Building Knowledge Repositories with Neural Networks and Orthogonal Designed Bases of Expert Holistic Judgments

Abstract

This paper proposes and illustrates a framework of building knowledge
repositories with Neural Expert Systems (NES) to overcome difficulties often
encountered by conventional experts systems technologies. The proposed
system uses Neural Networks (NN) to capture decision patterns / production
rules in a set of holistic judgments provided by experts on a minimal sample
of cases taken from a problem domain. A NN learns tacit knowledge from
implicit relationships between decision attributes and outcome. Holistic
judgments overcome the difficulty of explaining explicitly production rules and
heuristics of the experts. An orthogonal plan defines a minimal sample to
47
acquire the initial training set for NN and to alleviate experts from the cognitive
burden of specifying a complete set of production rules. Starting from this
initial knowledge base, counter-examples given by experts when the system
is in production are added to subsequent training. The knowledge base of a
NES and consequently the knowledge repository will grow over time with
additional patterns learning from its own production and expert opinions.