An Intelligent Agent Playing Generic Action Games based on Deep Reinforcement Learning with Memory Restrictions

Among the topics that increasingly gained special attention in Computer Science recently, the evolution of Artificial Intelligence has been one of the most prominent subjects, especially when related to games. In this work we developed an intelligent agent with memory restrictions so to investigate its ability to learn playing multiple, different games without the need of being provided with specific details for each of the games. As a measure of quality of the agent, we used the difference between its score and the scores obtained by casual human players. Aiming to address the possibilities of using Deep Learning for General Game Playing in less powerful devices, we explicitly limited the amount of memory available for the agent, apart from the commonly used physical memory limit for most works in the area. For the abstraction of machine learning and image processing stages, we used the Keras and Gym libraries. As a result, we obtained an agent capable of playing multiple games without the need to provide rules in advance, but receiving at each moment only the game video frame, the current score and whether the current state represents an endgame. To assess the agent effectiveness, we submitted it to a set of Atari 2600™ games, where the scores obtained were compared to casual human players and discussed. In the conclusion, we show that promising results were obtained for these games even with memory limitations and finally a few improvements are proposed.