Hu, Zhiyuan (Tsinghua University), HU, Jianming (Tsinghua University)

KDLight: A Traffic Signal Control Algorithm for Scenarios with Partially Observable Vehicle States

Scheduled for presentation during the Invited Session "S23a-Trustworthy AI for Traffic Sensing and Control" (TH-LM-T23), Thursday, November 20, 2025, 10:30−10:50, Coolangata 2

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

Traffic Signal Control (TSC) has become a critical issue in urban traffic management. Recently, researchers have applied reinforcement learning to TSC, leveraging its advantages of not requiring prior knowledge and enabling real-time control. However, many studies assume all vehicles are connected, ignoring the reality of mixed traffic flow, where connected and non-connected vehicles coexist. In such scenarios, the behaviors of non-connected vehicles remain unobservable, impacting the performance of traditional TSC methods. To address this, we propose KDLight, which employs a multi-unit self-attention mechanism as its network structure. To handle scenarios with partial vehicle information, KDLight uses a two-stage training approach: knowledge distillation with multi-teacher models and reinforcement learning fine-tuning. Comprehensive experiments on the SUMO simulator, using datasets with varying connected vehicle penetration rates, demonstrate the effectiveness of KDLight and its potential for real-world applications.