Kim, Hyunsoo (Korea Advanced Institute of Science and Technology), Lee, Jung Won (Korea Advanced Institute of Science and Technology), Yeo, Hwasoo (KAIST)

HyLight: Hybrid Approach of Reinforcement Learning and Rule-Based Optimization for Coordinated Traffic Signal Control Based on Average Phase Demand

Scheduled for presentation during the Regular Session "S34c-Data-Driven Optimization and Governance in Intelligent Urban Mobility" (FR-LA-T34), Friday, November 21, 2025, 16:20−16:40, Surfers Paradise 1

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Reinforcement learning (RL) has gained significant attention as an effective method for traffic signal control. While previous studies have explored either offset or phase split individually using deep RL techniques, their isolated optimization often results in suboptimal network-wide performance. Despite the interdependence between the two control variables, coordinated and simultaneous control of both has received limited attention. Moreover, many existing works assume simplified topologies, such as linear corridors, which overlook the complexity of urban networks. To address these limitations, this paper presents HyLight, a hybrid traffic signal control framework designed for real-world deployment in complex urban networks. HyLight integrates RL-based offset control with rule-based phase split optimization, jointly addressing two fundamental variables in signal control. HyLight adopts an asynchronous decentralized partially observable markov decision process (AD-POMDP) formulation, reflecting that decentralized agents make decisions independently at their own local time. To support RL training and decision-making, a novel traffic state indicator, average phase demand, is introduced, which aggregates fine-grained demand data over control intervals to capture temporally resolved traffic conditions. Experiments conducted on a real urban network in South Korea demonstrate that HyLight outperforms baselines in several evaluation metrics, while maintaining stable and balanced operation across the network.