ABSTRACT
This chapter describes a hybrid architecture for classification expert systems that combines semantic networks and neural networks for representing knowledge. A semantic network is used to describe the objects of the problem domain and their relations at the intensional and extensional levels (classes of objects and instances). The relation Influences is the most important in this semantic network and signals where evidential reasoning (defined here as the possibilistic evaluation of objects representing hypotheses, based on the possibility values of objects representing evidences) may be performed. The possibilistic evaluation is performed in fact by a fuzzy neural network based on the authors’ combinatorial neural model and genetic algorithms. This hybrid scheme allows the construction of fuzzy connectionist expert systems able to inherit desirable properties from both the fields of sub-symbolic neural networks and symbolic expert systems, such as: expert knowledge representation, integration of multiple expert knowledge sources, multiple problem domain views, heuristic learning from examples, incremental learning, feature selection, treatment of vague, imprecise and partial input data, cost conscious inquiry, and reasoning explanation.
