Evaluation of Multi-Agent Deep Reinforcement Learning Model with Fault-tolerance Attention Mechanism for Traffic Light Control System

Abstract

Managing urban traffic at intersections is a complex challenge. Traditional traffic signal systems struggle to adapt to real-time congestion and variable vehicle flow, particularly at roads with high traffic volume. These systems also lack coordination between neighbouring intersections, leading to inefficient vehicle movement, delays for emergency vehicles, and unsafe pedestrian crossings. This paper proposes a solution using Multi-Agent Reinforcement Learning (MARL) to model a traffic network as a multi-agent system. Specifically, it employs Fault-Tolerant Attention Multi-Agent Deep Deterministic Policy Gradient (FT Attn. MADDPG), where decisions are based on average queue lengths. The Fault-tolerance Attention mechanism allows agents to minimize the impact of malfunctioning agents, improving overall performance. The approach also supports various intersection types through a parametric action space. Simulation results show that FT Attn. MADDPG significantly reduces travel time by 16.21% under high, 26.97% under medium, and 6.89% under low traffic demand compared to standard MADDPG.