Abstract: In traffic networks,the target of intersection scheduling is to maximize the flow rates and minimize average waiting time of all concerned vehicles.Game relationships exist between each intersection and the other ones in the traffic scheduling.In the process of this traffic game,there is a constraint of mutual coordination among strategies so that the maximal profits of neighboring intersections can be achieved.In this paper,we focus on the complex traffic scheduling problem,and give the modeling approach for traffic systems based on the multi-agent multi-step game theory.Considering the practical characteristics of dynamic traffic game environments,in this paper we further propose an approach for learning traffic scheduling strategies from historical traffic scheduling data based on the reinforcement learning algorithm in game theory.Then the Nash equilibrium of multiple intersections that participate in the traffic game can be achieved ultimately.Therefore,the optimal scheduling strategies of traffic systems will be obtained.Experimental results show the feasibility and effectiveness of our method.