繰り返しゲームにおいて協調行動を生成する先読み型行動選択方法

概要

論文の詳細を見る
We have formalized a finite iterated game with change. The formalization extends a traditional framework, e.g., the prisoner's dilemma, by incorporating an influence on the payoff matrix at some future point by executing an action at the present time. This enables us to explain why a rational agent behaves cooperatively, because profit through cooperative behavior finally exceeds that through noncooperative behavior. This situation is seen in human interactions, where We often behave cooperatively, even though from a myopic view, cooperative behavior may not seem profitable. Next, we propose a new method for selecting an action in such a framework. It overcomes the drawbacks of previous methods, thus yielding cooperative behavior and consuming little time. A simple model is used to investigate when our method is able to select an appropriate action. Finally, we compare previous methods and our method by evaluating some example problems. Through experimental results, we show that our method is both efficient and simple.
社団法人人工知能学会の論文
1997-11-01