情報処理学会 第76回全国大会講演要旨

We investigate how the sequential actions influence the individual strategies in iterated prisoner’s dilemma game using the agent-based modeling and the genetic algorithm in agent strategy evolution. By sequential actions, the player performs the same action (cooperate or defect) for more than one round, and this intention of performing the same action again is known to the opponent. Each agent’s unique strategy determines how it responds depending on the past actions of itself and the opponent. We analyze the extent to which the sequential actions with and without conditional bonus and/or penalty are favored in local and non-local iterated prisoner’s dilemma game. The results shows that the strategies containing cooperative sequential actions are able to survive at a stable minority proportion under different situations, suggesting that the preference of non-sequential actions in strategy selection is greater unless the sequential cooperative actions are compensated with a potential bonus. We also find that the dominant strategies in games with sequential actions demonstrated some characteristics of “tit-for-tat” type of strategies in standard iterated prisoner’s dilemma game.