부산대학교 도서관

Reinforcement learning (RL)-based traffic signal control (TSC) optimizes signal switches through RL agents, adapting to intersection updates. Yet, existing RL-based TSC methods often demand substantial storage and computation resources, impeding real-world implementation. This study introduces a two-stage approach to compress the network, maintaining performance. Firstly, we identify a compact network via a removal-verification strategy. Secondly, pruning yields an even sparser network. In addition, Multi-task RL is adopted for multi-intersection TSC, reducing costs, and boosting performance. Our extensive evaluation shows a compressed network at 1/1432nd of original parameters, with an 11.2% enhancement over the best baseline. This work presents an efficient RL-based TSC solution for real-world contexts, offering insights into challenges and opportunities in the field.