Desenvolvimento de estratégias e fenômenos em dinâmicas de jogos de múltiplos agentes

Abstract

Recent developments in Reinforcement Learning (RL) methods are focused on models that can learn good policies in non stationary environments, such as multi-agent games, where agents must learn how to react to changes in other agent’s strategies or in the environment. Some development has been made by studying not only how one agent can develop it’s policy, but how a population of agents can evolve from initial distributions to stable states of strategies. Evolutionary Game Theory (EGT) is the theoretical framework that applies mathematical and economical knowledge from game theory and biological evolution inspiration to study how individuals from a population dynamically interact in an environment. In this paper, we first introduce EGT concepts and show how they can be applied to understanding a population’s learning dynamics in the context of RL. Then we link those concepts with learning algorithms and study how one can infer the behaviour of those methods from links with evolutionary dynamics. Finally, we study and evaluate a recently proposed algorithm derived from policy gradient model and EGT dynamics and discuss next steps.