Uma arquitetura de uso geral baseada em planejamento probabilístico para agentes completos em jogos de estratégia em tempo real

• Autor(a) principal: Naves, Thiago França
• Data de Publicação: 2017
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da UFU
• Texto Completo: https://repositorio.ufu.br/handle/123456789/20111; http://dx.doi.org/10.14393/ufu.te.2017.13
• topic: POMDP; POMDP.; Tomada de Decisão; Decision Making; Planejamento; Planning; Mineração de Dados Sequencial; Sequential Pattern Mining; Ações; Actions; Jogos RTS; RTS Games; Inteligência Artificial; Artificial Intelligence; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO

Real-time Strategy games, also known as RTS games, are characterized by acting in a dynamic environment, with uncertainties and various resources to be managed. This genre of games becomes a great testbed domain for artificial intelligence (AI) algorithms, in particular using planning and decision-making approaches, which are active AI research topics. This work aims to propose the development of a complete player agent for RTS games. In order for the agent to be considered complete, there are several tasks that it must perform, such as: data modeling between disputed matches; decision-making under uncertainty; resource management; planning against the opponent in real time; scheduling of actions. Thus, for the complete implementation of a successful player agent, an integrative approach is needed, which manages such tasks at different levels of abstraction. Among the main works in the field of RTS games, there are few references that propose an integrative approach, since the vast majority use only techniques based on predefined scripts or conditional rules. Thus, this thesis proposes a new approach, based on probabilistic planning, for complete control of players agents in RTS. This approach is proposed under an architecture that operates with sequential data mining algorithms, prediction trees; partially observable Markov decision process (POMDP), reactive planning and scheduling of actions. The approach manages all the tasks of the game with compatible answers, considering the real-time restrictions of these games. To validate the proposal, experiments against other agents, human players, with performance and quality tests are performed, and their results discussed.