Double Deep Q-Network with Dynamic Bootstrapping for Real-Time Isolated Signal Control: A Traffic Engineering Perspective

Abstract

Real-time isolated signal control (RISC) at an intersection is of interest in the field of traffic engineering. Energizing RISC with reinforcement learning (RL) is feasible and necessary. Previous studies paid less attention to traffic engineering considerations and under-utilized traffic expertise to construct RL tasks. This study profiles the single-ring RISC problem from the perspective of traffic engineers, and improves a prevailing RL method for solving it. By qualitative applicability analysis, we choose double deep Q-network (DDQN) as the basic method. A single agent is deployed for an intersection. Reward is defined with vehicle departures to properly encourage and punish the agent’s behavior. The action is to determine the remaining green time for the current vehicle phase. State is represented in a grid-based mode. To update action values in time-varying environments, we present a temporal-difference algorithm TD(Dyn) to perform dynamic bootstrapping with the variable interval between actions selected. To accelerate training, we propose a data augmentation based on intersection symmetry. Our improved DDQN, termed D3ynQN, is subject to the signal timing constraints in engineering. The experiments at a close-to-reality intersection indicate that, by means of D3ynQN and non-delay-based reward, the agent acquires useful knowledge to significantly outperform a fully-actuated control technique in reducing average vehicle delay.