Multi-agent Reinforcement Learning-based Control in a Yield-sign Controlled Intersection

Abstract

Both single- and multi-agent reinforcement learning have been widely used to solve different problems of autonomous driving. Single-agent methods have already outperformed traditional rule-based algorithms in several areas, but recent research shows that multi-agent algorithms have an even greater potential. Encouraging cooperative behavior and communication between the agents leads to their altruistic behavior, which results in safer and more reliable driving strategies while applying techniques such as parameter sharing and curriculum learning helps to deal with the increased complexity of the problem. This paper aims to compare the performance and reliability of single- and multi-agent reinforcement learning algorithms through the example of a yield-sign controlled intersection.