Multiple Opponent Optimization of Prisoner’s Dilemma Playing Agents

Abstract

Agents for playing iterated prisoner's dilemma are commonly trained using a coevolutionary system in which a player's score against a selection of other members of an evolving population forms the fitness function. In this study we examine instead a version of evolutionary iterated prisoner's dilemma in which an agent's fitness is measured as the average score it obtains against a fixed panel of opponents called an examination board. The performance of agents trained using examination boards is compared against agents trained in the usual coevolutionary fashion. This includes assessing the relative competitive ability of players evolved with evolution and coevolution. The difficulty of several experimental boards as optimization problems is compared. A number of new types of strategies are introduced. These include sugar strategies which can be exploited with some difficulty and treasure hunt strategies which have multiple trapping states with different levels of exploitability. The degree to which strategies trained with different examination boards produce different agents is investigated using fingerprints.