THE REFINEMENT OF PROBABILISTIC RULE SETS FOR CLASSIFICATION EXPERT-SYSTEMS - THE COMBINED OPTIMIZATION METHOD

Expert shells that support classification problem solving usually allo w the use of an uncertain reasoning method that assumes some degree of conditional independence between observations. As a consequence, prob abilistic rules can behave in an undesirable manner and thereby decrea se the expert system's performance with respect to classification accu racy. This paper describes one type of undesirable phenomenon, called sociopathicity, and describes results related to the induction and ref inement of probabilistic rule sets with this property. A knowledge bas e is said to be sociopathic if additions to the knowledge base degrade problem solving performance independent of computational resources. T he paper then presents an improved method of minimizing the error rate for sociopathic probabilistic rule sets. The refinement method, calle d Socio-Reducer2, is based on static and dynamic aspects of the probab ilistic rules. Experimental results in a medical diagnosis domain show that it can reduce the diagnosis error rate by a reasonable margin.