This paper introduces a multi-agent approach to adjust traffic lights based on traffic situation in order to reduce average delay time. In the traffic model, lights of each intersection are controlled by an autonomous agent. Since decision of each agent affects neighbor agents, this approach creates a classical non-stationary environment. Thus, each agent not only needs to learn from the past experience but also has to consider decision of neighbors to overcome dynamic changes of the traffic network. Fuzzy Q-learning and Game theory are employed to make policy based on previous experiences and decision of neighbor agents. Simulation results illustrate the advantage of the proposed method over fixed time, fuzzy, Q-learning and fuzzy Q-learning control methods.
