Abstract

Enhancing signal control efficiency through deep reinforcement learning constitutes a focal point of research in the field of traffic control. Existing methods primarily rely on copious state information and complex networks to attain precise timing schemes, which can result in decision biases and protracted training cycles. Addressing these shortcomings, a two-stage control algorithm is proposed. The algorithm employs discrete encoding of traffic states to optimize the exploration of the state space and enhances performance by utilizing a competitive Q-network structure. Concurrently, control is divided into two stages: in the first stage, traditional methods are employed to select a subset of traffic state information for phase decision-making, while in the second stage, reinforcement learning techniques are used to select another subset of state information for phase duration prediction. Experiments are conducted on the traffic simulation platform CityFlow, and results demonstrate that the proposed algorithm surpasses traditional control methods and reinforcement learning-based approaches, exhibiting higher traffic flow efficiency and reduced travel times.