Applying neural computing to expert system design: Coping with complex sensory data and attribute selection

Abstract

Recently the relation of subsymbolic (“neural computing”) and symbolic computing has been a topic of intense discussion. Our purpose is to focus this discussion to the particular application area of expert system design. We address some of the drawbacks of current expert systems and study the possibility of using neural computing methodologies to improve their competence. The topic can be discussed at various levels of integration: the higher the integration level, the more symbolic functionalities (such as an inference engine) are implemented directly at the level of the neural computational model. In this paper we address the lowest levels of integration: neural networks that can be used to implement feature recognizers which allow symbolic inference engines to make direct use of complex sensory input via so called detector predicates. We also introduce the notion of self organization as a means to determine those attributes (properties) of data that reflect meaningful statistical relationships in the expert system input space, thus addressing the difficult problem of conceptual clustering (“abstraction”) of information. The concepts introduced are illustrated by two examples: an automatic inspection system for circuit packs and an expert system for respiratory and anesthesia monitoring. The adopted approach differs considerably from the earlier research on the use of neural networks as expert systems, where the only method to obtain knowledge is learning from training data. In our approach the synergy of rules and detector predicates combines the advantages of both worlds: it maintains the clarity of the rule-based knowledge representations at the higher reasoning levels without sacrificing the power of noise-tolerant pattern association (“inference by memory”) offered by neural computing methods.