sudden) deceleration while approaching the intersection. More details of the statistics are summarized in Table 1. The difference in first-to-stop and going-through vehicles is not significant for peak hours and off-peak hours. Rather cycle-to-cycle or even driver-to-driver differences appear to be significant. Study at more intersections is still needed before any conclusions can be drawn on such details of RLR driver behaviors. Implementation of RLR Prediction Algorithm The RLR model and its prediction method presented in the previous section is an online algorithm that utilizes statistics of the measurements obtained from historical data. Data from a group of point detectors are collected, and those from d 1 and d 2 are used for online prediction purpose, whereas all the detectors need to work together to get the offline statistics. The offline supervised parameter-setting procedure to quantify the RLR prediction algorithm is described as follows: Input: a set of trajectories {Z n } and false-alarm requirement θ; Each Z n contains measurement for RLR detection X n and measurements from other detectors; For each historical trajectory Z n = {t(k), v(k), including the stop bar} if v (stop bar) in Z n and v (stop bar) > v min X n is H g ; If t (stop bar) > Y X n is RLR; end Else X n is the stopping vehicle (check memory to see if it is the first to stop); End End Construct the estimator tˆ I(k 2 ); For all {X n H g and t (stop bar) > Y}, calculate û g and Σˆ g; For all {X n H s, first to stop}, calculate û s and Σˆ s; Calculate Pˆ (H g ), Pˆ (H s ) and the constant c; For a in [a min, a max ] and v in [v min, v max ] Calculate the false-alarm rate P FA (g s) and the missing rate P M (s g); If P FA (g s) < (θ c)/pˆ (H s ), find the (a opt, v opt ) pair that minimizes P M (s g); Output (a opt, v opt ) that minimizes P M (s g) while P FA (g s) < (θ c)/ Pˆ (H s ). Once the optimized decision boundary has been obtained, the online prediction algorithm can be formulated as follows: Some details about the offline supervised parameter-setting algorithm are described next. Estimation of Mean and Variance of Two Hypotheses It is assumed that the number of historical samples for H g is N g and N s for H s, respectively. The mean and variance in Equation 8 can be estimated by using Equations 24 and 25, respectively: and ( )= ( )> ( )> D X if a k a and v k v and tˆ ( k )> Y, RLR 1 2 opt 2 opt I 2 1 uˆ g = ( Xn Hg) ( 24) N g ˆ 1 T g = ( Xn Hg)( Xn Hg) ( 25) N 1 g n 0 otherwise ( 23) For X n H s, the estimator is similar. Arrival Time Estimator The arrival time estimator tˆ I(k 2 ) needs to meet the assumption in Equation 15. A simple estimator can be formed as follows: ˆ d k2 ti k2 t k2 ( 26) v k ( ) = ( ) ( ) + ( ) n which assumes that the vehicle moves at constant speed. To incorporate the measured acceleration at d(k 2 ), a slightly different form is adopted: ˆ d k2 t k t k ˆ I 2 2 ρ ak2 ak2 ( v k 27) ( ) = ( ) ( ) + ( ) + ( ) ( ) where â (k 2 ) is the estimated mean acceleration and ρ < 0 is a constant coefficient that rejects some vehicles that go through legally by accelerating after the yellow onset; ρ should be kept small so that most RLRs are still accepted. Empirical data show that ρ= 0.05 s 3 /m is a good option (a is in meters per second squared and t is in seconds). Performance of RLR Predictor 2 2 The probability of correct prediction (defined as P D = 1 P M ) as a function of the probability of false alarms is frequently used to quantify the performance of detection systems. With radar systems, where

138 138 Transportation Research Record 2128 the NP criterion has been extensively used (19), this function is called the receiver operating characteristics. For an intersection collision avoidance system, a similar SOC function is defined as SOC θ where P FA ( )= max PD [ D( X) ] θ c [ D( X) ] ( 28) P H s ( ) DATA COLLECTION AND SIMULATION Field Setup and Configuration of Point Detectors The point detectors at the advance area are for RLR modeling and prediction. The ones at the stop area and exit area are for supervision of the algorithm results as well as for establishment of statistics of RLR occurrences. The aforementioned statistics determine the locations of the speed loops of the Autoscope cameras. Each individual intersection presents different characteristics in terms of approaching driver behaviors. Thus point detectors were placed at different locations, depending on the intersection. The constraint is that the advance detector closest to the intersection (with distance denoted as d min ) on a given approach must satisfy a preset system requirement. An important preset requirement in this system is that protection (or warning) is provided for 85% of the RLR occurrences. In other words, for at least 85% of the RLRs, the system should be able to detect the RLR (the RLR vehicle arrives at the last sensor) before it is too late (red onset). The time-into-red of RLR occurrence is denoted by t inred when the vehicle actually arrives at the stop bar (there is also a sensor at the stop bar to monitor RLR occurrences). The distance of the vehicle to the intersection at t inred after red onset is zero (the vehicle just arrives at the stop bar). The vehicle is v t inred m away from the stop bar at the time of red onset, when the vehicle should have passed the last sensor or it would not be captured by the all-red extension system. The following is the mathematical requirement of sensor placement: P( v i tinred < dmin RLR)> 085. ( 29) The minimum distances to the intersection for different approaches were obtained by using empirical statistics, which are shown in Figure 3. From Figure 3b, it was observed that for the eastbound approach of the field intersection, over 85% of the RLRs were 60 ft or less away from the intersection at the time of red onset, and this distance for the other two approaches was 90 ft. So d min was set to 60 ft for eastbound and 90 ft for the other two approaches. Details of the configurations of the field intersection are summarized in Table 1. The aforementioned method to determine the minimum location of the advance sensor was a two-step procedure. In the first step multiple sensors were placed at various locations about 200 ft from the intersection to the stop bar with the Autoscope configuration software tool. Then from the data collected from these sensors, the empiri- Percentage (histogram) Distribution of t inred v of RLR Minimum distance required to detect (feet) (a) Page Mill Rd EB Page Mill Rd WB Page Mill Rd SB Percentile (cumulative distribution) Page Mill Rd EB Page Mill Rd WB Page Mill Rd SB Minimum distance required to detect (feet) (b) FIGURE 3 Empirical distributions of t inred v of RLR: (a) minimum distance versus percentage and (b) minimum distance versus percentile.

139 Zhang, Zhou, Zhang, and Misener 139 cal distribution was built of v t inred (as shown in Figure 3) to get the minimum distance of the advance sensor. Then the next step was simply to move (pick up) two of the advance sensors to leave one at the minimum distance and the other further away. Since the acceleration at the advance area is needed, these two sensors should not be put far apart usually 30 to 60 ft is a good choice. Simulation Results Computer simulation was carried out to verify the RLR prediction algorithm. Data collected in the field were used to generate the parameters for each approach of the site. Additional data were collected to verify the algorithm. Data Collection and Processing Field data were collected with the Autoscope cameras. Nine cameras were installed at three approaches of an arterial intersection (Page Mill Road and El Camino Real, California State Route 82) in the San Francisco Bay Area. Three cameras were installed for each approach, covering respectively the exit area (inside intersection), stop area (90 to 0 ft to stop bar), and advance area (200 to 90 ft to stop bar). The data sets were divided into two parts: one was for parameter setting and the other for simulation and algorithm verification. The first part was collected January 18 to 21 and March 2 to 11, 2008, 14 days in total. The second part was collected May 15 to 28 and June 6 to 20, 2008, 29 days in total. The total number of interesting samples (as defined by Definition 1) is 37,894. Details are summarized in Table 1. Simulation Parameters Data sets were collected during March The offline supervised parameter-setting results are shown in Table 2. These results were used as prediction parameters. Sensor Imperfection It is quite common that sensors would miss the report of a vehicle or falsely report a vehicle that is absent. During the simulation, for sensors k 1 and k 2, if measurement at either of them is missing, a duplication from the previous sensor would be used (speed set to the same as the last sensor reports; thus acceleration is zero). So if the vehicle was actually stopping, this missing by the sensor would probably make the system falsely predict the vehicle as going through. The errors caused by sensor imperfection were included in the simulation results. Calibration of Sensors The speed loops of the Autoscope recorded the speed when a vehicle was at a specific location. This speed was calculated (by embedded Autoscope image-processing software) on the basis of the image sequence of the vehicle entering the virtual loop. To verify the accuracy of the engineering data generated from the Autoscope processors, a fully instrumented vehicle with a Global Positioning System device and data recorder was driven at the test site for 2 days. The calibration results showed that the errors of the Autoscope speed loops were within 10% most of the time. To validate the measurements from the Autoscope, part of the RLR data was also compared with the recorded video clips. The comparison showed that the engineering data of the Autoscope were consistent with the raw video data. Building Trajectories from Point Detectors To associate the data of the discrete point sensors with a vehicle, a multiple-hypothesis tracking (MHT) algorithm developed by Reid was employed (20). MHT is a widely used tracking method (21) that assumes that each discrete measurement could possibly be assigned to a previously found target (vehicle) with an a priori probability. Whenever a new measurement is reported by the sensor, it can be associated with (assigned to) any previous targets (vehicles), thus forming a set of new hypotheses. The MHT algorithm then calculates the evolving probability of each new hypothesis and finally (for example, when the measurement from the exit area has been associated with some track), the hypothesis with the highest probability is accepted. Details of the MHT tracking algorithm used are beyond the scope of this paper and can be found online at lpzhang/index_files/docs/ MHTalgorithm.pdf. Simulated System Performance Data sets collected during May and June 2008 were used to verify the algorithm. The actual SOC curves for the field intersection show that the simulated online performance of the prediction algorithm (Figure 4) is a good match with the theoretical SOC curves for the TABLE 2 Offline Supervised Parameter-Setting Results Page Mill Rd. at El Camino Real, Fully Actuated Signal Control Eastbound Westbound Southbound RLR occurrence H s occurrence 1, ,672 H g occurrence 5,453 3,920 8,319 Constant c Pˆ(H s ) Σˆ g û g Σˆ s û s

140 140 Transportation Research Record Probability of Correct Detection for RLR Simulated SOC curve with field data (90 and 60 feet) Theoretical SOC curve (90 and 60 feet) 0.1 Simulated SOC curve with field data (120 and 90 feet) Theoretical SOC curve (120 and 90 feet) Probability of False Alarm (a) P d Probability of Correct Detection for RLR Simulated SOC curve with field data Theoretical SOC curve based on training statistics P fa Probability of False Alarm for RLR (b) P d Probability of Correct Detection for RLR Simulated SOC curve with field data Theoretical SOC curve based on training statistics P fa Probability of False Alarm for RLR (c) FIGURE 4 SOC curves for RLR prediction with field data: (a) eastbound approach, (b) westbound approach, and (c) southbound approach.

141 Zhang, Zhou, Zhang, and Misener 141 low false-alarm rate region. However, the actual prediction algorithm cannot achieve an arbitrarily low missing rate. Since the optimization procedure here only involves the NP optimization of the stop-versus-go prediction parameters and the arrival time estimator is not parameterized, the missing error caused by the arrival time estimator is a constant in the predictor and cannot be traded off in the predictor. Comparison of the SOC curves shows that the closer the point sensors are, the better RLR prediction the system can achieve. An example is the eastbound experiment, where two setups with different distances were compared. The first pair of virtual speed loops was at 90 and 60 ft away, and the second pair was at 120 and 90 ft. With the same set false-alarm rate at around 3%, the first setup achieved a correct prediction rate of 80% compared with 63% in the other case. This result is mainly due to the better separation of the statistics of stopping and going-through vehicles and a smaller variance in the arrival time estimator. It is also noteworthy that the differences in the statistics for different approaches lead to different system performance. This difference was most likely caused by variations in traffic for different approaches. The application of the dynamic all-red extension also raises the concern of how drivers might adapt to this feature and in the long run undo the benefits of the system. The authors believe that by improving the hazard prediction algorithm, the triggering rate of the all-red extension could be kept to a very low level to minimize the influence on driver behavior. The combination of a dynamic system with enforcement, such as the RLR camera, could also be a requirement for successful deployment of such systems. ACKNOWLEDGMENTS This work was sponsored by California Business, Transportation and Housing Agency, California Department of Transportation, and by U.S. Department of Transportation as part of the California Cooperative Intersection Collision Avoidance System Project. The authors thank Bob Pang for review of the video clips and Meng Li for help in calibrating the sensors. They also thank Steve Candy from Econolite for his help in fine-tuning the Autoscope speed loops. CONCLUSION The prediction of RLR based on a probabilistic model was investigated. The event of stopping (after the yellow onset) was modeled as an evolving event while the vehicle is approaching the intersection. Speed measures of a set of discrete point sensors (a minimum of two) and their corresponding event time stamps are used to identify the RLR, where the decision boundaries were formed from the NP criterion based on the empirical statistics. By employing the principle of NP detection, the system was able to obtain a trade-off between two types of prediction errors. Moreover, the SOC function has been defined to quantify this trade-off. The simulation study using data collected from the field demonstrated the feasibility of the described approach, where the SOC curves match the performance predicted with the theoretical model. It was found that for the same intersection but different approach directions, the average and variance matrices of speed for the going-through and stopping scenarios were different, with each approach having a unique system of performance measures. Regardless of these differences, the closer the sensors, the better the performance. The main limitations of using this technique in the field were found to be as follows: Because of the prohibitive complexity of the multidimensional optimization involved in solving the NP detection, the suboptimum solution of two-dimensional observation vectors for given distances was used. This approach could limit the performance of the system. The offline parameter-setting procedure for the collision avoidance system involves a large amount of data collection for each individual approach and intersection. The operational impact of these data is to be determined; certainly, further data collection and analysis are needed to identify the similarity among different intersections and to find more general decision boundaries that can serve more intersections to standardize the setup of the RLR collision avoidance system. REFERENCES 1. Bellomo-McGee Incorporated. Intersection Collision Avoidance Study, Final Report. Report FHWA/TX-03/ FHWA, U.S. Department of Transportation, Ragland, D. R., and A. A. Zabyshny. Intersection Decision Support Project: Taxonomy of Crossing-Path Crashes at Intersections Using GES 2000 Data. Report UCB-TSC-RR University of California, Berkeley, Bonneson, J., D. Middleton, K. Zimmerman, H. Charara, and M. Abbas. Development and Evaluation of a Detection-Control System for Rural Intersections. Report FHWA/TX Texas Transportation Institute, Texas A&M University, College Station, Brittany, N., B. N. Campbell, J. D. Smith, and W. G. Najm. Analysis of Fatal Crashes Due to Signal and Stop Sign Violations. Report DOT HS NHTSA, U.S. Department of Transportation, Chang, J., D. Cohen, L. Blincoe, R. Subramanian, and L. Lombardo. CICAS-V Research on Comprehensive Costs of Intersection Crashes. Report NHTSA, U.S. Department of Transportation, Available from www-nrd.nhtsa.dot.gov/pdf/nrd-01/esv/esv20/ O.pdf. 6. Retting, R. A., A. F. Williams, and M. A. Greene. Red-Light Running and Sensible Countermeasures: Summary of Research Findings. In Transportation Research Record 1640, TRB, National Research Council, Washington, D.C., 1998, pp Chan, C.-Y. Characterization of Driving Behaviors Based on Field Observations of Intersection Left-Turn Across Path Scenarios. IEEE Transactions on Intelligent Transportation Systems, Vol. 7, Chan, C.-Y., D. R. Ragland, S. E. Shladover, J. A. Misener, and D. Marco. Observations of Driver Time Gap Acceptance at Intersections in Left-Turn Across-Path Opposite-Direction Scenarios. In Transportation Research Record: Journal of the Transportation Research Board, No. 1910, Transportation Research Board of the National Academies, Washington, D.C., 2005, pp Neale, V. L., M. A. Perez, S. E. Lee, and Z. R. Doerzaph. Investigation of Driver Infrastructure and Driver Vehicle Interfaces for an Intersection Violation Warning System. Journal of Intelligent Transportation Systems, Vol. 11, 2007, pp Zennaro, M., and J. A. Misener. A State Map Architecture for Safe Intelligent Intersections. Presented at 13th Annual Meeting of ITS America, Minneapolis, Minn., 2003.

142 142 Transportation Research Record Bonneson, J. A., K. H. Zimmerman, and M. Brewer. Engineering Countermeasures to Reduce Red-Light Running. FHWA, U.S. Department of Transportation, Aug Bonneson, J. A., M. Brewer, and K. H. Zimmerman. Review and Evaluation of Factors That Affect the Frequency of Red-Light Running. Report FHWA/TX-02/ FHWA, U.S. Department of Transportation, Sheffi, Y., and H. Mahmassani. A Model of Driver Behavior at High Speed Signalized Intersections. Transportation Science, Vol. 15, No. 1, Feb. 1981, pp Bonneson, J. A., and H. J. Son. Prediction of Expected Red-Light Running Frequency at Urban Intersections. In Transportation Research Record: Journal of the Transportation Research Board, No. 1830, Transportation Research Board of the National Academies, Washington D.C., 2003, pp Gates, T. J., D. A. Noyce, L. Laracuente, and E. V. Nordheim. Analysis of Driver Behavior in Dilemma Zones at Signalized Intersections. In Transportation Research Record: Journal of the Transportation Research Board, No. 2030, Transportation Research Board of the National Academies, Washington D.C., 2007, pp Rakha, H., I. El-Shawarby, and J. R. Setti. Characterizing Driver Behavior on Signalized Intersection Approaches at the Onset of a Yellow Phase Trigger. IEEE Transactions on Intelligent Transportation Systems, Vol. 8, Dec. 2007, pp Goh, P. K., and Y. D. Wong. Effects of Distance and Speed on Driver Stop-Versus-Cross at Traffic Signals. Road and Transportation Research, Vol. 13, Sept Kay, S. M. Fundamentals of Statistical Signal Processing, Vol. 2; Detection Theory. Prentice Hall, Upper Saddle River, N.J., DiFranco, J. V., and W. L. Rubin. Radar Detection. Prentice Hall, Upper Saddle River, N.J., Reid, D. B. An Algorithm for Tracking Multiple Targets. IEEE Transactions on Automatic Control, Vol. AC24, 1979, pp Blackman, S. S. Multiple Hypothesis Tracking for Multiple Target Tracking. IEEE Transactions on Aerospace and Electronics Systems, Vol. 19, No. 1, Jan. 2004, pp The contents of this paper reflect the views of the authors, who are responsible for the facts and accuracy of the data. The contents do not necessarily reflect the official views or policies of the State of California. The Traffic Signal Systems Committee sponsored publication of this paper.

143 Microscopic Modeling of Traffic Signal Operations Comparative Evaluation of Hardware-in-the-Loop and Software-in-the-Loop Simulations Aleksandar Stevanovic, Ahmed Abdel-Rahim, Milan Zlatkovic, and Enas Amin Currently, there are three different methods to model traffic signal operations in microscopic simulation models: the simulation model s controller emulator, hardware-in-the-loop simulation, and software-inthe-loop simulation. Although all three methods can be based on the same industry standard code, their different implementations suggest potential operational differences. This study investigates operational differences of the three methods by examining how each method operates in five experimental scenarios. Each of the scenarios differs from the others in network size (one intersection to five intersections) and operational strategies (pretimed, actuated, actuated coordinated, and two different signal transition logics). Ten 75-min simulation runs with 100-ms simulation resolution were executed for each experiment with the three signal control modeling alternatives. The results showed that for basic signal control operations, such as pretimed and isolated actuated operations, the three alternatives provided similar results as indicated by the average green time allocation and different operational measures of effectiveness. When advanced controller operations were used, such as signal transition logic, the simulation model emulator showed significantly different behavior than that observed in hardware-in-the-loop and software-in-the-loop simulations. Most of the current traffic microsimulation packages, which are used to simulate traffic operations on urban arterials, consist of two components: a simulator of traffic flows and a generator of traffic signal states. In the simplest case, a generator of traffic signal states sends the current signal status to the simulator (e.g., pretimed traffic control). Two-way communication is used when actuated traffic control is modeled. In this case, the traffic simulator records the activations of virtual detectors and sends the inputs to the traffic control generator. The traffic control generator processes the inputs through its traffic A. Stevanovic and M. Zlatkovic, Department of Civil and Environmental Engineering, University of Utah, 122 South Central Campus Drive, Room 104, Salt Lake City, Utah Current affiliation for A. Stevanovic: Department of Civil Engineering, Florida Atlantic University, 777 Glades Road, Building 36, Room 231, Boca Raton, FL A. Abdel-Rahim and E. Amin, Civil Engineering Department, University of Idaho, P.O. Box , Moscow, ID Corresponding author: A. Stevanovic, aleks.stevanovic@fau.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / control logic and returns current signal states to the simulator. Interaction between the two components can be more complex when advanced types of traffic control (e.g., adaptive) are modeled. This study focuses on the traffic control component. The first traffic control logics implemented in traffic microsimulation were generated by internal microsimulation programs, which simply changed traffic signal states at predetermined intervals. This concept was later enhanced to provide for actuated traffic control logic within the microsimulation model. The real enhancements on the traffic control side came when external field controllers were coupled with microsimulation through a concept called hardwarein-the-loop simulation (HILS) (1). Recently, the HILS concept has been advanced through its software version known as software-inthe-loop simulation (SILS). Each of the new approaches improved communications between traffic simulator and traffic controller. However, more sophisticated approaches tend to be more expensive and more computationally demanding and require more expertise than less sophisticated approaches. Sometimes when basic traffic control functions are tested, it is advantageous to use a simpler and faster approach. This study investigates three ways of connecting a generator of the traffic control logic with a traffic microsimulator. An internal microsimulation emulator of traffic control is compared with the HILS and SILS concepts. VISSIM microsimulation, which supports all three concepts, is used to test a variety of traffic control operations on a case study network. Operations of the three traffic control concepts and their impact on traffic performance in the network were analyzed. RESEARCH BACKGROUND Emulation-in-the-Loop Simulation As an internal source of traffic control, traffic simulation software uses its own emulator. The simulation software (e.g., CORSIM and VISSIM) may have emulators based on National Electrical Manufacturers Association (NEMA) standards. During a simulation a traffic simulator passes the status of its detectors and signal heads to the emulator of the NEMA controller each simulation second and the emulator returns the state of the signal heads for the next simulation second. Because this emulator does not have any counterpart in the field, it was decided to use the phrase emulator-in-the-loop 143

144 144 Transportation Research Record 2128 simulation (EILS) to refer to this concept. The major disadvantage of using EILS is that this concept does not provide sophistication and variety of control operations to a field controller. The advantages are higher simulation speed, ease in setting up signal timings, perfect coordination with the traffic simulation model, and low installation costs (the emulator comes together with the main software). HILS Concept The basic idea of HILS is as follows: a simulation model generates detector input data, which are then sent through the controller interface device (CID) to the actual traffic controller. The CID is a piece of hardware that provides the interface from the discrete logic levels of the control pins on the controller to the computer running the microscopic traffic simulation. The traffic controller analyzes the detector input data, determines the status of signal control, and sends the data back to the simulation model through the CID. This data exchange between the simulation model, the CID, and the simulation model is done every simulation time step. When a CID is used, the real traffic signal controller replaces the internal controller emulation logic of the simulation program. The CID functions as a bridge between the electrical signals of the computer and those of the traffic signal controller. The HILS concept was first developed as the package by the University of Louisiana (2). There is a long list of HILS deployments and testing studies performed since then (1 10). Currently, there are at least a half dozen university centers across the country where HILS is used for testing, research, and education. The major disadvantages of HILS are an inability to run faster or slower than real time, lack of synchronization between controllers and the computer s clocks, and separate controller hardware required for each intersection (2). SILS Concept Major HILS issues were addressed when SILS was developed. SILS consists of a microscopic simulation model, a virtual traffic controller running on the same computer, and an interface that allows for communication and exchange of information between the microscopic simulation and the virtual traffic controller. At least two SILS applications have been developed: Siemens s NextPhase, which is linked to CORSIM and VISSIM (2), and Econolite s ASC/3, which connects to VISSIM (the only SILS commercially available). PTV America and Econolite Control Products, in cooperation with the University of Idaho (the MOST Project), have developed an ASC/3 SILS controller embedded in VISSIM. The ASC/3 SILS concept enables the use of multiple virtual ASC/3 controllers, all capable of running signal timings faster or slower than real time. The ASC/3 controllers in SILS are compliant with the National Transportation Communications for Intelligent Transportation Systems Protocol; they run from the same code base as the ASC/3 hardware controllers, which makes them nearly identical. The major disadvantage of SILS is that it does not have the features of a real controller that support communications within a cabinet or centralized traffic signal system (2). Both EILS and SILS are essentially both emulators and software packages. The term software is used in the abbreviation SILS because it represents the same version of software that is deployed in a field controller. For EILS the term emulator is used because its emulation of a (NEMA) traffic controller is much less realistic than that of SILS. In the following experiments described, internal VISSIM EILS and ASC/3 Econolite controllers in SILS and HILS systems were used. Real-Time Systems Both HILS and SILS can be considered statically scheduled realtime systems. Real-time systems can be classified on the basis of their ability to tolerate failures to meet deadlines to hard systems and soft systems. Hard real-time systems cannot tolerate failures to meet deadlines. Examples of hard real-time systems are aircraft or missile flight control systems. A deadline failure in these systems could cause the loss of an aircraft or a missile. Soft real-time systems can tolerate some deadline failures and still function correctly. Both HILS and SILS are examples of soft real-time systems. Here, an occasional missed deadline may affect some measures of effectiveness (MOEs) but should not cause the simulation to fail completely. Therefore, if the HILS SILS system is overloaded during a specific simulation time step and unable to finish its tasks on time, the subsequent scheduled time step will be disrupted and will not be initiated at the prescheduled time. The latency and any other faults occurring during the simulation update may negatively affect the output of the simulation. Simulation models using EILS signal control and HILS SILS represent two different simulation systems. In the EILS system, the signal timing control and detector call response algorithm are included as part of the simulation software. In contrast, HILS and SILS are real-time systems with additional components, which were described earlier. When a simulation model utilizes EILS to provide signal timings, the phase-updating and detector responses have no latency since they are all handled within the simulation model environment. However, the standard HILS implementation of the algorithms requires the integration of four different components. First, the simulation model must be synchronized to run on a real-time basis. Next, the phase and detector information needs to travel back and forth between the four components of the standard HILS system. The data flow for the HILS and SILS systems is presented in Figure 1. Latency occurs during the generation of the data and the communication of the data between the various HILS and SILS components. Latencies in HILS can be attributed to either software or hardware. Hardware latency usually results from five sources: universal serial bus (USB) communication, CID signal conversion, traffic controller, signal transmission, and signal propagation. Software latency most likely results from four possible factors: simulation model, shared memory, CID interface software, and the USB driver. Latencies in SILS are attributed to software only. The major sources of latencies in HILS are 1. Propagation delay the time it takes a data packet to travel between one point and another (in HILS, via the CID cable to the traffic controller). 2. Transmission delay the delay introduced by the medium itself. The size of the data packet affects transmission delay. The USB maximum packet size for the isochronous in and out endpoints is 1,023 bytes. In the CID system, the maximum packet size is 73 bytes, which will not affect the USB communication. 3. CID signal processing delay the time that it takes each CID to convert data from digital to analog or from analog to digital. 4. Software processing the time that software modules take to complete their functions.

145 Stevanovic, Abdel-Rahim, Zlatkovic, and Amin 145 Microscopic Simulation Model Shared Memory CID Signal Convention Traffic Controller CID Interface Software Communication Media (USB Driver) HILS data flow Microscopic Simulation Model Interface Software Virtual Traffic Controller SILS data flow FIGURE 1 HILS and SILS data flows. RESEARCH METHODOLOGY Case Study Network To test the differences in EILS, HILS, and SILS, a 1.5-mi segment of roadway was used: from 4000 West to 5200 West of the 3500 South (SR-171) arterial in West Valley City, Utah. The segment, shown in Figure 2, represents one of the major east west arterials in the county. Levels of service and basic geometry are given for all intersections in Figure 2. All intersections have ASC/3 Econolite controllers, which were recently installed to support transit priority operations on the arterial. A VISSIM model of the study case segment was built, calibrated, and validated on the basis of data from the field. To calibrate the VISSIM model, timings from the field were used. These included speed limits, p.m.-peak 15-min turning-movement counts, and queue lengths at some intersections. To validate the model, travel times (floating car with Global Positioning System device) along the arterial were measured while passing times at each intersection were recorded. Figure 3 shows the results of the calibration and validation efforts. Experimental Scenarios To determine the consistency between different ways of modeling traffic control, various controllers operational scenarios were tested. Since EILS does not support advanced and vendor-specific features of HILS and SILS, only operations supported by EILS were tested. The Utah Department of Transportation provided signal timing databases from the ASC/3 Econolite controllers in the field. All experiments were performed by using VISSIM Version Both HILS and SILS used the same version of the ASC/3 controller software V HILS utilized the ASC/ hardware version of the controller. For EILS, a standard version of VISSIM s emulation of NEMA controllers was used. A variety of intersections were used to test five operational scenarios. The first scenario tested operations of an isolated intersection. The second and the third scenarios tested smooth and maximum dwell transitions on two and three intersections, respectively. The last two scenarios tested pretimed and actuated coordinated operations on four and five intersections, respectively. When one or more intersections were removed from the original VISSIM model, new traffic inputs were created to re-create the traffic flows observed 4155 W 4000 W 3500 S A D B B D 5200 W 4800 W 4400 W 0.0 mile 0.25 mile 0.5 mile FIGURE 2 Case study segment of 3500 South Street, West Valley City, Utah.

146 146 Transportation Research Record TABLE 1 Average Green Times (s) for Free-Running Intersection Model traffic counts [veh/h] R² = Field traffic counts [veh/h] (a) 4000 West EILS SILS HILS Signal Group Mean SD Mean SD Mean SD 1 EBL WBT SBL NBT WBL EBT NBL SBT NOTE: EBL = eastbound left; WBT = westbound through; SBL = southbound left; NBT = northbound through; WBL = westbound left; EBT = eastbound through; NBL = northbound left; and SBT = southbound through. Model travel times per section [sec] 75 R² = Field travel times per section [sec] (b) FIGURE 3 Validity of VISSIM model of 3500 South Street: (a) calibration and (b) validation. To identify this effect the set of VISSIM network outputs (Table 2) was examined. In addition to these outputs, the changes in cycle length were observed during the transition periods to identify how each system performs the transitions. Finally, VISSIM s signal change protocols and signal change and detector records were analyzed. This analysis was crucial to explaining differences and inconsistencies in the performance of the three control systems. The MOEs reported here represent 60-simulation-minute averages from 10 randomly seeded runs. All simulations were 60 min long with 15 min of warm-up time. Each scenario was simulated in EILS, SILS, and HILS for the same 10 random seeds. HILS took exactly 1 h and 15 min to finish each simulation. In contrast, EILS and SILS took on average 5.5 and 39 min to complete the same simulation, respectively. TABLE 2 MOEs for Free-Running Operations in reality. Each reduced model was checked for consistency and validity. To run a certain timing plan, the date and time in VISSIM were set up so that the SILS clock (which is synchronized with VISSIM) selects a corresponding timing plan. EILS was continuously synchronized with VISSIM. However, HILS had to be run at the corresponding time of day or its plan scheduler would need to be adjusted. In addition, controller and simulation model needed to be started manually at the same time to achieve their clock synchronization in HILS. Although this manual procedure does not precisely mimic synchronization of the SILS process, it was the only feasible way to start the two components at the same time. RESULTS There are many ways of testing the consistency of EILS, HILS, and SILS performance and their operations. Signal timings generated by these systems were examined through VISSIM s performance measures. Phase average green times (Table 1), as recorded by VISSIM, were compared to investigate whether signal timings from the three systems were identical. However, slightly different signal timings can sometimes have the same effect on traffic performance. Performance Measure Statistic EILS SILS HILS Total number of Mean 4, , , vehicles SD Average delay/ Mean a vehicle (s) SD Total delay (h) Mean a SD Average number Mean of stops per SD vehicle Total number of Mean 2, , , stops SD Average stopped Mean a,b delay/vehicle (s) SD Total stopped Mean a,b delay (h) SD Average speed Mean a (mph) SD Total travel Mean a time (h) SD a Value is significantly different from corresponding SILS value. b Value is significantly different from corresponding HILS value.

147 Stevanovic, Abdel-Rahim, Zlatkovic, and Amin 147 Free-Running Operations: Single Intersection TABLE 3 MOEs for Smooth Transition Operations To perform the basic test of the controller s timing-plan-free operations, the intersection of 4000W and 3500S was selected, which for this purpose was cut from the rest of the network. The controllers in EILS, SILS, and HILS used original free-running settings from the field. To support EILS (which operates only integer values) the fractional values for some of the field controller s settings were removed. More specifically, red clearance times for certain phases were modified from 1.5 to 2 s. This type of adjustment was later repeated for all experimental scenarios and other settings (amber, vehicle extension, etc.). The analysis of outputs for the free-running intersection, shown in Table 1, presents some differences in the way that each system reports phase greens. The phase greens for HILS and SILS are more similar than those reported by EILS. Table 2 shows intersection MOEs for the three traffic control systems. Two-tail t-tests for paired samples tested the null hypothesis (α=0.05) that all performance measures from Table 2 are the same for each pair of controls. The test results show that EILS MOEs are statistically different than MOEs from any other system for six of nine MOEs. Differences between HILS and SILS are not statistically significant for any MOE. Smooth Transition: Two Intersections To test the second scenario, the smooth transition strategy, two intersections (4400W and 4800W) were used. The purpose of this scenario was to investigate whether various ways of modeling traffic control make differences in the transition between the two signal timing plans. It was assumed that shortway offset seeking in VISSIM EILS functions the same as the smooth transition in the ASC/3 SILS operations (11). Shortway offset seeking in EILS runs the clock 20% slower or faster during any phase until correct offset is reached, meaning that at most 21/2 cycles are needed to get back in sync (12). In ASC/3, smooth transition is accomplished by adding a maximum of 20% or subtracting a maximum of 17% of cycle length per cycle (13). The two intersections were running actuated-coordinated operations with a 96-s cycle length for the first 30 min and a 120-s cycle length for the rest of the simulation. The results of the smooth transition experiments show great similarity between the SILS and HILS average green times, although the EILS average green times are within the same range. The differences in average green times between any two traffic control models were less than 1 s in most cases. When network performance is observed (Table 3), there is no statistical significance in differences between the three ways of modeling traffic control for all but one MOE. However, transition logic was not identical in the experiments. Figure 4 shows that shortway and smooth transition logics worked differently, especially for the intersection of 4800W and 3500S, in which case the EILS transition was inverted. Figure 4 also shows that all three systems at 4400W synchronized almost simultaneously, whereas the 4800W EILS took longer to synchronize. In spite of the differences in transitioning between cycle lengths, VISSIM s traffic performance measures and average signal timings were unaffected. Maximum Dwell Transition: Three Intersections Another operational strategy that was interesting for comparison (because it was implementable in all three systems) was the maxi- Performance Measure Statistic EILS SILS HILS Total number of Mean 4, a 4, , vehicles SD Average delay/ Mean vehicle (s) SD Total delay (h) Mean SD Average number Mean of stops per SD vehicle Total number of Mean 4, , , stops SD Average stopped Mean delay/vehicle(s) SD Total stopped Mean delay (h) SD Average speed Mean (mph) SD Total travel Mean time (h) SD Total distance Mean 3, , , traveled (mi) SD a Value is significantly different from corresponding HILS value. mum dwell transition. This strategy was tested at three intersections on 3500S: 4400W, 4800W, and 5200W. Maximum dwell transition adjusts the start of a cycle by extending the green time of the coordinated phase for a limited amount of time in each cycle (11). As appropriate, settings for maximum dwell times were adjusted in VISSIM EILS and ASC/3 SILS and HILS to 20 s. As with the previous experiment, the transition from 96-s to 120-s cycle length was executed at the 30th min of the simulation time for all three intersections. The maximum dwell transition experiment yielded some unexpected results. Analysis of green times does not show that differences in their averages are practically significant. Here, the reference to practical significance indicates that the difference is large enough to be observed as a significant difference even without statistical tests (e.g., differences in delay of HILS and others in Table 4). However, Table 4 reveals differences that are significant statistically and practically. Two-tailed t-tests show that most of the differences in MOEs between SILS and EILS are not statistically significant. All other differences show the opposite trend. Figure 5 provides additional information about these differences; it shows that of all three systems, HILS was consistently the first to get back in sync during the transition for any of the three intersections. For this reason, when in transition, HILS does not run the 120-s cycle as long as the other two systems do, which evidently reduces overall delay and improves traffic performance in the three-intersection network. Pretimed Operations: Four Intersections A segment with four intersections, from 4000W to 4800W, was selected to test the consistency of pretimed controller operations. For all three systems, the actual signal timings from the field were

148 148 Transportation Research Record Cycle Length (s) Simulation Time (s) EILS SILS HILS (a) Cycle Length (s) Simulation Time (s) EILS SILS HILS (b) FIGURE 4 Cycle lengths during smooth transition: (a) 4400 West and (b) 4800 West. modified to operate with a pretimed control. The motivation for this experiment was investigation of differences in the systems performances in the least-responsive traffic control environment. When pretimed control is selected in ASC/3 controllers, the ASC/3 automatically calls all pedestrian signal groups, with corresponding vehicular signal groups, and implements a rest-in-walk option for the pedestrian phases. Pedestrian signal timings in VISSIM s EILS needed to be adjusted to mimic the pedestrian operations from the ASC/3 controllers. The analysis of the pretimed operations at the four intersections shows no difference in average green times between EILS and SILS. There is a very small difference between any of these two systems and HILS. The difference in HILS average green times can be attributed to the latency of the HILS system. Table 5 shows a comparison of the network MOEs for the pretimed experiments. With the exception of two cases, when EILS is different from SILS, there was no statistically significant difference in MOEs between any pair of traffic control systems.

149 Stevanovic, Abdel-Rahim, Zlatkovic, and Amin 149 TABLE 4 MOEs for Maximum Dwell Transition Operations Performance Measure Statistic EILS SILS HILS Total number of Mean 4, a,b 4, b 4, vehicles SD Average delay/ Mean b b vehicle (s) SD Total delay (h) Mean b b SD Average number Mean b b of stops per SD vehicle Total number of Mean 4, b 4, b 4, stops SD Average stopped Mean b b delay/vehicle (s) SD Total stopped Mean b b delay (h) SD Average speed Mean b b (mph) SD Total travel Mean b b time (h) SD Total distance Mean 4, b 4, b 4, traveled (mi) SD a Value is significantly different from corresponding SILS value. b Value is significantly different from corresponding HILS value. Actuated Coordinated Operations: Five Intersections Finally, motivation for the fifth experiment was investigation of the basic actuated coordinated operations. All five intersections from the study network (4000W to 5200W) were used to evaluate the operations of the three control systems when actuated-coordinated operations were running. The controller settings were the same as those utilized in the field except for the aforementioned adjustments in fractional signal parameters. Average phase green times for each intersection again do not reveal enough information; each system looks very similar to the other two. Average MOEs in Table 6 show that, similar to the free-running operations, EILS is almost always significantly different from HILS and SILS. However, this time EILS yields better MOEs than the other two systems. DISCUSSION OF RESULTS The analyses of second-by-second signal and detector changes were performed for each of the five experiments to determine the potential causes for the observed differences in operations of the three traffic control systems. Three major reasons were identified for the discrepancies in the operations and generated VISSIM MOEs: 1. Noticeable differences were observed in the way that EILS, SILS, and HILS react on detector actuations. EILS, which works on a 1-Hz controller frequency, has a lower sensitivity of detector actuations than SILS and HILS, which work on 10-Hz controller frequencies. In free-running operations the consequence of this small delay in reaction will usually cause slightly worse EILS performance than that from SILS and HILS. 2. The three systems experienced different start-up processes. In EILS, the start of the controllers is well synchronized with the beginning of the simulation. Actual signal timings for each intersection start according to a provided input (e.g., starting phases and offsets). No significant initial delay between a simulator and a traffic control generator was observed, and there was no need for subsequent adjustments. In SILS, the controller starts simultaneously with the simulation but also requires an initialization process (e.g., placing calls on different phases), which can cause a small delay. Usually the SILS controller needed to adjust signal timings within the first few cycles to synchronize with signal timings from a time-of-day plan. In HILS, this initialization delay was even longer because the HILS controller needed to be powered up at the start of the simulation and then go through the initialization process similar to SILS. These small delays in controller initializations will often cause different signal timing sequences for the three systems. The differences in the signal timing sequences will often have some effect on overall system performance. For example, various signal timing sequences can cause transition events (e.g., maximum dwell) for the three systems to occur at different times throughout the coordinated phases. In such a case, depending on when the call for the transition event occurs with respect to cycle time, EILS may find that it is better to extend the cycle length, whereas SILS and HILS may decide the opposite. Such different behavior is observed in experiments with both transition strategies. Corresponding issues related to the initialization processes (and subsequent variations in the signal timing sequences) were the source of discrepancies in all experiments performed in this study. 3. There is latency in the SILS and HILS systems. For example, built-in detector actuations in SILS and HILS are 0.1 and 0.2 s, respectively. The latencies for other events within the SILS and HILS systems may be different and higher than those reported for detector actuations. The sources of latency in real-time systems are described earlier in the text. CONCLUSIONS This study investigated three methods for connecting traffic controllers with a traffic microsimulator. A five-intersection VISSIM model was used to test the following traffic control operations: free-running, shortway (smooth) transition, maximum dwell transition, pretimed, and actuated-coordinated. On the basis of average traffic metrics and signal setting outputs, the following conclusions were drawn: 1. The HILS and SILS approaches generate more realistic signal timings than does EILS. The EILS inability to work at a 10-Hz frequency affects vehicle actuations and can introduce delay in free-running operations. 2. EILS, HILS, and SILS initialization processes can have a significant impact on the sequence of signal timings implemented by each concept. Initialization is shortest for EILS and the longest for HILS, with SILS in the middle. The difference in sequence caused by initialization introduces randomness in the processes when a single event (e.g., transition) places a call on signal operations. In such a situation it is quite difficult to draw any conclusions about how various systems handle certain events. It seems that most of the differences observed should not be attributed to various ways of applying the same (NEMA) basic logic but to randomness caused by the initialization process. 3. Overall, SILS and HILS performed in a very similar way, whereas EILS occasionally performed differently than the other two. Measured traffic metrics show that operational differences were rarely significant, although statistical differences may be present because of small variations in VISSIM s outputs.

150 Cycle Length (s) Simulation Time (s) EILS SILS HILS (a) Cycle Length (s) Simulation Time (s) EILS SILS HILS (b) Cycle Length (s) Simulation Time (s) EILS (c) SILS HILS FIGURE 5 Cycle lengths during maximum dwell transition: (a) 4400 West, (b) 4800 West, and (c) 5200 West.

151 Stevanovic, Abdel-Rahim, Zlatkovic, and Amin 151 TABLE 5 MOEs for Pretimed Intersection Operations TABLE 6 MOEs for Actuated Coordinated Intersection Operations Performance Measure Statistic EILS SILS HILS Total number of Mean 6, , , vehicles SD Average delay/ Mean vehicle (s) SD Total delay (h) Mean SD Average number Mean a of stops per SD vehicle Total number of Mean 7, a 7, , stops SD Average stopped Mean delay/vehicle (s) SD Total stopped Mean delay (h) SD Average speed Mean (mph) SD Total travel Mean time (h) SD Total distance Mean 7, , , traveled (mi) SD a Value is significantly different from corresponding SILS value. Performance Measure Statistic EILS SILS HILS Total number of Mean 6, , , vehicles SD Average delay/ Mean a,b vehicle (s) SD Total delay (h) Mean a,b SD Average number Mean a,b of stops per SD vehicle Total number of Mean 7, a,b 7, , stops SD Average stopped Mean a,b delay/vehicle (s) SD Total stopped Mean a,b delay (h) SD Average speed Mean a, b (mph) SD Total travel Mean a,b time (h) SD Total distance Mean 8, a 8, , traveled (mi) SD a Value is significantly different from corresponding SILS value. b Value is significantly different from corresponding HILS value. REFERENCES 1. Urbanik, T. II, and S. P. Venglar. Advanced Technology Application: The SMART Diamond. In Compendium of Technical Papers, Institute of Transportation Engineers, 65th Annual Meeting, Denver, Colo., 1995, pp Urbanik, T. II, M. Kyte, and D. Bullock. Software-in-the-Loop Simulation of Traffic Signal Systems. SimSub Mid-year Newsletter, 2006, pp Koonce, P., T. Urbanik II, and D. Bullock. Evaluation of Diamond Interchange Signal Controller Settings Using Hardware-in-the-Loop Simulation. In Transportation Research Record: Journal of the Transportation Research Board, No. 1683, TRB, National Research Council, Washington, D.C., 1999, pp Abbas, M., D. Bullock, and L. Head. Real-Time Offset Transitioning Algorithm for Coordinating Traffic Signals. In Transportation Research Record: Journal of the Transportation Research Board, No. 1748, TRB, National Research Council, Washington, D.C., 2001, pp Kyte, M., A. Abdel-Rahim, and M. Lines. Traffic Signal Operations Education Through Hands-On Experiences: Lessons Learned from a Workshop Prototype. In Transportation Research Record: Journal of the Transportation Research Board, No. 1848, Transportation Research Board of the National Academies, Washington, D.C., 2003, pp Bullock, D., B. Johnson, R. Wells, M. Kyte, and Z. Li. Hardware in the Loop Simulation. In Transportation Research, Vol. 12C, No. 1, Feb. 2004, pp Yun, I., M. Best, and B. Park. Evaluation of Adaptive Maximum Feature in Actuated Traffic Controller: Hardware-in-the-Loop Simulation. In Transportation Research Record: Journal of the Transportation Research Board, No. 2035, Transportation Research Board of the National Academies, Washington, D.C., 2007, pp Abdel-Rahim, A., Z. Li, and M. Kyte. Hardware-in-the-Loop Simulation: What s the Difference? Presented at 83rd Annual Meeting of the Transportation Research Board, Washington, D.C., Kyte, M., A. Abdel-Rahim, and M. Dixon. Hardware in the Loop in Education. Presented at Summer Workshop on Advanced Evaluation Techniques and Hardware in the Loop Traffic Simulation, Traffic Signal Systems Committee, Transportation Research Board, Portland, Ore., July Smadi, A., and S. Birst. Use of Hardware-in-the-Loop Traffic Simulation in a Virtual Environment. In Applications of Advanced Technology in Transportation: Proc. 9th International Conference, ASCE, Chicago, Ill., Washington, D.C., 2006, pp Traffic Signal Timing Manual. Publication FHWA-HOP FHWA, U.S. Department of Transportation, NEMA Editor Manual, Version 5. Planung Transport Verkehr AG, Karlsruhe, Germany, Advanced System Controllers ASC/3 Programming Manual P/N Rev. 03. Econolite Control Products, Inc., Anaheim, Calif., Feb The Traffic Signal Systems Committee sponsored publication of this paper.

152 Lost Time and Cycle Length for Actuated Traffic Signal Peter G. Furth, Burak Cesme, and Theo H. J. Muller Time within an actuated signal cycle can be decomposed into time that is fully used, which is the saturation headway multiplied by the number of passing vehicles, and time that is wasted or lost. Activity network modeling is used to show the interaction between signal timing events and traffic flow transitions. Seven components of generalized lost time are identified: those associated with start-up, minimum green, parallel queue discharge (for simultaneous gap-out), extension green, parallel extension (for nonsimultaneous gap-out), the passing of the critical gap, and phase end. Simple formulas can be used to estimate all of these components for many practical cases, allowing one to estimate average cycle length without iteration. The modeling framework accounts for the dual-ring structure with minimum green and maximum green constraints and on off settings for recall and simultaneous gap-out. Experiments with microsimulation software verify the formulas developed. The formulas show the sensitivity of lost time, and therefore average cycle length, to parameters that a designer can control including detector setback, critical gap, gap-out settings, and number of lanes. They also show sensitivity to total demand and to the ratio of noncritical to critical phase volumes. At a fully actuated traffic signal, cycle length varies from cycle to cycle as an outcome of traffic demand and various physical and timing parameters. Average cycle length is an important measure of performance of an actuated signal because average delay for motorists and pedestrians is roughly proportional to cycle length. Traffic engineers need more transparent tools for predicting average cycle length in order to better understand the relationship of signal design to performance. Microsimulation software offers one such tool; however, many traffic engineers lack a facility with microsimulation and would benefit from an approach that is intuitive and simple. An activity network and a generalized concept of lost time are used to model the operation of a fully actuated signal with a dual-ring structure. The focus here is on time in a signal cycle that is wasted (lost) as a basis for determining cycle length, and seven components of lost time are identified. Compared with published methods of cycle length estimation, the model discussed here represents some features of actuated operation more accurately, such as joint gap-out logic and minimum green. It is shown that in some common situations, the expected cycle length cab be determined without iterative calculation. P. G. Furth and B. Cesme, Department of Civil and Environmental Engineering, Northeastern University, 360 Huntington Avenue, Room 400 SN, Boston, MA T. H. J. Muller, Transportation and Planning Department, Delft University of Technology, Stevinweg 1, Delft 2628 CN, Netherlands. Corresponding author: P. G. Furth, pfurth@coe.neu.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / EXISTING CYCLE LENGTH PREDICTION METHODS The seminal research on actuated operation by Newell and Osuma (1) showed that although a pair of one-way traffic streams could be easily analyzed and efficiently operated, complications arising when parallel streams were considered made the analysis muddy and removed much of the theoretical efficiency of actuated control. Their analysis assumed that controllers can detect the moment of queue discharge when in fact the performance of actuated controllers is driven by their inability to discern a sharp boundary between saturated and unsaturated flow. Lin (2) considered multiphase control with the gap-seeking logic of modern controllers together with their minimum and maximum green constraints. His work focuses on estimating the length of the extension green period and, for that purpose, modeling headway distributions for single and multilane approaches. He proposes an iterative, deterministic approach to estimating cycle length without considering interactions with noncritical phases. Akcelik (3) built on Lin s work as part of NCHRP Project 3-48; since 1997, his work has been adopted in the Highway Capacity Manual (HCM) (4). It includes further development of headway distributions and consideration of dual-ring control. However, interaction with noncritical phases is almost trivial because like Lin, a largely deterministic analysis is used, which predicts that noncritical phases never affect cycle length. LOST TIME IN CRITICAL CIRCUIT A four-leg intersection is considered with left and through right phases on each approach, with signal control following the standard dual-ring, eight-phase structure, as shown in Figure 1. Phase lengths (splits) vary from cycle to cycle depending on traffic. Barriers in the dual-ring structure divide the signal cycle into half-cycles, each containing two half-rings that must start and end simultaneously at the barriers but otherwise can run independently. In any given pass through a half-cycle, depending on traffic, either half-ring may be dominant. The half-ring with the greater average demand is considered the critical half-ring, and its phases critical phases. Because of the barriers, the cycle length can be treated as the sum of the times (splits) of the four critical phases. For each critical phase, the time is divided into used and lost time. For every vehicle that passes the stopline, a time (1/s i ), the saturation headway for phase i, is accounted as used; the remainder of the split is treated as lost time. Thus, if the saturation headway is 2 s and two cars pass with a headway of 3 s, 2 s is accounted as used and 1 s as lost. With the familiar terms L i, total lost time for phase i; v i, approach volume for phase i; s i, saturation flow rate for phase i; and C, cycle length and use of expected values of C and L i, the mean number of phase i vehicles served per cycle is Cv i, and so, summing over critical 152

153 Furth, Cesme, and Muller (240) SBT SBL NBL 4 EBT SBT (300) 150 EBL EBT WBT 225 (450) WBL WBL EBL 5 WBT 6 7 SBL NBT 8 NBL NBT 180 (360) x (y) = low demand, high demand volume E-W Street N-S Street FIGURE 1 Example intersection with eight-phase dual-ring control (SBT southbound through, SBL southbound left, WBT westbound through, WBL westbound left, NBL northbound left, NBT northbound through, EBT eastbound through, and EBL eastbound left). phases, the mean amount of time used per cycle is Σ Cv i /s i. At the same time, the mean amount of time used per cycle is C ΣL i. Equating these quantities leads to the familiar cycle time formula: C = 1 icritical icritical i L v s i The usefulness of Equation 1 depends on an ability to determine generalized lost time. ACTIVITY NETWORK FOR PHASE NOT ENDING AT BARRIER () 1 An actuated signal cycle can be modeled as an activity network (5, 6), in which each node represents a moment in time and each arc represents an interval. Figure 2 shows an activity network for a single phase that terminates independently of other phases (i.e., does not end at a barrier). The convention that square nodes represent events in the usual sense of the word is used, with the following meanings: SG = start of green (start of split); SQ = start of queue discharge; EQ = end of queue discharge; SX = start of extension green; UX = start of unsaturated part of green extension; LXD = moment at which front of last vehicle that extends green passes downstream edge of extension detector; LXS = moment at which front of last vehicle that extends green passes stopline; GD = moment at which gap long enough to end green is detected (gap-out); SY = start of yellow; SPX = start of parallel extension flow period (which appears only in Figure 3); SLA = start of late arrival period; when cars that arrive after gap-out but before clearance time are served; SCl = start of effective clearance time; E = end of split; and MaxG = maximum green expired (max-out). Round nodes are not events in the usual sense of the word; they are a device used to divide an interval into subintervals of interest. The network has three kinds of arcs: Solid lines represent intervals with externally determined length; Dashed lines represent slack time that ensures that every arc arriving at a common node arrives at the same time; the length of slack arcs must be nonnegative; when two slack arcs arrive at a common node, one of them usually has zero length in any given pass through SG L s G min SQ SX t qd used EQ LminG UX used t ux G min L x gap-out LXD max-out L x L gap t late LXS SLA SCI h* h* = h crit for gap-out h* GD 0 used SY t r +u/2a YAR L end MaxG E FIGURE 2 Activity network for phase not ending at barrier.

154 154 Transportation Research Record 2128 L gap used L px used L end used L x D ds /u LXS i SPX SLA SCI h crit G px t late UX i t i ux LXD i h i crit GD i G i p SY i E GD both h j crit GD j LXD j FIGURE 3 Nonsimultaneous gap-out. the cycle; however, averaging over many cycles, both may have a nonzero expected length; and Dotted lines represent subdivisions of an interval into used time and lost time. Activities and events concerned with signal timing are shown in the lower part of Figure 2, and activities and events concerned with traffic flow are shown in the upper part. In both the network and the equations that follow, phase indices are suppressed except where phases interact. Start-Up Lost Time, Queue Discharge, and Minimum Green In the upper part of Figure 2, arc SG SQ represents start-up lost time with length L s, as defined in the HCM. Arc SQ EQ represents queue discharge (t qd ) at a uniform rate s. Traffic volumes and maximum green settings are assumed to be such that an approach can discharge its queue before reaching maximum green. Expected queue discharge time is given by the following well-known formula: t SQ EQ = where r is the effective red, v is the arrival volume, and s is the saturation flow rate. In the lower part of Figure 2, arc SG SX represents a phase s minimum initial green period G min, after which the phase enters its extension green period. Controllers have a minimum green period for one or all of the following purposes: to allow traffic flow to get beyond start-up irregularities, to prevent gap-out until after the discharge of a queue that does not reach back as far as an upstream detector, and to provide sufficient pedestrian crossing time. UX, the start of the unsaturated part of extension green, has two predecessor events connected to it by slack arcs: with respect to timing, UX follows Event SX, and with respect to traffic flow, it follows Event EQ. When EQ precedes SX, a condition called early queue discharge, the minimum green constraint governs. In that case, the slack arc from EQ to UX has nonzero length. If n is the number of arrivals during the preceding red period, early queue discharge will occur when n < n 0 rv s v ( ) with n 0 = s v Gmin L s early queue discharge ( 2) in which case the unsaturated part of the minimum green has the following length: t Gmin Ls n n = s v EQ UX ( )( ) ( ) ( ) During this slack period, cars are expected to arrive at a rate v, and each arrival will use a period 1/s. The unused part of this slack period is L ming, lost time due to minimum green: v n Lmin G n = ( Gmin Ls) n n ( ) 1 s s < 4 0 Taking expectation over possible values of n, LminG = P( n < n0 ) Gmin Ls for n < n ( ) 1 Equation 5 applies when the phase is set to recall, meaning that it may not be skipped. If the phase is not set to recall, it will be skipped if n = 0, in which case there will be no lost time. For a phase without recall, LminG = P( 0 < n < n0 ) Gmin Ls 1 ( ) ( ) v E n n n s < s ( ) v E n n n s 0 < < s Akcelik (3) and the HCM (4) use a deterministic approach to model the impact of the minimum green. For a given cycle length, they determine E[n], and depending on whether it is less than n o, minimum green is assumed to govern either in every cycle or never. In spite of being iterative, this approach underestimates the impact of minimum green on cycle length for phases in which minimum green governs in some, but not all, cycles. If it is assumed that minimum green always governs, there is an error from ignoring those cycles in which the phase lasts longer than the minimum green, and if the minimum green constraint is ignored, there is an error for those cycles in which it actually extends the phase length. 0 () 5 0 () 0 3 () 6

155 Furth, Cesme, and Muller 155 Extension Green and Extension Lost Time During the extension green, if the maximum green constraint is not binding, the signal is kept green until the Event GD, when a gap longer than a controller parameter mingap is detected by a presence detector. This parameter is invariably set long enough that a phase will not gap out during queue discharge. mingap corresponds to a critical headway: h where Len d = detector length, Len v = vehicle length, and u = approach speed in the absence of queues. Using h crit rather than mingap allows one to treat detection as an instantaneous event occurring when the front of a vehicle passes the downstream edge of a detector. The time during which traffic flows between SQ and SCl can be divided into four periods with differing flow rates: Queue discharge with flow rate s, ending at Node EQ (which is simultaneous with Node UX when minimum green is not binding); An unsaturated flow period during which headways are subcritical that is, shorter than h crit ending with Events LXD and LXS when the last extending vehicle (last vehicle with subcritical headway) passes the detector and the stopline, respectively; Period of no flow while the critical headway passes, ending with Event SLA; and Late arrival period at flow rate v running from SLA to SCl. For a phase that terminates independent of other phases, the same general logic is followed as that of Akcelik and the HCM for determining the duration of the unsaturated extension green. Headways are assumed to be random and independent. The unsaturated flow period consists of a series of subcritical headways. On the basis of the chosen headway distribution, one can determine p = P(H < h crit ) and E(H H < h crit ), where H is headway length. Headways are treated as a sequence of Bernoulli trials, with the expected number of headways before gap-out equal to n crit ux = Lend + Lenv = min Gap + ( 7) u p 1 p ( ) ( ) (If a phase is prone to sometimes max out, that is, end its green because of the maximum green constraint, Equation 8 overestimates n ux.) Multiplying by E(H H < h crit ) gives t ux, the length of the unsaturated extension period, and subtracting the used time (n ux /s) gives the extension lost time, L x : 1 Lx = nux E[ H H < h ] crit s () 8 () 9 In the activity network in Figure 2, a distinction is made between when the last green-extending vehicle is detected (LXD) and when it crosses the stopline (LXS). In intersection flow models, arrivals are defined with reference to the stopline. Therefore, the time covered by the n ux subcritical headways runs from Node UX to Node LXS. With a shorter, snappier critical gap, extension lost time is diminished by two mechanisms: p is shortened, reducing n ux, and the lost time per subcritical headway is reduced as well. Unsaturated Headway Distribution The form of the assumed unsaturated headway distribution matters mostly with single-lane approaches on which very small headways do not occur. Akcelik and the HCM use a two-parameter modification of the exponential distribution. One parameter is h min, a minimum headway; the other is a bunching parameter that determines the proportion of headways that equal the minimum headway. A special case of the two-parameter model has only one externally specified parameter, h min. It assumes an underlying exponential distribution with parameter λ and transforms to h min all headways shorter than h min. With this distribution, and 1 E[ H]= hmin pmin + ( 1 pmin ) h + min ( 11) λ The value of λ is determined by equating E[H] to 1/v. Then p= P H < h 1 exp λh ( 12) 3, p E[ H H h ]= 1 crit h + pv p crit λ In numerical experiments with the simulation software VISSIM, the authors found that on one- and two-lane approaches, the one-parameter model yielded a slightly better fit than the two-parameter model with parameter values recommended in the HCM. Gap Lost Time and End Lost Time If the maximum green constraint is not binding, the final portion of a phase s split begins at LXD, when the last extending vehicle passes the detector. The signal timing track continues through Nodes GD and SY (which follows GD immediately for phases not ending at a barrier) and, after the yellow and all-red interval (YAR), Event E. On the vehicle flow side, LXS follows LXD by the travel time D ds /u, where D ds is the distance from the downstream edge of the extension detector to the stopline and u is speed. If the downstream edge of the extension detector lies beyond the stopline, this travel time is negative. Following LXS there is a period of no flow of length h crit as the critical gap passes, ending at Node SLA; this interval is a component of lost time called gap lost time, given by L gap ( ) = ( ) p = P H = h 1 exp λh ( 10) min min min ( ) = ( ) crit gap-out = h ( 14) crit crit ( 13) Equation 14 shows how the minimum gap setting, already shown to affect extension lost time in two ways, affects gap lost time as well.

156 156 Transportation Research Record 2128 The clearance interval begins at the moment within the YAR period after which vehicles no longer enter the intersection. Assuming that vehicles follow the stop if you safely can convention on seeing the yellow signal, Event SCl follows the onset of yellow by the interval u tr + ( 15) 2 a where t r is the reaction time, taken in these experiments to be 1.0 s, and a is the deceleration rate, taken to be 0.35 g = 11.3 ft/s 2. From the network diagram, the length of the late arrival window is t late The expected number of late arrivals is vt late, and the time they use is (v/s)t late. End lost time represents the clearance interval and the unused part of the late arrival interval: L Equation 17 shows how lost time can be reduced by using an upstream detector, a well-known result reported by, for example, Bonneson and McCoy (7). With a stopline detector, the critical headway crosses the stopline while the light is green, wasting a lot of time; when the detector is moved upstream, much of the critical headway passes after the signal has transitioned to its clearance interval. Maximum Green If the wait for a gap becomes too long, a maximum green constraint will trigger the start of the yellow (the phase maxes out ). This logic is shown in Figure 2 by the node MaxG positioned a time G max (the maximum green setting) from the start of the green, and a slack arc, whose length must be nonnegative, running from SY to MaxG. When the maximum green is not binding, the slack arc SY MaxG is harmlessly nonzero, but if the green period s length reaches G max, the nonnegativity constraint on the slack arc will force the start of the yellow. When the green ends because of max-out, Arcs LXD GD and LXS SLA can have any length less than h crit ; however, those two arcs will have equal length (h* in Figure 2), and therefore the position of SLA becomes fixed at G max D ds /u. When there is a max-out, the period of subcritical headways extends from UX all the way to SLA. Accounting for the fraction of that period that is used, and because it may be assumed that the minimum green will not be binding in any cycle that maxes out, L gap max-out = 0 ( 18) L end x u Dds = max tr +, 0 a u ( 16) 2 Dds v u Dds gap-out = YAR tr u s max +, 0 ( 17) 2a u Dds v max-out = G teq max u 1 s subcrit ( 19) where v subcrit = 1/E[H H < h crit ]. Equation 19 shows how maximum green constraints improve the efficiency of an actuated signal by eliminating gap lost time. Of course, that efficiency gain is negated if G max is too short to allow the queue to discharge, which prevents G max from being set too aggressively. In a deterministic approach, like that followed in the HCM, if the expected number of arrivals during an expected cycle results in max-out, max-out will be assumed to always occur, overestimating phase length because max-out will not occur in every cycle; and if expected values do not predict max-out, the maximum green constraint will be assumed nonbinding, again overestimating phase length because maximum green will limit phase length in some cycles. Instead of the always-or-never approach, the impact of the maximum green on expected gap lost time can be accounted for as follows: Lgap = pmax-outhcrit ( 20) where p max-out is the fraction of cycles in which the phase maxes out. Without detailed modeling, it may be difficult to estimate p max-out with confidence for a given phase; however, one can often make a rough estimate or prior guess of p max-out, in which case Equation 19 models the maximum green effect on a continuous scale rather than as all-or-nothing. JOINT TERMINATION OF PHASES AT BARRIER Controllers offer two options to ensure that the two phases that end each half-cycle (at the barrier) end simultaneously. One is simultaneous gap-out, a default setting in at least some American controllers, which requires that both phases must have a gap greater than or equal to mingap at the same moment to force an end to the extension green. In the other option, nonsimultaneous gap-out, normally used in the Netherlands, whichever phase gaps out first enters a green subphase of length G p called the parallel green until the other phase gaps out (or until max-out), and then both phases end their green. Nonsimultaneous Gap-Out An activity network representing nonsimultaneous gap-out of phases i (critical) and j is shown in Figure 3, where maximum green effects have been omitted for clarity. The start of the yellow waits until both phases have independently gapped out, represented by the Event GD both, which follows Nodes GD i and GD j with slack arcs representing parallel green time. The parallel green time on the critical phase s slack arc is G i px. In the upper part of the diagram, representing traffic flow, LXS i, is followed by an interval of lost time as the critical gap passes and then by a period of flow with unconstrained headways whose length is G i px plus the late-arrival period given by Equation 16. The lost time during the late-arrival period is incorporated into the end lost time formula (Equation 17). Lost time for the parallel green extension, L px, represents the unused time during the parallel green and is given by L px i v i = G i px 1 s ( 21) When gap-out processes are modeled deterministically, parallel extension lost time will always be zero for critical phases, biasing cycle length estimates downward. The error will be negligible when the demands of the noncritical half-rings are small compared with their parallel critical half-rings but not otherwise.

157 Furth, Cesme, and Muller 157 used L i ming EQ i used L pd UX i used L x SX i UX both t ux = f(v 1 + v 2 ) LXD EQ j UX j SX j FIGURE 4 Simultaneous gap-out. Simultaneous Gap-Out An activity network representing simultaneous gap-out of two phases, i (critical) and j, is shown in Figure 4. Assuming that mingap is large enough to prevent gap-out if either phase is discharging its queue, the search for a gap does not effectively begin until both phases have discharged their queues and passed their minimum green. This is Event UX both, which follows UX i and UX j by slack arcs representing the time during which the first phase that has discharged its queue and passed the minimum green waits for the other to do the same. The fraction of this average slack that is not used is called parallel discharge lost time, L pd, which is the average length of the critical phase s slack arc multiplied by (1 v i /s i ). Following Node UX both, the usual gap-out logic applies, but with one big difference: the headway distribution must represent the combined volumes of both phases, v i + v j, and should account for the number of lanes serving both phases combined, making the extension green and its associated lost time longer than it would be with nonsimultaneous gap-out, especially on multilane approaches. (It also makes max-out far more likely.) Using volume shares to determine the fraction of the critical phase vehicle represented by those headways, Equation 8 becomes modified as follows: n ux = p vi 1 p v + v ( ) ( i j) ( 8a) with p based on the combined (two-directional) headway distribution. Equation 9 for extension lost time is still valid with the headway distribution for phase i. RECALL AND PHASE SKIPPING If a phase has no queue when its turn in the cycle comes, it will be skipped unless it is set to recall. It is common for left-turn phases, and sometimes minor through phases, not to be set to recall. The impact of skipping can only be evaluated by using a probabilistic analysis, which is included in the HCM s recommended procedure. The probability of phase i s being skipped assuming a Poisson arrival process is exp( v i r i ), where r i is the expected length of phase i s red period minus the intergreen following the conflicting phase preceding phase i. (The controller decision of whether to skip is made just before that intergreen phase.) The skipping of a critical phase increases the chance of a noncritical phase s becoming dominant, increasing parallel lost time and thus reducing some of the apparent benefit of phase skipping. No attempt has been made to estimate this effect. DIRECT ESTIMATION OF CYCLE LENGTH Under certain simplifying conditions nonsimultaneous gap-out, low noncritical demands, no phase skipping, and minimum and maximum green constraints can be ignored the only lost time components are L s, L x, L gap, and L end, of which the first consists of data and the remainder can be determined by formulas (Equations 9, 14, and 17) without prior knowledge of the cycle length. In such a case, average cycle length can be determined directly from Equation 1. The ability to readily determine lost time components and cycle length makes it easy to explore the impact of demand and design parameters. Figure 5 shows expected cycle length under the simplifying S=0, H=4.6 S=0, H=3.6 S=120, H=4.6 S=120, H=3.6 S=0, H=3.6 S=0, H=2.6 S=120, H=3.6 S=120, H= S=setback(ft) H=h crit (sec) v/s (a) v/s (b) FIGURE 5 Expected cycle length as a function of demand, critical headway H, and detector setback S: (a) one through lane and (b) two through lanes.

158 158 Transportation Research Record End Lost Time Extension Lost Time Gap Lost Time S=0 H=4.6 1 lane S=0 H=3.6 1 lane S=120 H=4.6 1 lane S=120 H=3.6 1 lane S=0 H=3.6 2 lanes S=120 H=2.6 2 lanes FIGURE 6 Selected lost time components per cycle as a function of critical headway H in seconds and detector setback S in feet. conditions mentioned earlier for varying levels of demand (Σv/s), detector setback, critical headway, and number of through lanes per approach (number of left-turn lanes per approach is always 1). Critical phase volumes for the single-through-lane case are shown in Figure 1 for low- and high-demand scenarios; intermediate scenarios are linear interpolations. Through volumes are simply doubled to create the two-through-lane scenarios. Saturation flow rate, measured experimentally from the microsimulation model used to verify results, was 1,950 (veh/h)/lane; start-up lost time, also measured experimentally, was 1.5 s for each phase. Figure 5 shows how sensitive average cycle length is to detector setback and h crit. For a given level of demand, using an upstream detector instead of a stopline detector reduces average cycle length by about 20 s, and making h crit 1 s shorter reduces average cycle length by about 12 s. Reductions of this size make a substantial difference in level of service. Figure 6 shows how demand and design parameters affect extension, gap, and end lost time for the main street through movement. The main design effect is reducing end lost time by a combination of detector setback and critical gap; also substantial is the effect on gap lost time by reducing h crit. The impact on extension lost time is relatively minor. VERIFICATION To verify the model developed, simulation experiments were performed with VISSIM, using its default parameters for vehicle behavior. Each experiment consisted of a 1-h simulation after four cycles of start-up. Figure 7 shows the good agreement in expected cycle length between the model s formula-based results and the simulation experiments. The scenarios reported in Figure 7 include not only those covered in Figure 5 but also the scenarios described later involving minimum green constraints (with recall) and cases with nonsimultaneous gap-out, in which the demand of noncritical phases was as great as 85% of the critical phase demand. IMPACT OF MINIMUM GREEN CONSTRAINTS Two sources of randomness affect the distribution of n, which is needed to evaluate Equation 5 or 6: random arrivals for a red period of given length r and different values of r resulting from varying cycle lengths. In most practical situations, the second effect is small compared with the first and may be neglected Spreadsheet R 2 = RMSE = 2.56 n = Microsimulation FIGURE 7 Expected cycle length for spreadsheet model versus simulation experiments.

159 Furth, Cesme, and Muller with recall without Left-Turn Volume (veh/h) FIGURE 8 Minimum green lost time versus volume with and without recall. A two-pass procedure amenable to spreadsheet calculation was tested against both iterative spreadsheet calculation and full simulation in the case of 8-s minimum green constraints applied to left-turn movements with and without recall. In the first pass, Equation 1 was used to find cycle length and splits assuming that the left turns were governed by minimum green; in the second pass, n s distribution was Poisson over the phase s red period as determined in the first pass. For each possible value of n < n o, its probability and conditional impact were calculated to evaluate Equation 5 or 6. The two-pass process yielded cycle length estimates within 0.5 s of average cycle length as determined by using repetitive iterations, confirming the validity of the two-pass process when only minor movements are affected by minimum green constraints and when there is no phase skipping. Errors compared with microsimulation were similarly small. Figure 8 shows minimum green lost time determined by using simulation results for a single left-turn approach for various levels of demand when demand on other streams is such that the cycle length is about 60 s. (Results when cycle length is about 75 s or 90 s were similar.) When there is no phase skipping, minimum green lost time diminishes with volume, as expected, from a maximum of (G min L s ). When there is phase skipping, minimum green lost time varies little with demand, holding close to 2 s for a wide range of left-turn volumes. IMPACT OF NONCRITICAL PHASES AND SIMULTANEOUS GAP-OUT Until now, all of the results were obtained by assuming that noncritical phases do not affect the cycle length. For the nonsimultaneous gap-out setting, VISSIM and its application programming language VAP were used to determine average parallel extension lost time for a phase terminating at a barrier as a function of the ratio of noncritical to critical phase demand under different levels of overall demand. Results are shown in Figure 9a. For ratios near 1, lost time due to waiting for a parallel phase to gap out was as great as 6 s in the high-demand scenario. However, for ratios between 0.15 and 0.85, L px grew gradually from near 0 to about 2 s and was stable for a wide range of critical phase demands; an equation for predicting L px is shown in Figure 9. As an example, these results indicate that if noncritical demand rises from 40% to 80% of the critical phase demand, the added lost time will be only about 1 s per half-cycle. The formula for L px was implemented into the spreadsheet model and very good agreement was found between predicted and simulated cycle length; comparisons are included in Figure 7. Figure 9b compares simultaneous and nonsimultaneous gap-out, showing the sum of extension lost time (which for simultaneous gap-out is a function of combined demand in the two directions) and L p = 2.78*(R-0.1), R > R = ratio of non-critical flow to critical flow (a) Given v/s = R = ratio of non-critical flow to critical flow (b) Simultaneous Gap-out Nonsimultaneous Gap-out FIGURE 9 Lost time due to termination of green at barrier: (a) parallel extension lost time with nonsimultaneous gap-out and (b) sum of parallel extension or parallel discharge lost time and extension lost time.

160 160 Transportation Research Record 2128 either parallel extension or parallel discharge lost time for a case with moderate demand. Results are from VISSIM simulations, with VAP programming to mark key events. As the graph shows, simultaneous gap-out creates more lost time, especially when the noncritical phase volume is high. A simple formula for estimating L pd. has not been proposed here. CONCLUSION Modeling the interactions between signal timing and traffic flow by using an activity network allows one to better understand how actuated traffic signals perform and in particular to understand how demand and design parameters affect average cycle length. The network framework is especially useful for understanding interactions between parallel signal phases in a ring-and-barrier system. It was also shown how deterministic estimation of cycle length, even if iterative, has a downward bias with respect to minimum green and parallel green effects and an upward bias with respect to the maximum green effect. Seven lost time components were identified, and simple methods for estimating them were proposed for six of them lost time associated with start-up, minimum green, parallel queue discharge (for simultaneous gap-out), extension green, parallel extension (for nonsimultaneous gap-out), the passing of the critical gap, and phase end. Simulation experiments verify the simple formulas used. With lost time components and cycle length estimates less of a black box, it is the authors hope that this approach will allow engi- neers to focus more clearly on designs that improve actuated signal operation. REFERENCES 1. Newell, G. F., and E. E. Osuma. Properties of Vehicle-Actuated Signals: II Two-Way Streets. Transportation Science, Vol. 3, No. 2, 1969, pp Lin, F.-B. Estimation of Average Phase Durations for Full-Actuated Signals. In Transportation Research Record 881, TRB, National Research Council, Washington, D.C., 1982, pp Akcelik, R. Estimation of Green Times and Cycle Length for Vehicle- Actuated Signals. In Transportation Research Record: Journal of the Transportation Research Board, No. 1457, Transportation Research Board of the National Academies, Washington, D.C., 1994, pp Highway Capacity Manual. TRB, National Research Council, Washington, D.C., Head, L., D. Gettman, D. M. Bullock, and T. Urbanik II. Modeling Traffic Signal Operations with Precedence Graphs. In Transportation Research Record: Journal of the Transportation Research Board, No. 2035, Transportation Research Board of the National Academies, Washington, D.C., 2007, pp Head, L., D. Gettman, and Z. Wei. Decision Model for Priority Control of Traffic Signals. In Transportation Research Record: Journal of the Transportation Research Board, No. 1978, Transportation Research Board of the National Academies, Washington, D.C., 2006, pp Bonneson, J. A., and P. T. McCoy. Methodology for Evaluating Traffic Detector Designs. In Transportation Research Record 1421, TRB, National Research Council, Washington, D.C., 1993, pp The Traffic Signal Systems Committee sponsored publication of this paper.

161 Proposed Concept for Specifying Vehicle Detection Performance Dan Middleton, Ryan Longmire, Darcy M. Bullock, and James R. Sturdevant Specifications for procurement of vehicle detection systems have historically used performance comparisons with known accurate detectors, perhaps specifying one value of detection accuracy to represent several weather, lighting, and traffic conditions. In most cases, comparison with inductive loops has provided the necessary information by which to judge the performance of these detectors as either acceptable or unacceptable. However, differences in detection technologies are not adequately addressed with this method. A new concept for defining detection performance measures is proposed: it provides for some stochastic variation in sensor performance, within prescribed limits. In this case, video image vehicle detection systems (VIVDS) provide an example technology for applying this concept, but the concept is also appropriate for other technologies. Stochastic thresholds are defined that are consistent with field observation of three different VIVDS. Although these thresholds are quite large, it is believed that the combination of this new definition and a framework for defining stochastic performance will provide the basis for the detector industry to enhance its products. Vehicle detection must satisfy two objectives for actuated signal control: To extend green service to a phase until there is no longer demand or flow rates have been reduced to predetermined levels for phase termination and To call service to a phase when, and only when, there is demand. A third objective, when dilemma-zone protection is desired, is occasionally added: To detect the presence (and perhaps speed) at a precise location. The focus here is on the first two objectives. Detection for actuated traffic signals has traditionally been provided by inductive loops, but now many agencies are replacing failing loops at signalized intersections with nonintrusive detectors. Reasons for using nonloop options include the nonintrusive nature of the newer options, reduced delay to motorists during installation and maintenance, no damage to the pavement structure, and, in some cases, D. Middleton and R. Longmire, Texas Transportation Institute, Texas A&M University, 3135 TAMU, 2929 Research Parkway, College Station, TX D. M. Bullock, School of Civil Engineering, Purdue University, 550 Stadium Mall Drive, West Lafayette, IN J. R. Sturdevant, Division of Traffic Control Systems, Indiana Department of Transportation, 100 North Senate Avenue, Room N925, Indianapolis, IN Corresponding author: D. Middleton, d-middleton@tamu.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / reduced costs. In fact, even though the accuracy of most nonintrusive options is not on a par with that of inductive loops, many agencies are still choosing them because of offsetting advantages. The motivation for drafting this new concept for specifying vehicle detection performance is to define an improved framework for public agencies to use for procurement and testing and perhaps move the industry toward improved performance. BACKGROUND Specifications for nonloop detectors have often required comparing the performance attributes of these detectors with those of loops or other point detectors or with manual counts. However, that comparison is not always appropriate for a variety of reasons, and it does not necessarily provide all the critical information needed to make acceptance or rejection decisions. In the case of video image vehicle detection systems (VIVDS), comparison with loops has been difficult since the two systems have different perspectives on approaching vehicles. In all cases except those in which cameras are oriented vertically downward, cameras and loops (or other pavement-based detectors) detect vehicles at different points. Also, for VIVDS, factors such as camera height and horizontal distance to detection zones (defined as the aspect ratio) vary significantly, and these variables significantly affect the accuracy of camera processor systems. The purpose of this paper is to propose a new concept for VIVDS specification. Detection errors by any detection technology can be associated with either efficiency or safety, or both. Recent research activities have attempted to define and categorize the types of errors encountered by VIVDS and in some cases to compare the systems with inductive loops. MacCarley and Palen (1) developed a methodology using methods and metrics for evaluating detectors at actuated signalized intersections. They developed common definitions to describe the types of detector errors possible at these intersections. One part of the methodology penalizes the detector if it makes a mistake, whereas another part penalizes the detector if the controller makes incorrect decisions based on detector mistakes. Examples include failing to call or extend a phase or terminating a phase early. Rhodes et al. (2) defined incorrect detections as false positives (detection when there is no vehicle present) or missed detections. Under this methodology, each detection event could be classified into one of four different states. The first two states occur when the two detectors agree as when neither of them places a call or when both place a call. Rhodes et al. referred to these states as either L0V0 or L1V1, where L represents the loop and V refers to the video system. The numbers indicate whether the detector is off [0] or on [1]. The other two states occur when the two detection systems do not agree, designated as either L1V0 or L0V1. Abbas and Bonneson (3) described video performance in terms of discrepant call frequency. A discrepant call is an unneeded call 161

162 162 Transportation Research Record 2128 or a missed call, determined by comparing manual counts from recorded video. Rhodes et al. (4, 5) investigated detection differences by VIVDS between day and night periods and introduced a new metric for the evaluation of detectors at signalized intersections. They discuss the differences, based on field data collected during good weather, between day and night detection in the area of the stop bar. The researchers installed VIVDS cameras at four locations on each approach to the selected intersection and found that three of them resulted in premature detections at night compared with daytime due to headlight detections. The four camera locations were as follows: 1: 40 ft high on signal mast arm, far side (vendor recommended); 2: 40 ft high on a side-mounted pole, far side; 3: 25 ft high on the signal mast arm, far side; and 4: about 30 ft high near the stopline, near side. The data analysis used detector on and off times, or activation and deactivation times. Testing of sample means with Student s t-test indicated significant differences (at α=0.05) in activation times from daytime to nighttime for all but one of the 16 cameras. The differences for deactivation times from daytime to nighttime were less pronounced compared with the activation times, perhaps because the intersection had street lighting and deactivation times were probably based on detecting the rear of vehicles (same as in daytime). These findings clearly indicate the phenomenon of early detection at night due to headlight detection, even in good weather. Rhodes et al. conclude that consistent detector performance under different lighting conditions would require adjusting gap times by time of day and day of year. Also, improved consistency in activation times at the stop bar could be achieved by positioning cameras on the near side (Camera 4 position), although this assessment should be verified with additional research. With respect to dilemma-zone detection (not part of this research), this camera position would not allow monitoring of set-back detectors with the same camera. Recently, the Indiana Department of Transportation proposed the use of detection zones considering the stochastic variation that is inherent in video detection (6). In subsequent sections that concept is expanded on, a field evaluation of the concept is described, and a set of tables is provided that define thresholds that the current generation of video detection devices can achieve. STOCHASTIC VARIATION OF DETECTION ZONES Concept Definition Figure 1 shows the way detections might be conceptualized to account for stochastic variation of the activation and termination detection zones. There is an activation point (either temporally or spatially) at which video initially detects the vehicle and registers a call in the controller, and a termination point at which it no longer detects the vehicle and releases the call. The difference in time between the two points is the duration. On a vehicle-specific basis, this duration can also represent vehicle length when factored by the vehicle speed. During video setup, the installer tries to set activation points in the Detection Zone Lane Lines A u A A d T u T T d (a) On/Off Traces Detection Zone A u A A d T u T T d Durations (b) FIGURE 1 Stochastic variation in vehicle detection zone activation and termination points: (a) car approaching detection zone and (b) example spatial variation in activation and deactivation of detection zone (A activation, T termination).

163 Middleton, Longmire, Bullock, and Sturdevant 163 video system to match points on the approach at which vehicles need to be detected to satisfy detection needs either at the stop line or for dilemma-zone protection. As Figure 1 illustrates, a tolerance is necessary to account for the difference between the desired activation point and the actual activation point. There is some quantifiable distance upstream and downstream of the desired Point A where activations actually occur. These variations are due to a variety of factors such as camera quality, sun angle, shadows cast by the detected vehicle, and even color of detected vehicles. Terminations coincide with the end of detected vehicles. Since vehicle lengths and heights vary, terminations are more scattered than activations. One way to consider termination characteristics of video systems is to force terminations to occur around a known point. Effectively, the process used in this research forces the range of terminations to be centered on the actual ends of vehicles. To accurately determine and record the beginning and end of vehicles, the process requires an accurate baseline system. The most commonly used baseline system over time has been inductive loops, but loops were not available at this data collection site and are becoming more difficult to find because of replacement by nonintrusive technologies. In this case, researchers used two side-fire radar detectors to provide the true activation and termination points in time and space. In this case, positioning of the two radar detectors coincided with the specified positions of detectors for dilemma-zone detection. Communication between the radar detectors and the cabinet involved wireless devices and recording time stamps at the front and rear of every vehicle detected. The only issue with this setup was a modest amount of latency in signal transmission between radar detectors and the equipment cabinet where data storage occurred. However, researchers were able to accurately account for this latency in the data-processing step because of its consistency. Figure 2 shows the intersection layout and positions of data collection components. As noted earlier, terminations form a distribution of points that become more manageable if they are forced to occur around the actual termination points (as seen by the baseline system). Agencies might choose to ignore terminations altogether, but terminations could be an input in identifying gaps between vehicles, an important parameter in signal control. The length of a vehicle as seen by video is its effective vehicle length and includes the sum of the actual vehicle length and the distance behind the vehicle shadowed by that vehicle. Subtracting actual vehicle lengths (measured by the baseline system) from effective lengths (measured by video) on a per-vehicle basis and taking their mean value yields a parameter termed average excess detector on-time duration. Although it is beyond the scope of this study, thresholds could be developed for this parameter to use in a standard specification to supplement other values. One must realize that this value is a function of the vehicle mix within the sample, but the comparison still appears to have merit. In a later section, detector on-time is converted to length, and another variable is introduced termed average excess detector length. If a data sample has a higher percentage of large trucks, the length and height of the trucks force the duration (and average length difference) to be higher than if all vehicles were passenger cars. For purposes of a video specification, some agencies might choose not to use night data because of the additional challenges involved. Night activation points with video are affected by the level of street lighting and by the leading boundary of the headlight bloom. In most cases, night detection occurs well ahead of the actual vehicle because of this headlight bloom. Adjacent-lane detections are also more prominent at night compared with daytime. Night terminations using video are more challenging to track as well. Observed Performance Table 1 summarizes the data collected for this analysis. Manufacturer representatives were allowed full access during setup of each system in order to optimize performance. Daytime data for this analysis covered a time period of approximately 1 h, and the nighttime data Stone Oak Dr. 175' 378' VIVDS Research Cameras 516' 74' 34' 145' 365' R.M ' Wavetronix HD Camera Cabinet 20-ft detection zone Pole Span wire Mast arm FIGURE 2 Layout of data collection intersection (R.M. Ranch-to-Market; HD high density).

164 164 Transportation Research Record 2128 TABLE 1 Summary of Data Used to Tabulate Activation and Deactivation Histograms Missed False Detection Detection Linked Detection Processor Lighting Aspect Ratio Sample Size Events % Events % Events % V1 Night 4: V2 Night 4: V3 Night 4: V1 Night 10: V2 Night 10: V3 Night 10: V1 Day 4: V2 Day 4: V3 Day 4: V1 Day 10: V2 Day 10: V3 Day 10: covered approximately 4 h (from midnight to 4:00 a.m.). Table 1 indicates missed vehicle detections, false detections, and linked detections. Linked vehicles are simply those occluded by other vehicles, in many cases with a taller vehicle in the lead that occludes trailing vehicles. VIVDS sees one vehicle instead of multiples and counts the group as one vehicle. As the tabulated values indicate, the three VIVDS exhibited similar accuracy across all three categories in daylight and across two of the categories at night. False detections at night were more problematic than during daytime but were generally better at an aspect ratio of 4:1 than at 10:1. There were fewer missed vehicles at night than during the daytime in this data set. The overcounts represented by high false detections were largely due to headlight detections in adjacent lanes. The camera mounting location (beside the roadway with offset of 10 ft to the nearest lane) might have been a contributing factor to the higher number of false detections. In the data analysis, discrepancies as high as 10 to 15 s were observed in the missed-detection category, but most matches were significantly less, especially during the daytime. Therefore, during data reduction, if the time stamp of a vehicle detected by VIVDS was within 10 to 15 s of the baseline time stamp, it was considered a match. Manual verification indicated that this range was reasonable. The false detection discrepancy was typically 2 to 3 s, but some values as high as 10 to 15 s were observed to be accurately matched. Linked vehicles at 10:1 were usually not linked at 4:1, but their numbers were relatively low in either case. Daytime Temporal Histograms Explained with Photographs This research recorded the time of activation and time of termination from VIVDS processors as vehicles passed through the detection zone. These two points in time are then individually compared with the true activation and termination times provided by the baseline system. This procedure is followed for each vehicle in the sample and is summarized in a set of histograms that display the frequency of early and late activations and terminations by the VIVDS processors. Although the activation and termination histograms appear to both be centered on zero, they are actually separated by the average duration of detections as seen by the VIVDS processor. Figure 3a and b shows a truck traveling through the 10:1 detection zone (ideally, a point). Figure 4a indicates the detection of this vehi- cle, denoted by arrows. The vehicle activated the detection zone 905 ms earlier than average for Processor V1 and terminated the detection 258 ms earlier than average for Processor V1. The duration of the detection was 1.4 s longer than that of the baseline system. Figure 3c and d shows a sedan traveling through the 10:1 detection zone. Figure 4c also indicates the detection of this specific vehicle, denoted by the arrows. This vehicle activated the detection zone 495 ms later than average for Processor V2 and terminated the detection 607 ms earlier than average for Processor V1. The duration of the detection was 0.5 s longer than that of the baseline system. Figure 3e and f shows a pickup truck traveling through the 10:1 detection zone. Figure 5e indicates this specific vehicle, denoted by the arrows. It activated the detection zone 316 ms earlier than average for Processor V2 and terminated the detection 360 ms earlier than average for Processor V1. The duration of the detection was 1.1 s longer than that of the baseline system. Nighttime Temporal Histograms Explained with Photographs Figure 5a f contains histograms similar to those seen in Figure 4a through f but displaying data collected during nighttime hours only. The activations are dispersed over a greater time period and occur 2 to 3 s earlier than in daytime (on average) for all VIVDS processors at the 10:1 aspect ratio. The early and long detections at night are due to the effect of vehicle headlights on the detection zone. The impact from headlights and headlight reflections is lessened by improving the ratio of detection distance to camera height, as seen in the histograms for the 4:1 ratio (Figure 5b, d, and f ). Terminations at nighttime appear to group much better than activations at nighttime. This finding is caused by termination of the call by the VIVDS when it stops detecting the headlights. For most vehicles this results in detection termination relatively close to when the actual termination should occur. Logic Behind Spatial Plots The temporal histograms (Figures 3 and 5) were converted to spatial histograms (Figures 6 and 7) by using individual vehicle speeds provided by the baseline side-fire radar detection system. The acti-

165 Middleton, Longmire, Bullock, and Sturdevant :1 Detection Point (a) (b) (c) (d) (e) (f) FIGURE 3 Vehicles corresponding to activation and deactivation times in Figure 4: (a) V1 activation, (b) V1 termination, (c) V2 activation, (d ) V2 termination, (e) V3 activation, and (f ) V3 termination. vation and termination time differences multiplied by the speed of the same vehicle yield activation and termination distances relative to the actual detection point. This calculation assumes that vehicles are traveling at a constant rate of speed across the detection area, a reasonable assumption over short distances. The spatial histograms in Figures 6 and 7 provide an alternative visualization of where vehicles really are located when VIVDS detection activates and terminates. Ideally, 100% of the detections would be contained in a short region centered over zero. However, the majority of detections occur within an 80-ft window during the daytime. At night, the effect of headlights appears again in the dispersion of data points up to 400 ft early at the 10:1 ratio. Tables 2 and 3 derive their values from the histograms (Figures 4 through 7). A VIVDS specification should base its pass-fail criterion on a selected reasonable percentile; this analysis used the 85th percentile. The tabulated values summarize the data set results and provide a clearer picture of VIVDS performance than simply a comparison with point detectors such as inductive loops. The data collection

166 166 Transportation Research Record % 50% 40% 40% Frequency 30% 20% Frequency 30% 20% 10% 0% Activation Termination % 0% Activation Termination Time (seconds) Time (seconds) (a) (b) 50% 50% 40% 40% Frequency 30% 20% Frequency 30% 20% 10% 0% Activation Termination % 0% Activation Termination Time (seconds) Time (seconds) (c) (d) 50% 50% 40% 40% Frequency 30% 20% Frequency 30% 20% 10% 0% Activation Termination % 0% Activation Termination Time (seconds) Time (seconds) (e) (f) FIGURE 4 Daytime temporal variation of activation and deactivation of detection zones: (a) V1 at 10:1, (b) V1 at 4:1, (c) V2 at 10:1, (d ) V2 at 4:1, (e) V3 at 10:1, and (f ) V3 at 4:1.

167 Middleton, Longmire, Bullock, and Sturdevant % 50% 40% 40% Frequency 30% 20% Frequency 30% 20% 10% 0% Activation Termination % 0% Activation Termination Time (seconds) Time (seconds) (a) (b) 50% 50% 40% 40% Frequency 30% 20% Frequency 30% 20% 10% 0% Activation Termination % 0% Activation Termination Time (seconds) Time (seconds) (c) (d) 50% 50% 40% 40% Frequency 30% 20% Frequency 30% 20% 10% 0% Activation Termination % 0% Activation Termination Time (seconds) Time (seconds) (e) (f) FIGURE 5 Nighttime temporal variation of activation and deactivation of detection zones: (a) V1 at 10:1, (b) V1 at 4:1, (c) V2 at 10:1, (d ) V2 at 4:1, (e) V3 at 10:1, and (f ) V3 at 4:1.

168 168 Transportation Research Record % 50% 40% 40% Frequency 30% 20% Frequency 30% 20% 10% 0% Activation Termination % 0% Activation Termination Distance (feet) (a) Distance (feet) (b) 50% 50% 40% 40% Frequency 30% 20% Frequency 30% 20% 10% 0% Activation Termination % 0% Activation Termination Distance (feet) (c) Distance (feet) (d) 50% 50% 40% 40% Frequency 30% 20% Frequency 30% 20% 10% 0% Activation Termination % 0% Activation Termination Distance (feet) (e) Distance (feet) (f) FIGURE 6 Daytime spatial variation of activation and deactivation of detection zones: (a) V1 at 10:1, (b) V1 at 4:1, (c) V2 at 10:1, (d ) V2 at 4:1, (e) V3 at 10:1, and (f ) V3 at 4:1.

169 Middleton, Longmire, Bullock, and Sturdevant % 50% 40% 40% Frequency 30% 20% Frequency 30% 20% 10% 0% Activation Termination % 0% Activation Termination Distance (feet) Distance (feet) (a) (b) 50% 50% 40% 40% Frequency 30% 20% Frequency 30% 20% 10% 0% Activation Termination % 0% Activation Termination Distance (feet) Distance (feet) (c) (d) 50% 50% 40% 40% Frequency 30% 20% Frequency 30% 20% 10% 0% Activation Termination % 0% Activation Termination Distance (feet) Distance (feet) (e) (f) FIGURE 7 Nighttime spatial variation of activation and deactivation of detection zones: (a) V1 at 10:1, (b) V1 at 4:1, (c) V2 at 10:1, (d ) V2 at 4:1, (e) V3 at 10:1, and (f ) V3 at 4:1.

170 170 Transportation Research Record 2128 TABLE 2 Temporal Variation in Activation and Termination Zones Lower Bound on Upper Bound on Lower Bound on Upper Bound on Average Excess Activation Time, Activation Time, Termination Time, Termination Time, Detector On-Time Vendor A u (ms) A d (ms) T u (ms) T d (ms) Duration (ms) Time Variation, Day, 10:1 V , V , ,574.7 V , ,129.0 Time Variation, Day, 4:1 V V V Time Variation, Night, 10:1 V1 4, , , ,998.8 V2 4, , ,949.1 V3 3, , , ,660.7 Time Variation, Night, 4:1 V1 2, , ,238.7 V2 1, V3 1, , ,513.5 TABLE 3 Spatial Variation in Activation and Termination Zones Lower Bound on Upper Bound on Lower Bound on Upper Bound on Activation Point, Activation Point, Termination Point, Termination Point, Average Excess Vendor A u (ft) A d (ft) T u (ft) T d (ft) Detector Length (ft) Spatial Variation, Day, 10:1 V V V Spatial Variation, Day, 4:1 V V V Spatial Variation, Night, 10:1 V V V Spatial Variation, Night, 4:1 V V V

171 Middleton, Longmire, Bullock, and Sturdevant 171 TABLE 4 Parameters for Measuring Detector Performance per Respective Detection Zone That Current Generation of Video Detection Systems Can Meet Test Parameter Day Night Activation position, upstream tolerance 20.0 ft ft (A u - A) Activation position, downstream tolerance 20.0 ft 20.0 ft (A - A d ) Termination position, upstream tolerance ft ft (T u - T) Termination position, downstream 50.0 ft 50.0 ft tolerance (T - T d ) Activation response time, typical (R a85% ) 1 s 1 s Activation response time, maximum (R a100% ) 5 s 5 s Termination response time, typical (R t85% ) 1 s 1 s Termination response time, maximum (R t100% ) 5 s 5 s False call duration (F d ) 15 s 15 s methodology worked in the time domain, so Table 2 summarizes the initial results. Table 3 represents a conversion to the spatial domain. The essential results in Table 2 are the upper and lower bounds of the observed activation events (A u and A d in milliseconds) and the upper and lower bounds of terminations (T u and T d in milliseconds). Again, the distribution of activations resulting from comparing VIVDS activations with the more accurate baseline activations was forced to center on a zero point. Terminations of VIVDS lagged the baseline by some amount based mostly on vehicle height, with taller vehicles causing greater length error. Analysts called the resulting parameter average excess detector on-time duration, measured in milliseconds and calculated as the average of individual differences between VIVDS vehicle length and baseline vehicle length. The analysis forced the termination distributions to center on this average length difference. Conversion into a spatial summary (Table 3) may make the specification bounds more informative and helpful to users. The resulting parameter from this conversion is average excess detector length, which might be considered as a supplementary threshold to be used in a VIVDS specification. TABLE 5 Acceptance Criteria per Detection Zone That Current Generation of Video Detection Systems Can Meet Performance Performance During Amber During Green Test Criterion and Red Intervals Interval Number of missed calls (N mc24 ) in 24 h Number of missed calls (N mc01 ) 9 9 in 1 h Number of false calls (N fc01,d ) 20 in any 1-h (day) Number of false calls (N fc01,n ) 200 in any 1-h (night) Tables 4 and 5 propose a set of tolerances and performance thresholds that the authors subjectively defined with the goal of developing a specification that all three systems could realistically achieve. These tables could be used by a jurisdiction for defining VIVDS acceptance criteria. Some jurisdictions might choose to use different performance levels based on the signal phase, but that would require the detection system to be able to monitor the signal state. For example, the Indiana Department of Transportation has defined the thresholds shown in Tables 6 and 7. CONCLUSION A procedure is proposed for defining vehicle detection performance that considers the temporal and spatial stochastic variation of vehicle detection zones used by video detectors. A suite of field tests were performed with three different well-calibrated detection systems installed at 4:1 and 10:1 aspect ratios during both day and night conditions to obtain empirical observations of what one might expect for performance characteristics. On the basis of those results, a set of tolerances and performance standards were developed (Tables 4 and 5) that all three systems were capable of meeting. Although these tolerances and performance standards are loose, defining the evaluation model and parameters establishes a framework that the profession can use as a TABLE 6 Parameters for Measuring Detector Performance per Respective Detection Zone Defined by Indiana Department of Transportation (6) Low Performance Standard Performance During Amber During Green During Amber During Green Test Parameter and Red Intervals Interval and Red Intervals Interval Activation position, upstream tolerance (A u - A) 6.0 ft 6.0 ft 6.0 ft 6.0 ft Activation position, downstream tolerance (A - A d ) 6.0 ft 6.0 ft 6.0 ft 6.0 ft Termination position, upstream tolerance (T u - T) 6.0 ft 6.0 ft 6.0 ft 6.0 ft Termination position, downstream tolerance (T - T d ) 6.0 ft 6.0 ft 6.0 ft 6.0 ft Activation response time, typical (R a85% ) 2 s 1 s 1 s 100 ms Activation response time, maximum (R a100% ) 10.0 s 5.0 s 5.0 s 1.0 s Termination response time, typical (R t85% ) 2 s 1 s 1 s 100 ms Termination response time, maximum (R t100% ) 10.0 s 5.0 s 5.0 s 1.0 s False call duration (F d ) 5.0 s 5.0 s 500 ms 500 ms

172 172 Transportation Research Record 2128 TABLE 7 Acceptance Criteria per Detection Zone Defined by Indiana Department of Transportation (6) Performance Performance During Amber During Green Test Criterion and Red Intervals Interval Number of missed calls (N mc24 ) 0 10 in 24 hrs Number of missed calls (N mc01 ) 0 10 in 1 hr Number of false calls (N fc ) 20 in 24 hours Number of false calls (N fc ) 20 in 24 hours model for discussing and refining the performance expectations that users expect the detection industry to meet. In fact, the Indiana Department of Transportation is pursuing this same framework but will use more stringent performance characteristics. Last, although this study was conducted with video detection systems, the concept could easily be extended to incorporate other sensing technologies. REFERENCES 1. MacCarley, C. A., and J. Palen. Evaluation of Video Traffic Sensors for Intersection Signal Actuation: Methods and Metrics. Presented at 81st Annual Meeting of the Transportation Research Board, Washington, D.C., Rhodes, A., D. M. Bullock, J. R. Sturdevant, Z. Clark, and D. G. Candey. Evaluation of the Accuracy of Stop Bar Video Detection at Signalized Intersections. In Transportation Research Record: Journal of the Transportation Research Board, No. 1925, Transportation Research Board of the National Academies, Washington, D.C., 2005, pp Abbas, M., and J. Bonneson. Video Detection for Intersection and Interchange Control. Publication FHWA/TX-03/ Texas Department of Transportation; FHWA, U.S. Department of Transportation, Rhodes, A., E. Smaglik, D. M. Bullock, and J. R. Sturdevant. Operational Performance Comparison of Video Detection Systems. Proc., 2007 ITE International Annual Meeting and Exhibit, August 5 8, Rhodes, A., K. Jennings, and D. M. Bullock. Consistency of Video Detection Activation and Deactivation Times Between Day and Night Periods. Journal of Transportation Engineering, ASCE, Sept Procedure for Evaluating Vehicle Detection Performance. ITM No P. Indiana Department of Transportation, Indianapolis, Jan The Traffic Signal Systems Committee sponsored publication of this paper.

173 Local Synchronization Control Scheme for Congested Interchange Areas in Freeway Corridor H. Michael Zhang, Jingtao Ma, and Yu (Marco) Nie Congestion that initiates at closely spaced highway junctions and intersections, particularly freeway interchange areas, may spread and severely degrade the operational efficiency of the whole network if not handled in a timely and proper manner. A local synchronization traffic control scheme is proposed to manage queues at those critical locations through coordination of neighboring intersection traffic signals and freeway onramp meters. By reducing the amount of traffic feeding into and increasing the amount of traffic discharging from heavily queued sections, the scheme can prevent a queue from evolving into gridlock and thus improve overall system performance. With the help of a network kinematic wave traffic flow model, the local synchronization scheme is implemented and tested on a computer for two sample networks, one small synthetic corridor network and one large, real corridor network. The numerical results indicate that this control scheme can improve the overall operational efficiency in both corridors considerably, with as much as 50% travel time savings. This control scheme appears to perform best under incident conditions and, somewhat surprisingly, compares favorably with a more complex global optimal control scheme. Congestion that initiates at closely spaced highway junctions and intersections, for example, freeway interchange areas, could severely degrade the operational efficiency of highway networks because queues formed locally can easily spread from these critical areas to adjacent road sections and ultimately lead to local or even networkwide gridlock conditions. One important reason for this type of congestion is failure of a control system to identify the critical road sections that trigger this domino effect and to act promptly to prevent it. Examples of such critical sections include short, metered freeway on-ramps and short off-ramps that do not receive adequate green times from their immediate downstream intersections. When the arrival demand is high or the discharging flow is low for such sections, queues form and grow back to neighboring urban streets or the freeway mainline and block through traffic in both cases. As transportation corridors experience longer and more congested peaks in most urban areas (1), the need for management of trans- H. M. Zhang, Department of Civil and Environmental Engineering, University of California, Davis, 1 Shields Avenue, Davis, CA J. Ma, PTV America Inc., 1145 Broadway Plaza, Suite 605, Tacoma, WA Y. Nie, Department of Civil and Environmental Engineering, Northwestern University, Evanston, IL Corresponding author: H. M. Zhang, hmzhang@ucdavis.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / portation corridors in a holistic manner becomes more apparent and has received increased attention (2). A variety of integrated freeway corridor control studies have been performed in the last two decades, which can be divided into two classes: optimization-based, time-dependent, and feed-forward systemwide control and locally traffic-responsive feedback control. In the first class of studies (3 8) the control problems are often formulated as optimal control or mathematical programming problems, with traffic flow being modeled as kinematic waves or higher-order continuum fluids and traffic control actions including ramp metering, intersection signal control, and route diversions, or combinations of them. These problems were routinely solved by linear or constrained nonlinear programming techniques. Chang and Stephanedes (3), for example, applied a center-space scheme to approximate the kinematic wave model of Lighthill, Whitham (9), and Richards (10) and solved the formulated nonlinear program by the conjugate gradient algorithm to obtain a set of time-dependent ramp-metering rates for a freeway corridor. In contrast, Papageorgiou (11) and Diakaki et al. (12) formulated a linear network control program with the storeand-forward traffic flow model (13) to compute the optimal signal timing plan and optimal diversion rates for a road network. Although most of the aforementioned studies evaluated their proposed control methods with simulation (14), some were tested in the field (15,12), and all reported significant improvements. As the aforementioned studies indicated, systemwide, optimal control-based corridor control strategies can be quite effective in reducing traffic congestion, but they often present serious challenges to implementation. First, they require extensive data inputs, such as time-dependent origin destination (O-D) data, traffic measurements (flow, density, etc.) at a sufficient number of locations, and driver behavior parameters (e.g., diversion propensity). Second, they need a significant support infrastructure such as a central traffic management center, high-performance computing devices, a reliable detection and communication system, and supportive policies and the availability of experienced operating personnel (16). Last but not least, their centralized structure makes them less robust in the presence of component failures, which are very likely to occur in a large control system with hundreds or thousands of field elements. Considering the challenges facing the first class of control methods, the second class of methods, namely, local feedback traffic control, uses local traffic information and performs traffic management in a distributed manner. In the early 1990s a study sponsored by FHWA provided a list of operational tactics to deal with corridor congestion at various levels (17). One tactic is to quickly discharge freeway on- or off-ramp traffic in severely congested conditions. A similar treatment of priority control at freeway on-ramps was also carefully studied by Tian et al. (18). Their scheme considered both 173

174 174 Transportation Research Record 2128 the on-ramp meter and the adjacent traffic signals and set the rampmetering rates to one of two preselected values to deal with various levels of congestion on these ramp segments. A VISSIM (19) simulation study shows that this simple scheme is quite effective in reducing freeway congestion. An even more myopic local control strategy was suggested by Han and Reiss (20), in which ramp-metering rates were fine-tuned to accommodate the fluctuating traffic feeding onto the controlled on-ramp. More recently, Fang and Elefteriadou (21) extended the Online Public Access Catalog logic (22) to control closely spaced intersections within a diamond interchange area. Although the local feedback control strategies just mentioned are less rigorously formulated than the systemwide control strategies and may not steer the system to its most efficient operating state, they are nevertheless quite appealing because they are distributed and less data-intensive, they may be more robust with respect to component failures, and they require fewer resources to implement and operate. In light of these features, a local synchronization scheme is developed that integrates the operation of neighboring traffic signals and ramp meters within a corridor context. This control scheme uses a network kinematic wave model in the spirit of Daganzo (23, 24), Lo (25), Gomez and Horowitz (26), and Nie et al. (27). It detects critical sections based on the amount of queued traffic on those sections. Once queuing becomes critical on a section, flows into and out of it will be adjusted simultaneously through control of the signals in nearby sections. Computational experiments indicate that such a simple control scheme performed equally well compared with a more complex, systemwide control scheme. In the next section the network traffic flow model used in formulating the control scheme is briefly introduced, followed by a description of the local synchronization scheme. After numerical experiments for two corridor networks one small synthetic network and one large real network to evaluate the potential benefits of the developed local synchronization scheme, it is compared with a systemwide control scheme. MODELING TRAFFIC FLOW DYNAMICS UNDER CONTROL A network flow model based on a cell transmission model (CTM) is presented under which the control actions from traffic signals and ramp meters are coherently embedded. Flow Dynamics on Roadway Section The basic traffic flow model used in this study is the widely accepted Lighthill Whitham Richards (LWR) model (9, 10): q x ρ + = q = f ( x ρ ) t 0, ( 1) where q = flow rate on a road section, ρ=density, and x and t = space and time variables, respectively. Daganzo (23) developed a stable numerical approximation scheme that approximates the LWR model. The model discretizes the entire time horizon T (assignment period) into small intervals t, the loading interval. Conforming to the loading interval, the model divides every road section of the network into homogeneous segments called cells such that the cell length equals the distance traversed by one typical vehicle at free-flow speed in one loading interval. Daganzo shows that if the relationship between traffic flow q and density ρ is in the form where v = free-flow speed, q max = maximum flow rate, w = backward-moving wave speed, and ρ j = jam density, then the LWR model can be approximated by the following set of difference equations: y t min n t, q, δ N n t ( ) and ni t+ 1 ni t yi t yi+ 1 t ( 4) where y i (t), y i+1 (t) = fluxes entering cell i and i + 1 at time t, respectively; n i 1 (t), n i (t), n i+1 (t) = number of vehicles in cell (i 1), i, and i + 1 at time t, respectively; q i,max = capacity flow into i at t; and N i n i = space available in i, δ= w v. Essentially, Equation 4 states that the number of vehicles staying in cell i at loading interval t + 1 is the number of vehicles from interval t plus the incoming vehicles and minus the outgoing vehicles. Daganzo (24) extended the model to a general network by carefully dividing various roadway junctions into basic merges and diverges. Since control actions take place at junctions, the flow-updating rules at general junctions including signalized intersections as well as metered ramps will be introduced. Flow Updating at Signalized Intersections Lo (25) showed that the CTM can be deployed to model flow updates at urban intersections with a few changes. If the flow capacity q max in Equation 2 is replaced by flow that depends on the signal timing variable, q ()= { () ( ())} i i 1 i,max i i 3 ( )= ()+ () () max { ( )} q = min vρ, qmax, w ρ j ρ ( 2) q ( t) = 0 max t green otherwise () 5 where it switches between q max (green) and zero (red), the end cell of an intersection approach will serve as a functioning signal, and the flow dynamics still approximates the LWR model. A generalized four-leg intersection with all vehicular movements is shown in Figure 1. The traffic streams naturally form the merging and diverging flows, the movement of which will depend on the following demand capacity analysis rules.

175 Zhang, Ma, and Nie 175 i q imax (t) j as in Equation 3 with q i,max replaced by q i,max (t) in Equation 5. This simplified treatment was also used in Lo s studies (25, 28). The flow dynamics model with signalized diverges and merges embeds all four major signal control parameters, namely, cycle length C, phase sequencing, phase duration g, and offset Δ between two adjacent signalized intersections, and their values are measured in multiples of the loading interval t. Metered Freeway On-Ramp Signalized Diverges The diverging flows occur when the traffic stream on a single link splits into left-turn, through, and right-turn movements. It is a common engineering practice to enlarge the road section behind the stop line to store the incoming vehicles. To accommodate this feature, at the end cell h of a link j approaching a signalized intersection, the vehicle accumulation is updated as follows: m m n t+ n t y t h (a) m ( )= ()+, () yh () t 1 h 1 h 1 h m= L,R,T m= L,R,T The subscript m {L, R, T} denotes the left-turn, right-turn, and through movements, respectively. The cell C j s 1 is the preceding cell of C j s. The fluxes into and out of cell s are computed by y ( t+ 1)= min n () t, q () t, δ ( N n () t ) h 1, h h 1 h,max h h m hh, + 1( 1) = h ( ) h,max( ) δ ( h+ 1 h+ 1 ( t) ) m m t m y t+ min n t, q t, N n where the notation follows Equations 3 and 4. N hm, m = L, R, T in Equation 7 are the different storage capacities for various movements, ensuring that different sizes of turning bays can be modeled properly. Signalized Merges In this study, right turns are explicitly considered in the signal timing optimization. In this way, the flow updating at intersections is simplified to be the same as a pair of sequential links, which then reads ( )= ()+ () () j j j 1 j n t+ 1 n t y t y t () { } ( 7) { } h 1 m= L, R, T (b) FIGURE 1 General representation of cell-based intersection movements: (a) control for movement and (b) movement and conflicting flow. m = L, R, T ( 8) where n j 1 (t + 1) is the number of vehicles in the starting cell of the downstream link ( j) for the next time interval (t +1), and y h j 1 (t) is the number of vehicles outgoing from the upstream link ( j 1) into the downstream link ( j). The incoming flux y h j 1 (t) is determined by the signal timing plan but shares the same updating rules () 6 An on-ramp is modeled as a merge. There is only one control variable to consider when ramp metering is incorporated into the flow dynamics model: the ramp metering rate at an entry ramp. This step is accomplished by suitably modifying the demand function (29): yi t min ni t, qi,max t, δ Ni ni ( 10) t D D t t min, R ( 11) t t t D = D + D ( 12) t S S t t min, D ( 13) f t M f ()= { () () ( )} R = ( R ) t R = ( M ) t DM D S t = ( 14) t t DR D S t = ( 15) t where ramp metering R t is embedded in the demand function D t R, and the other variables are as follows: D t R M = ramp demand at time t, D t = demand on beginning cell of link downstream of ramp, D t M = competing demand on mainline, S t M = supply of beginning cell of downstream link, S t = total service flow rate, f t R = outflow from ramp, and f t M = outflow from upstream mainline. In this generalized merge model, the freeway mainline and ramp flows will be distributed proportionally to their relative demand (Equations 13 15) when traffic is congested on both the freeway and the entry ramp (29). Traffic Demand and Traffic Routing In this study, time-dependent demand between origin r and destination s, D r,s (t), is taken as given for all OD pairs (r,s) (R,S) and for the traffic demand that enters the network at time t via the starting cell of source link j: ()=, () (, ) {(, )} ( ) j rs Q1 t D t r s R S 16 r s R where Q j 1(t) is the sum of the time-dependent demand D r,s (t) that starts from the source link j.

176 176 Transportation Research Record 2128 Also, several traffic routing options have been implemented so that traffic diversion under incident conditions can be modeled, as indicated in the numerical examples. These include routing through fixed paths (such as the shortest path for each O-D pair under free-flow conditions or a set of externally specified paths) or dynamically updated paths obtained through reactive user equilibrium traffic assignment. O c = critical occupancy at bottleneck beyond which congestion will build up upstream of bottleneck, and O t = observed occupancy at bottleneck location. Within the CTM framework, the occupancy at a roadway section is calculated as the ratio of the number of vehicles within the corresponding cell to its holding capacity: DEVELOPMENT OF LOCAL SYNCHRONIZATION SCHEME O t i ()= n t i N () i ( 18) As shown in Figure 2, local synchronization control (LSC) is centered around monitoring the traffic operations and particularly the vehicle queuing on critical links. Usually queues that are not dissipated in a timely manner will evolve into local or even networkwide gridlock. When such a potential spillback is detected, normal traffic operations will be superseded by LSC schemes, in which the control actions are synchronized to discharge the queuing traffic and reduce the feeding traffic simultaneously. Normal operations will resume after the critical queue is cleared. In this study, a normal operation consists of two parts: (a) at signalized intersections, traffic signals are operated in vehicle-actuated (VA) control mode, with full-fledged features such as phaseskipping, gap-out, and max-out (30), and (b) at entry ramps, meters are controlled by ALINEA (31), an adaptive ramp-metering algorithm of the following form: R( c ) t R + 1 t t = R + K O O ( 17) where t = metering-rate updating cycle, R t+1, R t = metering rates for successive cycles, K R = modulation parameter for metering-rate adjustments, Initialize The critical occupancy O c would simply be the one corresponding to the capacity flow rate, as shown in the triangle fundamental diagram (25), where the free-flow and congestion-flow regimes are divided. In addition, K R is a parameter that has been found to affect the metering performance only marginally; a value of 70 vehicles/h is generally recommended (32). Moreover, under normal operations it is assumed that there is no coordination between intersection traffic signals, between ramp meters, and between traffic signals and ramp meters. This scheme will serve as the base case for comparison with other modes of operation, including LSC and globally optimized operation. Typical road sections that require synchronization treatment are short ones that are controlled on either end or both ends and carry large interfacing flows. Example sections are on-ramps with either ramp meters downstream or traffic signals upstream, off-ramps leading to signalized intersections, and roadway sections that have signals on both ends. Three synchronization operations are then developed to address the different control needs at these different bottleneck sites (Figure 3): On-ramp priority (Type a, Figure 3a). The queue detector is positioned at the upstream end of the metered on-ramp. When spillback is detected, the synchronization operation will be triggered to (a) turn the meter off (i.e., ramp traffic can compete freely with the freeway traffic for the right-of-way) or (b) reduce the maximum green time (or duration) of the phases that discharge traffic to the ramp. For example, in Figure 3a, the maximum green Traffic monitoring YES Queue detected/persistent? YES A B Initiate virtual LSC cycle (a) (b) NO Virtual cycle reached? NO Synchronization operations Normal operations C D (c) queue detectors FIGURE 2 Flowchart of LSC scheme. FIGURE 3 LSC operations.

177 Zhang, Ma, and Nie 177 times of Phases 1 and 4 at Intersection A will be reduced by an adjustment factor to be explained later. Off-ramp priority (Type b, Figure 3b). The queue detector is positioned at the upstream end of the off-ramp. Once the detector is triggered, the maximum green time (or duration) that discharges the off-ramp traffic (Phases 3 and 4 of Intersection B) will be increased by the adjustment factor. Intersection internal metering (gating) (Type c, Figure 3c). This operation considers the queues building up on a critical section (e.g., C D). Typical sections are the cross-street links within interchange areas that have signals controlling the interfacing traffic between the freeway and the surface streets. The queue detector is positioned on the upstream end of the critical section; once it is triggered, the feeding phases (Phases C3 and C4) will be reduced by the adjustment factor, whereas the discharging phases (Phases D1 and D2) will be increased. This synchronization scheme is rule based and requires a set of parameters to fine-tune the control. These parameters are as follows: Queue detector position. To set up any LSC scheme, the foremost step is to determine the critical locations at which queue spillback could be frequent and deleterious. The beginning of vehicle queuing, which is characterized as the traffic state that transitions from free flow to congested flow, is conveniently defined under the CTM modeling framework (Equations 1 18) as when the critical occupancy is reached. Detection of the vehicle queues for an LSC scheme is thus the detection of the occupancies at the selected roadway section. Once the corresponding cell occupancy reaches O c (Equation 18) at a certain loading interval t, LSC actions will take over the control measures to function on a virtual cycle basis. Virtual cycle (Figure 2). A virtual cycle specifies the duration of synchronization operations. Once a virtual cycle is triggered, the synchronization operation will continue until it reaches the end of the virtual cycle. If queues persist after one virtual cycle, another one will be initiated until queues subside to a level at which normal operation would be adequate to handle them. Adjustment factor of synchronization intensity. The adjustment factor determines how much the LSC unit will meter the affected phases, that is, how long the discharging phase will be increased and the feeding phases will be decreased. It is set to avoid too aggressive LSC actions, such as reducing the feeding phases to its minimum green time all the time (18). Except for the first parameter, that is, where to place the queue spillback detectors, the other two can be adjusted and tuned in real time in field operations. The virtual cycle is determined in conjunction with the queue detection: it equals the time period required to discharge the queue under the synchronization scheme. In this simulation platform, this parameter is set equal to the number of intervals for the queued traffic to traverse all the queued sections that have triggered the queue detector. As for the adjustment factor, the numerical experiments found that one-third of the maximum phase green time duration produces satisfactory synchronization results. The placement of queue detectors, however, relies a great deal on engineering judgment and good knowledge of the local traffic network conditions. A rule of thumb is to place them on short links with control on both ends and high flows, where queue spillback is more likely to occur. CASE STUDIES Two case studies were carried out to investigate the effectiveness of the proposed LSC scheme in managing queues in corridor networks and to compare it with other types of control schemes. The first case study concerns a small, synthetic corridor network with four ramps and two signalized intersections, and the second case study involves a real-life congested corridor network in Fresno, California. Case Study 1. Simple Freeway Frontage Road Corridor LSC Versus Global Optimal Control The synthetic network keeps all major features of a freeway corridor but removes those elements that are not essential to the study. For instance, only one direction of the freeway and the arterial in the same direction are considered, but the major connections (links numbered 15 and 18 in Figure 4) are treated as two-way streets. A bottleneck is created at the downstream end of the freeway, namely, on Link 6, where the number of lanes drops from three to two. D 3 D 4 O D O D 2 14 O 3 17 O 4 O 2 : Origin D 2 : Destination 12 : Link No. Signalized node Ramp meter Phasing group FIGURE 4 Network layout and control elements for Case Study 1.

178 178 Transportation Research Record 2128 TABLE 1 Hourly Demand Structure for Case Study 1 Destination Origin D 1 (13) D 2 (8) D 3 (19) D 4 (21) O 1 (4 a ) 2, O 2 (20) O 3 (24) O 4 (25) a Node ID corresponding to the O-D. Demand Scenarios Based on the hourly flow rates shown in Table 1, a number of traffic demand scenarios were generated for a 4-h analysis period: Off-peak traffic demand pattern. All the hourly link volumes are under the link capacity values. Peak traffic demand pattern. Some facilities, especially the bottleneck sections, will be overloaded with the hourly link volumes exceeding the link capacity values and will hence potentially experience congestion. Off-peak traffic demand pattern with incident starting 1 h after the beginning of the off-peak and lasting for half an hour. The location of the incident is on freeway Link 3 between two interchanges, and it reduces the capacity of that link by 60% before the incident is cleared. Three demand release patterns uniform, triangle, and reversed triangle (see Figure 5) are used to evaluate how demand patterns affect the control algorithms. All three demand patterns share the same shape shown in Figure 5, with different demand levels. For example, the peak demand pattern is obtained by increasing the off-peak demand proportionally by a factor of 1.1. In reference to Figure 5, the hourly trip rates as shown in Table 1 are translated into different release patterns by keeping the total number of trips constant but varying the flow profile in an interval of 0.5 h. Drivers Route Choice and Traffic Routing Under both off-peak and peak demand scenarios, all travelers are assumed to take the shortest paths determined by the free-flow travel times of each link. For example, trips of O 1 D 2 will take the route In the incident case, however, a predetermined portion of traffic is diverted from the freeway route to the parallel arterial route for the trips of three O-D pairs, O 1 D 1, O 2 D 4, and O 3 D 4. Since rerouting in the presence of traveler information systems can significantly change traffic conditions in the network, dynamic routing is not permitted in this case study so that the effects of different control algorithms can be more clearly distinguished. It is, however, used in the second case study. Control Strategies Only two synchronization actions are taken in this simple network: onramp priority (Links 8 and 10) and off-ramp priority (Links 7 and 9). No internal street metering is provided for surface street intersections. In addition to the LSC scheme, two more control strategies are implemented for comparison purposes: Isolated adaptive (IA) control (VA signal control plus ALINEA ramp metering) and Genetic algorithm based offline global optimal (GO) control algorithm. The IA control strategy does not coordinate ramp meters and surface street signals, and its results are used as the basis for comparison. In this control scenario, all intersection signals are vehicle actuated and ramp meters are controlled by ALINEA Uniform Releasing Pattern 500 Number of Vehicles to Be Released Triangle Releasing Pattern Reversed Triangle Simulation Horizon FIGURE 5 Traffic demand release patterns: uniform, triangle, and reversed triangle (O 2 D 1 ). (Simulation horizon is measured in number of loading intervals.)

179 Zhang, Ma, and Nie 179 The GO control plan was obtained by applying a genetic algorithm, an often-used global optimization algorithm in traffic control (33, 34), for the entire control period. Both the metering rates and signal timing plans are updated every 30 min in this control operation (35). Numerical Results In total, 27 combinations of control (IA, GO, LSC), time periods (off-peak, peak, off-peak with incident), and demand patterns (uniform, triangle, and reversed triangle) were evaluated. The results in terms of system efficiency (measured in total network travel time, TNTT) are given in Table 2. The local synchronization control would function as the IA control if none of the three synchronizing actions (on-ramp priority, off-ramp priority, internal metering) is triggered. From the test results, except in the uniform demand patterns, where all three control strategies achieved similar performance, in all other cases local synchronization is triggered. Overall, both LSC and the GO strategy improve network performance over the IA strategy, which is somewhat expected. The improvement ranges from 1% (off-peak, reversed triangle) to 51% (incident, triangle) in travel time savings. Surprisingly, local synchronization performed remarkably well when compared with the more sophisticated GO control plan, despite its simple logic and limited use of traffic information. It performs particularly well when traffic demand and traffic conditions fluctuate more quickly, as in the incident case. Table 2 also reveals that when the total demand remains the same, how this demand is released over time can drastically affect network performance. For example, under a triangle loading pattern, the total network travel time is twice that under uniform loading. This finding concurs with everyday experience in that a concentrated peak can cause significantly more delays than a spread-out peak. Case Study 2. California State Route 41 Corridor at Fresno The network in Case Study 2, as shown in Figure 6a, consists of 25-mi long SR-41 and the system interchanges with SR-180 and SR-99. Two frontage roads, Blackstone Avenue (to the left) and Fresno Street (to the right), are major arterials that are parallel to SR-41. In the network 83 intersections have VA control and 16 on-ramps have meters. The VA signals and ramp meters provide the existing control, serving as the basis for local synchronization. The network was calibrated by various groups against the evening peak period (36), and the most congested period (5:00 6:30 p.m.) was selected in this study to investigate the performance of the proposed LSC schemes. The microsimulation indicated that the network is near saturation, where the local congestion could spread quickly if not handled properly. Implementation of LSC on SR-41 Preliminary simulation analysis of this network indicates that traffic congestion occurs at the upper portion of the network (between Herdon Avenue and Shields Avenue in Figure 6a). Closer observation of the simulation reveals that congestion starts at the surface streets within some of the interchange areas, where the freeway traffic and the surface street traffic interact frequently. Therefore, four of these key interchange areas are selected for potential LSC: Herdon Avenue, Bullard Avenue, West Shaw Avenue, and East Shields Avenue. Since the congestion forms mainly on the southbound end of Blackstone Avenue and the feeders to the southbound end of SR-41, only two types of synchronization actions are taken in these areas: intersection internal metering (Type c) and on-ramp priority control (Type a). These interchanges share the same geometric characteristics and control settings: the major intersections of Blackstone at these four cross streets and the adjacent SR-41 southbound on-/offramps. Figure 6b illustrates the geometry and proposed LSC settings at Shaw Avenue. One queue detector (A) is for internal metering and the other (B) is for on-ramp priority. Similar LSC settings are used in the other three locations as well. Test Scenarios As mentioned earlier, this network was modeled and calibrated by using microsimulation (36). The calibration found that about 50% to 70% of the drivers in this network are familiar with the local traffic patterns, which means that some rerouting would occur when there was heavy congestion in the network. To take into account the effects of dynamic routing, the LSC scheme was tested under different rerouting scenarios: No rerouting. All drivers stick to their static shortest paths with no en route changes, Reroute 0.3. Thirty percent of the drivers can update their routes when the network conditions are updated, TABLE 2 TNTT Under Three Control Strategies Control Strategies Isolated Adaptive Local Synchronization GA-Optimized Demand Scenarios (IA) (LS) (%) (GO) (%) Off-peak traffic (U a ) (U) (0.2 b ) (U) (2.2) 1,876.3 (T) 1,674.5 (T) (10.8) 1,613.6 (T) (14.0) 1,188.4 (R) 1,174.0 (R) (1.2) 1,108.9 (R) (6.7) Peak traffic (U) (U) (0.1) (U) (2.0) 4,192.3 (T) 3,609.4 (T) (13.9) 3,579.2 (T) (14.6) 2,135.5 (R) 1,869.0 (R) (12.5) 1,809.9 (R) (15.2) Incident (U) (U) (0.1) (U) (3.7) 3,845.4 (T) 1,932.3 (T) (49.8) 1,860.9 (T) (51.6) 1,535.7 (R) 1,208.0 (R) (21.3) 1,181.5 (R) (23.1) a U, T, R: traffic releasing patterns in uniform, triangle, or reversed triangle shapes, c.f. Figure 5. b The value shows the travel time reduction compared to the IA case in the same row.

180 180 Transportation Research Record 2128 (a) (b) FIGURE 6 Case Study 2, SR-41 corridor in Fresno, California: (a) network in SR-41 corridor and (b) geometry and proposed LSC settings at Shaw Avenue.

181 Zhang, Ma, and Nie 181 Reroute 0.5. Fifty percent of the drivers can update their routes when the network conditions are updated, Reroute 0.7. Seventy percent of the drivers can update their routes when the network conditions are updated, and All reroute. All drivers can update their routes when the network conditions are updated. The network conditions are updated every 2 min in this study; hence travelers who reroute can change their routes in response to changing traffic conditions every 2 min. When travelers reroute, their routes are computed on the basis of traffic conditions at that time by using reactive user-equilibrium dynamic traffic assignment (37). Test Results Figure 7 shows the network performance under existing control (IA) and LSC. It can be seen that total network travel time (TNTT) and total network delay can be further reduced with LSC. The largest improvement, at 5% to 6%, occurs when there is no rerouting in the network, that is, when the system solely relies on IA control and LSC to dissipate the congestion in the network. As the percentage of travelers who can reroute increases, the advantage of LSC over IA control diminishes. It is clear that rerouting has a significant effect on network performance under both control strategies, since rerouting allows travelers to avoid spots with heavy congestion and hence reduce the 60, , % 4.53% 5 TNTT (veh-hr) 40,000 30,000 20, % 1.55% 0.03% Percentage of Change 10, No Re-route Re-route 0.3 Re-route 0.5 Re-route 0.7 All Re-route Routing Types Existing Control LSC Percentage of change (a) 0 45, , % 6.00 Total Delay (veh-hr) 35,000 30,000 25,000 20,000 15,000 10, % 5.45% 4.33% 0.27% Percentage of Change 5, No Re-route Re-route 0.3 Re-route 0.5 Re-route 0.7 All Re-route Routing Types Existing Control LSC Percentage of change (b) 0.00 FIGURE 7 Network performance under existing control (IA) and LSC schemes: (a) TNTT and (b) total delay.

182 182 Transportation Research Record 2128 Speed (vph) % 0.18% -0.97% No Re-route Re-route 0.3 Re-route 0.5 Re-route 0.7 Routing Type % % All Re-route Percentage of Change Existing Control LSC Percentage of change (a) Speed (vph) % % % % % 0 No Re-route Re-route 0.3 Re-route 0.5 Re-route 0.7 All Re-route Routing Type Percentage of Change Existing Control LSC Percentage of change FIGURE 8 Average speed on freeway (SR-41 southbound) and on parallel arterial under existing control and proposed LSC schemes: (a) freeway and (b) arterial streets. (b) demand pressure at bottlenecks. The results also show that the marginal value of traveler information is decreasing. The largest drop of total delay or travel time occurs when 30% of the traveling population receives real-time traffic information and can rereroute, when the network experiences a total TNTT reduction of about 40%. Figure 8 shows the average speed on the freeway (SR-41 southbound) and on the parallel arterial (Blackstone between Herdon and Shields Avenues). Although the average speeds on the freeway under both control strategies are comparable, those on the arterial street show dramatic improvement under LSC when less than half of the travelers reroute. As with travel delay, the improvement diminishes when more than half of the travelers can change their routes. CONCLUSIONS As traffic in many urban transportation corridors has become more congested during the last several decades, the need for integrated corridor traffic management becomes more urgent. Although many sophisticated integrated traffic control strategies have been proposed over the years, few of them have been implemented because they usually require extensive supporting infrastructure (sensors, communications, and computing) and numerous inputs. A simple integrated control strategy is proposed, the LSC scheme, to manage queues in a corridor network in a distributed fashion. This control scheme involves three elements on-ramp priority control, off-ramp priority control, and internal metering (gating) and requires little infrastructure support (queue detection and local communication). Since the role of control in congested corridors is essentially queue management, LSC attempts to distribute the queues over a wider area through coordinating the control actions of neighboring signals and ramp meters so that queues do not get concentrated at a few locations to become the seed for local or networkwide gridlock. The tests of the LSC scheme on a synthetic and a real network show that it consistently outperforms the isolated control strategy and compares favorably with even a GO control strategy in the synthetic network. The LSC scheme appears to work best when traffic demand or traffic conditions fluctuate quickly, such as with an incident. The tests

183 Zhang, Ma, and Nie 183 also show that traffic diversion (rerouting) can significantly affect the performance of the network but its effect diminishes as more and more travelers divert. The placement of the queue detectors is vital to the LSC scheme. In this study, the locations of the queue detectors were determined after several rounds of experimentation in the simulation. A rule of thumb is to place them on those links that have small queue storage yet carry high flows. A more systematic investigation is needed to develop a set of guidelines for placing the queue detectors, and this study is left for future research. ACKNOWLEDGMENTS This research is jointly sponsored by California Partners for Advanced Transit and Highways and a dissertation grant from the Sustainable Transportation Center at the University of California, Davis. REFERENCES 1. Schrank, D., and T. Lomax. The 2003 Annual Urban Mobility Report. Technical Report. Texas Transportation Institute, Texas A&M University System, ICMS Work Plan. FHWA, U.S. Department of Transportation, workplan.htm. 3. Chang, K., and Y. J. Stephanedes. Optimal Control of Freeway Corridors. Journal of Transportation Engineering, ASCE, Vol. 119, No. 4, 1993, pp Chang, G.-L., P.-K. Ho, and C.-H. Wei. A Dynamic System-Optimum Control Model for Commuting Traffic Corridors. Transportation Research, Vol. 1, No. 1, 1993, pp Messmer, A., and M. Papageorgiou. METANET: A Macroscopic Simulation Program for Motorway Networks. Traffic Engineering and Control, Vol. 31, 1990, pp Messmer, A., and M. Papageorgiou. Automatic Control Methods Applied to Freeway Network Traffic. Automatica, Vol. 30, 1994, pp Kotsialos, A., and M. Papageorgiou. Optimal Coordinated and Integrated Motorway Network Traffic Control. In Transportation and Traffic Theory: Proc. 14th International Symposium on Transportation and Traffic Theory, Jerusalem, Israel, July 1999, pp Kotsialos, A., M. Papageorgiou, M. Mangeas, and H. Haj-Salem. Coordinated and Integrated Control of Motorway Networks via Nonlinear Optimal Control. Transportation Research, Vol. 10C, 2002, pp Lighthill, M. J., and G. B. Whitham. On Kinematic Waves. I: Flow Movement in Long Rivers; II: A Theory of Traffic Flow on Long Crowded Roads. Proceedings of the Royal Society, Vol. 229A, 1955, pp Richards, P. I. Shockwaves on the Highway. Operations Research, Vol. 4, 1956, pp Papageorgiou, M. An Integrated Control Approach for Traffic Corridors. Transportation Research, Vol. 3C, No. 1, 1995, pp Diakaki, C., M. Papageorgiou, and T. McLean. Integrated Traffic- Responsive Urban Corridor Control Strategy in Glasgow, Scotland: Application and Evaluation. In Transportation Research Record: Journal of the Transportation Research Board, No. 1727, TRB, National Research Council, Washington, D.C., 2000, pp Gazis, D., and R. Potts. The Oversaturated Intersection. In Proc., Second International Symposium on Transportation and Traffic Theory, 1963, pp Chang, G.-L., J. Wu, and S. L. Cohen. Integrated Real-Time Ramp Metering Model for Nonrecurrent Congestion: Framework and Preliminary Results. In Transportation Research Record 1446, TRB, National Research Council, Washington, D.C., 1994, pp Papageorgiou, M., J. Blosseville, and H. Hadj-Salem. Modeling and Real- Time Control of Traffic Flow on the Southern Part of Boulevard Périphérique Paris, Part I: Modeling. Transportation Research, Vol. 24A, No. 5, 1990, pp McLean, T., C. Brader, C. Diakaki, and M. Papageorgiou. Urban Integrated Traffic Control in Glasgow, Scotland. Road Transport Information and Control, Vol. 21, 1998, pp Pooran, F. J., and H. C. Lieu. Evaluation of System Operating Strategies for Ramp Metering and Traffic Signal Coordination. Proc., IVHS America Annual Meeting, Tian, Z. Z., K. Balke, R. Engelbrecht, and L. Rilett. Integrated Control Strategies for Surface Street and Freeway Systems. In Transportation Research Record: Journal of the Transportation Research Board, No. 1811, Transportation Research Board of the National Academies, Washington, D.C., 2002, pp VISSIM 5.0 North America Release Manual. PTV America, Corvallis, Ore., Han, B., and R. A. Reiss. Coordinating Ramp Meter Operation with Upstream Intersection Traffic Signal. In Transportation Research Record 1446, TRB, National Research Council, Washington, D.C., 1994, pp Fang, C. F., and L. Elefteriadou. Development of an Optimization Methodology for Adaptive Traffic Signal Control at Diamond Interchanges. Journal of Transportation Engineering, ASCE, Vol. 132, No. 8, 2006, pp Gartner, N. H. OPAC: A Demand-Responsive Strategy for Traffic Signal Control. In Transportation Research Record 906, TRB, National Research Council, Washington, D.C., 1983, pp Daganzo, C. F. The Cell Transmission Model: A Dynamic Representation of Highway Traffic Consistent with Hydrodynamic Theory. Transportation Research, Vol. 28B, 1994, pp Daganzo, C. F. The Cell Transmission Model, Part II: Network Traffic. Transportation Research, Vol. 29B, 1995, pp Lo, H. K. A Novel Traffic Signal Control Formulation. Transportation Research, Vol. 33B, 1999, pp Gomez, G., and R. Horowitz. Globally Optimal Solutions to the Onramp Metering Problem, Part I. Proc., 7th International IEEE Conference on Intelligent Transportation Systems, 2004, pp Nie, Y., J. Ma, and H. M. Zhang. A Polymorphic Dynamic Network Loading Model. Computer-Aided Civil and Infrastructure Engineering, Vol. 23, No. 2, 2007, pp Lo, H. K. A Cell-Based Traffic Signal Formulation: Strategies and Benefits of Dynamic Timing Plans. Transportation Science, Vol. 35, No. 2, 2001, pp Zhang, H. M., and W. L. Jin. On the Distribution Schemes for Determining Flows Through a Merge. Transportation Research, Vol. 37, No. 6, 2003, pp Prassas, E. S., W. R. McShane, and R. P. Ross. Traffic Engineering. Prentice Hall, Saddle River, N.J., Papageorgiou, M., H. Hadj-Salem, and J. Blosseville. ALINEA: A Local Feedback Control Law for On-Ramp Metering. In Transportation Research Record 1320, TRB, National Research Council, Washington, D.C., 1991, pp Papageorgiou, M., and C. Diakaki. Simulation Studies of Integrated Corridor Control in Glasgow. Transportation Research, Vol. 5, No. 3, 1997, pp Hadi, M. A., and C. E. Wallace. Hybrid Genetic Algorithm to Optimize Signal Phasing and Timing. In Transportation Research Record 1421, TRB, National Research Council, Washington, D.C., 1993, pp Lo, H. K., E. Chang, and Y.-C. Chan. Dynamic Network Traffic Control. Transportation Research, Vol. 35A, 2001, pp Ma, J., Y. Nie, and H. M. Zhang. Solving Integrated Corridor Control Problem Using Simultaneous Perturbation Stochastic Approximation. Presented at 86th Annual Meeting of the Transportation Research Board, Washington, D.C., Zhang, H. M., J. Ma, S. P. Singh, and L. Chu. Developing Calibration Tools for Microscopic Traffic Simulation, Final Report Part III: Global Calibration O-D Estimation, Traffic Signal Enhancements and a Case Study. PATH Research Report UCB-ITS-PRR University of California, Berkeley, Calif., Friezs, T. L., D. Bernstein, N. J. Mehta, R. L. Tobin, and S. Ganjalizadeh. Dynamic Network Traffic Assignment Considered as a Continuous Time Optimal Control Problem. Operations Research, Vol. 37, 1989, pp The authors are solely responsible for the contents of this paper. The Traffic Signal Systems Committee sponsored publication of this paper.

184 Distributed Ethernet Network of Advanced Pedestrian Signals Dustin DeVoe, Sanjeev Giri, and Richard W. Wall For more than 60 years, traffic signals have used direct wire connections between the traffic controller cabinet and the signals and detectors dispersed throughout the intersection. A network-based approach is reported for distributed control and sensing of traffic signal devices. The motivation of this research has been to improve safety and performance while reducing the cost of a signalized intersection installation. An Ethernet-based network architecture is presented that uses the National Transportation Communications for Intelligent Transportation Systems Protocol (NTCIP) for real-time signal control combined with the IEEE 1588 precision time protocol for robust operation of safety-critical applications synthesized into the term smart signals. The investigation focuses on improving accessibility and safety for pedestrians through the use of smart signals. The smart signal paradigm is the basis of an enabling technology that permits complex functioning signals and detectors with self-test capability. The bidirectional communication provides the ability to monitor the operational status of signals and detectors that are currently unobservable by automated traffic controls. The elements of network messaging that enable NTCIP to be used for real-time intersection signal control are presented. The concerns of signal outputs being generated outside the scope of observation for malfunction management units are addressed by implementing a deterministic network. The time performance of the system is evaluated for Simple Network Management Protocol and Simple Transportation Management Protocol network messages. The results of testing on the global network time synchronization are presented and show that inexpensive microprocessors can achieve stable long-term time division multiplexed operation with 100-µs accuracy. The fundamental means for controlling traffic signal lights is and has been since its inception over 60 years ago dedicated wires for controlling signal lights and detecting vehicles and pedestrians in a binary fashion (1). In 2005, Wall and Huska (2) and Wall et al. (3) presented a networked architecture for controlling traffic signal lights based on the IEEE 1451 family of transducer interface standards that results in the ability to exchange more complex information than simple binary states of indication and detection. The plug-and-play (IEEE 1451) standard was initially chosen because it supported the capabilities to potentially sim- R. W. Wall, National Institute for Advanced Transportation Technology, and D. Devoe and S. Giri, Department of Electrical and Computer Engineering, University of Idaho, P.O , Moscow, ID Corresponding author: R. W. Wall, rwall@uidaho.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / plify intersection signal installations, operation, and upgrades. A path to integration was presented in 2007 that demonstrated how the IEEE 1451 based signals and detectors can be integrated with modern TS2 controllers (4). Since the networked signals and detectors now contain innate intelligence, the system is called smart signals. In 1992, the National Electrical Manufacturers Association (NEMA) approved the TS2 standard, which specified an open architecture. This standard provided more robust methods of fault tolerance and information distribution inside traffic controller cabinets. In the standard, the TS2 Port 1 synchronous data link control (SDLC) replaced direct point-to-point wiring methods that often cluttered cabinets with proprietary solutions. The SDLC represents a distributed control environment in which the traffic controller, load switches, detectors, and malfunction management unit (MMU) share a common communication bus. Each component has particular message types defined in TS2 as frames that must propagate the network in a time-regulated network (5). Described here is how pedestrian signals and buttons can be controlled in a fault-tolerant distributed Ethernet network by using standardized protocols. The National Transportation Communications for Intelligent Transportation Systems Protocol (NTCIP) defines objects that can provide control and state information for efficient operation of a compliant traffic system (6). Research was conducted that pairs NTCIP with the IEEE 1588 Precision Time Protocol (PTP) to encompass the fault tolerance and synchronization needed for a distributed Ethernet network. It is anticipated that the improved quality of information there will benefit pedestrians by providing advanced indication of the transfer-to-walk interval, acknowledgement that a request was received from the traffic controller, and persistent updates to the countdown timing through the duration of the Walk and Pedestrian Clear interval. BACKGROUND The concept behind smart signals uses a distributed processing architecture for controlling signals and acquiring service requests from pedestrian buttons, vehicle loop detectors, and special service sensors. Internet technology is used for communication between smart devices and the traffic controller. This operation is in contrast to conventional methods used for traffic control, in which signals and sensors use dedicated wires routed from the controller cabinet. One goal of the smart signal approach was to lower construction costs by reducing the number of wires required for signalized intersections and to permit development of new devices by allowing increased complexity of information exchanged between the controller and the signals or detectors. 184

185 DeVoe, Giri, and Wall 185 National Transportation Communications for ITS Protocol In the past, each manufacturer of microprocessor-based traffic control devices and software either developed or adopted a different, proprietary protocol for data communications. This property required extensive integration projects to incorporate different systems from different manufacturers as well as to communicate between systems operated by adjacent agencies. NTCIP provides common standards for protocols that can be used by all manufacturers and system developers to help ease control network assimilation. A communications protocol defines a set of rules for messaging and how to encode the data contained in those messages. The NTCIP establishes the rules that allow bytes, characters, and strings to be organized into messages that can be decoded by other NTCIPcompliant devices. NTCIP is a family of communications standards for transmitting data and messages between microcomputer-controlled devices used in ITS. An example of such a system is a computer at a traffic control center monitoring and controlling the operation of microprocessor-based roadside controllers at signalized intersections. The computer may send instructions to the traffic signal controllers to change signal timings as traffic conditions change, and in return the intersection controllers send status and traffic flow information back to the traffic control center (7). the object. The syntax defines the data type, such as an integer or a string of octets. The encoding describes how the information associated with the managed objects is serialized for transmission between machines. SNMP uses a subset of Abstract Syntax Notation One (ASN.1) (9). The SNMP compilation rules for encoding data types into bits and bytes are defined by basic encoding rules. Definition of the MIB conforms with the SMI specified in Request For Comments (RFC) The latest Internet MIB is established in RFC 1213 and is called MIB-II (10). Traffic controller manufacturers compile their MIBs by using standardized tools. NTCIP Standards 1201, global object definitions, and 1202 (Object Definitions for Actuated Traffic Signal Controller Units) contain definitions of standardized objects with ASN.1 notation (10, 11). Proprietary objects must be defined by the manufacturer but are still defined and compiled in the same standardized manner. The successful completion of a MIB compilation results in generation of a text file that provides links to the OIDs of all addressable objects contained within the traffic controller. The text file will be referred to as OidNamesOut.txt here. Each OID is written as a sequence of decimal digits separated by periods. This sequence is generally around 17 bytes long when encoded. An object instance is identified by appending the instance number to this base OID. Thus, each instance of data within the device has a unique number associated with it. Simple Network Management Protocol for NTCIP Since its initial development in 1988, the Simple Network Management Protocol (SNMP) has become the de facto standard for internetwork management. NTCIP recognized the wide use of SNMP and adopted this protocol as a communications standard for use in the ITS industry. Because of its flexibility, it provides management stations with the ability to define their message content through the simplicity and robustness of the protocol. Even though there were concerns about the large amount of encoding overhead, it was decided that the protocol provided a core set of functionality and that companion protocols could be developed to reduce large overhead issues. SNMP is typically applied to managing network devices. Contained within the traffic controller software are managed objects, or variables, that contain parameters directly relating to the current operation of the intersection. These objects are arranged in a virtual information database, called a management information base (MIB). SNMP allows traffic controllers to communicate objects in their MIB to management centers for the purpose of status and control. In the manager agent paradigm for SNMP, managed network objects must be logically accessible. Logical accessibility means that management information must be stored somewhere and therefore that information must be retrievable and modifiable. SNMP actually provides the means for retrieval and modification by using a get-set paradigm to exchange individual pieces of data. It is also possible for the traffic controller to provide an unsolicited SNMP message that is similar to a get message but driven by an internal event; it is known as a trap. Each piece of data stored within a device that is accessible via SNMP is called an object. Objects are organized hierarchically within the MIB according to the rules set in the Structure of Management Information (SMI) Protocol (8). The SMI organizes, names, and describes information so that logical access can occur. The SMI states that each managed object must have a name, a syntax, and an encoding. The name, an object identifier (OID), uniquely identifies Dynamic Objects The NTCIP technical working group developed the Simple Transportation Management Protocol (STMP) for application layer bandwidth reduction. STMP uses a similar get-set paradigm to that of the SNMP without the protocol data unit (PDU) overhead of OIDs and error codes. The content of every data packet requires each protocol entity to have prior knowledge of the configuration of that message. Every message is built from a user-defined structured collection of variables known as a dynamic object. The process of building a dynamic object is a run-time operation that requires communication using SNMP and a list of OIDs included in the MIB. NTCIP dictates that up to 13 dynamic objects can be defined within the traffic controlling device. The size of a dynamic object is limited by the maximum packet size of the communications network (12). SYSTEM ARCHITECTURE In response to comments from practitioners and reviewers pertaining to the initial research on the smart signal concepts, the authors moved their focus to using NTCIP for communications between the smart signal devices and the TS2 cabinet, MMU, and traffic controller. Because NTCIP is not currently designed for time-critical applications, IEEE 1588 PTP time synchronization was applied for added system reliability and supervision (13). NTCIP and PTP Communications For implementation using dynamic objects within user datagram packets, the NTCIP manager controller shown in Figure 1 sends a single-byte packet to the TS2 controller, which responds in turn with the dynamic object data. The smart signal controller then rebroadcasts this packet to all smart signal devices. The configuration of the

186 186 Transportation Research Record 2128 Econolite: ASC3 MMU sync message status message Dynamic Object Response Dynamic Object Request NTCIP MANAGER PTP event port SSN port NTCIP port IEEE 1588 clock Time Triggered Ethernet status message status message receive sync message and adjust clock receive sync message and adjust clock SNMP Pedestrian Request Dynamic Object Broadcast Smart Signal FIGURE 1 Communications architecture for distributed smart pedestrian traffic signal system (SSN smart signal network). dynamic object is a non run-time configuration program executed by the NTCIP manager; it occurs at the initial start-up and is imperative for operation of the system. The dynamic object is composed of pedestrian signal status, pedestrian and vehicle detector status, Walk and Pedestrian Clear timing, and next-phase indication. The IEEE 1588 Time Sync Protocol is applied to low-level device management to reduce nondeterminism, packet collisions, and coordinate transitions by using a common synchronized clock. Time division multiple access is employed to provide each node with a specified time slot to communicate with the master coordinating device. Because the coordinating device can determine which devices are synchronized, it then acts as a distributed control MMU, referred to here as a network management unit (NMU). Figure 1 demonstrates all common intranet messages inside the PTP and NTCIP communications stack. Smart Signal Network Devices The architecture of the smart signal system is based on an Ethernet backbone. It is made up of four critical devices: the NMU, pedestrian signals, networked accessible pedestrian stations (APS), and a pedestrian smart signal controller. The smart signal network (SSN) represents all communications to allocate, manage, and control the distributed pedestrian architecture. The design shown in Figure 2 includes an NMU, which implements the MMU-type functions for the networked devices. The NMU maintains the master clock and time slot management. It represents an interpretation of how the Traffic Signal 2 (TS2) standard could evolve into a network-based model; it does not represent full TS2 compatibility. The intent is that in the event that a node behaves unexpectedly, the intersection would be put into a fail-safe state. The networked APS place pedestrian calls and receive the pedestrian signal display status by network communications. This procedure eliminates the need for control wiring between the pedestrian signals and APS that is common practice today. Besides conventional control and status functionality, the smart devices can communicate the results of self-diagnostic tests back to the controller or even to the traffic control center. The smart pedestrian signals display status information received indirectly from the traffic controller. Because of the limitations of traffic controller devices, it is not currently possible to request a dynamic object broadcast from the traffic controller. Therefore, the pedestrian smart signal controller device must request and rebroadcast the NTCIP message. METHODS AND MATERIALS Hardware To the extent possible, the hardware used for this investigation is based on standard industry products to meet the requirements of the Manual on Uniform Traffic Control Devices (MUTCD) and is organized as shown in Figure 2 (14). The added equipment inside the traffic controller cabinet includes the NTCIP manager or pedestrian smart signal controller, PTP MMU or NMU, an Ethernet switch, and an Ethernet-over-power-line (EoPL) modem. An Econolite Model ASC was used, which is a NEMA TS2 Type 2 traffic controller (15). The smart signal controller and smart PTP NMU are microcontroller units that use a Rabbit Semiconductor RCM 3000 series microprocessor with a 10-Mbps Ethernet controller (16). A network switch is needed to communicate Ethernet messages with the ASC3 traffic controller for data retrieval and

187 DeVoe, Giri, and Wall 187 WAN Maintenance Laptop Networked Pedestrian Signals Wire Ethernet Ethernet over Power Line Ped Smart Signal Controller Network Management Unit New Traffic Cabinet Hardware TS2 Traffic Controller Cabinet Conventional Traffic Cabinet Hardware Load Switches MMU/ Conflict Monitor Networked Accessible Pedestrian Stations Existing Signal Hardware FIGURE 2 Smart signal integration to TS2 traffic controller cabinet (WAN wide-area network; ASC actuated traffic signal controller unit). management. The NTCIP messages must be rebroadcast by the smart signal controller to all smart signal devices connected to the Ethernet over the Essentials of Programming Language (EoPL) network. EoPL modems function as an Ethernet hub distributed over the existing power-line infrastructure. A single EoPL modem was connected to the switch, which basically extends the capacity and ease of connectivity on the local network. Future designs will operate at low alternating-current (AC) voltage. The tests reported here used commercial models that operate at 120 V AC. Equipment outside the cabinet consists of four Econolite 12-in. polycarbonate countdown pedestrian signals that were customized to interface with the smart signal network using Netgear HDX Mbps EoPL adaptors (17, 18). The proprietary controller is based on a Rabbit Semiconductor RCM 3000 series microprocessor with 10-Mbps Ethernet controllers. The microprocessor manages signal LED illumination, networking, signal status, and pushbutton activity. Network Software The experiment requires a TS2-type traffic controller with an NTCIPcompliant MIB. Wireshark (19), a network protocol analyzer program available at no cost, was used to make the observations. Each smart signal node operates on a minimized Transmission Control Protocol/Internet Protocol (TCP-IP) Ethernet stack. COMMUNICATIONS User Datagram Protocol Ports, Size, and IP Addressing For this project the User Datagram Protocol (UDP)/IP internet transport profile was chosen for system communications, as defined in NTCIP This protocol incorporates placing the data stream into a UDP datagram and then placing the UDP datagram into an IP packet. For STMP communications, the NTCIP standard specifies that all communications be directed on Port 501, whereas SNMP typically uses Port 161. There is a significant savings in message size of STMP over SNMP. For example, the size of a single OID SNMP get-request for one object is 42 bytes. In comparison the STMP request message contains just 1 byte and can request as many objects as desired. Continuing this example, an SNMP get-response for a single integer-type OID is 43 bytes and the STMP response for that same variable would be 3 bytes. The sequence graph shown in Figure 3 demonstrates the communications for the dynamic object data within the smart pedestrian signal network. The device initiating the dynamic object request is IP address and the traffic controller is Once the response is received by the device initiating the request, it is rebroadcast to all nodes on that network using IP address on a new unique port where all of the smart signal pedestrian devices are listening. The broadcast port assignment for the data is arbitrary; however, it is critical that the pedestrian devices all be listening to the predetermined port. Use of Dynamic Objects The sequence of SNMP commands shown in Table 1 must be sent to the controller by the client to create a dynamic object. The diagram uses the term SET as an abbreviation for an SNMP set operation followed by the variable type in parentheses. The object name has a corresponding OID, which can be found in the OidNamesOut.txt. The sequence begins by setting the configuration OIDs to Invalid and then undercreation status, so that the new dynamic object

188 188 Transportation Research Record 2128 Smart Signal/ NTCIP Manager Smart Signal Button Press SNMP-SET: PedCall UDP 161 On the Next Dynamic Object Request Interval STMP-GET: Dynamic Object STMP-GET: Response: (PedTimer, PedCallStatus, PhaseWalk,...) UDP 501 STMP Rebroadcast Ped Call Acknowledged / FIGURE 3 Sequence graph for dynamic objects in smart signal network. can be created. Next the new dynamic object is specified in Steps 3 and 4 with the particular OID and the appended dynamic object number desired. Step 4 should be repeated in the order of the desired list of objects by using the OID and the incremented value n after each set. The sequence is concluded by setting the configuration status to Valid. Retrieval of Dynamic Objects A device can retrieve a dynamic object by sending 1 byte of data within a UDP message. The Transportation Management Protocol (TMP) defines all dynamic object get-requests to be composed of the hexadecimal value 0x80 masked with the number of the dynamic object. A response message contains the header 0xC0 followed by the raw serialized data. An example query with five OIDs programmed into Dynamic Object 1 is detailed in Table 2. In the observations from the experiment with the Wireshark protocol analyzer, extraneous data were included in the transmitted response packet from the controller. The origin of this anomaly comes from the definition of the UDP header information, which in this case was how the UDP minimum packet size was defined; the extraneous data should be excluded when decoded. TABLE 1 Sequence of Steps to Configure New Dynamic Object SNMP (type) Object Name Status (value) 1. SET (int) dynobjectconfigstatus Invalid (3) 2. SET(int) dynobjectconfigstatus undercreation (2) 3. SET(string) dynobjectowner.# SmartPed 4. SET(OID) dynobjectvariable.#.(1... n) SET(int) dynobjectconfigstatus Valid (1) PTP Communication Scheduling A global time base generated by using PTP synchronization is divided into mutually exclusive time slots dedicated for data transmission. In the current application, 15 time slots allocated to 10 spatially distributed devices were used. Figure 4 shows the communication scheduling scheme (20) used for distributed traffic control system applications. In Figure 4, the non-idle slots are ms long and the idle slots are ms long. Idle slots are used to allow smart signal network devices to process sync messages and NTCIP update control messages. The summation of all of the slot times results in the requirement for each scheduling cycle interval of ms. With this scheme, the NMU can detect critical and noncritical faults within a span of 203 ms. Slot zero is dedicated to the NMU, which is essentially the PTP master device. This slot is used to broadcast PTP sync messages to the network devices and contains necessary information for TABLE 2 Definitions of Packet Fields for Example Dynamic Object Payload Details 81 Request Packet of 1 Byte: STMP-GET- Response for Dynamic Object 1 C1 Seven Byte Response Header: STMP- GET-Response for Dynamic Object 1 Included Information Field 14 Variable 1 = phasepedestrianclear.2 02 Variable 2 = phasestatusgrouppedcalls Variable 3 = asc3pedtimer.1 02 Variable 4 = phasestatusgrouppedclears.1 09 Variable 4 = phasestatusgrouppedclears.1

189 DeVoe, Giri, and Wall 189 Non-idle Slot Interval = ms Idle Slot Interval = ms Cycle Interval = ms Cycle # 0 1 Slot # Slot Owner NMU IDLE IDLE Smart PED 1 Smart PED 2 Smart PED 3 Smart PED 4 Smart PED 5 Smart PED 6 Smart PED 7 Smart PED 8 NMU NTCIP IDLE IDLE NMU IDLE IDLE Time FIGURE 4 Smart signal PTP communication scheduling scheme. synchronizing PTP slave clocks. Slots numbered 3 through 10 are allocated to spatially distributed smart signals. During these time slots, the smart signals broadcast their status messages. If a request for service is initiated by a pedestrian, the pedestrian set detector NTCIP message is also sent to the traffic controller within the same slot interval. The pedestrian request is acknowledged in the NTCIP slot interval when the intersection status broadcast verifies the pedestrian detector set. Slot 11 is allocated to the NMU to transmit status messages, such as timing updates. Slot 12 is allocated to the NTCIP traffic controller timer and status update. During this slot, the smart signal controller sends intersection status messages, which could include the flashing condition or information sharing between signals and buttons. After sending the status message, the smart signal controller sends an STMP control request message to the TS2 traffic controller and rebroadcasts the received control message to the remaining smart signal devices. Fault Detection and Handling Failures occur when a component fails to meet system data integrity or time response requirements. Hardware or software failures within a distributed smart signal device that inhibit reliable operation of the overall system cause critical faults. Detection of status message omission helps the system to identify faulty nodes. A device in the active state is responsible for consistently transmitting status messages during its allocated time slot. Failure to receive status messages within the scheduled sender node s time slot is interpreted as a device fault. Faults within the smart signal controller or the NMU are considered to be critical faults. Occurrence of a critical fault triggers the system to enter the safe-fail mode. In this mode, all pedestrian signals would turn off as per MUTCD specifications (14). A failed pedestrian device is treated as a noncritical fault as long as it can be determined that the pedestrian signal is not displaying the Walk light. On occurrence of noncritical faults the system maintains normal operation; however, there is some loss of functionality. RESULTS AND ANALYSIS NTCIP 1103 defines an SNMP or STMP request as timed-out if the time from when a request is received and responded to exceeds 100 ms plus 1 ms for each byte contained in the variable-binding field (OID and data), unless otherwise specified by a communication standard. This reference will be used as SNMP and STMP responses from the ASC3 traffic controller are considered. The experiment compared two SNMP and STMP response packets that are similar in total UDP data size. Message Timing: SNMP The timing test results shown in Table 3 were generated by polling the traffic controller with unique SNMP requests every 200 ms. The average response time was around 12 ms with a maximum of 170 ms. The experiment duration was 215 min. The same test was conducted with a 225-ms polling interval, and in general the response times were between 20% and 56% more consistent and responsive. Unfortunately, response times were still erratic as signified by the relatively high standard deviation for both intervals. However, a common response time between 4 and 50 ms was observed. Message Timing: STMP Two request intervals were tested for the STMP. The timing test results shown in Table 3 were generated by polling the traffic controller with a dynamic object every 225 ms and 200 ms. The dynamic object was constructed to respond with 47 objects, which is equivalent in data packet size to the two objects queried with the SNMP tests. With the initial test of a 200-ms polling interval, the response delay actually increased at a linear rate as time progressed. The NTCIPdefined time-out condition was reached approximately 200 min into TABLE 3 Traffic Controller Response Time Statistics in Milliseconds SNMP SNMP STMP STMP 200 ms 225 ms 200 ms 225 ms Max Min SD Average

190 190 Transportation Research Record 2128 the test. At this interval, the STMP has a high standard deviation, which exceeds that of the SNMP. Overall performance was worse because the average and maximum were much higher than desired. In the 225-ms polling interval, the response times were significantly improved over those of the SNMP. The average response time was around 2 ms. The most significant improvement was the standard deviation, which was down to = from = and much better than the SNMP at = The standard deviation at the 225-ms polling interval was found to fit the behavior necessary for PTP time slicing. Advantages and Disadvantages of STMP When the timing and packet size per object payload are compared, using dynamic objects has significant advantages. The reduced processing for the protocol overhead was evident when the SNMP was compared with the STMP response standard deviation along the same polling interval. The SNMP protocol overhead seemed to cause a more erratic response delay within the tested traffic controller. In a time-critical environment, such as the PTP-constrained smart pedestrian signals, the SNMP would need to provide a more consistent reply. The STMP proved to be a much more efficient means for extracting specified pedestrian information from the traffic controller. The disadvantages of dynamic objects are few, but message integrity and abnormal timing should be considered. Part of the initial testing of dynamic objects yielded inconsistent data because of a lack of knowledge of the objects contained in the STMP response. The receiving agent must have prior knowledge of the dynamic object structure because it is not encoded within the message. Response delay was also an issue. The abnormal behavior of the traffic controller at a 200-ms polling period was reason for concern. It is concluded that in these test conditions, instances in which a dynamic object is polled at a period of less than 225 ms for an extended time will result in an increasingly undesirable response delay. Synchronization Error Factors such as clock drift, clock resolution, clock variance, communication latency, and communication delay fluctuations must be taken into account in order to synchronize clocks across an Ethernet network. Each of these factors contributes to the synchronization error. Figure 5 shows a frequency distribution plot for the percent synchronization error of this system. The plot was obtained by running a synchronization test for 12 h. During this test, samples were taken from the PTP master and the slave clocks every 500 ms. The clock samples were then compared to generate the data for the plot in Figure 5. During the majority of the test period interval the synchronization error was limited to ±100 µs. The worst-case synchronization error for this 12-h test was 808 µs. It is important to minimize this error in order to develop a communication scheduling scheme with mutually exclusive time slots. Omission Fault Detection The smart signal controller loses its time slot after its Ethernet cable is removed. The rising edge on the NMU failure detect signal represents the omission failure detection. The NMU detects the omission fault within 10 ms because the communication slot in which the dynamic object is broadcast was expected shortly afterward. On detecting this omission failure, the NMU puts the smart signal network into a fail-safe mode of operation. This mode does not occur until the status update slot, which occurs shortly before the smart signal detects the fault condition. FUTURE WORK Throughout the development of the smart signal concept, feedback from a diverse group of traffic industry practitioners was seriously considered. To resolve their concerns, the following topics were identified for further research: Extended NTCIP managed communications. The smart signals currently rely on a predefined dynamic object. In future iterations it would be useful to have an interface that allows for additional configuration of dynamic objects and the managed nodes in the field. NTCIP-compliant accessible pedestrian buttons. Current operation has pedestrian button management inside the smart signal software. For field testing and future municipal acceptance of Ethernet-networked traffic devices, development of an advanced pedestrian button designed for special needs users would be the most prudent. Intersection fault modes. For the PTP-managed MMU it is now necessary to research more advanced fault conditions beyond the omission fault documented here. 25 Frequency Distribution (%) Synchronization Error (µs) FIGURE 5 Synchronization error distribution using PTP IEEE 1588.

191 DeVoe, Giri, and Wall 191 Alternate traffic controllers. Although this experiment works well with the Econolite ASC3, in order to achieve industry adaptation, the system should be tested on other (Ethernet-enabled) traffic controllers that conform to NTCIP standards. NTCIP traps. The implementation of NTCIP-defined traps would significantly reduce network traffic and possibly increase reliability. Unsolicited messages from the traffic controller preempting signal and timing transitions would be a next step in integrating smart signals into the traffic controller. Traffic controller operations efficiency supervisor. Development of an independent application that analyzes the efficiency of the traffic timing plans could be implemented by using dynamic objects to extract the information. CONCLUSION An efficient method for distributing signal control information to pedestrian traffic signals is described. The overall pedestrian system effectiveness was improved by implementing an Ethernet infrastructure. The system described has the capability for improved communications that can result in more reliable operations and a higher degree of functionality. Experiments demonstrated that lowcost microprocessor-based devices are readily able to communicate by using NTCIP and PTP. Both SNMP and STMP objects were investigated, and it was shown that there was a performance threshold with a 200-ms polling interval. The measured delays were unique to the software of the TS2 controller selected for the experiment. Given that traffic controllers are designed for real-time operation, the authors believe that software modifications to any NTCIP-compliant controller would lessen the observed delays. The STMP has distinct performance advantages over the SNMP for managing real-time control because of its inherent protocol efficiency and simplicity. The PTP communication scheduling scheme was integrated with eight smart signals that were synchronized and managed as an MMU would within the TS2 standard. The devices were synchronized to ±100 µs. However, future distributed traffic control applications that require more devices may use multiple smart MMUs to handle multiple clusters without increasing fault detection time. Omission failure was demonstrated and detected within 10 ms, and the system could be placed into flash mode within 203 ms. ACKNOWLEDGMENTS Funding was provided to the National Institute of Advanced Transportation Technology at Idaho University by RITA, U.S. Department of Transportation. Equipment and engineering support were provided by Econolite Control Products, Inc., Anaheim, California, and Campbell Company, Boise, Idaho. REFERENCES 1. Sessions, G. M. Traffic Devices: Historical Aspects Thereof. Institute of Traffic Engineers, Washington, D.C., 1971, pp Wall, R. W., and A. Huska. Design Platform for Plug-and-Play IEEE 1451 Traffic Signal. Presented at 31st Annual IEEE Industrial Electronics Conference, Raleigh, N.C., Nov 6 10, 2005 (Paper RD ). 3. Bullock, D. M., R. W. Wall, and A. Huska. Application of Plug-and- Play Distributed Signal Technology to Traffic Signals. Presented at 85th Annual Meeting of the Transportation Research Board, Washington, D.C., Wall, R. W., T. Urbanik II, D. M. Bullock, S. Allen, M. Busby, D. DeVoe, A. Huska, and T. Rallens. Intelligent Traffic Signal Control: Adding Pedestrians to the System. Presented at 86th Annual Meeting of the Transportation Research Board, Washington, D.C., Traffic Controller Assemblies with NTCIP Requirements. NEMA Standards Publication TS National Electrical Manufacturers Association, Rosslyn, Va., NTCIP 9001: The NTCIP Guide Version 3.02b. AASHTO, ITE, NEMA, Washington, D.C., Dubuisson, O. ASN.1 Communication Between Heterogeneous Systems. Morgan Kaufman Publishers, Sept html. Accessed July 31, RFC1902: Structure of Management Information for Version 2 of the Simple Network Management Protocol (SNMPv2). Association for Computing Machinery, RFC1902&dl=ACM&coll=portal. 9. Simple Network Management Protocol. Cisco Systems, Inc., Oct #wp NTCIP 1202: Object Definitions for Actuated Traffic Signal Controller Units (ASC) Version 1.07b. AASHTO, ITE, NEMA, Washington, D.C., NTCIP 2202: Internet (TCP/IP and UDP/IP) Transport Profile version AASHTO, ITE, NEMA, Washington, D.C., NTCIP 1103: Transportation Management Protocols (TMP) version 2.10b. AASHTO, ITE, NEMA, Washington, D.C., Giri, S. Application of a Safety Critical Network for Distributed Smart Pedestrian Signals in a Road Traffic Intersection System. MS thesis. Electrical and Computer Engineering Department, University of Idaho, Jan Manual on Uniform Traffic Control Devices for Streets and Highways. FHWA, U.S. Department of Transportation, dot.gov/pdfs/2003r1r2/mutcd2003r1r2complet.pdf. 15. Econolite Control Products, Inc., Anaheim, Calif. pdf/controllers/asc3.pdf. 16. Rabbit Semiconductor, Davis, Calif. products/rcm3000/rcm3000.pdf. 17. Econolite Control Products, Inc., Anaheim, Calif. docs/signals/signals_12inch_pedestrian_polycarbonate_data_sheet.pdf. 18. Netgear Inc., Santa Clara, Calif. Networking/PowerlineEthernetAdapters/HDX101.aspx. 19. Wireshark. A Network Protocol Analyzer. GNU2. Wireshark Foundation Pop, P., P. Eles, and Z. Peng. Scheduling and Bus Access Optimization for Time-Driven Systems. In Analysis and Synthesis of Distributed Real-Time Embedded Systems, Kluwer Academic Publishers, Boston, Mass., 2004, pp The Traffic Signal Systems Committee sponsored publication of this paper.

192 Comparison of Before After Versus Off On Adaptive Traffic Control Evaluations Park City, Utah, Case Study Cameron Kergaye, Aleksandar Stevanovic, and Peter T. Martin An adaptive traffic control system, the Sydney Coordinated Adaptive Traffic System (SCATS), was installed in Park City, Utah, to improve traffic performance at its network of signalized intersections. A field evaluation of the previous time-of-day actuated coordinated signal timings was conducted before SCATS installation to compare the two systems. However, the post-scats field evaluation could not occur until two additional signals were installed and several other changes were made to the network. Two years after the original pre-scats field evaluation, the network was reevaluated with an off on technique analogous to a before after study. The signal timings and parameters in the off condition forced SCATS to use time-of-day actuated coordinated control, similar to the before study but with timings readjusted for the additional signals and changed traffic conditions. The performance gains with SCATS on were measurably greater than those with SCATS off for travel time and number of stops and greater overall for stopped delay. These data provided the transportation agency with results anticipated by its traffic engineers. However, the original field evaluation data from 2 years earlier provided a less distinct conclusion. A methodology is presented to determine the relevance of an off on study in place of a before after study. The results show that before and off data sets behave consistently 62.5% of the time. This value provides a basis of support for using off data, which better represent before signal timings on an after network. It also quantifies sensitivity to changes in the network, which are substantial in this case. Usually, after new corridor traffic management strategies are applied, such as improvements to road geometry or installation of a new traffic control regime, evaluations are performed to measure the effects of those strategies. A before after study is usually conducted to measure traffic performance or to survey commuters to quantify implemented improvements. When an adaptive traffic control system (ATCS) or any centralized system is installed with the ability to implement various signal timing plans, evaluators of the new system C. Kergaye, Utah Department of Transportation, 4501 South 2700 West, Salt Lake City, Utah A. Stevanovic and P. T. Martin, Department of Civil and Environmental Engineering, University of Utah, 122 South Central Campus Drive, Room 104, Salt Lake City, Utah Current affiliation for A. Stevanovic: Department of Civil Engineering, Florida Atlantic University, 777 Glades Road, Building 36, Room 231, Boca Raton, FL Corresponding author: A. Stevanovic, aleks@trafficlab.utah.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / reapply the old traffic control timings within the new conditions for the sake of a fair comparison. This procedure is followed when a new ATCS is installed to replace old time-of-day (TOD) plans. Most ATCSs have the ability to turn off their adaptive control algorithms and implement the TOD signal timings that were in effect previously. In this way, both old and new traffic control strategies are exposed to approximately the same traffic conditions as opposed to a beforeand-after study, in which traffic conditions could significantly change between completion of these studies. This approach to evaluating new and old traffic control is often called an off on study. Off refers to the new traffic control being switched off (and instead the old TOD plans are running) and on refers to the new traffic control being implemented. Both methods, before after and off on, have some advantages and disadvantages. An off on study represents an assessment of the two traffic control modes on the same system. Although signal timing implementation strategies such as ATCS and TOD are different, there are a number of signal settings that are common (e.g., minimum green, amber, and red clearance). In the event that these settings are extensively changed, the TOD strategy running under the new ATCS umbrella may not truly represent the old TOD plans. However, a weakness of the before after study approach is that road and traffic conditions may change between study periods. The new conditions may include changes in road and intersection geometry, traffic demand and distribution on the network, speed limits, on-street parking, signal phasing, and signal controller settings that are not perceived as major changes. If significant changes do occur, the before-and-after study may be compromised. The problems with significant changes in traffic conditions are mostly avoided when off on studies are utilized, in which the time interval is usually short enough to minimize those changes. Evaluations of ATCSs have been performed in both before after and off on studies. More recent U.S. ATCS evaluations were performed using the off on approach (1 6). However, there is no study in which these two approaches are compared. Yet there is a need to know more about the difference in the two approaches. How much do these changes, which occur during installation of a new ATCS, affect the reported performance of the new system? Can it be said confidently that there is no difference between before after and off on evaluation studies? A case study answers these questions for an ATCS installation in Park City, Utah. All the important changes in the Park City network that could have biased the outputs are discussed. Those traffic measurements that were affected by changes in traffic conditions were eliminated from the analysis, and results from data collections done for both studies are compared. 192

193 Kergaye, Stevanovic, and Martin 193 BACKGROUND A network of signalized intersections operating TOD actuatedcoordinated timing plans in Park City was converted to an ATCS in order to maximize traffic performance benefits from a system that continually adjusts or adapts to changes in traffic flows. After several years of investigating adaptive traffic control in microsimulation environments, the Utah Department of Transportation (UDOT) decided to install the Sydney Coordinated Adaptive Traffic System (SCATS), a system developed by the Road and Traffic Authority (RTA) of New South Wales, Australia. ATCSs, also known as realtime traffic control systems, have been used widely since the early 1980s. Although still not extensively used in American cities, these systems have been deployed in more than 30 U.S. locations (7). Park City was selected as a deployment site because it was a fast-growing area that often experiences significant shifts in traffic demand due to its frequent recreational and artistic events. The objective of the SCATS deployment was to reduce travel times, vehicle delays, and number of stops in the network. SCATS has been evaluated many times (1 6). These previous ATCS field evaluations in North America showed that most of these systems improve traffic performance on the networks under control. These improvements vary between 5% and 45% although deployment of ATCSs does not always improve every performance measure. ATCSs are usually evaluated by comparing their impacts on traffic with the impacts from conventional traffic control systems. The impacts are captured through various performance measures, which are either collected in the field or delivered by microsimulation tools. Conventional traffic control is usually represented either by signal timings from the field or optimized signal timings from offline optimization tools (e.g., TRANSYT-7F, SYNCHRO, and PASSER). Conventional signal timings are often called TOD plans because multiples of them are implemented throughout the day, thus reflecting diurnal fluctuations in traffic demand. The TOD plans are implemented through either fixed-time traffic control or vehicle-actuated traffic control. In fall 2005 UDOT authorized researchers to conduct an evaluation of the network operating under TOD actuated-coordinated signal timings in order to provide a baseline for comparison to the succeeding ATCS. The intent was to perform a before after SCATS evaluation to verify and perhaps justify the new signal system. The performance measures were to be collected before SCATS was installed and again after its installation and fine-tuning. In this way, performance measures under the original actuated-coordinated field control would be compared with performance measures under SCATS control. The before data were collected in August 2005 and prepared for comparison with the data from the eventual after-scats evaluation. However, two new intersections were signalized and added to the existing system of 12 signalized intersections. Also, one of the existing signalized intersections was both realigned and redesigned from a three-leg into a four-leg intersection. These modifications occurred after the before data were collected but before SCATS was fully deployed. The three modified intersections were eventually brought under the SCATS umbrella and fine-tuned within a year and a half after the start of the SCATS installation. These modifications complicated the original before after evaluation study. Therefore, UDOT opted to perform an evaluation similar to a before-and-after study by substituting performance data obtained while SCATS was turned off. When adaptive logic within SCATS is turned off, a fallback regime known as Flexilink runs TOD plans under actuated coordinated control. Through By programming of TOD plans to those of August 2005, 12 of the intersections resemble conditions before SCATS. With the addition of two signalized intersections, modifications to the third one, and changes in traffic volumes and posted speeds, UDOT decided not to precisely match the previous timings. So the off timings were fine-tuned for the current conditions and presumed to represent before timings if the previous timings had been updated to accommodate the network changes. Although conventional traffic signal systems across the United States may not be regularly updated, using updated off timings provides an optimal performance baseline. Any advantage from SCATS on over SCATS off could be viewed as an expected minimum. The SCATS off on study was conducted in September 2007 with results indicating superiority for SCATS on (2). The performance measures included travel times, number of stops, and intersection stopped delay. These are the same measures used in the original 2005 field evaluation for the before conditions. The objective of this research is to compare before after and off on studies of SCATS and determine the adequacy of an off study as a substitute for a before study despite significant changes over time. A methodology is also presented to provide statistical equivalency between before after and on off studies. Overall relevance is discussed with recommendations to transportation agencies and researchers. FIELD EVALUATIONS The first field evaluation of SCATS was conducted by the RTA and known as the Parramatta Experiment. Field performances were measured for various traffic control types on both open and closed networks (6). On the closed (central business district) network, SCATS was mostly better in terms of stops and journey times than isolated vehicle-actuated control or fixed-time control. Similar results were seen on the open (arterial) network. In the few cases in which journey times under SCATS control were not reduced, they were of similar performance to the ones measured under other traffic controls. Oakland County in Michigan has the most installations of SCATScontrolled intersections in the United States and was field-evaluated extensively (3). The findings showed that SCATS reduced the number of stops, side-street delays, and left-turn delays when compared with fixed-time control. More recently, SCATS has been evaluated in Cobb County, Georgia, and Gresham, Oregon. In Cobb County, SCATS was compared with recently updated signal timing plans for semi-actuated traffic control. The evaluation had two components: technical performance and driver satisfaction. A rigorous comparison of technical performance showed that neither control system was superior (4). Results of the driver satisfaction study were similar. They were decidedly unbiased toward either control with the explanation that travel times, speeds, and delays were similar for each signal system (5). In Gresham s field evaluation, SCATS was compared with 3-yearold coordinated actuated signal timing plans (1). The study focused on traffic performance on the main corridor, and investigation of side-street performance was planned for the future. Initial results showed that SCATS reduced travel time by 16% and delay by 42% on weekdays. The impact on weekends was even greater. In summary, there have been a variety of field evaluations of SCATS, in which SCATS was shown to be much better when compared with conventional signal timings, especially when the timings have not been updated recently. It also seems that newer and more rigorous studies show fewer benefits than the initial evaluations. This

194 194 Transportation Research Record 2128 finding is perhaps due to the fact that conventional traffic control itself has become more adaptive in recent years. Vehicle-actuated traffic control (usually coordinated for a system of closely spaced intersections) has become the predominant type of U.S. traffic control. Moreover, new features of signal controllers support vehicle-actuated operations that are more responsive than ever before. DESCRIPTION OF STUDY CORRIDOR SCATS was deployed at 14 intersections along the Park City road network (Figure 1) before the off on study in The Park City road network consists of two suburban arterials, State Routes (SR) 224 and 248, and many cross streets. SR-224 is a five-lane major arterial, and SR-248 is a three-lane minor arterial. Both arterials have a median left-turn lane. Posted speed limits on SR-224 range from 55 mph along sparse sections with widely spaced intersections to 35 mph in the downtown area. Speed limits along SR-248 are 35 mph in the downtown area and 50 mph as the road leaves the network. Recent traffic counts have shown that SR-224 and SR-248 carry around 30,000 and 16,000 vehicles per day, respectively. This corridor is the primary route for recreational traffic from Salt Lake City and other areas to ski resorts and recreational areas in Park City and serves as a connector between Interstate 80 and U.S. Route 40. The network can be divided into three distinct parts (encircled in Figure 1) on the basis of prevailing traffic conditions: Kimball Junction is a single-point urban interchange for SR-224 and I-80 with neighboring signalized intersections (Landmark Drive and Olympic Park). This area hosts many local businesses generating work-related and shopping traffic. It has the highest traffic demand throughout the day with level-of-service (LOS) C at the three intersections during the p.m. peak. The close proximity of the intersections dictates the need for coordination. Bobsled Drive, Bear Hollow, Sun Peak Drive, Canyons Drive, Payday Drive, and Thaynes Canyon Drive make up six intersections in the middle of the network and at the beginning of the downtown area, providing access to residential and recreational areas. Intersections in this area have LOS A or B during the p.m. peak. Spacing between intersections allows some signals to run uncoordinated. Park Avenue and SR-248, Park Avenue and Deer Valley, Deer Valley and Bonanza Drive, Bonanza Drive and SR-248, and SR km FIGURE 1 SCATS intersections in Park City, Utah.

195 Kergaye, Stevanovic, and Martin 195 and Comstock Drive are intersections that form a small gyratory system providing circulation for traffic to access cultural and historical downtown Park City. The demand for these intersections is LOS B to A during the p.m. peak. Spacing warrants coordination between these intersections, but it is not mandatory because traffic flows can be very low. CORRIDOR CHANGES A number of changes occurred to the Park City network between August 2005 and September Intersections were added or modified, speed limits were adjusted, traffic volumes fluctuated, and signal timings were changed. The following sections specifically describe these changes and any mitigating measures taken to align the evaluation studies. Road Geometry In August 2005, there were 12 signalized intersections in Park City. Bobsled and Sun Peak became signalized later. Another intersection (Canyons Drive) was realigned and redesigned from a three-leg to a four-leg intersection. The new spacing between Bear Hollow and Sun Peak or Sun Peak and Canyons Drive became particularly short. In order to compare similar performance between evaluation studies, only similar intersection and segment data are used in the analysis. Intersection stopped delay data from Bobsled and Sun Peak are not used for the comparison. Travel time and number of stops are similarly not considered for any segment containing these two intersections. This change affects four segments in each direction. Speed Limits The long stretch of SR-224 from Olympic Park to Payday Drive was posted at 55 mph in August With the new signals, the three closely spaced intersections of Bear Hollow, Sun Peak, and Canyons were changed to 45 mph (after April 2006). There were no changes to speed limits along SR-248 or in downtown Park City. Traffic Volumes Automatic traffic count data were obtained from UDOT in the form of hourly vehicle counts. Two permanent traffic counters were located along the corridor: Automatic Traffic Recorder (ATR) 605 was located along SR-224 and ATR 606 was located along SR-248. The monthly average for ATR 605 was 16,400 vehicles in August 2005 and 14,446 vehicles in September 2007, for a decrease of nearly 2,000 vehicles in the 2-year period. The monthly average for ATR 606 was 7,773 vehicles in August 2005 and 7,677 vehicles in September Traffic levels in Park City also have a strong seasonal nature, as shown in Figure 2. Gaps in the data shown in Figure 2 are due to ATR malfunctions or damage. However, there are enough data on both roadways to reflect surges in demand. The obvious impact in comparing evaluation periods is the effect from the start of the academic year. Just a single month can make a large difference. Had the 2007 study taken place in August, traffic volumes (and possibly performance results) may have been statistically closer. The difference in hourly traffic levels for the two evaluation periods is shown in Figure 3. Volumes are averaged over the data collection period to represent a typical day. Although the before traffic levels 45,000 40,000 35,000 Monthly Average Daily Traffic 30,000 25,000 20,000 Before Data Collection On & Off Data Collection 15,000 10,000 5,000 SR 248 SR 224 FIGURE 2 Average monthly traffic, Park City.

196 196 Transportation Research Record Traffic Flow (veh/hour) Time of Day Before Off On (a) Traffic Flow (veh/hour) Time of Day Before Off On (b) FIGURE 3 Daily profiles of traffic flows, Park City: (a) SR-224 and (b) SR-248.

197 Kergaye, Stevanovic, and Martin 197 TABLE 1 are quite distinct from the off and on levels for both roadways, the difference is small. Signal Timings Olympic Park and SR-224 Intersection Signal Timings Phase Split Before Off Minimum Before Off Extension Before Off Maximum Before Off Yellow Before Off Red Before Off Walk Before Off Ped. Clear Before Off The signal timings representing TOD actuated coordinated plans were collected for all comparable intersections employing beforeand-after and off-on timing plans. Table 1 contrasts the timing parameters for one intersection that experienced some of the network s heaviest peak-hour traffic. The cycle length was 128 s and the offset was 16 s for both before and off timing plans. It should be noted that all before timings were taken from field cabinet controller information at the time. There are often inconsistencies between parameter values and controller functions, such as imbalances between ring cycle lengths, in which case the longest cycle length governs. This discrepancy occurred in at least one instance. Nearly 600 individual signal timing parameters for all common intersections are used in both evaluation periods. The correlation coefficient for all these pairs of similar values is 0.84, which indicates that before and off parameters have a strong linear relationship, as seen in Figure 4. This finding supports the acceptability of an off study as a substitute for a before study. It also supports, to some extent, the interchangeability of their performance data if changes between study periods were negligible. PERFORMANCE MEASURES Network changes and all performance data collected for the before after and off on evaluations, where after data are essentially on data, ultimately determine the substitutional value of before and off evaluations. Drastic changes to the physical corridor or traffic demand may explain differences in measured performance. However, the performance of each type of control is actually less important than their relative performance and the underlying disaggregated values. Two types of traffic performance data collections were conducted. Vehicle probe data were collected to measure traffic performance along major roads in the Park City network. Stopped-delay studies were conducted at intersections to investigate the impact of SCATS performance on both side-street and main-street traffic during the peak traffic periods. Before data were collected between August 9 and August 30, Traffic data during the off period were collected from September 9 to September 15, The data collection period for the on study was the following week, from September SCATS Off Before SCATS FIGURE 4 Signal timings, Park City network.

198 198 Transportation Research Record 2128 to September 22. All data were collected between 7:00 and 9:00 a.m. for the morning peak and between 4:00 and 6:00 p.m. for the afternoon peak. Routes for the vehicle probe data collection were based on major traffic flows on the Park City corridor. Northbound and southbound routes were traversed to collect travel times, stopped delays, and number of stops. The southbound route starts at a stop sign before Kimball Junction and goes along SR-224 until Deer Valley is reached. The route then turns left onto Deer Valley and then turns left again at Bonanza Drive. The final turn is made at SR-248, where the route goes right toward Comstock Drive and ends shortly afterward. The same path is used to collect the performance measures in the opposite direction. The length of the route is around 7.35 mi. The minimum required number of 13 test runs (for each direction) was computed according to guidelines from NCHRP Report 398 for the suggested sample size for data collection on arterial streets (8). This number was exceeded for each peak period and directional run. The vehicle probe runs were conducted with the floating-car technique as described in the ITE Manual of Transportation Engineering Studies (9). Each probe vehicle was driven to pass as many vehicles as those that passed the probe vehicle. Approximately the same number of runs were made in the median and curb lanes. Drivers of the probe vehicles used handheld computers with Global Positioning System devices to log vehicle latitude and longitude at 1-s intervals. In total more than 500 vehicle runs were collected. All of the collected data were exported to a spreadsheet by a customized computer application for further analysis. Vehicle travel time and stop data are shown in Figure 5. Each set of bars is a comparison of average segment data for all directions and peak periods for the three evaluation studies. In the case of off on performance, SCATS is superior, as originally expected by transportation agency personnel. In comparison with before performance, the superiority of SCATS is less obvious. Intersection stopped-delay studies were conducted to assess the influence of the SCATS system on stopped delay for major traffic movements at the intersections. These studies were done by two observers according to the procedure specified in the 2000 Highway Capacity Manual (10). A series of 5-min studies of each major lane group at the intersections (through and left-turn movements) were conducted with 16-s sampling intervals. Stopped-delay data were collected for eight major movements (four through movements and four left-turn movements). In total, more than 50 peak traffic hours of stopped-delay data were collected. Intersection stopped delay is the average delay for vehicles stopping at any intersection. The data summarized in Figure 6 show stopped delay for through and left-turn movements along the main and side streets averaged over four intersections. In comparing off on data, the largest average reduction in stopped delay occurs for through movements. Both the main and side roads experienced less delay by 2 s or more during SCATS operation. Stopped delay for left turns was less consistent. On the main roads, left-turn stopped delay was reduced approximately 7 s during SCATS operation, whereas on the side street left-turn stopped-delay differences were negligible. The comparison with before data is also in favor of SCATS operation for each road type and turning movement. EVALUATION OF STUDIES To determine the adequacy of using an off evaluation in place of a before evaluation, the effects from changes in the network must be minimized. Segments that included either of the two new intersections, for example, were omitted from any performance comparisons rather than allow excess incompatibility. Other segments that were affected because of pedestrian or school crossings, which were not a factor in the before study because school was not in session, were also eliminated (e.g., Comstock Drive). It is then necessary to formulate a methodology for comparing before and off evaluations. A direct comparison of performance measures would be appropriate if performance per se was being compared. So an indirect comparison of their performance with that of another system, in this case one running SCATS, was used as the basis. Assuming that external factors for all three evaluations are similar, if before and off performance data behave consistently compared with on performance data, it can be assumed that the two TOD actuated coordinated timing plans are equivalent. In addition, any possible advantage to using SCATS can also be determined. Figure 7 illustrates the process of comparing performance measures from the three evaluation studies. The first step is to compare population means of the performance measures by using the t-test. The on data are used as the basis of comparison with a null hypothesis (α =0.05) that the means of all performance measures are the same for each pair of data sets. Comparing the t-test results for before on and off on yields four possible outcomes as shown in the first column (T Tests) of Figure 7. If both t-test comparisons meet the hypothesis (before on and off on) then before and off evaluations may be considered equivalent. If neither t-test comparison meets the hypothesis (before on and off on), then secondary comparisons of means are used to determine equivalencies as described in the following. If only one t-test comparison meets the hypothesis (before = on and off on, or vice versa), then before and off evaluations cannot be considered equivalent though secondary comparisons of means may be used to determine possible SCATS advantages. There are cases in which both t-test outcomes do not meet the hypothesis (B4 on and off on). In those cases means are compared with the outcomes shown in the second column of Figure 7 (Comparisons of Means). If both before and off performance are better than on performance (B4 < on and off < on), before and off evaluations may be considered equivalent, with SCATS on proving to be a disadvantage (a score of 0). If both before and off performance are worse than on performance (B4 > on and off > on), before and off evaluations may be considered equivalent, with SCATS on proving clearly advantageous (a score of 100). If before performance is better than on performance, whereas off performance is worse than on performance (B4 < on and off > on), or vice versa, before and off evaluations may not be considered equivalent and it cannot be concluded that SCATS has any advantage (a score of 50). Seventy-two performance measures were processed through the logic of Figure 7 representing travel times and number of stops collected from eligible segments of the network. A histogram of performance result weightings is shown in Figure 8. The center bar consists of 47 occurrences, only 3 of which prove inconsistent SCATS advantages. Coupled with the single SCATS advantage occurrence, 62.5% of all tests indicated consistency between before and off evaluations. CONCLUSIONS Field evaluations before an ATCS installation provide realistic baseline data for postinstallation performance comparisons. The ability of an ATCS such as SCATS to replicate conventional signal timings affords an alternative baseline that may be used in lieu of a before-

199 Kergaye, Stevanovic, and Martin Travel Time (seconds) NB AM NB PM SB AM SB PM Combined Before SCATS SCATS Off SCATS On (a) 0.8 Average Stops per Travel Time Segment per Vehicle NB AM NB PM SB AM SB PM Combined Before SCATS SCATS Off SCATS On (b) FIGURE 5 Before SCATS versus SCATS off versus SCATS on: (a) travel times and (b) average number of stops (NB northbound; SB southbound).

200 200 Transportation Research Record Average Stopped Delay (sec/veh) Main Street Through Main Street Left Side Street Through Side Street Left Before SCATS SCATS Off SCATS On FIGURE 6 Average intersection stopped delay. T Tests Comparisons of Means Results Weight Begin T-tests B4 = On AND Off = On Yes B4/On Study = Off/On Study. Neutral SCATS advantage. 50 No B4 < On AND Off < On Yes B4/On Study = Off/On Study. SCATS disadvantage. 0 B4 On AND Off On Yes No B4 > On AND Off > On Yes B4/On Study = Off/On Study. SCATS advantage. 100 No No B4/On Study Off/On Study. Inconclusive advantage. 50 B4 On AND Off = On Yes B4 < On Yes B4/On Study Off/On Study. Possible SCATS disadvantage. 25 No B4/On Study Off/On Study. Possible SCATS advantage. 75 No Off < On Yes B4/On Study Off/On Study. Possible SCATS disadvantage. 25 No B4/On Study Off/On Study. Possible SCATS advantage. 75 FIGURE 7 Determination of performance data correlation.

201 Kergaye, Stevanovic, and Martin Frequency FIGURE 8 0 Points 25 Points 50 Points 75 Points 100 Points SCATS Advantage Performance correlations. and-after evaluation. Well-conducted before and off evaluations should provide statistically equivalent performance measures if their signal timing parameters and field conditions are approximately identical. The Park City SCATS evaluations tested the limits of a before and an off comparison because of the many network changes. The addition of two new signals, modifications of another adjacent signal, speed limit adjustments, and a decrease in traffic volumes along one route were among the obvious changes to the physical network. However, the numerous changes to signal timings also contributed to limiting the substitutional value of before and off evaluations. Outputs from before on and off on studies were compared and it was found that results behaved consistently 62.5% of the time. In terms of SCATS performance it was found that SCATS is slightly better than conventional control when assessed through both before on and off on approaches. Formerly, when SCATS was assessed only through an off on approach, SCATS results were more favorable. Further research may be useful to determine a correlation between the magnitude of parameter changes and the subsequent impacts on performance. REFERENCES 1. Peters, J. M., J. McCoy, and R. Bertini. Evaluating an Adaptive Signal Control System in Gresham. Presented at the ITE Western District Annual Meeting, SCATS%20ITE%20Paper.v2.pdf. 2. Stevanovic, A., C. Kergaye, and P. T. Martin. Field Evaluation of SCATS Traffic Control in Park City. Utah Traffic Lab report. University of Utah, Salt Lake City, Wolshon, B., and W. C. Taylor. Impact of Adaptive Signal Control on Major and Minor Approach Delay. Journal of Transportation Engineering, ASCE, Vol. 125, No. 1, 1999, pp Hunter, M., S. K. Wu, and H. K. Kim. Evaluation of Signal System Improvements: Cobb County ATMS Phase III. vt.edu/documents/jan2006annualmeeting/sundayworkshop/hunter_ signal_workshop_jan_22.pdf. 5. Petrella, M., S. Bricka, M. P. Hunter, and J. E. Lappin. Driver Satisfaction with Urban Arterial After Installation of Adaptive Signal System. Presented at 85th Annual Meeting of the Transportation Research Board, Washington, D.C., Luk, J. Y. K., A. G. Sims, and P. R. Lowrie. SCATS: Application and Field Comparison with a TRANSYT Optimised Fixed Time System. Proc., International Conference on Road Traffic Signaling, Institution of Electrical Engineers, London, United Kingdom, 1982, pp Stevanovic, A., and P. T. Martin. Integration of SCOOT and SCATS in VISSIM Environment. Presented at PTV Vision International Users Group Meeting, May 24 25, Park City, Utah, utah.edu/documents/integration%20of%20 SCOOT%20&%20SCATS %20in%20VISSIM.pdf. 8. NCHRP Report 398: Quantifying Congestion. TRB, National Research Council, Washington, D.C., Robertson, D. H., J. E. Hummer, and D. C. Nelson. Manual of Transportation Engineering Studies. ITE, Prentice Hall, Englewood Cliffs, N.J., Highway Capacity Manual. TRB, National Research Council, Washington, D.C., The Traffic Signal Systems Committee sponsored publication of this paper.

202 Determination of Significant Critical Movements to Generate Traffic Scenarios for Large Arterial Networks Sherif Lotfy Abdelaziz, Montasir M. Abbas, and Catherine C. McGhee Any traffic signal control mode requires generation of multiple traffic scenarios for design purposes as well as for validation of the control system. The overall system performance depends on the traffic scenarios used to design the entire system. Therefore, all traffic patterns that might exist in a control system should be considered according to their probability of occurrence on a daily basis. However, when it comes to large networks, considering all combinations of traffic movement levels becomes time-consuming and impractical. A new approach to generate traffic scenarios for large networks is proposed. This approach is based on the selection of significant critical movements controlling the network. Selection of these critical movements is performed by using statistical correlation analysis of actual detector data and synthetic origin destination analysis for the entire network. The proposed approach was implemented in the design of a traffic-responsive control mode for the Reston Parkway arterial network, in Northern Virginia, which has 14 intersections. Detector data were then used to validate the results of the proposed approach. The validation shows no significant difference between the actual detector data and the proposed procedure s results and therefore proves that the traffic system was correctly modeled and sufficiently represented with the proposed approach. Generation of different traffic scenarios is one of the required steps to conduct different traffic engineering analyses. It is important to generate realistic traffic scenarios representing the actual network characteristics. The generated scenarios should cover a wide range of possible traffic variations in the network. The accuracy of the analysis being performed is directly related to the traffic scenarios used in the analysis. There are no clear guidelines or approaches for generating reasonable scenarios. Most of the existing techniques consider all traffic patterns based on the analysis to be conducted, even those that might not exist in reality. Inaccurate traffic patterns lead to inadequacies in the conducted analysis. For example, traffic scenarios for any network should reflect the actual patterns and major movements on such a network; otherwise the analysis will not represent the real network. S. L. Abdelaziz and M. M. Abbas, Department of Civil and Environmental Engineering, Virginia Polytechnic Institute and State University, 301 Patten Hall, Blacksburg, VA C. C. McGhee, System Operations and Traffic Engineering, Virginia Transportation Research Council, 530 Edgemont Road, Charlottesville, VA Corresponding author: M. M. Abbas, abbas@vt.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / Considering all traffic patterns has the advantage of including the effect of all patterns in the analysis. However, it is not acceptable in some cases to include traffic scenarios that rarely exist. For example, traffic scenarios being used to design signal control modes should be representative but not comprehensive. Moreover, for large networks, considering all possible traffic patterns becomes time consuming and computationally impossible. A new approach to generate traffic scenarios is proposed. This approach uses correlation between different movements in the network accompanied with synthetic origin destination (O-D) analysis to determine the traffic movements controlling the whole network. These movements are called significant critical traffic movements. With these critical movements, different traffic scenarios are generated. The procedure ensures that the generated scenarios cover the actual traffic variations in the network as well as the actual traffic combinations. The proposed approach is implemented in the Reston Parkway network in Northern Virginia. The project scope includes designing of traffic-responsive control for the Reston Parkway network, which consequently requires that multiple traffic scenarios be considered during generation of timing plans for signal control. BACKGROUND There is limited literature about traffic scenario generation. A report on implementing traffic-responsive control in different networks published by the Texas Transportation Institute (TTI) indicates that all traffic scenarios for each network should be considered when optimum timing plans are determined for each network. TTI networks have between three and five intersections. The study used clustering of detector data to determine traffic levels for traffic entering the network from all origins and then distributed it equally over all destinations (1). The timing plans obtained for each network show improved network performance. Although the TTI approach provides a good methodology for the networks being considered, equal traffic distribution is not a good assumption for all networks. In order to generate more realistic traffic scenarios for large networks, O-D matrices should be estimated, and since these O-D matrices differ with the time of day and are costly for large networks because of the prohibitive number of required sensors, the synthetic dynamic demand module (SDDM) is proposed to generate dynamic O-D matrices for traffic networks (2 4). In this method, the static detector counts are used to generate the O-D matrix to be used in different analyses. The only drawback of this method is that it requires 202

203 Abdelaziz, Abbas, and McGhee 203 driver behavior data, which are not available for all networks. Another way to estimate an O-D matrix is by using the path flow estimator (PFE) to generate and improve the O-D matrix for traffic networks (5 7). However, this method requires a large amount of data to generate accurate results. A new approach based on significant critical movements is proposed to generate traffic scenarios on traffic networks. In this approach, the major traffic movements are considered by generating an O-D matrix and then relating it to the correlation between detector counts. STUDY NETWORK Figure 1 shows the Reston Parkway network, located in Northern Virginia. The network consists of 14 intersections with a total length of 16,572 ft. The spacing between intersections ranges from 524 ft to 3,309 ft. The speed limit for the main arterial is 45 mph and ranges from 15 mph to 45 mph for the side streets. Eleven intersections are four-leg intersections, Intersection 13 is a three-leg intersection, and Intersections 6 and 7 are four-leg intersections with one-way side streets. Also, Intersection 10 has only right-turn movements for Node 02 Node 01 Node 03 Node 04 Intersection # 13 (3-legs) Node 05 Node 06 Intersection # 10 (Only side street right turns) Node 08 Node 10 Node 36 Node 07 Node 37 Node 09 Node 12 Node 11 Node 25 Node 24 Node 14 Intersections # 06 and 07 (One-way side streets) Node 16 Node 13 Node 15 Node 18 Node 17 Node 20 Node 22 Node 19 Node 21 Node 23 FIGURE 1 Reston Parkway network in Northern Virginia.

204 204 Transportation Research Record 2128 the side streets, that is, no through movements or left turns from the side streets. Nodes of the Reston Parkway network have equal attractiveness for traffic. Actual detector data from this network are taken for a period of 1 month starting from April 5, 2008, to May 6, These detectors cover almost the whole network and have a record every 15 min. The following sections present the details of the proposed approach to deal with large arterial networks considering Reston Parkway as an example. PROPOSED APPROACH ANALYSIS STEPS The proposed approach is based on four analysis steps, each of which affects the others significantly. 1. Traffic Level Determination Step 1 includes clustering of detector counts for both the main arterial and side streets. The purpose of this step is to determine the traffic levels for the movements entering the network. 2. Correlation Analysis for Detector Data In Step 2, the correlation between different movements on the network is determined. The purpose of this step is to come up with a good understanding of the relationship between different movements in the network. This step is proposed in the TTI approach. Clustering can be based on the volumes of different link movements (i.e., left, through, and right) or on link flow volumes. In this study traffic levels are based on the link flow volumes. The clustering should be done for the main arterial and separately for the side streets since it is believed that the side streets do not have traffic levels as high as those on the main arterial. K-means clustering is proposed to be used for the traffic level determination. MATLAB (8) is used to perform the entire analysis. K-means clustering uses an iterative algorithm that minimizes the sum of the distance from each object to its cluster centroid over all clusters. This algorithm moves objects between clusters until the sum cannot be decreased any further. The number of clusters should be provided to the k-means function as an input so that it attempts to minimize the distances over this given number of clusters. This step is repeated for different numbers of clusters. The best number of clusters that represents the number of traffic levels is then determined by using the silhouette value. This value determines how good the clustering of data using the given number of clusters is. Finally, to determine the best number of clusters, a graph presenting the number of clusters versus each cluster silhouette value is drawn and the number of clusters having the maximum silhouette value for the main arterial and side streets is selected (Figure 2). On the basis of the k-means analyses performed for the main arterial (Reston Parkway) and all side streets, five traffic levels for the main arterial and three traffic levels for the side streets are recommended to be used to design traffic-responsive control. K-means results are in the form of a vector including each object and the cluster to which this object is assigned. This vector is used to determine the flow limit for each level. Figure 3 shows the limits for the main arterial clusters and the side-street clusters. Table 1 summarizes the cluster limits for both main arterial and side streets. 3. Synthetic O-D Analysis Traffic entering the network from each origin node is distributed over all destinations. The distribution percentages for different traffic levels at each origin node are determined. This is a very important step to generate realistic traffic patterns. 4. Determination of Significant Critical Movements Step 4 combines the results of the previous three steps to determine the significant critical movements that control the entire network. After that, traffic patterns are generated. TRAFFIC PATTERN GENERATION FOR LARGE ARTERIAL NETWORKS In this section, the details of the proposed approach to deal with large arterial networks are presented. In order to simplify and clarify the approach steps, the Reston Parkway arterial network is considered as an example. The four steps described in the previous section are performed on this network. Traffic Level Determination (K-means Clustering) The first step to generate traffic patterns for any network is to determine the traffic levels for different movements in the entire network. Correlation Analysis for Detector Data As discussed previously, the main idea in the proposed approach is to determine the significant critical movements controlling the whole network. The significant critical movements can be defined as those that do not have any correlation or have low correlation with other movements. At the same time, they have considerable variation in traffic levels. These movements will be used to generate traffic patterns for the required design. On the basis of this definition, traffic movements in any traffic network can be classified into critical movements and noncritical movements. The noncritical movements are the highly correlated movements, which means that if the traffic level for one of them increases, all traffic levels for other movements that are highly correlated with such a movement increase as well, and vice versa. The situation is different for the critical movements; if the traffic level for one of the critical movements increases, it does not mean that other levels in the network increase. Using these concepts of critical movements and noncritical movements to generate traffic patterns will ensure that the generated patterns are only the patterns that take place in reality. In other words, there will be no pattern with zero probability. It is therefore obvious that a correlation analysis should be done for the detector data so that correlation factors between each movement and all other movements can be obtained. The SAS statistical package (9) was used to perform the entire correlation analysis. This analysis should only concern movements entering the network either from the main arterial or from the side streets. However, side-street through

205 Abdelaziz, Abbas, and McGhee Silhouette Value Number of Clusters (Log Scale) (a) Silhouette Value Number of Clusters (Log Scale) (b) FIGURE 2 Silhouette value corresponding to different number of clusters: (a) main arterial traffic and (b) side-street traffic.

206 206 Transportation Research Record Silhouette Value LEVEL 1 LEVEL 2 LEVEL 3 LEVEL 4 LEVEL Traffic Volume (mph) (a) Silhouette Value LEVEL 1 LEVEL 2 LEVEL Traffic Volume (mph) (b) FIGURE 3 Link flow limits for clusters: (a) main arterial traffic and (b) side-street traffic.

207 Abdelaziz, Abbas, and McGhee 207 TABLE 1 Cluster Limits for Main Arterial and Side Streets Link Flows: Main Arterial Link Flows: Side Streets (vph) (vph) Traffic Level Minimum Maximum Minimum Maximum , , ,493 2, ,276 4,900 NOTE: The sign indicates that the traffic level does not exist for the side streets. movements are also considered because these through movements affect the final timing plans. Thus, two correlation runs should be performed: one for all movements entering the network and another for all side-street through movements. For the Reston Parkway arterial network, the two correlation runs were performed; Table 2 is the correlation table for movements entering the network and Table 3 is that for side-street through movements. These correlation factors alone cannot determine which movements can be considered highly correlated with other movements. Therefore, it is important to define a threshold between the highly correlated movements and the uncorrelated movements. This threshold is found by running k-means clustering with only two clusters over all correlation factors for both SAS runs. The threshold is found to be 0.50 for movements entering the main arterial and 0.60 for side-street through movements. For the movements entering the network, if the correlation factor between any two movements is more than 0.50, these two movements are highly correlated. In Tables 2 and 3 the red (shaded) cells represent the correlation factors that are more than 0.50 and 0.60, respectively. It is seen from Table 2 that five movements are not correlated with any other movements. Therefore, these movements can be considered as independent and critical: northbound through and the westbound right at Intersection 1, eastbound left at Intersection 4, westbound left at Intersection 5, and southbound through at the last intersection, Intersection 14. Movements entering the network from the first and last intersections are considered critical because they represent boundaries of the entire network. Also it is clear from Table 3 that all side-street through movements are highly correlated. The question now is which one of these five critical movements is significantly critical. In other words, does each one of these movements affect the network significantly? The answer for this question would be based on (a) the maximum actual observed traffic level on links where these movements exist (this maximum observed level should be assigned as a constraint for the level of all movements on such links) and (b) at which level such movement is significant (if any one of these movements is found to be on a link with significant variation in its traffic level). Synthetic O-D analysis gives a clear answer. Synthetic O-D Analysis Synthetic O-D analysis aims to determine the distribution percentages for traffic entering the network from each origin node to all destination nodes. Since the traffic entering the network from any origin node is constrained by the maximum observed traffic level at this node, synthetic O-D analysis is performed for each of the possible traffic levels at each node. It is obvious that for the noncritical movements obtained from the correlation analysis, the levels of all these movements should increase or decrease together. If any of these noncritical movements has a maximum of one observed level, this level should not be affected by the variation in other movements. Moreover, if one of these movements is found to be on a link with a number of observed traffic levels that is less than the maximum number of levels for such links (Reston Parkway has three levels for side streets and five for the main arterial), the maximum level on this link should be applied as a constraint for higher traffic levels of other links. For the critical movements, the traffic level for the links on which these movements belong varies regardless of what is happening to other link flow levels. This level is also constrained by the maximum observed traffic level for links containing these critical movements. This is an important point since low correlation between certain movements and other movements in the network does not mean that such movement affects network performance significantly. Low correlation might be found for any movement while it has only one level, which should not be considered as significant as other movements with wide traffic level variation. When this constraint is applied to the five critical movements obtained from the correlation analysis, it is found that the eastbound left-turn movement at the fourth intersection belongs to a link with only one traffic level. This means that variation of this movement is not expected to have a significant effect on the network. Thus, the eastbound left-turn movement at the fourth intersection is excluded; it is not a significant critical movement. The other four movements are found on links with a wide range of traffic level variation. Therefore, they are expected to affect the network significantly. Synthetic O-D analysis provides more verification. It provides distribution percentages for each movement in the network, which confirm the significance of such movements. The correlation analysis does not include all possible movements because it is based on the actual detector data, which sometimes cannot show the traffic volume for right turns and left turns when there are shared lanes. Some of these shared right or left movements can be significant critical movements as determined by their distribution percentages. Synthetic O-D analysis provides a good tool to determine such missing movements. Synthetic O-D analysis for the Reston Parkway arterial network is performed on the basis of the actual detector data. QueensOD software (10) is used to perform the required synthetic analysis. One run is done for each traffic level combination of all links subjected to the maximum observed link flow, as mentioned earlier. The validity of the QueensOD package was discussed in previous research efforts (11, 12). Synthetic O-D analysis not only provides the distribution percentages for each movement but also gives the distribution percentages for traffic entering the network from each origin node over all destination nodes. Figure 1 includes node numbers used in the QueensOD runs. These O-D percentages are important since they indicate which major movements in the network are significant. These distribution percentages are expected to change during the day, and since time of day is represented here with the traffic level on the network, these distribution percentages are determined for each traffic level. Tables 4 through 7 present the distribution percentages for different traffic levels.

208 208 Transportation Research Record 2128 TABLE 2 Correlation Factors Between Movements Entering Reston Parkway Arterial Network Movement NBT EBL WBR EBR EBL WBR WBL EBR EBL WBR WBL EBL WBR WBL WBR WBL 1 NBT EBL WBR EBR EBL WBR WBL EBR EBL WBR WBL EBL WBR WBL WBR WBL EBL WBL EBR EBL WBL EBR EBL WBR WBL EBR EBL WBR WBL EBR EBL SBT WBL NOTE: Red cells contain correlation factors more than 0.50, which is considered the threshold between highly correlated movements and uncorrelated movements. Headers in the first columns and first row indicate intersection number, then movement; for example, 8 EBR means Intersection Number 8 and eastbound right movement. NB, SB, WB, and EB refer to northbound, southbound, westbound, and eastbound, respectively. L, T, and R refer to left movement, through movement, and right movement, respectively.

209 Abdelaziz, Abbas, and McGhee EBL WBL EBR EBL WBL EBR EBL WBR WBL EBR EBL WBR WBL EBR EBL SBT WBL

210 210 Transportation Research Record 2128 TABLE 3 Correlation Factors Between Side-Street Through Movements for Reston Parkway Arterial Network Movement EBT WBT EBT WBT WBT EBT WBT EBT WBT EBT WBT EBT WBT EBT WBT 1 EBT WBT EBT WBT WBT EBT WBT EBT WBT EBT WBT EBT WBT EBT WBT NOTE: Red cells contain correlation factors more than 0.60, which is considered the threshold between highly correlated movements and uncorrelated movements. Headers in the first columns and first row indicate intersection number, then movement; for example, 8 EBT means Intersection Number 8 and eastbound through movement. TABLE 4 Distribution Percentages for Side Streets: Traffic Level 1 Destination Nodes Origin nodes NOTE: = unavailable movement.

211 Abdelaziz, Abbas, and McGhee 211 Determination of Significant Critical Movements To finally determine the significant critical movements for the Reston Parkway network, both the correlation analysis and the synthetic O-D analysis were considered and performed at the same time. Each one of these two analyses has an advantage that the other one lacks. For Reston Parkway, it was found that five movements were not correlated with any other movements. One of those five movements, the eastbound left-turn movement at Intersection 4, was found to be on a link with only one traffic level. Thus, it was excluded. The other four movements were considered critical. Synthetic O-D analysis confirms that the remaining four movements have a great effect on the network since each one of them has a high distribution percentage for the traffic coming from the link to which it belongs. Moreover, from Tables 4 through 6, another movement, westbound right at Intersection 7 (not shown in the correlation analysis because it is a shared right turn), is found to be significant to the network. The distribution percentage for this movement is found to be very high (85% and more). In addition, it belongs to a link with three traffic levels. This movement is then confirmed with the Virginia Department of Transportation to be a significant critical movement. Synthetic O-D analysis also shows other movements that belong to links with three or two traffic levels and have high distribution percentages, such as the eastbound right-turn movement at the third intersection, westbound left-turn movement at the fourth intersection, and westbound right-turn movement at the fifth intersection. Although these movements satisfied the required two conditions for significant critical movements, they were not considered as significant as they appear in that the correlation analysis found all these movements to be highly correlated. In other words, their levels increase and decrease together, and therefore they are considered noncritical movements from the start. Thus, the final significant critical movements for the Reston Parkway network are as follows: Intersection 1, northbound through and westbound right turn; Intersection 5, westbound left turn; Intersection 7, westbound right turn; and Intersection 14, southbound through. Figure 4 shows the major movements for each of the selected critical movements. These major movements are determined from the distribution percentages obtained from the synthetic O-D analysis. VALIDATION OF PROPOSED APPROACH With the five significant critical movements obtained by combining the correlation analysis results with synthetic O-D results and the distribution percentages for traffic entering the network from each Total % per Movement L T R

212 212 Transportation Research Record 2128 TABLE 5 Distribution Percentages for Side Streets: Traffic Level 2 Destination Nodes Origin nodes TABLE 6 Distribution Percentages for Side Streets: Traffic Level 3 Destination Nodes Origin nodes

213 Abdelaziz, Abbas, and McGhee 213 Total % per Movement L T R Total % per Movement L T R

214 214 Transportation Research Record 2128 TABLE 7 Distribution Percentages for Main Arterial: Different Traffic Levels Destination Nodes Main Arterial Level 1 Origin nodes Main Arterial Level 2 Origin nodes Main Arterial Level 3 Origin nodes Main Arterial Level 4 Origin nodes Main Arterial Level 5 Origin nodes WBR 5WBL 14SBT origin node over all destinations for all possible traffic level combinations, different traffic patterns were generated. Traffic levels for the noncritical movements movements that are highly correlated with each other or do not affect the network significantly increase at the same time to form the traffic background to be added to different levels of the significant critical movements. The traffic level for each noncritical movement is limited to the capacity of the link to which it belongs. Traffic patterns were generated by using all possible traffic backgrounds, traffic levels for noncritical movements, and all combinations of traffic levels for critical movements. These patterns were to be used to obtain timing plans to be implemented in the entire network. The selection of timing plans is beyond the scope of this paper. The obtained traffic patterns should be verified before being used to obtain timing plans. This verification aims to make sure that these patterns are not significantly different from the actual patterns on the network. The validation process was performed by comparing the obtained traffic patterns with the actual detector data. Tables 8 and 9 show some of the obtained traffic patterns and how they match with the actual detector data. These data show very little difference between the traffic patterns obtained with the proposed approach and the actual traffic patterns obtained from detector data. 1WBR 1NBT FIGURE 4 Major distributions of critical traffic movements on Reston Parkway network. SUMMARY Generation of traffic patterns is an important step for different traffic engineering applications. For small arterial networks with up to five or six intersections, all possible traffic patterns can be considered since the computations are not too time consuming and the assumption of equal traffic distribution for traffic generated from each origin over all destinations is acceptable for some analyses. For large arterial networks, it is impossible to consider all traffic combinations because the computations will be time consuming and

215 Abdelaziz, Abbas, and McGhee 215 Total % per Movement L T R moreover the equal traffic distribution assumption is not acceptable. The proposed approach is to simplify the process of generating traffic patterns for large networks. Correlation analysis and synthetic O-D analysis were used together to obtain what is termed significant critical movements, defined as the movements that are not correlated with any other movements and span a wide range of traffic variation. The significant critical movement concept was implemented in the generation of traffic scenarios for the Reston Parkway network in Northern Virginia, and the obtained traffic scenarios were validated with actual detector data. This approach shows great results when compared with the actual traffic data. A methodology is presented here for the generation of design traffic scenarios on large arterial networks. The generated scenarios can be used to determine the timing plans on the network for either time-of-day or traffic-responsive control modes. ACKNOWLEDGMENTS This work was sponsored by the Virginia Transportation Research Council (VTRC). The materials and methods presented were developed as part of the VTRC project Evaluation of Traffic Responsive Control Mode in Northern Virginia. TABLE 8 Example for One Traffic Pattern Generated with Proposed Approach Eastbound Westbound Northbound Southbound Intersection L T R L T R L T R L T R , , , , , , , NOTE: Values are expressed in vehicles per hour. The dash indicates movement that does not exist.

216 216 Transportation Research Record 2128 TABLE 9 Actual Traffic Pattern Obtained from Detector Data Eastbound Westbound Northbound Southbound Intersection L T R L T R L T R L T R , , , , , , , NOTE: The dash indicates movement that does not exist. REFERENCES 1. Abbas, M. M., N. A. Chaudhary, G. Pesti, and A. Sharma. Guidelines for Determination of Optimal Traffic Responsive Plan Selection Control Parameters. Research Report Texas Transportation Institute, College Station, Marchal, F., and A. De Palma. Dynamic Traffic Analysis with Static Data: Guidelines with Application to Paris. In Transportation Research Record: Journal of the Transportation Research Board, No. 1752, TRB, National Research Council, Washington, D.C., 2001, pp Ran, B., N. M. Rouphail, A. Tarko, and D. E. Boyce. Toward a Class of Link Travel Time Functions for Dynamic Assignment Models on Signalized Networks. Transportation Research, Vol. 31, No. 4, 1997, pp Wu, J., and G.-L. Chang. Estimation of Time-Varying Origin Destination Distributions with Dynamic Screenline Flows. Transportation Research, Vol. 30B, No. 4, 1996, pp Chen, A., P. Chootinan, and W. W. Recker. Examining the Quality of Synthetic Origin Destination Trip Table Estimated by Path Flow Estimator. In Journal of Transportation Engineering, ASCE, Vol. 131, No. 7, 2005, pp Ashok, K., and M. E. Ben-Akiva. Dynamic Origin Destination Matrix Estimation and Prediction for Real-Time Management Systems. Proc., 12th International Symposium on Transportation and Traffic Theory, Berkeley, Calif., Ashok, K., and M. E. Ben-Akiva. Estimation and Prediction of Time- Dependent Origin Destination Flows with a Stochastic Mapping to Path Flows and Link Flows. Transportation Science, Vol. 36, No. 2, 2002, pp MATLAB 14 version Mathworks, Inc., Natick, Mass., SAS Qualification Tools User s Guide. SAS Institute Inc., Cary, N.C., QueensOD Rel 2.10 User s Guide: Estimating Origin Destination Traffic Demands from Link Flow Counts. M. Van Aerde and Assoc., Ltd., Kingston, Ontario, Canada, Van Aerde, M., H. Rakha, and H. Paramahamsan. Estimation of Origin Destination Matrices: Relationship Between Practical and Theoretical Considerations. In Transportation Research Record: Journal of the Transportation Research Board, No. 1831, Transportation Research Board of the National Academies, Washington, D.C., 2003, pp Van Aerde, M., B. Hellinga, G. MacKinnon, and J. Wang. QUEENSOD: A Practical Method for Estimating Time-Varying Origin Destination Demands for Freeway Corridors/Networks. Presented at 72nd Annual Meeting of the Transportation Research Board, Washington, D.C., The Traffic Signal Systems Committee sponsored publication of this paper.

217 Evaluating Green-Extension Policies with Reinforcement Learning and Markovian Traffic State Estimation Zain M. Adam, Montasir M. Abbas, and Pengfei Li Several protection algorithms strive to reduce the number of vehicles trapped in the dilemma zone. These algorithms use some arbitrary policies such as terminating the green when only one vehicle is present in the dilemma zone and the dilemma zone has not cleared after a certain period of time. The research proposes a control agent that is able to develop and adapt an optimal policy by learning from the environment. The agent incorporates a Markovian traffic state estimation into its learning process. A novel approach is presented for controlling traffic signals so that the number of vehicles trapped in the dilemma zone is reduced in an optimal fashion according to changes in traffic states. A comparison between the proposed optimal policy and the emerging detection-control system two-stage policy was conducted, and it was found that the policy based on reinforcement learning reduced the number of vehicles caught in the dilemma zone by up to 32%. Vehicle accidents lead to more than 40,000 deaths and 2,780,000 injuries in the United States each year and many more worldwide (1). Vehicle accidents in fact account for more than 22% of the total fatal injuries (2). Intersection-related crashes make up more than 25% of all traffic crashes, resulting in many fatalities and severe injuries (3). In 2000, more than 2.8 million intersection-related crashes occurred, representing 44% of all reported crashes. About 8,500 fatalities (23% of the total fatalities reported) and almost one million injury crashes occurred at or within an intersection. In 2003, 8,569 people died and more than 1.4 million suffered injuries as a result of intersection-related crashes (4). The probability of a crash at a signalized intersection is associated with the concept of the dilemma zone, usually defined as between 2 and 6 s from the stop bar at the onset of the yellow indication on any given approach. If trapped in the dilemma zone, some drivers may decide to proceed, and some drivers may decide to stop, resulting in rear-end collisions. The dilemma zone is also associated with red light running incidents, when some drivers who decide to proceed enter the intersection during the red indication, which can lead to right-angle collisions. Limited, experimentalstage research has examined the dilemma-zone problem. The state of the art so far includes simple green-extension systems, enhanced Department of Civil and Environmental Engineering, Virginia Polytechnic and State University, 301-D Patton Hall, MC 0105, Blacksburg, VA Corresponding author: M. M. Abbas, abbas@vt.edu. Transportation Research Record: Journal of the Transportation Research Board, No. 2128, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp DOI: / green-extension systems, and green-termination systems (5 8). Because of limited information about the expected traffic variation characteristics, these control systems typically make simplified assumptions, such as constant speed between the point of detection and the intersection stop bar, as well as simplified driver decision modeling (driver response to the dilemma zone with acceleration or deceleration, lane changing, etc.) that does not take into account different traffic stream characteristics. Unfortunately, these simplified assumptions and lack of better traffic state estimation can lead to inadequate control actions that jeopardize intersection safety. In addition, there is a lack of theoretical development of signal control strategies for intersection safety. Current signal control strategies were developed to work with archaic traffic-detection simple binary logic (information on vehicle presence or lack of presence). The logic was originally developed to provide input for old electromechanical controllers that were developed in the early 1920s and was sufficient for only that control (9). The current state of the art in intersection control is not adequate to reverse the intersection-related crash statistics; therefore, a more comprehensive and novel approach is needed. Meeting these objectives will require some radical changes in the current intersection control logic that not only focus on intersection safety but also incorporate traffic state variation knowledge in control actions. EMERGING U.S. DILEMMA-ZONE PROTECTION SYSTEMS Texas Transportation Institute (TTI) developed a detection-control system (D-CS) that assumes a constant dilemma zone, defined as the time period between 2.5 and 5.5 s upstream of the intersection. The D-CS uses advance detection, at about 1,000 ft upstream of the intersection, to predict vehicle presence in the dilemma zone. D-CS implements a two-step gap-out strategy to reduce the probability of max-out, therefore reducing the probability of trapped vehicles in the dilemma zone. The D-CS provides dilemma-zone protection to vehicles at highspeed isolated signalized intersections. The system implements a twostep gap-out strategy to reduce the possibility of sudden green-phase termination by max-out. During the first step, the D-CS holds the green until the dilemma zone is clear. If the dilemma zone is not cleared after a certain time, the D-CS applies the second-step relaxed criterion, which allows the green to end when there is only one car in the dilemma zone (but not a truck). This two-stage operation somewhat mimics the operation of the LHOVRA and self-optimizing signal control (SOS) systems, developed in Sweden (6, 7). Compared 217

218 218 Transportation Research Record 2128 with the conventional systems with multiple advance detectors, D-CS requires fewer detectors per through lane (two versus three or more) and is much less likely to experience max-out according to theoretical analysis done by TTI researchers for the same maximum allowable headway and maximum green. The D-CS installation cost may exceed that of the system with multiple advance detectors because of the extra length of trenching and cabling. Field studies conducted by TTI concluded that the D-CS provides significant reductions in red light running compared with traditional advance detection (5). Although the D-CS appears to be developing in the right direction with the dual-stage approach, there is no theoretically based reasoning to determine the length of each stage so far. Moreover, variations in traffic states are not accounted for in the logic. Presented here is an approach that addresses these issues and can greatly improve the potential of dilemma-zone protection systems. RESEARCH OBJECTIVE AND SIGNIFICANCE The objective of this research is to develop an optimal policy for control decisions that accounts for changes in traffic states or patterns. The control agent (the controller) presented is capable of learning and adapting to the environment by using reinforcement learning (RL) techniques. Application of this concept can determine when a dilemma-zone protection algorithm should start accepting a certain number of vehicles in the dilemma zone and deciding whether it is optimal to end or extend the current phase at time t. This procedure allows the controller to adapt its policy on the basis of changes in traffic characteristics. In the following sections, insight into the RL concept and the Q-learning algorithm (QA) is provided. After an explanation of Markovian process model development, the simulation experiment, coding architecture, and setting of testing scenarios are presented. REINFORCEMENT LEARNING RL is a new area of machine learning that has achieved much success in recent years (10). RL tackles the problem of how an independent agent that senses and acts on its environment can learn to choose optimal actions to reach its long-term goals. The view of the exploration and exploitation dilemma colors the entire literature of the RL approach since the agent deliberately delays its expected reward by selecting actions with low expected rewards in order to learn more about the environment. Simply, RL is interesting in the trade-off between long-term reward and short-term reward. Crucially, this method is significant over the long term since it keeps learning from the observations, actions conducted, and rewards received. Unlike supervised or unsupervised learning, RL strives to learn the optimal policy by perceiving states of the environment and receiving nondeterministic rewards from the environment. This method considers learning to be a trial-and-error process. With no prior instructions to do so, the agent will converge to the optimal policy by using information of states, state transition probabilities, actions, and rewards received. Application of RL in Traffic Control RL seems to offer significant advantages for traffic signal control applications, in which real-time, adaptive, and responsive control are curtailed to enhance effectiveness, efficiency, and safety. Ou et al. (11) suggested a multiagent Bayesian learning procedure as the basic framework for traffic signal control. Each agent represents a single intersection in the network coordinates with other agents to reduce traffic congestion. Ma et al. (12) implemented a QA to control a traffic signal at an isolated intersection. By observing the queues and traffic flow trends, the controller chooses whether to end or extend the green phase. The received reward is indicated by throughput volumes. Thorpe (13) employed the SARSA RL algorithm in traffic signal control. A discrete states actions space was defined by the queue length, vehicle position in the queue, and elapsed time in the current phase. The negative rewards were assigned on the basis of stopline detector actuations. Abdulhai et al. (14) applied a QA to improve the performance of an isolated traffic signal control. They considered queue lengths and elapsed phase times as state parameters. A negative reward was given in terms of total delays incurred and the procedure was tested under variable demand. Wiering (15) applied multiagent RL to control large signalized networks. States were defined by nodes with vehicles waiting, destination of vehicles, and queues. The goal was to minimize total vehicle waiting time. Jacob and Abdulhai (16) used a QA to examine the integration of freeway and arterial traffic control. PARAMICS software was used to train and evaluate the RL agent offline within a simulated environment. Bingham (17) proposed combining RL with a neuro-fuzzy traffic signal control. RL was used as a learning algorithm for the neural network. The RL rule is to use the information that the supervised neural network s learning algorithms fail to handle. The system objective is to minimize the delay of vehicles constrained by a certain green extension. Learning Phase in RL In the RL approach there are two ways to scheme an agent s life: Consider the agent s life as sequences of observations, actions, and rewards (i.e., observation action reward observation action reward...); an optimal policy can be acquired online. Consider the agent s life as a history of experiences, that is, (s, a, r) 1, (s, a, r) 2, (s, a, r) 3,..., (s, a, r) n, where s is the state, a is the action, and r is the reward; the agent consumes a history of experiences in learning the optimal policy offline. Q-LEARNING ALGORITHM The QA is considered the most important breakthroughs in RL and was developed by Watkins in 1989 (11). The QA is the first direct adaptive optimal control method that does not need a model for a system. It maps state action pairs to Q-values without learning the model. It evaluates and learns from the visited states only. The algorithm can be implement offline (i.e., learning from the units of experience). The QA maintains an estimate of the Q-value of every state action pair at each time step and keeps updating the Q-values for the next steps. The updated Q-values are a mixture of the old Q-values and new guesses. The new guess is the immediate reward plus the discount of the best value of the next state. The QA has approved its convergence under certain conditions to an optimal policy.

219 Adam, Abbas, and Li 219 QA Anatomy The QA is a stochastic approximation version of Q-value iteration. It is defined as follows: Q Q + α r + γ.maxq 1 Q ) () 1 ( sa, ) ( sa, ) ( sa, ) ( sa, ) ( sa, ) where γ=discount factor, 0 γ 1, α=learning rate, 0 α 1, r (s,a) = reward or penalty received in the process of taking action a at state s, Q 1(s,a) = previous value of Q-estimate of optimal policy in the process of taking action a at state s, and Q (s,a) = current value of Q-estimate in the process of taking action a at state s. The Q-value iteration is a deterministic algorithm that converges by contraction. The reward r (s,a ) in Equation 1 is the stochastic term that transforms the deterministic Q-value iteration algorithm to its stochastic version (i.e., the QA). Since the reward is a real observation, it has a mean-zero noise. From applied mathematics (stochastic approximation), if a deterministic algorithm converges to a value and it has an equivalent stochastic version that adds mean-zero noise at each time step, under certain conditions the stochastic version will converge to a value with probability (10). QA fits this characterization exactly. The conditions that ensure QA convergence are α = and α 2 is a finite quantity. ( The perception of the first condition is that α (step size) must be a very small value, giving the algorithm enough time to learn. In other words, the algorithm will never stop learning. The second condition emphasizes that the noise (variance) has shrunk all the time, and it does not keep adding as the algorithm progresses. Tuning Learning Parameters Two control parameters in the QA affect the acceleration and stability of the convergence: the discount factor and the learning rate. The discount factor regulates the effect of the next state s optimal Q-value on the emerging policy. Thus, if the discount factor is small, it gives more weight to the current reward than the expected Q-value of the next state. Learning rate α quantifies the importance of the new learned information on the updated Q-value. If the learning rate is small, the learning process will be slow but steady. If the learning rate is set too high, the agent will act in an erratic fashion and information may be lost, resulting in some amount of instability in the learning process. Statement of Algorithm Initialize Q(s, a) arbitrary Repeat (for each episode): Initialize s Repeat (for each step of episode): Choose a from s using policy derived from Q (e.g., e-greedy) Take action a, observe s, R(s ) Q(s, a) Q(s, a) +α[r(s ) +γmax Q(s, a) Q(s, a)] S s Continue until s is terminal Modeling Dilemma-Zone Problem The state of the environment (a high-speed intersection) at any instant of time can be modeled as the total number of vehicles trapped in the dilemma zone if the green were to terminate at that time (t). Thus, the environment consists of nine states (0, 1, 2,..., 8, 8 representing 8 or more vehicles in the dilemma zone). In addition, one inherent variable is the time to max-out. This variable is included in the reward function, as will be described later. Figure 1 shows the state structure of the modeled environment. MARKOVIAN STATE ESTIMATION AND RL CONTROL DECISIONS Two major factors affect the total number of vehicles in the dilemma zone at any time (t): 1. The change in the entering traffic characteristics (i.e., the average and variance of the traffic volume entering the link) and 2. The traffic flow mechanism down the link (i.e., the carfollowing and lane-changing behaviors). Any change in one of these two factors will result in different transitioning probabilities from one state to another. In this context, the transitioning probability refers to the probability of changing from any state (state S is the number of vehicles in the dilemma zone; S = 0 to 8) to another. This transitioning probability can be expressed in a Markovian transitioning probability matrix (18). The Markovian matrix captures the structure of the traffic in the system and can gradually change to reflect any change in the traffic characteristics. An adaptive system that utilizes the Markovian state estimation should have a mechanism to update the Markovian matrix in real time as new data are obtained. The transition matrix is time independent, yet it is state dependent as Equations 2 and 3 show; Equation 4 shows that the target state is dependent only on the preceding state: { ( ) = ( ) = }= ( ) Pr N t x N t 1 y Pr x, y ( 2) { ( ) = ( ) = } ( ) Pr N t x N t 1 y P t, x, y ( 3) { ( ) ( ) = }= { ( ) ( ) = ( ) = } Pr N t N t 1 y Pr N t N t 1 y, N t 2 z =... ( 4)

220 220 Transportation Research Record 2128 Q-L Policy 100% Regret 400 vehicle / hour / lane 80% 60% 40% End Extend Q-L 20% 0% Vehicles in DZ Regret End Regret Extend Vehicles in DZ Mint Green Variable Green Reward FIGURE 1 Environment state structure of modeled dilemma-zone protection system (Q-L Q-learning; DZ dilemma zone). where x, y, z = vehicles trapped in dilemma zone, N(t) = targeted state, and N(t i) 0<i<t = ith preceding state. If a Markovian matrix that predicts the state probabilities in the next second is obtained, the matrix itself can be used to predict the state probabilities after n seconds by raising the Markovian matrix to power n. This process is also known as the Markovian chain process (18), which is essential to estimate the number of vehicles in the dilemma zone, say, 10 s from now, given that there are currently two vehicles in the dilemma zone. The intelligent RL agent should be able to use the Markovian state matrix, the existing number of vehicles in the dilemma zone, and the time into green (or time to max-out) to come up with an optimal decision of whether to accept the current number of vehicles in the dilemma zone and terminate the phase or to extend the phase in expectation of a lesser number of vehicles in the dilemma zone before the max-out. In fact, the agent should even know whether the max-out is actually as dreadful as it sounds, on the basis of the existing Markovian state matrix. Actions and Rewards The reward function is defined in Equations 5 through 7. The reward for ending the phase takes into account the total expected number of vehicles in the dilemma zone in the whole hour by dividing the total number of seconds in an hour by the cycle length if the major phase ended at time t. The reward for extending the phase takes into account the average expected value for equal or lower states from time t to G max. The rationale behind pre- ferring to end the phase later even when the same number of vehicles are trapped in the dilemma zone is the fact that this action will result in a longer cycle and therefore fewer vehicles in the dilemma zone over the whole hour. r r S end = extend = 3, 600 G + G min + t 2 1 3, 600 G + G min + t 2 1 S t S e G S G t M i t ( j SS ) 1 1Max Max e = = 0 t j e = G1max t () 5... ( 6) ( 7) where r end = reward (penalty) for ending phase, r extend = reward (penalty) for extending phase, G 2 = green duration for minor phases combined, G 1 min = minimum green for major phases, t = current time step in variable green, S t = traffic state at time t, S e = average Markovian traffic state if phase is extended, G 1 max = maximum green for major phases, and M St,S j = Markovian transitioning probability from state S t to state S j. The QA allows the agent to learn from experience. For example, the agent might wait in a certain cycle and extend the phase in expectation of a lesser number of vehicles in the dilemma zone, only to end up

221 Adam, Abbas, and Li 221 Observed State Actions Reward Learning Rate Q-Value Discount Rate Markovian Traffic State Estimation FIGURE 2 Q-learning training processes for single epoch. with a higher number of vehicles trapped during that cycle. The unit of experience will be used by the agent in the next cycle, and so on, as shown in Figure 2, where the agent senses the current state, takes an action, and observes the reward (r). The Markovian state matrix is multiplied by a discount rate (so as to emphasize experience versus new observation) and added to the old Q-values with a learning rate to come up with a new Q-value for each state action combination. The process continues until convergence occurs. At that point, a policy emerges for each given state for the given traffic structure. The policy can be adaptive as the learning is continually updated. This policy is referred to here as the optimal policy. Experimental Simulation Framework The experiment was conducted using VISSIM (19), a time-step behavioral-based microscopic traffic simulation model. VISSIM uses an extensive range of parameters that directly affect vehicle interactions. Instead of actively engaging in a wide range of live experiments, the training data that were acquired from a simulated environment can be passively used as the base knowledge for the agent s exploration phase. VISSIM was used to generate the units of experience needed to train the agent. It was run under the (.NET) framework through the VISSIM COM interface, which was used to obtain information at particular points in time. Different traffic volumes generated different arrival patterns at the intersection. Each data unit required an entire simulation run that included a warm-up time to ensure that the network was occupied, 15 s for minimum green, plus 40 s for variable green time. The successive runs were generated with exclusively different seed numbers. Vehicles were expected to be released with behavioral characteristics that are independent and identically distributed in order to eradicate any correlation or bias within the simulated data. The VISSIM COM integration with the RL framework is shown in Figure 3. The simulated network consists of an isolated intersection of a twolane arterial. Only through movements were modeled in the network. In order to illustrate the adaptability of the RL agent, three different volume scenarios were simulated: 400, 600, and 1,000 vehicles per hour per lane. Analysis and Discussion of Results The Markovian transitioning matrix obtained from the VISSIM simulation for the 400 veh/h/lane is shown in Table 1. This matrix was obtained by counting the frequency of each state occurrence immediately following the subject state and then dividing by the total. Figure 4 shows the S e -value used in the reward function for the extension decision starting from each second of the variable green. The graph shows that the S e -value converges to a single value (about 1.7 in this case) for longer times from max-out. A higher FIGURE 3 COM.NET Framework Rewards Units of experience States-action Data generation framework. VISSIM

222 222 Transportation Research Record 2128 TABLE 1 Markovian State Transitioning Matrix Based on 400 veh/h/lane Next State (vehicles in dilemma zone) Current State average volume will obviously correspond to a higher average value of S e. Figure 5 shows the convergence of the Q-values for each state. Each graph shows both the Q-values and the action probability progression through iterations. The agent bases its decision on the lower Q-value (i.e., higher probability). So for a given Markovian matrix, a policy can be defined either to extend or to end the phase if a given state is encountered. By taking a second look at the data and knowing the determined decision in each cycle, one can calculate the regret probability of each state. Regret is defined in this context as the possibility of finding a lower state before the max-out when the decision was to terminate or the possibility of ending up with a higher state when the decision was to continue. A policy is defined in the direction with the lower regret. So in Figure 6, the policy is to extend whenever there is one vehicle or more in the dilemma zone. Figure 6 shows that for such a low traffic volume and the Markovian matrix, the RL agent decides to extend the phase for States 1 or higher, knowing that there is a high probability of encountering the zero state before max-out. When the same procedure was followed for heavy traffic volume [1,000 veh/h/lane], the resulting policy (shown in Figure 7) changed since the RL agent preferred to end the phase whenever State 2 or less is encountered. Apparently, the agent knew from the Markovian matrix that encountering a higher state is more likely than not. COMPARISON OF PROPOSED APPROACH AND D-CS TWO-STAGE POLICY Figure 8 shows the evolving policy obtained from the RL agent s use of the Q-learning technique. The policy adapts to changes in volume, unlike arbitrary and stationary policies derived by using traditional approaches. Figure 9 shows the frequency of ending the cycle at each state out of 1,000 cycles used in the analysis. It can be seen that as the volume increases, the RL agent tends to accept one and two vehicles in the dilemma zone, whereas the traditional policy extends the phase and ends up with a max-out and ultimately more trapped vehicles in the dilemma zone. Finally, Figure 10 shows the total number of trapped vehicles by policy and volume. The RL agent was found to be constantly superior, with a reduction in trapped vehicles increasing as the volume increases. In the case of 1,000 veh/h/lane, the RL agent policy resulted in the reduction of vehicles trapped in the dilemma zone from 4,685 to 3,177 vehicles, or 32%. This large improvement was possible because the RL agent learned to accept up to two vehicles in the dilemma zone when there were lower probabilities of finding a lesser number of vehicles in the cycle. S e Time to max-out (sec) FIGURE 4 zone). Markovian chain of system states (vehicles in dilemma CONCLUSIONS AND FUTURE RESEARCH A novel approach is presented for developing an intelligent algorithm for dilemma-zone protection based on RL and Markovian traffic state estimation. The proposed algorithm can adapt to changes in traffic states and produce an optimal policy for phase termination or extension in real time. A comparison between the proposed optimal policy and the D-CS two-stage policy was conducted, and it was found that the RL-based policy reduced the number of vehicles caught in the dilemma zone by up to 32%. Future research should include other objective functions such as vehicle delay and stops in the learning mechanism. It is also recommended that a sensitivity analysis for the RL parameters be conducted.

223 Adam, Abbas, and Li 223 (a) (b) (c) (d) FIGURE 5 Evolution of Q-values for each state by number of vehicles in dilemma zone: (a) no vehicles, (b) one vehicle, (c) two vehicles, and (d ) three vehicles.

224 224 Transportation Research Record % 80% 60% 40% End Extend 100% 80% D-CS-2 stage Q-L 20% 0% FIGURE 6 volume Vehicles in DZ Regret End Regret Extend Policy derived from Q-learning for low traffic Probability 60% 40% 20% 0% Vehicles trapped in DZ (a) 100% 80% 60% 40% 20% End Extend Probability 60% 40% 20% D-CS-2 stage Q-L 0% Vehicles in DZ Regret End Regret Extend 0% Vehicles trapped in DZ (b) FIGURE 7 volume. Policy derived from Q-learning for heavy traffic 80% 60% D-CS-2 stage Q-L Traffic Volume (veh/hour) FIGURE 8 Q-learning. End Extend Vehicles in DZ Evolving policy derived from Probability 40% 20% 0% Vehicles trapped in DZ FIGURE 9 Frequency of ending phase at any given state by policy: (a) 400-veh/h/lane case, (b) 600-veh/h/lane case, and (c) 1,000-veh/h/lane case (Q-L Q-learning). (c)

225 Adam, Abbas, and Li DCS-2 stage Q-L Traffic volume veh/hour/lane 1000 FIGURE 10 Vehicles trapped in dilemma zone by different volumes. REFERENCES 1. Safe Car Guide htm. Accessed June World Report on Road Traffic Injury Prevention. World Health Organization; World Bank. publications/road_traffic/world_report/en/index.html. Accessed May National Automotive Sampling System (NASS): General Estimates System (GES): Analytical User s Manual NHTSA, U.S. Department of Transportation, 2001, p Taking Action to Reduce Intersection Fatalities. Safety Compass Newsletter, Vol. 1, No. 2, March April May dot.gov/newsletter/safetycompass/2007/mar07_v1/index.cfm#story1. Accessed June Bonneson, J., D. Middleton, K. Zimmerman, H. Charara, and M. M. Abbas. Intelligent Detection-Control System for Rural Signalized Intersections. Report FHWA/TX-03/ FHWA, U.S. Department of Transportation, Kronborg, P., F. Davidsson, and J. Edholm. SOS-Self Optimising Signal Control: Development and Field Trials of the SOS Algorithm for Self- Optimising Signal Control at Isolated Intersections. TFK Report 1997:2E. Transport Research Institute, Stockholm, Sweden, Kronborg, P., and F. Davidsson. MOVA and LHOVRA: Traffic Signal Control for Isolated Intersections. Traffic Engineering and Control, Vol. 34, No. 4, April 1993, pp Peterson, A., T. Bergh, and K. Steen. LHOVRA: A New Traffic Signal Control Strategy for Isolated Junctions. Traffic Engineering and Control, Vol. 27, No. 7/8, July/August 1986, pp Bullock, D. M., and T. Urbanik II. Traffic Signal Systems: Addressing Diverse Technologies and Complex User Needs. In Transportation in the New Millennium: State of the Art and Future Directions, TRB, National Research Council, Washington, D.C., Sutton, R. S., and A. G. Barto. Reinforcement Learning: An Introduction. MIT Press, Cambridge, Mass., Ou, H., W. Zhang, and X. Xu. Urban Intelligent Traffic Control System Based on Multi-Agent Technology. Acta Electronica Sinica, Vol. 28, No. 12, 2000, pp Ma, S., Ying Lee, and Leopard Liv. Agent-Based Learning Control Method for Urban Traffic Signal of Single Intersection. Journal of Systems Engineering, Vol. 17, No. 6, 2002, pp Thorpe, T. Vehicle Traffic Light Control Using SARSA. MS thesis. Colorado State University, Fort Collins, Abdulhai, B., R. Pringle, and G. Karakoulas. Reinforcement Learning for True Adaptive Traffic Signal Control. Journal of Transportation Engineering, ASCE, Vol. 129, No. 3, 2003, pp Wiering, M. Multi-Agent Reinforcement Learning for Traffic Light Control. In Machine Learning: Proceedings of the Seventeenth International Conference, Stanford, Calif., Morgan Kaufmann Publishers, San Francisco, Calif., 2000, pp Jacob, C., and B. Abdulhai. Integrated Traffic Corridor Control Using Machine Learning Proc., IEEE International Conference on Systems, Man and Cybernetics, Vol. 4, 2005, pp Bingham, E. Reinforcement Learning in Neuro-Fuzzy Traffic Signal Control. European Journal of Operational Research, Vol. 131, No. 2, June 2001, pp Stewart, W. J. Introduction to the Numerical Solution of Markov. Princeton University Press, Princeton, N.J., VisSim 4.30 User Manual. Planung Transport Verkehr AG, Karlsruhe, Germany, The Traffic Signal Systems Committee sponsored publication of this paper.