Brainstorming Multi-Agent Reinforcement Learning for Multi-Vehicles Games

Every year, traffic accidents in the world cause economic losses equivalent to 600 billion dollars. And autonomous driving technology can improve driving safety and traffic efficiency. Therefore, unmanned vehicles are the development direction of future transportation, and decision-making control is an important issue that needs to be faced in the development of unmanned driving technology. The existing reinforcement learning algorithms are mostly limited to the research of single agent. Combining the reality of multi-vehicles driving on the road at the same time, in this work, we propose brainstorming multiagent reinforcement learning (BMARL) to guide the decision-making of multi-vehicles autonomous driving. The basic framework of BMARL is based on the actor-critic network structure. It adopts a centralized training and decentralized execution structure. A Critic network and two Actor networks are trained for each agent. This article uses the intelligent driving simulation environment commonly used in the field of artificial intelligence, and the open source racing simulator (TORCS) simulates the algorithm to verify the effectiveness of the above algorithm in the field of automatic driving decision-making control.