FIT: Fairness-Aware Intelligent Traffic Signal Control with Deep Reinforcement Learning

Abstract

The fast development of learning techniques make it possible to intelligently adjust the traffic signal, also the promotion of the future network(e.g. 6G) provides the possibility to adjust the signal in real time. Several studies have proposed learning-based methods for signal control achieving superior performance compared with traditional methods. However, these methods mainly focus on improving the total efficiency for passing the intersection. To this end, a higher priority is usually given to the lane with larger traffic, which could cause the lane with few vehicles to be starved for a long time. Although traditional methods have better performance in preventing starvation, the efficiency still needs to be greatly improved. In this paper, we consider both the fairness and efficiency in intelligent traffic signal control and try to relieve such starvation while improving the total efficiency. Specifically, a deep reinforcement learning based approach is proposed to dynamically control the traffic signal according to real-time traffic information. Inspired by the proportional fair scheduling (PFS) in wireless networks, a new fairness-aware traffic signal control model is designed to maintain a good trade-off between efficiency and fairness. Extensive experiments are conducted to demonstrate that our method achieves better fairness while also provides a good efficiency guarantee.