Evolving novelty strategies for the Iterated Prisoner's Dilemma in deceptive tournaments

Abstract

This thesis proposes that the concept of deception brought forward by novelty search research can be applied to the Iterated Prisoner's Dilemma problem, and in doing so simultaneously fights the claim that Zero-determinant strategies can outperform any evolutionary opponent. Zero-determinant strategies are a special class of strategies where its moves are probabilistically conditioned on the previous outcome through careful mathematics. When compared with behaviors that merely attempt to obtain the highest score possible through objective search, more complex and above all unique behaviors generated from novelty search allows us to transcend the deception problem that come with certain configurations of an Iterated Prisoner's Dilemma tournament.