NOT FOR QUOTATION WITHOUT PERMISSION OF THE AUTHOR
EXPERT 1 5 I,'LAhTAGEEEICT SUPPOF3' S Y S m S : SEKAI4TIC ISSUES
Ronald Id. Lee
Working Papers a r e interim reports on work of the International Institute for Applied Systems Analysis and have received only limited review. Views or opinions expressed herein do not necessarily represent those of the Institute or of its National Member Organizations.
INTERNATIONAL INSTITUTE FOR APPLIED SYSTEMS ANALYSIS 2361 Laxenburg, Austria
Expert systems hold great prorr~ise for technical aplication aress such as medical diagnosis or engineering design. They are, we argue, less promising for management applications. The reason is that manage-- b a r e not experts in the sense of possessix a forrnal body of knowledge w h c h they apply. The limitations of artificial intelligence approaches in managerial domains is explained in terms of semantic change, motivating attention towards management (decison) support systems.
Keywords: expert systems, management decision support systems, knowledge representation, formal semantics, applied epistemology.
INTRODUCTI ON
BUREAUCRACIES JrS KACXINES ORG-4NIZATI ONAL _4D..ZPT4Tl OK
MECHANICAL VS Iv~ANAGEZIAL COC-NITiOii A. T h e E g o P r o b l e m
B. T h e S e m a n t i c P r o b l e m
EXPERT SYSTEMS VS DECISIOK S U P P O R T =STEMS SUMMARY REMARKS
POST SCRIPT
ACKNOWLEDGMENTS REFERENCES
by Ronald
K.
LeeINTRODUCTIOK
AI
is getting market appeal. Expert systems, robotics and 5th gen- eration technology are ge ttisg serious recog nition in the economic plans for 1984 and beyond. The attempt here is to assess the potential impact ofAI
future technology on commercial organizations and other social institutions. Technology assessment suffers the lack of a convincing methodology. Hence the strategy here is not to try to predict t h e actual course of A1 innovations, but rather consider what would be the theoreti- cal limits to the technology.Our concern is mainly with
Al
technology in organizations, i.e., with groups of people working in cooperation. These remarks are not intended to apply to industrial robots, nor to single user expert systems, but r a t h e r t o what might be called a 'knowledge-based information system'(KEiS). Such a p ~ l i c a t i o n s ~ ~ o i l l c i s e e n to be the eventual result of a con- vergence of database macagement with
Al
knowiecige representation.To simplify the argmilent and cvoid large literature surveys, we take the liberty of imagining a future IBIS as a large scale theorem prover operatiw on a dateSase of logizal assertions about the organization 2nd its environment. This trend might be discerned from the litereture on 'logic and databases' (Gallzire e t nl. 1978, i98l) and the logic progrnm- ming discussions of relational datebases (Clocksin m d Mellish 1981, Coelho 1980, K o ~ ~ a l s k i 1979b).
The question is, what could such a ?1?31S do?
The principle function of an information system in organizations is to facilitate communication betv~een individuals thct are geographically and/or temporally separated. Unlike e.g., telephone or electronic mail, the advantage offered by a n information system accessing a s t r u c t u r e s database is that i t offers the possibility of making inferences o a the com- munications it intermediates. lnferencing facilitates the chunking of information (Miller 1956) necessary as communications flow upward in the management h e r a r c h y (Jacques 1976).
Jay Galbraith (1973, 1977) observes that h e r a r c h y itself is an infor- mation processing device, helping the organization to cope with the con- flicting pulls of a complex environment vs the limited attention and bounded rationality of management (Simon 1955). Knowledge-based information systems would, we expect, reduce the complexity by taking over more and more managerial problem solving.
But is there a limit? Ycouldn't the future, super-powerful, knowledge-based i n f o r m a t i ~ n system evzntual1~- eliminate the need for menagement? The arguments which f0110'~ lead to a negative conclusion.
A1 will make an important contribution to management problems, but the brave new world of the future M-ill not only be built with technology.
Arguments of t h s sort tend to rely 03 the 'unsiructuredness' of the managerial tesk as the basis for a view that i n f o r ~ a t i o n technology will a t best aid, but not replace management (Gorry ~ n d Scott-%!orton 1971, Keen and Scott-Xhorton 1978). Eut that argument eventually encounters a circularity if by 'structured problems' is meant those that have a decision algorithm. Technology has an untidy habit of advancing beyond problems that were previously thought impossible.
The arguments given here are based on two interconnected themes.
One is the problem of preferences (goals, values, free will), v h c h we argue that c o n p u t e r s don't have. (Computers don't intrinsically prefer chocolate to vanilla.) The other theme involves basic issues in semantics which, especially for organizations in dynamic, uncertain environments, provide fatal difficulties for even an idealized A1 system.
The arguments, interestingly, have a certain parallel with issues of bureaucracy. Various insights can perhaps be exchanged be tween A1 knowledge representation topics and the apparent limitations to bureau- cratic rationalization.
A characteristic of m a c h n 2 intelligence is that it is 'rule based'. If we consider only this softw~-re aspeci (and i g n o ~ e differences in processor hardware), then the most ubiquitous and successf~d ~ e x a m p l s of mechani- cal cognition are bu:-eaucracies. Yet while the projects to create various types of x t i f i c i ~ l iztelligence have a certain r o E a n c e and intellectual adventure aboui t h e n , the t e r m 'bureaucracy' seerns 2t best dreary and more often spite!,&. It is leden r:ith negative connotations cf plodding, brutish orgenizetions, insensitive to the indil~dual; indifferent to the exceptional.
Yet in more scientific usage, 'bureaucracy' is used neutrally as merely one form of administrctior,. The negative associations it has in popular usage gives empirical evidence that people's encounters with bureaucracies are often un2leasant. The definition of bureaucracy used here is based on I'feber (195S/1978), indiceting crganizations whose administration is based on explicit rules and procedures. This contrasts with a n idiosyncratic form of management based on persona! interest and the w h m s of the moment. Bureaucracies, then, a r e organizations whose behavior is 'rationalized' to eliminate such idiosyncratic tendencies. This gives rise to a concept of organizational r o l e , and explicit, detailed job descriptions. Personnel become substitutable; t h e organization takes on a mechanical consistency and permanence that outlives its members. In Weber's words,
Bureaucracy develops the more perfectly, the more it is "dehu- manized," t h e more completely it succeeds in eliminating from official business love, hatred, and all purely personal, irrational, and emotional elements which escape calculation (Weber
Consider h o ~ i this vieiv compares with standard models of compuia- tion. In automata theory (e.g., HopcroR and Ullmhn 1969), we vie?: a computer abstractly as a lenguage processor, transforming a n input string of sjrmbols t o output symbols (see Figure 1 a). In information sys- tems applications we can regard these cjrmbols as part of a commori language, cali it
LAW,
~ h i c h are assertions a b m t the 'rea! world' (organi- zation.~! eiiviromient). These assertions are normally stored in the organization's d a t ~ b a s e and the processor is invoked by queries, calls t o application programs, e t c . Hence, what we call the 'automaton' here is meant to include the entlre set of application programs, DBKS software,-
query interfaces, etc. (in hat ever future software designs you like).The a u t o ~ i a t o n , as l a w u z g e prccessar, is regarded as a grammar.
This grammar is itself de:ined in a notatior-, call it
LC.
Practically,LC
correspor~ds t c an arbitrary programming language." Ignoring efficiency considerations, we might regard
LC
as reducing to a set of production rules (Davis and King 1975) of the formIF <condition>
THEN
DO <action>If none of the various conditions are m e t , that is, if no rule is actuated, the default is inaction. The machine doesn't do anything it's not instructed t o do by one of its rules.
* It is common for LISP users and others to deny the disthction between dete and program.
Distinctions however depend on expository purpose. We could of course consider the language formed in t h e union of LRw and LC. ,The two languages are distinguished semantical- ly. The semantics of LRg i s all those expressons in the information system which denote real world phenomene. The semantics of L is machine operations. These are of course hopeless ly intertwined in all present day irnpgmentations, which is why we resort to talking about idealized machines.
A currently popular view of organizational management (e.g., March and Simon 1958) regards managers as information processors. Taking the metaphor literally, we might replace the automaton with a person (Figure l b ) . The 'programming' of this person might be in another language,
LB,
expressing the various bureaucratic rules and procedures this person is to follow.But if we regard LB (bureaucratic programming) abstractly in the way we did
LC
(computer programming), we encounter a problem if we use only production rules. As observed in a body of literature inorganizs.tiona! psy chclogy arid sociology ( e . g . , l~Caslo~!< 1 9 4 3 , hIcGreg2:- 1960, Cyert and hl:arch 11953, Karch a:id Olsen 1979) people are not nsic- ally id!e. They have their o ~ m indivi6u.l interests, goals, aspirations, eic.
which they are seekin2 to satisfy through their participation in the orgap- ization.
When these col-respoa2 to the interests snZ g o d s of the organization itself, we tend to rega1-d their independent behabior as 'initiat;ivel, other- wise it is considerec more as the dysfunctional pursuit of ' p e r s o n ~ l interest'. LB (bureaucratic programming) therefore contains another basic aspect. It not only orders the execution of desired behavior, but restrains the performance of undesired behavior. In Lee (18BO), we sug- gest t h a t a primitive structure of bureauzrztic softyare would therefore include the basic operators of deontie logic (von Itfright 196B), namely, (for q an arbitrary action):
O, q is obligatory P, q is permitted
'%
q is forbidden.Using negation, these operators are interdefinable. Permission to do q is equivalent to not being obligated not to do it (P,
-
"OWq), whlle forbid- d~ng q is being obligated not to do q (Fq-
OWq). Likewise, permission and prohbition (forbidding) a r e negates (Fq- "Pq ; Pq -
-Fq).
To be adequate as a language for bureaucratic procedures, these operators need to include an aspect of contingency (correspon&ng to the conditions in production rules). Unfortunately, contingency is not straightforward in deontic logic, and a number of proposals appear
(Hi!pinen 1031a, 1951b). Nate that discretiozary actions a r e those no', forbidden, hence permitted. A 'perfezt' bm-eaucl-acjr, in the sense of being complete!y riitionalized and determined, M-ould eliminate per-=is- sions
entire!^.
Everytl-iing ~vould be e i t h ~ r (contirigently) obligatorjr or forbiddea.Ths is of course a mazabre end unvrorkable design for any human organization. As Xorbert Wiener (i9G7j argued in the early days of ccjrn- puting, such e x t r e r ~ e regimentation is an inhuman use of human beirigs;
activities zre not only ecoriomlcally but moral!y better left to machines.
Jay Galbraith extends the info:-mation processing vie747 of orgenlzz- tions by classifying the environments they face on a two dimensional scale of 'complexity' and 'uncertainty' (Figure 2).
Complexity might be measured in terms of the number of informa- tion processing steps (mferences) required to plan the organization's actions.
Uncertainty is essentially the amount of surprise or unpredictability in the environment. This is different from simple contingencies, where the alternatives are foreseen, though the particular outcome is unknown.
Uncertainty involves completely surprising events. Thus, as uncertainty increases, planning, even contingent planning, becomes less effective.
The organization has to do more and more revision and adaptation whle the task is being performed. As a n analogy, consider planning a road trip.
uncertainty
9 ORGAN1 (discretionary) C organizaticns
C 0
rn
P 1 e
li
i t Y
7 I
MECH.LV1 ST1 C (Bureaucratic)
organizations
You take along a spare tire, extra oil, etc. for the foreseeable contingen- cies. Then there is a n earthquzke, which you didn't expect and so you have to completely revise your plans.
Rationalization, whether by bureaucratic or computer programs, is most effective in situations where complexity is high but uncertainty is low. Surprise requires re-programmw, and that tends to be time con- suming for either type of software. Left to their own, however, human beings can be quite adaptable. So, Galbraith observes, a counter-strategy in highly uncertain environments is to rely more on individual discretion, rather than trying to pre-program the individual's behavior. This leads to what Burns and ~ t a l k e r ' ( l 9 6 1 ) call 'organic'
-
as opposed to 'mechanis- tic' -forms of organization.T h s seems to be effectivs in ~ n c e i - t a i n eiiviron~ei=ts v;here complec- ity is lo^. However, beyond certain fzirly modest levels, cqai6ed hum?>
cognition suffers memory lirnitatioLx and computational biases (Slm.33 1955, Miller 1956, Tversky 2nd K a h e m a n 197;). Hov;, then, should an organization which faces a n enri-ironrilent that is both complex and uncEr- tain be administ e r ed?
It is in response to this question t h a t A1 research seems most promising. The appezl is thai .rvhile bureaucr2tic procedures are ge2- erally m i t t e n to be deterministic an6 inflexible,
AI
problen so!ti;lg research has led to approaches where numerous heuristics can be trie2 for a particular problsm sariznt. If one strategy 60esr~'t v~ork, we back- track and look for another. Several strategies may in fact be setisfactory in which case we can wisen Lhe scope to include less consequential p r o 5 lem variables and so provide adaptive, responsive solutions thzt simple, deterministic bureaucratic rriethods don't uncover.So far we have considered only the character of the instructions given to the problem processor, automaton vs human administrator. The instructions were expressed in languages LC and LB respectively. We now consider the language LRv; which these entities process. Typically t h e input stream includes some description of the problem, while the output stream is a course of action (to be followed by other entities in the organ- ization, whether machine or human or both).
Managerwnt texts ty?ic~.!ly dicdo th? activities of managers inio plcnnt:-~g a d C P ~ Z L T O ~ . In a platmirig p r ~ S ! e ~ the in2ut is current end predicted ir3orma;ioil a C ~ c t the external environment and the output is a plan to be fo!lorr.:ed by suSordinate entities (departments, people, mechines) in the organization. In 2 control problem the input is curreni and predicted facts about the internzl environment as compzre6 to a n existing plan. The o u t p ~ t is a revision to the plcn.
This view is q ~ l t e compatible with the conception of plaming i11 -41.
There a r e however two differences which eventuzlly limit the degree to which A1 technology cen tzke over m a n z g e r n e ~ t tzkes in the o~ganization.
We refer to these as the 'ego' and the 'semantic' problems.
A. The E p P r o b l s ~
People 1-lave preferences, c o n p u t e r s don't. Computers (as w-e knov;
them) will never prefer chocolate to vanilla. By preference we mean basic or intrinsic values, as opposed to instrumental or intermediate goals. Chess programs, for instance, have intermediate goals leading to the winning of the game. The goal of winning itself, however, is presumed prior to the system design.
The argument here is not absolute, but rather political. We could for instance imagine a robot with high priority heuristics for survival. This might lead down eventually to a sub-goals such as a taste for sweets or a compulsion to win a t chess. However, we aren't likely to allow such machines to indulge these preferences if they compete with our own.
(Note how Asimov's robots (1978) are programmed to be socially inferior.) Robot suffrage is not forthcoming.
The conve;-se conce7t to the sGcia2 r i g h t t o have an2 indlLge o r e ' s prefel-ences is i - ~ ~ 7 0 7 t ~ i b i l ; l Q The outcome of a cornpilter frecd triel is never to put tho com2uter in jail. I n t e r e s t l ~ g l y , not only people but zlsr, organizations ere granted this social status. k corporation (as wel!, a sovereign state) has inZependent legal resimnsibilitp; it can sign COP- t r a c t s , can be sued, etc.
The prefel-ewes (goals, va!ues) of an orgatlization are generdly regarded as deriving from t h e preferences of individu~ls. Capitdis:, econonics asscmes these to be the values of investors, Socizlist econorn- ics presumes these are imposes by the society a t large. Theories of organization, however, tend to ascribe a larger role to the preferences of people within the organization. Cyert and March (1963) note that the influence of stcckholders in large corporations has come t o b e m i n i m ~ l , and regard the preferences of managers as more significant in a predic- tive theory. Earlier, bureaucracies were characterized as orgznizatioiis where the influence of individual preferences was minimized. Mmagers fill prescribed roles and are substitutable over time. The organization's life is not limited t o the life of its members. On the other hand, the mechanistic character of bureaucracy which gives it permanence, also fixes i t s value structure. Hence railroads, post offices and the military continue to pursue ends that no longer coincide with social interests (Bouldmg 1978).
In the other extreme, March and Olsen (1979) discuss the nature of organizations where the goals expressed in the organization's formal c h a r t e r are vague and difficult to measure
-
e.g., universities, research institutions, charity organizations, e t c . Here the organization's goals areheavily influenced by those of indl:.idu?l mmem5ers, an? shift ir, a flilis Ts:aj7 in what they call a 'garbage c 2 s process'.
Deal and Kennedy (i9C2) provide an interesting intermediate viewpoint in their concept of 'corporzte culture' (see also Peters 1933).
In numerous case examples, for instance IBS!, Genera! Electric, Dupont, and 'Japan, Inc.', they observe coorcIinateZ, cohesive behavior yet v:ithout heavy bureaucratic regulation. The differentiating veriable, they argue, is that these organizatior~s have b ~ l t a s t r o n ~ organizetionz! c d t u r e which influences ar-d molds individual &ives and interests to coincide with the organizatim a t large. Conversely, individud preferences and values also exert influence on those of the organization. The dual membership of the individual in the corporate cult~u-e as ~veli as the cul- ture a t large ensures that the organization maintains goals and values compatible with its larger social context.
The point is that indivisual preferences play a n im2ortant role irL tn e adaptation and goodness-of-fit of the organization to its social en-?iron- ment. While we might conceive of e scenario where a robot or information system also displayed intrinsic preferences, this would be socially inad- missible (and has been in all the science fiction to date). It is of course not the preference itself but the tendency to indulge that preference t h a t matters. Having the right to indulge one's preferences (within socially defined bounds) amounts to political participation, a right still not won by all human beings, let alone robots.
We observed in the beginning of this section that an important func- tion of managers is planning. Planning is also a n important AI topic.
However, one limitation of
AI
systems to do organizational planning is inthe selection of the ultimate PI-eferences an5 v d u e s to which the p l a t s are di:-ected. Another limitation, e se-r;laritic cne: is discusse6 next.
B. The Semimtic PrubIeiz
Everyone knows that computationz! s e x a n t i c s is hard. vD'e argue that for management applicktions semeatics is in?ossible, so long as compaters don't have a social life.
Semantics is a rather touchy subject, since there c r e a n r m b e r of definitions that circulate and t h e y are rather hard. to sepzrate. Gen- erally, semantics is the corresponden-e between a symbol system (language) and its referrents.
In the first section we distinguished between LC, the language refer- ring to the computer and its operation, from Lxyi, which referred to the organizational environment. Tn cu-rent terminology this might be phrased as programming language semantics vs database semanti.cs. As before, we attempt to avoid the present debztes (e.g., various data management models vs semantic network representations) by skipping over aspects of psychological modeling, retrieval efficiency, etc, and assume that LRw can be characterized as a (first-order) pre&cate cal- c ulus l ang uag e
.
The other advantage of this assumption is that it helps to focus the immense literature on formal semantics without computational distrac- tions. In the p r e h c a t e calculus (data management and semantic nets as well) we typically make the assumption that semantics follows syntax.
That is, the semantics of complex expressions is constructible from the
sercantics of its s j ~ n t a c t i c c o n ~ t i t u e ~ t s . ( D o ~ t j 7 e t zl. 18EI:Zh. 3 ) . T h s :s Frege's 'Principls of Com~ositionslity'." The ro!e of the usec!; lo,-iczl cox- nectives an6 quantifiers in coristructirg the s e m ~ s t i c s of first order assertions is well studied (van Fraassen 1971). What remains is t h e semantics of the o7en vocabulzry of the logic, namely ii~dividuz! and predicate nernes. The epproaches a t t h s point divide roughly into t v . : ~ c a m p s , what we ivil call the eztoi?sias;zl a n d i n t e n s i g n a l vie?,? oinis.
Extensional Semnntizs
The extensional viewpoint is doinizant in forrilal logic, originating mainly from the model theory of Tzrski (1955). Here, indlvidunl objects a r e regarded as primitive, leslving generic properties and relationships t o be defined s e t theoreticelly. An interpretation or m o d e l , of a giver, (first o r d e r ) predicate logic therefore begins with the assumption of a domain of individuals, D, a n d a n interpretation function, F, which m a p s individual n a m e s t o individuals i n D, 1-place predicates t o s a b s e t s of D, n-place predicates t o relations on D, e t c . Hence a model Id of a language L has t h e form
This is entirely satisfactory as long a s t h e population of individuals i n D c a n b e clearly specified, a n d t h e y don't change.
* Here we are speaking of formal, constructed languages. The principle of compodtionelity doesn't always hold in natural language, e.g., for proper nouns like 'Marilyn Monroe' or norni- nal compounds like 'red herring' where the referrent of the expression is not constructable from the referents of it's component words.
Ho~::e:~er, a pro5:em for ~ i l c n ? . ~ ement applications is the? o r g ~ n i z a - tions a.xd their eni7iro,m.e-is d c chznge. Change is fundernental to econon~ic growth; it c m ' t be ignored. An obvious step is to extend the model to indude a time Zirnension, T, so t h a t D includes all individusis existing a t different tirr.es. X02els of the language are then of the form:
T h s , however, encounters difficu!ties when we consider aspects of
the f u f u r e . Much of management is concerned with p!anrf-ng. Since
there may be a variety of a!ternate or contingent plans, we must likewise consider multiple futures. This leads to another extension t o t h e model including so-called possible ~vorlds, W, hence adopting models of t h e form:
T h s is essentially t h e ontology proposed by Kontague (see Do~vty e t al: 1981, Lee 1981). While t h s enables a
mathematical!^
elegant solution, the question is whether it is still semantics. If semantics is the correspondence between symbols and the world, but if the world is not merely t h e actual world (past and present) but also future and hypotheti- cal worlds, we have to consider how it is we know about these other worlds.Strawson (1959) points out t h a t the principle basis for our shared epistemology is reference within a common spatial/temporal framework.
Possible worlds are mental constructions, Gedanken experiments. They are outside the framework of external reference and so a r e questionable as a basis for mutual understanding. We r e t u r n to this problem shortly.
Intensions1 Semantics
The intensione! viewrpo;nt is more c h ~ r a c t e r i s t i c of the A! paradigm (especially semantic net representations). Here, it is not indivi?-ual objects that & r e primitive, but rether generic properties and relation- ships. Particular objects and events zre seen as instances of these gen- eric concepts. For e:;atrqle, we postulate primitiv? conce?ts, MALE, FEKALE, SPOUSE,
CHILD
a.nd from these are able tc define the entire vocabulary of kinship relations. Pzrticular cases o: family trees, e t c , are regarded as 'ins'iantiations' of these generic co:lcepts.The i n t e n s i o n ~ l a p p r ~ a c h is entirely satisfactory for whet we might call idealized or artificial subject domeins, M-here the scope of variation is fixed theoretically or by explicit rules. However, the intensional approach also has difficulties, especial!^ in describing real world domeins where no theoretical foundation exists. For example, suppose we want t o develop a concept, LEMON. We then seek to e1abo:-ate the essential pro- perties of lemons. T h s might be a property list s o m e t h n g like:
COLOR: YELLOW
SHAPE: OVAL
TEXTURE : BUMPY TASTE : ACID
The problem, typically, with real world domains is t h a t we can't simply define what a LEMON is, but rather our definition has to correspond to what the users of the system conceive lemons to be. Now we run into the so-called 'criteria1 properties' problem. We want a s e t of properties t h a t in conjunction uniquely selects out lemons and only lemons from the
various objects in the er1i7ironment. The 2roblern here is tv-sfo!?: that too many tliinzs walify (e.g., yellow limes) and the Sefirdti~n e~:cludes atypical lemons (e.g., green lemons, lemons that a r e n ' t ovsl, etc.).
Yr'ittgenstein (1853/i958) is a classic eleboration of these diffici~lties.
There is a n interesting relationship between the effez'iiver,ess of the intensional appronzh and the status of- the science o? that subject domain. Chemistry, f o r i~x'iecce, provides a criteria1 defiaition for water (as H20). Psycholc,gy, by contrast, has no critel-id definitions for such phenornenz as intelligence or creativi:;;.
The problem seems all the worse in the social/econ3mic domains that are most common to management problems. Teke for i n s t a x e the mudane exam2le of chairs. Is there a s i ~ z l e physical characteristic t h a t chairs have in common? Consider such exzmpies as rocking chzirs, stuffed chairs, been-bag chairs, plestic inflata5le chairs. I t seems t h a t what is common to them all is not what they are, but what we do vyith them, namely sit. But this is no longer a 2 actual property, but rather a propensity or disposition, w h c h leads to similar epistemological difficul- ties as with possible worlds. (Rescher ( 1975:Ch.7) comments on disposi- tional properties and possible worlds.)
A Sociological View of Semantics
- --
Both the extensional and intensional approaches to semantics suffer epis temologic a1 difficulties, especially in the social/ec onomic domains typical for management. This leads to an examination of the mechanisms by which we come to know and use the terms of our everyday language.
If we fo1lo~- the extensiozzl a?prczch, then c e r main focus xill b e 3 2
our kno~vledge 2nd icier-tiflcation of indivic'-uzis (pzopk arrd objects). This brings zttentio-1 to the ser:iantics of pl-o>er nzriles and the identifica:ion codes we assign to machir~es and oiher objects. As Kent (1978) points out, these zre of f ~ l ~ d z m e n i a ! concern in datz precessing applications, map- pirg database recorss to invento~y, equipment, personnel, customers, su?plie:s, e t c .
How a r e these names associated to individuals? In the case of mznufactured objects, quite often the identifying name is stamped directly on the object. In the case of names of persons znC compaaies, the identification relies hezvily on honest reportip2 of their narnes by the entities themselves, e .g., on emp!oyment applications, sales orders, etc.
The point is t h a t the organization doesn't have to r e c o g n i z ~ these indivi- duals through some collection of identifying properties, it is simply fold, e.g., "I am John Doe," "Here is the
XYZ
comparijT."The point applies much more broadly. Most of vrktat we knou- about other individuals (people, places, things) that are temporally or geo- graphcally distant is what we have been told. The proper name provides a tag to which various characteristics are attached. The names them- selves are passed from one person to the next in a series of 'causal chains' of reference, leading back to a direct identification of the indivi- dual. Sometimes, in the case of multiple names for the same individual, the causal chains may separate, leading to assertions like
Mark Twain
=
Samuel Clemenshaving a n infc:.:xakisre content rather ikan b e i r i a tauto!cglnal Identity.
Kripke (1971, 1972) applies this concept of causal chains in a forv~ai-6 fashion in c h a r a c t o r i z i ~ g possible worlds. "Possible w?orids c:-e r ~ o t far- away planets," they are rather c o ~ ~ s t r u c t e d , based on known, a c t u ~ l refer- ences.
Cocsider, for ir~stance, a scenario beginning with the supposiiisn that Ronald Reagan is bald. The question arises, how do you knov: it's Ronald Reagan if, in this possible w ~ r l d , he has different properties. ( Y e can exaggerate the case - suppose Ronald Rezgan is really a robot, mznu-fac- tured on Kars, etc. -this is called the problex of 'trans-world identifica- tion of individuals'.) Ikipl-:e's point is that v;e don't have to r e c o g x i z e Ronald Reagan in this v,iorl<, we s t i p u l a t e that he is the s e x e in our con- struction of the scenario. The proper name Ronald Reagan is a 'rizid designator'.
Putnam (i970, 1978) suggests a some:.vSat similar explanation to our understanding of generic concepts like 'lemon' and 'chair'. Consider the first example of 'lemons'. Being a poor cook, my concept of lemons is fairly rudimentary. I surely couldn't tell a lemon from a yellow lime. Yet I don't often make mistakes in shopping for them. How do I manage? 1 go t o the supermarket and look for the fruit section. There, typically, is a case labeled 'lemons', where 1 draw my selection. I rely heavily on the supermarket's knowledge to know what lemons are. But how does the supermarket know? They make purchases orders to a distributor requesting s h p m e n t of 'lemons'. How does the distributor know? They order 'lemons' from certain fruit growers. How do t h e fruit growers know? Eventually the chain goes back t o a botanist or agronomist who
has ce:-tain sciantific criteria f2r 1eno;-is.
Kosv consider the coiicept, chair. Again sv~: car: foiloiv; the chain of refereace back, t h s t l m to c e r t ~ i n chair mazufacturing com2anies. But how do they knov:. :$:hat a chair is? They specify that their products are chzirs. Thus one eiite~prising company may stuff burlap bags % i i h shred- ded st;yrofoam an< mcrke: it i s 2. 'pil!o~:. cheir'. Another might fold and paint pieces G? cardboer2 sellng them as 'throv?-avxy chairs'. The suc- cess of their markziing also succeeds in modifying the concept of chair.
The eifect of these 2:-gments is to iiltroduce a sociological concep- tion of semactics, whzt Schvartz (1977) calls the ' n e ~ v theory of refer- ence'. It gives a convincing accomi of wk~j7 semantics is so difficdt t o do computationally: ser,lantizs isn't fuzzy, it's social. For many of our terms, e.g., lemori, c h i i r , the extensicn of the concept is quite exacting.
A t h n g is a lemon [chair) or it is not. Hosi-ever, the cclgniiion t h a t makes this discri~~iinztioii is not nn indiviZucl one, but rather a cooperation of a broad social netwo-k. As Putnzrn observes, we tend to regard words like hand tools i h a t we use insiiidually. For many words, a more fitting meta- phor is to compare them to a big ocean liner that requires a crew of hun- dreds for its operation.
EXPERT
SYSTEMS VS DECISION SUPPOHT SYSTEMSExpert systems are typically built to model individual expertise, e.g., a doctor, a travel agent, an automechanic. The view, generally, is of an independently operating problem solver.
Maczgers dcYi1t q p e s r to be e>:>eris in this same sense. Jv!intzSerg (1973), in an empiricel study of the activities of high leire! e:recutives:
notes that a great pol-tion of marlagenial activity is spent in c o m m ~ ~ i c e - tion, observation and diitn gathering. K<oreover, some 70% of their time is spent in informa.! meetings and committees. Indeed, in this s a m ~ l e , managers 01117 spent about 22% of thzir time in isolated concentration.
The s ~ g e s t i o z here is that m a a g e r s , rather then possessing an indicidu- aljzed expertise, are more like specialized nodes in a larger 'organizn- tional cogniticn'. Organizntions in turn, react and pcrticipate in a larger 'social cognition' in their attempts to mzrket nev; products and/or nove!
services.
An important part of the manager's activity is to observe and under- stand changes and trends in the marlcet, the economic, legal and sociel environments. Much of this Is not simply s h f t s in magnitlude on pre- defined dimensional szales. (Were this so, mathernnticz! models would surely have a bigger impact on manzgerial practice.) Instead, mznagerid cognition often involves the modification of primitive concepts. For instance, the range of phenomena we call an 'automobile' changes from year to year. Each competitive innovation, each new marketing angle.
each special interest group expands and re-organizes the phenomena the manager includes in his/her conceptual framework. And, given t h a t his/her contact with the world is primarily through linguistic interac- tions, the semantics of organizational language is constantly shifting.
Because mechanical inference relies on a stable, fixed semantics, the utility of an idealized, fully integrated, knowledge-based inference system will be limited to organizations in completely stable environments.
Similar criticisms can be made of bureaucratic rationalizati~n (Loe 19E3).
The c o ~ c l u s i o n to be d r a v ~ r ? is that integrated information systems will only be of use for those aspects of the or~an;zation's activities where semantic stability can be mzintained. This conciusion corresponds to the empirical observations rilade by Gorry and Scott-KIortor? (1971), which 1 ~ d to the conception of 'decision s u p p r t systems' (e.g., Keen and Scott- Morton (1978), Bonczek et a:. (lSEi), Fick aad S p r a g ~ e (19SOj, So1 (1993).
The underlying idea in the DSS work is to promote t h e dsvelo2ment of technology which, raiher t h . n r e p l ~ c e humzn cognition, sseks to assist and augment it. The trend seems to be to~;ards developing DSS 'genera- tors' which provide computational building blocks v2~ich can be variously structured for different ad-hoc, decision situations.
Interestingly, despite the iiedely recognized importance cf group decision making, nearly all CSS packages are oriented t o ~ r a r d s assisting the individual manager in isolation. The e x ~ l a n a t i o x may be semantic:
an individual can assign an interpretation to a particular syntactic representation (s)he invents. In a group setting however, the semantics is n e g o t i a t e d , and our technology so far seems to have had little effect on these socio-linguistic processes.
SUMMARYREMARKS
The preceding arguments can be summarized i n t h e following state- ment: we make words m e a n what w e want. Three aspects a r e emphasized.
Semantizs is plestic. A s Tarskian model theory so bluntly poilits out, the semzatics cf e lznguage is ar- inierpretation assigned to it. Certain truths (logical truth) are t a u t o l o g ~ u s i s that they h.olZ under any interpretation (true in all possijre m ~ < e l s ) . In organizational app!ications, ho~vever, we are more concerned with specific interpretations (synthetic truths, true in scjrne models, no: true in others). The validity of the inferences drawn depends on the stability of this interpretation. For example,
LEKON (x) 4
YELLOW(x)
is true if in fact all lemons zre yello?~.~, but fails if some botmist succeeds in g e n e r z t i q a strain with different colors an6 declares that they, too, are lemol-is.
We mkke words meari what we want.
Semantic change has a pragmatic component, depending on t h e interests, preferences and values of its users.
We make words mean what we want.
Semantics is plastic, pragmatic, but also the product of social consensus.
Indeed, i t is not only socially determined, but socially understood.
POST SC,R3"
The pw-pose of- this p q e r has beer, mainly to elaborste a p r o b l e : ~ rather than proaose spszific solutions. The point certainly hzs n o t besn to discourage furth2r A! research. Rather, it may serve to explain some of the frustratioa felt in meny of attempts a t kno~vledge representation, particularly in mar~sgerial applications. As we suggest here, the p r o b k m may be overif~helmiaglj~ difficult, requiring dtimately a forEal explicztion of all of s o c k t y . If that is the case, we ~ ~ o u l d do ~l;ell to seek out more achievable goals and strztegies.
Likewise, vre have to be careful not to overstate our claims. -4s pointed out in the beginning, A1 is getting market appeal. Big mccey is shifting. But the people b e h n d those big deeisions aren't techniciens r,or theoreticians. They a r e n ' t accustomed to our tendency to e x t r q o l a t e world shaking impl',caticns from toy-sized implementetions. They may actually believe us. And the plans for 1 9 8 4 are in the making now.
ACKNOKWDGEfEWTS
The author gratefully acknowledges the stimulating interactions with Steven Kimbrough, Eckehart Kohler and Amilcar Sernadas on these topics. As well, Werner Shimanovich and other members of the 'Vienna Quadrangle' (our humble remake of the Vienna Circle) provided a general background of discussion linking artificial intelligence and formal philoso- phy.
Asimov, I. 1978. I, R o b o t . Kew York: Fawcett.
Bonczek, R.H., C.MT. Hoisapp!e 2nd A.B. IYhnston. i961. F o u z d c t i o n s of D e c i s i o n S u p p o r t S y s t e m . New York: Academic Press .
Boulding, K.E. 1978. S t a b l ~ P e a c e . Austin: University of Texas Press.
Burns, T., and G.M. Stalker. 1961. t h ~ Idanagenzent of I n n o v a t i o n . Lon- don: Tavistock.
Clocksin, W.F. and C.S. Mellish. 19Bi. P r o g r a m m i n g in Prolog. New York:
Springer-Verlag.
Coelho, H., J.C. Cotta and L.M. Pereira. 1980. H o w t o S o l v e I t W i t h Prolog.
2nd Edition. Lisbon: Laboratdrio Nacional de Engenharia Civil.
Cyert, R.M. and J.G. March. 1963. A B e h a v i o r a l T h e o r y of t h e f i r m . Englewood Cliffs, New Jersey: Prentice-Hall.
Davis, R., and J. King. 1975. An Overview of Production Systems. Stan- ford
AI
Lab Memo AIM-271, Stanford Computer Science Report.STAN-CS-75-524. Stanford, California.
Deal, T.E. and Kennedy A.A. 1982. C o r p o r a t e C u l t u r e s . Reading, Mas- sachusetts: Addison-Wesley.
Dowty, D.R., R.W. Wall, and S. Peters. 1981. I n t r o d u c t i o n t o M o n t a g u e S e m a n t i c s . Boston: Reidel.
Fick, G . and R.H. Sprague, Jr. e2s. 1980. Dac.lsi3n S u ~ ; ~ o r f Systen-&s:
Issues and Ch~Llenaes. Oxford: Pergainon Press.
Galbraith, J. 1973. D e s i ~ r ~ i n g CcrnpLzz O T ~ u z i z a t i ~ n s . ReaCing, 'n'ass. : Addison-Wesley P u b l i c e t i ~ z s .
Galbraith, J. 1977. Qrganizatio:~ Design. Reading, Kassachusettj::
Addison-Mresley,
Gallaire, H., and J. Minlter. eds. 1978. Logic and Data B ~ F E S . Nev: York and London: P l e n m Press.
Gallaire, H., J. Kinker and J.K. K i c ~ l a s . eds. 1981, A d v a n c e s i n DatczBase Theory Volume 1. Kevi York a n d Lcndon: P l e n ~ n P r e s .
Gorry, G.A. and M.S. Scott-Xorton. 1871. A Frameu-ork for Xanagernent lnformation Systems. SLom. Kanaj,oment Revisw, 13(1):55-70.
Hilpinen, R. 1961a. ed. Deor~tic Logic: I n t r c d u c f o r ? ~ and S y s t e m a t i c Readings. Dordrecht: D. Reidel.
Hilpinen,
R.
1981b. ed. Mew S t u d i e s in Deontic Logic. Dordrecht: D.Reidel.
Hopcroft, J.E. and J.d. Ullman. 1963. F ~ r m a l L ~ n g u a g ~ s cr,d their Rela- t i o n t o Automata. Reading, Idassachusetts: Addisoc-Wesley.
Jacques, E. 1976. A G ~ ~ E Y z L Theory of Bureczzcracy. London:
Heinemann.
Keen, P.G.W. and K.S. Scott-Xsrtcn. 1976. Dzcision Sup?ort S y s t e m s . Reading, Massachusetts: A66ison-I'fesley.
Kent, W. 1978. Data and Reality. Amsterdam: North-Holland.
Kowalski, R. 1979a. Algorithm = Logic
+
Control. Communications of t h e ACM, 22(7):424-436.Kowalski, R. 1979b. Logic for Problem Solving. New York and Oxford:
North Holland.
Kripke, S. 1971, Identity and Necessity. In M.K. Munitz, ed. I d e n t i t y a n d Individuation., pp. 135-164. New York: New York University Press.
Also reprinted in: S.P. Schwartz, ed. 1977. N a m i n g , Necessity, and Natural Kinds, London: Cornell University Press.
Kripke, S.A. 1972. Naming and Necessity. In Davidson and Harman, eds., S e m a n t i c s of Natural Language pp.253-355.
Lee, R.M. 1980. Analyzing Red Tape: The Performative vs Informative Roles of Bureaucratic Documents. WP-80-164. Laxenburg, Austria:
International Institute for Applied Systems Analysis.
Lee,
R.M.
1981. CANDID Description of Commercial and Financial Con- cepts: A Formal Semantics Approach to Knowledge Representation.WP-81-162. Laxenburg, Austria: Interns.iioz&! Institute for Ap2lied Systems Ana!ysis
Lee, R . X . 1923. Eureaucracies, Eilreaucrkts and Info~ination Technology.
VP-e3-20. Laxenburg, Anstria: International Institute for Applies Systems Analysis.
March, J.G, and J.P. Olsen. 1979. Ar~bigzi-lty rznd Ciioice in OrganizatCons.
2. Edition. Bergen: I/'niversitetsforle.get.
March, J.G. an$ H.A. Simon. 1958. &ganieatim-ci. Nevi York: Y(i1lep.
McGregor, D. 1980. The - W i m s n S i d e of EnL'~rp7-ie. Kevi York: KcC-:-a?&:- Hill.
Mill.er, G.P.. 1956. The Magical h'urnber Sevez, P l u or Einus %TO: Soms Limits on Our Capacity for Processing Information. P s y c h o l o g i c ~ l Review, 63(2):81-97.
Mintzberg, H. 1973. The Nlzture of J t a n ~ ~ e r - i l T'/~rk. New York: Harper
& Roa.
Peters, T.J. 1930. Xanagement Systems: The Lznguege of Organizations!
Character and Coinpetence. Organizstignal D j n a m i z s . Summer.
Putnam, H. 1970. Is Semantics Possible'? In: H . E . Keefer and X1.K. Kun- itz, eds, L a n g u c g e , B e l i z f , and A!eCaphysics. New York: State University of Nev: TYork Press. Also reprinted in: S.P. Sch.~-artz, ed, N a m i n g , Necess2,ty, a n d N a t u r a l Kinds, 1977. London: Cornell University.
Putnam, H. 1978. Meaning a n d t h e idoral Sciences. Bosion: Routledge 8:
Kegan Paul.
Rescher, N. 1975. A Theory of Possibility. Pittsburgh: University of Pittsburgh Press.
Schwartz, S.P. ed. 1977. N a m i n g , N e c e s s i t y , a n d N a t u r a l K i n d s , London:
Cornell University Press.
Simon, I-I.A. 1955. A Behavioral Model of Rational Choice. The Quarterly J m m a l of Economics, Vol. LXlX.
Sol, H.G. 1983. Processes a n Tools for Decisi.cn S u p p o r t . Proceedings of IF'IP/IIASA Working Conference on Processes and Tools for Decision Support, July 19-21, 1982, Laxenburg, Austria. Amsterdam: North- Holland.
Strawson, P.F. 1959. I n d i v i d u a l s . Garden City, New York: Anchor Books.
Tversky, A. and D. Kahneman. 1974. Judgment under Uncertainty:
Heuristics and Biases. Science, 185(September): 1124-1 131.
van Fraassen, B.C. 1971. Form .EL S~n-iir,~~tins and Logic. New York: Y L C - millac.
von Kright,
G.H.
1068. Ari Essay in DeoEtic Logic and t h e General T'neol-y of Action. k c t r Pfililoscphiz~ F m n i r ~ , Fas:. XCI, Amsiwdem: Korth- Holland.Weber,
K..
1956/ 1578. E c o n ~ m y and Socicfy. Berkeley, CaliforrLa:University of California Press, transiated from nFi-tscha;t u 7 ~ 3 Gesellschaft, Tuebingen: J.C.B. Mo-b, i 956.
Wiener,
N.
1967. Htmcn Use o; h-umcz. B~ingrs: C y b e r n ~ t i z s a n d S o c i ~ t ~ New York: Avcn Books.Wittgenstein, L. 1953/1953. Philoss?lzcZsal J m u e s t i ~ ~ t i n : z s . Trmslzted b y G.E.M. Anscornbe. Third E d i t i ~ n . New York: Xacmilian.
