基于深度强化学习的单路口交通信号控制

Deep Reinforcement Learning for Traffic Signal Control of Isolated Signalized Intersections

合理的交通信号灯控制方案能减少交叉口处的排队长度,缓解交通拥堵问题.路口交通流具有非线性、时变性、不确定性等特点,对其建模困难,从而导致无法借助其精确的数学模型来优化交通信号控制方案.本文将深度强化学习方法应用到交通信号控制问题,深度强化学习Agent以减少路口处的排队车辆总数为目标,通过观察交叉口处所有入口车道的状态进行相位控制;使用SUMO仿真平台对本文提出的控制方法进行了仿真实验.实验结果表明,相较于定时控制方法,本文提出的基于深度强化学习的控制方法能显著减少交叉口处的排队车辆数,缓解交通拥堵.

Well-designed traffic signal control strategies can help decrease vehicle queue length and alleviate traffic congestion at intersections.It is hard to model traffic dynamics at intersections accurately due to its nonlinearity,time-varying,and uncertainty.Thus it is extremely difficult to optimize traffic signal timing plans mathematically.In this paper,deep reinforcement learning is employed to examine traffic signal control problems.The traffic Agent observes traffic states in all the approach lanes at the intersection,and decides the control plan which aims to reduce the total vehicle queue length at the intersection.The proposed method is tested with the SUMO simulation platform.Experimental results show that the proposed method can reduce the queue length at the intersection significantly compared with the fixed-time signal control method.