D-MA-Draughts: um sistema multiagente jogador de damas automático que atua em um ambiente de alto desempenho

• Ano de defesa: 2013
• Autor(a) principal: Tomaz, Lídia Bononi Paiva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; BR; Programa de Pós-graduação em Ciência da Computação; Ciências Exatas e da Terra; UFU
• Palavras-chave em Português: Sistema jogador multiagente distribuído; Damas; Aprendizagem por reforço de máquina; Redes neurais artificiais; Busca distribuída; Representação de estados por características; Algoritmos de clusterização; Tabelas de transposição; Jogo de damas por computador; Redes neurais (Computação); Distributed multi-agent player system; Draughts; Machine reinforcement learning; Artificial neural networks; Distributed search; Representation of states through features; Cluster algorithms; Transposition tables; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufu.br/handle/123456789/12544; https://doi.org/10.14393/ufu.di.2013.255

The objective behind this study is the proposal of a Draughts learning system, D-MA-Draughts (Distributed Multi-agent Draughts): a multi-agent Draughts player that operates in high performance environments. D-MA-Draughts is based on the integration and refinement of two other successful automatic Draughts architectures: MP-Draughts and D-VisionDraughts. The first is a multi-agent system, which does not operate in a high performance environment, and the second is a mono-agent system, which operates in a high performance environment. The D-MA-Draughts player, in particular aided in enhancing the MP-Draughts player with respect to the dynamics of the interaction between the agents. Similarly, it increases the performance of D-VisionDraughts by augmenting the accuracy with which it represents the boards along with an increase in the number of processors pertaining to the environment. The D-MA-Draughts player is composed of 26 agents, with one denominated as IIGA (Initial Intermediate Game Agent), which is a specialist in initial and intermediary game states and 25 agents specializing in end game states. Each of these agents consists of a multi-layer neural network (MLP), which was trained through the use of Temporal Difference Methods (λ), in a distributed processing environment, in order to achieve a high performance level with the minimum of human intervention possible, unlike the automatic player of Draughts world champion, Chinook. The search for the best move is guided by the distributed search algorithm denominated as, Young Brothers Wait Concept (YBWC). The board states are represented by NET-FEATUREMAP mapping, which is composed of functions that describe features inherent to the game of Draughts. The specialist agents in the end game states were trained to deal with a specific type of board profile\" of end game states (cluster). These clusters were obtained through Kohonen-SOM networks from a data base with end game board states obtained from real life matches. Once trained, the D-MA-Draughts agents are able to operate in a match by following two dynamics of the game. In both of these dynamics the IIGA agent will start the match. From this point they will vary as follows: in the first, once the board state characterized as the end game has been reached the Kohonen-SOM network will appoint the agent (from within the end game agents), which represents the cluster whose profile\" is closest to the current board state. From here on, agents of this type will substitute the IIGA and conduct the match until the end. In the second dynamic, each time that the automatic system executes a movement in an end game situation, the Kohonen-SOM network will be activated to indicate, which of the end game agents has the most adequate profile\" for the current end game state. This study shows that by using the proposed architecture D-MA-Draughts managed to improve the performance achieved by its predecessors. Such improvements are due to the greater accuracy with which the board states are perceived by the agents allowing them to make more precise decisions. On the other hand, improving the board representation leads to an increase in the system\'s processing load. In this context, the increase in the number of processors circumvented this situation besides contributing to a deeper player future vision or look-ahead during the search, which made it more suitable in choosing the best moves.