Extract intelligible and concise fuzzy rules from neural networks

The advent of artificial neural networks has contributed significantly to the field of knowledge engineering. Neural networks belong to a family of models that are based on a learning-by-example paradigm in which problem solving knowledge is automatically generated according to actual examples presented to them. The knowledge, however, is represented at a subsymbolic level in terms of connections and weights. Neural networks act like black boxes providing little insight into how decisions are made. They have no explicit, declarative knowledge structure that allows the representation and generation of explanation structures. Thus, knowledge captured by neural networks is not transparent to users and cannot be verified by domain experts. To solve this problem, researchers are interested in developing a humanly understandable representation for neural networks. This can be achieved by extracting production rules from trained neural networks. Current rule extraction approaches can successfully deal with problems with discrete-valued inputs but are less efficient when dealing with problems with continuous-valued inputs. This paper presents a novel approach to represent continuous-valued input parameters using linguistic terms (discretization) and then extract fuzzy rules from trained binary single-layer neural networks. An algorithm was developed to extract the most dominant fuzzy rules. The results are very simple rules that can achieve high predictive accuracy. The algorithm was applied to a couple of benchmark pattern recognition problems and a real-world manufacturing problem with promising results. (C) 2001 Elsevier Science B.V. All rights reserved.