Fuzzy Naive Bayesian classification in RoboSoccer 3D: A hybrid approach to decision making

Abstract

We propose the use of a Fuzzy Naive Bayes classifier with a MAP rule as a decision making module for the RoboCup Soccer Simulation 3D domain. The Naive Bayes classifier has proven to be effective in a wide range of applications, in spite of the fact that the conditional independence assumption is not met in most cases. In the Naive Bayes classifier, each variable has a finite number of values, but in the RoboCup domain, we must deal with continuous variables. To overcome this issue, we use a fuzzy extension known as the Fuzzy Naive Bayes classifier that generalizes the meaning of an attribute so it does not have exactly one value, but a set of values to a certain degree of truth. We implemented this classifier in a 3D team so an agent could obtain the probabilities of success of the possible action courses given a situation in the field and decide the best action to execute. Specifically, we use the pass evaluation skill as a test bed. The classifier is trained in a scenario where there is one passer, one teammate and one opponent that tries to intercept the ball. We show the performance of the classifier in a test scenario with four opponents and three teammates. After a brief introduction, we present the specific characteristics of our training and test scenarios. Finally, results of our experiments are shown. © Springer-Verlag Berlin Heidelberg 2007.