Learning and Generalization in Atari Games


This thesis describes the design of agents that learn to play Atari games using the Arcade Learning Environment (ALE) framework to interact with them. The application of machine learning in video games, given its high complexity, is considered to be a bridge towards real-world domains such as robotics. The goal in Atari games is to achieve the highest possible score. To solve this task, reinforcement learning and search techniques are used. These algorithms outperform humans in 30 of the 61 games supported by ALE. Since humans are very good at making generalizations between games, special emphasis is/ngiven to evaluating how well an agent learns from multiple games simultaneously. These experiments usually result in a higher score for specific pairs of games. Besides, there are games that tend to increase their score when playing with other games, whereas there are games that help others to perform better.

Bachelor’s Thesis, Universitat Pompeu Fabra