perm filename NEWELL.PLN[LET,JMC] blob
sn#187284 filedate 1975-11-17 generic text, type T, neo UTF8
∂16-NOV-75 1044 FTP: host CMUA
From: ALLEN NEWELL(A310AN02)@CMUA
Date: 16 Nov 1975 1329 EST
Subject: First rough draft of DNL proposal to NSF
To: HERB SIMON, FEIGENBAUM@SUMEX, HART@SRI-AI,
LICK@DMS, JMC@SU-AI, NILSSON@SRI-AI, WINSTON@MIT-AI
- - - -
To: Feigenbaum, Hart, Licklider, McCarthy, Nilsson, Simon, Winston
From: Newell
Date: 16 Nov 75
Re: Proposal for the Distributed National Laboratory
It has been, as always, hard to find the time to work on this.
However, it seems essential that we get something to NSF soon
(somewhere a 1 Dec deadline sticks in my memory). I have at
least gotten started and so will transmit this to you for
comment while I try to proceed on the missing sections. Please
excuse the typo's. I find they come out easier on hardcopy and I
have seen only softcopy through a small window on this. This is
very rough draft and all parts of it are up for grabs.
I do think we want to circulate something very broadley to the
community as soon as we can. I would hope to do it next week,
and the same copy would go to Atkinson, Pasta, etc. as to
various people in the field. It would be labeled ROUGH FIRST
DRAFT and would disclaim that anything at all was firm, but that
this was just a trial balloon. We have two options on
authorship: to have it come from the little committee we set up
((One of Hart, Nilsson), McCarthy, Newell (ch), Winston) or to
expand it to, say, the set of people in the address list above.
My bent is slightly for the first of these, since that is what
we set up and it is easy to explain. A covering memo would go
out that explained it all.
The present partial draft does take a position on a number of
issues you should attend to:
(1) DNL for AI or DNL for Cognitive Science? I tried one version
with DNL for the study of the mechanisms of intelligence, but
that got too clumsy. There appear to be two issues. First, the
issue of making the focus be scientific-problem centered, not
discipline centered. Second the liabilities that many people
feel about the term AI. The existing version makes it easy to
coelesce with cogntive psychology, which I think will
increasingly happen. It makes it harder to imply the membership
in Computer Science. In fact it even raises the question of why
one should not be worried about other types of laboratory
facilities for studying human behavior. Anyway, I decided to
send this one around to let you see what it would be like to try
to move off the name of AI. Dick Atkinson didn't feel that the
"AI" issues was critical.
(2) The major consequence of the scheme here is the willingness
of the AI Labs to go this route and give up the autonomy. It is
what will be inferred if we send this document out in draft
form. SUMEX has not quite gone this far, I discover. The other
alternative is to make the DNL a creature of the existing Labs.
That then has problems of whether it is really for the
community.
(3) Most of the positive prose about why-AI (why Cognitive
Science) is slated for the missing sections. What is here
concentrates on the vehicle itself. My intent is to fill those
sections with things that tout Cognitive Science as a
fundamental science, that describe all the familiar areas in
terms of exciting research done, and talk about next scientific
steps. I do not intend to make a case for any one substantive
thrust rather than another ("now is the time for theme X")
because I think the DNL should be basic for the whole field. One
cannot in truth talk about great new things that will occur with
the acquistion of the resources, since the first task is to
obtain the current+ levels of basic funding.
(4) Though there is a fair sized promise in the notion of a
genuine community of scientists tied together by shared
facilities over the network, my tendency is to treat this rather
warily and not as the main point. The possibility should
certainly be discussed, but to make this a key selling point
runs afoul of too clear a perception by the community of why
this proposal is here at this time (ARPA => NSF transfer). My
belief is that a document of this kind loses points if it
appears to disemble, ie, to create a picture of reality that
differs radically from the picture all believe. Whether it gets
positive points for being straightforward and honest is much
less clear and too deep a proposition for me. It also run afoul
of the cubic miles of talk about such things through the last
many years, so that only rather concrete near-term things can be
taken seriously. I do think we can capitalize strongly on the
fact while others must point to the ARPA net as substantiating a
style they wish to try, we can claim to be the original ARPANET
community and thus our opinions are to be taken seriously about
what can be achieved.
(5) Is the separtion of description of the institution and
argument about it (Sections 4 & 5) right, or should they be
combined somehow.
(6) Is the flavor of Section 5, which tries to state concerns
directly and them meet them with argument, rather than leaving
all the problems unsaid, a good thing? One hopes to convey an
impression that the arrangements have really been thought about,
rather than just being something half assed. What problems have
we missed?
PROPOSAL FOR A DISTRIBUTED NATIONAL LABORATORY FOR COGNITIVE SCIENCE
(DNLCS)
File DNLP1.1C [A310AN02] @ CMU-10A 15 Nov 75 Newell
> Tries the name Cognitive Science, rather then Mechanisms of
Intelligence
OUTLINE
> [...] means not written yet
> Lack of section numbers means it hasn't even been located
within section and might be decomposed in some other way
> 1. Intro
> 2. Cognitive Science
> [Covers why we call it cognitive science rather than AI
or Intelligence]
> [Describes the substance briefly - nothing knew here]
> [Describes the current frontier]
> [Describes the history of development of the field,
especially the ARPA support]
> 3. The Need for Facilities and Institutions
> [The current experimental style]
> [Something about how much computing is needed per person]
> [Give a description of the computational resources
at AI Labs]
> [Something about size and composition of the field]
> [By example (speech/vision?) makes clear that not
cycles alone]
> 4. The Basic Proposal
> 4.1 General Functions
> 4.2 Types of Facilities
> 4.3 Institutional Structure of DNLCS
> 4.4 Nature of Arrangements with the Research Center
> 5. Problems and Promise
> 5.1 Why a National Laboratory at all?
> 5.2 [What will be Gained?]
> 5.3 Why a Distributed Laboratory?
> 5.4 What Assures that the Entire Scientific Community Benefits?
> 5.5 What Prevents Perpetuation of Favored Research Centers
> 5.6 What Prevents Isolation of the Cognitive Science Community?
> 5.7 Why Environments rather than just Facilities?
> 5.8 Why isn't DNLCS just a Disguised Granting Agency?
> 5.9 Why won't DNLCS become a Bureaucratic Nightmare?
> 5.10 [How will the Scientific Community Evolve with DNLSC?]
> 6. Funding Calculations
> 6.1 The DNLCS Fraction
> 6.2 [The Operating Budget]
> 6.3 [Typical Facilties Budgets]
> 6.4 [The Initial Situation: The AI Laboratories]
1. INTRODUCTION
This document proposes the establishment of an institution to
carry forward the basic scientific study of the mechanisms of
cognition. We call this institution the Distributed National
Laboratory for Cognitive Science. The meaning behind the name
will become clear presently. We will refer to it simply as the
DNLCS.
How is it that intelligent action is possible? How is it that
the world can be perceived in an intelligible way? How is it
that a wide array of knowledge can be organized so that it
becomes available when necessary to solve some problem? How is
it that an action can be guided skillfully in the face of a
noisey or even a hostile environment and still succeed? How is
it that some task can be almost impossible to perform at one
time, yet with further attempts at performance becomes much
easier? How is it that knowledge and intellectual skill can
increase continually -- for decades and maybe even without bound
-- without producing an immense confabulation? In sum, how is it
that cognitive activity occurs and is effective in our
universe?
These questions delineate an area of scientific inquiry. Until
recently they would have been asked only of humans. But within
the last two decades, with the development of the digital
computer, we have learned that they should be asked more
generally. The answers we seek should provide a basis for the
construction of intelligent systems. They should provide an the
understanding of intelligence in humans and also in the rest of
the biological world. Most generally, they should yield an
appreciation of the role of cognition in our Universe -- of its
capabilities and limitations.
The scientific study of cognitive activity resides today
primarily within Computer Science and Psychology -- in
Aritificial Intelligence and Cognitive Psychology respectively.
Concern with these problems can be found, of course, in parts of
many fields: in Pattern Recognition, Biology, Control Theory,
Economics, Operations Research, Mathematics, Philosophy -- even
now in Astronomy with its concern for the possibilities of
extraterrestial intelligence. The proposed institution is
centered on a scientific domain. Though this maps strongly into
two disciplinary sectors, namely Artificial Intelligence and
Cognitive Psychology, the intent is to cover the scientific
problem no matter under what rubric it happens to reside.
We propose the establishment an institution because we see that
the fruitful growth of the scientific study of cognition
requires resources and styles of investigation that we do not
know how to meet in other ways. We propose this now -- not
yesterday and not tommorrow -- because we see drawing to a close
a pattern of support for this field, which has been highly
successful up to now and which appears to be changing. The
field will not prosper, indeed will quiesce, without new
arrangements for its investigations. Thus, history enters into
the story. We propose a particular type of institution -- what
we will call a distributed national laboratory -- because we
think if offers the possibility for good science and it is well
matched to some of the resources requirements of the field, as
it exists today.
A request for novel support from society (here, from NSF) must
call forth a certain network of justifying arguement. Why is the
field worth supporting? What will be gained by the type of
support so proposed? Who are to be the recipients of the support
and why should resources be devoted to them, rather than to
others, who also have claims of importance and need? Given that
support per se is justified, why are the institutions and
devices proposed the preferred ones? We will attempt to answer
all these questions here. We have preferred to be quite definite
about many features of the proposal, feeling that sharp edges
make for clear perception. However, variations on these
mechanims are to be explored as considerations emerge that
somehow slipped by us or that we did not give due weight to.
This document is an initial proposal.
Though we will do so, it is always uncomfortable to argue the
importance of a scientific field. It implies comparison with the
importance of other scientific fields -- which only a mild
conspiracy of silence keeps decently muted. Science is all of a
piece. Each aspect of the universe calls forth the attempt to
study it and yields its unique type and quota of scientific
knowledge, which in the long run lets us understand ourselves
and find ways to carry out the business of living and improve
the commonweal.
We believe that understanding the nature of cognition -- of
perception and intelligence -- is one of the great keys to the
understanding of nature and one that is peculiarly relevant to
the attainment of applied ends -- almost by its very nature. We
believe that the development of the computer has permitted the
discovery of the physical nature of symbolic systems and with
that of some of the main mechanisms by which intelligence and
cognition are attained. We believe that the gain to be made to
society by a strong continued thrust on the nature of cognition
will be immense, both fundamentally in understanding ourselves
and the universe, and practically in constructing intelligent
systems. But we recongize both that what we study is emersed in
a larger scientific enterprise (Computer Science and
Pyschology), so that some of the importance and potential is
shared by the larger field. And also that there is no absolute
way to ajudicate our claims relative to the claims, potential
and fascination of other sciences.
We provide in Section 2 a very brief sketch of the current
geography and state of the attempt to understand cognition.
This will include something about the pattern of existing
support and enough history to provide the context for the
proposal. We then, in Section 3, discuss the research styles
appropriate to the field and thus the needs of the field for
resources of various kinds. Section 4 describes the proposed
institution. Enumeration of all the requirements on such an
institution, and how DNLCS is designed to meet them, is held to
the Section 5. Section 6, the last, provides some funding and
budgetary estimates.
2. COGNITIVE SCIENCE
3. THE NEED FOR FACILITIES AND INSTITUTIONS
4. THE BASIC PROPOSAL: A DISTRIBUTED NATIONAL LABORATORY
We propose to create a separate non-profit institution, DNLCS,
as the primary means of achieving our stated goals. We present
here a description of the structure of DNLCS and its operation,
without attempting to argue directly the advantages of the
proposal and how it deals with the various problems that attend
such a national laboratory. Section 5 provides this analysis in
a systematic way.
4.1 General Functions
DNLCS will arrange for the creation, maintenance and enhancement
of environments and facilities for doing research in
intelligence. It will do this by a two sided operation. On the
one side it will make contracts with research institutions for
the provision of these environments and facilities, and for
access to them by the scientific field at large. On the other
side, it will arrange with scientists for the use of these
facilities and envrionments which it caused to be created.
To describe the operation of DNLCS requires keeping separate two
types of participants. The first is the Research Center, which
provides facilities and envrionments; it might be a university,
a non-profit research organization, or even a commercial
organization for certain specialized types of facilities. The
second is the Scientist, who is an individual or group of
individuals who wish to use some facilities or environments to
pursue some particular research; they may exist at any
institution and are not restricted to Research Centers. DNLCS
deals directly with both sets of participants.
DNLCS engages in a transaction with a Reseach Center of the
following general nature. It decides that is wishes to acquire
facility, X, costing $C. It contracts with a Research Center to
provide these. They become operationally part of the Research
Center. The DNLCS acquires an interest in the total environment
of the Research Center of $D, where D is some fraction of the
total cost, C, of the specific facility. The remainder of the
value of X (C-D) is in effect the payment to the Research Center
to integrate X into its total facility and to maintain an
environment in which X can be effectively used to pursue
research by the community.
Given that DNLCS has an interest, D, in some environment, it
engages in a negotiation with scientists throughout the nation.
This would occur in the following familiar way. The DNLCS has a
review committee which accepts proposals to engage in research
at a specific Research Center. This research would involve some
mixture of travel to the Reseach Center to work in the
environment, with consequent use of the facilities there, and of
use via networks from the home bases of the scientists involved.
The DNLCS would have some additional functions of helping to
create and maintain the total scientific community, such as
supporting cooperation and communication, but these are
ancilliary to the main two above.
4.2 Types of Facilities
The facilities that DNLCS deals in are those proper to an
advanced computer science laboratory: computing power, primary
memory, secondary and tertiary memory, networks, engineering and
fabrication facilities, etc. There are three basic criteria for
when DNLCS should obtain a facility, rather than, for instance,
individual research projects obtaining it.
(1) The expense of the facility precludes many research
centers from obtaining it through normal funding
channels.
(2) The facility is a unique one with a low enough
demand so that only one or two such facilities are
needed by the total community.
(3) A cooperative effort is needed to agree upon a
development that will benefit the community, even though
the unit cost may not be too large for individual
research efforts.
A few examples will flesh out the general statements:
Computing systems with very large power
Either because of economies of scale or because
of peak power demands of particular research
efforts
Systems with very large primary memory (10↑8 words).
Terabit tertiary store for backup file storage and large
data bases.
Availability to the community via network is
especially important here
High bandwith networks
Large specialized processing engines
Eg systems for processing visual images with
1-10 billion operations per second (Bips)
Terminals for the network with appropriate local power
and storage to access the network facilities
The terminal systems themselves should not be
too expensive for individual research projects.
However, the development of such a terminal
might be. In addition, assuring the
availablility of such terminal systems on short
notice or for limited time periods may be
important and not easy to accomplish through
individual projects.
Image transduction facilities
Personal Lisp computers
This is again an item in which the final system
is inexpensive enough for individual projects
(their very point!), but development is not.
The operation of an Enineering Lab capable of designing
and fabricating digital systems.
Examples of such systems are multiple computer
systems, special interfaces, etc. Such
laboratories do not exist apart from a research
environment which they serve.
4.3 Institutional Structure of DNLCS
This aspect of the proposal is especially open to modification.
The intent is to be sure that a structure is posited that does
deal with all the considerations.
DNLCS will be a separate non-profit corporation. It will be
formed by a consortium of universities that are representative
of the entire spectrum of interest in the area of research. The
Board of Trustees will be drawn from these organizations and
will be broadly representative of the total scientific and
technical community and not limited to active scientists in the
area of intelligence.
There will be a full time Director. His duties, as outlined
below, though of critical importance and demanding of a person
of stature, do not require the whole of the Directors time. The
Director, therefore is free to engage in his own personal
research for such time as is available (it appears that it might
be about half time). Thus, the Director is to be an active,
though senior, scientist. Needless to say, the facilities in
which DNLCS has an interest are available to him.
The Director will have a staff sufficient to carry out the
administrative and support functions of DNLCS. These consist of
the management and support of the committees, the preparation of
proposals and reports to the funding agencies, the negotiations
with the Research Centers and the monitoring of the facilities
in which the DNLCS has an interest. A quite small staff would
appear to be adequate for this.
There will be a Proposal Review Committee, to consider and
decide upon proposals from individual scientists (and groups of
scientists). These proposals are for the use of the array of
facilites and environments in which DNLCS holds an interest. The
members of the Proposal Review Committee are selected from the
scientific community at large without respect to institutional
affiation. They will be active researchers in the area of
intelligence, so that the committee is a committe of scientific
peers. Selection of proposals will be based solely on scientific
merit, due regard being taken for the amount of resources being
requested relative to the supply. Membership will be for a
fixed term, and selection is made by the Board of Trustees. The
terms will be for 3 years and will be staggered.
There will be a Facilities Committee to formulate and consider
proposals for new facilities. It will accept and solicit
informal prosals and suggestions from the community and will
also create its own proposals. The proposals and suggestions
from the community are not to be acted upon, ie, approved or
disapproved qua proposals, but are inputs for the formulations
by the committee itself. It will be empowered to create ad hoc
study groups to consider specific facilities. The committee is
advisory to the Board of Trustees, who make the final decision
to proceed with the acquisiton of a particular facility.
Membership on this committee is, like that of the Proposal
Review Committeee, drawn from the relevant active scientific
community, for fixed terms, and made by the Board of Trustees.
The terms will be for 3 years and will be staggered. The Ad Hoc
Study Groups are formed from whoever in the total scientific
community is appropriate to the facility under investigation,
due account being taken of potential conflicts of interest.
There will be a number of Fellows of the DNLCS. These will be
appointed for fixed terms, as negotiated, with renewal possible.
The Fellows will conduct research at the various facilities and
will thus be active scientists. They will also have the function
of representing the DNLCS in the day to day interactions with
the Research Centers at which they reside. They are paid
directly by the DNLCS and are not to be members of the host
Research Center, though they may be on leave from other centers.
They may be provided suitable means and resources for whatever
small adminstrative tasks they have to perform. They provide
direct and continuous knowledge about the state of DNLCS's
interests in the local environment. They are to be recruited and
appointed by the Director. The number of Fellows will be
determined by the number and diversity of interests held by
DNLCS.
The primary functions of the Director, over and above the normal
conduct of business of DNLCS, are as follows. (1) He would
carry out the negotiations with the various Research Centers,
within the guidelines set down by the Board of Trustees in their
approval of the acquisition of specific facilities. He will
prepare and present for decision to the Board of Trustees the
final arrangements that commit the financial resources of DNLCS.
These will also go to the Facilities Committee for comment. (2)
He will prepare the proposals to the funding agencies with whom
the DNLCS negotiates for the funds with which to carry out the
acquisitions. In doing so he will be responsible for carrying
out these negotiations, and for interacting with whatever
evaluation and monitoring structure the funding agencies have
created. (3) He will formulate the policy for the conduct of
the committees of the DNLCS, to be approved by the Board of
Trustees. He will be a member ex officio of the Facilities
committee, but will not be a member of the Proposal Review
Committee.
The negotiations with venders who will supply the facilities in
which DNLCS has an interest will be carried out primarily by the
Research Center, with the DNLCS being a party to the
negotiations at the level of financial committment and final
approval relative to the extent of its interest involved.
The funding agencies would create (at their discretion) an
Advisory Committee of scientists and others to oversee the
activities of DNLCS. Presumably, such a committee would be
independent of the people involved in the operations of DNLCS
(ie, of its various committees) and of the research centers in
which DNLCS has interests.
Nothing would prevent the DNLCS from acquiring wholly owned
facilities. This would however be a device of last resort, if
the facilities could be obtained by any reasonable means through
existing research centers.
4.4 Nature of Arrangements with the Research Center
Each negotiation between DNLCS and a Research Center involves
the establishment of a fraction of the funds provided for the
facility to be under the control of the Research Center with the
remainder to be under the control of DNLCS. We can think in
terms of the DNLCS-fraction, which is the fraction that DNLCS
retains. The size of this fraction will vary considerably with
the situation and must be considered part of the negotiation in
establishing the facility.
In consideration of the fraction provided to it, the Research
Center agrees to provide an appropriate research environment in
which the facility in question resides and to integrate this
facility into his total system and reserach environment to
permit appropriate scientific activity. This includes
installing, maintaining and managing the facility. It also
includes developing it in terms of programming and engineering
improvements. It includes, finally, making the facilities
accessable to members of the scientific community, in
particular, to members who have been authorized by the DNLCS to
make use fo the Research Center. What accessibility means may
vary according to the type of facility. In general it will
include network access, where feasible. It will also include
residence access, ie, the provision of offices and other
amenities appropriate to visiting scientists. Most important,
the fraction provided to the Reserach Center covers the use of
the facility by the scientists of the Research Center. Thus,
scientists do not have to apply to DNLCS for the use of
facilities at their home institutions. Thus, although one
aspect of the funds provided to a Research Center is a
management fee in the usual sense, that is a quite modest aspect
of the total consideration.
DNLCS is not limited to acquiring new facilities in which DNLCS
provides the complete funding, simply locating these within
various Research Centers. DNLCS may purchase an interest in part
of an existing facility, or it may join with other funding
available in the Research Center to obtain a given facility.
Similarly, what DNLCS obtains for the community by its provision
of funds is an interest in the research environment as a whole
created and maintained by the Research Center. Occasionally,
DNLCS's interest will be rather narrowly defined in terms of
access to a specific facility. (The paradigm example being a
large tertiary store when all users are equally remote and the
Reseach Center is really providing a service to the total
community including itself.)
The rights of the DNLCS are expressed either in terms of amount
of service (measured in appropriate terms) to be available on a
specific facility or in terms of the right to have a visitor of
DNLCS's choosing in residence at the Research Center. The
normal course is that such visitors will be treated in no way
different that would any other scientific visitor to the
Research Center invited there by a member of the faculty or
staff. In particular, there will not be a special class
distinction of DNLSC visitors versus visitors of the Research
Center itself. Office space and amenities of the sort normal
for the Research Center will be provided.
A Research Center has the same rights with respect to a DNLSC
visitor as it would have with respect to any visitor it invited
itself. It may expect the visitor to conform to the
laboratories culture, it may request that he leave under
appropriate circumstances, etc. But except for these sorts of
circumstances, the Research Center does not have a veto power
over the people selected by the DNLSC, providing the selection
is made by the Proposal Review Committee as described above.
The DNLCS in its agreement with a Research Center may commit for
extended periods of time, thus providing the Research Center
with a measure of stable funding with respect to its facility.
The length of time of such commitment must be a direct
reflection of the duration of committement that the DNLCS can
obtain from its Funding Agency. With respect to those parts of
the DNLCS proposal to its Funding Agency that represent
agreements with a Research Center, the DNLCS can expect the
Research Center to help both in preparation and defense.
5. PROBLEMS AND PROMISE
Given the institutional structure laid out in Section 4 we need
to analyse how it attains the stated goals and how it deals with
the various problems that dog any attempt to create cooperative
arrangements for the use of shared resources.
5.1 Why a National Laboratory at all?
The fundmental reason for a national laboratory is that the
research facilities needed by the field are not available
through currently operating funding mechanisms for basic
research into cognitive science. They must be called into
being. There size would seem to preclude this happening by the
individual proposals of individual researchers through the
project grant mechanism.
As described earlier, facilities of modest size have been made
established by ARPA over the last decade at several places,
notably the four so-called AI Labatories (CMU, MIT, SRI and
Stanford). These facilities appear to be no longer supportable
as basic reserach environments by ARPA. If any of these places
were to attempt to obtain alternative funding on the same scale
as currently exists (and with a charter for basic research) it
would simply not be forthcoming. Even if such a proposal was
successful in getting reconsideration of the level of funding of
Cogntive Science within NSF, it would raise large issues about
the discrepency between the support for such an individual
Laboratory and what is available for the field as a whole. Thus,
it seems appropriate to raise the question of proper support for
the science in the context of the entire field rather than of
individual research projects or even specific research
institutions.
It is a required corallary that the facilities required by the
science demand the assembly and maintenance of expensive
technical environments which are subject to economies of scale.
Otherwise, funding at whatever level is justified should
properly be funneled directly to each individual researcher. But
this is manifestly not the case, as can be seen from the
existing laboratories and has been discussed earlier in
connection with the types of large facilities that are needed.
5.2 What will be gained?
5.3 Why a Distributed Laboratory?
There are two sets of reasons for proposing a distributed
laboratory rather than a unique facility, which is the expected
solution to facilities becoming too large for individual reserch
groups to command. One set of reasons concern our preferences
for multiple scientific environments. The other set concerns
the enabling conditions that make our preference attainable.
Science as we know it always tends to concentration -- to a few
strong environments with a number of smaller ones. But the
existence of unique (ie, really one-of-a-kind)centers runs
strongly counter to the need for a broadly based science capable
of supporting growth along many fronts. In fact, the United
States has developed a broadly based scientific community
located at many univerisities. Concentration remains, but there
may exist as many as a dozen really strong places in a given
scientific field or even a subfield. In consequence of this, we
observe, when unique facilities are required, the demands of
competative equality force that facility not to be located at
any univeristy. The AEC Physics National Laboratories give
testmony to this, as does NCAR. Only when unique history
intervenes, as in Scripps or Woods Hole, do these considerations
get over ridden. We believe that the university is the right
place for scientific environments to be created. While we are
in no position to comment on the effect in High Energy Physics
of most of its laboratories being located away form the campus,
we would prefer if at all possible to keep the work in Cognitive
Science both at multiple strong centers as well as at
univeristies.
There is also an immediate issue in that much of the value of
the present AI Laboratories lies in the environments that have
grown up there. To create a unique facility (a single National
Laboratory of Cognitive Science) is to throw away the existing
environments and attempt to create one anew. This follows since
the present situation is one in which the existing laboratories
cannot, in the expected future, maintain themselves without some
new substantial support.
The reasons why a distributed laboratory is feasible for
Cognitive Science lies in two aspects of the nature of computer
technology. The first is that the equipment, while large and
expensive for a single research group under current funding
patterns, is not so large as to preclude any but a single
installation. A large computer of the type appropriate for
Cogntive Science research costs 1-2 M$, not 25-50 M$, with
corresponding operating costs of .2 to .3 M$ rather than 4-6 M$.
The second is the ability to gain remote access to facilities
via networking. As the scientific community that has lived most
intimately with networking (on the ARPA Net), we can testify
both to its basic viability and to its currently limitations.
It is not a massive substitute yet for direct local computing
environments, but much computing, including interactive
computing, can be done via the network, and the situation is
bound to get better if technical development is continued. The
existance of network access permits another feature of
computational facilites to come to prominance, which is their
modularity. Many facilties can be developed and maintained in
separate environments, yet used by all. Thus specialization of
development is possible in a Cogntivie Science community, with
the entire community receiving the benefits. This works for
facilities such as large tertiary files, special transduction
facilities, special algorithm machines (under certain
circumstances).
Thus, a distributed community is probably the preferred style of
operation in the present technological domain. When Cognitive
Science requires an engine that costs 50 M$, so that at most it
can justify only one, then indeed it will have to live with the
difficulties of a unique facility.
5.4 What Assures that the Entire Scientific Community Benefits?
The critical feature that guarantees that the facilities of the
DNLCS are available to the entire community is that the
community controls the rights of access to them and disposes of
these rights in its own way. This is the ultimate import of the
contract between the Research Center and DNLCS and of the role
of the Proposal Review Committee in assigning pieces of the
DNLCS interest to individual investigators for their own use.
Thus, it is a critical feature of this proposal that the
individual Research Centers genuinely become "nationalized".
They agree to have visitors in their Laboratories who have the
same rights of local access as do the members of the Research
Center themselves, visitors who are selected by the DNLCS
Proposal Review Committee. This loss of autonomy is real and one
part of the fraction of the DNLCS supported facility assigned to
the Research Center is a payment for becoming thus nationalized.
5.5 What Prevents Perpetuation of Favored Research Centers
Given the existing situation, the initial Research Centers will
probably be drawn from the four existing AI Laboratories. This
is true, not only because this proposal is being generated in
part in repsonse to a problem these Laboratories face, but
because these AI Laboratories are the location of the main
environments by any objective assessment. Note, by the way,
that the phrase is "probably", since whether they do become loci
of DNLCS facilities depends on the negotiation between the DNLCS
and the facilites in the light of a program of facility
acquistion made by the DNLCS.
In any event, it becomes an important problem whether this
initial position of these Research Centers produces a
self-perpetuating situation in which these Research Centers
become the exclusive locus of ever larger facilities.
We believe it to be critical that this not happen, not only as a
matter of equity, but because we believe that there should be a
number of additional substantial environments of high quality in
the Cognitive Sciences. Furthermore, we believe that
self-perpetuation will not occur with the structure we have
proposed.
The mechanism that prevents this is the independence of the
DNLCS from the Research Centers and the nature of the
negotiation between them. Nothing prevents the DNLCS from
deciding to put funds at another Research Center. Nothing
prevents the DNLCS from deciding not to continue an association
at a Research Center in which it has had an interest.
Disengagement with a Research Center in which much there has
been much investmant and a long standing relation will of
course, be attended by all of the difficulties the always attend
(and should attend) the withdrawal of support. But, especially
if the given Research Center no longer provides facilities of
interest to the community, it can certainly happen and we
predict that it would. On a more positive note, we would expect
it to be highly likely that the DNLCS would attempt, as a
postive goal, to establish facilities at additional Research
Centers in order to help create a large number of substantial
environments. Of course, such a concern will come into play
only after the primary requirements exist for establishing the
new facility at all.
An important secondary condition in keeping the participating
Research Centers from becoming a closed set, is that DNLCS
obtains interests in total facilties, rather than simply dealing
in wholly owned facilities that it purchases and locates. Thus,
any Research Center that through other means builds itself up so
that members of the scientific would like to make use of its
facilities becomes a highly likely candidate for joining the
group of participating Research Centers.
5.6 What Prevents Isolation of the Cognitive Science Community?
A problem that may not have occurred to the reader is the twin
ones of isolation and obsolescence. If the DNLCS were indeed a
unique physically localized laboratory, then it faces the
problem of its facilities becoming less that state of the art
for the conduct of research. It also faces the problem of
becoming so isolated from the main stream scientifically as to
become second rate. These are both very real problems. They both
have an analogue in the scheme we are proposing, if the DNLCS
dealt in total systems, rather than in interests in
environments. In such a scheme, it would purchase a system and
locate it at some Research Center for management and development
purposes, and the scientific community would have rights of
access to that system, but not to the Research Center as a
whole. That is, a boundary would be drawn around the DNLCS's
facilities to keep them separate.
This would pose a problem of obsolesence in that this facility
would be going on in the midst of the rest of the Research
Center, but would gradually become less than state of the art.
This would be particular invidious in that the visiting
scientists would find themselves at a comparative disadvantage
and at short range. The boundary would also induce the second
feature of isolation, making the DNLSC visiting scientists a
socially isolated group. This need not happen under skillful
management, but the potential and the tendency would be there.
The proposed arrangement tends to solve both of these by buying
interests in the total environment as well as permitting the
interests to be bought in whatever facilities the DNLCS in its
wisdom desires and can negotiate. In particular, it is in the
interests of the Research Center to make its facilities as good
as possible, and there is no incentive to ignore some delimited
"DNLCS Facility", which becomes seen only as management
obligation. Furthermore, the DNLSC is always in the position of
trying to buy in to the most advanced facilities and
environments in the entire community.
5.7 Why Environments rather than just Facilities?
The notion of purchasing interests in environments rather than
just specific facilities may seem unncessarily esoteric. As the
discussion in the previous section on isolation and obsolescence
showed, there may be strong advantages to the scheme proposed,
and indeed fatal disadvantages to working just with physically
demarked facilities.
There are two other reasons that are also important for using
the environment, rather than the facility. The first is that one
of the things that scientists most want is the envirnment. They
do not want just cycles. Thus, the goals of the DNLCS could not
be achieved at all by a machine placed somewhere on the net
simply delivering service to whoever made connection. This is
so, not just because the network have not grown up to where they
deliver really rapid reliable response at high enough bandwidth,
though that it part of the issue (and a part that may gradually
fade as networks improve). But increasingly at the frontier of
cognitive science one is dealing with modifications of
architecture and with interface devices (both input and output)
that do not allow one to be isolated in total from the physical
computer systems on which one runs. Thus, an environment is also
a place with know how, engineering laboratories, and physical
computing systems.
It might be objected that such things are at the frontier of
computer science, but not of cognitive science, which should
just be considering the model of cognitive behavior on the
computers. But, as the earlier sections were at some pains to
show, this is simply not true. Artificial Intelligence reserach
has consistently pushed the boundaries of Computer Science. And
indeed what distinguish Cogntive Science from Computer Science
is only the extension into a concern with cognitive behavior in
man (and presumably other biological organisms). The press to
understand the space of computing structures remains, as it does
for the rest of Computer Science.
The second additional reason for dealing with envirionments is
that is keeps permeable the boundary between Cognitive Science
and the rest of Computer Science. In some of the existing AI
Laboratories, for instance, there is no boundary between those
concerned with AI and those working in the Computer Science.
They work on the same experimental machines, get involed in the
same software research efforts, etc. It would be a grave
diservice if creating the DNLCS were to create boundaries
between Cognitive Science as a whole and the the rest of
Computer Science. By working with environments one can accept
whatever situation exists in each particular the Research
Center.
5.8 Why isn't the DNLCS just a Disguised Granting Agency?
[THERE APPARENTLY ARE STIPULATIONS ON THE CONDUCT OF GRANTING
AGENCIES THAT DO NOT PERMIT THEM TO GIVE MONEY TO AN
ORGANIZATION FOR THE PURPOSE OF THE ORGANIZATION PROVIDING THE
MONEY TO OTHERS TO PERFORM ACTIVITIES NOT STATED IN ADVANCE IN
THE ORIGINAL GRANT OR CONTRACT. IS THIS TRUE? WHAT ABOUT THE
MATHEMATICAL SOCIAL SCIENCES BOARD, SUPPORTED BY NSF? IN ANY
EVENT LET US CONDUCT THE DISCUSSION ANYWAY]
The DNLCS performs a number of functions that would not be
performed by a granting agency. It acquires specific resources
and then makes these available to the scientific community. The
scientists who make proposals to DNLSC are not asking for
general funds to conduct their research, they are asking for a
portion of the specific resources that DNLSC already own and
maintains (though agreements with the Research Centers).
Presumably the investigators have research grants of their own,
which pays the normal costs of doing resarch. The costs, if any,
that DNLSC pays directly to the investigator are all those that
are completely incident upon the use of the facilities, namely,
special arrangments for network access, provision of special
terminals, etc. A rather clear policy needs to be set down here
on allowable aid, so that there is no misperception. But such
arrangements have been standard on many special resources
provided by Research Centers under grants from the Federal
Government, so it should be straighforward. The principle is
that in so far as possible, the scientist should obtain funds
for his research from his own grants and institution. He should
apply to DNLSC only for that portion that has to do with the
special computational facilities or access to environments that
DNLSC controls.
The negotiations with the Research Center are perhaps another
matter. Here the funds cannot be given by the government to
DNLSC without a fairly specific notion of what they are to be
spent for. [WHY NOT? -- BUT LETS ASSUME NOT] Thus, each year
there will be a proposal from the DNLSC to the funding agency
requesting specific amounts for specific projects. These will
include standing amounts from past agreements, which are in
effect the operating costs of the DNLSC, and also proposals for
new acquisitions. These latter could range from situations in
which the Board of Trustees has approved only the Acquistion
Proposal, to situations in which the negotiations with the
Research Center and perhaps even with Venders have been
essentially carried to an advanced state (due regard being taken
for procurement requirements on open bidding, etc.). In either
case, DNLCS is never the recipiant of unspecified funds from the
government with which it is free to give to others to conduct
their research.
5.9 Why Won't the DNLCS become a Bureaucratic Nightmare?
The institutional apparatus of the DNLCS is constructed to
provide clear representation of the interests of all parties
involved and to asssure the existence of appropriate
accountability and approval mechanisms. The intent is to create
a smoothly functioning system in which appropriate research
facilities that meet the needs of the specific scientific
community can come into being in a timely fashion.
One important feature of the DNLCS is that its administrative
staff is to be small. It will be working directly with the
committees, so that there should few delays there. Another is
that the DNLCS will be dealing with relatively large and long
standing arrangments with Research Centers, so that these should
not be continual sources of bureaucratic friction. In so far as
their have to be yearly modifications in expenses of facilities
this will get reflected in the proposal submitted by DNLCS to
the Funding Agency.
Another important feature of the DNLCS is that it does not
require that scientists at a Research Center get approval for
use of their home facilities. Thus, once the general agreement
to place and maintain a facilities at a Research Center is made,
the DNLCS gets out of the loop as far as the internal operation
of the Research Center is concerned, only being involved in the
assignment of the resources to those external to the Research
Center. Thus, it should not interpose at all on the style of
operation or control within a Research Center.
There are two places where there will be an interposition of a
process that may require extra time. One of these is the
determination by DNLSC, by means of study groups and the
Facilities Committee of what resources are required. Typically,
this is likely to stretch out in proportion to the size of the
facility. Whether DNLCS will add any additional time remains to
be seen. Indeed, it can be agued that DNLCS will shorten the
time involved substantially, since to get any of these
facilities would require in any event some form of cooperative
action between groups in the field as well as opening an
extensive dialogue with the Funding Agency. With DLNCS a smooth
organization for cooperative action will be in existence,
permitting this part of the negotiations to all be carried out
with much greater dispatch.
The other place is that DNLCS approval is in series with the
Funding Agency approval, so that a more extensive process must
go from the Research Center's point of view to get each years
funding settled down. This we think is a potential weak spot and
must be watched carefully. The antedotes seem only to be to keep
the entities to be negotiated as large and as long term as
possible (thus diminishing the absolute amount of negotiation to
be done) and to keep the DNLCS as small as possible to avoid any
extra layers of bureaucracy.
The provision of DNLCS Fellows is a device to maintain contiuous
high grade contact of the DNLCS with the facilities and
environments in which it has an interest, without interposing
any structure of official monitoring (eg, periodic visit of the
DNLSC to the Reserach Center site, which must be treated as
serious site visits to be prepared for). It is presumed that
the Director of DNLSC will be a still active scientist so that
he will in fact be intimately familiar with all the major
Research Centers in any event.
5.10 How will the Scientific Community Evolve with DNLCS?
6. FUNDING CALCULATIONS
The structure outlined here is meant to continue in being as an
institutional device for aiding the course of Cognitive Science.
Thus, what is appropriate at this early point is to describe the
continuing budget and to given a purely hypothetical schedule of
facilities that might be acquired. This should be done in a
sufficiently analytical way, so that some feeling can be had for
why funding of a specified level is appropriate.
6.1 The DNLCS Fraction
As described, for each facility that the DNLSC funds or helps
fund, it acquires an interest in the facilities of the Reseach
Center of specified sorts, depending on the particular type of
facility in question. The fraction will vary considerably with
the type of facility, and no hard and fast rules can be laid
down in advance of some experience. Also, one must take account
of the size of Research Center at which the facility must be
placed, since the fraction to the Research Center covers the use
by its own scientists.
[MORE]
6.2 The Operating Budget
6.3 Typical Facilities Budgets
6.4 The Initial Situation: The AI Laboratories
-------