Non-Monotonic Reasoning with Connectionist Networks using Coarse-Coded Representations

Abstract

This paper, describes a connectionist fault-tolerant non-monotonic reasoning system, which uses coarse-coded distributed representations. Distributed representations are known to give the advantages of fault tolerance, generalization and graceful degradation of performance under noise conditions. A semantic network is designed, using a novel approach, with connectionist networks using coarse-coded representations to perform non-monotonic reasoning. The system performs non-monotonic reasoning using the property of inheritance. The system also supports the feature of cancellation of inheritance, whereby more specific information associated with the nodes lower in the ‘isa’ hierarchy is given precedence over default information associated with the nodes higher in the hierarchy. System has exhibited good generalization ability on unseen test inputs. System’s performance with regard to its ability to exhibit fault tolerance under noise conditions is also studied. The system offers very good results of fault tolerance under noise conditions.