Computational stability of expert systems

It has been shown that propositional expert systems are equivalent to decision tables, and therefore equivalent to classification systems. In many cases, the elementary facts for the classification may not be accurately known. Even if they are, frequently the expert system reasons on the basis of qualitative descriptors of quantitative measurements, which may be subject to borderline effects. This paper considers the computational stability of the classification in the presence of errors in the data, using concepts derived from error-correcting codes, in particular Hamming distance. It suggests a number of methods of analysis of the decision table to identify potential instabilities, and suggests methods of correcting or avoiding these problems.