perm filename AUGBAC.MSG[MSG,JMC] blob
sn#175876 filedate 1975-08-29 generic text, type C, neo UTF8
COMMENT ⊗ VALID 00002 PAGES
C REC PAGE DESCRIPTION
C00001 00001
C00002 00002
C00104 ENDMK
C⊗;
∂16-JUL-75 1052 1,MSW @ USCT
Ed Fredkin can be reached at the VC lodge untill July 22 75
Tel. (303) 944 2211
∂15-JUL-75 1517 1,PAW
Chowning and his group would like to meet with you on Thursday at 3:00, let me
know if that is not agreeable....patte
∂14-JUL-75 2234 ESS,JMC
Call Glassmire 503 753-2527
∂14-JUL-75 2033 VCG,DCL
John- Please let me know what the status of the adjunct
faculty position is.
Also I don't know if you have a copy of the final version
of the NSF proposal that was submitted.
-David
∂14-JUL-75 1444 network site RAND
M E M O R A N D U M 8 July 1975
To: F. Corbato, J. Markowitz, J.McCarthy/L. Earnest, M. Pirtle,
T. Stockham, I. Sutherland, B. Woods
From: Peter Weiner (WEINER @ RAND-RCC)
Subject: Report on Computing Resources
--------
The body of the (hopefully) final version of our report
follows. Please let me know, by July 16, if you agree to our
submitting it to ARPA. If you want other changes, please be very
specific.
Regards,
Peter
--------
SUMMARY OF RECOMMENDATIONS
------- -- ---------------
1. Provide network-based reliable computing
2. Develop "personal computing"
3. Develop a network-wide file/archive system
4. Add special computing devices and software services to the network
5. Continue PDP-10 TENEX service during 1975-80
6. Support near-term TENEX/Network improvements
7. Induce DEC to assume TENEX responsibility
8. Centralize common network resources in 2 or 3 sites
INTRODUCTION
------------
Our committee was formed to consider the computational
needs of the ARPA/IPTO research community and how they might best
be fulfilled in the next decade. We take this charter in its
broadest sense, assuming that our committee should provide advice
and suggestions not only about extensions to the existing facility
but also alternatives to it, even though radical. In a time when
the state of digital electronics is changing the capabilities of
computers by an order of magnitude every five years and ARPA itself
is tending towards direct involvement in military applications, a
general review of the computing facilities required by ARPA/IPTO
research contractors seems warranted.
A View of ARPA/IPTO Computing Ten Years Out
- ---- -- --------- --------- --- ----- ---
In this section, we briefly describe a scenario of the
ARPA/IPTO computing environment ten years from now. We do this for
two reasons. First, we believe that ARPA can and should strive
for a significant impact on the future style of DOD computing;
and second, because we want our recommendations to be consistant
with a clearly stated long term goal.
It is clear that the ARPA/IPTO research community will
continue to rely heavily on networking into the indefinite future,
but the role of the network may change somewhat. The most important
function we see for networking is to provide an absolutely
reliable and solid network-wide file system for the research
community. Of particular note is the implementation of an
automatic archival sytem that integrates local on-site storage
with the network file system. Computational resource sharing
will continue to be an important function of the network, but
mainly with respect to unusual (typified today by the Illiac IV)
or varied (typified today by Multics, the IBM 360/91, the B6700,
etc.) resources; tasks now being handled by time-shared
medium-size computers -- text creation and modification,
compilation and small/medium scale computations, etc. -- will
revert to the province of relatively small on-site processors.
--------
An important prelude to making recommendations about the
directions of future facilities is first to define the objectives.
ARPA's mission, it seems to us, should be two fold: on the
one hand to fill the needs of advanced researchers so as to
advance the state of the art as much as possible at as little
cost as possible, and on the other hand to provide for the
maximum useability of that research for military purposes. To
some extent these goals conflict. Consider, for example, the
conflict between standardization and individuality. Unique
facilities tend to promote unique and potentially interesting
results. They also tend to make those results hard to pass on
to others.
The remainder of this report provides, in this context,
our advice to ARPA/IPT0 management; eight specific recommendations
are offerred, each with some discussion. The first four of our
recommendations are concerned with realizing the long-term
senario described above. When applicable, suggestions are made
for associated research and development that needs to be carried
out over the next few years. The second four of our recommendations
are concerned with ways to strengthen the short term situation.
Recommendation 1: Provide network-based reliable computing.
-------------- -
Over the next five to ten years, ARPA/IPTO should provide
to its research contractors a network-based computing environment
that is a) absolutely reliable and always available, b) consistent
in performance, and c) highly responsive.
As a first step, ARPA should initiate a 5-year research
and development program in reliable computing.
Discussion:
----------
By "abolutely reliable and always available," we mean to
suggest performance comparable to that of the telephone system;
service outages are extremely infrequent, information is rarely
lost, file integrity is maintained, etc. By "consistent in
performance" we mean that the user's expectations of system
response, based on past usage, are uniformly satisfied. "Highly
responsive" implies a high bandwith information flow from the
computer to the user, i.e. high speed two-dimensional output devices,
supported with appropriate software.
The principle component of ARPA computing facilities has
been the ARPA network. This network, now comprising more than
75 host computers in a nationwide hookup, has provided the bulk
of the computing services used by ARPA contractors.
The ARPA network has been a major success technically.
It is reasonably reliable. It is in widespread use. It has
clearly demonstrated the technical feasibility of packet switching
as a communications technique.
Perhaps the greatest success of the network, however, has
been a psychological one: a sense of community has developed among
ARPA researchers which makes collaboration and the use of other
people's work more professionally acceptable than ever before. The
ability to pass programs and data files quickly and easily from one
researcher to another has made collaborative research practical.
The network mail service has enhanced communication capabilities;
it takes less time to send a message to a colleague through the network
than it takes to call him by phone and he can pick up the message
at his convenience. The sense of community among ARPA researchers
fostered by the network is having a major effect on the number and
intensity of collaborative efforts now underway and is thus improving
the quality of research available to ARPA.
If the ARPA network has fundamentally changed our view of
what to expect from computing resources, there is still much that
can be improved upon. Consider the researcher who experiences
a system crash on a remote system and then finds all his files
have gone away. Consider the manager who mails a message to a
subordinate, only to be told later that the system lost the message.
Or consider the researcher who has to wait an increasingly anxious
and uncertain ten minutes for an overloaded computer to respond
to a request that got done yesterday in only two minutes. In
all these cases the productivity of the computer user suffers
and he spends all too much of his time fighting the system
instead of using it. He may even reject the system entirely
and choose instead to depend on more reliable, though less effective,
means.
The successful demonstration of packet switched network
theory in the form of the working ARPA network has had a profound
influence on DOD thinking about computer communications and resource
utilization. We believe that a working demonstration of reliable
consistant, responsive network-based computing will greatly facilitate
acceptance of ARPA developed research ideas. For this reason we
consider Recommendation 1 to be the key finding of our investigation.
Recommendation 2: Develop "personal computing"
-------------- -
ARPA should support the development of "personal computing."
One attractive initial step would be to contract for the development
of a minicomputer-priced version of the PDP-10 that runs TENEX.
For the 1980-85 time period, ARPA should provide each member
of its research community with a (logical) personal computer.
Discussion:
----------
As we have already indicated, we believe that computing
should be highly responsive to the user, with predictable responses
to repeated requests. This leads us to a computing senario that
includes a dedication of computing resources to individual users --
to "personal computing."
An important concept we have evolved is that of a
"logical personal computer" that we feel should be provided
each researcher. We argue as follows: The requirement for
responsive computing suggests a dedication of memory to each
connected computer user, certainly for the function of refreshing
a two-dimensional terminal screen. In addition, adequate
editing response can only be achieved by having some local
file storage together with local processors. Within ten years,
low-cost local processors will have power exceeding today's
medium sized machines. The requirement of consistent performance
suggests that these processors be dedicated to connected users.
But the resulting "personal computer" need not be co-located
with the user, nor need it sit idle when the user is disconnected
from the system -- it need only be within a distance that allows
a high bandwidth connection to the user's office.
One view of the personal computing environment is as
follows. Each person will be provided with a two dimensional
display for output (we conjecture that TV technology will continue
to dominate), and a keyboard and mouse for input. Audio input and
output will be available as options to be used when the added expenses
can be justified. A processor, with the power of, say, a KI-10
will be dedicated to each active user. The processor will run an
operating system which is designed to support display terminals. The
collection of (real or logical) personal computers at each facility
will be connected through a local network. This network will be
provided with enough file storage to hold all working files, and will
be connected to the ARPA network (or its sucessor) and the telephone
system. The ARPAnet will contain a collection of "computational
crunchers": unique resources such as the ILLIAC IV, Datacomputer, a
machine with large primary memory tuned to running large LISP programs,
etc. The network file system will give authorized users access
to any file generated in the community. Communication connections
to other networks will be provided.
We believe that it is only a matter of time before the
computing environment just described can be cost-effectively
supplied to the entire research community. (We assume that as costs
come down the size of the research community does not grow
substantially.) It is less clear whether manufacturers will of
their own initiave develop personal computing in the style described.
It may be necessary to initiate a research and development program
and to allow prototypes to be built and tested.
Recommendation 3: Develop a network-wide file/archive system
-------------- -
ARPA should initiate the development of a network-wide file
system. Emphasis should be on providing ultra-high reliability,
together with complete archival capabilities.
Discussion:
----------
One of the major discoveries of ARPA research in information
processing so far has been the extreme importance of the file system
associated with a multi-access computer. The file system in early
time-sharing systems rapidly became a means of communication between
users of the system, enabling them to much more easily build on each
other's efforts. The same importance of files has been evident in
the ARPA network. Files can be transmitted through the network from
one researcher to another and thus form again the major communication
mechanism between people who use the network.
At the present time, the network consists largely of a
collection of computing centers, each with its own file system.
While we believe that the future will see local on-site processors
satisfying most computing requests, we see all the more need for
a standardized network-wide file system (with physical file storage
centralized in 2 or 3 locations.) The cost of a centralized bulk
store is expected to be much lower on a per bit basis than that of
relatively small sized local stores. Another advantage of such a
system is that it would allow for uniformity of file management as
well as uniformity of file addressing. Moreover, reliability through
redundancy becomes practical through standard network-wide policies.
We believe that the network file/archive system needs
effective and flexible user control of file sharing and access,
denial of access, write protection, etc. This should be specifiable
by user and user subgroups. The net-archive, in addition to needing
access controls also needs to strive for processor independence,
language independence, and needs to solve the multiple copies for
reliability and the single copy for consistency dilemma.
Except where large amounts of data are needed close to the
computing power (such as in signal processing research involving
images, sonar, and very large data bases), only small amounts of
file storage holding active files need be local. Cost considerations
suggest the use of very large bulk storage devices to hold inactive
files. We believe the largest of these devices will be cheap enough
to allow all work produced by a researcher to be kept in permanent
archives, and that this facility should be an automatic feature of
the system.
Recommendation 4: Add special computing devices and software services
-------------- - to the network.
ARPA should remain alert to opportunities for placing special
computing devices and facilities into the network.
ARPA should remain alert to opportunities for supporting
special software services.
Discussion:
----------
Because the hardware technology available for fabrication of
computing devices is undergoing rapid change, many advances in
computing capability can be expected to come from new configurations
of hardware. These advances can sometimes lead to accelerated
progress in areas of prime concern to ARPA. For example, intelligent
systems research is now being limited by the address space and amount
of primary memory accessible in today's machines. Similarly, progress
in signal processing is likely to be greatly enhanced in the near
future by special collections of hardware able to implement the fast
Fourier transform (FFT).
Such devices may at first be very costly, and might thus be
considered only as special facilities within the network. The network
provides the important capability of making such facilities available
to researchers in spite of widespread geographical distribution. By
introducing such devices on a trial basis as network resources ARPA
can achieve the maximum impact on the future of DOD computing.
Illiac IV and the Datacomputer are two examples of facilities
which now exist. Two possible new candidates are: 1) a machine with
large primary memory, tuned for intelligent systems research, and
2) a special machine designed to search very large data bases quickly.
Similarly, special software packages developed by ARPA have
the potential for greatly affecting the way DoD uses its computational
resources. The development of message technology as a communications
mechanism which was preceded by the introduction of network mail
services is a case in point. Cross compilation services for small
machines are possible candidates for inclusion.
Improved methods are needed to disseminate information about
the network and the capabilities of the various sites. We suggest
that a network-wide documentation protocal be developed and
implemented at each site. The protocol would enable outsiders to
interact both in top-down and directed search modes to locate
information about equipment and services available, key personnel,
etc. Public research reports, position papers, bibliographies, etc.,
should also be maintained on-line. The incremental cost of
maintaining this kind of service would be quite modest.
Recommendation 5: Continue PDP-10 TENEX service during 1975-80
-------------- -
ARPA should initiate immediate and detailed discussions
with DEC about the "TENEX version" of the KL-10. If preliminary
assertions about the cost-effectiveness of this machine stand up,
ARPA should seriously consider depending on the KL-10 for the next
five years.
Discussion:
----------
The PDP KA-10 computer running TENEX has been the mainstay
of the ARPA research community. While this system has some
deficiencies (notably a lack of facilities for support of display
terminals, a fragile file system, inefficient handling of real
time devices, etc.) it has served the community well. The newer
KI-10 computer, however, using its own paging device rather the BBN
paging box, has proven to run LISP programs very slowly and is not
cost effective for the ARPA work load. (Changes to TENEX code
are expected to alleviate this situation.)
There exists a shortage of PDP-10 cycles at this time, and
demand for additional cycles is likely to grow rapidly. Fortunatly,
DEC is about to introduce a new KL-10 processor and is capable of
modifying the paging mechanisms and/or TENEX to produce a version that
will efficiently run TENEX. The new system is projected to run about
five times as fast as the KA-10, even for the typical ARPA job mix,
while costing just a little more than the KA-10. If these assertions
hold up under closer examination (DEC was understandably vague in our
discussions), then ARPA should seriously consider depending on the
KL-10 over the next five years.
If the choice must be made between leasing several new
KL-10 computers soon, or scheduling the purchase of these computers
over a longer span of years, we favor the leasing mechanism to
provide greater amounts of computing power sooner, and to allow
for a controlled phaseout of KL-10 service in favor of personal
computing. Indeed we would prefer to see ARPA arrange for the
block-purchase of PDP-10 cycles; if ARPA were to guarantee a
certain amount of business over the next five year period,
private capital to finance purchase of the required computing
equipment could be attracted.
Recommendation 6: Support near-term TENEX/Network improvements
-------------- -
ARPA should provide funds for changing TENEX and TIP
software where significant improvements in machine or network
performance can be projected. In particular, changes to Telnet,
host-host, and host-imp protocols are needed. Plans should be
made to produce a "final" ARPA supported version (1.35) of TENEX.
All TENEX sites should be strongly encouraged to move quickly to
the latest version.
Discussion:
----------
There are currently 15 machines on the ARPAnet that run TENEX.
If a software improvement results in a 10% increase in thruput,
then this improvement is worth the cost of 1 1/2 TENEX systems.
For example, it is estimated that the scheduler in version 1.31 used
up to 30% of the CPU, while the new scheduler uses only about 10%.
The largest inefficiency remaining relates to the full-duplex
character-at-a-time treatment of network teletype traffic.
Perhaps as much as 10% of each TENEX machine can be recovered by
replacing the Telnet protocols in both TENEX and the TIPs
to allow for a line-at-a-time discipline. Other changes to the
host-host and host-imp protocols would increase reliability and
significantly increase the effective bandwidth of host-host file
transfers.
Another potentially cost-effective approach to the same ends
would provide for a front-end processor and interface between the
PDP-10 and the network.
Recommendation 7: Induce DEC to assume TENEX responsibility
-------------- -
A vigorous effort to get the Digital Equipment Corporation
to assume responsibility for the TENEX operating system and its
successors should be undertaken. The systems programming group
involved should be on the network.
Discussion:
----------
Although DEC purchased rights to TENEX from BBN some years
ago, it is only recently that DEC has been willing to commit its
future time-sharing activities to the foundation established by
TENEX. Having a commercial supplier willing and able to take over
continued development of the prime operating system used in the
network would be a very important step towards transfer of network
ideas to industrial and military application. As long as systems
developed by ARPA contractors are not directly available from some
commercial supplier there will be unnecessary resistance to its
application in other areas.
Our contacts with DEC indicate that they are considering
adopting and supporting TENEX as a commercial product. ARPA
should encourage this decision by making it part of the negotiations
for new machines. Specific contractual agreements are recommended
as opposed to informal understandings. In particular, arrangements
should be made to insure that TENEX machines run efficiently in
the network environment.
It is important to the network users that the operating
system development group be faced with the same problems of
machine use that face the users. "If you want a reliable operating
system, use the one that the operating system developers use!" is
an absolutely true statement. It is important to the ARPA research
community, therefore, that the operating system group at DEC be
made an integral part of the network.
It is important that the operating systems development group
be on the network for technology transfer reasons as well. If this
group is an active participant in the network they will become
familiar with problems facing network users as well as solutions
developed with ARPA support. Such an intimate familiarity would
greatly speed the transfer of ARPA-developed technology into
commercial availability.
Recommendation 8: Centralize common network resources in 2 or 3 sites
-------------- -
Common network resources should be centralized in at least two
and at most three different geographic loactions. This recommendation
is a compromise between the savings in operational maintenance
and system programming costs to be achieved by centralization and
the risk of total system loss to natural disaster if only a single
location were used.
Discussion:
----------
The phrase "common network resources" is meant to denote
differents things during different time periods. For the next five
years, we mean general time-shared computing, particularly TENEX
service. Beyond that time, as personal computers take over this
computing load, we mean file storage and archival resources.
At the present time, the bulk of general time-shared computing
is provided to ARPA contractors by similarly configured PDP-10
installations geographically dispersed across the nation. This
situation is historically rooted from a time prior to the introduction
of the ARPA network.
More recently, the trend has been to accumulate computing
power at but a few locations, such as at ISI and BBN. This has
resulted in low cost management, operation, and maintenance, both
of hardware and software. This trend should be continued. As
existing isolated KA-10 processors lose their cost-effectiveness,
they should not be upgraded in place; rather, additional capability
should be added to centralized resource centers.
It should be observed that not all computing can be provided
from centralized resources. Much image processing research, for example,
requires high bandwidth connection between the input/output devices
(which must be local) and the computing devices as well as between
file system and the computing devices. Such computing requirements
will continue to demand that the network contain some smaller
computer centers. Indeed, important research could be seriously
curtailed if special cases are not given adaquate support;
centralization must not be pressed too far.
An opportunity for technology transfer exists in the choice
of operators of the resource centers. The initial transfer of the
ARPAnet responsibility to DCA, for example, may prove to be an
important step in technology transfer of the network technology to
military communication needs, since it virtually forces DCA to learn
the new technology. It remains to be seen what the cost, if any, in
deterioration of network services results.
Similarly, one might choose the operators of the main resource
centers with several objectives in mind. For example, if the Digital
Equipment Corporation operated a center we would expect increased
technology transfer of ARPA sponsored research to the commercial
sector. We believe that in addition to these considerations,
ARPA should choose center operators with ultra high reliability and
lowest cost objectives in mind.
∂14-JUL-75 1331 FOL,RWW
RE: SET THEORY MINI COURSE
It seems that from an incomplete survey that WED and MON afternoons
is a good time for some and particuliarly good for me. It is my
intension to meet in my office (unless we don't all fit) regularly on
WED afternoons at 4:00 til about 6:00. I would like to start this week
so some immediate feedback on this choice would be nice. A note
finalizing this WED will be sent sometime Thusday. I hope to cover the
material below and depending on what the group seems most interested
in to weigh it in different ways. If there is enough interest I might
extend this into the fall as there is a lot of things to do.
For factual information about sets I intend use Cohen's
book as a text and recommend you getting a copy. Each week I expect
to cover some section of this book. In addition each week I will assign
some "philosophical" reading in an attempt to shed light on what questions
were being asked by the people who found the currently popular axiom
systems. This reading is supposed to be fairly simple and basically
non technical. People will be free in addition to "volunteer" to
present the contents of slightly more technical papers.
The main interest of this course for AI will involve looking at the
various axiomatizations of set theory and finding out what features
of each are useful in a computational environment. This hopefully
will lead to a set of set theoretic principles that make doing
mathematics on a computer more "straightforward", that is, more in
line with mathematical practice.
Hopefully by summers end you will know the basic differences between
Zermelo,Zermello-Fraenkel,Godel-Bernays-von Neumann,and Kelly-Morse
set theory, as well as something about FOL.
∂14-JUL-75 1328 FOL,RWW
COURSE.INF
∂13-JUL-75 1429 THE,AJT
I've taken the liberty (again!) of giving yr. name as a reference for
a chap called Garey who will probably write to you from the
Universite de Lausanne. As I mentioned to you, Sloman has offered me
a job, but if it can be worked out, I would prefer to take a job with
Garey (a) because I want to return to working on neurophysiology
(Garey is interested, as am I, in applications of vision techniques
to biological images), and (b) because the job probably wouldn't
start until Jan.1976, which would mean more time for me to do a
better job on my thesis.
If/when Garey writes, I'd appreciate it very much if you
could reply just as soon as possible, since Sloman is breathing down
my neck for a decision.
Thanks a lot. Arthur.
∂13-JUL-75 0922 100,100: NARC @ SAIL
COMING SOON, TO YOUR NEIGHBORHOOD FOODPUSHERS, TURKEY JERKY!
∂13-JUL-75 0635 S,WD (reminder)
6:00PM Garbage goes out tonight.
∂11-JUL-75 1637 network site SRI
Date: 11 JUL 1975 1634-PDT
From: RAPHAEL at SRI-AI
Subject: IJCAI
To: PHW at MIT-AI, EJS at SU-AI, JMC at SU-AI, LES at SU-AI
cc: RAPHAEL
I just received a letter from Regina Bublil, Director, BACSJ.
She writes that she has discussed the 4IJCAI situation with other groups
in the East, and they "feel that both the 'boycott' and the 'action in
Tbilisi' are good actions, but there is no question that going to the
Soviet Union would be much more constructive for all the reasons that we
discussed." Although they will take no immediate action to oppose the
Minker letter, the BACSJ (and their colleagues elsewhere in the country)
would appreciate an opportunity to invite proposed IJCAI attendees to
"briefings" about the current situation in Russia, possible actions and
their likely effects, and so on. Having taken such a briefing, I believe
these briefings are well worthwhile, presenting a reasoned, practical view
of how best to preserve humanistic values without engaging in any illegal
activities and without endangering the scientific goals of the conference.
Can you (Erik or Pat) send me a mailing list of Americans likely to attend
the conference--e.g., authors and preregistrants-- that I may deliver to
Regina? I'm sure she would agree to clear any material with us before
sending it to the list, if you thought such a condition were necesary.
She also asks if we can find out what scientific publications,
if any, are likely to be represented at 4IJCAI. (She will find out about
the wire services.) 'Science' and 'Datamation' come to mind as
possibilities. Would you call them to find out? Are you aware of any other
possible press interest?
For your information, Lerner's biography:
"Alexander Lerner, born 1919, married with 2 children. Doctor of
Technical Sciences, was employed at Institute of Control Sciences of the
Academy of Sciences of the USSR, he applied for emigration in June 1971."
Bert
-------
∂11-JUL-75 0840 network site RAND
JOHN,
I GOT YOUR LETTER JUST AS I SENT OUT THE εINALVERSION
OF OUR COMMITTEE REPORT TO EVERYONE. (I SENT THE STANFORD
COPY TO LES.) I PRESUME YOU ARE BACK AT STANFORD NOW.
DO YOU WANT ME TO MAIL ANOTHER COPY TO YOU DIRECTLY?
REGARDS,
PETER
∂10-JUL-75 1308 S,LES
I have a copy of the letter that Minker plans to send to SIGART members,
if you are interested.
∂10-JUL-75 1126 network site CMUA
From: ALLEN NEWELL (A310AN02)@CMUA
Date: Wednesday, July 9, 1975 at 9:23 P.M. EDT
Subject: Response
To: JMC@SU-AI
- - - -
John: Welcome back. Hope trip was a good one.
(1) Heilmeier seems to be firmly in the helm at ARPA. He
gradually seems to eventually approve things but it is hassle,
hassle, delay, delay.
(2) For instance, we (CMU) have finally been approved (about mid
June for a 1 July contract date) by Heilmeier and thus are
transiting the rest of the process after our contract has
expired. We took about a 20% cut on our proposed amount which
was somewhat in excess of our last years proposal level, but did
get a bit more funding for some work in multiple micro
computers. However, we went in for 18 months (re extended
government fiscal year) but Heilmeier would approve for only 12
months. It appears that this was due primarily to our AI
component, hence related to your 6 months bit. However, no
specific instructions came with it. Arpa is no good at feedback.
(3) There is a JASON study on AI going on at LaJolla this
summer. Saul Amarel is involved heavily; Ed F is also and you
can get more info from him. Heilmeier specifically requested it.
(4) There is no collective action going on that I know of. I
suspect part of the reason is that the spring made everyone
relatively refractory. I have been fighting with a set of
analogous problems with Heilmeier and the SUS program. It is, in
my humble estimation, the very model of a good APRA program. But
Heilmeier is hassling it continuously. (Eg, why do we need three
system contractors, why not cut back to two. this gets laid to
rest, then re-emerges, then laid to rest, then ...) I have had
essentially neither time nor energy to devote to creative
strategies for AI.
(5) As you know there is an AI panel for COSERS, which Ed F is
nominal chairman and Terry W is a member. COSERS is presenting a
descriptive (non evaluative and non proposing) resport to NSF.
But it will be something NSF will use as a resource document.
Hence AI report plays some role in going to NSF for much more AI
funds.
(6) As far as I know, nothing has happened at ARPA or IPTO on
contacting NSF relative to transfer of some AI research to NSF.
(7) Lick seems to be coasting in the sense that he no longer
believes he can influence Heilmeier in a major way. So he told
me he is simply disagreeing with him at all the junctures,
trying to tell him why and letting the chips drift downward as
they wist.
(8) Lick is returning to MIT in September as far as I know. It
appears that Heilmeier will obtain someone from industry to
replace him. I think the result is totally unpredictable. By
that I do not mean necessarily bad, I mean unpredictable.
A.N.
-------
∂09-JUL-75 2253 ACT,REG
By the way, a file of yours, VOR1[1,JMC], is clobbered, and
has been clobbered since at least November 1972, since the DART tape of
that date has only a clobbered version.
∂09-JUL-75 1104 network site SRI
Date: 9 JUL 1975 1102-PDT
From: RAPHAEL at SRI-AI
Subject: IJCAI
To: EJS at SU-AI, JMC at SU-AI, EARNEST at SU-AI,
To: PHW at MIT-AI
cc: RAPHAEL
I spoke yesterday with a representative of the Bay Area Council on
Soviet Jewry (BACSJ), and with Minker's assistant in Maryland. Minker is
on vacation and apparently not returning calls. Here is my perception of
the situation, as a result of those conversations:
1. BACSJ, certain related organizations elsewhere in the country,
and Minker's "Concerned Scientists" group, seem to have a long standing
(and somewhat uneasy) agreement not to interfere with each other's
activities. BACSJ think that Minker's proposed letter is a bad idea,
but they will take no action to prevent or oppose it. They hope that
"the scientists" (we) can convince him not to send it.
2. The Concerned Scientists do not expect, and probably do not
even hope, that their campaign to move the conference will result in the
conference actually being moved. Their goal is to get publicity, to
increase awareness of the plight of Soviet Jews and the irresponsibility
of official Soviet promises, in the minds of both
the general public and the scientific community. They believe they can
best achieve this goal by having some people (them) oppose going to Russia,
while others (us) go and protest there. Although this plan sounds
vaguely inconsistent to me, I'm afraid they may be right.
3. The letter was being reproduced yesterday, and will go out
today assuming ACM provides them with the SIGART mailing list. Since
ACM Council recently passed an anti-Russian resolution, I expect that
the mailing list will be provided, the letter will go out, and Erik and
Pat may have to answer a flood of resulting inquiries (sigh). But I doubt
whether a significant number of travel plans will be changed.
4. BACSJ recommends that the wire John is now drafting express
our deep concern, but not explicitly threaten any protest action. The
rationale is that too strong a statement now, aside from damaging the
general atmosphere for the conference, could result in Soviet steps
to bar reporters and restrict conference activities, thereby reducing
the effectiveness of any potential protest that does take place.
I'm still trying to reach Minker, but don't expect any success-- at least
until after the letter is mailed.
Bert
-------
∂09-JUL-75 1043 ESS,JMC (reminder)
ARPA network mail to newell at CMU-10A -- ok
∂07-JUL-75 1904 S,LES
Discussion of ongoing plans for IJCAI on Wednesday, July 9, at 2pm
in A.I. Conference Room.
CC: RF;RCB;TOB;DRB;DBL;CCG;BLF;JMC;DES;EJS
∂07-JUL-75 0932 1,MSW @ USCT
I HAVE RECEIVED YOUR MESSAGE. THINGS ARE STILL THE SAME
I AM ALSO TOLKING TO WINSTON ON THE SAME SUBJECT, AND DELAYED
ANY DECISION UNTIL I WILL HEAR FROM YOU. WINSTON WILL BE DOWN AT
CALTECH ON THE 16 TH TO DISCUSS THE MATTER.
I WILL CALL WED. 10AM, UNLESS I HEAR FROM YOU BEFORE THEN.
∂06-JUL-75 2342 1,MSW @ USCT
I WOULD LIKE TO SEE YOU, OR TALK OVER THE PHONE IN THE NEAR FUTURE.
PLEASE LET ME KNOW IF YOU WANT ME TO COME TO STANFORD, OR ELSE,
PLEASE SET A TELEPHONE
APPOINTMENT. THANKS.
∂06-JUL-75 2228 105,SGK @ AI
My plans are to start for Calif by auto about July 16. Do I still have
a room, and am I still welcome around the lab. I plan to finish the CUSP manual, do a little
more work on the new Dired, and relax a bit. That should about take care of the summer.
∂06-JUL-75 1517 REV,REF
-- APPROXIMATLEY 2650 STEPS IN THE OLD PROOF
∂01-JUL-75 2035 ACT,REG
Philip Shaw, a former student of yours, wants information about the early
development of computer games (i.e., spacewar). (When you're back from
Japan) could you call him at 433-4150 (in San Francisco)? Ralph
∂01-JUL-75 1317 1,EJS
Received your telegram and talked to Pat W about it. He has polled the
conference committee and the majority recognizes the seriousness of the
situation butwants to proceed more carefully. We therefore postponed
sending the proposed telgram to Keldysh until we can all talk it over.
Erik
PS Welcome home!
∂30-JUN-75 0952 S,WD
Miss Mitchell of American+Express called (602)248-1424.
∂21-JUN-75 1348 S,WD
A fellow name A. F. R. (Tony) Brown, who says he met you at MIT in 1958,
called to say that he can offer us a 360 based programming system for machine
translation work. His address is: 536 Lowell Ave. Palo Alto, 327-0436.
∂04-JUN-75 2112 227,DEW
Please read JMC[REP,DEW]. It tells what I'm up to.
∂16-APR-75 0955 network site CMUA
**** FTP mail from [A310AN02] (NEWELL)
o: Licklider @ ISI
rom: Newell @ CMU-10A
ate: 16 Apr 75
e: AI ROAD MAP MEETING
C: Carlstrom @ ISI, Fields @ ISI, Russell @ ISI, Simon @ CMU-10A
JMC @ SU-AI, LES @ SU-AI, CCG @ SU-AI, PHW @ MIT-AI,
Feigenbaum @ ISI, Amarel @ ISI, Nilsson @ SRI-AI,
Sacerdoti @ SRI-AI
ick: Twenty-four hours has permitted some reflections on
onday's session and the problems attendent thereto.
1) To repeat what was generally accepted, implicitly and
xplicitly, by the AI scientists at the meeting: It is
ppropriate, given the current general context and specific ARPA
ontext, for the AI field to attempt a series of applications.
uch applications would be a good thing for AI generally, as
ell as for its specific relations with DoD.
2) Also to repeat: Application opportunities must be
iscovered, verified and exploited. Some institutional means
ust be found to expedite this. For it is clear that the AI
ommunity by itself does not have the expertise nor the
onnections to find high payoff activities. Dave Russell, at
he end of the day, strongly posited a mechanism of a Rand-like
gency with the mission of finding application opportunities,
erifying them, building a bridge to the AI Labs, etc. There
ould also exist, as an adjunct to this, an AI Applications
echnical Group (or some such title), consisting of
epresentatives of the various Labs, CMU, MIT, SRI-AI, SU-AI,
U-HP, plus maybe others in related programs, such as BBN-SUS,
DC-SUS, Amarel, etc. This group would be a prime forum and
nitiation point for these applications. I expressed some
oncern that such an agent could come into being in short enough
rder to satisfy the needs of the day (implying that some
emporary vehicle would have to be erected), but Russell seemed
onfident that such expedients were unnecessary. It would be
etter his way.
3) It is extremely important to be sure that the payoffs of a
pecific application are real. It is too easy to get
andbagged to have a seeming application turn to dross. Given
hat ARPA is prepared to spend large fractions of its AI
ommunity (a relatively precious resource) on producing some
pecific applications, it is critical to substantiate the need
nd acceptability of an application. ARPA itself, though
nside the DoD and much closer to the application sites than the
I community, does not itself have the expertise and,
mportantly, the time to examine the situations enough to make
eliable assessements.
4) To be concrete, on reflection I am not at all convinced that
he items on Heilmeier's list are all really in the category of
enuine application opportunities. I cannot speak about the
SW problem, since that is not a single problem, but many --
amely, the question of where to apply AI techniques all up and
own a complex system. But the other two are bitty problems
imed at highly specific targets. I did not get any sense that
PRA really knew in detail whether the payoffs were real or
imply ephemeral, momentary opinions of one or two high level
eople in the organizations connected with the applications.
eilmeier's carriage-trade philosophy requires a really good
arketing and marketing research arm if it is to succeed. It
ill do ARPA no good if it squanders its substance on a bunch of
rrelevant mirages. The military scene is littered with the
ead bones of expensive solutions which were only monuments to
omeone's folly.
5) It appears that MIT-MAC is already spending substantial sums
relative to the apparent size of the problem) on the Morse-code
roblem. Surely it would not be fruitful for the AI community
o get further involved in that one. I am not quite sure why
his one showed up on Heilmeier's list, but maybe it was just
eant to be illustrative of what ARPA had no decided to do.
6) On the language-spotting task, I need to reiterate what I
aid at the meeting. First, I do not think the AI
abs (in counterdistinction to the SUS Labs) should take on the
roblem, they simply would have to build up much of the
xpertise that the SUS Labs have, which would be a genuine
uplication. Second, I consider that the SURG has not been
sked to consider taking on that task. I agree that you possibly
entioned it to the SUSC (though I have forgotten it
ompletely), but since it was not brought up in the context of a
erious confrontation with the SUS 5-year goals, I do not take
t as a serious proposal. It would have been derilict of any SUS
roup to take on the task, given how tightly the SUS program is
trapped down to the 5 year goals.
do believe we can consider this task and, if it is important
nough, we can consider folding it into the present contractors
n some way. But we do have to face the potential effect on the
-year goals and to see how to work around them. As chairman of
he SUSC, I am quite willing to go around on that issue, but I
eed a signal from you or Dave Carlstom that indicates you want
o do that. This is a serious point and proposal I am making,
ince I do not want to be accused later of having fiddled while
ome burned -- of having not picked up this problem when it was
mportant to do so. I would like some feedback on this
pecific matter. The problem itself seems rather
traightforward, given the current art. I would estimate half a
an year for the technical work, if done at CMU given all the
acilities. To this must be added the whole custumer interface,
hich might be as much again. Much depends on details which I do
ot know, of course (eg, how must it be packaged and how much
ust it cost). If the problem could wait until after Nov76, you
ould surely get it taken on by the SUS Labs if it were as
mportant as Heilmeier stipulates.
7) Another example of a SUS-related application is the Korean
n-line communication aid, which you raised as a problem and
hich I suggested a solution approach to some time ago. I do
ot know what became of that. Again, it would be derilict for
he SUS community to fold that in without at least explicitly
acing the 5 year goals. All this stem, of course, from the fact
hat we (read: ARPA and possibly Newell-cum-SURG-initiator)
anted a program tightly fixed on impressive goals, and
herefore not with much slack for such things. Again, I am
illing to consider this.
8) I cannot believe that the CBC is not on the track of an
mportant application. It has two things wrong with it: (1)
here is not an immediate customer eager and ready to pay; (2)
otwithstanding SRI's search, there does not seem to be one
iding out there quite yet (though I do not know how intensive
hat search was or is). Yet, it does not seem to me profitable
- for ARPA, even on its own current terms -- to jerk that
ffort up at the roots and radically redirect it. Rather it
eems to me critical to widen the scope to "Real-time operations
onsulting" (namely, how to help someone carry out an operation
n real time) and to search for applications within this wider
phere. The core of work on the CBC remains in fact strongly
elevant; and the new applications can be grafted on.
9) What should go in the Road Map? It seems clear to me that
he Road Map for Friday has its action component defined
ndependently of its substantive component. To wit, the
ormation of the application-finding mechanism, defined above,
ill not be justified, nor require justification, from the a
tatement of the current art or a statement of future scientific
oals. However, this application proposal will differ from all
ther such attempts by the promise, implicitly extracted at the
eeting, by the AI Labs to enter into such an application-search
holeheartedly.
he substantive mode must perhaps still be there by Saturday.
ou don't have much to work with, in terms of what was generated
efore and during the meeting. Thus, I would attempt to get the
ction component to stand in for the rest. Let me discuss each
f the substantive components a little, and then come back to
his.
10) When a set of the worlds best scientists, being asked about
he their very own scientific domain, becomes tongue-tied and
roduces answers unsatisfactory in a first year qualifier, then
he conclusion is not that the science doesn't exist, it is that
he question was posed wrongly or the situation inhibited
dequate response. You asked us, I think, to do something
nder constraints that communicated: (1) that none of our prior
ttempts was to be considered satisfactory -- that something new
nd different was required; (2) that we adopt a form of
pecification of results and of expectations that is foreign to
S and AI, and largely foreign to science (namely, to state in
dvance the content of the scientific results to be expected up
o several years in the future, so that the questions would only
e whether or when the result would be attained. This pre-empts
he science and leaves us tongue-tied).
or instance, in the CMU proposal I have just finished writing a
tatement about the basic scientific questions of AI and the
igh level propositions that characterize what we have found out
n AI. Apparently that is to be discarded as not adequate or
ppropriate to the task -- and I am to find yet another
tatement, different from that, that is to be adequate to the
ew (yet identical) task.
or instance, Nils has just finished writing a paper (IFIPS 74)
evoted to a summary of what AI has done and what areas it has
orked in. Apparently that is to be discarded as not adequate to
he task -- and Nils is to find yet another statment, different
rom that, that is to be adequate to the new (yet idenical)
ask.
et me strongly suggest, for instance, that as far as
haracterizing the present state is concerned, you take a copy
f Nils IFIPS paper and underline in red the items in the
ibliography that are done in the ARPA AI Labs, and on the many
harts that draw a map of the area, circle in red these same
tems. This will give (1) a direct picture of the coverage and
cope of the field of AI that ARPA has given birth to; and (2) a
irect picture of the extent to which ARPA is responsible for
hese results and for the important ones.
et me further suggest that you put in front of Nils paper the
irst section on AI goals from the CMU proposal, as giving a
igh level coherent picture of what AI as a science is striving
o achieve and what in global terms it has found out.
hese two items answer only the question: What are the results
n AI in its own scientific terms. They do not answer it fully,
ut they will do as well as what you can put together in yet one
ore attempt in a few hours.
12) I do believe that several additional descriptions of AI
cientific results are possible that will appear to be more
atisfactory to upper ARPA than the two above items (if, indeed,
nything is satisfactory). I cannot carry out these
escriptions in the time available, indeed I think it would take
couple of months of very hard work (maybe more). But I can
ketch and illustrate one part of it (which is indeed based on
ast efforts to systemitize).
rogress in AI proceeds in terms of increases in scientific
nowledge about the various components of the intelligent agent,
omponents that are defined functionally. A standard division,
hich corresponds in part to Nilsson's core areas, is:
> Recognition and description (Perception)
> Vision
> Speech
> Language
> Representation
> Problem Solving Methods
> Control Structure
> Assimilation & Accomodation (Learning)
ithin each component one can describe a series of structures
or mechanisms) that are possibilities for this component. The
iscovery of each such structure and mechanism is an advance for
I and a result. Verification, of course, is required; it
omes, usually, from incoroporation in several total systems.
nowledge about each mechanism grows with experimentation and
heoretical sharpening. Such knowledge, again when verified
xperimentally, constitutes scientific results for AI. It
onsists mostly of statements of adequacy or sufficiency in
pecific task environments.
hus, the statement "What are the results of AI" at a given date
s a listing of the various mechanisms (usually described by
onventional technical names), plus the associated statements of
dequacy. This list grows over time, and it, rather than a
arametrization of how good are the systems that can be produced
onstitutes the core transferable knowledge of AI. This core is
ndeed transferrable, precisely because it consists of the
bstracted mechanisms which have been shown experimentally to be
seful in several task environments.
cannot produce the lists of results for the total field,
ostly because they have not been extracted, labelled and
rganized in this way. I can do it for one subpart, that of
roblem solving methods. Here, much that we know can be given
y specific methods (analagous to the methods of numerical
nalysis). A fairly good list is:
> Generate and test
> Hill climibing
> Heuristic search
> Search stragegy:
> Depth first, Breadth first, Best first,
Progressive Deepening
> Evaluation
> Evaluation functions, level of aspiration,
duplication avoidance, external limits
> Matching
> Hypothosize and match
> Means ends analysis
> Substitute & eliminate
> Range restriction
> Abstraction planning
o find a short way to say what we know, eg, about Hill
limbing, takes more energy than I have at this wee hour. We do
now the major things to beware of (Multi-modality, Mesas,
idges, Cliffs), we do have some empirical things to say about
hen Hill Climbing seems to work and when it doesn't. We do
ave a way of classifying the refinements of the method (as
implifed models of the hill, which are used to predict the
ptimum hill-step to take). And so on.
he existence of this list implies a large kit of tools
vailable to be used in applications, and indeed, when Ed
eigenbaum says they used "standard AI" in Dendral, he means
hat the techniques in Dendral pretty much are drawn from the
ist of such known and characterized methods.
o plot the success of the Problem Solving Methods component
ver time is to watch this list grow and/or the amount of
nowledge about each such component grow.
do not know how much each of the other components can be so
haracterized, though I expect it could be pushed quite far. But
ot tonight!
13) Future goals of AI must be in general to extend the
echanisms and structures of each area and to show that they are
dequate to wider and more difficult problems. The discover of
new method is not to be asserted in advance as a goal, if it
ould then the new method would have been found! Sometimes one
nows enough about a method or structure to specify as a goal
hat it is to be explored. To find the true scope of the range
estriction method, is such a (small) goal.
oals, in the sense that you (read: ARPA) want them, are only to
e associated with systems or with instruments (Physics has such
oals for the energy of interactions its accelerators will
each, or the resolution of microscopes). There has not been
ny difficulty, as far as I know, in determining how to
arametrize the structure and performance of specific narrow
lasses of systems (such as SUSs or Dendral-like systems) when
he task comes close enough to feasibility to make it worth
dopting a system as an AI goal. It can surely be done for
arious other specific classes of systems, though it will not
over, thereby, all of AI's goals.
I have run out of gas here -- I believe more can be said about
tating AI pure goals, but it just escapes my fog-bound mind. I
ove on.]
14) Applied goals, I believe are to be stated in one of two
ays. These provide other ways of describing what AI has done
n terms of how it can contribute to these goals.
ne way is how we started out to do it at the meeting. An
pplied system is posited (ie, a total military system, such as
n ASW system). Then, within that some points of AI application
re found, from which one attempts to derive the AI that might
ake a difference. This is a form of systems analysis, and one
hat can lead to a backward chain of available relevant research
esults and, as well, of still needed research -- methods,
tructures, knowledge, experimentation, , etc. needed to do the
ob. I think we should do a substantial amount of this, and I
elieve quite satisfactory road-map results would come out of
t. Unfortuneately, I believe that the effort per complex
otal system must be a summer study-group sort of thing, ie,
bout what we put into the SUS initial report. But this is
xactly what is to be done by this Rand-like agency (plus some
f us).
he second way is to specify applied technologies. The two
rototypic examples at hand are the notion of a SUS technology
nd (more pertinent) the natural language front-end technology
hat we all were talking about. One can take the development of
uch a technology as a goal and describe both what existing AI
lready provides and what new research is needed to get it.
his can be carried out much more within the AI community,
hough some sense for what is really required to make specific
pplications go is important. But again, it takes a fair sized
ffort to lay out such a technological alternative. We could
ommission such explorations. (It would depend, I guess, on
pper ARPA being prepared to consider such expansions.)
15) We did assert rather strongly that there have been a number
f civilian applications of AI, eg, in management science, in
esign, etc. Ferreting these out and asking whether any of them
ould be applied to military systems would be an additional
mportant task for this applications organization (along with
ome of us). This, of course, is yet one other way of stating AI
esults.
******
am really to the end of my rope tonight (this morning) and I
ill send this out after Herb gets a chance to look it over. I
m willing to work on expanding or modifying any piece of this.
am pretty much around from here through the weekend.
.N.
.S. Recall that I am expecting feedback on the Language
potting issue.