Bluetooth Positioning Optimization by Genetic Algorithm

Transcription

1 Bluetooth Positioning Optimization by Genetic Algorithm ALESSANDRO GENCO Dipartimento di Ingegneria Informatica Università degli Studi di Palermo Viale delle Scienze, Palermo ITALY Abstract: This paper discusses a strategy to optimize the disposition of Bluetooth base stations in a subset of locations which are compliant with site characteristics and constraints. To this end we propose a genetic algorithm and a layout chromosome to evolve up to a minimal gene configuration. This allows us to deploy a minimal number of Bluetooth devices, as well as to achieve an accuracy degree as high as possible in evaluating people position. As for case studies, some experiments and results are shown dealing with a castle in Sicily where we carried out many trials. Keywords: Bluetooth, Positioning, Context Aware, Pervasive System, Ubiquitous Computing 1 Introduction One of the main goals of pervasive systems is to provide context aware services which take into account the dependence of what to do on context elements such as: who, why, where, when. Some elements typically deal with semantics and allow a system to arrange services according to what someone may expect in a given reality [1], [2], [3]. Other elements can help in better instantiating a context by time and space information. Time information is often quite simple to get; differently, the position of people who are the receivers of context aware services, is almost unpredictable and difficult to detect. Positioning can be performed by means of a wide variety of technologies suitably installed in the environment. Depending on specific technology, some additional devices may need to be carried by people to be localized. Bluetooth is a wireless communication technology which is also implemented in cellular phone for data interchange, and therefore, it can be considered somewhat wearable and very easy to own. Bluetooth is currently investigated as a positioning technology in research laboratories by means of home made Bluetooth devices which can provide optimal performances. Nevertheless, at present, they are not available as standard accessories in either cellular phones, or PDA, or personal computers. Hereafter we discuss our approach to the positioning problem [4] [5] and propose solutions for making low cost Bluetooth devices effective for high definition positioning. 2 Positioning Technologies and Methods Existing positioning systems use different technologies and methods in order to locate a mobile device either indoor or outdoor. GPS [6] is the most famous system for positioning, undoubtedly. Nevertheless, GPS is based on radio time-of-flight lateration from satellite sources, and therefore, it can be used just for outdoor applications. As far as indoor positioning is concerned, there are many systems. Among these, ActiveBats [7] and Cricket [8] use ultrasound timeof-flight lateration, and provide an accuracy of a few centimeters. SpotOn [9] and MSR RADAR [10] are based on the Received Signal Strength Information (RSSI) of ad-hoc Bluetooth or frameworks, thus giving an accuracy of 3-4 meters. The Ascension Technology MotionStar system [11] uses magnetic sensors moving in a magnetic field around a source. This system provides very high accuracy, but it requires very expensive hardware as much again. As far as methods are concerned, AOA (Angle Of Arrival) is based on the direction of received signals. Location is estimated by triangulation when two reference devices at least measure the signal angle of arrival from a mobile device [12]. The Time Of Arrival (TOA) method is based on delay measurement. The roundtrip time (RTT) of the signal between mobile and reference device is measured, thus leading to a circle whose centre is the reference device and radius is half the RTT. TOA requires accurate clocks, because 1 µs error in timing leads to 300 mt error in distance estimation.

2 The Cell Identity (CI) method treats a network as divided into cells, each cell being the radio coverage of one reference device. A mobile device connected to a given reference device is assumed to be inside its cell; therefore, accuracy can be improved, either by the option of reducing cell size, or by cell overlapping and connectivity-induced geometric constraints as in [12]. The CI method only needs to know whether a mobile device is in the range of a given reference node or not. The RX power level positioning method is quite similar to TOA positioning in that both methods locate mobile devices on the intersection of three or more circles. In the case of RX power level positioning, circle radius is evaluated on the measured strength of received signals, thus assuming a relationship between signal strength and distance. 2.1 Bluetooth Positioning Hallberg et al. [13] developed two different methods based on Bluetooth (BT) RSSI values: a direct method, which requires a BT device to be programmed, and an indirect method, with no programming required. The direct method gives good accuracy, but it needs some programmable hardware. The indirect method is cheaper, but its accuracy is very poor with a worst-case error of 10 meters. The BT Local Positioning Application (BLPA) [14] uses RSSI to feed an extended Kalman filter for distance estimation. A good accuracy is achieved only assuming theoretical RSSI values; actual experimental values give unreliable distance estimation. The Bluetooth Indoor Positioning System (BIPS) [15] is designed for tracking mobile devices in a building. BIPS mainly deals with timing and device discovering, thus obtaining an accuracy of 10 meters. Michael Spratt [16] proposed the Positioning by Diffusion method which works on information transferred across short-range wireless links. Distance estimation is achieved by geometric or numeric calculations. 3 Experiments on BT Positioning Most of the existing approaches to BT positioning agree on the unreliability of RSSI values for distance estimation. RSSI can only be useful if no high accuracy is required. Actually, BT devices give two kind of measures, namely RSSI and Link Quality (LQ), which can be read through HCI (Host Controller Interface) [17]. LQ is a measure in a range from 0 to 255 of a link between two nodes. We found that LQ is a quite reliable parameter for distance estimation; RSSI only allows us to know whether the device is in a given BT base station (hereafter BTBS) power range or not. The use of LQ is recommended by BT standard specifications [17], and therefore LQ is available on most commercial devices and can be used in position estimation frameworks. A correlation can be assumed as a linear or a logarithmic regression between LQ values and distances up to 30 meters. We used three different positioning methods to translate LQ data into space coordinates. Namely, we used triangulation, fuzzy logic and neural networks. 3.1 Positioning by Triangulation Distances estimated by linear or logarithmic regression can be processed by simple trigonometric laws. For instance a 2D triangulation algorithm can be used to locate a mobile device by processing the distances from three different BTBS whose coordinates are known. We assumed in our early trials that each LQ value draws a perfect circle around a BTBS, thus allowing a triangulation to be used. First we selected a suitable initial arrangement of the BTBS with a minimal distance not less than 10 meters between them. Then we carried out our experiments and a lot of measurements to improve the initial arrangement. Finally, we achieved our best triangulation results with a radius average error in a range from 1.9 to 3.7 meters. 3.2 Positioning by Fuzzy Logic Unfortunately, LQ values are affected by high degree of uncertainty. Actually, we can draw an iso-lq curve whose points give the same LQ value from a base station. However, this is a cloud-shaped region around a BTBS and not a circle. Under this condition, geometrical triangulation cannot be used, and furthermore, we would need a great number of BTBS to cover all points in an area and achieve a good positioning accuracy. As said before, RSSI and LQ are the only measures we can get from cellular phones or low cost Bluetooh devices in general. Probably, higher quality low cost BT devices will be on the market in next future, which will turn out to be less affected by obstacles and environmental noises. In the mean time, some solutions can come from methods which are capable of dealing with uncertainty. Our first approach in such direction was fuzzy logic. More precisely, we used a Mamdani-type fuzzy [18] system with M

3 inputs and two outputs, where M is the number of base stations in the environment. Inputs are the normalized LQ values (LQ/maxLQ) of BTBS links to the mobile device to be tracked. Outputs are estimations of mobile device rectangular coordinates on a given reference system. We carried out our experiments within our department, where we placed ten BTBS. What appears from results is that each BTBS participates in the calculus for position detection with a weight which depends on both its LQ measure and its distance from other BTBSs. We achieved an absolute error of 0.78 mt. and a relative error of 5.17% for a device at a point covered by all BTBSs; the absolute error was 4.96 mt. and the relative error was 21% for the device not completely covered by BTBSs. These experimental results show that the relative positioning error is quite small if all the base stations receive the mobile device signal. A higher relative error is due to the number of base stations that do not receive the BTMD signal. Due to its high computational complexity, the proposed fuzzy algorithm turned out to be unsuitable for position tracking of fast moving devices. It takes about 10 seconds in a Pentium 4, 2GHz PC to return position coordinates. 3.3 Positioning by Neural Networks The high degree of uncertainty entailed by LQ values leads to the high complexity of their relationships with mobile device position; so we tried to approach the problem by an adaptive solution. A positioning problem needs most to relay on site dependent solutions and not to deal with geometric laws. Each site has its own obstacles and environment noises which affect LQ measurement; therefore, we need to assume each LQ value as specific for a given place. A positioning system need to learn LQ values as they are, without any concern with distances between BTMDs and BTBSs. A Neural network can be a solution in such a direction. A neural network can learn the LQ distribution, tune its weights, and then, be ready to provide real time position fast estimates. We carried out some experiments on a singlelayer neural network with n inputs, one output, a linear activation function and no hidden layer. Input is a m-dimensional array of LQ raw values, and no preliminary processing is required. We started our experiments with 15 base stations which we placed within our department randomly. This initial arrangement just was to cover the area where to detect mobile devices. The training set was given by LQ values read by mobile devices placed in 5 known positions. Once trained, the network gave a good accuracy with a relative error of 3% for positions covered by the whole base station set. When we moved mobile devices far from some base stations, relative error raised to unacceptable values. At present we are arranging a multi-layer neural network with the end of improving accuracy. To this end we are also trying to refine weights in order to improve the dependence of output coordinates on input array. 4 Bluetooth Base Station Layout Optimization A common result of our experiment is that positioning accuracy can be heavily affected by erroneous base station arrangements. This problem can be summarized as follows: when arranging a positioning system in a site, there are some constraints to be taken into account as for the selection of the places where to install base stations. For instance, we are unlikely to be allowed to put a base station in the middle of a room. Further constraints are to be observed when we are dealing with heritage sites: base stations should be invisible and just selected places are available. Therefore, the problem can be expressed in the following terms: given the total number of places where a BT base station can be put, select a minimal subset which allows the system to evaluate the position of a BT mobile terminal in any part of the site, with an accuracy degree as high as possible. 4.1 Case Study: Positioning in a Middle-Age Castle The site we are investigating is the Manfredi s Castle in Mussomeli - Italy (fig. 1a), whose map is sketched in Fig. 1b. We started assuming that a maximum of 10 suitably placed BTBS had to be enough to cover the whole area of the castle. According to the assumption in subsection 3.2, we assume that LQ dependence on walls, noises, and obstacles in general, is likely to be low within small regions. For instance, we can assume the LQ values read in one room mostly depend on BT mobile device position in that room rather than on the obstacles standing from the room to a BTBS. That is especially true if the room has its longest dimension in the direction of the BTBS. Site partitioning in small regions also increases the probability to detect a base station, among the others, that provides LQ values almost linearly

4 dependent on the distance from BTBS to BTMD. 1. External wall door 2. Stable 3. Castle entrance 4. Courtyard 5. Big arc 6. Vestibule 7. Gothic portal 8. Barons Hall 9. Destroyed bodies 10. Three women room 11. Chimney hall1 12. Cross vault halls 13. Semicircular Turret 14. Maid lodging 15. Male 16. Chapel 17. Polygonal external wall 18. External wall stable 19. Hayloft 20. Defense fencing 21. Double lancet windows Fig. 1a: Castle planimetry Fig. 1b: Castle map Fig. 2: Castle cell arrangement First we split the area to be visited in 70 cells, each representing a room, an area around an artifact, a path (or even a part of it). In detail, there are 50 rectangular cells for rooms and paths and 20 circular cells for areas around columns, portals, or other interesting artifacts. Fig. 2 sketches the 50 rectangular cells, with blue points being centers of circular cells. Then we selected 50 suitable positions where BTBS were allowed to be put. These are the red points in Fig. 2. We conducted a lot of measurements to collect a significant set of LQ values. To this end we placed a BTBS in each red point of the map and read LQ values from a BTMD moving around. Then we stored the collected data in a 50x70 matrix (Table. 1). A generic (i,j) matrix element represents the LQ value range which was measured by the BTBS at the i th position (i = 1, 2,..., 50) as for a BTMD moving within the j th cell (j = 1, 2,, 70). If the generic (i,j) element contains the [0..0] range, the BTBS placed at the i th position did not read significant LQ values for a BTMD in the j th cell.

5 We assumed a width and a maximum error (in meters) for each cell: a diameter of 2 mt. with a maximum error of 0.5 mt. was initially assumed for cells around artifacts; the longest internal distance (typically a diagonal) and a maximum error of 1 mt. was initially assumed for other cells (rooms, paths, or parts of them). Table. 1: LQ ranges : Layout Optimization by Genetic Algorithm Starting from the above assumptions, we implemented a genetic algorithm to achieve the best BTBS arrangement upon the condition the area under investigation was fully covered, the requested precision achieved, and a minimum number of BTBS deployed. Our genetic algorithm implements the evolution of a population of chromosomes with 50 binary genes each. The i th gene represents one of the 50 points where a BTBS is allowed to be placed. Our idea was to work with a constant number of chromosomes and to attach the same coupling probability to each of them. We started from generating a population of 50 chromosomes and assigning each gene a probability of 25% to be included in a chromosome. This way each chromosome includes a variable number of genes which converges towards an almost stable number produced by the evolutionary process. Once calculated the q i,j chromosome quality values, evolution starts. Depending on age, fitness and coverage, a maximum of 10 chromosomes are killed each step. The same number of surviving chromosomes are then coupled to generate new chromosomes thus replacing the killed ones. The evolution steps are then repeated 500 times. Coupling is performed according to one-pointcrossover and alternating gene exchange (one time the initial part and one time the final part. A randomly selected chromosome is candidate to be killed if it is old, or if it does not cover the whole area. As for age, an equal percentage probability to die is assumed. So, if a chromosome is 15, it has a 15% probability to die, being 100 the maximum age for each chromosome. As a further condition, if a BTBS is the only one to cover two different cells by measuring the same LQ range, the whole chromosome cannot be a solution. In any case each cell must be covered by two different BTBS at least, with a requested precision given by the q i,j ratios; otherwise the considered solution must be rejected. As said before, a width W j and a maximum error p j are assumed for the j th cell. A LQ range [LQ min.. LQ max ] i,j is measured by a BTBS placed at the i th position as for a BTMD moving within the j th cell. As far as the requested precision for the j th cell is concerned, we divide the W j width in n j parts, with n j = W j / p j (rounded to the next integer). This precision can be achieved only if one BTBS at least gives a LQ range greater than, or equal to, n j. Next we consider the ratio [LQ max - LQ min ] i,j / n j = q i,j which gives the values to be stored in a matrix. We adopted two kind of refinements in our experiments. One deals with the requested accuracy, and one with the number of genes to be included in a chromosome. As far as artifact cells are concerned (see Table. 2), we carried out experiments by setting the maximum error from 0.5 to meters, while for other cells we set the maximum error from 1 to 0.33 meters. No solutions were found for the precision. As far as the minimization of the number of BTBS is concerned, we achieved a solution with 10 BTBS deployed, a maximum error of 0.4 for the artifacts cells and 0.8 for other cells. In this case, the site coverage is the one represented in Fig. 3. Lower accuracies allowed us to find solutions with a lower number of BTBS.

6 Fig. 3: Best solution site coverage 5 Conclusions This paper demonstrates the relevance of Bluetooth base station layout optimization for high definition positioning. To this aim we proposed the use of a genetic algorithm whose solution succeeded in minimizing the number of base stations to be deployed. Our best result, in the case study on positioning in a castle, was 10 base stations and accuracy better than 0.5 meters. Even better accuracy can be achieved according to different problem setup, for instance, by increasing the number of base stations to be deployed. We are currently working on improving the genetic algorithm by the use of more effective evolutionary strategies and of more complex fitness function to be capable of providing even better results. The used approach gave solutions which were specific for the castle problem; nevertheless, it can be used for detecting optimal arrangements of Bluetooth base stations for positioning in any area. As a final remark, we can say layout optimization should be considered a preliminary step for positioning. Once the base station layout has been optimized, various methods can be adopted for actual positioning. In this paper we discussed our positioning approaches by the use of triangulation, fuzzy logic and neural networks. Table. 2: Experimental results Gene probability to Maximum artifact Maximum generic BTBS Number Evolution steps be in a chromosome cell allowed error cell allowed error 25 0, ,4 0, ,375 0, ,25 0, ,166 0, , ,4 0, ,375 0, ,25 0, ,166 0, , ,4 0, ,375 0, ,25 0, ,166 0, , ,4 0, ,375 0, ,25 0, ,166 0,33 - -

7 References [1] A. Genco, HAREM: The Hybrid Augmented Reality Exhibition Model, WSEAS Trans. On Computers, Issue 1, vol. 3, 2004, ISSN [2] A. Genco, Augmented Reality for Factory Monitoring, Proc. of int. conf. ISC Industrial Simulation Conference, June 7-9 Malaga, Spain, 2004 (2 nd Best Paper Award) [3] F. Agostaro, A. Genco, S. Sorce, A Collaborative Environment for Service Providing in Cultural Heritage Sites, Proc. of International Conference on Embedded and Ubiquitous Computing (EUC-04) August 26-28, Aizu Wakamatsu, Japan 2004 [4] F. Agostaro, A. Genco, S. Sorce, A Fuzzy Approach to Bluetooth Positioning, WSEAS Transactions on Information Science and Applications (ISSN ), issue 1, vol. 1, July. 2004, pag Best Paper Award on Computers at the 8 th int. Conf. CSCC Circuits, Systems, Communication and Computers, July Vouliagmeni, Grece, 2004 [5] F. Agostaro, F. Collura, A. Genco, S. Sorce, Problems and solutions in setting up a low-cost Bluetooth positioning system, WSEAS Trans. On Computer Science, Issue 4, vol. 3, 2004, ISSN [6] Getting, I.A., The Global Positioning System, IEEE Spectrum, Volume 30, Issue 12, Dec. 1993, Pages: 36-38, [7] A. Harter, A. Hopper, P. Steggles, A. Ward, P. Webster, The Anatomy of a Context Aware Application, Proceedings of 5th Annual Int. Conference on Mobile Computing and Networking, ACM Press, New York, 1999, pp [8] N. B. Priyantha, A.K.L. Miu, H. Balakrishnan, S. Teller, The cricket compass for context-aware mobile applications, Proceedings of the 7th annual international conference on Mobile computing and networking, July [9] J. Hightower, C. Vakili, G. Borriello, R. Want, Design and Calibration of the SpotOn Ad-Hoc Location Sensing System, University of Washington, Seattle, August 2001, on/ [10] P. Bahl and V. Padmanabhan, RADAR: An InBuilding RFBased User Location and Tracking System, Proc. IEEE Infocom 2000, IEEE CS Press, Los Alamitos, California., 2000 [11] Technical Description of DC Magnetic Trackers, Ascension Technology Corp., Burlington, Vt., [12] I. Jami, M. Ali, R. F. Ormondroyd, Comparison of Methods of Locating and Tracking Cellular Mobiles, IEE Colloquium on Novel Methods of Location and Tracking of Cellular Mobiles and Their System Applications (Ref. No. 1999/046), 1999 [13] Hallberg, J.; Nilsson, M.; Synnes, K.; Positioning with Bluetooth, Proceedings of 10th IEEE International Conference on Telecommunications ICT 2003, Volume: 2, Pages: [14] Kotanen, A.; Hannikainen, M.; Leppakoski, H.; Hamalainen, T.D.; Experiments on Local Positioning with Bluetooth, Proceedings of IEEE International Conference on Information Technology: Coding and Computing [Computers and Communications] (ITCC) 2003, Pages: [15] Anastasi, G.; Bandelloni, R.; Conti, M.; Delmastro, F.; Gregori, E.; Mainetto, G; Experimenting an indoor bluetooth-based positioning service, Proceedings of 23rd International Conference on Distributed Computing Systems Workshops, 2003, Pages: [16] M. Spratt, An Overview of Positioning by Diffusion, Wireless Networks, Volume 9, Issue 6 Kluwer Academic Publishers, November 2003 [17] Specification of the Bluetooth Core System 1.1, [18] E.H.Mamdani, Applications of fuzzy logic to approximate reasoning using linguistic synthesis, IEEE Transactions on Computers, Vol. 26, No. 12, 1977.