• Aucun résultat trouvé

Japanese attrition Japan

N/A
N/A
Protected

Academic year: 2022

Partager "Japanese attrition Japan"

Copied!
15
0
0

Texte intégral

(1)

Multiagent Environments

Brahim Chaib-draa

Computer ScienceDepartment, PavillonPouliot,

LavalUniversity, Ste-Foy, PQ,Canada G1K 7P4.

email: [email protected]

Abstract

Analyticaltechniquesaregenerallyinadequatefordealingwithcausal

interrelationshipsamong asetofindividual andsocialconcepts. Inthis

paper,wepresentasoftwaretoolcalledCM-RELVIEWbasedonrelational

algebrafordealingwithsuchcausalinterrelationships. Then weinvesti-

gatetheissueofusingthistoolinmultiagentenvironments,particularlyin

thecaseof:(1)thequalitativedistributeddecisionmakingand,(2)theor-

ganizationofagentsconsideredasawholisticapproach.Foreachofthese

aspects,wefocusonthecomputationalmechanismsdevelopedwithinCM-

RELVIEWtosupportit.

Index Terms Decision support, cognitive maps, knowledge base manage-

ment,tools andsupports,causalmaps,agentandmultiagentsystems.

1 Introduction

Cognitivemapsfollowpersonalconstruct theory,rstputforwardbyKelly[9].

Thistheory providesa basis forrepresentinganindividual's multiple perspec-

tives. Kellysuggeststhatunderstandinghowindividualsorganizetheirenviron-

mentsrequiresthat subjectsthemselvesdenetherelevantdimensions of that

environment. Heproposedasetoftechniques,knowncollectivelyasarepertory

grid, in order to facilitate empirical research guided by the theory. Personal

construct theory has spawned many elds and has been used as a rst step

in generating cognitive maps. Hu [8] has identied vegeneric families of

cognitivemaps:

1. Mapsthatassessattention,associationandimportanceofconcepts: With

thesemaps,themapmakersearchesforfrequentuseofrelatedconceptsas

ThisworkhasbeensupportedbytheNaturalSciencesandEngineeringResearchCouncil

ofCanada,NSERC.

(2)

nization,forexample,andlookfortheassociationoftheseconceptswith

otherstoinfermentalconnectionbetweenimportantstrategicthemes. She

alsomightmakejudgmentsaboutthecomplexityoftheserelationshipsor

dierencesintheuseofconcepts.

2. Maps that show dimension of categories and cognitive taxonomies: The

mapmakerinvestigatesheremorecomplexrelationshipsamongconcepts.

She might dichotomize concepts and construct hierarchical relationships

amongbroadconceptsandmorespecicsubcategories. Mapsofthistype

have been used to dene the competitive environment, and to explore

therangeand natureof choicesperceived by decisionsmakers ina given

setting.

3. Mapsthat show inuence, causalityand systemdynamics(causalmaps):

Thesemapsallowthemapmakertofocusonaction,asforexample,how

therespondentexplainsthecurrentsituationintermsofprevious events,

and what changesshe expects in thefuture. This kindofcognitivemap

iscurrently,hasbeen,andisstill,themostpopular mappingmethod.

4. Maps that show the structure of argument and conclusion: This typeof

map attemptsto showthe logicbehindconclusionsand decisionsto act.

The map maker includes here causal beliefs, but looks more broadly at

the text, as a whole, to show the cumulative impact of varied evidence

andthelinksbetweenlongerchainsofreasoning.

5. Maps that specify frames and perceptual codes: This approach suggests

that cognitionishighlyconditionedbypreviousexperience,andthat ex-

perienceisstoredinmemoryasa setofstructuredexpectations.

In this paper, we extend the causal map by providing a tool called CM-

RELVIEW, which has a precise semantics based on relation algebra [6]. Then

weinvestigatetheissueofusingthistoolinmultiagentenvironments,specially

for(1)thequalitativedistributed decision makingand,(2)theorganizationof

agentsconsideredasa wholisticapproach.

1.1 Causal Maps: Motivations Through an Example

Muchattentionhasbeenrecentlydirected towardtheproblemofstrategicde-

cisionmakingin dynamic and open environments. Theissues ofthis problem

tendtobecomemore complicated,unstructured, andnotalwaysreadilyquan-

tiable. Theyparticularlyinvolveinteractingvariablesthatmakethemdicult

todealwith. Analyticaltechniqueshavebeenusedtohandlewell-denedprob-

lems,but theyare inadequateforthis typeofproblem. A causalmap maybe

employedtocopewithcomplicatedproblemsbecauseitoerstomodelinterre-

lationshipsamongavarietyofconcepts. Causalmaps(CMs)makethefollowing

threeassumptionsaboutcognitionin thecontext ofdecision[8]:

1. Causal associations are a major way in which decision problems can be

describedandunderstood;

(3)

comes;and

3. Choiceamongalternativedecisionactionsinvolvescausalrelations.

Causal maps are usually based on human communications collected by

interviewingor found in documents such as corporate reports or memos. We

arelookingthereforexpressionshavingthegeneraltype:

Entity/phenomenon/conceptAleads-to/causes/inuences/etc.

/Entity/phenomenon/conceptB or,

B iscaused/eected/inuenced/etc./ byA.

Thesearecausal assertions,which aretakento indicatethat theconcerned

subjects: (1) use concepts (A,B) to refer to some phenomena or concepts in

theirdomains,and(2)think(believe,assume,know,argue,etc.) thatthereare

certainrelationships betweentheseconcepts.

Wegenerallyusecausalmapsfordealingwithsuchcause-eectrelationsem-

bedded in deciders' thinking. Thesesmaps are representedas directed graphs

where the basic elementsare simple. The conceptsan individual (a decision-

maker or a group of decision-makers) uses are represented as points and the

causal links between these concepts are represented as arrows between these

points. This representation givesagraph ofpointsandarrows,called a causal

map(CM). Thestrategicalternatives,allofthevariouscausesandeects,goals,

andtheultimateutility 1

ofthedecision-makercanallbeconsideredasconcept

variablesand representedas pointsin the CM. Causalrelationships can take

ondierent values basedon themost basic values + (positive), (negative),

and0 (neutral). Logical combinationsof these three basic values give thefol-

lowing: neutral ornegative ( ), neutral orpositive (), non-neutral (),

ambivalent (a) and, nally, positive, neutral, or negative (i.e.,universal)

(?) [1,5,13].

The real powerof this approach appears when a CM is pictured in graph

form. Itisthenrelativelyeasytoseehowconceptsandcausalrelationshipsare

relatedtoeachotherand tosee theoverall causalrelationshipsof oneconcept

withanother,particularlyifthese conceptsaretheconceptsofseveralagents.

Forinstance,theCM ofFig. 1,takenfrom[11], explainshowtheJapanese

made the decision to attack Pearl Harbor. Indeed, this CM states that re-

maining idle promotes the attrition of Japanese strength while enhancingthe

defensivepreparedness of the United States, bothof which decrease Japanese

prospectsforsuccessinwar. Thus,aCM isasetofconceptsasJapanremains

idle, Japaneseattrition, andsoforth, anda setof signededgesrepresenting

causalrelationslikepromote(s), decrease(s),andso forth.

Notethattheconcepts'domainsarenotnecessarilydenedpreciselybecause

therearenoobviousscalesformeasuring US preparedness, successinwar,,

andsoforth. Nevertheless,itseems easyto catchtheintendedmeaningofthe

1

Utilitymeanstheunspeciedbestinterestsofadecisionmaker.

(4)

Japanese attrition Japan

remains idle

preparedness

Japanese success

in war US

+

+ _

_

Figure1: Anexampleofcausalmap(from[11]).

signedrelationships in thismodel [19 ]. As any causal map, theCM ofFig. 1

canbetransformedin amatrixcalledanadjacencyor valencymatrixwhichis

asquare matrix,withonerowand onecolumn foreach concept. Thus,ifa, b,

canddrepresentsJapanremainsidle, Japaneseattrition,Japanesesuccess

inwar, andUSpreparedness, respectively,weobtainthefollowingadjacency

orvalencymatrix:

0

B

B

@

a b c d

a 0 + 0 +

b 0 0 0

c 0 0 0 0

d 0 0 0

1

C

C

A

Inferencesthat wecan drawfrom a CM arebasedona qualitativereason-

ingsimilartofriend'senemyisenemy, enemy'senemyisfriend,andso forth.

Thus,inthecaseofFig.1,remainingidledecreasestheprospectsforJapanese

successin a war along two causal paths. Noticethat therelationship between

idlenessandwar prospectsisnegativebecausebothpathsagree. Inthese con-

ditions,Japanhasaninterestin startingwarassoonas possible ifshebelieves

thatwar isinevitable.

Thus,causalmapsandthequalitativereasoningthatitsustainsservegener-

allyas themodelinglanguage forproblemresolutionthrough decisionmaking,

particularlyinmultiagentsystemswheredecisionemergesgenerallyfrominter-

relationshipsamongagents'concepts. Suchisthecaseforthepreviousexample

that reects a multiagent system in the sense where Japan and USA are

individualagents.

1.2 Related Work

As previously stated, causal maps appear to be the most widely used cogni-

tive maps in practice. Causal maps have been used for decision analysis in

theeldsofinternational relations[1,5],administrativesciences[15],manage-

mentsciences [7, 17], and distributed group decision support [21]. Inthe lat-

ter context, Zhangand his colleagues provided thenotions of NPN (negative-

positive-neutral) logic, NPN relations, neurons, and neural networks. Zhang

[20 ]extendedhisapproachbyanNPNfuzzysettheoryandanNPNqualitative

(5)

of CM, based on particular techniques for associating intervals with edges of

directedgraphs. Intheirmodel,quantitiescombinebypropagationalongpaths,

butthere is nootherconnectionto theoriginalspirit ofCMs, (i.e.,thecausal

reasoning).

Exceptfor Zhang'swork [21,20], all other approachesto CMs were based

onsimple inference mechanisms about the consequences of a CM. Thus, the

denition of a precise semantic interpretation of qualitative causality has re-

ceivedverylittleattention. Theonlywork thatweareawareofinthiscontext

is Wellman's [19 ] and Nadkarni's [12] approaches. The rst author used an

approach based on graphical dependency models, for probabilistic reasoning,

andsignalgebras, forqualitativereasoning. Thesecond authorused Bayesian

network approachtomakinginferencesincausalmaps. Thus, bothusedprob-

ability theory as is the case usually in Articial Intelligence. However, their

approachesareapplicableonlyintheacycliccase,becausecircularrelationsor

causal loops, common in causal maps, violate the acyclic graphical structure

requiredin aBayesian network.

In this paper, wepresent an implementation of a formal model which has

been implemented in a system used as a computational tool supporting the

relational manipulations. This tool is called CM-RELVIEW and is built over

the RELVIEW software 2

, a freeware package developed by Berghammer and

Schmidt[2].

2 CM-RELVIEW: An Implementation of a Relation

Model of CMs

Withthis tool, alldata are representedas binaryrelations, which the system

visualizesin two dierentways. For homogeneous relations, CM-RELVIEW of-

fersa representation as CMs, includingseveraldierentalgorithmsforpretty-

printing. Asanalternative,anarbitraryrelationmaybedisplayedonthescreen

asaBooleanmatrix. Withmatrixrepresentation,wecanvisuallyeditandalso

discover various structural properties that are not evident from the CM rep-

resentation. CM-RELVIEW system can manage as many graphs and matrices

simultaneously as memory allows and the user may manipulate and analyze

the relations behind these objects by combining them with the operators of

relational algebra. Theelementary operationscan be accessed through simple

mouse-click,buttheycanalsobecombinedintorelationalexpressions,mapping,

andimperativeprograms. CM-RELVIEWallowsalsouserstostorerelationsand

CMs.

CM-RELVIEWoersamenuwindow(aswecanshowinFigure2)thatcanbe

dividedintothreeparts: therstpart,thatistherstthreebuttonrows,deals

withsystemadministrationtaskslike(a)FILES:opensthele-chooserwindow;

2

This software can be obtained by anonymous ftp from http://www.informatik.uni-

kiel.de/progsys/relview.html.

(6)

editor; (d) GRAPH: Pops thewindow of CMs editor; (e) XRV/PROG: display

thedirectorywindowshowingthestateoftheworkspaceandthereasoningon

cognitivemaps; (f)LABEL: Opensthelabeldirectorylistinglabelset whichis

inourcaseC:=fa;+; ;0;; ;;?g.

The buttons in the second part of the menu window cover actions which

are mostlyneeded while working with the systemlike (g) EVAL: Pops upthe

evaluation window for entering a relational term (a relational term can be a

relation,afunction,orarelationalprogram);(h)TESTS:Popsupawindowfor

invokingtests. Withthelatterwindowwecanperformthefollowingactions(1)

TEST-1-R:toexecutevariouskinds oftestsona relation(isitempty,injective,

symmetric? etc.); (2) TEST-2-R: to execute tests on two relations (are they

equal, included? etc.); (3) SUBJECTIVE VIEWS: to do some tests on CMs

in the case of the reasoning on subjective views (comparison, prediction and

explanation) discussed in details in [6]; (4) WHOLISTIC-CM to execute some

strategies of changes on the particular CMs representing an organization of

agentsasdiscussedin Section4.

Inthethirdandlastpartofthemenuwindow,anumberofrelationalopera-

tionsaredirectlyaccessibleviapushbuttons. Thereisalsoadirectorywindow

thatpresentstheactualstateoftheworkspaceto theuser.

FILES INFO QUIT

RELATION GRAPH

LABEL XRV/PROG

DEFINE EVAL ITER

| & - * ^

S/R R\S SYQ

TRANS REFL SYMM

DEF 1st 2nd ORD TESTS

Figure2: ThemenuwindowofCM-RELVIEW.

TherelationeditorcanbeopenedbyclickingontothebuttonRELATIONin

themenu window (seeFigure 2). Selecting a relation in the rstscroll list of

thedirectorywindowloadsthisrelationintotheeditor. Evaluationofrelational

termsperformsa load commandimplicitly, i.e., theresultof theevaluation, a

relation,is automatically displayedin therelation editor window. There isno

explicitsave commandbecauseanymodicationoftherelationshowninthe

relationeditordirectlychangestherelationstoredin theworkspace. Typically,

the window of the relation editor looks as a grid network in which a single

(7)

is represented byone ofthe set C :=fa;+; ;0;; ;;?g. An exampleof a

relationisshowninFigure3.

+

+ +

- -

-

+

- Name : R1 Dim : 8 x 8

Figure3: Therelationeditor.

Ifthemousepointer islocatedonanitemof arelation,themousebuttons

invoke thefollowing dierent actions: (1) theleft mouse button sets theitem

ifit was cleared, or clearsitif itwas set; (2) themiddle mousebutton allows

tochooseone relation (which isused bythe leftmouse toset) ofthe set C:=

fa;+; ;0;; ;;?g;nally(3) therightmousebuttonpopsupamenu that

appearsinFigure 4.

RELATION-MENU

RENAME DELETE NEW

CLEAR RANDOM FILL RELATION -> GRAPH PRINT

LABEL DRAW_MODE

Figure4: Thepop-upmenuoftherelationeditor.

Themostinterestingmenusofthepop-upmenuare: NEW:Itcreateanew

relation,DELETE:Itdeletestherelationdisplayedintherelationeditorwindow

from the workspace (the cognitive map associated with the deleted relation

is also deleted), RELATION ! GRAPH: It creates a CM from homogeneous

relationwiththesamenameastherelation(suchCM isdisplayedinthegraph

(8)

GRAPH-MENU

DELETE NEW

COPY

GRAPH -> RELATION PRINT

MARK NODES MARK EDGES UNMARK GRAPH FIT IN WINDOW TOGGLE GRID GRAPH-DRAWING

Figure5: Thepop-upmenuofthecognitivemapeditor.

Finallythewindowofthe grapheditor (i.e.,CM editor) canbe openedby

pressing the button GRAPH in the menu window. Similarly to relations, all

actions within this menu are selected with the same right mouse button. By

pressingsuchbutton,wereachthegraph-menuthatappearsinFigure5. Within

thismenu,wecanparticularlyinvokethefollowingactions: DELETE:Itdeletes

all nodes of a CM; NEW: It opens a dialog window which allows to enter a

name for a cognitive map; GRAPH !RELATION: It creates a relation from

a cognitivemap, GRAPH-DRAWING:It opens a submenu from which dierent

graph algorithms can be chosen, particularly: LAYER which places the edges

vertically;FORESTwhichdrawsadirectedforestandWHOLISTIC-APPROACH

whichdrawsa particularCM that wewilldetaillaterin section4.

3 CMs as a Tool for Qualitative Distributed De-

cision Making

CMs canalsohelpanagentoragroupofagentsconsideredasawholetomake

a decision. Given a cognitive map with one or more decision variables and a

utilityvariable, which decision should be taken andwhichshould be rejected?

To achieve this, the concerned agent should calculate the total eect of each

decisionontheutilityvariable. Thosedecisionsthathavea+ortotaleect

onutilityshouldbechosen,anddecisionsthathavea or totaleectshould

berejected. Generally,noadvicecanbegivenaboutdecisionswitha,thatisan

ambivalent,total eectonutility,whereasthat a or ? totaleectonutility

should not be rejected because it raises the undetermined decision problem.

To solve such undeterminated decision,wepropose here anoriginal algorithm

whichisbasedontheprincipleofsuperpositionadoptedforCMs. Thisprinciple

stipulates that the result of applying together two concepts C

1 and C

2 is the

sameasapplyingC

1 andC

2

in sequence.

(9)

1. Foranyconcept C thathas anundeterminedresult onthe utility U,calculate all

theindirecteectsbetweenCandU;thenseparatethoseindirecteectsinpositive

andnegativepaths;i.e.,pathswith+ and totalindirecteect respectively;

2. Cutoallthenegativepathsandevaluate theeectofpositivepathsonU,then

noteP

1

thisevaluation;

3. RepeatthepreviousstepfortheeectofnegativepathsonU (withouttakinginto

accountthepositivepaths)andnoteP2 thisevaluation;

4. CompareP

1 andP

2

(a) ifP

1

ismorevaluablethanP

2

thenthesignbetweenCandU is+;

(b) elseif P

1

islessvaluablethanP

2

thenthesignbetweenCandU is ;

(c) elseif P

1

iaasvaluableasP

2

thenthesignbetweenC andU is0.

We will showbelow how thisalgorithm operates with a concrete example.

Before that, we now illustrate the decision-making process in the context of

multiagent environments using CMs. To achieve this, consider for example

the causal map of a professor P

1

(considered as an agent) shown in Figure 6

whosupervisesaresearchgroupcalledG

12

)andwhohastochoosebetweentwo

coursesD

1 andD

2 (D

1 andD

2

aredecisionsvariables). Thequestionnowishow

P

1

canchoosebetweenD

1 andD

2

knowingthefactsreectedbythecausalmap,

showninFigure6. ThiscausalmapincludesthefollowingP

1

beliefs. D

1 favors

thetheoreticalknowledgeofG

12

'sstudents. Greatertheoreticalknowledgegives

agreatermotivationto students. Greatermotivationofstudentsgivesa better

qualityofresearchforgroup G

12

,which givesagreaterutilityofG

12

which, in

turn,hasapositiveresultonutilityofP

1

. Finally,theseconddecisionvariable

D

2

isan easycoursethat decreasestheworkloadofP

1

. Obviously, decreasing

P

1

'sworkloadincreasesherutility.

Inthiscase,howcanP

1

makeherchoicebetweenthetwocoursesD

1 andD

2

?

Noticethatinthecontextofourexample,P

1

shouldreasonaboutanotheragent

which isthe groupG

12

to makeherdecision. Inothercontexts, and forother

decisions, she can also collaborate with her group to develop herdecision. In

thissense,thedecision-makingprocessconsideredhereisamultiagentprocess.

Torunthisprocess,itmightbeusefultoconvertthecausalmapbeinganalyzed

totheform ofa valency matrixV. Withthevalencymatrix, P

1

can calculate

indirectpathsoflength2(i.e.V 2

),3(i.e.V 3

),etc.,andthetotaleectmatrix

V

t

. Infact,V

t tellsP

1

howthedecision variables D

1 and D

2

aect herutility

and G

12

's utility. This gives the following matrix of size 22 (keeping only

therelevantentries)involvingtwodecisionconcepts(DC),D

1 andD

2

,andtwo

utilitiesconsideredasvalueconcepts(VC),namely, UtilitiesofG

12 andP

1 .

(10)

Theoreticalknowledge

ofstudents

Student

motivation Researchquality

P

1

workload

Utility

ofP

1

+

+ +

+

ofG

12

ofG

12

+ D1

D2

Figure6: Anillustrativeexamplefordecision-makingin amultiagentEnviron-

ment.

DCnVC UtilityofG

12

UtilityofP

1

D

1

+ ?

D

2

+

Thus, P

1

perceives (1) decision D

1

as having a positive eect on Utility

ofG

12

andan undeterminedeect onher utility; (2) decision D

2

as having a

negative eecton Utility of G

12

and a positive eect onher utility. In these

conditions,itisimportanttoremovetheundeterminedresultofD

1

decisionon

P

1

utility. Toachievethis,weapply theprevious algorithmasfollows:

1. ToseetheimpactofgivingthecourseD

1

onutilityofG

12

wecutothe

negativepathproduced by Studentmotivation (+)>Workloadof

P

1

( )>UtilityofP

1

. Practically,thismeansthatP

1

evaluatesthe

followinghypotheticsituation: ifthecourseD

1

willbe givenbyanother

colleaguewhatwillbetheimpact(I

1 )ofD

1

onmyutilitywithouttaking

intoaccounttheworkloadinducedbyD

1

?

2. Similarly, wecuto thepositivepathproducedby Studentmotivation

(+)>ResearchqualityofG

12

(+)>UtilityofG

12

(+)>

UtilityofP

1

. Bydoingso,wecanseetheimpact(I

2

)ofgivingthecourse

D

1

ontheworkload(W2) ofP

1

without thepositiveimpactinducedby

the group G

12

. Practically, this means that P

1

evaluates the following

hypotheticsituation: Whatwillbetheimpact(I

2

)onmyutilityifIgive

thecourseD

1

toanothergroupthathasnoconnectionwithme?.

3. Finally,If theimpactI

1

compensatesI

2

then D

1

(0)>utilityof P

1

;

(b) is more valuable thanI

2

then D

1

(+)> utility of P

1

; (c) is less

valuable thanI

2

thenD

1

( )>utilityofP

1

(11)

1 1

eectsonher utility, viahergroup ofresearch, whichare more valuablesthan

whatthiscoursegivesheras workload. Intheseconditions,wehave

DCnVC UtilityofG

12

UtilityofP

1

D

1

+ +

D

2

0 +

Itis clearhere thatdecision D

1

wouldbepreferredondecision D

2

because

thisdecisionhasapositiveimpactonP

1

'sutilityandonG

12

utility. Conversely,

D

2

hasonly limitedimpact becauseit only positively inuences theutility of

P

1 .

How the CM-RELVIEW tool can be used by Decision Makers

(DMs) for their QDM

Decision makers (DMs) can elicit causal knowledge about their decision and

utilityvariablesfrom dierentsources,includingdocuments(suchascorporate

reportsor memos),questionnaires, interviews,grids,and interactionand com-

munication between other agents. After that, they use the relation editor of

CM-RELVIEWforlling matricesrelativeto thiscausalknowledge. Then, they

use the GRAPH button for transforming those matrices, into graphs (causal

maps). Finally,theyanalyzethosecausalmapsusingtheTRANSbutton.

Here,howa decisionmaker(DM)canusethistool. Bypressing thebutton

TRANSin the menu window(Fig. 2), CM-RELVIEW, a decision maker (DM)

cancalculatethetransitiveclosure,(i.e.,thetotaleectthat adecision hason

the utility variable). Inthe case where there is an undeterminedresult, CM-

RELVIEWappliesthealgorithmintroducedpreviouslyandaskstheDMtogive

it some guidance to solve the undetermined result. In particular, the DM is

askedtosupply(1)theimpactofpositiveandnegativepathsand,(2)themost

valuableimpact. Afullyautomatedprocessforsolvingtheundeterminedresult

problemisscheduledintheagendaofourfuturework.

4 CMs as a Tool For Studying Changes in Orga-

nization of Agents

Inmultiagentsystems,thestudyof anorganizationofagentshasgenerallyfo-

cusedonsome structuralmodels such: (1)centralized andhierarchicalorgani-

zations,(2)organizationsas authoritystructure,(3) market-likeorganizations,

(4) organizations as communities with rules of behavior. All these structures

misseddynamicaspectsandinuencesthat existin anorganizationofagents.

Weick[18]suggestedtochangetheprevalentstaticviewofanorganizationof

agentstoa dynamicviewwhichissustainedbychange. Precisely, heproposed

that organization and change were two sides of the same social phenomena.

His reasoning was that an organizationis a process of co-evolution of agents'

(12)

organizationandchangebasedonthegraphsofloopsinevolvingsocialsystems.

Inthelast decade,additionalinvestigationguidedbythisapproach [3 ,4]tried

to articulate how CMs provide a way to identify theloops that produce and

controlanorganization.

Asanexample,considertheorganizationthatbindsresearchers,grantagen-

ciesand qualiedpersonnel inany(scienceandengineering)department. The

causalmap representingthis organization isshown in Figure7. Themeaning

ofthisCM isclearandweshallexplainitnomore.

Personnal

Qualication satisfaction

Retentionof

bestresearchers

Quantityof Research

quality

Resourcesfor

research Salaries

Grantagencies

satisfaction

qualied

personnel

+

+

+

+

+ +

+

+

+

+ +

(2)

(3) (4)

(5)

(6)and(7)

(1)

Researcher

Figure7: Anorganizationofagentsasloops.

Inthiscausalmap,conceptslinktogetherto formloops,someofwhichare

numbered(1)to(7). Loops(1),(4)(7)aredeviation-amplifyingloops. Change

in theorganizationis theresult ofsuch loops,becauseany initial increase(or

decrease)in any conceptloops back to that concept as anadditional increase

(ordecrease)which, inturn,leads tomoreincrease(ordecrease).

Loops (2) and (3) are deviation-countering loops [4]. The stability of the

organizationis the result of such loops. In thecase of loop (2), for instance,

anincrease ofresources for research can leadto an increaseof salaries which,

in turn, reduces the resources allowed to research. If this reduction is not

enoughtocompensatetheinitialincreaseofresources,thena residualincrease

ofsalaries takesplacewhich,in turn, reduces theresources, andso on,until a

balancebetweentheinitialincreaseofresources andsalaries isreached. Thus,

(13)

organization.

Noticethat ina wholistic approachthewhole constraintstheconceptsand

therelationshipsbetweenthem. Withanorganizationofagentsrepresentedasa

wholisticapproach,weobtainadynamicsysteminwhichdeviation-amplifying

loopsareresponsibleforchangeanddeviation-counteringloopsareresponsible

forstabilityoftheorganization. Usingtheseloops,anindividualstrategistcan

direct strategic change in the desired directions. This can be achieved by 1)

choosingandchangingaloopor 2)choosing andchanginga setofloops.

How the CM-RELVIEW Tool can be Used by DMs for the

Reasoning on Organization Changes

Herealso,DMselicitcausalknowledgeabouttheirorganizationsfromdierent

sourcesasreports,memos,questionnaires,interviews,etc.. Afterthat,theyuse

theCM-RELVIEWforconstructingcausalmapsreectingthiscausalknowledge.

Finally,theyusetheCM-RELVIEWtoolforanalyzingthosecausalmaps.

AsstatedinSection2,thesubmenuofthegraphmenucalledWHOLISTIC-

APPROACH allows DMs to draw a wholistic causalmap, whereasthe menu

WHOLISTIC-CM of TEST allows them to test it by choosing and changing a

loop. Obviously,thelooptobechangedshould beaweaklooplooselycoupled

to the system. CM-RELVIEW oers DMs thefollowingactions for changing a

loop(from deviationamplifying to deviationcountering, or vice versa): ADD-

NODE:adding a node;REM-NODE: removinga node; REP-NODE:replacing a

node;CHG-LABEL:changing thelabelofalink.

5 Conclusion and Future Work

We have rstly proposed a tool for qualitative reasoning based on cognitive

mapsrepresentingrelationships betweenagents'beliefs. Thistoolallowsusers

todeterminecertainquantitativeandqualitativefeaturesofanycognitivemap.

Then, we have argued for the use of this tool in the context of multiagent

systems, particularly for the reasoning on interrelationships among a set of

individualandsocialconcepts.

Therearemanydirectionsinwhichtheproposalmadeherecanbeextended.

Thefullpossibilitiesofrelationalgebrahaveyettobeexploited. Another

option is to study fuzzy relations between agents' concepts [20]. Our

approach mightbe extendedin this direction to takeinto accountmany

degreesandvaguedegrees ofinuencebetweenagentssuchas none,very

little,sometimes,a lot,usually,more orless,andsoforth[10,14].

Applications such as the following ones must be investigated in greater

depth:

1) negotiation and mediationbetweenagents in the caseof reasoning

aboutsubjectiveviews;

(14)

causalmaps;

3) reasoningaboutthewholisticapproach; and

4) reasoningonsociallaws,particularlyforqualitativedecisionmaking.

References

[1] Axelrod, R. Ed. Structure of Decision: The Cognitive Maps of Political

Elites.PrincetonUniversityPress,1976.

[2] Berghammer,R.andSchmidt,G.RELVIEWAcomputersystemforthe

manipulation ofrelations.InAMAST93, 1993,pp.405406.

[3] Bougon, M. G. Uncovering Cognitive Maps: The Self-Q Technique, Priv.

PrintHandb.,Penn.StateUniv.,1986.

[4] Bougon,M.G.andKomocar,J.M.Directing strategicchange: adynamic

wholistic approach, in Mapping Strategic Thought, A. S. Hu, Ed, Wiley

andSons, 1990,pp.135163.

[5] Buede,D.M.andFerrell,D.Convergencein ProblemSolving: APrelude

to Quantitative Analysis, IEEE Trans. Syst., Man, Cybern., vol. 23, pp.

746765, 1993.

[6] Chaib-draa, B. andDesharnais, J. A Relationalmodel ofcognitivemaps.

Inter.Jour.of Human-ComputerStudies,49,1998,pp.181-200.

[7] Eden, C. J.and Sims, D.Thinking in organizations, Macmillan, London,

1979.

[8] Hu, A.S.Mapping StrategicThought,Wiley&Sons,Eng.,1990.

[9] Kelly,G.A.ThePsychologyof PersonalConstructs,New: Norton,1955.

[10] Kosko,B.Neural NetworksandFuzzySystems,PrenticeHall. 1992.

[11] Levi, A. and Tetlock, P. E. A cognitiveanalysis of Japan's 1941 decision

forwar.Journ.of ConictResolution,24,1980,pp.195-211.

[12] Nadkarni,S. andShenoy, P. P.A Bayesian Network Approach toMaking

InferencesinCognitiveMaps.Eur.Jour.OfOperationalResearch,vol.128,

no.3,2001.

[13] Nakumara, K.Iwai,S. andSawaragi,T.Decisionsupport usingcausation

knowledgebase,IEEE Trans.Syst.,Man, Cybern.vol.SMC-12,1982,pp.

765777.

[14] Park,K.S.andKim,S.H.Fuzzycognitivemapsconsideringtimerelation-

ships,Int.J. Human-ComputerStudies,42,1995,pp.157168.

(15)

Admin.Sci. Qua.,vol.25,pp.5771, 1980.

[16] Schmidt,G.andStröhlein,T.RelationsandGraphs,MonographsonThe-

oreticalComputerScience, Springer-Verlag,Berlin,1993.

[17] Smithin, T.andSims, D.UbiCaritas?Modelingbeliefsaboutcharities,

Eur. Jour.Opl Res.,vol.10,pp.273243,1982.

[18] Weick,K.E.The socialPsychologyof Organizing,Reading, MA:Addison

Wesly,1969.

[19] Wellman,M.P.Inference incognitivemaps,Mathematicsand Computers

in Simulation,36,1994,pp.112.

[20] Zhang,W.R.NPNFuzzySetsandNPNQualitativeAlgebra: AComputa-

tionalFrameworkforBipolarCognitiveModelingandMultiagentAnalysis.

IEEE Trans.Syst.,Man, Cybern.,vol.26(4),pp.561574,1996.

[21] Zhang,W.R.Chen,S.S.andKing.R.S.ACognitiveMapbasedApproach

totheCoordinationofDistributedCooperativeagents.IEEETrans.Syst.,

Man, Cybern.,vol.22(1),pp.103114, 1992.

Références

Documents relatifs