A Hybrid Rule-based System with Rule-reenement Mechanisms

In this paper we present HTR, a hybrid trainable rule-based system. The key features of the system include: the heterogeneous integration of multiple control regimes, rule induction and reenement mechanisms, easy and concise speciication of knowledge, treatment and support for uncertainty. All these features are obtained thanks to the introduction of a new formalism for rule speciication that extends the behaviour of a traditional rule based system. The formalism utilises predicates as an intermediate layer between the conditions and the actions of rules. Diierent kinds of predicates support various sorts of control, distributed connict resolution, and learning from examples. In particular we describe the characteristics of truth-table-based, collection-based and artiicial-neural-net based predicates which are pre-deened in HTR. We also report on a method for automatic rule induction using genetic algorithms. Experimental results exploiting the afore-mentioned features are described.