Tos Lteklider @ ISI

From: Newell @ CMUe1OA

Date! {6 Apr 75

Re: AT ROAD MAP MEETING

CCt#r Carlstrom @ ISI, Fields #@ ISI» Russel) @® ISI, Simon @ CMUnMIOA
JMC @ SUMAT, LES @® SUWAI, CCG @ SUMAT, PHW @ MIT#AT,
Feigenbaum ® [SI-e Amare] @ ISI, Nitssom ® SRI@AI,
Sacerdoti @ SRI#AT

Lick: Twentyefour hours haS Permitted some reflections on
Monday's session ard the problems attendcent thereto,

(1) To Pepeat what was generally accected, implicitiy and
explicitivs by the AI seiemtists at the meeting! Tt is
aoprepriater, givem the current general comtext and smecifice ARPA
comtext, for the AI field to attempt a series of applications,
Such apelications would be a good thing for AI generally, as
well as for its specific relations with Ded,

(2) Aliso to repeat? Apolication opportunities must he
discovered, verified and exploited, Some fnstitutional means
must be found to expedite this, For it is clear that the AI
commumity by itself does mot have the expertise mor the
commectioms ta find high pavoff activities, Dave Russell, at
the end of the day, strongly posited a mechanism of a Randelike
agency with the mission of fimdina apelication opportunities,
verifying them, nutilding a bridge to the AI Labs- etc, There
would also exists aS am adjumct to this, am AI Applications
Tecmnical Group (or same such titled, consisting of
rPepresemtatives of the various Labs, CMU, MIT, SRIWAI, SuUMAT,
SU*HP, plus maybe others in related programs, such as FBN@SUS,
SOC=SUS, Amarel, etc, This arove would be aprime forum and
{nitiation point for these applications, I expressed some
concern that such am agent could come tnto being in short eneyah
orde* to satisfy the meeds of the day (implying that some
temporary verkicle would have to be erected), but Russell seemed
confident that such expedients were unnecessary, It would tke
better his way,

(3) It 48 extremely tmoortant to be sure that the payoffs of a
specific apriication are real, It is too easy to get
Sandbagged to have a seeming application turm to dross, Given
that ARPA is prepared to spend large fractions of its Al
eommunity fa relatively precious resource) om opreducing some
soecific apollcationss it is critical to substantiate the need

amd acceptability of am application, ARPA itself, though
imside the Dod amd much closer to the apptication sites than tre
Al commumity, does not itself have the expertise and,

imoorptantive the time to examine the situations enough to make
3 <FEIGENBAUMSNEWELL 3] SUN 20#APR=75 6:29PM

reliaple asSesgements,

(4) To be concreter om reflection I am mot at al! convinced that
the ¢tems om Heilmetier's list are all really in the categeary of
genuine apolication opportunities, I cannot speak about the
ASW problem, simce that is mot a single problem, but many «=
mamelvs. the ayestion of where to apply AI techniques all up and
dowm acomplex system, Sut the other two are bitty problems
aimed at highly snmecific targets, I did mot get any sense that
APRA really knew im detail whether the payoffs were real oF
simoly ephemeral, momentary opinions of one or two high level
peoole in the organizatioms connected with the anpnlications,
Heilmeier'’s carriagestrade philosophy reauires a really geod
mapketing and marketing research arm if it is to succeed, It
will do ARPA mo good i¢# it saquanders its substance om a bunch of
irrelevant mirages, The military scene is littered with the
dead bones of exoensive solutions which were onty monuments to
someone's folly,

(5) It appears that MIT=MAC {5 already spending substantial sums
(relative to the apparent size of the croblem) on the Morseecode
Problem, Surely it would not be fruitful for the Al community
to get further invelved in that one, I am not quite sure why
this one showed up on Heilmeter's list, but maybe it was just
meamt to be illustrative of what ARPA had mo decided to do,

(6) Om the language=spotting tasks, I need to reiterate what I
satd at the meeting, First,» I do not think the Al
Laos (im counterdistinction to the SUS Labs) should take on the
PrPoplem, they simoly would have to bulld yup much of the
expertise that the SUS Labs haver which would be a genuine
duplication, Seeonds, tT consider that the SURG has not been
asked to comsider taking om that task, I agree that you possibly
mentioned it to the SuUSC (Cthough I have forgotten it
comoletely)s but since tt was mot brought up im the context of a
Serious comfrontation with the SUS Sevear goals, I do wnat take
{t a8 a Serious orepesal, It would have been derilict of amy SUS
group to take on the tasks, givem how tightly the SUS program is
Strapped dowm to the 5 year goals,

I do believe we can consider this task and, if it §8 important
emoughs we cam consider folding ft fnmto the present contractors
{nm some way, But we do have to face the potential effect on the
S*yea"™ goals and to see how to work around them, As chairmam of
the SUSC,» I am autte witting to go around on that issue, But I
Peed a signal from you or Dave Caristom that indicates you want
to do that. This is a serious poimt and proposal I am makimg,
since I do mot want to ne accused later of having fiddled while
Rome burPnNed == of having not picked up this oroblem whem it was
importamt to do so, I would tike some feedback on this
Specific matter, The problem itself seems rather
Straightforwards given the current art, I would estimate half a
mam year for the techmical work, if dome at CMU givem all the
facilities, To this must he added the whole custumer interface,
which might be as much again, Much denends om details which I do
mot kmowr Of course (eq, how must ¢t be packaged and how much
must it cost), If the preblem could wait until after Nov76, you
$ <FEIGENBAUMSNEWELL od SUN 20"APRe75 &:29PM

could surelv get it taken on hy the SUS Labs if it were as
important a8 Heilmeter stimulates,

(7) Amother example of a SUSsrelated application is the Korean
onelime communication aide which you patsed as a problem ard
whieh I suggested a solution approach to seme time ago, I do
mot kmow what became of that. Aqaim, it would be derilict for
the SUS community to fold that im without at least explicitiy
facing the 5 year goals, A!) this stem, of course, from the fact
that we (read: ARPA and possibly Newellecum=SURG=Iinitiator)
wanted @ Program tightly fixed om <dImpressive gealss. and
therefore mot with much slack for such things, Agaim,y, I am
willimg to consider this,

(8) I cannot believe that the CAC is mot om the track of an
important application, It has two things wrong with dt: (1)
there 15 mot am immediate customer eager and ready to pay; (2)
motwithstanding SRI's search, there does not seem to be one
hiding out there quite yet (though I do mot know how intensive
that search was or 18), Yetr it does mot seem to me profitable
=e for ARPA, evem om its owm current terms == to Jerk that
effort up at the poots amd radically redirect it, Rather it
Seems to me critical to widen the scope to "Realetime operations
consulting" (mamely, how to helo someone carry out an operation
im repeal time) and to search for applications within this wider
sphere, The core of work on the CBC remains im fact stronaty
relevamt? and the new applications cam be grafted on,

(9) What should go im the Road Map? It seems clear to me that
the Road Map tor Friday mas its actinn camponent§ defined
imdependemtly of its substantive component, To wit, the
formation of the anplicationestfinding mechanism, defined above,
will mot be justifieds nor require Justification, from the a
Staterent of the current art or a statement of future scientific
goalSe However, this application propesal will differ from all
other sucm attempts by the promises implicitly extracted at the
meetimags by the AI Labs to enter into such am application=search
wholeheartedty,

The Substantive mode must oserhaos still he there by Saturday,
You dom't have much to work with, im terms of what was generated
before and during the meeting, Thuse I would attempt to get the
actiom comporment to stand in for the rest, Let me discuss each
of the Substantive components a tittler and then come back to
this,

(10) Whem a set of the worlds best scientists, being asked about
the their very own scientific domain, becomes tonquestied and
Produces anSwers unsatisfactory in a fipst year aualifier, then
the conclusion is mot that the science doesn't exist, tt is that
the auestion was posed wrongly or the situation inktboited
adequate response, You asked us», I thinks, toa do something
under constraints that communicated: (1) that none of our prior
attemots was to he considered satisfactory == that something new
and differemt was required: (2) that we adopt a form of
Specification of results amd of expectations that is foreign to
CS amd Al+r and largely foreign to science (mamely, to state in
s <FETGENBAUMDNERELL 31 SUN 20"APR#75 6:29PM

advamce the content of the scientific results to be expected ue
to several years in the futures 80 that the questions would only
be whether or whem the result would be attained, This prewenpts
the science and leaves us tomguestied),

For inatamces im the CMU proposal I have Just finished writing a
Statement about the basic sctemtific questions of AI amd the
high level propositions that characterize what we have found out
im AI, Apparently that is to he discarded ss mot adequate or
appropriate to the task == ang [| am to find yet another
Statement, different trom thate that is to pe adeavate to the
new (yet identical) task,

For imstamces Nils has just finished writing a pacer CIFIPS 74)
devoted to a summary of what AI has dane and what areas it kas
worked §m, Apparently that is to ce disearded as mot adequate to
the task == and Nils is to fimd yet amother statment, different
fron thats that is to be adequate to the mew fyet idenical)
task,

Let me Stramgly suggest, far instance, that as far as
characterizing the oresent state is cancermedcs, yoy take a copy
ef Nils JIFIPS paper and underline in pred the items in the
bibliography that are done im the ARPA AI Labs, amd om the many
charts that draw a map of the area, circle im red these same
items, This will give (1) a direct pieture of the coverage and
scone of the tield of AI that ARPA has given birth toy and (2) a
direct picture of the axtemt to which ARPA is responsibite for
these results and for the important omes,

Let me further Suggest that you pyt im frost of Nils paper the
fiest section om AI qoals from the CMU proposals, a8 givina a
high level coherent picture of what AI as a science is striving
to achieve and what im global terms it has found out,

These two items answer only the ayestion: What are the results
im AIT tm ft83 own scientific terms, They do not answer it fully,
but they will do as well as what you cam put together fm yet one
more attemot mia few hours,

(12) I do believe that several additianal descriptions of Al
Scientific results are possible that will appear to be more
Satisfactory to upper ARPA tham the two above itens (ifs indeed,
amything 8s Satisfactory), ] cammot carry out these
descrigetions jm the time avallable, indeed I think {t would take
a couple of months of very hard work (maybe more), But I can
Sketch amd <iliustrate one part of it (which is indeed based on
Past efforts to systemitize),

Progress im AI proceeds im terms of imereases im scientific
knowledge about the various components of the intelligent agente
compoments that are defined functionally, A standard division,
which corresponds in part to Nilsson's core areas, isi

> Recognition and description (Perception)
> Visien
> Speech
$  SFEIGENBAUMONEWELL 31 SUN 20"APRe75 o129PM

> Langquaae

Reoresentation

Prohtem Solving Methods

Centro! Structure

Assimilation & Accomodation (Learning)

vvvyv

Withim each compoment ome can describe a series of structures
Cor mechanisms) that are possibilities for this compoment, The
discovery of each such structure and mechanism {8 an advance for
AI amd a pesult, Verification, of course, if requireds it
comes, usually, from i{nmcoroporation im several total systems,
Knowledge about each mechanism grows with experimentation and
theoretical Sharpeming, Suen knowledger again whem verified
experimentally, constitutes scfentific results for Al, It
cOomsists mostly of statements of adequacy or sufficiemey in
specific task environments,

Thus,e the statement "what are the results of AI" at a givem date
is a listing of the various mechanisms (usually described by
conventional technical mames), plus the associated statements of
adequacy, This list grows over timer, amd it, rather than a
Oarametrizatian of mow good are the systems that cam be produced
constitutes the core tramsferable kmowledge of AI, This core ts
indeed transferrahla,s, precisely because it comsists of the
abstracted mecramiams which have beer Shown experimentally to be
useful fm several task environments,

T cannot praduce the lists of results for the total field,
mostiy because they have mot teen extracted, labelled and
organized im this way, T cam do it for ome subpart, that of
problem solving methods, Heres, much that we know cam be given
by specific methods (analaqous to the methods of mumerical
amalysis), A fairly good list iss:

> Generate and test
> HETT climibing
> Heuristic search
> Search stragegys
> Depth firsts, Breadth first, Best first,
Progressive Deepening
> Evaluation
> Evatuatton functions, level of aspiration,
Gupltcation avoidance, external limits
Matcning
Hypethestze and match
Meams ends analysis
Sunstitute & eliminate
Range restrictian
Abstraction planning

vvvvveyv

To find a short way to Say what we kmow, @gre about Hil]
Climbingr takes more energy than I have at this wee hour, We do
know the major things to beware of (Multi=modality, Mesas,s
Ridges» Ciiffs', we do have some empirical things to say about
whem Hd11 Climbing seems to work and when it doesn't, we do
have a wav of classifving the refinements of the method (as
Simplifed models of the hills which are used to predict the
' <FREYGENBAUMPNE WELL Pd SUN 20*APR#®75 6:29PM

optimum kill=step to take), And so on,

The existence of this list imolies a large kit of tools
available to be used im applications, amd indeeds when Ed
Feigenbaum says they used “Standard AI" in Dendrals he means
that the tecrniaques im Demdral oretty much are drawn from the
list of sSueh “Own amd characterized methods,

To plot the success cf the Peablem Solving Methods component
over time 1s to watch this list aqrow and/or the amount of
knowledge about each Such caompoment grow,

I do not know how much each of the other componmerntsS cam be so
characterized, though I expect it could be pushed ayite far, But
mot toniqhkt:

(13) Future goals of AI must be {nm general to extend the
mechanisms and Structures of each area and to show that they are
adequate to wider and more difficult preblems, The discover of
a mew method is mat to be asserted im advance as a goal, if it
could them the mew method would have been found! Sometimes one
kmOowS enough about a method of structure to specify as a goal
that it i8 to be explored, Te find the true scope of the ranae
Pest™ictionm methods, is such a (small) goal,

Goals» im the sense that you (read: ARPA) want them, are only to
be associated with systems or with instruments (Physics has such
goals for the eneray of interactions its accelerators will
Feache of the resolution of microscopes), There has mot pneen
any oifficultyv, as far as I «mow, Im determinina how to
Parpametrize the structure and nrerformance of specific narrow
classes of systens (such as SUSS of Demdral=like systems) when
tme task comes close enough to feasibility to make ¢t worth
adooting a syste as an AI goal, It cam surely be done far
various other specific classes of systems, though ft will not
covers thereby, all af Al's goals,

{I Rave Pum out of gas here *~ I believe more cam be said about
stating AI pure goals, but it Just escapes my fogebound mind, I
move on,)

(14) Applied goals, I pelieve are to be Stated in ome of two
WaYS,. These provide other ways of describing what AI has done
im terms of how ft cam contribute to these goals,

One way $8 how we started out to do it at the meeting, An
applied system is posited (ie, a total military system, such as
am ASW system), Then, withim that some potnats of AI application
are founde from which ome attempts to derive the AI that might
make a difference, This is a form of systems amalysisSs and ome
that can lead to a backward chain of available relevant research
results ands as well, of still meeded research == methods,
Structures, kmowledger experimentation, e etc, Meeded to do the
Job, I think we should deo a substantial amount of this, and I
belteve quite satisfactory roademap results would come out of
it, Unfortumeately, I believe that the effort per complex
total system must be a summer study=qroup sort of thina, ie,
py <FETGENBAUMONEWELL rt SUN 2O=APRe75 6329PM

apout what we peut imto the SUS initial report, But this is
exactly what is te be done by this Randelike agency (plus some
of us),

The Second way is to specify applied technologies, The two
prototypic examples at hand are the motion of a SUS technology
and (more pertinent) the matural language fronteend technol oay
that we all were talking about. One cam take the develonment of
SUch &@ technology as a goal and describe both what existing Al
already Provides and what mew fesearch is meeded to aet it,
This cam be carried out much more within the AI community,
thougs some sense for what {3 Peally fequired to make specific
aoplications qo is important, But agains it takes a fair sized
effort to lay out such a teeknological alternative, We could
commission Such explorations, (It would depend, I quess, on
Upper ARPA being prepared to consider such expansions, )

(15) We did assert rather strongly that there have been a number
of civilian applications of AI, eger im management science, in
design, etc, Ferreting these out and asking whether any of them
could be applied to military systems would he am additional
{mportant task for this applications organization (along with
some of us), Thiss of course, iS yet ome other way of stating Al
results,

wkkkkkks

IT am Pealiv to the end of my rope tomight (this morning) and I
will sand this cut after Herb gets a chance to look ft over, I
an willimg to work om expanding or modifying any piece of this,
IT am pretty much around fron here through the weekend,

ALN,

PS, Recall that I an expecting feedback om the Larnquace
Spotting issue,