Intelligent Traffic Signal Control Based on Critical Intersection Identification

Abstract

To address the high training cost, excessive computational burden, and scalability limitations of deep reinforcement learning (DRL)-based large-scale traffic signal control (TSC), this paper proposes KeyNode-Light, an intelligent framework that strategically deploys DRL agents based on critical intersection identification. Specifically, the framework first fuses a local structural constraint indicator derived from structural hole theory with a global topological importance indicator obtained via improved K-shell decomposition to accurately identify core nodes. Dedicated DRL agents are then deployed exclusively at these critical intersections, leveraging a spatiotemporal attention network that integrates both graph attention and self-attention mechanisms to capture complex traffic dependencies. Simulation experiments on a 45-intersection traffic network using SUMO demonstrate that KeyNode-Light significantly decreases the number of stopped vehicles and fuel consumption, achieving superior performance compared with several baseline methods.