Using the Critiquing Approach to Cope with Brittle Expert Systems

Abstract

We have conducted a series of studies aimed at understanding how to design cooperative problem-solving systems to deal with situations in which the computer is not fully competent (the “brittleness” problem). Results from an initial empirical study showed that an expert system acting as an automated assistant induced new errors in such situations, but that this did not occur when the system was designed to critique user's performance on the task. Guided by these results, a more complete critiquing system was then designed and evaluated. On cases where the computer was fully competent, outcome errors were completely eliminated (a 30–60% reduction in errors). On a case for which the system was brittle (less than fully competent), misdiagnosis rates were still reduced by 31%, giving an overall significance of p < 0.000003 across all Post-Test Cases, further supporting our initial study regarding the potential value of critiquing as an effective role for an expert system. A detailed analysis of the influence of performance further indicates the ways in which different classes of critics detect errors and influence the user's problem-solving.