The Research Of Urban Traffic Signal Control Based On Multi-Agent System

Traffic signal control is an important method to alleviate traffic congestion,improve travel efficiency and reduce traffic accidents.Currently,the problem is that the traffic light control system is usually time-fixed,the light is difficult to control in real time according to real situation.Therefore,searching for a more efficient and intelligent traffic light control has attracted researchers' close attention.Since the traffic control system is a complexly uncertain system with nonlinear structure,though the traitional adaptive method is effective,it's still difficult to adapt to the changeable traffic flow,and it relies heavily on the traffic model.Reinforcement learning method does not need to establish traffic model,but can realize the improvement of control scheme through continuous interaction with traffic environment.In this paper,the traffic signal control system is viewed as a multi-agent system where each traffic signal controller controls one intersection,and the reinforcement learning is introduced to realize traffic signal control in urban traffic network.First of all,this paper investigates a single-junction which is the minimun traffic control unit.Technological advances have made it possible for traffic system to collect large volums of varied data.In order to make full use of the traffic data collected,the deep reinforcement learning algorthim is introduced to realize the real-time control of the intersection,and a new state space design method is proposed.The simulation of a single intersection by deep reinforcement learning algorthim in SUMO verifies the effectiveness of this method.Then the structure of urban traffic signal control system based on Multi-Agent is studied.Then this paper studies the max-plus algorthim based on coordination graph.Further,on the basis of deep reinforcement learning algorthim,the knowledge of game theory is introduced.Next this paper establish the Multi-Agent interaction model based on the n-person non-zerocooperative game,and the corresponding game table is established to solve the Nash equilibrium under the current environment.On the basis,the Nash Q learning method for solving the four phases of multiple intersection is proposed.The simulation shows the effectiveness of the proposed methods.