The Prisoner ’ s Dilemma Computer Tournaments Revisited

under the title, " Breeding hybrid strategies: the Prisoner's Dilemma computer tournaments revisited. " Page 2 ABSTRACT: The repeated Prisoner's Dilemma has proved to be a rich area of study for examining the tension between competition and coöperation. Computer tournaments—such as Axelrod's and Fader and Hauser's—have sought to obtain " effective " strategies in repeated games from diverse entrants, as a way of generating robust results from a broad range of competing strategies. The development of Holland's Genetic Algorithm has enabled researchers to dispense with open tournaments to obtain such diversity, and to derive strategies almost certainly closer to globally optimal, given the environment of rivals. This paper generalises a process reported by Axelrod in describing how the genetic algorithm has been used to generate winning strategies of specified complexity in a repeated 2-person Prisoner's Dilemma, against two specific " niches " of competing strategies, from Axelrod's second computer tournament. It also derives winning strategies in " noisy " PD, in which competitors are not certain of their opponents' moves. These strategies exhibit some of the characteristics seen in Rapoport's famous Tit for Tat. In an analysis of strategic complexity and its relationship to strategic success, the paper examines finite automata in an attempt to simplify the strategies obtained, which are conditioned on 1-, 2-, and 3-round memory of the moves and outcomes of previous rounds. Winning strategies are represented as finite automata.