ee DS ae De FF 00 OF OFS OF Oe ETD FE OE Um OF OE COE SOF Oe FH OS Gm GE SF Os OH WO FS Oe Oe Oe Ewe FE OF AR OS OH FF WO OF Gn OF ne
«< Reading: <CREARY2AI/SM,DOC33 6220[7)
Bel 7eJUL@e77 220578 29,1 7eJUL O77 22357829,189IUL@77 07302825

PDI

a a * ® « : ® « ® * VV, a % s a e ® & 9 ®
THE IMPACT OF HPP@STYLE AI WORK ON SCIENTIFIC METHODOLOGY

We May distinguish between two rather different forms that an
interest im scientific methodology might taket

1, The attempt to provide an accurate, systematic, and
i!luminating characterization of the selentific practice
of an historically given group of selentists, past or present,

2. The attempt to transform and improve selentific Practice in a
Given area, prompted by the availability of powerful mew tools,
and guided by the intuitive sense of methodological rightness
of scientists practicing in the given area,

The first of these approaches to the study of methodology {ts deseriptive,
while the second is primarily normative (though the new, {mproved methods
surely are to be deseribed, with both completeness and preeiston),

While each of these approaches is reflected to some degree in
the AT work done at SUMEX#AIM, it seems to me that the second one is
Preponderant, no less interesting philosophically than the first, and
Potentially much more sianificant from a practical point of view, Some
Serious description of existing scientific practice is attempted, but
this is mot pursued beyond the level accessible to introspection, and
tends to be used merely aS a starting point for improvements, Even in
Situations where the initial intent is only to make existing mathods
more fully explicits the HPP emphasis on high performance is likely to
lead to transformed methods designed to take advantage of the special
abilities of the eomouter,

Im the instances of the second approach that concern uss the
"powerful new tools" are of course the computer and asscciated AT
Programming techniques, New tools generally call for changed methods;
the correct approach to a given problem for an investiaator armed with a
slide rule and two graduate assistants wil! in general be quite
different from the right approach to the same problem for an
investigator eauicped with a POP=10 and a research staff of 20, In
effects, one engaged in the effort to develop appropriate new methods
must somehow elicit from his informed "methodological conscience" or
"scientific intuition" the best way to proceed in the mew situation,
This is nO easy tasks howevers Since the intellectual component here
Peferred to as "methodelogical conscience" seems to be more of a tester
than a generator of mew methods, Also, re rest {tself {s not a simple
a priori ome, but involves a strong empiri eaomponent, One plausible
approach to the generation problem {s to v A possible mew methods as
modified versions of existing methods,

Pursued at the jowest level, such methodological enterprises are
hardly distinguishable from the main body of selentifie research which,
with regularity if mot routine, results in the development of new methods,
But the HPP methedologist typically {s looking for the general im the
particulary he {@ especially interested in developing methods whose ranges
Of apelicability extend well bevond the contexts of thelr development,
So our methodoloaist ig faced with his own practical problem of induction,
If he sueceeds in finding methods with demonstrable applicability over a

__ Very Wide range of subject matters, then he will have contributed not

Just to (say) chemica) methodology, but to the methodology (Cand oerhaps
even the logic) of selence,

Consider now, for example, the thesis that "selentific discovery
ie heuristic search," in the context of work om DENDRAL and
Meta=DENDRAL, What {a the status of this thesis in the light of the
foregoing discussion? The thesis js not the descriptive conclusion of a
careful empirical investigation of human discovery processes, since no
BUeh investigation has been undertaken by the Heuristic Programming
Group, Nor is it merely the assertion of a Possibility, the result of a
"eufffeiemcy analysis" in Newell's Sense, Rather, it {8 (Cor at least
aspires to be) a normative methodological preserinption, Whether it {s
Peqarfded prior to the DENDRAL investigations as a preliminary working
hypothesisr or afterwards as a conelusion supported by those
nvestigations, the thesis asserts thats, in view of the information
Processing capabilities that are now becoming available, scientific
discovery OUGHT to be a matter of (computerwaided) heuristic search,
[The 'ought! occurring here is that of rationality, mot morality,]

Tt musty of course, be considered what qualifications Cif any)
shou}d be made to this thesis =* of equivalently» how broad an
application of the thesis js Justified by the experience with DENDRAL
and MetaeDENDRAL., It must also be considered whether more specitie
Versions of the thesis can be SUPported =" versions that relate certain
kinds of discovery problem with specific forms or organizations of
heurist{e search, And {t must be asked whether a reasonable range of
alternatives to the versions being defended have been conceived,
developed: and allowed to compete fairly, But these inductive {ssues do
not change the basic nature of the thesis as a prescription af normative
methodofogy, To the extent that investigators generally become
convinged that this prescription is correct and widely applicable on a
Practical basis, scientific discovery will in fact become a matter of
Ccomputer=alided) heuristic search,

I think that similar analyses would be appropriate for most

(though perhaps not auite all) of the methodological theses that AI work
at SUMEX=AIM might be found to supcort, If this 4s so, then the main
Impact of HPPestyle AI work on Scientific methodojogy {s the development
of news Improved methods that are precisely formulated, with seme

_ Improved understanding of current methods being produced as a —
byeproduct, At the minimum, the mew methods may be of quite limited
applicability, and not clearly better than many plausible but untried
alternatives, At the maximums they may be contributions to the logle of
science, Needless to savy this impact of AI om methodology is in its
earliest stages: being at present far more potential than actual,