Genetic algorithms in repeated matrix games: the effects of algorithmic modifications and human input with various associates

Abstract

In many real-world systems, multiple independent entities (or agents) repeatedly interact. Such repeated interactions, in which agents may or may not share the same preferences over outcomes, provide opportunities for the agents to adapt to each other to become more successful. Successful agents must be able to reason and learn given the dynamic behavior of others. This is challenging for artificial agents since the non-stationarity of the environment does not lend itself well to straight-forward application of traditional machine learning methods. In this paper, we study the performance of genetic algorithms (GAs) in repeated matrix games (RMGs) played against other learning agents. In so doing, we consider how particular variations in the GA affect its performance. Our results show the potential of using GAs to learn and adapt in RMGs, and highlight important characteristics of successful GAs in these games. However, the GAs we consider do not always perform effectively in RMGs. Thus, we also discuss and analyze how human input could potentially be used to improve their performance in RMGs.