Reinforcement learning approaches for acquiring conflict avoidance behaviors in multi-agent systems

This paper proposes a new reinforcement learning approach for acquiring conflict avoidance behaviors in multi-agent systems. Multi-agent systems are able to construct orderly systems autonomously through interactions with autonomous agents. We expect to be able to construct flexible and robust systems for the environmental changes by multi-agent system approaches. However, it is difficult for designers to preliminarily embed appropriate behaviors to avoid conflict because complex dynamics emerge by interactions with many agents. We demonstrate the effectivity of the proposed method in narrow road problems that many agents go by each other in narrow roads. In narrow road problems, it is the optimal strategy that agents select behaviors different from other agents. However, agents are faced with a dilemma because they cannot predict other agents' behaviors beforehand. The proposed method can differentiate into agents preferring to pass through a road and agents preferring to give way, by Q-learning that can adjust discount rates. We show that agents differentiate into two type of agents, and acquire stable conflict avoidance behaviors with higher probability than conventional Q-learning algorithms, through experimental results.