Técnicas de aprendizado de máquina aplicadas em jogos RTS

• Ano de defesa: 2020
• Autor(a) principal: Antonio Luis Cardoso Silva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ICX - DEPARTAMENTO DE CIÊNCIA DA COMPUTAÇÃO; Programa de Pós-Graduação em Ciência da Computação; UFMG
• Palavras-chave em Português: Inteligencia Artificial; Jogos; Redes Neurais; Computação – Teses; Jogos digitais – Teses; Inteligencia artificial – Teses; Redes neurais – Teses.
• Link de acesso: http://hdl.handle.net/1843/46487

The Real Time Strategy Games (RTS) domain presents great challenges for the artificial intelligence area because it is a dynamic, real time, adversarial and uncertain environment. One method for addressing these challenges is through the use of machine learning algorithms where an intelligent agent learns using data of games played by humans. In this work, several machine learning techniques were implemented in the creation of an agent capable of playing Starcraft and predicting the outcome of the match. To take care of the strategic and economic part of the game by the agent, online case-based planning was used. To take care of the combat module, influence maps were used. To predict the match result we used recurrent neural networks. The STARDATA database was also used, which contains information on more than 6500 Starcraft games. The prediction module was able to obtain an accuracy between 67% and 86% according to the game time. Also, the strategic, economic and combat modules were more accurate than the works we used as reference. The intelligent agent competed against other agents in the AIIDE 2017 competition and it was observed that it manages to adapt to different situations in the game.