マルチエージェント強化学習を応用したサッカーエージェントの行動生成

抄録

The objective of this study is to acquire behaviors for controlling multiple soccer robots using multi-agent reinforcement learning. For the experiments, we constructed a simulation environment for soccer using RoboCup Humanoid League as a reference, and used it as a base for future application to real environments. For the experiments, MA-POCA was used as the reinforcement learning algorithm. We also designed rewards to ensure that the robot learns to play soccer appropriately. As a result of learning, the soccer robot’s policy acquired behaviors necessary for soccer, such as dribbling, passing, and defense. This paper has contributed to the automation of the soccer robot’s behavioral decisions.