Omer F. Rana Lu Moreau
Department of ComputerSiene Department of Eletronisand ComputerSiene
CardiUniversity Universityof Southampton
NewportRoad, CardiCF243XF, UK Southampton SO17 1BJ UK
Abstrat
We emphasise and brieyreview existing infrastruturerequired to realisethe
ComputationalGrid,anddenesuhGridwithrefereneto Knowledge and
Infor-mation Grids. We then propose an agent-based approah for the Computational
Grid,whih isentered on providing \servies"formanagingresoures.
1 Introdution
The\Grid" isan emerginginfrastruture thatonnets multiple regionaland
national grids to reate a universal soure of omputing power | the work
\Grid" was hosen by analogy with the eletri power grid, whih provides
pervasive aess to power. We believe that by providing pervasive,
depend-able,onsistentandinexpensiveaesstoadvanedomputationalapabilities,
databases, sensors, and people, omputational grids will have a transforming
eet similar to the eletri power grid, allowing new lasses of appliations
to emerge.
Ian Foster and Carl Kesselman [14℄
Sine the publiation of the Grid Manisfesto [14℄, researh ativity has dramatially
inreasedinthis area. Theterm \Grid"has beomeareferenetoalargesale pervasive
infrastruture into whih hardware or software omponents an be plugged, and whih
permits easy onguration and reation of new funtionality from existing omponents.
TheGridisthereforetheunderlyinginfrastruturethatenablesnewinformationservies
tobedened, ativated, supported and managed ina uniform way.
Researh in\Grids"has beomeanarea of ativeinterest, withommunitiessuhas
the GridForum [1℄intheUS, theE-Grid [2℄foruminEurope,and theAsia-Pai Grid
Forum inAustralia and Japan, working to unify ommoninterest aross the world.
Due to the very generi nature of the Grid, the term \Grid" has started to have
dierent meanings for dierent people. An organisation based on three layers has been
very dynami servies required by eah of these layers. In this paper, we fous on the
ComputationalGrid, for whih we identify existing approahes, we disuss their
limita-tions and suggest researh diretions for agent-based solutions.
2 A Three Layer Model
A threelayermodelforthe Grid infrastruturewasdesribed byJeery [5℄inastrategy
doument earlier this year. This model has been adopted widely by various researh
ommunitiesin the US and Europe, and we nowdesribeit.
The Computational Grid, the lower layer, is primarily onerned with large-sale
pooling of omputational and data resoures. (Alternatively, this lower layer is alled
the Data Grid.) Suh pooling requires signiant shared infrastruture to enable the
monitoring and ontrolof resoures in the resultingensemble. The ComputationalGrid
generalisesideasundertaken inearlywork onMetaomputing[11℄, onernedwith
reat-ingagiantomputationalenvironment out of adistributedolletionof les, databases,
omputers, sienti instruments and devies.
The Information Grid onstitutes the middle layer, allowing uniform aess to
het-erogeneous information soures and providing ommonly used servies running on
dis-tributed omputationalresoures. Uniformaesstoinformationsoures relieson
meta-data to desribe information and to help integrating heterogeneous soures. The
gran-ularity of the oered servies an vary, from subroutine or method alls to omplete
appliations. Hene, in sienti omputing, servies an inlude the availability of
spe-ialisednumerialsolvers,suhasmatrix solvers andpartialdierentialequationsolvers,
to omplete sienti odes for appliations suh as wheather foreast and moleularor
uid dynamis. In ommerial omputing, servies an be statistial routines based on
existingsoftwarelibraries,suhasSPSSorSAS,orpreditionservieswhihoeroarser
grained funtionality, suh as database proling or visualisation servies. In
hyperme-dia appliations, servies an be multimedia ontent analysis algorithms or hyper-link
servers [21℄. Servies anthereforebeoeredbyindividualproviders orby orporations;
they may be speialised for spei appliations, suh as genomi databases, orgeneral
purpose, suhas numeriallibraries.
The Knowledge Grid is the top most layer and provides speialised servies whih
anlookforpatterns inexistingdatarepositories,andmanageinformationservies. The
knowledgegridisaimedatreatingnew,valueaddedservieswhihannotbedenedas
asingleservie: theyinvolveanaggregationofmanydierenttypesofservies,providing
a orrelation between data sets generated by dierent servies, or ombining results of
existing servies in novel ways.
Itisintendedthateahoftheselayersprovideserviestovariousappliations,ranging
from support for mobile devies, to large sale single appliations suh as modelling
proteinfoldingand onurrent engineering.
AsubstantialpartoftheresearheortdediatedtotheGridhasonentratedonthe
ComputationalGrid. Theprimaryreasonisthatthe Physisommunityisanimportant
Security
I/O
RSL
LDAP
Schedulers
MetaComputing Services
Data
Management
JVM
...
MDS
User
Support
...
...
Mobility
Service
Definition
Service
Decomp.
Service
Match
...
...
Service
Service
Aggreg.
Service
Discovery
Device and Resource Management
Services
Host and Network
Computational Grids
Information Grids
Knowledge Grids
Figure1: Grids and the Computational Grid
generated by the LargeHadron Collider(LHC) experimentstobeundertaken by CERN
in Switzerland. The data volumes are of the order of Petabytes per year, and must be
distributed tosientists and engineers arossthe world.
However, initiativessuhasthe\SemantiWeb"[3℄, orresearhinollaborative
envi-ronments[22℄,learlyinvestigatefundamentalissues thatarerelevanttotheInformation
and KnowledgeGrids.
Agents [28℄ anprovidea useful abstrationat eah ofthe threeGrid layers. Indeed,
bytheirabilitytoadapttotheprevailingirumstanes, agentswillprovideserviesthat
are very dynami and robust, and therefore suitable for a Grid environment. However,
in this paper, we onentrate on the rst layer, outlining how agents an support and
extend existing omputationalinfrastrutures.
Our emphasisonthe lowerlayerisprimarilyduetothe lakofagreementabout
stan-dards for servies to be supported within Information and Knowledge Grids at present.
The resoure management problem has been investigated in detail, and a onsiderable
researh eort has been expended to reateand support standards in this area (suh as
LDAP). However, what is laking in urrent Computational Grid eorts, is the lak of
servie disovery and management, whih an agent based paradigm an oer. We feel
that a smallimprovementat this layer an havea signiant impaton the higher level
InformationandKnowledgeGrids;therefore,weshowhowagentsanbeusedtosupport
In this setion,we identify someof the urrent appliations ofthe Grid and derive some
requirements. We disuss some of the servies of the omputationalgrid whih have to
meet these requirements.
3.1 Overview
TheComputationalGridaimstoprovidetheinfrastrutureforintegratingomputational
resoures to form a single virtual mahine, or enable solving large sale problems that
annot be solved on a single system. Examples inlude distributed interative
simula-tion suh as military simulations, and simulation of physial proesses suh as limate
modelling. Itmaybeusefultonotethatsuhresoureintegrationmaynotleadto
perfor-maneimprovements,insteadprovidinghigh throughputomputingtoimprove resoure
utilisation. Appliationsthatouldbenetfromthisapproahinludetheuseofmultiple
distributed workstationsto solve hard ryptographi oromplex design problems.
The ComputationalGrid may also be used to provide on-demand omputing and to
meetshort-termrequirementsforresouresthatannotbeost-eetivelyoronveniently
loated loally. The resoures may provide omputation apabilities, or may inlude
software and data repositories, speialised sensors and other devies. Unlike existing
appliations based on distributed superomputing, these new appliation domains are
oftendriven byost-performane onerns ratherthan absoluteperformane. Partiular
hallenges in these appliationsare the existene of dynamiresoure requirements, and
the potentially large population of users and resoures involved; in partiular, loation,
sheduling, ode management, onguration, fault-tolerane and seurity deserve some
investigation. To support some of these requirements, we dene a servie layer that
an benet from existing monitoring and management servies within Metaomputing
systems. Suh a servie layer an inlude (1) management servies, suh as loating
devies,migratingjobstodevies,loadbalaningondevieset,or(2)apabilityservies,
suhas matrix solving orother linearalgebra operations,runningamoleulardynamis
oruiddynamis odeet. Hene, weouplemanagementserviesurrently availablein
Metaomputing systems, as illustrated in gure 1, with a layer that an aggregate and
ombinethesewithsoftwarelibrariesthatoerapartiularfuntionalityforappliations.
The ComputationalGrid may also provide support for data-intensive omputing, in
whihthegrid isused tosynthesisenew informationfromdatamaintainedin
geographi-ally-distributed repositories, digital libraries, and databases. Challenges in this lass
of appliations are the shedulingand onguration of omplex, high-volumedata ows
through the network and multiple levels of proessing. Collaborative and onurrent
engineering provides one example of suh design ativities, enabling groups of users to
ollaborate within virtual environments. In many ases, these appliations involve
Figure1providesourperspetiveontheComputationalGrid,withrefereneto
Informa-tion and Knowledge grids identied above. As illustrated, our modelof Computational
Grid omprises threelayers of servies: (1) host and network servies, (2)meta
omput-ingserviesand(3)devie andresoure management. Theselayersare basedonexisting
implementationinfrastrutures,suhasJava/Jini[12℄ and Metaomputingsystemssuh
as Globus[9, 14℄orLegion [10℄.
At the lowest level, we nd a set of resoures and loal shedulers. Resoures an
be single proessor mahines managed by an operatingsystem, ormultiproessor
super-omputers managed by bath queueing systems suh as LSF or Codine. Resoures may
alsoinludemobiledeviesontainingaJavaVirtualMahine (JVM)orotherembedded
ontrollers, suh as Digital Signal Proessor (DSP) hips. Eah individual resoure or
luster isresponsible for managingtasksloally through the use of aloalsheduler.
The seond layer is omposed of a set of Metaomputing servies relying on loal
resoures and sheduler. In partiular, the Metaomputing Diretory Servie (MDS)
is responsible for translating devie harateristis into a searhable tree, omprising
devie identiers, parameters suh as CPU speeds, loal memory et. Devies an be
managed within administrative domains. Seurity servies an also be provided at this
level, oering authentiation based on ertiates, or lower level enryption tehniques
suh asSSL.
At the third level, we then nd higher level servies, for servie disovery, servie
advertisement and servie management.
4 Agent Based Servies
Grids useexisting infrastrutureswherever possible,espeiallyifstandardisation workis
already in progress or ompleted. We review some of the tehnologies whih ould
on-tributetothe ComputationalGrid,analyse theirlimitations,and suggestagent-oriented
solutions.
4.1 Resoure and Servie Disovery
Resoure disovery exists within many Metaomputing toolkits, suh as Globus and
Legion,andisprimarilyonernedwithloatingdeviesbasedontheirIPaddresswithin
a domain or sub-domain. The Lightweight Diretory Aess Protool (LDAP) based
hierarhialnamingshemehasbeenwidelyusedtoregisterdevies,enablingnewdevies
to be registered, and available devies to be disovered at run time. Suh a resoure
disovery servie an be used by a resoure management system to identify possible
devies for task plaement and exeution.
Existing resoure disovery servies an only be used to loate available devies for
exeuting tasks, and generally do not provide details of the software available on these
devies. Insomelimitedases,softwaredetailsaremadeavailable,butthesearegenerally
levelinformationisalsoinluded inaresoure speiation. An agentframeworkan be
usedtoprovidesuhserviedisovery,basedonaServieAdvertiseragent,aMathMaker
agent,andaServieRequestagent[6,7℄. Servieadvertisementsanrangefrommethods
or proedure alls, to more omplex signatures speifying servie types and ategories.
The mathmaking proess an alsovaryin omplexity fromexat mathing of servies,
to mathing based on ategory or parameters. Furthermore, agent frameworks provide
various mehanisms by whih omponents may be kept informed, suh asasynhronous
notiations orthe publish/subsribe paradigm[24℄.
Jini The Java programming language provides many features onduive to the
de-velopment of the Computational Grid, most signiant of whih is the Jini API. The
`Lookup-Disovery'protoolinJini[15℄,alongwiththe `Leasing',`Transations'and
dis-tributed`Events'serviesanbeusedasunderlyinginfrastruturefordevelopingresoure
andserviedisovery,serviemanagementandmigration,andserviedeompositionand
delegation.
For instane, the Leasing servie an be used to grant liensing periods, over whih
a user or a servie an invoke another servie. On the expiry of this period, aess
to the servie is withdrawn, requiring the user or servie to aquire a new lease. Suh
funtionalityinoreinfrastrutureanbeusefulforbuildingmoreomplexservies,suh
asmathmakingandservieadvertising. AgentframeworkssuhasSoFAR[24℄usesuh
a onept of lease inthe semantis of registration and notiation.
However, issues suh as stability,seurity and salability are laking in the Jini API
atpresent, and these needto beaddressed before itan be deployed within the
Compu-tationalGrid. Furthermore, servie disovery and mathing in Jiniis based onan exat
syntax math between servie requestand servie interfae, andbased onJava primitive
orderived types. This may not beadequatewhen servie requestsmust bedeomposed,
ordelegated toother providers.
The Jini API an also be used in ollaborationwith the JavaSpaes API toprovide
distributed shared memory support for sharing servies. JavaSpaes is based on the
onept of tuple spaes in Linda,and enablesservie types tobegrouped, and mathed
on type signatures. A similar servie from IBM alled TSpaes, also provides a similar
funtionality, and may be used in ollaboration with Jini. This idea of tuple spae has
been reused to oordinatethe ativities of mobile agents [20℄.
4.2 Servie Speiation
A shared data model or framework is required to enable resoures and servies to be
desribed. This is neessary to math servie requests with servie availability, for
ser-vie deomposition, and for new servies to be made available. The DARPA Agent
Markup Language (DAML) [19℄ provides a onstraint language to enable the
speia-tion of servies, and oers auseful tool for enoding servies. No standard data models
or ontologies exist at present, however, for speifying resoure apabilities and servies.
lan-takle the more diÆulttask of dening servies.
Itisunlikelythataommonontologywillemerge,suggestingthatnegotiationbetween
servie providers and servie users is more likely to be the prevalent senario. Dynami
ontology reation is important in this ontext, to enable users and providers to use
ommontermsorthemestonegotiateand propose servies. Animportantonsideration
inthis ontext isthe timetoreahagreement: negotiationprotools shouldbe hosenin
funtion of their ost and their impat onthe salabilityof the Computation Grid.
4.3 Metaomputing Frameworks
Metaomputingframeworks generallyinvolve supportfor aggregating various aspets of
resoure management, suh as sheduling, seurity, ommuniation, resoure loation,
resoure alloation, proess management, and data aess. Resoures an be tightly
oupledparallel mahines, orloosely ouplednetworks of workstations. Metaomputing
frameworks do not provide loal shedulers or resoure managers, but enable multiple
resoure managers to be integrated at a higher level. In this way, loal ontrol and
resoure management poliies an be respeted, whilst enablinglarge sale appliations
tobeexeuted ondistributed omputing infrastruture.
The Globus projet [9, 14℄ is one example of a Metaomputing framework, and is
based on the assumptions that, (1)grid arhitetures should provide basi servies, but
not presribe partiular programming models or higher-level arhitetures, (2) grid
ap-pliationsrequire servies beyond those provided by today's ommodity tehnologies.
Legion [10℄ provides another suh environment, where a olletion of workstations,
vetor and parallel mahines onneted by loalarea and larger-sale networks appears
totheuserasasingleomputer. Legionuses objet-oriented designtehniquestospeify
resoure apability, their loation and deployment information,and means of aessing
them. The Legion arhiteture denes a omplete objet model that inludes objet
abstrations for ompute resoures (alled host objets), storage systems (alled data
vault objets), aswellasother objetlasses. Users an use inheritanetospeialise the
behaviour of these objets to support spei requirements, as well as to develop new
objets. The use of reetion (the representation of parts of the underlying system as
objetsthatanbediretlyoperatedontoaessandhangesystembehaviour)is
parti-ularlyimportantin Legion. For example,host objets represent Legion proessors. One
ormorehost objetsrunoneahomputing resouresinluded inLegion. These objets
reate and manage proesses for appliation-levelLegion objets. Objet lasses invoke
theoperationsofhostobjetstoativatetheir instanesonthe omputingresoures that
thehostobjetsrepresent. RepresentingomputingresouresasLegionobjetsabstrats
the heterogeneity of dierent host omputing platforms, and allows resoure owners to
manageand ontrol theirresoures withinthe ontext of the system.
Metaomputingframeworksavailableatpresentareprimarilyonernedwithresoure
disovery and management. There is little or no support available to speify servies.
They provideuseful infrastruturefor buildingagent basedservie disoveryand
Mobileagentsalsoprovideauseful abstrationfor supportingservie migrationand
mo-bileserviesintheComputationalGrid. ExistinglibrariessuhasVoyager(ObjetSpae)
[4℄, Aglets (IBM Researh) [8℄, D'Agents (DartmouthCollege) et, provide welldened
alls toenableaservie tobemigrated,hekpointed,and runonremotemahines.
Vir-tually allof these librariesstill rely onthe existene of a \plae" proess on the remote
host to reeive in-oming ode. No support is provided to bootstrap or initialise new
hosts at run time intoa luster inany of these libraries.
Additional work is needed to enable mobile agents to be deployed within the
Com-putational Grid, and ranges fromsupport for managing\forwarders" duringservie
mi-gration [25℄, maintaining state onsisteny after migration, hekpointing the omplete
exeutionstate,seletivedata migrationafterserviemigration,and mehanismstodeal
with seurity, related both to the host reeiving the mobile ode, and the integrity of
the mobile agent itself. All of these requirements will bring seurity, adaptability and
robustness tomobile agent systems; their deployment inthe ComputationalGrid is still
onditional tothe salabilityof theirsupporting infrastruture.
We an identify ommon agent servies in the ontext of the Computational Grid,
whih may be shared aross appliations. These servies an be oered entrally and
shared, or servies may be federated, with one servie in eah resoure luster. The
objetive of dening suh agent servies is to identify ommon usage patterns aross
resoures, whih an vary in type and apability, and to ombine expertise at a single
plae.
4.5 High-Level Multi-Agent Interations
All the servies desribed beforewould benet of some agent-based approahes, but we
rekon that the benet of agents would further be visible in higher-level servies that
involve the ollaborationof several nodes in the omputationalGrid.
Two forms of interations between autonomous agents are generally distinguished.
Cooperation (ooperative problem solving) is the proess by whih a group of agents
hoose to work togetherto ahieve theirgoal[27℄. Negotiation is the proess by whiha
groupofagentsommuniatewith oneanothertotryandometoamutuallyaeptable
onlusion [26℄.
In thegeneralase,resoure alloationanbeseenasaprobleminvolvingthe
ollab-orationof several nodes, whihhave toagree onasatisfatoryalloationof resoures for
the urrent problem and the prevailing irumstanes. We an foresee a very dynami
market-based approah where idle nodes bid for omputations to be started on or even
migrated tothemselves. A similar approah is adopted in the design of new generation
mobilephonenetworks,wheresoftware agentsare usedtotradeommuniationservies,
e.g. QoS [23℄. Multi-agent high-level interations suh asautions seem therefore worth
The DARPA CoABS (Control of Agent Based Systems) projet [18℄ aims to develop
agentsystems for oering speialisedservies, suh as omponentinteration managers,
databasewrappers,traders/brokers,toresoureplannersandinterationsmanagers. The
remit of the projet is very wide, and a diverse set of requirements have been identied
withintheprojet. TheapproahproposedinthispaperaddressesasubsetoftheCoABS
funtionalities. The omputationalgridismorewelldened,andontainsamorepreise
set of requirements. There are also more well dened standards whih operate at the
level ofthe ComputationalGrid. However, as theCoABS projet matures, itislikelyto
impat various aspets of the ComputationalGrid also.
5 Conlusion
Establishing and implementing Grids is an importantundertaking, and oers new
hal-lenges in developing robust and salableinfrastrutures. Wehave identied three types
ofGridsinthispaper, aComputationalGridasoreinfrastrutureforsupportinghigher
level servies in Informationand Knowledge Grids.
Sofar,theComputationalGridhasbeenonsideredprimarilyasresourespeiation
and management. We have extended this notion to also inlude servie disovery and
speiation; wehave suggestedthe use ofagents,asoeringservies withinsuhagrid,
and alsoas means to manageexisting resoures and integrateexisting frameworks. We
havepresentedinfrastruturetoolsandframeworksthatouldbeofbenetinestablishing
the Computational Grid, and we have proposed an agent based arhiteture for the
ComputationalGrid.
Referenes
[1℄ The US Grid Forum, see Web site at: http://www.gridforum.org
[2℄ The European Grid Forum, see Web site at: http://www.egrid.org
[3℄ The Semantis Web http://www.semantiweb.org
[4℄ ObjetSpae. Voyager. http://www.objetspae.om /.
[5℄ Keith G Jeery, \Knowledge, Information and Data", A brieng to
the OÆe of Siene and Tehnology, UK, February 2000. Available at:
http://www.itd.lr.a.uk/A ti vity Publ iat ions /23 9
[6℄ K. Syara, J. Lu, and M. Klush. \Interoperability among Heterogeneous Software
Agents on the Internet", Tehnial report, Carnegie Mellon University, Otober
1998. Researh reportno. CMU-RI-TR-98-22.
[7℄ OmerF. Rana,\DynamiResoure Disoverythrough MathMaking",Proeedings
Ams-Agents with Aglets. Addison-Wesley,1998.
[9℄ GregorvonLaszewski,\CoG-Kits: ABridgeBetweenCommodityDistributed
Com-puting and High-Performane Grid",Mathematis and Computer Siene Division,
Argonne National Laboratory, see Web site at: http://www.globus.org,2000.
[10℄ S. Chapin, J. Karpovih and A. Grimshaw, \The Legion Resoure Management
System",Proeedings ofthe5thWorkshoponJobShedulingStrategiesforParallel
Proessing atIPDPD99, San Juan, Puerto Rio, April1999.
[11℄ G. Fox and W. Furmanski, \Petaops and Exaops: Superomputing on the Web",
IEEE InternetComputing 1(2), Marh-April 1997.
[12℄ R. Ashri and M. Luk, \Paradigma: Agent Implementation through Jini",
Pro-eedings of the EleventhInternational Workshop onDatabase and Expert Systems
Appliations, IEEE Computer Soiety Press, 2000.
[13℄ R. Raman,M. Livny,and M. Solomon.\Mathmaking: DistributedResoure
Man-agement for High Throughput Computing", Proeedings of the Seventh IEEE
In-ternational Symposiumon HighPerformane Distributed Computing,July 1998.
[14℄ I. Foster and C. Kesselman (eds.), \The Grid : Blueprint for a New Computing
Infrastruture", Morgan Kaufmann,1999.
[15℄ W. Keith Edwards, \CoreJini", Addison Wesley, 1999.
[16℄ OmerF. Rana andDavidW.Walker, \Agentbased Resoure Integration inPSEs",
Proeedings of HICSS34,January 2000. (toappear)
[17℄ Frank Manola, \Charaterizing Computer-Related Grid Conepts", Tehnial
Re-port, Objet Servies and Consulting, In, Deember 1998. See Web site at:
http://www.objs.om.
[18℄ The CoABS DARPA Projet, \Control of Agent Based Systems", see Web site at:
http://oabs.globalinfotek. om /.
[19℄ The DARPA Agent Markup Language (DAML), see Web site at:
http://www.daml.org.
[20℄ G.Cabri, L. Leonardi,and F. Zambonelli. Reative Tuple Spaesfor Mobile Agent
Coordination . In Proeedingsof the 2nd International Workshop on Mobile Agents
(MA'98), number1477 in LNCS, 1998.
[21℄ David De Roure, Nigel Walker, and Leslie Carr. Investigating Link Servie
Stairmand. An open arhiteture forsupportingollaborationonthe web. In WET
ICE '98 | IEEE Seventh International Workshops on Enabling Tehnologies:
In-frastruture forCollaborativeEnterprises, June17-19,Stanford University,
Califor-nia, USA, pages90{95, 1998.
[23℄ Gwenael Le Bodi, Demessie Girma, James Irvine, and John Dunlop. Dynami
3GNetworkSeletion forInreasingtheCompetitionintheMobileCommuniation
Market. inPro.IEEE Vehiular Tehnology, Boston, MA, September 2000.
[24℄ LuMoreau,NikGibbins,DavidDeRoure,SamhaaEl-Beltagy,WendyHall,Gareth
Hughes, Dan Joye, Sanghee Kim, Danius Mihaelides, Dave Millard, Sigi Reih,
Robert Tansley, and Mark Weal. SoFAR with DIM Agents: An Agent Framework
for Distributed Information Management. In The Fifth International Conferene
andExhibition onThe PratialAppliationof Intelligent Agentsand Multi-Agents,
pages 369{388,Manhester, UK,April 2000.
[25℄ Lu Moreau. DistributedDiretoryServie and MessageRouter for MobileAgents.
Siene of ComputerProgramming, Aepted forpubliation.
[26℄ P. Faratin, C. Sierra, and N. R. Jennings. Negotiation deision funtions for
au-tonomous agents. Int. Journal of Robotis and Autonomous Systems, 24(3-4):159{
182, 1998.
[27℄ M. J. Wooldridge and N. R. Jennings. Cooperative problem solving. Journal of
Logi and Computation, 9(4):563{592,1999.
[28℄ M.WooldridgeandN.R.Jennings.IntelligentAgents: TheoryandPratie.
