16-JAN-81 13:12:10-PST,3319;000000000001
Date: 16 Jan. 1981 1:11 pm PST (Friday)
From: Bobrow.PA
Subject: Your Paper on Heuristics
To: Lenat
Reply-To: Bobrow
cc: Dake, Bobrow


The basic problem with your paper is that it reads like an ARPA proposal rather
than a report of technical results.  There is so much advertising that exactly
what the results are, and their exact nature is obscured by the rampant
verbiage.  I have a strongly annotated version of the paper in my office, but
here is an overview of the contents.


Your argument on page 2 about severing the umbilicus raises severe questions of
the form of the data input, etc.  In addition, we rarely leave an individual
human in such a state, not talking to other experts, trying out ideas, etc.

One of your main points, time after time, is that Heuristics and Knowledge
Representation are fields JUST LIKE ANY OTHER.  What properties are you
imputing to fields other than the fact they have names.  If you name something,
does it become a field?  Are all fields studiable?

You argue on one hand for heuristics being compiled hindsight, and on the
other that they spring into being (see p 10).  Neither description seems relevant,
but it does confuse the reader.

The central assumption underlying heuristics seems misstated.  It is not
continuity at all that is at issue (Dividing by a number very near zero is not the
same situation at all as dividing by 0).  It is that there exist characterizations of
situations which are not dependent on all the properties of the situation which
can give one clues for appropriate actions.  The heuritician's job is to find those
characterizations.

The curve drawing exercise is overblown.  Do you have ANY heuristics for
which you can really plot such curves?  Isn't it all really a metaphor?   If
presented that way (as a metaphor) it would be equally illuminating, and not
raise questions such as "What is on that task axis?" 

On page 17 you say there is always a Generality/Power tradeoff, but your earlier
results on pruning the AM heuristics argues strongly otherwise.

A significant problem in the latter half of your presentation is your nonstandard
use of the words slot and subslot.  Each clearly carries lots of meaning to you,
but seem an implementation term to me.  Perhaps you mean a relationship an
entity can participate in.  Perhaps there are special proeperties of slots as they
express relationships.  This needs to be made clear before metadescription of slots
is attempted.  I was confused, and I know your work reasonably well.  Perhaps a
good strategy would be to provide a brief description of the epistemology of AM
and its basic processing.  

The example at the end of the paper is confounded by interjections of
generalizations.  Explain the example first, and then draw generalizations; and
please tell how much of this is proposed scenario, and how much is real program
running now.  If there is a program, how general is it?

My bottom line is that your paper needs drastic revision to appear in the AI
journal.  I will be happy to have a new version to consider, but I think this one
should not wait further reviewing for a decision (since it has been reviewed by
Bruce and me).  I am glad to talk to you about the paper at some length if you
want.

danny