Neural-networks designed on Approximate Reasoning Architecture

Abstract

We propose Neural-networks designed on Approximate Reasoning Architecture (NARA), and show that it becomes easy to debug network in terms of the rule structure and to improve the performance. It is important to implement the knowledge into artificial neural network (NN) for improving its performance. Most conventional knowledge implementations had very few possibility of help using relatively high level knowledge structure. They were mainly concentrated to statistical analysis of input data or prewiring. NARA is constructed by approximate reasoning as framework and by NNs as component. Because of the high level knowledge structure (approximate reasoning architecture), NARA has many advantages. The most advantageous point of NARA is to be able to reconstruct the structure or component NNs easily to improve its performance. NARA is constructed by small NNs that correspond to approximate reasoning rules -- this enable us to analyze the relationship between NN structure and its performance, and can modify the small NNs or NN structure related to the performance. The second point is to shorten the learning time of NN. We show the algorithm for model construction and evaluation by simulation.