GA-Critic: A Traffic Signal Control Strategy Under Incident Conditions for Urban Networks

Non-recurrent congestion is typically caused by traffic incidents in urban networks, particularly when a valid Traffic Incident Management (TIM) strategy is lacking. This paper proposes a Traffic Signal Control (TSC)-based TIM strategy to mitigate severe congestion following incidents. We utilize a Twin Delayed Deep Deterministic Policy Gradient (TD3)-based Deep Reinforcement Learning (DRL) algorithm to optimize phase splits under incident-free conditions. However, the trained DRL agent faces challenges when sudden environmental changes occur due to incidents. To address these unexpected variations in traffic capacity and demand, we introduce GA-Critic, a genetic algorithm-based gradient-free optimizer. When incidents occur and road segments are closed for diversion, we present two approaches to help the TD3 agent adapt to the new environment swiftly. Through multiple experiments, we demonstrate the effectiveness of the online TIM strategy.