A method for integrating end-user preferences for design evaluation in rule-based systems

Expert systems for design often include provisions for comparison of preliminary design alternatives. Historically, this task has been done on an ad hoc basis (or not at all) due to two difficulties. The first difficulty is design evaluation of multiple attributes. The second is that of taking into account highly subjective end-user preferences. Design experts have developed techniques which have enabled them to deal with these two difficulties; weighted average methods for the former and heuristic "rules of thumb" which categorize end-users for the latter. Limitations of these techniques are that the accuracy and precision of weighted average methods is inadequate, and that the "rules of thumb" might be reasonable and valid for most end-users, but not for some others. This paper brings quantitative rigor to the modelling of end-user preferences which is equal to that used in other phases of engineering analysis. We present a technique by which a heuristic rule base derived from technical experts can be analyzed and modified to integrate quantitative assessment of end-users' subjective preferences. The operations research tool of multiattribute utility analysis is integrated with artificial intelligence techniques to facilitate preliminary evaluation of design alternatives of multiple attributes specific to individual users. The steps of the methodology are: develop the heuristic rule base, analyze the rule base to separate subjective from objective rules, add a subjective multiattribute utility assessment module, add an uncertainty assessment module, make objective rules explicit, and express performance attributes in terms of design decision variables. The key step is making the distinction between subjective and objective aspects of rules, and replacing the former with utility analysis. These steps are illustrated through an expert system for materials selection for a sailboat mast. Results indicate improved expert system performance for both "typical" and "atypical" end-users.