LOCATION SENSING AND PRIVACY IN A CONTEXT-AWARE COMPUTING ENVIRONMENT ASIM SMAILAGIC AND DAVID KOGAN, CARNEGIE MELLON UNIVERSITY

Transcription

1 C ONTEXT-AWARE C OMPUTING LOCATION SENSING AND PRIVACY IN A CONTEXT-AWARE COMPUTING ENVIRONMENT ASIM SMAILAGIC AND DAVID KOGAN, CARNEGIE MELLON UNIVERSITY Pervasive computing is an emerging field in computer systems research. This paradigm introduces smart spaces which interact with users in a natural and flexible manner. For many of these interactions to occur, the system must know the locations of users. ABSTRACT This article presents and evaluates the performance of a location sensing algorithm developed and demonstrated at Carnegie Mellon University. We compare our model with various others based on different architectures and software paradigms. We show comparative results in accuracy, the complexity of training, total power consumption, and suitability to users. Our method reduces training complexity by a factor of eight over previous algorithms, and yields noticeably better accuracy. The algorithm uses less power than previous models, and offers a more secure privacy model. INTRODUCTION Pervasive computing is an emerging field in computer systems research. This paradigm introduces smart spaces that interact with users in a natural and flexible manner. For many of these interactions to occur, the system must know the locations of users. Given the importance of position as a component of context-aware computing, our efforts have focused on developing a location service. Our implementation is based on signal strength and access point information from the IEEE 82.11b WaveLAN wireless network that covers the entire Carnegie Mellon University (CMU) campus [1]. This is in contrast to a Global Positioning System (GPS)-based approach, which has poor indoor accuracy and requires each client to be equipped with a GPS receiver that adds weight and consumes power. Our goal was to develop a power-efficient, easily trained, scalable, and secure location service. It had to be available to a variety of clients including wearable and laptop computers. Within the scope of the WaveLAN wireless network infrastructure, there are two fundamental algorithmic approaches: client-centric and server-centric. In the client-centric model, each client obtains the signal strengths to multiple WaveLAN access points, and that set of data is used to calculate a location. In the server-centric approach, a centralized server determines which clients are in contact with which access points, and uses this data to calculate position. In this article, we discuss a client-centric triangulationbased remapped interpolated approach (CMU- TMI). We evaluate this approach in comparison to our previous client-centric algorithm based on pattern matching (), Microsoft s RADAR system [2], a server-centric approach (CMU-SC), and statistical approaches (e.g., in [3, 4]. We also relate these WaveLAN-dependent systems to proprietary infrastructure-based systems such as GPS [5], radio frequency (RF), and ultrasound [6] systems based on cellular technology [7, 8], and proprietary non-infrastructure-based systems such as video-based systems utilizing pattern matching techniques as described in [9, 1]. We outline the relative merits, and focus specifically on comparing the WaveLAN-based systems in accuracy, complexity, training requirements, power capacity, and the inherent acceptability to end users. PERVASIVE ENVIRONMENT MODEL The location service fits in the server layer of our model for a pervasive computing environment. This architecture, called Handy Andy, is shown in Fig. 1. The bottom of the figure has a range of nonhomogeneous mobile and fixed devices. The second and third layers translate requests coming from devices and associate them with unique users and user states. The fourth layer represents shared applications, utilities, and servers, and is the layer where a location service would exist. The position of the location service in the system makes it clear that a device could request its own position or someone else s position. This is integral to the system s functionality, since it means that location information has to be available to people other than the user. This means that the service must scale well to large numbers of users and large areas, and must be secure to prevent users from misusing the information it contains. An example application of a location service is Portable Help Desk (PHD), a context-aware application prototyped at CMU. This service allows a user to determine the location of other users on campus as well as information about them. This data is used within a class environment for the purposes of arranging meetings, finding people for help or discussion, and finding resources such as printers. Only a partial list of users and resources is available at any given time, since users may not wish to be visible at all times, /2/$ IEEE IEEE Wireless Communications October 22

2 Database Service User proxy Device proxy Device Infrastructure login/logout Device proxy Speech encode/decode Itsy Waldo Figure 1. Handy Andy pervasive computing architecture. Ph.D. User proxy Device proxy Handhelds Virtual whiteboard Device proxy Other devices Location Most applications for position tracking indoors deal with directions as applicable to a person on foot. This means several meters of error is acceptable, but anything beyond the scope of a single room is a problem. and devices may not all be available. Figure 2 illustrates a visual user interface for the PHD application. People and resources are selected in the left pane, the results of the queries are presented in the middle pane, and locations of people and resources are displayed in the right pane. It is clear that the PHD would not be possible without a robust location service. LOCATION SENSING In this section we summarize aspects of previous implementation of location sensing, and go into details of Carnegie Mellon s Triangulation, Mapping and Interpolation algorithm (CMU-TMI). ARCHITECTURES Our choice of architecture stems from the resources available and the requirements of the service. We considered a variety of architectures that can be used for location sensing. GPS-based [5] systems are the most well known. For our purposes, these have two downsides: they require each user to carry specialized hardware, and they are almost useless indoors, particularly in the vertical axis. Since we wish the service to be easily available, and need accuracy within a few meters in a campus environment, this system is unacceptable for our use. Most applications for position tracking indoors deal with directions as applicable to a person on foot. This means several meters of error is acceptable, but anything beyond the scope of a single room is a problem. Proprietary systems based on RF and ultrasound [6] have been implemented in business environments. These have good indoor accuracy, but require a great deal of investment in infrastructure, and special hardware has to be worn by all users. Due to the constraint of making the location server easily available to a large number of users on the CMU campus, this also was not acceptable. Video-based systems utilizing pattern matching techniques [9, 1] require that the user be equipped with a wearable camera. These systems require a great deal of training and suffer from lack of accuracy. Systems based on the inverse approach, identifying users by stationary video cameras, are advantageous because they do not require any hardware to be worn by users. However, this needs a prohibitive investment proportional to both the area covered and accuracy required, since a camera has to be at any location where a user can be tracked. These systems require a great deal of training on each user, so new users are difficult to add, and the accuracy is unclear when dealing with a large population. Systems based on cell phone signals [7, 8] have downsides in accuracy. While cell phones are widely used throughout the world, and the signals provide fair accuracy outdoors and in Figure 2. Portable help desk screen shot. IEEE Wireless Communications October 22 11

3 Frequency Mean: 75.6 Median: 75 Mode: 74 StdDev: Signal strength (dbm) Figure 3. Stationary signal strength distribution certain cases good quality indoors, they suffer from the reflection of the RF signal deep within buildings, and tend to have poor vertical resolution, not being able to distinguish between floors. These systems can be used to complement an based location service. Our ultimate choice was an architecture based on CMU s wireless network. This provides a functional client-server relationship between users with portable devices using WaveLAN network cards and Lucent WaveLAN access points. Anyone using the wireless network has the necessary hardware, so no investment is necessary beyond the development of software. Because the network information is available within the client devices and accessible at the access points, there is no need for contracts with external systems. Furthermore, high levels of accuracy can be achieved due to the density of access points within the campus. SOFTWARE MODEL Several algorithms have been implemented to take advantage of wireless network data, and our algorithm builds on this previous work. The simplest system is a sever-centric approach prototyped at Carnegie Mellon (CMU- SC). This approach gathers data from the wireless access points and returns the position of the access point to which the user is connected. The benefit is the simplicity of the algorithm and independence from the user. The only data required is a mapping of access point addresses to physical access point positions. The downside is that the accuracy is directly proportional to the density of access points, which results in a resolution of approximately 25 m, which is the average range of a wireless access point. A second, less obvious, downside is that because the data is independent of the user, the security involved is poor, since anyone could have access to this part of the network. A final issue concerns implementation. The Lucent WaveLAN access points on the CMU campus cache the medium access control (MAC) addresses of the client s radios. This cache is persistent, even after the client has left the wireless network. This ghosting of MAC addresses prevents accurate location information from being obtained directly from the access points. Two algorithms based on pattern matching have been developed for a greater degree of accuracy. The RADAR [2] system and CMU s Pattern-Matching algorithm () use multiple signal strengths to determine a user s position. Each algorithm gathers the strengths of signals between a user and the wireless access points in the user s vicinity. RADAR gathers this data from the access points, while CMU- PM has the clients gather the data. This data is then matched against a pregenerated table of signal strength vectors mapped onto positions, and the closest match is returned. The advantage of the is that given a sufficiently dense table of patterns, it is possible to pinpoint a user s position to within 1.52 m in the majority of the measurements. The disadvantage is the large amount of training necessary. The algorithms return the trained position that best matches the current signal strength vector, and due to their structure, interpolation is ineffective. So in order to get accuracy within 2 m, there must be trained positions every 2 m throughout the area in which location determination is to be performed. The two algorithms as implemented have different training point density and result in different accuracies, but are quite similar otherwise. Due to its serverside nature, RADAR may not scale as well to large numbers of users, and does not give users complete control over their privacy. When designing our new algorithm, CMU- TMI, within the wireless infrastructure, we had two specific goals. The algorithm had to be accurate on par with the pattern-matching algorithms, which meant accuracy within a few meters. It also had to be scalable, such that there did not have to be a trained data point for every returned value. In other words, for an area A with N training points, the estimate has to be closer than (A/(N p)) with a large probability p. That means there must be interpolation between trained values. In order to achieve these goals, we went with a different model of the world than the previous algorithms. We perform two distinct transformations on the raw data of signal strengths. We first calculate the client s position on a continuous coordinate grid, assuming signal strengths map directly to distance. We then map the resulting coordinates onto real space coordinates using a set of trained values. Because both transformations are continuous, the result is interpolative on the trained data, and we achieve accuracy much greater than that of the area divided by the number of trained values. For the sake of clarity, we call the domain of the initial mapping Signal Space, and the range of the final mapping Physical Space. The definition of Signal Space is a mapping of positions such that the distance between any two points is correlated with the signal strength between those points. Although Signal Space is discontinuous, triangulations based on it are quite accurate. A more detailed discussion of the algorithm follows. 12 IEEE Wireless Communications October 22

4 Figure 4. CMU-TMI possible locations for all access points. CMU-TMI The CMU-TMI system requires an investment of initial data. The physical position of all the access points in the area needs to be known, much like what was needed for CMU-SC. In order to generate the Signal Space positions, a function mapping signal strength to distance has to be generated. This function was calculated empirically from observations. Finally, to map the Signal Space positions onto Physical Space positions, a set of trained points have to be generated. These trained points are nothing more than offset vectors from Signal Space positions onto Physical Space positions, and are calculated by performing the first two steps of the algorithm at a known location and recording that location. The algorithm can be divided into four distinct steps. The first two steps are identical for training the system and tracking a device s position. The third step is different for training and tracking, and the fourth step is only performed when tracking a user. Step 1 (Training or Locating): Scanning The client device initiates this stage. The device scans all the access points within range to determine their signal and noise levels. The algorithm for scanning uses previous well-developed work, accessing the wireless network card hardware on the device. Five scans are performed and averaged to reduce the variance in signal strength due to noise. The number of points within range varies depending on location, but three are generally sufficient for calculating a position, and four to six are commonly available. It is simple to determine the building and building floor from the strongest signal because signals travel poorly between buildings and floors. From there, the distance to three access points is needed to pinpoint a position in two dimensions. The signal strength data can be sent to a central server for processing. In the prototype system, the rest of the calculations are performed locally on the client. Step 2 (Training or Locating): Triangulation The signal strengths are used to infer the distance between Figure 5. CMU-TMI possible locations for strong access points. the client and the access points. Although it is possible to use a principled function to approximate distance [11], this requires considerable data on the building, and it was simpler in our case to use an empirical approach. Prior to implementing the algorithm, several dozen line-of-sight measurements of signal strength were taken to generate the approximate relationship between signal strength and distance. A polynomial function turned out to approximate the data quite well. The following equation, where d = distance in meters and s = signal strength in decibels, describes the relation: d =.163 s2 2.3 s + 8. Circular contours are generated around each access point at the calculated distances, and intersections between contours are found. Because of the inconsistencies of Signal Spaces and the effects of noise, contours of two access points do not necessarily intersect. Nonintersecting contours are grown or shrunk to estimate an intersection, with the change in size linearly proportional to the initial estimated distance, so the contours from nearby points are judged to be more accurate than those from distant points. Noise and the nonlinearity of Signal Space cause the intersections to be fairly scattered. In general, the number of calculated points is order n 2 the number of visible access points. These points are usually clustered, and are averaged for the user s position in Signal Space. It is at this stage that the amount of noise involved in the RF-signal-based system becomes evident. The data generated from a single scan varies a great deal over time, with the outliers particularly widespread. Figure 3 illustrates the distribution of signal strength from a client to a single access point. Figure 4 illustrates the large distribution of calculated possible positions taken from a single location over several minutes. Steps are taken to minimize the noise in this data. Measurements based on weak signal strengths from distant access points are filtered out. Using only the five strongest signals results in a distribution with far fewer outliers. Figure 5 shows the cleaner distribution of points when the distant weak access points are removed. When designing our new algorithm, CMU-TMI, within the wireless infrastructure, we had two specific goals. The algorithm had to be accurate on par with the patternmatching algorithms and had to be scalable. IEEE Wireless Communications October 22 13

5 The idea of Signal Space is just a way of performing fast interpolation. One could achieve a similar result by dividing a map into a mesh with the trained positions, and then interpolating between these positions within the scope of a few of the nearest trained values. This is the final step performed for training. Averaging signal strength over five scans in the scanning stage and then averaging the resulting 5 1 position points results in fairly good accuracy. Noise is discussed further a later section. Step 3 (Training): Training If the previous two steps were executed for training, the next step is to train the correlation between Signal Space and Physical Space. It is important to understand the distinction in Signal Space, the distance between points is correlated with the signal attenuation between the points. Physical Space is the actual map of the area. When we generate the user s position as described earlier, we do so assuming that signal strength can be converted directly to distance. This assumption is clearly wrong, because solid matter absorbs and reflects RF signals, making the previous distance calculations significantly wrong. However, it is reasonable to assume that points significantly distinct in Physical Space will likewise be distinct in Signal Space. That is, a device will return a position in Signal Space that is different from a device at another physical location. Since changing a position in Physical Space changes the value of the position in Signal Space, it is possible to map the positions in Signal Space onto positions in Physical Space. In order to keep the advantage of interpolation, this mapping has to be continuous, and because the algorithm will be executed often, the mapping must be fast. We therefore chose the mapping to be a linear distortion of the space, such that Signal Space is stretched at distinct points, much like a rubber sheet, to generate the positions in Physical Space. The mapping is trained by using Signal Space positions calculated at known positions and then setting those to be anchor points for the distortion. The idea of Signal Space is just a way of performing fast interpolation. One could achieve a similar result by dividing a map into a mesh with the trained positions, and then interpolating between these positions within the scope of a few of the nearest trained values. This is the final step performed for training. Step 3 (Locating): Mapping When the algorithm is used to determine position, applying the mapping function is quite simple. The nearest set of trained mappings from Signal Space to Physical Space are found. An average of the offsets of the trained values is applied to the Signal Space positions to calculate the Physical Space positions. The average is weighed linearly on proximity, and only those trained values within twice the distance of the nearest trained value are applied. Step 4 (Locating): Smoothing Noise and the finite granularity of a mapped physical location causes the resulting position to jump with consecutive scans, regardless of whether the device is moving or not. To alleviate this jitter and to minimize the error caused by a single poor calculation, the results of previous calculations are averaged with the new calculation. This time-averaged location converges slowly toward new data. The speed of the convergence is linear with the distance between the smoothed value and the new calculated point. That way, motion is minimized while the device is still, but permits the value to react quickly to actual motion. Figure 6 shows sample output from this algorithm. The leftmost point is the Signal Space location. The rightmost point is the computed Physical Location. The middle point represents the time averaged smoothed location, which is almost still because the computed value is very close to the old value. In this example, the time averaged smoothed location is indist inguishable from the actual location of the user on this scale. RESULTS AND EVALUATION Different location sensing techniques each have distinct advantages; likewise, they have their weaknesses. We briefly mentioned the problems with using alternative infrastructures earlier, so here we focus on comparing CMU-TMI against, CMU-SC, and RADAR, which all work within the scope of a wireless network. Accuracy is very important, but we also discuss training complexity, power consumption, and usability, which have not been discussed in detail in previous publications. The results were obtained by traversing an area approximately 5 m 2 in size inside a building and automatically gathering signal strength at regular intervals. Approximately 35 data points were gathered throughout the area. This data was then split into training and testing sets. The algorithm was trained on the training set and tested on the testing set to calculate accuracy. The size of the training set is reflected in Fig. 8. ACCURACY The accuracy of the methods is measured by the probability of a measurement being incorrect by more than a given threshold, measured over 35 data points. Figure 7 shows this cumulative distribution function vs. accuracy for the four methods discussed. The accuracy of the nearest base station approach used by CMU-SC is significantly lower than any other method, since it returns the nearest access point, which may be as far as 25 m away from the user. The accuracy of more precise methods, which use signal strengths for position determination, is greatly influenced by the inherent noise in the system. As discussed in previous publications [1], wireless networks within buildings are inherently noisy, since the RF signal reflects and is attenuated by moving objects, and is influenced by minor dynamic changes in the environment. Data gathered at a single position varies commonly by 5 percent in the strength of the signal in decibels, and signal strength is completely erratic when the direction of the wireless network card is changed. The goal of systems such as, RADAR, and CMU-TMI is to give accurate location data despite the noise, which results in the need for multiple scans, caching and smoothing. It is important to note how the systems handle noise. The pattern matching approaches of CMU- PM and RADAR result in accuracy correlated approximately with the density of their trained values. The data for appears biased toward positions near the trained values, which 14 IEEE Wireless Communications October 22

6 A A 121D Figure 6. CMU-TMI location output. results in the sharp slope at small distances. These approaches give good results near the trained values. When the result returned by the two pattern matching algorithms is corrupted by noise, the error is often quite large, since an entirely different trained value returns. This vulnerability to noise is one large downside to pattern matching, and there is a significant probability of large error. For both pattern matching algorithms, 1 percent of the sample data was incorrect by more than 9 m, a significant distance when trying to pinpoint a user. The CMU-TMI algorithm generates results between the two pattern matching accuracies for low distances, but generates better results for errors greater than 2 m. The greater accuracy at high distances is due to the way the different algorithms make errors. When or RADAR make an error, it is generally large, since an incorrect training point is returned. CMU-TMI s error, however, is smooth, so while there is a good chance of the returned position being slightly incorrect, it is very unlikely that the returned position will be far from the actual location of the user. In the experiments performed, only 5 percent of the sample readings were incorrect by more than 5 m, and none of the readings were incorrect by more than 9 m. TRAINING COMPLEXITY The training complexity is important when it comes to implementing the system. Someone has to actually go through the area being scanned to take readings and map them onto corresponding physical locations. Furthermore, because the RF signal strengths can be influenced greatly by small changes in the environment such as minor remodeling or furniture rearrangement, this mapping task would have to be repeated on a regular basis to maintain accuracy. If the number of readings to be taken is overly large, this task can be daunting. Algorithms based on pattern matching such as RADAR and require a great deal of training, since data must be sampled in a regular covering pattern throughout the area where Cumulative probability CMU-TMI Error distance (m) Figure 7. Cumulative distribution function (CDF) accuracy of location sensing algorithms. the location sensing is to be performed. The density of the training points places an upper bound on the accuracy of the location service. RADAR has a training point approximately every 23 m 2, as extrapolated from the work done by Microsoft on a limited scale. has a higher density, with a point approximately every 1 m 2. The accuracies of the algorithms are bounded by those numbers, since there is no way to pinpoint a user s position to more than an area of that size. CMU-TMI relies on interpolation for the accuracy of its results, so even a relatively sparse distribution of training points generates good results. To generate the values in Fig. 7, about one training point every 19 m 2 was used. That means that there can be a distance of 14 m between every two training positions, but because the user s position is interpolated within that space, more than 5 percent of the results are accurate to within 2 m. Figure 8 shows the number of training samples used to achieve the accuracies shown in Fig. 7. The plot shows required training points for a given area, with the estimated size of the CMU 3 RADAR Nearest base station IEEE Wireless Communications October 22 15

7 Required training points 3, 25, 2, 15, 1, 5 RADAR CMU-TMI Nearest base station RADAR 12, training CMU academic building approximate area (28, m 2 ) 28, training points Target location dimension (sqrt(area)) Figure 8. The training complexity of the algorithms. CMU-TMI 15 training points campus highlighted. CMU-TMI would require about 15 points to cover the campus, while RADAR would require 12,, and would need 28, points to achieve similar accuracy. Although each algorithm could use additional points to improve accuracy, the amount of accuracy per concentration of data varies. Figure 8 only shows the number of points needed to achieve the accuracy in Fig. 7. Different densities of training would change the results. POWER CONSUMPTION Location sensing is primarily useful for mobile users, who have significant power constraints. The amount of power consumed by a given location algorithm affects its acceptability, since users are not likely to take advantage of a service that quickly renders their device unusable. The battery life for an HP Jornada 68 handheld computer was evaluated while running the and CMU-TMI location sensing algorithms. These algorithms force the wireless network card on the device to perform rapid scanning of nearby access points, which is a drain on the battery. For our experiment, the device was fully charged, and run until fully discharged. Location calculations were performed once per second. Figure 9 shows the results of this testing. The control group, performing no scans, ran for almost 4 min. averages 1 scans per calculation, and decreases battery life by 75 percent. The interpolation of CMU-TMI, and the smoothing performed in its final step, is more robust to noise, so only 5 scans are done per calculation to achieve the accuracy in Fig. 7. As a result, CMU-TMI has less of an impact, decreasing battery life by 6 percent. It is worthy of note that the reason for the great impact is the frequency of calculation. Changing the scanning rate from once per second to once per 1 s reduces battery life impact to less than 6 percent for CMU-TMI and less than 8 percent for, although this reduces the information known about the user. We did not have access to the RADAR algorithm to perform power tests, but since the data is gathered from the access points rather than the network card, we assume the power consumption was less than or CMU-TMI. However, some response from the network card still needs to take place, so RADAR would consume more energy than the control. Naturally, a device using CMU-SC, which does not force additional scans, would last as long as the control group. USABILITY AND PRIVACY There are several usability issues integral to a location service. The architectural side was discussed earlier, so here we are interested in the human side of the system. A usability study was conducted at Carnegie Mellon to determine what users desired in a system, and it was found that in addition to accuracy, privacy was a great concern. For the most part, users wished to have complete control over the visibility of their location. Although a simple set of Boolean rules is sufficient from a software side to provide privacy, the architecture of the system has to be inherently secure. The advantage to CMU-TMI and is that the gathering of data occurs on the client machine. This provides users the greatest amount of control over their personal information, since the user can choose whether their location is transmitted to the server for others to access. The RADAR system provides the possibility of security, but does not give users direct control over their data. The information is gathered from access points, so any influence a user has over the availability of his or her information must be done through remote accesses to the central server. Finally, in the current state of the implementation, CMU-SC provides no privacy at all, since its data is freely available to anyone who knows the locations of the access points and cares to access the appropriate protocols over the network. This demonstrates a wide gap in the security of the system. In the general case, there is no guarantee that remote systems cannot access private systems, and the only absolute way to ensure invisibility is to make certain one s network card is powered down. However, controlling the data sent by the client, as in the case of CMU-TMI, provides more immediate control than having a server-based system. SUMMARY Each method has its strengths and weaknesses. The nearest base station approach used by CMU-SC is the simplest to implement, requires no field training, and does not drain the target device s batteries substantially. However, it provides only very coarse location results, has implementation issues, and provides little to no privacy. Pattern-matching-based systems such as RADAR and provide good accuracy, but require a great deal of training data. RADAR has less of an impact on battery power than, but is potentially not as secure, and may not scale well. Our latest algorithm, CMU-TMI, reduces training complexity over RADAR by a factor of eight, and generates more accurate results. It also does not drain a mobile device s battery as much as, and offers a secure pri- 16 IEEE Wireless Communications October 22

8 vacy model. Its advantages make it much more suitable to implementation on a large scale than previous algorithms, without sacrificing accuracy. FUTURE RESEARCH Context information is useful for generating more intelligent behavior in systems, but providing this information is greatly taxing on the system s architecture. Location information needs to be scalable, accurate, and yet secure and private. While systems such as CMU-TMI are already highly precise, these systems need to be tested on a wide scale, and, if at all possible, the amount of training required should be reduced further, since it is still quite significant. In addition, it is yet to be determined exactly how well such a system will be accepted by end users, and how easy it will be to integrate into existing systems. More information on the theory and practice of privacy systems needs to be researched as well. Another important piece of research is an assessment of the data needed to retrain the system. We have not presented results on this within this article, but a method for automatically gathering data to simplify retraining is being developed at CMU using a dead reckoning device [12]. At this time, however, retraining the system would require adding data points as in the training step. It is likely that not as much data would be needed for retraining, and a function could be used to age out old data. CONCLUSIONS CMU-TMI Control Runtime (min) Figure 9. Power consumption of CMU algorithms. [8] [9] B. Clarkson and A. Pentland, Recognizing User Context via Wearable Sensors, Proc. 4th Int l. Symp. Wearable Comp., Atlanta, GA, 2. [1] W. Rungsarityotin and T. E. Starner, Finding Location Using Omnidirectional Video on a Wearable Computing Platform, Proc. 4th Int l. Symp. Wearable Comp., Atlanta, GA, 2. [11] Siedel and Rappaport, 914 MHz Path Loss Prediction Model for Indoor Wireless Communications in Multifloored Buildings, [12] A. Hills, Large-scale Wireless LAN Design, IEEE Commun. Mag., vol. 39, no. 11, Nov. 21, pp ADDITIONAL READING [1] J. Small, Location Determination in a Wireless LAN Infrastructure, Master s thesis, Dept. of Elec. and Comp. Eng., CMU, This article presents and evaluates CMU-TMI, a new model of a location service, and compares its implementation against other existing algorithms. We introduce attributes for characterizing location sensing systems: accuracy, complexity of training system, power consumption, and usability. Our results provide guidelines as to under which circumstances one location sensing algorithm may be used in preference to another. Our method reduces training complexity by a factor of eight, and yields noticeably better accuracy. ACKNOWLEDGMENTS We would like to acknowledge the support provided by IBM Research, Hewlett Packard, Lucent, the National Science Foundation, Pennsylvania Infrastructure Technology Alliance, and the Defense Advanced Research Projects Agency. REFERENCES [1] A. Smailagic and D. P. Siewiorek, User-Centered Interdisciplinary Design of Wearable Computers, ACM Mobile Comp. and Commun. Rev., vol. 3, no. 3, 1999, pp [2] P. Bahl and V. N. Padmanabhan, RADAR: An In-Building RF-Based User Location and Tracking System, Proc. IEEE INFOCOM 2, Tel Aviv, Israel, Mar. 2. [3] P. Castro, T. Kremenek, and R. Muntz, Nibble: A Component-based Location System, IEEE Comp., vol. 34, no. 6, 21. [4] T. Roos, P. Myllymaki, and M. Tirri, A Statistical Modeling Approach to Location Estimation, IEEE Trans. Mobile Comp., vol. 1, no. 1, 22. [5] [6] [7] S. Tekinay, Wireless Geolocation Systems and Services, Special Issue IEEE Commun. Mag., Apr BIOGRAPHIES ASIM SMAILAGIC (asim@cs.cmu.edu) is a faculty member at Carnegie Mellon University s Institute for Complex Engineered Systems, College of Engineering, and Department of Electrical and Computer Engineering. He is also director of the Laboratory for Interactive Computer Systems at CMU. This Lab has designed and constructed 24 generations of novel mobile/wearable computer systems over the last decade. He received the Fulbright Postdoctoral award at Carnegie Mellon in Computer Science in He was a visiting professor at Cambridge University and University of Leeds. He has been Program Chair of six IEEE conferences. Also, he has been Associate Editor and Guest Editor in leading archival journals, such as IEEE Transactions on Mobile Computing, IEEE Transactions on VLSI Systems, IEEE Transactions on Computers, Journal on VLSI Signal Processing, and Journal of Computing and Information Technology. He was a recipient of the 1992 Tempus European Community Award for scientific cooperation resulting in new curriculum development, and the 2 Allen Newell Award for Research Excellence from Carnegie Mellon s School of Computer Science. He has written or edited several books in the areas of mobile computing, computer systems design and prototyping, and VLSI system design. He made major contributions to several projects that represent milestones in the evolution of advanced computer systems: from the CMU s Cm*Multiprocessor System and Edinburgh Multi-Microprocessor Assembly (EMMA), until the CMU s current projects on Wearable Computer Systems, Smart Modules, Communicator, Wireless LAN Location Sensing, and Aura Pervasive Computing. His research interests include pervasive and wearable computing, wireless communication, advanced computer architectures, and rapid prototyping. DAVID KOGAN is a software engineer at Oracle Corporation. He received his B.S. and M.S. in electrical and computer engineering from Carnegie Mellon. While at CMU, he participated in the Aura Pervasive Computing project. His current interests are in advanced algorithms and artificial intelligence. He is a member of the Sigma chapter of Eta Kappa Nu. IEEE Wireless Communications October 22 17