Newsgroups: comp.ai
Path: cantaloupe.srv.cs.cmu.edu!rochester!cornellcs!newsfeed.cit.cornell.edu!newsstand.cit.cornell.edu!news.kei.com!simtel!lll-winken.llnl.gov!uwm.edu!cs.utexas.edu!news.eas.asu.edu!ennews!news
From: rao@parichaalak.eas.asu.edu (Subbarao Kambhampati)
Subject: Re: Is EBL still a research issue ?
Message-ID: <RAO.95Nov9092830@parichaalak.eas.asu.edu>
Sender: news@ennews.eas.asu.edu (USENET News System)
Reply-To: rao@asu.edu
Organization: Dept. of Computer Science, Arizona State University, Tempe
In-Reply-To: Champciaux Laurent's message of 9 Nov 1995 14:34:40 GMT
Date: Thu, 9 Nov 1995 16:28:30 GMT
References: <47t3i0$cnf@wfn.emn.fr>
Lines: 83
X-Nntp-Posting-Host: parichaalak.eas.asu.edu

In article <47t3i0$cnf@wfn.emn.fr> Champciaux Laurent <lchamp> writes:
> 
> Hi,
> 
> Is Explanation-Based Learning still a research issue ? If so, does somebody
> have any recent references about EBL ?
> 
> Thanks in advance for your help.
> 
> Laurent.
> 

I believe EBL is very much a research issue, except that there is a
more mature understanding of EBL's rightful place among the canon of
learning methods. Specifically, any background knowledge can be used
as a strong form of bias to focus inductive learning. EBL can be seen
as an extreme form of this bias, where the bias is so strong that
the hypothesis space is narrowed drastically. This realization has
made people abandon the EBL or Induction dichotomy and concentrate
more on hybrid learning approaches which use EBL to pre-process the
training examples that are given to the induction algorithm. The
advantage of such an approach is that unlike traditional EBL methods,
we don't need to insist on complete and correct explanation. Any
reasonable domain theory may be able to give us _some_ idea about what
the relevant attributes of an example are and this helps 

As for references, you can start with the chapter on knowledge based
learning in Russell and Norvig's book to get an idea of how EBL
relates to knowledge based induction.

You can then consider reading papers by Mooney and his students on
combining explanation based learning with inductive logic programming
to learn search control information (essentially, you use EBL to
pre-process the examples given to your induction algorithm such that
they don't contain irrelevant attributes; this increases the rate of
convergence of induction). 

A similar use of EBL is made by Dietterich and Flan on combining
reinforcement learning and EBL approaches in a paper that appears in
the recent MLC proceedings. (try Dietterich's homepage accessible
through http://www.cs.orst.edu). Dietterich also has started the
hybrid approaches originally through his induction over explanations
algorithm (appeared in MLJ several years back).

Jonathan Gratch just completed a nice thesis that provides a sound
statistical basis for handling the "utility problem" in EBL. Papers on
his work can be found in AAAI-94 and earlier. 

Since I don't want to be accused of not having any bias of my own, let
me also put in a blurb for our recent work in the area -- application
of explanation based search control rule learning to partial order
planning (UCPOP/SNLP type planners). This work provides a relatively
clean formalization of EBL in terms of regression and propagation, and
makes it easier to see connections between the no-good learning
techniques used in constraint satisfaction adn those used in
planning/problem solving. The paper, to appear in Artificial
Intelligence, is available at http://rakaposhi.eas.asu.edu:8001/yochan.html

You may also look at
http://rakaposhi.eas.asu.edu:8001/planning-class.html for some online
notes on use of EBL and inductive learning technqiues in planning.


Hope this helps.

Regards
Rao
[Nov  9, 1995]
----------
Subbarao Kambhampati Dept. of Comp Sci. and Engg. Arizona State University, 
Tempe, AZ 85287-5406  rao@asu.edu (email) 602-965-0113 (Phone)
602-965-2751 (FAX) 
WWW: "ftp://rakaposhi.eas.asu.edu/pub/rao/{rao.html, papers.html}"
     "http://rakaposhi.eas.asu.edu:8001/{rao.html,papers.html}"
-- 








