Uncertainty handling and cognitive biases in knowledge engineering

The extent to which cognitive biases have an impact on knowledge engineering, particularly in the area of uncertainty handling, is explored. Subjects wrote rules to predict the color of objects drawn from an urn, based on several indicators. They also assigned confidence values to their rules. This task, though admittedly simplistic, was taken as representative of a large class of diagnosis and prediction problems. The results suggest that, using an automated aid, subjects can write a highly accurate, nearly unbiased set of rules that will predict event type based on three indicators: weight, size, and shape. Estimates of certainty extracted from rules generated by subjects using a prototype automated knowledge elicitation tool, as well as certainties that were later directly estimated, are shown to be highly correlated with the true probabilities for each event type. However, both types of certainty estimates were significantly less than the actual baseline probabilities. Also, in contrast to some results in the judgment and decision making literature, there was no evidence that the subject´s judgements or rules could be closely approximated with a linear model