Evaluation of the Utility of Control Rules

Evaluation of the Utility of Learned Control Rules

Since PRODIGY learns control rules via Explanation-Based Learning, it must address the utility problem. PRODIGY purposely looks for very useful control rules that have low match cost. To do this, PRODIGY uses the following utility metric for evaluating a control rule:

Utility = (Average-Savings x Application-Frequency) - Average-Match-Cost

This metric addresses the three factors that can attribute to performance degradation via EBL. When learning a new rule, PRODIGY estimates the rule's utility based on a comparison of match cost versus the cost of directly exploring the search tree. If the rule appears to be useful, it is learned and included among the control rules. Statistics are then kept to determine the continuing usefulness of the rule. If the rule's utility becomes negative, the rule is deactivated and removed from the active control rules. In PRODIGY, this is known as selective forgetting.

Press UP to return to the property list.

Press HOME to return to the top of PRODIGY.