The power of knowledge acquisition systems that employ failure-driven learning derives from two main sources: an effective global credit assignment process that determines when to acquire new knowledge by watching an expert’s behavior, and an efficient local credit assignment process that determines what new knowledge will be created for completing a failed explanation of an expert’s action. Because an input (e.g., observed action) to a failure-driven learning system can generate multiple explanations, a learning opportunity to extend the incomplete domain theory can go unobserved. This paper describes a failure-driven learning with a context analysis mechanism as a method to constrain explanations and thereby increase the number of learning opportunities. Experimentation using a synthetic expert system as the observed expert shows that the use of context analysis increases the number of learning opportunities by about 47%, and increases the overall amount of improvement to the expert system by around 10%.
[1]
Thomas G. Dietterich,et al.
The Role of the Critic in Learning Systems
,
1984
.
[2]
Tom M. Mitchell,et al.
LEAP: A Learning Apprentice for VLSI Design
,
1985,
IJCAI.
[3]
Judea Pearl,et al.
On Evidential Reasoning in a Hierarchy of Hypotheses
,
1990,
Artif. Intell..
[4]
D. C. Wilkins,et al.
Apprenticeship Learning Techniques for Knowledge Based Systems
,
1987
.
[5]
John R. Anderson.
Causal Analysis and Inductive Learning
,
1987
.
[6]
David C. Wilkins,et al.
Knowledge Base Refinement Using Apprenticeship Learning Techniques
,
1988,
AAAI.
[7]
Y. Kodratoff,et al.
WHAT IS AN EXPLANATION IN DISCIPLE
,
1987
.
[8]
William J. Clancey,et al.
Acquiring, representing, and evaluating a competence model of diagnostic strategy
,
1993
.