Refining classifiers with neural networks

Abstract

This article is devoted to presentation of two approaches to the task of combining classification using decision rules with that of neural networks. Main stress is put on applying the knowledge derived from data with methods coming from area of rough sets and Boolean reasoning to construction of feedforward artificial network architecture. Such an approach allows for both easier construction of neural network based decision support system and better classification results for previously unseen cases. One more paradigm presented within the paper is the application of a simplified neural network to the task of resolving conflict that occur in rule-based decision support systems. The approaches presented in the paper are illustrated with examples and results of actual numerical experiments.