Mapping Explanation-based Learning onto Soar: the Sequel

In past work, chunking in Soar has been analyzed as a variant of explanation-based learning. The components and processes underlying EBL have been mapped to their corresponding components and processes in chunking. The cost and generality of the resulting rules have also been compared. Here we extend that work by analyzing an implementation of EBL within Soar as a sequence of transformations from a problem solving episode to a learned rule. The transformations in this sequence, along with their intermediate products, are then evaluated for their eeects on the generality and expensiveness of the rules learned, and compared with the results of a similar analysis previously performed for chunking. The analysis reveals that EBL, as implemented for Soar, yields the same sources of expensiveness and overgenerality as does chunking | and that, in fact, these problems stem more from other aspects of Soar than from the details of either learning algorithm. (However, some of these aspects may appear in other AI architectures, even though the analysis is based on Soar.) Moreover this analysis reveals that the diierences between EBL and chunking are localized within a single transformation, where chunking overspecializes with respect to EBL.