Compositional Design of a Generic Design Agent1

Compositional Design of a Generic Design Agent

Frances M.T. Brazier, Catholijn M. Jonker, Jan Treur, and Niek J.E. Wijngaards*

Vrije Universiteit Amsterdam, Department of Mathematics and Computer ScienceArtificial Intelligence Group

De Boelelaan 1081a, 1081 HV Amsterdam, The NetherlandsEmail: {frances, jonker, treur, niek}@cs.vu.nl

URL: http://www.cs.vu.nl/~(frances, jonker, treur, niek}

AbstractThis paper presents a generic architecture for a designagent. The design agent is based on an existing genericagent model, and includes a refinement of a generic modelfor design, in which strategic reasoning and dynamicmanagement of requirements are explicitly modelled. Thegeneric architecture has been designed using thecompositional development method DESmE, and has beenused to develop a prototype design agent for automatedagent design.

1 Introduction

Design is a task often performed by one or more specialisedagents. Other agents interact with such ’design agents’ by,for example, providing qualified requirements, initial(partial) design object descriptions, and design processobjectives. Specialised agents ate often encountered inhuman society: e.g.. architects are specialised agents: theirarea of expertise is the design of buildings. A design agentgenerates a design object description on the basis of theinformation received from other agents, and makes resultsof the design process available to other agents.

In this paper a generic architecture is introduced for adesign agent, which was modelled using the compositionaldevelopment method for multi-agent systems DESIRE(Brazier, Dunin-Kepliez, Jennings and Treur, 1997). Section 2 this compositional development method isbriefly introduced. The design agent is based on an existinggeneric agent model (Section 3), and includes a refinementof a generic model for design (Brazier, Langen, Ruttkay andTreur, 1994), in which strategic reasoning and dynamicmanagement of requirements are explicitly modelled(Sections 4 and 5). Knowledge of different types can included, for example, knowledge that can be u.~d to derivewhether a given design object description satisfies givenproperties (e.g., requirements), and knowledge that can used to derive how to refine requirements into more specificrequirements. In Section 6 the application of the genericdesign agent to compositional system design is described.

* Currently at: Software Engineering Research Network,Department of Computer Science, University of Calgary, Canada

In Section 7 the results are discussed and a perspective issketched of the use of the design agent in ElectronicCommerce applications.

2 Compositional Design of Agents

The design agent described in this paper has been developedusing the compositional development method DESIRE formulti-agent systems (framework for Design andSpecification of Interacting Reasoning components; of.(Brazier et al., 1997). The development of a multi-agentsystem is supported by graphical design tools within theDESIRE software environment. Translation to an operationalsystem is straightforward; the software environmentincludes implementation generators with which formalspecifications can be translated into executable code of aprototype system. In DESIRE, a design consists ofknowledge of the following three types: processcomposition, knowledge composition, and the relationbetween process composition and knowledge composition.These three types of knowledge are discussed in more detailbelow.

2.1 Process Composition

Process composition identifies the relevant processes atdifferent levels of (process) abstraction, and describes how process can be defined in terms of (is composed of’) lowerlevel processes.

Identification of Processes at Different Levelsof Abstraction. Processes can be described at differentlevels of abstraction; for example, the process of the multi-agent system as a whole, processes defined by individualagents and the external world, and processes defined by task-related components of individual agents. The identifiedprocesses are modelled as components. For each process theinput and output information types are modelled. Theidentified levels of process abstraction are modelled asabstractiow’specialisation relations between components:components may be composed of other components or theymay be primitive. Primitive components may be eitherreasoning components (i.e., based on a knowledge base),

CopyriRht ~ 1998, Ameri,.:an Association for Artificial Intelligence.

30 AIMW-98 ^,, ri~lts reserved.

From: Proceedings of the Artificial Intelligence and Manufacturing Workshop. Copyright © 1998, AAAI (www.aaai.org). All rights reserved.

or, components capable of performing tasks such ascalculation, information retrieval, optimisation. Theselevels of process abstraction provide process hiding at eachlevel.

Composition of Processes. The way in whichprocesses at one level of abstraction are composed ofprocesses at the adjacent lower abstraction level is calledcomposition. This composition of processes is described bya specification of the possibilities for information exchangebetween processes (static view on the composition), and specification of task control knowledge used to controlprocesses and information exchange (dynamic view on thecomposition).

2.2 Knowledge Composition

Knowledge composition identifies the knowledge structuresat different levels of (knowledge) abstraction, and describeshow a knowledge structure can be defined in terms of lowerlevel knowledge structures. The knowledge abstractionlevels may correspond to the process abstraction levels, butthis is often not the case.

Identification of knowledge structures atdifferent abstraction levels. The two main structuresused as building blocks to model knowledge are:information types and knowledge bases. Knowledgestructures can be identified and described at different levelsof abstraction. At higher levels details can be hidden. Aninformation type defines an ontology (lexicon, vocabulary)to describe objects or terms, their sorts, and the relations orfunctions that can be defined on these objects. Informationtypes can logically be represented in order-sorted predicatelogic. A knowledge base defines a part of the knowledgethat is used in one or more of the processes. Knowledge isrepresented by formulae in order-sorted predicate logic,which can be normalised by a standard transformation intorules.

Composition of Knowledge Structures.Information types can be composed of more specificinformation types, following the principle ofcompositionality discussed above. Similarly, .knowledgebases can be composed of more specific knowledge bases.The compositional structure is based on the different levelsof knowledge abstraction distinguished, and results ininformation and knowledge hiding.

2.3 Relation between Process Composition andKnowledge Composition

Each process in a process composition uses knowledgestructures. Which knowledge structures are used for whichprocesses is defined by the relation between processcomposition and knowledge composition.

3 A Generic Agent ModelAgents are often designed to perform their own specifictasks, for example the design of an artefact. In addition, anumber of generic agent tasks can be identified. Thissection describes a generic agent model in which suchgeneric agent tasks are modelled. This model abstracts fromthe specific domain of application and can be (re)used for large variety of agents. The model is based on the abilitiesassociated with the notion of weak agency OVooldridge andJennings, 1995).

Instead of designing each and every new agentindividually from scratch, a generic agent model can be usedto structure the design process: the acquisition of a specificagent model is based on the generic structures in the model.

r’" f .................................................. .~ "’~. .................. Aoent task control /

’"~.’: .............. :~ ........... World Info tO OpC --

mi Own .m aOont info to opc[ Ill Proams :ll ........... I; -I 1 :1-I i: ""lw"t..z~ iL-~- Ovm pmcese I

pro~m~irffo to win1"~ proceu Info to............................ Ir~o to maJ

........ ~ ~--~,i~o-zo ~m m~ _~ZLtO_m_amm~d’: . f .............. ~ commurlcated if

I I Ageot ;I agent Ib~l--. Maintenance ~. __

"1 lqinf : :-’-’- of Aoeflt --, -,:-::"111: Management U " 1’ i l "BDP*~%- - J’ "~- .......... J /

communicated world Info //¯ ~ agent Info to aim

--; I ~;.d~-~ - -~:~ 1 -:I _.[__1 o~.m..~l.fo ......... -I I --II~iL ..... --,,rt/- /m-~ ...... - "°~!11 -- --, I -

’ towlm i’ imo ’/ . ~.. wo~ - 11~.4M.,.~. !..J I

l-- / --I I’ [n~orac~ion I : [[¢~[ I o~World I /._ I¯ -- I ’~ --’ "/’---i~ ’ Management , ~:~:::.:.J I r Infomlation ,!11~ ]~ ,I ~.: .... ~_~ t, I

I’ woed~lowim "J wondinfo" ........... ~h

~IHn~ ~ ~m ----observed coromunlcsdod /

""}action and ,nfo to ost ,hie to ost ..i~- ~ ~[,~.j.]=bservalionnfo from mt -- Task

i communication Info from Mr[’,,." .............. ....-~ "..::’~T.:’-TT.T.:" ..... "--" ": J

Figure 1. A Generic Agent Model.

The characteristics of weak agency provide a means toreflect on the tasks an agent needs to be able to perform.Pro-activeness and autonomy are related to an agent’sability to reason about its own processes, goals and plansand to control these processes (own process control).Reactivity and social ability are related to the ability to beable to communicate with other agents (agent interactionmanagement) and to interact with the external world (worldinteraction management). The ability to communicate withother agents and to interact with the external world oftenrelies on the information an agent has of the world(maintenance of world information) and other agents(maintenance of agent information). The generic agent

Brazier 31


model also includes an empty generic component to modelthe agent specific task. The tasks related to the genericabilities and agent specific tasks may be modelled bycomponents within an agent as depicted in Figure 1. Inaddition to the sub-components, the model includesinformation links that specify which information isexchanged between components; these information links arenamed.

The exchange of information within the generic agentmodel can be described as follows. Observation results aretransferred through the information link observation resultsto wim from the agent’s input interface to the componentwodd interaction management. In addition, the componentworld interaction management receives belief informationfrom the component maintenance of world informationthrough the information link world info to wim, and theagent’s characteristics from the component own processcontrol through the link own process info to wim. Theselected actions and observations (if any) are Iransferred the output interface of the agent through the informationlink observations and actions.

The component maintenance of world informationreceives meta-information on observed world informationfrom the component world interaction management, throughthe information link observed world info and recta-information on communicated world information (throughthe link communicated wodd info) from the componentagent interaction management. Epistemic informationfrom maintenance of world information, epistemic worldinfo, is Iransferred to input belief info on world of thecomponents world interaction management, agentinteraction management and own process control, throughthe information links world info to wire, world info to aim andworld info to opc.

Comparably the component maintenance of agentinformation receives recta-information on communicatedinformation from the component agent interactionmanagement, through the information link communicatedagent info and meta-information on observed agentinformation (through the link observed agent info) fromthe component world interaction management. Epistemicinformation, epistemic agent info, is output of thecomponent maintenance of agent information, becomesinput belief info on agents of the components worldinteraction management, agent interaction managementand own process control, through the information linksagent info to wim, agent info to aim and agent info to opc.

4 A Generic Model of Design

The generic model of a design agent is based on both thegenetic agent model discussed in Section 3, and a genericmodel of the design task, used to model the agent specific

task component. In this section the structure of this agentspecific task component for a design agent is described.

A generic model of design, in which reasoning aboutrequirements and their qualifications, reasoning aboutdesign object descriptions and reasoning about the designprocess are distinguished, has been introduced in (Brazier,Langen, Ruttkay and Treur, 1994). This model is based ona logical analysis of design processes (Brazier, Langen andTreur, 1996) and on analyses of applications, includingelevator configuration (Brazier, Langen, Treur, Wijngaardsand Willems, 1996) and design of environmental measures(Brazier, Treur and Wijngaards, 1996). The model not onlyprovides an abstract description of a design processcomparable to a design model, e.g., (Coyne, Rosenman,Radford, Balachandran and Gero, 1990; Smithers, 1994),but also a generic structure which can be refined for specificdesign tasks in different domains of application.Refinement of the generic task model of design, byspecialisation and instantiation, involves the specificationof knowledge about applicable requirements and theirqualifications, about the design object domain, and aboutdesign strategies.

An initial design problem statement is expressed as a setof initial requirements and requirement qualifications.Requirements impose conditions and restrictions on thestructure, functionality and behaviour of the design objectfor which a structural description is to be generated duringdesign. Qualifications of requirements are qualitativeexpressions of the extent to which (individual or groups of)requirements are considered hard or preferred, either inisolation or in relation to other (individual or groups of)requirements. At any one point in time during design, thedesign process focuses on a specific set of requirements.This set of requirements plays a central role; the designprocess is (temporarily) committed to the currentrequirement qualification set: the aim of generating a designobject description is to satisfy these requirements.

During design the considered sets of requirements maychange as may the design object descriptions: they evolveduring design. The strategy employed for the co-ordinationof requirement qualification set manipulation and designobject description manipulation may also change during thecourse of a single design process. Modifications to therequirement qualification set, the design object descriptionand the design strategy, may be the result of straightforwardimplications drawn from knowledge available to a designsupport system. Modifications may also be the result ofspecific knowledge on appropriate default assumptions (,seealso (Smith and Boulanger,1994)), or the result interaction with an outside party (e.g., a client or designer).

32 AIMW-98


Design task control

Deslgn

Co-ordination . .............I .................. __ ~ ""L ..../ overall design

---’- .... . i -J ’ :-1 i strategy to DOOM

! overall design strategy to RQSM .--t ..~:.. ..... "Y’. ̄ ..H- ...... DeiSM process..................................... I i evaluation report~-" --"- .... ~" DODM

I .... RQSM process mevaluation report

I I.... | I._.~-- ROSm

l l l ~mm

’: $ Mani’-ulation intermediate RQS information-¢ ...............

) .............................RQ_S. information I

intermediate DODM results

DODI resuJt~. ~

""_ t

Manipulation ’:

~.,,.

....... i -=-.0-- !ntermediate~Uon JIi =..................... i

L ..................................................... .j’

~’-DODM DOD history query’ " ..................................... curmntl~D replacement requestI ’,,

DOD modification results to hisl......... . ......................... :::::: .......o~50 m~r~i~,~,,~_~!=o.~.q~,~.

RQS information query................................. DODl~history query

.................................... DOD-modification progress............ . ...... "--:..-~- .............................

I ............-~""""’"" .......... - ...... overall design strafi)~ l...... I i ~ DODM ~ ~ DOD

~i ) history DODM.h!s!oW.cluery msu~ llLb modificationoverall des!~ln ’: ). ~.. maintenancestrategy tO ~i.s~ i" ~ .... DODM history information. ~p.

’ DOD modification results histor- ........~..,; rf msul~dlcinfo rmatien- ~1¢ ...........

information.l__.-~’~~ormation & ... .............................

query resul..~

i I):::::::::::I -I- ~o h~to~_?uo~ ~.~.= ~.................Doe basis 4 ......

DOD history information ~Jinformation "~,~ --r]2::]~ " ........... ~ .... I- i_ ! ...... . " .......i l-

new ................. current DOD focus i Icurrent ...... l~5modificati~ns .................. i !

; i........................... . ........... ".’.’.’.’.’." ". ......DOD current DOD to DOD refinement focus

- rrent ~- be stored ,,;,,- -cur ~ current DOD to i.-~ deductive_..~:: . .DOD j be analysed / OOD -7, .::.:::::::IIF mmntenanc6._...:l --. ~L__. refinement

JI . I .....~---~-de$ign-It.~ ~ object information ~-.-~

" I design object information! ...................................

..................................................................... j’

Figure 2. A generic model of design: two process abstraction levels.

DODMprocessevaluation

._ status

-. OOD~modifica .ion

result~ ....

Brazier 33


Figure 2 shows two levels of composition of the genericmodel tor design. Three processes are shown at the toplevel, together with the information exchange. Fourprocesses and information exchange are shown at the secondlevel for DODM.

The four processes (see Figure 2) related to the processrequirement qualification set manipulation ( RQSM) are:

¯ RQS modification: the current requirementqualification set is analysed, proposals formodification are generated, compared and the mostpromising (according to some measure) selected,

¯ deductive RQS refinement: the current requirementqualification set is deductively refined by means ofthe theory of requirement qualification sets,

¯ current RQS maintenance: the current requirementqualification set is stored and maintained,

¯ RQSM history maintenance: the history ofrequirement qualification sets modification is storedand maintained.

The four processes related to the process of" manipulation ofdesign object descriptions (DODM ) are:

¯ DOD modification: the current design objectdescription is analysed in relation to the currentrequirement set, proposals for modification arcgenerated, compared and the most promising(according to some measure) selected,

¯ deductive DOD refinement: the current design objectdescription is deductively refined by means of thetheor7 of design object descriptions,

¯ current DOD maintenance: the current design objectdescription is stored and maintained,

DODM history maintenance: the history of designobject descriptions modification is stored andmaintained.

The process design process co-ordination is composed in asimilar manner.

5 Specialisation of the Design AgentModel

The generic model for design described in the previouschapter can be refined by specialisation and instantiation.This section addresses a specialisation of the generic modelof design which has been applied in the domain of re-designof compositional (knowledge-based or multi-agent)systems. For reasons of space limitation only thespecialisation of the process of requirement qualification setmodification and the process of design object descriptionmodification are presented.

5.1 Specialisation of requirement qualification setmodification

The process R¢~ modification determines modifications toa requirement qualification set (RQS). To this purpose number of sub-processes are distinguished as shown inFigure 3. The process RQS modification process co-ordination is responsible for the co-ordination of the entireprocess within RQSM: this process determines whether,when and by which means a specific RQS is to be modified.

The global phases within RQS modification resemble aprocess control model (e.g., controlling a chemicalprocess). A process control task usually relies on a feedbackloop within which sub-tasks such as analysis, planning,and execution are distinguished.

RQS modificationhlst0clCal RQS modification state results

ioverate (lesion so’stagy

I.~J~ mSlOry rosults.r- ....

.i

; I

~.---

ImCU~¢l RQS contents tOa~on\.

¯ ...~HCJ~ mo0mcnon proceos co-or0matton results

....... ",, historical RQ8 modification stateRQS ~ information

_.~.~-- modification 1

~’:j ,uamto~re,.=an~rn,,on

process I =~..~ .......[ coordination L;-7-~-I~ modlflcalon result.¯ ~. ...... J ~:~UO0~ I

va/i~bb°t~ . :c°a~_~nnX3n~ugts stmte~yUt~ mo~lnr’ca s~l~e~y t°~o"~o~usn=~,...o. . ........ ~ f

- / focus mulls to

/ , ¢o-ocdnn~l/" ~t,RQS (" " .... ,,RQS [ , mes=,ent RQS! validaUon i *-LtoJ -°u ̄ modification, ,, .; .~ . .- OCUS = ̄ ¯’, ....

_ ~... ..: .., identification I ! "=,.-=~=.~.r ............. ~ ....... _ .¯ "

velidatioc0l [~ .... ,’~"

..-~,,. HUU moamc~on vocUito modlficiafon

/ ¯ RQScurrent ROSl RC~ I --modification

to rnoc.caUo= -- g determination ! sekactodROSmocliflcdon

RQS refinement .g~_s ~ ......... P"~- . ....... =~:~ ......... --

Figure 3. Specialisation of RQS modification.

34 AIMW-98


C ~b--modmca#on ....."~ UCK/iTli0OIIC6g0nN,om,, coo mom~.,.te.,.....,~k ~ ..~_ .=~o,,~,... ~,~P="

i ovul d+mn mmw ~] ! nnn ’ hlModcal DOD nmlIIcagon .Ioi .....P.r" _v_ I .. i~eIdln~emI1onn Lm.~ ,’xm mommm ~ ~um,~ ... modificationI-~ ~ mommm..-Jm

! uuv n,m,yr,~ ...... ~ coordination ....-- -~ i"-

: ----I,eJU rlietoly Inlomldon: ........

~---~. 1 rolalodrolIIo¢

:,

~ , )-;, l-nJ~’p~l~....~

VOIlemnon i I moclllCllon rUullS ~ mom¢~od / ~ moomcalmm

,,, -:.-.---- - .L_ o r ...... -, ......... L-, , y....... veMddon J’" "1 validation .... : _ E: modification...-~." P"

I ~: : ’-’ i focus.......... l ’. "’~-L ~ t-I [ ~-~:AI ,denUficationi..!-~ I |~

i’ curt’ant DOD ". ......... --’~’.-- ..........---4-

~"’ "*/ ~IP

....................... -~@T-~i~r~ems....... contonte to ...................... -..., . ..,..,-.,+, I,, ...... , ’!¯ " " " f--’" -- ..... ~ UIt~J ~I TOCll .....l...... ...,I..ooo1:~0 F PV’~PI ’~ Iomo¢IIkIlll~n [__IL~ IcontmU to Iocul ’ ~mod / m" "+" i-- [¯ "’J t°m°dmcammI L~’" modificationI.--._ .................... :

! ..--.--:::I¢ determ nationI ’, ~ uuumoommon , ~ ,"~--I curroMDODcontolito I~’~~ ~-’ . r ~IP I~ t mo+m+.. + ~_ _.~ DOD ~gom j "

’,,.. .......................... " --- ":- .................................. --TI .....

Figure 4. Spccialisation of DOD modification.

Similarly, within RQS modification analysis is performedby RQS validation, planning is performed by RQSmodification focus identification and RQS modificationdetermination, and execution is Performed by effectuatingmodifications to a RQS, resulting in a new RQS in currentRQS maintenance.

5.2 Specialisation of design object descriptionmodification

The process DOD modification determines modifications toa design object description (IX)D) in such a fashion that DOD is constructed which adheres to the designrequirements given to DODM. To this purpose a number ofsub-processes are distinguished as shown in Figure 4. Theprocess DOD modification process co-ordination isresponsible for the co-ordination of the entire processwithin DODM: this process determines whether, when andby which means a particular DOD is to be modified.

The global phases within DOD modification alsoresemble a process control model: analysis, planning,execution. Similarly, within DOD modification analysis isperformed by DOD validation (including . assessingrequirements in the current IX)D), planning is performed DOD modification focus identification and DOD modificationdetermination, and execution is performed by effectuatingmodifications to a IX)D, resulting in a new DOD in currentDOD maintenance.

6 Application: a Design Agent for AgentDesign

The generic model of a design agent introduced can, in

principle, be used for any domain of application. To obtaina proof of concept it was applied to a specific domainwithin compositional system design: the domain ofcompositional agent design. For this domain a prototypeapplication has been designed and implemented. For thisprototype system a formal ontology of requirements onagents has been developed. Moreover, knowledge has beenidentified that can be used to reason about theserequirements, to derive more specific requirements byrefining the original requirements. These more specificrequirements play a crucial role in the design process: theyguide the direction in which solutions are sought.

Requirements are formulated in terms of abilities andproperties of agents and the external world. Abilities andproperties can be assigned to

¯ individual agents,

¯ the external world,¯ an individual agent in relation to the agents and the

world with which it interacts,¯ the world in relation to the agents with which it

interacts, and¯ a multi-agent system as a whole.

Abilities of agents such as co-operation, bi-directionalcommunication, and world interaction are often needed foragents to jointly be able to perform a certain task. InFigure 5 the ability of bi-directional communication and itsrefinements are depicted. For a description of other agentabilities see Brazier, Jonker, Treur and Wijngaards (1998).

Brazier 35


unidirectional realisations

~communication from otherd%.... ; ...... j

unidirectional realisa~’~ ./communication to others "-.. ) .Y... .

......~’~- " :

¯ executing bi-directional ~.~..~.’~ "communication -"

r

reasoning about unidirectionalcommunication from others

reasoning about unidirectionalcommunication to others

executing unidirectional~mmunication from others

executing unidirectionalcommunication to others

Figure 5, Refinements of the ability of hi-directional communication.

The ability of bi-directional communication can be refined,both with respect to its specialisation (refinement of theability into more specific abilities) and with respect to itsrealisation (refinement of the ability into more fine-grainedabilities related to reasoning about the ability, and morefine-grained abilities abilities related to the effectuation ofthe ability).

Figure 5 shows the refinement relationships for theability of bi-directional communication. The more specificabilities related to bi-directionai communication arc theability to communicate to others (unidirectionalcommunication to others) and the ability to receivecommunication from others (unidirectional communicationfrom others). The abilities related to the realisation of the

ability of bi-directional communication are the ability toreason about bi-directional communication, and the abilityto execute bi-directional communication.

These more specific abilities are further refined, andrelated to the ability to reason about unidirectionalcommunication from others, the ability to reason aboutunidirectional communication to other, the ability toexecute unidirectional communication from others, and theability to execute unidirectional communication to others.

Knowledge on refinements of the ability of bi-directionalcommunication can be formally represented as shownbelow. Meta-reasoning is employed to decide whichrefinement alternative should be employed for whichability.

Example

Ifend

and

and

then

and

end

Representation of requirements refinement knowledge

is_qualilied_requirsment_selected_as_focus( QR: qualilied_requirement_name holds( is_qualified_requirement( QR: qualilied_requiremsnt_name,

Q: requirement_quali6ca~n,R: rsquirsment_name )

holds( refers_to_requirement( R: requirement_name,has_property( A: agent .name,

is_ capable._ of_ bidirectional _communica~onwith( ~.: a~nt_name ))

poe)rsflnement..alterna’dve( specialisations

addilon_to currsnt_RQS(is_qualilied_requiremont(

eddltion_to..curront_RQS(refers_D_requirement(

adddion_to_cu rront_RQS(is_qualmed_rsquirement(

new_name( QR: quali§ed_requirement_name, =a" ),Q: requirement..qualificaSon,now_name( R: requirement_name, "a" ) )

new_name( R: requirement_name, "a" ),has_property( A: agent_name,

is_capable._of_unidirectional_communication_from( A2: agent_name ) )

new_name( QR : qualifiedrequirement_name, "o" ),Q: requirement_quallfica’don,new_name( R: requirement_name, "o" ) )

Con~hued on ~e next page.

36 AIMW-98


and

and

and

addition_to_current_RQS(refers_to_requirement(

addrdon_to_current_RQS(Is_qualified_requirement(

addrdon_to_current_RQS(refers_to_requirement(

new_name( R: requirement_name, "b" ),has_property(A: agent_name,

is_capable_of_unidlreclkxlaUcommunication_to( A2: agent_name ) )

new_name( QR: qualified_requirement_name, %" ),Q: requirement__qualificalion,new_name( R: requirement_name, =d’ ) )

new_name( R: requirement_name, "c" ),has..property( A: agent_name,is_capable_of_comblnlng_unidireclional_communicalion_frorn_end_to( A2: agent_name ) )

Top-level requirements are refined into more specificrequirements during a design process. The result is theconstruction of a specific hierarchy of requirements, whichadheres to the requirements ontology and refinementknowledge. Figure 6 shows an example of (part of) such requirements refinement hierarchy. The current prototypedesign agent makes extensive use of the requirementsontology, generic models and design object buildingblocks. The design process is fairly linear, in the sense thatfew options are generated and selected. The most i~inedrequirements are almost directly operationaiisable bybuilding blocks for design object descriptions. A specific

design requirement, currently in focus in DOE) modification,is broken up (i.e., refined) into smaller properties:assessment points. These assessment points can be testedfor, and when not yet reaiised, building blocks related to anassessment point can be ~ to the current design objectdescription.

The generation of options for sets of qualifiedrequirements and design object descriptions involvingexplicit strategic knowledge can be incorporated in thedesign model, as described by (Brazier, Langen, and Treur,1998).

I has_property( agent_D, is..capable, of_active_observation_in( wodd_W ) . has_property( agent_D, is_capeble_of_procossing_observation results_from( world_W )

I has_property( agent_D, is_capeble_of_reasoning_about_procassing_observation_rosults_from( world_W ) has_property( agent_D, is_capable_of_executing_procossing_observation_rasults._from( wodd_W )

[ has_property( agent_D, s_capab e_of_combining_reason ng_about_and_executing_processing_observation_results( world_W ) has_property( agent_D, obsentation_initiation_in( woddW )

has_property( agent_D, is_capable_ol_recsoning_about_observation_initiation_in( wodd_W ) has_property( agent_D, is_capable_of_executing_observation_initiation_in( world_W )

: has_property( agent_D, is-capab~e-~-c~mbining-mas~ning-ab~ut-and-executing--~bsenIati~n-initiati~n-in( wodd_W ) has_property( agent_D, is_capeble_of_combining_processing_observation_results_and_ observation_initiation_in( world_W )

Figure 6. Requirement refinement hierarchy constructed by the prototype design agent.

The implication of designing (parts of) a multi-agentsystem, is that a multi-agent system is the object ofdesign, and as such should be formally represented in adesign object description. In this paper the design objectdescription is assumed to be a compositional objectdescription. The assumption underlying this decision is thata compositional structure facilitates the process of (re-)design. The compositional formal specification languageunderlying DESIRE forms an adequate basis for such adesign object description representation.

7 Discussion

To design an agent capable of designing, insight is requiredin the agent model and the design process model to be used.

The architecture of the design agent is based on an existinggeneric agent model, and includes a refinement of a genericmodel of design. It combines results from the area ofMulti-Agent Systems and the area of AI and Design.

The generic agent model has been developed on the basisof experience with agent models of different kinds; forexample, models for information gathering agents, co-operative agents for project co-ordination, BDI-agents,negotiating agents, broker agents, and agents simulatinganimal behaviour. The generic design model was developedand evaluated on the basis of experience with designapplications in a number of domains; for example, designof sets of measures for environmental policy, aircraftdesign, and elevator design.

Brazier 37


The generic design agent, which combines the genericagent model and the generic design model, was successfullyapplied to compositional agent design. The design agentmodel for this application includes formalisations of agentdesign descriptions and requirements on agents, andtbrmalisations of agent design knowledge. After this proofof concept, the approach introduced will be applied in thedomain of Electronic Commerce.

Electronic Commerce necessarily involves interactionbetween human users in different types of organisations,and very dynamic automated environments, in which theparties involved are not known beforehand, and oftenchange. In such environments human users can besupported by Personal Assistant Agents, which in turnmake use of existing broker agents and other task-specificagents. Co-operation between these (human and computer)agents is to the advantage of all. To cope with the dynamiccharacter of the environment, frequently new agents need tobe created, or existing agents need to be modified, forspecific purposes. Such frequent modification of anenvironment necessitates almost continuous maintenance.

Recently a few applications of broker agents have beenaddressed for this area; see, for example (Chavez and Maes,1996; Chavez, Dreilinger, Guttman and Maes, 1997;Kuokka and Harada, 1995; Martin, Moran, Oohama,Cheyer, 1997; Sandholm and Lesser, 1995; Tsvetovatyyand Gini, 1996). However, these applications have beenimplemented without an explicit design at a conceptuallevel, and without taking into account the dynamicrequirements imposed by the domain of application and themaintenance problem implied by this dynamic character.

On the basis of the approach introduced here, a genericmulti-agent Electronic Commerce environment will bedeveloped in which a broker agent can dynamicallyreconflgure (parts of) the multi-agent system by adding modifying Personal Assistant agents, broker agents andadditional agents. More specifically, the aim is to develop amulti-agent broker architecture with a number of co-operative broker agents, Personal Assistant agents, and taskspecific agents. Each broker agent can dynamicallyconfigure and implement new agents or modify existingagents as part of the multi-agent system as follows:

¯ if new users (clients) subscribe to a broker agent,Personal Assistant agents tuned to the requirementsimposed by this user, may be created, or existingPersonal Assistant agents may be modified, due tochanged requirements.

¯ if required in view of the load of an existing brokeragent, new broker agents can be ~ to distributethe load (and avoid overload of the existing brokeragent), or existing broker agents can be modified.

¯ if opportune, or requested, new agents may becreated to pertbrm specific tasks, fulfilling certain

dynamically imposed requirements; for example, forsearching the Internet for specific types ofinformation, or shadowing information at a specificsite.

A principled approach to the design of the architecture is ofcrucial importance: a generic conceptual architecture of abroker agent is needed to support the (re)design processneeded for dynamic creation or modification of agents basedon dynamically imposed requirements. An approach inwhich conceptual design is the basis for structure-preserving (formal) detailed design and operational design,can provide the means to model, specify and implement theflexible structures required.

AcknowledgementsThe authors wish to thank Pieter van Langen for hiscontributions to the generic model of design, and Lourensvan der Meij and Frank Cornelissen for their support of theDESIRE software environment. This research has been(partially) supported by NWO-SION within project 612-322-316: ’Evolutionary design in knowledge-based systems’(REVISE).

References

Benjamins, R., Fensel, D., and Straatman, R. 1996.Assumptions of Problem-Solving Methods and their Rolein Knowledge Engineering. In: Wahlster, W. (ed.).Proceedings European Conference on AI (ECAI ’96), 408-412, Wiley and Sons, Chichester.

Brazier, F.M.T., Dunin-Keplicz, B.M., Jennings, N.R. andTreur, J. 1997. Formal Specification of Multi-AgentSystems: a Real World Case, In: Lesser, V. (ed.),Proceedings First Intenu2tional Conference on Multi-AgentSystems, ICMAS’95, 25-32, MIT Press. Extended versionin: Huhns, M. and Singh, M. (eds.), International Journalof Co-operative Information Systems, IJCIS vol. 6(1):67-94, (1997), special issue on Formal Methods in Co-operative Information Systems: Multi-Agent Systems.

Brazier, F.M.T., Jonker, C.M., Treur, J. and Wijngaards,N.J.E. 1998. The Role of Abilities of Agents in Re-design.In: Gaines, B. and Musen, M. (eds). Proc. of the llthKnowledge Acquisition Workshop, KA W’98. Banff.University of Calfgary. URL: http://ksi.cpse.ucalgary.ca/KAW/KAW98/brazier2.

Brazier, F.M.T., Langen, P.H.G. van, Ruttkay Zs., andTreur, J. 1994. On formal specification of design tasks. InGem, J.S., and Sudweeks, F. (eds.) Artificial Intelligencein Design ’94, 535-552, Kluwer Academic Publishers,Dordrecht.

Brazier, F.M.T., Langen, P.H.G. van, and Treur, J. 1996.A logical theory of design. In: Advances in Formal Design

38 AIMW-98


Methods for CAD, J.S. Gero (ed.), 243-266, Chapmann Hall, New York, 1996.

Brazier, F.M.T., Langen, P.H.G. van, Treur, J.,Wijnganrds, N.J.E. and Willems, M. 1996. Modelling anelevator design task in Desire: the VT example. In:Schreiber, A.Th., and Birmingham, W.P. (Eds.), SpecialIssue on Sisyphus-VT. International Journal of Human-Computer Studies, 1996, 44:469-520.

Brazier, F. M. T., Treur, J., and Wijngaards, N. J. E. 1996.Interaction with experts: the role of a shared task model. In:Wahlster, W. (ed.). Proceedings European Conference on A!(ECA! ’96), 241-245,Wiley and Sons, Chichester.

Breuker, J.A. 1997. Problems in indexing problem solvingmethods. In: Fensel, D. (ed.) Proceedings of the ProblemSolving Methods for Knowledge Based Systems workshop,IJCAI’97, 19-35.

Cetnarowicz, K., KisieI-Dorohinicki, M., and Nawarecki,E. 1996. The Application of Evolution Process in Multi-Agent World to the Prediction System. In: Tokoro, M.(Ed.) Proceedings of the Second International Conference Multi-Agent Systems, ICMAS’96, 26-32, AAAI press,Menlo Park CA.

Chavez, A., Maes, P. 1996. Kasbah: An AgentMarketplace for Buying and Selling goods. In: Proc. of theFirst International Conference on the Practical Applicationof Intelligent Agents and Multi-Agent Technology,PAAM’96, 75-90, The Practical Application CompanyLtd, Blackpool.

Chavez, A., Dreilinger, D., Gutman, R., Maes, P. 1997. AReal-Life Experiment in Creating an Agent Market Place.In: Proc. of the Second International Conference on thePractical Application of Intelligent Agents and Multi-AgentTechnology, PAAM’97, 159-178, The PracticalApplication Company Ltd, Blackpool

Coyne, R.D., Rosenman, M.A., Radford, A.D.,Balachandran, M., and Gero, J.S. 1990. Knowledge-baseddesign systems. Addison-Wesley Publishing Company,Reading.

Fensel, D., Motta, E., Decker, S., and Zdrahal, Z. 1997.Using Ontologies for Defining Tasks, Problem-SolvingMethods and their Mappings. In: Plaza, E. and Benjamins,R. (Eds.). Knowledge Acquisition, Modeling andManagement, proceedings of the 10th European workshop,EKAW’97, Lecture Notes in Artificial Intelligence,1319:113-128, Springer, Berlin.

Jonker, C.M., and J. Treur 1997. Modelling an Agent’sMind and Matter. In: Proceedings of the 8th EuropeanWorkshop on Modelling Autonomous Agents in a Multi-Agent World, MAAMAW’97, Lecture Notes in AI,1237:210-233, Springer Verlag.

Kuokka, D., Harada, L. 1995. On Using KQML for

Matchmaking. In: V. Lesser (ed.), Proc. of the FirstInternational Conference on Multi=Agent Systems,ICMAS’95, 239-245, MIT Press, Cambridge, MA.

Martin, D., Moran, D., Oohama, H., Cheyer, A. 1997.Information Brokering in an Agent Architecture. In: Proc.of the Second International Conference on the PracticalApplication of Intelligent Agents and Multi-agentTechnology, PAAM’97, 467-486, The PracticalApplication Company Ltd, Blackpool.

Numaoka, C. 1996. Bacterial Evolution Algorithm forRapid Adaptation. In: Van de Velde, W. and Perram, J.W.(Eds.) Proceedings of the 7th European Workshop Modelling Autonomous Agents in a Multi-Agent World,MAAMAW’96, Lecture Notes in Artificial Intelligence,1038:139-148, Springer.

Sandholm, T., Lesser, V. 1995. Issues in AutomatedNegotiation and Electronic Commerce: Extending theContract Network. In: V. Lesser (ed.), Proc. of the FirstInternational Conference on Multi-Agent Systems,ICMAS’95,328-335, MIT Press, Cambridge, MA.

Smith, I.F.C., and Boulanger, S. 1994. Knowledgerepresentation for preliminary stages of engineering tasksKnowledge Based Systems, 7:161-168.

Smithers, T. 1994. On knowledge level theories of designprocess. In Gero, J.S., and Sudweeks, F. (eds.) ArtificialIntelligence in Design ’96, 561-579, Kluwer AcademicPublishers, Dordrecht.

Tsvetovatyy, M., Gini, M. 1996. Toward a VirtualMarketplace: Architectures and Strategies. In: Proc. of theFirst International Conference on the Practical Applicationof Intelligent Agents and Multi-Agent Technology,PAAM’96, 597-613, The Practical Application CompanyLtd, Blackpool.

Wooldridge, M.J. and Jennings, N.R. 1995. IntelligentAgents: Theory and Practice. In: Knowledge EngineeringReview, 10(2): 115-152.

Brazier 39


Documents

Compositional Design of a Generic Design Agent1