perm filename TU[NOT,DBL] blob
sn#215677 filedate 1976-05-17 generic text, type C, neo UTF8
COMMENT ā VALID 00002 PAGES
C REC PAGE DESCRIPTION
C00001 00001
C00002 00002 Comments on:
C00018 ENDMK
Cā;
Comments on:
THE PSI TRACE EXPERT, by J. V. Phillips
**************************************************
GENERAL comments
**************************************************
There are some new, interesting ideas contained in this paper, plus a
few ideas which (although not novel) have never been explicitly
stated in print before. This paper should definitely be accepted for
the 2nd Int'l. Software Engineering Conference.
Although the paper is acceptable in its present form, I advise the
author to spend more time editing any future papers. If possible, he
should try to rewrite this paper, although I don't recommend that as
a condition of acceptance. The major defect in the paper is that of
very poor editing. By that, I mean:
(i) frequent use of undefined terms and jargon [e.g., INT
form]. Most of this stems from the obvious fact that TU is just one
module in a larger system. This paper should be edited so that it can
"stand on its own".
(ii) frequent occurences of very awkward English sentences.
Although I, as reviewer, had to take the time to read a sentence 3
times if necessary, the average reader may simply move on without
comprehending it. [e.g., the following two senteces follow each
other: "The scene fits. Output fits." I parsed the second one just
like the first, i.e., to mean "the output does in fact fit some
criteria", when in fact it was supposed to mean "print out the output
which deals with how x fit y"]
(iii) occasional lapses into naivete', ignorance of earlier
sources, and exaggerated claims. [e.g., "The most common way of
transferring knowledge between humans, be it in the form of concepts
or procedures,, is by using examples"].
(iv) Failure to make clear what the current state of TU is.
The reader is never quite sure whether this paper reports on a
running program, or merely an idea for a program.
(v) Poor placement of sections. [e.g., Section 3.3.1 and
3.3.2 are excellent sections, buried in the middle of the paper.]
(vi) Poor abstract (very hard to understand what this paper
is all about), and poor conclusions section (the material in Section
5 is more a "summary" of the design). What (if any) ARE the
conclusions?
**************************************************
SPECIFIC
**************************************************
Abstract: very hard to comprehend. Is this obfuscation synergetic?
p1: "The most common..." is pretentious and proabably inaccurate.
p2: Does TU presume that the program can naturally be written as a
procedural one (linear sequences of actions, loops, and tests)?
Could TU really infer, say, MYCIN -- a production-system-oriented
program? Of course the LISP compiler does produce such a procedural
form, but it would be very awkward to infer from traces. If the
inferrer were told to try to infer a production system, his task
would be greatly simplified. Have you thought about what assumptions
TU is making re its target program? How does a small bit of global
information -- like `it's a production system' help constrain the
inference process? It's perfectly OK that TU will write a certain
kind of program only, but you should state that explicitly.
p2: You claim that "A single trace is not enough" to infer TF. Is a
trace just one cycle of the outermost loop? Surely if a long enough
session were presented, a HUMAN could infer the program. Why isn't
this enough for TU?
p3: This sounds like TU can infer ANY program from its trace. Are you
really claiming that? How about the interesting and useful program
URAND? When claims of universality are involved, the conclusions are
necessarily trivial from a pragmatic standpoint (e.g., see the work
by Blum&Blum).
p3: Similar uses of the phrase "natural language descriptions" imply
that TU (or at least PSI) can speak English and other tongues
fluently. At best, I imagine that TU (PSI) claims to be usable by an
untrained, unprepared human, who will be led to conduct a dialogue in
a limited subset of English.
p3: Why is Natural Language capitalized in only some places? Do you
really mean a particular fixed subset of ENglish, when you capitalize
N.L.?
p4: Section 1.2.2 is either trivial or unfathomable, depending on
one's familiarity witht he rest of the PSI system. (To me,
unfathomable).
p5: What exactly is an "INT form"? What is a "form-changing rule"?
Give example of these things when you state them, if not sooner.
p6: This makes TU sound like a small, nonessential module in a larger
system. If TU is this unimportant, why isn't it just described in
the overall PSI paper? Probably, you should just skip all this
material.
p8: Why do you quote from an as-yet-unpublished reference for this
dialogue excerpt and for the TF program? The same excerpt (and target
program) are given in Lenat's "Beings" paper. See the 4th IJCAI
proceedings, pp 130-1.
p8: "The scene fits the concept so output fits to the user" is hard
to parse. Can PSI really handle these sophisticated constructions?
Even if it can, is this done by TU? If not, it is merely a
distraction in this paper. Keep the Natural Language easy to parse,
please.
p11: "concept" has hiterto been used in a specific technical sense.
Now it seems to be used in a new, colloquial sense. The first,
technical use (part of TF) was never stated explicitly. Considering
the emotional charge of that word, this is not a good policy. Even if
YOU call both of these things concepts, make up a separate name to
keep the reader from confusing them.
p12-13: What observation?! This is crucial! Was it by introspection,
by interviews with experts, by guessing, by careful psychological
studies, etc.? There may be some new ideas buried here, and you're
doing them a disservice. "Concept" here is not related to "Concept"
as a data structure in TF, is it? By the way, what does TF stand for
(True/false? Theory-former?...)?
p14: What if more than 1 template is suggested as relevant? e.g.,
what about this sentence: "The initial values are input"?
p14: What are "slot-filling rules"? Give examples. How many are
there? Do they exist yet, or are they merely planned? Who wrote
them, and how? How can they be checked, extended, etc.? How general,
powerful, useful are they? THis "fan-dancer" style of yours is quite
frustrating.
p14: Section 3.1.3: Does "value" mean "the name, the identity of"?
p15: What is a collection? plex? correspondence? This is obviously
PSI jargon. If you must use it, at least explain it. I infer that a
colllction is a 1-level list, a plex is a property list, and a
correspondence is a list of ordered pairs. Is this even close?
p.17: Section 3.3: What if the target program uses "processes", and
the only orderings on the events are PARTIAL orderings? WIll spurious
control branches be inferred?
p.17: Does `isomorphic to' really just mean "sufficient to serve as"?
p.17: Surely you don't mean to say that TU is designed to infer THE
control structure of the particular target TF program(s) you've been
looking at? Ther might be many "natural" ways to code up the same
programin LISP, i.e., to write a program having a simialr trace but a
quite different control structure. If 1 TF programis somehow
"distinguished", that is worth explaining carefully. It's OK if TU
has a specific ability, as long as that is made clear. In other
words, could TU turn out several very different control structures
from the same trace, or is there really just one family of control
structures that TU "assumes" the target program will fall into?
Section 3.3.1: an excellent section.
Section 3.3.2: an excellent section. Here you might mention
explicitly that TU would not work well for Operationg systems, or for
production-based programs (MYCIN), or for programs using the
agenda-like mechanism for control (a la Winograd and Bobrow's KRL),
where jobs get stored, and get executed inorder of priority.
p.20: Section 3.3.3: a collection of unintelligible PSI jargon,
again. Explain this, give examples, or OMIT.
p23: I disagree that the 2nd is more likely. The 1st is more likely,
since it's shorter. If A and A-B are equally plausible, A is simpler,
hence (by Occam) preferable to A-B. Of course you know that A-B is
"right" for getting TF, and this has been pre-programmed into TU.
Make these hacks clearer. Why doesn't TU consider the symmetric
difference between A and B, in fact? That works, in this case, and is
just as valid as the set-difference between A and B. [above, A and B
represent "scene" and "concept model"].
Section 3.6: does this (albeit short) section add anything to this
paper? It rates perhaps 1 sentence, in an appendix called "how TU
fits into PSI".
p25: What in the world is an "AAU Actor"? Does Carl Hewitt know? Does
such an actor belong to some union?
Section 4: After reading this far in the paper, why can't I follow
section 4 in detail? Perhaps some comments are needed in the margins.
Section 5: These ar not conclusions. Are there any so far? It's OK if
not.
Appendix A: As long as you're using up all these lines, why not add
little comments along the right margin, to explain what the slots
mean?