Abstract

Artificial intelligence has traditionally favored a reductionistic approach, study-ing intelligent behavior by analyzing each component independently: Knowledge representation, search-intensive problem solving, knowledge-intensive exper-tise, concept acquisition from examples, performance improvement due to experience, and so forth. Such a divide-and-conquer approach has been historically quite appropriate, providing many useful results. However, artificial intelligence is evolving from an exploratory endeavor to a quantitative science, and part of the maturation process is the emergence of unifying theories and integrated computational architectures. This chapter describes one such investigation, the PRODIGY system, an integrated architecture unifying problem solving, planning and multi-ple learning methods. The learning methods in PRODIGY encompass learning control rules through explanation-based learning (EBL) and static search-space analysis, learning plan knowledge through analogical transfer, learning abstrac-tion hierarchies through domain-definition analysis, and acquiring new domain knowledge through goal-oriented experimentation and dynamic interaction with a human expert.
Before endeavoring to describe PRODIGY in depth, let us situate it in the space of integrated computational architectures. There are multiple dimensions one can use to contrast and compare the architectures. We list each dimension, situating PRODIGY and contrasting it to SOAR (Laird, Newell, & Rosenbloom, 1987; Rosenbloom, Newell, & Laird, chap. 4 in this volume), THEO (Mitchell et al., chap. 12 in this volume) and occasionally ICARUS …

Date

October 14, 1991

Authors

Craig A Knoblock

Journal

Architectures for Intelligence

Pages

241

Publisher

Psychology Press