Análise e Previsão das Formações das Equipas no Domínio do Futebol Robótico.

Abstract

This study proposes a definition of one methodology of classification that let identify the formations of the teams, in domain of robotic soccer, in the simulation league of two dimensions (2D) league. To reach the goal of this study it was used techniques of Data Mining for classification problems. To explain the operation and the characteristics of robotic soccer simulated, with emphasis on multi-agent systems, is described: the constitution of the system simulation of soccer (football) with the respective rules, the communication between the simulator and the players and the respective protocols, the perceptions and agents actions, the heterogeneous players, the coach agent, their functions and their language of communication. Posteriorly, is presented the stages of Data Mining process: data preparation, data reduction, modeling and solution analysis, In this work the first stage data preparation presented: the selection of the test teams, the configuration of the simulation environment in Linux, the configuration of FC Portugal team, used in this study, and their training in order to make a game of robotic soccer simulated with ten different formations. After the completion of the six games, using four distinct teams was made the conversion of the log files, of these games, in a dataset with the typical format (matrix form). In the second stage was carried out the data reduction of the attributes in the empirical way, based on the knowledge of formations process in the real world soccer and in the robotic soccer simulated. In modeling were selected too in the empirical way, the classifiers with potential to produce the best forecast model of the formations. In the stage for solution analysis, the main indicators for evaluation were the error rate and the statistical test t-Student for paired samples. The results in the set of experiments demonstrated that it was possible to identify, with great accuracy, the formations used by the team FC Portugal in distinct games using techniques of Data Mining. Analysing the results it is possible to deduce that the classifiers Sequential Minimal Optimization (SMO) and the k-Nearest Neighbor (IBK) obtained the best performance in the experiments performed. Finally it was concluded that the most appropriate classifier to generate a forecast model before the games in robotic soccer simulated is the SMO.