[comp.ai] Article by Tadepalli

larsaam@idt.unit.no (Lars AAsmund Maele) (04/30/91)

I am writing a thesis where I am comparing Explanation-based learning (EBL) and
partial evaluation in logic programming (partial deduction). I am looking for 
an article written by Tadepalli at Oregon State University about the use of training examples
used in partial evaluation. Do any of you know where I can find it?

AAsmund Maehle
Trondheim
Norway

tadepall@godel.orst.EDU (Prasad Tadepalli) (05/02/91)

> Path:
orstcs!ogicse!mintaka!think.com!snorkelwacker.mit.edu!bloom-beacon!eru!h
agbard!sunic!ugle.unit.no!sigyn.idt.unit.no!larsaam
> From: larsaam@idt.unit.no (Lars AAsmund Maele)
> Newsgroups: comp.ai
> Subject: Article by Tadepalli
> Keywords: Training examples in Partial Evaluation
> Message-ID: <1991Apr30.142338.5652@ugle.unit.no>
> Date: 30 Apr 91 14:23:38 GMT
> Article-I.D.: ugle.1991Apr30.142338.5652
> Sender: news@ugle.unit.no
> Reply-To: larsaam@idt.unit.no (Lars AAsmund Maele)
> Organization: Div. of CS & Telematics, Norwegian Institute of Technology
> Lines: 8
> 
> I am writing a thesis where I am comparing Explanation-based learning
(EBL) and
> partial evaluation in logic programming (partial deduction). I am
looking for 
> an article written by Tadepalli at Oregon State University about the
use of training examples
> used in partial evaluation. Do any of you know where I can find it?
> 
> AAsmund Maehle
> Trondheim
> Norway


        You must have picked up the reference from the Readings in ML. The
        report that was referenced there underwent a few revisions, and
        will be appearing as an IJCAI-91 paper titled "A Formalization of
        Explanation-Based Macro-operator Learning".

        This paper may not exactly be what you are looking for because it
        does not explicitly address the relationship of EBL to PE. However,
        it shows that EBL benefits from examples in two ways: First, it
        learns the distribution of problems it has to tune itself to; and
        second, it also uses the examples to avoid searching for solutions.

        One of the main results is that EBL can achieve exponential
        speedup WITHOUT using the algorithm of Kedar-cabelli and McCarty,
        which is the critical link between the PE work and the EBG work in
        the paper by van Harmelen and Bundy (AI Journal, 36). The key reasons
        for the success of EBL seem to be  (a) the ability to decompose a
        solution to a small set of macros which can be composed, and/or
        (b) the sparseness of the solution space. Since neither of these
        reasons is emphasized in PE work, it provides a counter-point to
        the thesis "EBG = PE", and, in this sense, might be relevant
        to your work.

        I must also point out that my work is based on Korf's pioneering
        work on macro-operator learning.

        I will be glad to send a pre-print of the paper if you are interested.

        Thanks,

        Prasad Tadepalli
        Department of Computer Science
        Oregon State University
        Corvallis OR 97331-3202