EBL in Prodigy

EBL in Prodigy

EBL produces control rules by examining the complete problem-solving trace generated by the planner. This process is divided into two steps. First, knowledge gained from a trace is compressed into control rules. Second, these rules are evaulated to determine their utility. Utility is expressed as the rule's value in probable search-reduction minus the probable matching cost of the rule. Only rules with a high utility are stored. If the evaulation is accurate, a stored control rule will be easy to match, and will save search time. The purpose of this evaluation is to keep only those rules that will most likely result in faster problem solving when utilized in future planning.

Learning from failure is an important capability of EBL. Because EBL has access to the complete trace, it can exploit failures as well as successes. Dead ends in the search tree may provide EBL with knowledge it can compress into control rules.


Return to the top of this architecture.

Go to a discussion of this capability for multiple architectures.


Current Location: Prodigy-Capabilities-Multi-Method Learning-EBL

Go to NEXT page.