A structural theory of explanation-based learning

摘要

The impact of explanation-based learning (EBL) on problem-solving efficiency varies greatly from one problem space to another. In fact, seemingly minute modifications to problem space encoding can drastically alter EBL's impact. For example, while prodigy/ebl (a state-of-the-art EBL system) significantly speeds up the prodigy problem solver in the Blocksworld, prodigy/ebl actually slows prodigy down in a representational variant of the Blocksworld constructed by adding a single, carefully chosen, macro-operator to the Blocksworld operator set. Although EBL has been tested experimentally, no theory has been put forth that accounts for such phenomena. This paper presents such a theory.The theory exhibits a correspondence between a graph representation of problem spaces and the proofs used by EBL systems to generate search-control knowledge. The theory relies on this correspondence to account for the variations in EBL's impact. This account is validated by STATIC, a program that extracts EBL-style control knowledge directly from the graph representation, without using training examples. When tested on prodigy/ebl's benchmark tasks, static was up to three times as effective as prodigy/ebl in speeding up prodigy.