The authors introduce the subject of their paper by distinguishing between the models of decision making designed by knowledge engineers and psychologists, which are models of actual human cognitive processes (descriptive), and the models of decision making designed by operations researchers, such as optimality strategies (prescriptive). It has been found that both types of models outperform experts.
An extensive discussion of descriptive models ensues. Such models can be developed from an expert’s verbal reports--which are often incomplete, inaccurate, or poorly articulated--or from statistical techniques applied to a sample of actual expert judgments. Stewart and McMillan prefer the latter technique, called judgment analysis, and elaborate on it: the information available to the experts becomes the independent variables, the decision is the dependent variable, and the technique typically applied is multiple regression analysis. An extensive and interesting example is then given. However, the authors seem unaware of inductive inference techniques for debriefing experts, techniques that are in some central ways similar to the statistical techniques they discuss [1,2].
A brief discussion of prescriptive models of inference follows. These models use the analytic techniques of operations research (OR) to obtain optimal decisions with respect to quantitatively given criteria. A brief digression follows in which the possibility of “managerial robots” is discussed together with some difficulties.
The authors conclude that the best approach is to use both judgment analysis in conjunction with verbal description by the expert and OR techniques. Thus one gets the benefits of each technique, while avoiding its pitfalls.
The paper is interesting, particularly the material on judgment analysis. It fulfills the purpose implicit in its title and is reasonably well written. It will interest both knowledge engineers and cognitive scientists, as well as AI workers in expert systems and learning. No mathematical knowledge is needed, although the rudiments of basic statistical techniques would be helpful background material.