The representation of rules in connectionist models

Abstract

Three models of connectionist rule processing are presented and discussed: Shastri and Ajjanagadde's SHRUTI system, which uses connectionist mechanisms to overcome the inefficiency of classical rule-based models; Sun's CONSYDERR model, which combines rule-processing and similarity-matching in a two-level architecture; and Sumida and Dyer's Parallel Distributed Semantic (PDS) Network, which adds generalisation and inferencing capabilities to a semantic network. The Sun and Sumida and Dyer models are illustrated with implementations and example runs. Aspects of rule-based reasoning not addressed by these models (rule learning, the encoding of rules wuthout hard-wired structure, and holistic rule application) are discussed in the context of a proposed new model.