Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.2 9/18/84 +MULTI+2.11; site brueer.UUCP
Path: utzoo!linus!philabs!cmcl2!seismo!mcvax!ukc!reading!brueer!holte
From: holte@brueer.UUCP (Robert Holte)
Newsgroups: net.ai
Subject: Re: inductive learning in expert systems
Message-ID: <200@brueer.brueer.UUCP>
Date: Mon, 12-Aug-85 02:59:49 EDT
Article-I.D.: brueer.200
Posted: Mon Aug 12 02:59:49 1985
Date-Received: Thu, 15-Aug-85 08:32:59 EDT
References: <770@utcs.UUCP>
Organization: Dept of EE & E, Brunel University, Uxbridge, U.K.
Lines: 45
Xpath: reading gateway.cs


The only system I know of which generates rules for a knowledge base
on strictly probabilistic grounds is the RX system (R.L. Blum,IJCAI,1983).

The use of AI machine learning techniques to generate a knowledge base from
examples is strongly advocated by Donald Michie, who has made commercially
available knowledge engineering tools (e.g. EXPERT EASE, and RULEMAKER) for
this purpose. There has not been much published on the use of these tools,
and I would very much like to hear from anyone who has experience using them.

Michalski reported a striking case in which the knowledge base induced
from examples significantly outperformed those which had been engineered
in close collaboration with a human expert (International Journal of
Man-Machine Studies, vol.12, pp.63-87, 1980, and elsewhere), but he does not
appear to have carried out any similar studies since. A rule-learning module
was included in his ADVISE system for knowledge engineering, but I
have not heard anything about ADVISE since it was announced at IJCAI in 1983.
The successful domain was soybean disease diagnosis, and I believe a recently
begun ESPRIT project is attempting to extend this work to a wide class
of European plants and plant diseases.

I attended the International Machine Learning Workshop in June.
The topic of generating knowledge-bases from examples was discussed only
once, by Bruce Buchanan in his overview of learning research at Stanford.
Of the sixty-plus research summaries included in the workshop proceedings,
only a handful were concerned with this topic, namely those coming under
the general heading of Learning Apprentices. Learning Apprentices learn new
control (meta ?) knowledge or correct existing knowledge by observing a human
expert solving problems with the aid of a reasoning assistant (e.g. an ordinary
knowledge-based system): they are not intended to create new knowledge bases.
Most of the existing LAs will be described at IJCAI in Los Angeles (1985).

On the other hand, I have seen a proposal (not yet approved) for a joint
European project which specifically includes the general topic of automatic
knowledge acquisition for expert systems, and I have heard a rumor that there
will be a (relatively small) commercially-backed project on this topic in
Britain starting this year.


-- Rob Holte	holte%brueer@ucl-cs
		{...ENGLAND}!ukc!reading!brueer!holte
   Tower C
   Brunel University
   Uxbridge
   England   UB8 3PH