Machine learning agents for computer games

• Autor(a) principal: Araújo, Miguel Diogo Ferraz
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/31371
• topic: Machine learning; Reinforcement learning; Deep learning; Deep reinforcement learning; Agents; Computer games

In recent years, new Reinforcement Learning algorithms have been developed. These algorithms use Deep Neural Networks to represent the agent’s knowledge. After surpassing previous Artificial Intelligence (AI) milestones, such as Chess and Go, these Deep Reinforcement Learning (DRL) methods were able to surpass the human level in very complex games like Dota 2, where long-term planning is required and in which professional teams of human players train daily to win e-sports competitions. These algorithms start from scratch, do not use examples of human behavior, and can be applied in various domains. Learning from experience, new and better behaviors were discovered, indicating a lot of potential in these algorithms. However, they require a lot of computational power and training time. Computer games are used in an AI course at the University of Aveiro as an application domain of the AI knowledge acquired by students. The students should develop software agents for these games and try to get the best scores. The objective of this dissertation is to develop agents using the latest DRL techniques and to compare their performance with the agents developed by students. To begin with, DRL agents were developed for a simpler game like Tic-Tac-Toe, where various learning options will be addressed until a robust agent capable of playing against multiple opponents is created. Then, DRL agents capable of playing the version of Pac-Man used in the University of Aveiro course, in the 2018/19 academic year, were developed through the realization of various experiments where the parameters used in the learning process were modified in order to obtain better scores. The developed agent, that obtained the best score, is able to play in all game configurations used in the evaluation of the course and reached the top 7 ranking, among more than 50 agents developed by students that used hard-coded strategies with pathfinding algorithms.