Fairness Control of Traffic Light via Deep Reinforcement Learning
Chenghao Li, Xiaoteng Ma, Li Xia, Qianchuan Zhao, Jun Yang
Abstract
Traffic congestion is a severe issue of a developing world. Recently, many researchers are attempting to utilize deep reinforcement learning algorithms to bring intelligence to traffic lights. To the best of our knowledge, most prior researchers only consider the average criterion of all vehicles while training. However, fairness is another important metric but ignored. In this paper, we study the fairness control of traffic light and propose a deep reinforcement learning algorithm to optimize the fairness of all drivers' waiting time. The objective is to minimize the maximal waiting time of drivers during a light time loop, which also partly reflects the optimization of the average waiting time. We conduct experiments for a 4-lane crossroad in SUMO. Simulation results show that our algorithm can efficiently optimize the fairness criterion. Meanwhile the average criterion is further improved. We wish to shed light on complementing the entire framework of reinforcement learning with our research on fairness control.