Traffic Signal Control Using Deep Reinforcement Learning with Multiple Resources of Rewards

Intelligent traffic signal control is an effective way to solve the traffic congestion problem in the real world. One trend is to use Deep Reinforcement Learning (DRL) to control traffic signals based on the snapshots of traffic states. While most of the research used single numeric reward to frame multiple objectives, such as minimizing waiting time and waiting queue length, they overlooked that one reward for multiple objectives misleads agents taking wrong actions in certain states, which causes following traffic fluctuation. In this paper, we propose a DRL-based framework that uses multiple rewards for multiple objectives. Our framework aims to solve the difficulty of assessing behaviours by single numeric reward and control traffic flows more steadily. We evaluated our approach on both synthetic traffic scenarios and a real-world traffic dataset in Toronto. The results show that our approach outperformed single reward-based approaches.