Michael L Donnell

Human cognition and the expert system interface: mental models and inference explanations

Variables affecting user/expert system interaction are evaluated empirically, and theoretical development of the relationship between human cognition and the use of intelligent machines is addressed. The hypotheses are that a good mental model will lead to increased user/computer interaction and performance and that graphic inference explanations will lead to higher performance than textual explanation. The examination of these hypotheses, as well as the cognitive processing underlying them, contributes directly to expert system interface design principles, as well as to theoretical development of the role of mental models in the understanding and use of dynamic, complex systems. Based on Newell and Simons (1972) theory of problem solving and the concept occurring in a problem space, human interaction with expert systems is quantified, relevant search strategies and directions are identified, and a theory of mental models of complex, dynamic systems is developed. The basic findings are that the user must have a good understanding (i.e., good mental model) of the expert systems reasoning process, and that the user must effectively understand the information present in the expert systems inference explanations. >

The role of task properties in determining the relative effectiveness of multiattribute weighting techniques

Abstract Most studies comparing the relative effectiveness of multiattribute weighting techniques have found few differences. These studies have, however, failed to systematically vary attribute properties such as number of attributes and distribution of correct attribute weights. It was hypothesized that any weighting technique would be more effective at arriving at the correct weights the smaller the number of attributes because of lessened information processing requirements. In addition, it was hypothesized that the relative effectiveness of different weighting techniques would depend on the peakedness of the distribution of correct attribute weights because different weighting techniques should generate more peaked distributions than others. Two experiments were conducted to test these hypotheses. The first experiment focused on the accuracy of the weights assigned to attributes by individuals; the second on the accuracy of groups. Both experiments confirmed the first hypothesis regarding the number of attributes, but only the first experiment confirmed the second hypothesis regarding the peakedness of the distribution of attribute weights.

Intrasubject comparison of four response modes for “subjective probability” assessment☆

Abstract Despite much research on subjective probability or personal uncertainty (PU) little is known about how well various measures of PU agree on a within subject basis. Thirty Subjects made odds (0), probability (P), bets (B) and null datum (ND) responses to indicate their opinion about which of two normal distributions was being sampled. O and P responses were made verbally. For the B responses subjects played an active part in a betting procedure which allowed PU to be inferred. In the ND condition subjects responded by offsetting the probabilistic effect of a sample of two data with a third datum. PU was inferred from the offsetting datum. On both group and individual analyses the O , P , and B modes agreed well with one another and the ND mode differed significantly only from the P mode. Subjects may have used strategies to make their estimates but if so the strategies were in remarkable agreement.

The effect of Bayesian feedback on learning in an odds estimation task

Abstract Numerous studies have compared human and optimal uncertainty revision, but few have considered learning and transfer effects. The present study provided subjects with 60 feedback trials in an odds revision task using normal distribution data generators. Learning and transfer were tested by comparing pretraining and posttraining odds estimates for data generators differing in diagnosticity (higher and lower) and kind (binomial) from those used in training. Subjects showed rapid learning and a moderate amount of transfer. What subjects seem to do after training is to increase their initially too small odds by some factor related to the perceived diagnosticity of the data generators. The fact that transfer of training is so closely tied to the actual stimuli used during feedback poses problems for training operators of real world diagnostic systems.