ルールセット間距離によるマルチエージェントシステムのグループ形成過程の可視化

概要

論文の詳細を見る
Global behavior of multi-agent systems, which is caused by agents' learning and local interaction, is highly complex. Analysis methods for the behavior are necessary to make application of multi-agent systems. Takashina, et. Al proposed a framework, individual level analysis, that focuses on the individual level characteristics of agents. The current method utilizes not the content of actions but just the entropy of its probability distribution. Therefore, the current method cannot deal with a organizing process such as autonomous role assignment in a system. In this paper, we propose ruleset distance which is a metric for the difference between two rulesets in a generic manner. It employs the difference of probability for each action when an input is given. As for its application, we employ the minority game with reinforcement learning agents. We conclude that the method is effective for the intuitive understanding of group formation process in the minority game.
日本シミュレーション学会の論文
2004-03-15