The Dynamics of Selection: Game Theory, Evolution, and Machine Learning

Abstract

Inspired by the principle of natural selection, evolutionary algorithms are population-based stochastic search techniques that are known to be effective for many difficult real-world optimization problems. When used to solve games of strategic interaction, evolutionary algorithms can again yield impressive results, but they can also exhibit a variety of pathologies which fail to deliver game-theoretically justifiable outcomes. Many of these pathologies concern the dynamics of selection. In this talk, I will review my research concerning the relationship between selection dynamics and game-theoretic solutions. In particular, I will present novel results that explain why several selection mechanisms are incapable of converging onto polymorphic Nash equilibria. Research into new algorithms that avoid these pathologies will also be discussed.