David R. Collins


Introduction: The use of machine learning tactics such as the Moran process and evolutionary finite state machines have the potential to outperform classic strategies for the iterated prisoner’s dilemma.

Methods: Applying a genetic algorithm approach to the iterated prisoner’s dilemma while modeling strategies with finite state machines proved to be an efficient method in which the produced strategies were able to cooperate unilaterally with their opponents.

Results: Varying parameters in the evolution process such as the amount of generations, population size and bottleneck size were shown to directly contribute to the success of the strategies produced. In comparing various optimization methods, the genetic algorithm utilizing finite state machines outperformed the Moran process with respect to the highest scoring strategies produced by each.

Discussion: These results can be explained due to the benefits of genetic recombination that was made possible with the use of finite state machines, where crossing over of state / action pairs resembling choices to make depending on the state environment occurred from generation to generation. Due to natural selection and recombination, the genes of the strategies with the highest fitness levels were bred into the next generations. Population bottlenecks and gene mutation tactics were used to recreate the naturally occurring phenomenon of gene variation, resulting in the creation of new species (which are iterated prisoner’s dilemma finite state machine strategies) within the generations.

Conclusion: Tangible tactics can be extracted from these strategies, which were evolved using standard genetic algorithm tactics utilizing gene crossover and mutation techniques.

Abstract 490 | PDF Downloads 947