Designing Reward Functions in Multi-Agent Reinforcement Learning for Intelligent Intersection Control

Road transportation is a significant contributor to the global greenhouse gas emissions. The number of traffic jams, thus the number of hours spent in cars increases year by year. In addition, traffic in an increasing number of cities grew to unbearable levels, significantly increasing emissions globally and locally. The main traffic hubs in cities are intersections with traffic lights, generally supplied with a fixed cycle control, thereby impeding optimal traffic flow and thus optimal emissions. With the spread of reinforcement learning, more and more research addresses this problem. One approach to the problem concentrates on Traffic Signal Control, from which various research emerged in recent times. However, most research uses classical literature rewarding, such as queue length or average speed. This study presents and compares a novel rewarding method using Multi-Agent Reinforcement Learning. The results show that this technique outperforms its rivals in diverse metrics.