Continuous learning and cooperative prediction for traffic dynamics by Adaptive Multi-Agent Systems

Prediction of traffic dynamics plays a significant role in many Intelligent Transportation Systems (ITS). Nonetheless, accurate and real-time traffic prediction is always a difficult task. The classical models are challenged by the complex spatiotemporal relationships of the road network that raises unsolved questions relating the reliability and the feasibility of prediction models. Nowadays, the development of localization and communication technologies in transportation has led to massive data collected by on-board sensors known as floating car data (FCD). These data sets open up a new direction for traffic prediction using big data analysis methods. In this paper, we propose to address the traffic dynamics prediction problem using a self-adaptive multi-agent system that aims at continuously processing vehicle trajectory data to detect and learn different traffic dynamics and thus predict traffic evolution. The proposed system includes two processes: local learning, which distributes learning tasks at the agent level, and prediction process, which enables accurate traffic prediction using cooperative interactions among agents. The conducted experiments underline the performance of our system compared to the well-known models in the traffic prediction domain.