[3I2-GS-5e] エージェント：応用

[3I2-GS-5e-03] Acquiring Cooperative Behavior through Reinforcement Learning of Winning Strategy on Kick the Can Game

〇Ryosuke Takata¹, Takafumi Sakamoto¹, Yugo Takeuchi¹ (1. Shizuoka University)

Keywords:Kick the Can, Cooperation, Reinforcement Learning

In our daily activities, most of the cooperative systems observed in interactions with others are not necessarily planned, but rather emerge emergently through interactions between humans. The purpose of this study is to examine the bottom-up process and conditions for agents to acquire cooperative behavior using multi-agent reinforcement learning with a constructivist approach. As a learning task, we used a traditional game called "kick the can", which is both a cooperative task with the players and a competitive task with the co-adaptive tag. The results suggest that whether the players acquires a selfish or cooperative strategy depends on the difficulty of the task, and that cooperative behavior can be acquired bottom-up at certain difficulty levels. This research clarifies the conditions for cooperative interaction, and contributes to the design of an environment for the emergence of cooperative systems.