Improving expert system performance in solving problems with incomplete information

This dissertation determines the improvement in performance by an expert system expanded with deep knowledge over a heuristic based expert system when complete information for diagnosis is not available. The two expert systems are validated by Turing test and the results are compared statistically. There is a statistical difference in the mean scores for accuracy, usability, sensitivity, robustness and overall performance between the quality control supervisor and the two expert systems. Therefore, the two expert systems are less able to diagnose real world problems than the quality control supervisor for visual defects in coated paper products. There is no significant difference between the two expert systems in the five validation categories for the preplanned statistical comparisons. There is statistical evidence to show an improvement in the accuracy scores of the expanded expert system over the heuristic expert system. The scores for the expanded expert system are most noticeably higher when deep knowledge is applied to diagnose a defect.