Extraction of compact rule sets from evolutionary designed artificial neural networks

While artificial neural networks (ANNs) have proven their almost universal applicability in a broad variety of problem domains, many scientists (theorists and practitioners as well) are still worried by the opacity of neural problem solvers. As well a network may perform, it is often desirable, if not necessary, to know at least some general concepts the network bases its decisions upon. During the last years a number of ANN rule extraction (RE) algorithms have been proposed addressing the problem of ANN opacity. Most RE approaches demand specific preconditions on ANN type, structure, or training methods which often prohibit the use of these algorithms for extracting rules from ANN architectures generated by evolutionary algorithms (EAs). However, evolved ANN topologies with structures no human designer would consider might represent problem knowledge very efficiently. Therefore, we utilize two RE methods having (almost) general purpose properties, namely MofN and VIA (validity interval analysis), for the extraction of rules from evolved generalized multilayer perceptrons being trained by error-back-propagation. In order to derive compact rule bases, ANN evolution is governed by a fitness function favoring networks of low complexity without loss of accuracy. We compare ANN classification with the performance of the extracted rule bases, and analyze network and rule structures by solving the MONK's problems which have become an RE benchmark.