Aprimorando o processo de aprendizagem e alocação de agentes inteligentes em plataformas multiagentes: aplicação no domínio do jogo de damas

Abstract

One of the fundamental requirements for a unsupervised multiagent system to reach its objectives is that the agents that make up the system possess specific and complementary abilities, which allow them to act as specialists in the environments where they were trained. The adequate representation of these environments is fundamental to both the learning and to the good performance on the part of the agents, mainly when these act in competitive environments that possess an elevated state space. Likewise, the decisions from multiagent systems, through their allocation of adequate agents into particular situations that occur in these environments, are crucial in order that these successfully reach their objectives. In this sense, the present work presents three new approaches to optimize the performance of multiagent systems, which improves: the architecture and the learning process of the agents that make up the multiagent system; the representation of relevant information of the environments where these agents perform, as well as the process of allocating the adequate agent for performing in distinct situations that occur in these environments. Due to the spatial and technical complexity, the game of Checkers was used as the developmental and evaluative environment for these approaches, which were implemented onto the automatic player MP-Draughts. This player corresponds to a unsupervised multiagent system composed of specialist player agents in distinct phases of a game. In order to implement the proposed approaches onto the MP-Draughts architecture, the following work sequence was adopted: initially, an adaptive neural network was developed, ASONDE, which was used in the MP-Draughts architecture to define the knowledge profiles (clusters) necessary for representing the endgame phase, on which the specialist agents should be trained. Following on, an automatic features selection approach based on the frequent pattern mining was implemented, which extracts the most adequate features to represent the different environments (boards) that can occur during the performance of the multi-agent. Finally, a method for the allocation of agents was developed, which combined clustering artificial neural networks and exception rules, which together are responsible for indicating the most suitable agents to act in the different situations of a game. The partial results obtained from the implementation of each approach, as well as the final result, which applies all these into the MP-Draughts architecture, confirm that these were efficient in dealing with the problems for which they were proposed, in addition to contributing to the general performance of the multi-agent system.