Pairwise Comparison of Hypotheses Coverings as a Natural Mean Against Undesirable Niching in Evolutionary Inductive Learning
Krzysztof Krawiec · 2001
This report summarizes the results of research on the use of evolutionary learning for solving pattern recognition problems. The general idea consists in evolutionary search in the space of pattern recognition programs. The whole body of results described here was obtained in the improved version of GPVIS environment [15]. In particular, this report describes the 2.0 version of the environment and is devoted in a great part to the extensions beyond the standard genetic programming introduced into GPVIS, including the novel method of hypothesis evaluation proposed for evolutionary learning. This work focuses on reasoning from pictorial information based on evolutionary computation, or, to be more precise, on the paradigm of genetic programming [12]. The outline of the method is as follows. The genetic search engine performs the search through the space of image processing and analysis programs. The programs have the form of expressions formulated in a specialized language called GPVISL (Genetic Programming for Visual Learning language). The genetic search engine realizes the selection of parent solutions (individuals), which are then crossed over and mutated to obtain the next generation of solutions. The selection is done w.r.t. the value of evaluation (fitness) function. A solution is evaluated by testing its behavior on a set of fitness cases, which are equivalent to images in this context. The fitness function is the percentage of hits[14], i.e. of the correct decisions (recognitions) made by the system. Table of contents 1