Desenvolvimento de um agente inteligente para jogar jogos genéricos de Atari 2600

Abstract

Among the areas addressed in Computing, the evolution of Artificial Intelligence has been highlighting, especially in the area of games. The aim of this paper is the development of an intelligent agent capable of multiplayer games. For this, a study was first made on the necessary concepts, which constitute all the theoretical foundations of this work. For the abstraction of machine learning and image processing concepts, we used the Keras and Gym libraries. This paper presents the definition of models of Artificial Intelligence, Machine Learning, Deep Learning, Reinforcement Learning, Deep Reinforcement Learning, Q Learning, and Asynchronous Advantage Actor-Critic adapted to the context of this work. This paper surveyed the state of the art in the area of algorithms for playing generic games, presenting the works that are most relevant to the current context. Agents designed to effectively play multiple games must have the ability to adapt to the environment in which they find themselves, thus determining which actions best suit each situation. The quality of an agent is measured by the distance between its scores and the scores obtained by human players. This paper presents the development of an agent capable of playing multiple games without the need to define rules in advance, receiving only the game screen image, the current score and whether the current state represents an endgame state. For this, the diagrams of agent development are presented, together with their respective information. To prove the effectiveness of the agent, the scores obtained in the Space Invaders, Breakout, Demon Attack, and Kung Fu Master games were compared to the scores obtained by human players.