goldstn@evax.arl.utexas.edu (David Goldstein) (01/10/91)
A few comments would be in order on the discussion verifying and validating knowledge bases. While V&V procedures have not been AS FORMALLY developed in AI as they have been in other computing domains,knowledge engineers ARE expected to provide tracibility of the information in a system to the sources of expertise from which the information was derived. Validation, or demonstrating that the system does what the user wants, is facilitated by such tracability. Verification, or doing what the system is contractually specified to do, is a much more scientifically provable concept. Software engineers and knowledge-based systems programmers both work towards demonstrating the completeness and consistency of their products, as well as for tools to demonstrate how products satisfy requirements and specifications. Sure, if V&V is simply determining what experts are right and how their knowledge should be used, then on general problems - everyone has an opinion, so induction says the problems are impossible. Hence, V&V under such criteria, for projects such as Cyc, would be provably impossible!