User Interaction with Self-Learning Systems

This research investigated how users interact with an expert system in which underlying values change as a function of the situation or of the planning time horizon. The problem context was the prioritization of tactical air strike targets, and Air Force targeteers were the experimental subjects. Their task was to explain why the system made the recommendations it did. The expert system was simulated in storyboard form. It was found that users who were given a good conceptual model of the expert system, in the form of a brief summary of its step-by-step processes, performed better than those whose model of the system was relatively poor. Users whose displays were relatively user-oriented (geographic, top-down, and simplified) did not consistently differ in performance from users whose displays were relatively aid-oriented (data-intensive matrices). However, the aid-oriented displays seem to encourage users to make more frequent reference to three tables (available to all subjects) containing important information about the expert assessments on which the aid's algorithms operated. Users would typically generate their own solutions, using criteria and heuristics that often differed from those used by the aid, and question the expert system solution. Keywords: Training devices.