AUTOMATIC KNOWLEDGE BASE REFINEMENT FOR CLASSIFICATION SYSTEMS

被引：46

作者：

GINSBERG, A ^{[1
]}

WEISS, SM ^{[1
]}

POLITAKIS, P ^{[1
]}

机构：

[1] RUTGERS STATE UNIV,DEPT COMP SCI,NEW BRUNSWICK,NJ 08903

来源：

ARTIFICIAL INTELLIGENCE | 1988年 / 35卷 / 02期

关键词：

The knowledge acquisition problem can be divided into two phases. In phase one the knowledge engineer extracts an initial rough knowledge base from the * This research was supported in part by the Division of Research Resources; National Institutes of Health; Public Health Service; Department of Health; Education; and Welfare; Grant P41 RR02230. * Present address: AT&T Bell Laboratories; Holmdel; NJ. ** Present address: Digital Equipment Co; Hudson; MA. Artificial Intelligence 35 (1988) 197-226 0004-3702/88/$3.50 © 1988; Elsevier Science Publishers B.V. (North-Holland) expert; rough in the sense that the overall level of performance of this knowledge base is usually not comparable to that of the expert. In the second phase; the knowledge base refinement phase; the initial knowledge base is progressively refined into a high performance knowledge base. In terms of a rule-based knowledge base; phase one involves the acquisition of entire rules; indeed entire sets of rules; for concluding various hypotheses. The refinement phase; on the other hand; is characterized not so much by the acquisition of entire rules but by the addition; deletion; and alteration of rule-components in certain rules in the existing knowledge base; in an attempt to improve the system's empirical adequacy; i.e; its ability to reach the correct conclusions in the problems it is intended to solve. Obviously the foregoing description of knowledge base construction is an idealization. In practice the line between these two phases is not as sharply drawn. 1 A knowledge base refinement problem can be thought of as an optimization problem in which we start with a proposed general solution to a given set of domain problems and the goal is to refine it so that a superior solution is obtained. The proposed solution is a working knowledge base that is in need of minor adjustments; but not a major overhaul; one assumes that the rules given by the expert are basically sensible propositions concerning the problem domain. The refinements applied to the rules of this knowledge base must not only meet the obvious requirements of being syntactically and semantically admissible; they must also be conservative; in the sense that they tend to preserve; as far as possible; the expert's given version of the rules. Employing rule refinements that meet these requirements makes it more likely that the construction of a refined knowledge base will not simply be a matter of curve fitting; but will result in a knowledge base with genuinely improved empirical adequacy; that at the same time remains close to the actual knowledge of the expert. Thus when we speak of optimizing the performance of a knowledge base; we mean improving performance on sample case data as much as possible; subject to constraints of conservatism;

D O I：

10.1016/0004-3702(88)90012-4

中图分类号：

TP18 [人工智能理论];

学科分类号：

081104 ; 0812 ; 0835 ; 1405 ;

摘要：

引用

页码：197 / 226

页数：30

共 21 条

[1] THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM - A FRAMEWORK FOR EXPLANATION [J].

CLANCEY, WJ .

ARTIFICIAL INTELLIGENCE, 1983, 20 (03) :215-251

[2] HEURISTIC CLASSIFICATION [J].

CLANCEY, WJ .

ARTIFICIAL INTELLIGENCE, 1985, 27 (03) :289-350

[3] INTERACTIVE TRANSFER OF EXPERTISE - ACQUISITION OF NEW INFERENCE RULES [J].

DAVIS, R .

ARTIFICIAL INTELLIGENCE, 1979, 12 (02) :121-157

[4]

Doyle R. J., 1986, Proceedings AAAI-86: Fifth National Conference on Artificial Intelligence, P538

[5]

Duda R. O., 1973, PATTERN CLASSIFICATI

[6]

Eshelman L., 1986, Proceedings AAAI-86: Fifth National Conference on Artificial Intelligence, P950

[7]

GINSBERG A, 1986, THESIS RUTGERS U NEW

[8]

GINSBERG A, 1986, P AAAI 86 PHILADELPH

[9]

Hall R. J., 1986, Proceedings AAAI-86: Fifth National Conference on Artificial Intelligence, P568

[10]

Kahn G., 1985, P 9 INT JOINT C ART, P581

← 1 2 3 →