Adaptive traffic signal control based on Dueling Deep Q-Learning Network

Abstract

The intersection is the place where traffic flow converges and traffic accidents are most likely to occur, and the traffic conditions at intersections is also complex, which is easy to cause congestion and reduce the traffic efficiency. With the development of intelligent transportation system, adaptive traffic signal control is proposed to improve the traffic efficiency of the intersection by adjusting the phase sequence of the signal in real time. Meanwhile, with the extensive application of reinforcement learning in the traffic field, dueling DQN (Deep Q-Learning Network) is introduced to realize the adaptive traffic signal control in this paper. By comparing the traffic flow conditions of intersections under different signal control methods, we find that dueling DQN can effectively improve the traffic efficiency of intersections under the condition of random change of traffic flow. Through experiment results, we can conclude that the expected reward of dueling DQN is 41.1% and 4.8% higher than that of fixed phase method and DQN method. Meanwhile, simulation result shows that dueling DQN can react to the traffic flow of the road network more timely, and has higher stability than DQN, which can effectively improve the traffic efficiency.