Distributed Learning Agents in Urban Traffic Control

Automatic learning techniques stand as promising tools to respond to the need of higher efficiency of traffic network, even more so at times of mounting pressure from economic and energy markets. To this end, this paper looks into the operation of a traffic network with distributed, intelligent agents. In particular, it casts the task of operating a traffic network as a distributed, stochastic game in which the agents solve reinforcement-learning problems. Results from computational experiments show that these agents can yield substantial gains with respect to the performance achieved by two other control policies for traffic lights. The paper ends with an outline of future research to deploy machine-learning technology in real-world traffic networks.