COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	@make(letterhead,Phone"497-4430",Who"John McCarthy",Logo Old, Department CSD)
C00010 ENDMK
C⊗;
@make(letterhead,Phone"497-4430",Who"John McCarthy",Logo Old, Department CSD)
@style(indent 8)
@blankspace(4 lines)
@begin(address)
Professor Edward Teller
The Hoover Institution
Stanford University
Stanford, California  94305
@end(address)
@greeting(Professor Edward Teller)
@begin(body)
Dear Edward:

	I'm sorry it took so long to get to the Chandrasekaran
proposal.  Somehow
it got into a pile of mail I didn't see until I met your secretary
and she mentioned that it had been sent.  Then it still took me
two weeks.

	The proposals seem to me to be state-of-the-art with regard
to the @i(expert system) and @i(knowledge-based) techniques proposed.
That is, it seems to me to be similar to what Stanford expert
system people would propose if they were working on nuclear
power systems - which they aren't.

	The proposal is clear and well written.

	The authors have taken the nuclear problems seriously
and made a real effort to become acquainted with them.  I don't
know enough about nuclear power plants to judge their success
in this.

	Their equipment and proposed equipment is adequate
for the level of work they propose.

	Their previous medical AI work is well regarded.

	However:
	
	In my opinion, the hierarchical methodology they propose
has serious limitations, especially with regard to carrying
out the Teller proposal to answer questions like, "What if
I turn valve X to the left?".  Namely, the hierarchy is of
properties of the system as a whole.  This corresponds to their
work and other AI work in medical diagnosis, wherein the reasoning
usually involves properties of the system as a whole.

	It seems to me that their formalism isn't primarily designed
to represent relations between different parts of the system, e.g.
"valve 3 being stuck open is causing excess pressure of 5000 lbs
per square inch in line 5 preventing valve 119 from closing."
In general, their representation doesn't emphasize quantitative
relationships, and I don't remember their giving any examples
of quantitative relationships.

	All this makes it difficult for their work to interact with
ordinary simulation studies, of which I presume there are a great
number in the reactor industry.

	Perhaps they could use their methodology for diagnosis
and interface their program to a more conventional simulation
program.  Thus their program might diagnose that a certain valve
is stuck open, and this would set a suitable parameter in the
simulation program to enable it to answer questions about the
result of turning other valves.

	However, prediction requires that the diagnosis phase
also make quantitative estimates, e.g. "line 312 is leaking
forty gallons per minute, and therefore it must have a leak 
of diameter 0.03 and will therefore continue to leak at a rate
which is .07 times the pressure.

	If they can extend their methods to give quantitative
diagnoses and combine this with a conventional simulation program
that has qualitative parameters for such conditions as blockages
in each line and quantitative parameters for such conditions
as leaks, then I think the program has a good chance of being
useful.

	I doubt that a much better proposal is likely
to originate from the AI community in the near future.
They are likely to find out the limitations of their methodology in
the course of carrying out the proposed study.

	I would not claim that this proposal or any based on the
present state of artificial intelligence is sure to make a major
increase in the ability of operators to control reactors in
emergencies.  On the other hand there is a reasonable probability
that the work will lead to progress toward that end.

	If asked for a definite recommendation, I would offer the following:

	1. Support the proposal at the scale proposed.

	2. Recommend that they study how to incorporate conventional
simulation in their system and offer to support an additional
person expert in this.  Such people are available and hopefully
the interaction will produce light as well as heat.

	3. Recommend that they broaden their AI model to include
relations between parts of the system and relations between
quantitative functions of the states of parts at different times.

	I have consulted three colleagues concerning the proposal
and the reputation of the investigators.  Their opinions are similar
to mine, except that one reaches more negative conclusions.  They
are wary about making excessive claims that artificial intelligence
is in a position to solve such problems in the near future, although
they regard the problem as interesting for AI research.

	The authors are welcome to read these comments, and I
don't request anonymity.
@end(body)
Sincerely,



John McCarthy
Professor of Computer Science