Abstract: "The impact of Explanation-Based Learning (EBL) on problem-solving efficiency can vary greatly from one problem space to another. Although EBL has been tested experimentally, no theory has been put forth which predicts and explains when EBL will work and why. This dissertation present...

Ausführliche Beschreibung

Bibliographische Detailangaben

1. Verfasser:	Etzioni, Oren (VerfasserIn)
Format:	Buch
Sprache:	English
Veröffentlicht:	Pittsburgh, Pa. School of Computer Science 1990
Schriftenreihe:	School of Computer Science <Pittsburgh, Pa.>: CMU-CS 1990,185
Schlagworte:	Machine learning Problem solving Maschinelles Lernen Erklärungskomponente Hochschulschrift
Zusammenfassung:	Abstract: "The impact of Explanation-Based Learning (EBL) on problem-solving efficiency can vary greatly from one problem space to another. Although EBL has been tested experimentally, no theory has been put forth which predicts and explains when EBL will work and why. This dissertation presents such a theory. The theory makes use of Problem Space Graphs (PSGs), a graph representation of problem spaces that makes their structure explicit. The theory consists of four hypotheses: The Nonrecursive Hypothesis: EBL is effective when it is able to curtail search via nonrecursive explanations. The Recursive Hypothesis: EBL is not effective, on sufficiently uniform problem distributions, when it relies on recursive explanations The Structural Hypothesis: EBL's explanations are isomorphic to PSG subgraphs. The Complementary Target Concepts Hypothesis: Target concepts whose proofs do not mirror the problem solver's plans are essential for EBL's effectiveness in recursive problem spaces. When applied to PRODIGY/EBL, a state-of-the-art EBL system, the theory yields a number of surprising predictions. For example, the recursive hypothesis predicts that PRODIGY/EBL will be foiled in a representational variant of the Blocksworld which has been stripped of its nonrecursive explanations. The structural hypothesis suggests that a static analyzer of problem space definitions can match EBL's performance To test this prediction, a program (called STATIC) was written that extracts nonrecursive explanations from problem space definitions by constructing and analyzing PSGs. Unlike PRODIGY/EBL, STATIC does not require training examples, utility evaluation, or rule compression. STATIC was tested in each of PRODIGY/EBL's problem spaces. In the experiments, static generated control knowledge from twenty six to seventy seven times faster than PRODIGY/EBL, and was up to three times as effective in speeding up PRODIGY.
Beschreibung:	Zugl.: Pittsburgh, Pa., Univ., Diss., 1990
Beschreibung:	X, 193 S. graph. Darst.