Generation of good training data for extracting DTs from evolved NN robot controllers

Abstract

Neural networks (NNs) have been widely accepted as a good model of robot controllers. One reason is that NNs are good both for batch learning and for incremental learning. Batch learning is important for obtaining an initial controller using existing data, while incremental learning is useful for refining the controller using newly observed data. One drawback in using NN controller is that the knowledge learned by an NN is difficult to understand and to re-use. The goal of this study is to interpret an evolved NN controller using a decision tree (DT). For this purpose it is necessary to generate a good training set from which the most consistent DT can be induced. This paper introduces several simple methods for generating the training set. The efficiency and efficacy of these methods are verified through experiments.