aot a ,
i fe fer Lote by

Discussion of the paper
“artificial Intelligences Themes in the second decade"
by ZA. Feigendaun,
delivered at the IFIP Congress 1966,
Edinburgh, Seotland, Aug. 8, 1965.

by
Saul Amarel
RCA Leboratories

Princeton, NeJ. USA

I am in strong agreement with Feigenbaum's assessnent of the present
state of the art in A.J. and also with his views about the trends of vork
in this area and the nature of research problems that are now central and
deserve more of our attention, I would like to add a few comments that
are mainly intended to emphasize some of the points made by Feigenbaun.

Firat on the relationship betwaen the subject matter of A.I. and work
in "conventional" computer programaing. It is importent to renlize that
there is no sharp dividing line between the procedures of AeI. and the
procedures of today’s conventional software = both systems software and
applications software, Systems software ia concerned with problems of
language analysis and interpretation (macro= ox onblers, compilerg} and vith
a variety of problons of control and optinization of resources that are
becoming increasingly complex with the advent of tine sharing. The procedures
that are being developed and used for these problems strongly resenble
in thei: overall logical structure = end also in their technical detail «

to heuristic procedures for theorem proving and optimization of the
oo 2 =

type studied in A.I. research, Also, there are many application packarces
today in engineering and management that ere constructed on the basis of a
combination of aystematic and heuristic methods (e.g. viring and location of
componenta in integrated electronic modules, stock cutting, route scheduling).
Today's outstanding exauple of the application of A.I. ideas to an important
"real life” problen is Feigenbaum and Lederbderg's computer-based system for
spectromatry.

It ia becoming increasingly clear that the advanced procedures of A.I.
and the procedures that direct today's “useful” work of computing lie on some
sort of continuum. The key variables in thia continuum ara the amount of
systematic knowledge available about a problem class, and the deerce to which
this knowledge can be efficiently exploited for the solution of apecific
probleas in the class, the latter depends on the forn of the available
knowledge. At the one end of the continuum vhere emount of formal hmowledze
and its grede of utilisation are high we have most of today's “conventional”
programs. At the other end, where the enount of systematic knowledge and
its grade of utilisation are low, wa have the general, flexible, only partly
validated procedures of AeI., where a set of relatively weak problem-specific
principles are bombined with several powerful heuristic methods for organizing
search processes.

One of the important goals in the development of problem solving procedures
in A.I. is to enable a user to specify directly his problen to a computer = in
its “initial” high-level functional form - without having to apecify to the
computer an explicit procedure for solving it. This possibility would be a
major step in the road tovarda programming automation, It will bring the vast
information processing power of computers much closer to the man-vith-the-

problen, and it vill permit the application of computers to a much larrer domain
-3~

of intellectual tasks. Hovever, oven in this case the man is left with the

responsibility of formulating his problem to the computer in a marner which

promises a reasonably efficient solution-finding process. Hero we are
confronted vith the problem of problem representation, which, I fully agree

with Feigenbaum, is today's central provlem of AeIle In represonting 6

problem to a machine the man provides at present all the knowledze about the

problem that the machine can vork vith, end elso bk provides it in a specific
form which reflects his specific point of view = a point of view which may or
may not be fruitful for the solutionesearching process that he is forcing upon
the machine, ‘The question arises naturally whether it is posaible to endow
machines vith capabilities to shift problem reprosentations in an “appropriate”
direction. Such a capability will indeed provide us with problem solving
machines that combine great generality and power. I have been concerned with
this problem in the last few ycare, and I feel at present that in order to
realize beneficial shifte in problem representation we need to know more about
the following two general questions,

(1) How to choose the basic concepts for a language in which problem
aituationsa, rules for transitions betveen situations, and general
knowledge about the problem can be expressed. This is of particular
importance in "real life” problems, where the problom is not formulated
at the outset within a formal system but it is given verbaliy or it
includes information obtained from physical sensors. I think that this
4a the fundamental problem in the vork on “robotics”, i.0., how to
formulate descriptions of a physical environment — among the nultitude
of possible descriptions = that are most appropriate for the tasks on
hand. The question of choice of descriptions for vroblem-relevant

knowledge is also fundamental for the design of question-answering
systens with complex data bases, Choosing descriptions for data
bases in information systens is certainly an important part of the
problem of problem representations; in my opinion, this is a problen
that deserves much attention, and its study ia likely to produce many
fruitful resulta in the art of computing.

(2) How to proceed in the discovery of useful properties of a problem space
that can be used to transform it inte a space where sesrch for solution
is less difficult, and hov to use this knowledge in the formulation of a
better problem solving procedure. This involves the detection of
drrelevancies and redundancies, the recognition of regularities (such
as symmetries) in the space, and the ability to form more powerful rules
of action (say formation of macromoves from moves) that incorporate the
newly discovered knowledge. It is conceivable that the formation of
more powerful rules of action on the basia of new problem-specific
knowledge is mechanizable with ideas and techniques avallable at present,
To obtain non-trivial advances in this area we must know more about
problens of formation type.

The question (1) and the knowledge-creation part of the question (2) are
outside the realm of machines at present. However, I think that it is important
to start exploring them with a view to possible mechaniszations, via agpropriately
chosen case studies,

As a last comment I would like to indipate that most of the pregress (in
technique and theoretical understanding) in heuristic problem solving to date
has centred on problems of derivation type, where the objective is roughly to
construct a path between given boundaries. (e.g., theorem proving problems),
Problems of formation type have recedved so for leas attention, These problens

are more difficult then derivation problems, and they involve reasoning from
possible solutions to the problem conditions, Many “real life” problens,
notably design problems end diagnostic problems, are of this type.
Feigenbaun's spectrometry problem is largely a formation problem, Many
probleme of shift in problem representation are of formation type. I think
that as we move more and more {nto useful applications of Aes to complex
problems, and as we attempt to attein more problez solving generality via
computer handling of problem representations, we shall have to do much more

work on formation probleas.