Application of Deep Reinforcement Learning in Optimization of Traffic Signal Control.

Traffic signal control is the most basic and pivotal task in road traffic problems. The optimization of traffic signal control is for improving vehicle passing efficiency of each intersection in order to improve the traffic capacity of the whole road net. Currently, traditional traffic control methods are not able to satisfy the needs of continuously expanding traffic conditions. Thus, with ideal self-adapting and exploring abilities, deep reinforcement learning methods are widely applied in model traffic signal control. However, due to the problems of overestimation and instability during the learning process of Deep Q-learning (DQN), we came up with a model based on Double-Dueling-Deep Q Network (3DQN). This model combines dueling network, double learning method and target network by natural DQN to relieve overestimation and improve data using efficiency. According to the experimental results, compared with traditional traffic control and DQN, the model of 3DQN has better outcomes in traffic signal control.