Approximate Dynamic Programming for Traffic Signal Control at Isolated Intersection

As a new optimization technique for discrete dynamic systems, approximate dynamic programming (ADP) for the optimization control of a simple traffic signalized intersection is proposed. ADP combines the concepts of reinforcement learning and dynamic programming, and it is an effective and practical approach for real-time traffic signal control. This paper aims at minimizing the average number of vehicles waiting in the queue or the vehicles average waiting time at isolated intersection by using the action-dependent ADP (ADHDP). ADHDP signal controller is designed with neural networks to learn and achieve a near optimal traffic control policy by measuring the traffic states. As shown by the comparison with other traffic control methods, the simulation results indicate that the approach is efficient to improve traffic control at a simple intersection.