Information Filters and Their Implementation in the {SYLLOG} System
Shaul Markovitch and Paul Scott. Information Filters and Their Implementation in the {SYLLOG} System. In Proceedings of The Sixth International Workshop on Machine Learning, 404-407 Ithaca, New York, 1989.Morgan Kaufmann
Abstract
The study of knowledge has always been one of the central issues of the AI research. The general belief has been that "the more you have it the better you are." It was well understood that incorrect knowledge is harmful and should be avoided, but correct knowledge had been mostly considered beneficial. The potential harmfulness of correct knowledge received attention in very few early works but has become a key issue in several recent works. Knowledge is harmful if the costs associated with retaining it are greater than its benefits. Irrelevant knowledge and redundant knowledge are two types of knowledge that very often is turned to be harmful. When the knowledge that a problem solver uses is supplied by a human, we can hope that it does not contain harmful elements. However, when the knowledge is acquired by a learning program, it is desirable that the harmfulness of the knowledge will be eliminated, or at least reduced by the program itself. We introduce in this paper the notion of information filters. Information in a learning system flows from the experiences that the system is facing, through the acquistion procedure to the knowledge base, and thence to the problem solver. An information filter is any process that removes information at any stage of this flow. We call filters that are inserted between the experience space and the acquisition procedure data filters, and filters that are inserted between the acquisition procedure and the problem solver knowledge filters.
Keywords: Utility Problem, Active Learning, Forgetting, Selective Learning, Information Filtering, Explanation-Based Learning
Secondary Keywords:
Online version:
Bibtex entry:
@inproceedings{Markovitch:1989:IFT,
Author = {Shaul Markovitch and Paul Scott},
Title = {Information Filters and Their Implementation in the {SYLLOG} System},
Year = {1989},
Booktitle = {Proceedings of The Sixth International Workshop on Machine Learning},
Pages = {404--407},
Address = {Ithaca, New York},
Url = {http://www.cs.technion.ac.il/~shaulm/papers/pdf/Markovitch-Scott-icml1989.pdf},
Keywords = {Utility Problem, Active Learning, Forgetting, Selective Learning, Information Filtering, Explanation-Based Learning},
Abstract = {
The study of knowledge has always been one of the central issues
of the AI research. The general belief has been that "the more you
have it the better you are." It was well understood that incorrect
knowledge is harmful and should be avoided, but correct knowledge
had been mostly considered beneficial. The potential harmfulness
of correct knowledge received attention in very few early works
but has become a key issue in several recent works. Knowledge is
harmful if the costs associated with retaining it are greater than
its benefits. Irrelevant knowledge and redundant knowledge are two
types of knowledge that very often is turned to be harmful. When
the knowledge that a problem solver uses is supplied by a human,
we can hope that it does not contain harmful elements. However,
when the knowledge is acquired by a learning program, it is
desirable that the harmfulness of the knowledge will be
eliminated, or at least reduced by the program itself. We
introduce in this paper the notion of information filters.
Information in a learning system flows from the experiences that
the system is facing, through the acquistion procedure to the
knowledge base, and thence to the problem solver. An information
filter is any process that removes information at any stage of
this flow. We call filters that are inserted between the
experience space and the acquisition procedure data filters, and
filters that are inserted between the acquisition procedure and
the problem solver knowledge filters.
}
}