Certainty Factor Algebras: Comparing Conventional Mappings and Experimental Results

Among the various means that have been advanced for handling uncertainty in expert systems, certainty factor CF approaches have been widely adopted in both individual expert system applications and shells for building such systems. Nevertheless, CF approaches have been challenged on analytical grounds. This paper examines the descriptive efficacy of CF approaches from an empirical standpoint. Specifically, it compares conventional CF methods for certainty-combining situations in a rule's premise with experimentally observed behaviors. The experimental data are subjected to 'hit' analyses that present the percentages of experts adhering to conventional algebraic means for combining CFs. The findings indicate that anywhere from about one-fourth up to nine-tenths of subjects combine certainties in a way consistent with CF algebras conventionally mapped into various certainty-combining situations. Put another way, conventional CF algebra mappings are inadequate for describing anywhere from 10% to 75% of human behaviors across the situations studied here. This empirical evidence suggests that knowledge engineers should be cautious about adopting conventional CF algebras, particularly under certain circumstances identified in this study. It also suggests the prudence of considering uncertainty treatments other than conventional certainty factor algebras.