Developing Expert Systems for Validating Numerics

Developing Expert Systems for Validating Numerics. The selection of an appropriate method out of a software library and its application to a given problem require a lot of expert knowledge. Usually there are more than one mathematically equivalent self-validating methods with completely different behavior on a Computer. Therefore only a quantitative valuation of the available software modules rather than qualitative can lead to a satisfactory result. First the design of a shell suitable for the derivation of expert systems dealing with different domains of numerics is discussed. The main idea is the decoupling of the logic and the domain-specific part allowing an easy adaptation to a new field or extension with additional methods and selection criteria. Then a functional model for knowledge acquisition and representation under quantitative aspects is elaborated and enhanced by considering execution time problems arising from portations. The procedure is demonstrated by a comprehensive example dealing with two methods for systems of linear equations.