In this research, the Q-learning reinforcement method is used as an approach that interacts with the environment and undergoes the learning process in real-time, aimed at improving the scheduling of intersections. The criterion considered is the cumulative waiting time of vehicles on the incoming routes to the intersection, which is used as a benchmark for evaluating the proposed method. This study attempts to critically analyze the performance and shortcomings of conventional discrete-state reinforcement learning methods in problems with a very high number of states. The main contribution of this research is the definition of a new state space in traffic signal control problems and the examination of the impact of reducing the resolution of the state space and considering the uncertainty in action selection on improving the quality of reinforcement learning. Furthermore, it is shown that using reduced resolution leads to improved agent performance despite the use of discrete-state methods. To evaluate the performance of the proposed method, the results obtained from the proposed method are compared with those from the Q-learning method in continuous states and fixed-time from the articles [1, 2]. For the implementation of the designed control algorithms and generating the traffics data, a traffic simulator called SUMO has been used. The results indicate a reduction in the total waiting time and also a decrease in computation time in our proposed method compared to the methods under comparison.

کلمات کلیدی