Book Description

"This report describes the second experiment in a series devoted to estimating the effectiveness of automated hypothesis selection in man-machine systems in which threat evaluations or threat diagnoses are being performed. In the experiment an eight-man team produced evaluations of various threats posed by a hypothetical aggressor. These evaluations were made on the basis of intelligence information gathered on simulated reconnaissance overflights of the homeland area of the aggressor. IBM 1401 and 7090 computer facilities provided the means for generating the complex stimulus environment or data base. The primary output from this threat evaluation team was a series of a posteriori probabilities estimations produced by the team's commanding officer (CO). These estimations represented the CO's judgments as to the most likely of the four response alternatives available to the aggressor in deploying his forces along a border of contention. In three of the four experimental conditions the CO was provided with a hypothesis-selection aid based upon a modification of Bayes' theorem (MBT). In these three conditions the CO was permitted to exert an increasing amount of control over the MBT-aid mechanism. He exerted control either by adjustment of certain parameters in the MBT model or by direct insertion of conditional probabilities into the model. The purpose of the experiment was to observe whether increasing control over the MBT-aid mechanism would increase the user's acceptance of the aid and improve his threat-diagnosis performance. The CO's threat-evaluation performance did improve during the course of the experiment but independently of the MBT-aid configuration. Throughout the experiment, solutions of a posteriori probabilities based upon the MBT were calculated by the experimenter for comparison with the human estimates. These two sets of estimates were strikingly similar. The CO's estimates, although quite conservative in early trials, became noticeably less conservative as the experiment progressed. The overall difference between the accuracy of the CO and MBT estimations was negligible."--Abstract