Browse Prior Art Database

PATTERNS OF INDUCTION AND ASSOCIATED KNOWLEDGE ACQUISITION ALGORITHMS

IP.com Disclosure Number: IPCOM000147924D
Original Publication Date: 1976-May-13
Included in the Prior Art Database: 2007-Mar-28
Document File: 30 page(s) / 2M

Publishing Venue

Software Patent Institute

Related People

Hayes-Roth, Frederick: AUTHOR [+2]

Abstract

PATTERNS OF INDUCTION AND ASSOCIATED KNOWLEDGE ACQUSSTION ALGORITHMS~ Frodorick Hayea-Roth Computer Science Department2 Carnegie-Mellon University Pittsburgh, Pa. 15213 May 13,1976 The common need of both Artificial Intelligence and Pattern Recognition for effective methods of automatic knowledge acquisition is considered. A pattern of induction is defined as a framework which relates a theory of behavior generation, underlying knowledge structures, end e learning methodology. One particular learning theory, called interference matching, suggests that knowledge structures which underlie behavior descriptions can be directly abstracted from those &script ions. Because of the close connection between descriptions and inferences in such a framework, the strengths and weaknesses of several types of descriptions are considered. Algorithms which exploit this theory are presented for three classes of problems: pat tern learning and classification; induct ion of quantified production rules; and the induction of syntactic categories end phrase structure rules. Preliminary results are presented and directions for future research are outlined.

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 5% of the total text.

Page 1 of 30

PATTERNS OF INDUCTION AND ASSOCIATED KNOWLEDGE ACQUSSTION ALGORITHMS~

Frodorick Hayea-Roth

Computer Science Department2 Carnegie-Mellon University

Pittsburgh, Pa. 15213

May 13,1976

    The common need of both Artificial Intelligence and Pattern Recognition for effective methods of automatic knowledge acquisition is considered. A pattern of induction is defined as a framework which relates a theory of behavior generation, underlying knowledge structures, end e learning methodology. One particular learning theory, called interference matching, suggests that knowledge structures which underlie behavior descriptions can be directly abstracted from those &script ions. Because of the close connection between descriptions and inferences in such a framework, the strengths and weaknesses of several types of descriptions are considered. Algorithms which exploit this theory are presented for three classes of problems: pat tern learning and classification; induct ion of quantified production rules; and the induction of syntactic categories end phrase structure rules. Preliminary results are presented and directions for future research are outlined.

1 This paper will appear in the Academic Press publication of invited papers of the

  Workshop on Artificial Intelligence and Pattern Recognition held at Hyannis, Massachusetts, June 1 -3, 1 976.
2 This research was supported in part by the Defense Advanced Research Projects

Agency under contract no. F44620-73-C-0074 and monitored by the Air Force Offica of Scientific Research

[This page contains 1 picture or other non-text object]

Page 2 of 30

[This page contains 1 picture or other non-text object]

Page 3 of 30

Hayes-Roth

    Despite many impressive advances in the areas of knowledge representation and engineering, Artificial Intelligence (AI) has made virtually no progress on general learning problems in the last 20 years. Both A1 and Pattern Recognition (PR)

currently

experience a pressing need for automatic methods of knowledge acquisition, but their problems are somewhat different. Current efforts in A1 aimed at building large-scale knowledge-based systems (e.g., for speech, vision, text understanding) are virtually overwhelmed by the task of knowlcad~e enstineering. The goal of this task is the implementation of all potentially valuable "knowledge sources," problem solving modules which exploit the known physical, syntactic, contextual, and semantic relations to constrain the search for solutions. The cost--in terms of people, time, and machine resources--of translating this human knowledge into computer programs is nearly insupportable. Furthermore, even after these handcrafted knowledge sources are developed, they are difficult to evaluate comparatively because each tends to be "one- of-a-kind," a body of code specially tailored to operate in one specific system and to employ only one particutor subset of the many potentially relevant problem solving techniques. Thus, to o large e~tent, tho immediate need for general learning procedures in A1 is to automate much of tho work of knowledge progrrmming. The field of PR, on the other hand, needs general learning procedures becaus...