Coordination control of area traffic networks with multiagent architecture based on deep recurrent Q-learning networks

Traffic control is very important in dealing with urban traffic problems, and its purpose is to provide smooth flow and an orderly traffic environment. In this paper, we deal with the problem of area traffic control based on the coordination of agents. Each intersection pertinent to one agent consists of a Deep Recurrent Q-learning Network (DRQN). Each agent takes the traffic state of its own intersection as the input of the convolutional neural network (CNN). Also, the green signal time of the intersection, which is the action of agents and the traffic characteristics that are low-dimensioned by CNN of other agents, is the input of LSTM. This multi-agent architecture can successfully implement the coordination control of the area traffic network, while solving the dimensional explosion problem and fully recognizing the influence of other agents’ states. We simulate algorithm in several possible cases and control strategy for the coordination control by means of VISSIM. The simulation results show that when the proposed method can reduce the total traffic time and queue length under a congested traffic environment.