This study elaborates on the expert system implementation process outlined by Assad and Golden [ 1, section 41. In particular, we expand on the issue of selecting an approp~ate microcomputer-based expert system generator for a given problem (Step 2). Before taking this step, it is important to decide which, if any, of the existing knowledge processing methods is appropriate for the problem at hand. As Assad and Golden emphasize, there are several methods for constructing expert systems with currently available microcomputer tools: rule-based deduction, statistical pattern classification, etc. It is natural to wonder which of these methods is “best.” Several performance studies have been done to shed light on this issue. These studies have typically implemented multiple expert systems for the same application problem, using a different method for each system. The accuracy of each expert system’s predictions was then measured for the same set of test cases. Examples include comparative evaluation of different statistical pattern classi~~tion techniques [2,3], rule-based deduction versus statistical pattern classification [4], branching logic versus statistical pattern classification [5], and rule-based deduction versus prototype/description-based hypothetico-deductive inference [6,7]. The striking feature of these studies has been the repeated failure to demonstrate a clear performance advantage of one method over another. Even where minor differences do exist, it can be argued that they are attributable to differences in the problem-specific information in the knowledge bases as opposed to fundamental differences in the methods being used. In spite of such empirical findings, some authors remain unconvinced: “. . . pattern directed inference systems based on antecedent-consequent rules make a strong claim to being the best available scheme for knowledge representation . . .” [8]; or, “There is only one language suitable for representing information, whether declarative or procedural, and that is first-order predicate logic. There is only one intelligent way to process info~ation and that is by applying deductive inference methods” [9]. Such arguments are typically supported not only by examples of good performance by decision support systems, but also by arguments that the method in question has other advantages (ease of use, understandability, etc.). In contrast with these opinions, we would suggest that a more appropriate issue for consideration is nor which method is best, but rather which method is best suited for a specific problem There are several good methods, each with certain advantages and disadvantages, and the problem is really how one should select the most reasonable one to use in a given situation. There is very little practical information in the literature about this issue. However, certain observations can still be made about method selection. While these observations should be viewed as preliminary, they provide useful initial guidelines for the potential expert system developer, and they serve as a starting point for further discussion and research on this important topic. The first observation is that there are a variety of factors that go into selecting an appropriate approach to knowledge representation and processing. These factors include the pre-encoding format of the relevant knowledge, the type of classification that is desired, and the amount of context-dependent inherently present in a problem. One important selection factor is the form in which the knowledge is already organized, i.e., the pre-encoding format of the knowledge. That pre-existing format is often very natural for the particular application involved from the viewpoint of individuals working in that application area. Therefore, keeping the knowledge in the same form has a certain intuitive
Read full abstract
Jan 1, 1989
W.D Batchelor + 3
Open Access