Characterising Domains for Constructive Expert Systems

Expert systems technology arose initially from problems of diagnosis or pattern recognition but there have been fewer examples of successful synthesis systems. Constructive expert systems address such problems, particularly the problem of design. Two example expert system are presented, one of which (DOLMEN) supports the preliminary design of high rise buildings, and the other (JSD Advisor), which prescribes the design of a software system in a very narrow domain. DOLMEN uses a generate and test approach and incorporates an overall evaluation function which is used to limit the number of alternatives generated. The JSD system uses a generic domain model which is specialised during interactive dialogue with the user. The construction and features of the two systems are then analysed so as to identify the effect of domain characteristics on the type of expert system that can be built. It is argued that the critical dimensions are those of novelty, standardisation and monotonicity. The degree of novelty varies from highly creative to routine ; standardisation reflects the maturity of the design discipline and monotonicity determines the degree to which backtracking must be expected and supported. Whereas DOLMEN’s domain seems well suited to design support, effective expert systems in the field of software design will require standard object libraries and, even then, will concentrate on relatively mundane applications for the foreseeable future.