Traffic Light Controller by Reinforcement Learning Method with Local States

Abstract:
In this paper we describe an efficient method for traffic light controllers. This method is based on enhanced reinforcement learning with local states around each traffic light on crossings. It uses many independent controllers depending upon the number of crossings. Each controller can learn the best actions by a supervised method that is based on reward and punishment policy. The supervisor system monitors the outcome of actions on the specific state. The reward policy is based on this monitoring and tries to minimize the traffic on crossings. Each node or agent tries to self-organize for getting the best decisions (actions) for each state. It is possible to implement this project by the minimum infrastructure and accessories. At the end some result and benchmarking with classical method has shown.