Urban Traffic Data Mining and Neural Network Models

1 URBAN TRAFFIC DATA MINING The development of Intelligent Transportation Systems and the diffusion of urban traffic control and monitoring by Autonomous Agents have created very large databases of realtime and historic traffic data. Therefore, data mining approaches are necessary to design effective Decision Support Systems for traffic management and user information generation. In particular, the completion of urban traffic data in a spatial-temporal context represents a relevant problem particularly suited for traffic states estimation and forecasting, real-time traffic control, dynamic OD matrices updating and so on. As regards the traffic data completion, a constraints propagation approach has been proposed and implemented in the framework of the KITS European Research Project. However, more efficient and flexible methods have to be investigated and tested in order to support traffic data mining problems. 2 MULTILAYER NEURAL NETWORKS Neural Networks have been widely applied for traffic forecasting and they can be used for spatial data completion as well. To this aim, mobile detectors are able to collect necessary traffic data relative to those variables involved in the extrapolation. In this way it is possible to correlate traffic flows collected by fixed and mobile detectors for the training phase of the neural network model. Then, it will provide in output flow values relative to arcs no monitored on the basis of real data in input, relative to arcs with fixed detectors. The aim of this paper is to investigate and demonstrate the capabilities of neural network models for the estimation and completion of traffic data by the correlation among some variables of the traffic process, i.e. the arc flows but also speed and queue variables. Training methods based on optimization techniques, both in terms of accuracy (memory and generalization) and training time (cpu and epocs number) will be discussed. Further promising trends concern the implementation of new methods based on fast training algorithms (quasi-Newton) by testing different neural network architectures such as those developed with back-propagation techniques. 3 CASE-STUDY APPLICATIONFrom an application point of view, a neural network model has been designed and tested ona small road network of the Rome city. In particular, a main itinerary has been selected andthe traffic data completion on some arcs has been considered. A back-propagation algorithmwith momentum terms has been implemented and the results obtained show a good matchamong simulated and real traffic data. REFERENCESBielli M., Ambrosino G, Boero M. (Eds), Artificial Intelligence applications to trafficengineering, VSP, 1994 Bielli M., Caramia M., Carotenuto P., Genetic algorithms in bus network optimization,Transportation Research C, Vol. 10C, n.1, 19-34, 2002 Dougherty M.S., Applications of Neural Networks in Transportation, TransportationResearch C, Vol.5C, n.5, 1997 Bielli A., Un modello di reti neurali per la simulazione di traffico stradale, Doctoral Thesisin Mathematics, Roma Tre University, Italy, November 2001 Bielli M., Reverberi P., New Operations Research and Artificial Intelligence approaches totraffic engineering problems, EJOR, n.92, pp.550-572,1996 Faghri A., Hua J. Evaluation of Artificial Neural Network applications in transportationengineering, Transportation Research Record 1358, TRB, 1992Proceedings of the 11 Mini-EURO Conference on Artificial Intelligence in TransportationSystems and Science, and the 7 EURO Working Group Transportation, Helsinki, August1999 Bielli M., Carotenuto P. (Eds), Proceedings of the Rome Jubilee 2000 Conference onTransportation, Rome, September 2000