Incorporating Planning into BDI Systems

(1)

INCORPORATING PLANNING INTO BDI SYSTEMS

FELIPE RECH MENEGUZZI,AVELINO FRANCISCO ZORZO,

MICHAEL DA COSTA MÓRAAND MICHAEL LUCK

Abstrat. ManyarhiteturesofautonomousagenthavebeenproposedthroughoutAIresearh. Themostommon

arhite-tures,BDI,areproeduralinthattheydonoplanning,seriouslyurtailinganagent'sabilitytoopewithunforeseenevents. In

thispaper,weexplorethe relationshipbetweenpropositionalplanningsystemsandthe proess ofmeans-endsreasoningusedby

BDIagentsanddene amappingfromBDImentalstatestopropositionalplanningproblemsandfrompropositionalplansbak

tomentalstates. Inordertotesttheviabilityofsuhamapping,wehaveimplementeditinanextensionofaBDIagentmodel

throughtheuseofGraphplanasthepropositionalplanningalgorithm. Theimplementedprototypewasappliedtomodelaase

studyofanagentontrolledprodutionell.

Keywords. PropositionalPlanning,AgentModelsandArhitetures,BDI,X-BDI

1. Introdution. Developmentofautonomousrationalagentshasbeenoneofthemaindriversofartiial

intelligeneresearhfor sometime[37℄. Initial eortsfousedon disembodied means-endsreasoningwith the

development of problem-solving systems and generi planning systems, suh as STRIPS [15℄, later evolving

into theidea ofembodied problem solvingentities (i.e. agents) [37℄. Inthis line of researh, oneofthe most

widelystudied models of autonomousagentshas been that supported by themental statesof beliefs, desires

andintentions[7℄,ortheBDImodel. WhileeortstowardsdeningBDIarhitetureshavebeensustainedand

signiant,resultingin boththeoretial[34℄andpratialarhitetures[14℄,theyhavealsoledtoadisonnet

betweenthem.

Inpartiular, theoriesofautonomousBDI agentsoftenrelyonlogimodelsthatassumeinnite

omputa-tional power,while arhitetures dened forruntime eieny haveurtailed anagent's autonomybyforing

the agent to rely on a pre-ompiled plan library. Although simple seletion of plans from a plan library is

omputationally eient, at ompile time an agent is bound to the plans provided by the designer, limiting

anagent'sabilityto ope withsituations notforeseen at design time. Moreover,even ifadesigner is ableto

deneplansforeveryoneivablesituationinwhihanagentndsitself,suhadesriptionislikelytobevery

extensive, osetting someof the eieny benets from the plan library approah. Theabsene of planning

apabilitiesthusseriouslyurtailstheabilitiesof autonomousagents. Inonsequene,wearguethat planning

isanimportantapabilityofanyautonomousagentarhitetureinordertoallowtheagenttoopeatruntime

withunforeseensituations.

Though the eieny of planning algorithms has been a major obstale to their deployment in

time-ritial appliations, many advanes have been ahieved in planning [43℄, and developmentsare ongoing [2℄.

Consideringthat planningis an enablerof agentexibility, and that there havebeensigniantadvanes in

planningtehniques, it is valuable and important for autonomous agent arhitetures to employ planningto

allow an agent to opewith situations that the designer wasnot able to foresee. This artile desribesand

demonstratesonesuharhiteture,whihintegratespropositionalplanningwithBDI,allowingagentstotake

advantageof the pratialreasoning apabilities (i.e. seletingand prioritising goals)of the BDI model, and

replaingtheBDI means-ends reasoning(i.e. seletingaourseof ationto ahievegoals)withthe exibility

ofgeneri planning. Ourapproahis underpinned by amappingamong BDI mental statesandpropositional

planningformalisms,allowinganyalgorithmbasedonasimilarformalismtobeusedasameans-endsreasoning

proess for a BDI agent. In order to demonstratethe viability of suh an approah we take a spei BDI

agentmodel,namelytheX-BDImodel[27℄,andmodifyittousepropositionalplanningalgorithmstoperform

means-endsreasoning[30℄.

Thepaperisorganised asfollows: Setion 2ontainsanoverview oftherelated work and main onepts

used throughout this paper; Setion 3 desribes X-BDI and the extensions that allow it to use an external

planningalgorithm; Setion 4ontainsa asestudy used to demonstrate theimplemented prototype;nally,

Setion5ontainsonludingremarksabouttheresultsobtainedin thiswork.

2. Agents and Planners. Inthis setionwereviewbakgroundwork onagentsand planningsystems,

and onlude with a disussion of the integration of these tehnologies in an agent arhiteture, laying the

groundworkfortheremainderofthisartile. Setion2.1providesanoverviewofomputeragentsandtheBDI

(2)

algorithms and problem representation; Setion 2.3 desribes the partiular planning algorithm used in the

prototypedesribedin Setion 3;nally, wedisusshowthese tehnologiesanbepieedtogetherin order to

addresssomeoftheirindividual limitations.

2.1. Agents. Thegrowingomplexityofomputersystemshasledtothedevelopmentofinreasinglymore

advaned abstrationsfortheirrepresentation. An abstrationof growingpopularityforrepresentingparts of

omplexomputersystemsisthenotionofomputer agents [13℄, sofarastobeproposedasanalternativeto

the Turing Mahine as anabstration for thenotion of omputation [19, 42℄. Althoughthere is avariety of

denitions for omputer agents, mostresearhers agree with Jennings' denition of an agentas enapsulated

omputersystem,situatedinsomeenvironment,andapableofexible,autonomousationinthatenvironment

inorder tomeetitsdesign objetives [19℄.

In the ontext of multi-agent systems researh, one of the most widely known and studied models of

deliberative agents usesbeliefs, desires and intentions (BDI) asabstrationsfor thedesriptionof asystem's

behaviour. The BDI model originated from a philosophial model of human pratial reasoning [6℄, later

formalised[11℄andimprovedtowardsamoreompleteomputationaltheory[34,44℄. Thoughotherapproahes

tothedesignofautonomousagentshavebeenproposed[16℄,theBDImodelorvariationsofitareusedinmany

newarhiteturesofautonomousagents[13,31,4,40℄. Morespeially,theomponentsthatharaterisethe

BDImodelanbebrieydesribedasfollows[28℄:

•

beliefsrepresentanagent'sexpetationregardingtheurrentworldstateorthepossibilitythatagiven ourseofationwillleadto agivenworldstate;

•

desiresrepresentasetofpossiblyinonsistentpreferenesanagenthasregardingasetofworldstates; and

•

intentionsrepresentanagent'sommitmentregardingagivenourseofation,onstrainingthe on-siderationofnewobjetives.

TheoperationofageneriBDIinterpreteranbeseenasaproessthatstartswithanagentonsideringits

sensorinputandupdatingitsbeliefbase. With thisupdatedbeliefbase,asetofgoalsfromtheagent'sdesires

isthenseleted,andtheagentommitsitselftoahievingthesegoals. Inturn,plansareseletedasthemeans

to ahievethegoalsthroughintentionswhih representtheommitment. Finally,these intentionsare arried

outthroughonreteations ontainedinthe instantiated plans(or intentions). Thisproessisillustratedin

theativitydiagram ofFigure2.1,whihshowstheomponentsofanagentthat areusedin eahof themain

proessesofBDI reasoning,namely: obtainingsensorinput andupdatingbeliefs; seletingagoalfrom among

thedesires;andadopting intentionsto arryouttheationsrequiredtoahievetheseletedgoal.

This last proess of seleting and adopting intentions to ahieve a goal is one of the most important

proessesoftheBDImodel,sineitaetsnotonlytheationsanagenthooses,butalsotheseletionofgoals,

as an agent must drop goals deemed impossible. This problem of determining whether an agent is apable

of satisfying its objetivesthrough somesequene of ations given an environment and a set of objetives is

sometimesharaterisedastheagentdesign problem [45℄. ThemostwidelyknownBDIagentimplementations

[18,33,14℄bypassthisproblemthroughtheuseofplanlibrariesinwhihtheoursesofationforeverypossible

objetiveanagentmighthavearestoredasenapsulatedproedures. Agentsusingtheseapproahesaresaidto

pursueproeduralgoals. However,thetheories ommonlyusedtounderpinthereationofnewplansofationat

runtimeassumeanagentwithunlimitedresoures,thusmakingtheiratualimplementationimpossible[37,34℄.

When an agent selets target world-statesand then uses someproess at runtime to determine its ourse of

ation,itis saidtopursue delarative goals. Reenteortsseek todealwiththis problemin variousways,for

instane bydening alternate proof systems[27, 31℄ orusing model heking in order to validate the agent's

plan library[5℄. An alternativeapproahto solvingtheproblem istheuse ofplanningalgorithmsto perform

means-endsreasoningat runtime[37,26,47℄.

2.2. Planning Algorithms. Means-ends reasoning is a fundamental omponent of any rational agent

[6℄and isuseful in theresolution ofproblems inanumber ofdierent areas,suh assheduling[38℄, military

strategy [39℄, and multi-agent oordination [12℄. Indeed, the development of planning algorithms has been

oneof the main goalsof AI researh[35℄. In moredetail, aplanningproblem isgenerially dened by three

omponents[43℄:

•

aformaldesriptionofthestartstate;

•

aformaldesriptionoftheintendedgoals;and

(3)

Beliefs

Actions

Sensor Update

Goal Selection

Desires

Intentions

Intention Selection

Action

Fig.2.1.AtivitiesofageneriBDIinterpreter.

Aplanningsystemtakestheseomponentsandgeneratesaset ofationsorderedbysomerelationwhih,

whenappliedtotheworldinwhihtheinitialstatedesriptionistrue,makesthegoals'desriptiontrue. Despite

the high omplexity provenfor the generalase of planningproblems 1

, reent advanes in planning researh

haveled to thereationofplanning algorithmsthat performsigniantlybetterthan previousapproahesto

solvingvariousproblemlasses[43,2℄. Thesenewalgorithmsmakeuseoftwomaintehniques,eitherombined

orseparately:

•

expansionandsearhinaplanninggraph[3℄;and

•

ompilationoftheplanningprobleminto alogialformulato betestedforsatisability(SAT)[20℄. Onesuhplanningalgorithmis Graphplan,whihweonsider inmoredetailbelow.

2.3. Graphplan. Graphplan[3℄isaplanningalgorithmbasedontherstofthesetehniques,expansion

and searh in a graph. It is onsidered to be one of the most eient planning algorithms reated reently

[43,38,17℄,havingbeenrenedintoaseriesofotheralgorithms,suhasIPP(InterfereneProgressionPlanner)

[22℄ and STAN (STate ANalysis) [24℄. The eieny of Graphplan was empirially demonstrated through

the very signiant results obtained by instanes of Graphplan in the planning ompetitions of the AIPS

(InternationalConfereneonAIPlanning andSheduling) [21,25℄.

PlanninginGraphplanisbasedontheoneptofagraphdatastruturealledtheplanninggraph,inwhih

informationregardingtheplanningproblemisstoredin suhawaythat thesearhforasolutionanbe

ael-erated. Planninggraphonstrutionis eient,havingpolynomialomplexityin graphsizeand onstrution

timewithregardtoproblemsize[3℄. Aplanintheplanninggraphisessentiallyaow,inthesenseofanetwork

ow,andthesearhforasolutiontotheplanningproblemisperformedbytheplannerusingdatastoredinthe

graphtospeeduptheproess. ThebasiGraphplanalgorithm(i.e.withouttheoptimisationsproposedbyother

researhers[21,25℄)isdividedintographexpansionandsolutionextration,whihtakeplaealternatelyuntil

ei-therasolutionisfoundorthealgorithmanprovethatnosolutionexists. ThewaythesetwopartsofGraphplan

areusedthroughoutplanningissummarisedintheativitydiagramofFigure 2.2,andexplainedbelow.

Sineaplanisomposedoftemporallyorderedationsand,inbetweentheseationsthereareworldstates,

graphlevelsaredividedinto alternatingpropositionandationlevels,makingitadiretedandlevelledgraph,

(4)

Graph Expansion

Consistent Goals?

Yes

Solution Extraction

No Solution

No

Solution Impossible

Plan Found

Fig.2.2. Graphplanalgorithmoverview.

asshown in Figure 2.3. Proposition levelsare omposed of proposition nodes labelled withpropositions, and

onneted to the ations in the subsequent ation level through pre-ondition ars. Here, ation nodes are

labelledwithoperatorsandareonnetedtothenodesinthesubsequentpropositionnodesbyeetars.

Everypropositionleveldenotesliteralsthatarepossiblytrueatagivenmoment,sothattherstproposition

level representsthe literals that are possibly trueat time 1, the nextproposition level representsthe literals

thatarepossiblytrueattime2andsoforth. Similarly,ationlevelsdenoteoperatorsthatanbeexeutedata

givenmomentintimeinsuhawaythattherstationlevelrepresentstheoperatorsthatmaybeexeutedat

time1,theseondationlevelrepresentstheoperatorsthatmaybeexeutedattime2andsoforth. Thegraph

alsoontainsmutualexlusionrelations(mutex)betweennodes(at thesamegraphlevel)so thattheyannot

besimultaneouslypresentat the samegraphlevelfor thesamesolution. This givesthem afundamental role

in algorithmeieny, astheyallowthesearh forasolutionto ompletely ignorealargenumberof possible

owsinthegraph.

Level 0

Mutex

Action

Proposition

Level 4

Level 3

Level 2

Level 1

Fig.2.3. Aplanninggraph example.

Aftergraphexpansion,thegraphisanalysedbythesolutionextrationpartofthealgorithm,whihusesa

(5)

unless allthegoal propositions are presentand onsistent, sinetheyannot be mutually exlusive at the last

graphlevel. Fundamentalto Graphplanisitsassuranethat,wheneveraplanfortheproposedproblemexists,

thealgorithm willndit,otherwisethealgorithmwilldeterminethat theproposedproblemisunsolvable[3℄.

2.4. Disussion. Whenoneonsiders howBDI reasoningoperates,itisstraightforwardtopereivethat

propositionalplanninganbeusedasameans-endsreasoningomponent. Fromarepresentationalpointofview,

BDImentalstatesanbeonvertedtoplanningproblemswithoutompliation: beliefstranslateintoaninitial

statespeiation,ationsandapabilitiestranslateintooperatorspeiationsandseletedgoalstranslateinto

agoalstatespeiation. Atthissimplelevel,thedelegationof means-endsreasoningtoanexternalplanning

proess an improve the runtime eieny of existing BDI interpreters by leveraging advanes in planning

algorithmsresearh.

3. Introduingproedural planninginto X-BDI.

3.1. Introdution. GiventheshortomingsoftraditionalBDIarhiteturesintermsofruntimeexibility,

andthe performane problemsof alternativearhitetures, wedeneanextended versionoftheX-BDIagent

model[27℄,modiedtoaommodatetheuseofanexternalplanningomponent. Here,wefousonSTRIPS-like

(STanford ResearhInstitute ProblemSolver) formalisms[15℄. Ourformalismisbased ontheoneintrodued

byNebel [30℄, and, aordingto theauthor,isa

S

I L

formalism, i.e. thebasiSTRIPSplus thepossibilityto useinompletespeiationsandliteralsinthedesriptionofworldstates. Itisimportanttopointoutthatthe

formalismdenedbyNebel[30℄ismoregeneral,butsinewedonotaimtoprovideadetailedstudyofplanning

formalisms,weuseasimplerversion. Inpartiular,weuseapropositionallogiallanguagewithvariablesonly

in thespeiationofoperators, andwith operators notbeing allowedto haveonditional eets. InNebel's

desriptionofthetheSTRIPSformalism,oneannotiethatitdealsonlywithatoms. Nevertheless,withinthis

papermoreexpressivityisdesirable,inpartiular,thepossibilitytouserstordergroundliterals. Itispossible

toavoidthese limitationsthroughtheuseof syntatitransformationssothat planners anoperateoverrst

order ground literals. The main ontribution of ourwork lies in the eieny improvement of a delarative

agent arhiteture. The fat that this type of agentarhiteture hastraditionally beennotoriously ineient

highlightstherelevaneofthis eienygain.

3.2. X-BDI. AnX-BDIagenthasthetraditionalomponentsofaBDIagent,i.e. asetofbeliefs,desires

andintentions. TheagentmodelwasoriginallydenedintermsoftheExtendedLogiProgrammingwithexpliit

negation (ELP) formalismreatedbyAlferes and Pereira [1℄,whih inludes arevisionproedure responsible

formaintaininglogionsisteny. Wedonotprovideadesriptionoftheformalismhere,thoughweassumethe

preseneofitsrevisionproedurein ourdesriptionofX-BDI.Givenitsextended logidenition,X-BDIalso

hasasetoftimeaxiomsdenedthroughavariationoftheEvent Calulus [27,23℄.

Theset of beliefs is simply aformalisation of fats in ELP, individualised for aspei agent. From the

agent'spointofview,itisassumedthatitsbeliefsarenotalwaysonsistent,andwheneveraneventmakesthe

beliefs inonsistent,theymustberevised. Thedetailsofthisproessareunimportantin theunderstanding of

theoverallagentarhiteture,butanbefoundin[1℄. Thebeliefrevisionproessin X-BDIistheresultofthe

programrevisionproessperformedinELP.

EverydesireinanX-BDIagentisonditionedtothebodyofalogirule,whihisaonjuntionofliterals

alled

Body

. Thus,

Body

speies the pre-onditionsthat must be satised in order for an agent to desire

aproperty. When

Body

is an empty onjuntion, some property

P

is unonditionally desired. Desires may

be temporally situated, i.e. an be desired at a spei moment, orwhenever their pre-onditionsare valid.

Moreover,adesirespeiationontainsapriorityvalueusedintheformationofanorderrelationamongdesire

sets.

There are two possible types of intentions: primary intentions, whih refer to the intended properties,

andrelative intentions, whihreferto ations abletobring aboutthese properties. An agentmaynot intend

somethinginthepastorthatisalreadytrue,andintentionsmustinpriniplebepossible,i.e. theremustbeat

leastoneplanavailablewhoseresultisaworldstatewheretheintendedpropertyistrue.

Now, we diverge from the original X-BDI arhiteture in several respets. First, the original reasoning

proessveriedthepossibilityofapropertythroughtheabdutionofaneventalulustheoryto validatethe

property. Inbrief,thelogirepresentationofdesiresintheoriginalX-BDIinludedlausesspeiallymarked

(6)

for the implementation of X-BDI planningframework in extended logi, but the implementation of the logi

interpreter was notably ineient for abdutive reasoning. In this work, the planning proess is abstrated

outfrom theoperationaldenitionofX-BDI,allowinganyplanningomponentthat satisestheonditionsof

Setion2.2tobeinvokedbytheagent. Thus,thenotionofpossibilityofadesireisassoiatedwiththeexistene

ofaplantosatisfyit.

The reasoning proess performed by X-BDI begins with the seletion of eligible desires, whih represent

unsatiseddesireswhose pre-onditionsarevalid,thoughtheelementsofthissetofdesiresarenotneessarily

onsistentamongthemselves. A setofeligibledesiresthatarebothonsistentandpossibleisthen seletedas

andidate desires, to whih theagentommitsitself to ahievingby adoptingthem asprimary intentions. In

ordertoahievetheprimaryintentions,theplanningproessgeneratesasequeneoftemporallyorderedations

thatonstitutetherelativeintentions. ThisproessissummarisedinFigure3.1.

Consistency

Maintenance

Perception

Planning

Action

Primary

Intentions

Relative

Intentions

Candidate

Desires

Elligible

Desires

Beliefs

Deliberation

Fig.3.1. X-BDIoperationoverview.

EligibledesireshaverationalityonstraintsthataresimilartothoseimposedbyBratman[6℄overintentions

in the sense that an agentwill not desiresomethingin thepast orsomethingthe agent believeswill happen

withoutitsinterferene. Agentbeliefsmustalsosupportthepre-onditionsdenedin thedesire

Body

. Within

theagent'sreasoningproessthesedesiresgiveriseto asetofmutuallyonsistentsubsetsorderedbyapartial

orderrelation.

(7)

P

thatanbesatisedthroughasequeneofations. Inordertohooseamongmultiplesetsofandidatedesires,

theoriginal X-BDIusesELPonstrutsthat allowdesiresto beprioritisedin therevisionproess. Although

we depart from the original abdution theory, we still use these priority values to dene a desire preferene

relation. Throughthispreferenerelation,adesirepreferenegraphthatrelatesallsubsetsofeligibledesiresis

generated.

Candidatedesiresrepresentthemostsigniantmodiationmadein thispaperregardingtheoriginal

X-BDI[27℄. Originally,X-BDIveriedthepossibilityofadesirethroughtheabdutionofaneventalulustheory

inwhihthebeliefinthevalidityofadesiredproperty

P

ouldbetrue. Suhanabdutionproessis,atually,

aform of planning. Sine ourmain objetivein this paperis to distinguishthe planning proess previously

hard-oded within X-BDI, thenotionof desirepossibilitymust bere-dened. Therefore, wedene the set of

andidatedesirestobethesubsetofeligibledesireswiththegreaterpreferenevalue,andwhosepropertiesan

besatised. Satisabilityisveriedthroughtheexeutionofapropositionalplannerthat proessesaplanning

problemin whih theinitial stateontainstheproperties thattheagentbelievesatthetimeofplanning. The

P

propertiespresentin theandidatedesiresareused togeneratethesetof primaryintentions. Themodied

reasoningproessforX-BDIisillustratedin Figure3.2.

Consistency

Maintenance

Action

Perception

Mapping

Elligible

Desires

Relative

Intentions

Propositional

Planning

Candidate

Desires

Primary

Intentions

Desires

Beliefs

Deliberation

Fig.3.2. Modied X-BDIoverview.

Primaryintentionsanbeseenashigh-levelplans,similar totheintentionsin IRMA[7℄,andrepresenting

(8)

itsgoals. Relativeintentionsthenorrespondtothetemporallyorderedstepsoftheonreteplansgeneratedto

satisfytheagent'sprimaryintentions. Thusthenotionofagentommitmentresultsfromthefat thatrelative

intentionsmustnotontraditorannulprimaryintentions.

3.3. Intention Revision. Theomputationaleortand thetime requiredtoreonsiderthewholeset of

intentionsofaresoure-boundedagentisgenerallysigniantregardingtheenvironmenthangeratio. Intention

reonsideration should thereforenot our onstantly, but only when the world hanges in suh a way as to

threatentheplansanagentisexeutingorwhenanopportunitytosatisfymoreimportantgoalsisdeteted. As

aonsequene,X-BDIusesaset ofreonsiderationtriggers generatedwhenintentionsareseleted,andauses

theagenttoreonsideritsourseofationwhenativated.

ThesetriggeronditionsaredenedtoenforeBratman's[6℄rationalityonditionsforBDIomponents,as

follows. Ifalloftheagent'sprimaryintentionsaresatisedbeforethetimeplannedforthemtobesatised,the

agentrestartsthedeliberativeproess,sineithasahieveditsgoals. Ontheotherhand,ifoneoftheprimary

intentionsisnotahievedat thetimeplanned forit,theagentmust reonsideritsintentionsbeauseitsplans

havefailed. Moreover,ifadesirewithahigherprioritythantheurrentlyseleteddesiresbeomespossible,the

agentreonsidersitsdesiresin orderto takeadvantageofthenewopportunity. Reonsiderationis ompletely

basedon integrityonstraintsoverbeliefs, and sinebeliefs arerevisedat everysensoring yle,it is possible

thatreonsiderationours duethetriggering ofareonsiderationrestrition.

3.4. Implementation. Theprototypeimplementedforthisworkisomposedofthreeparts: theX-BDI

kernel,implementedinProlog;aplanningsystemontainingaC++implementationofGraphplan;andaJava

graphialinterfaeused to easetheoperationof X-BDIandto visualiseitsinteration withthe environment.

Thearhitetureisoutlinedin Figure3.3.

Socket

2

_BDI

_Graphplan

Agent Viewer

Java

Prolog

C++

Plan

Beliefs

Desires

X

Intentions

Fig.3.3. SolutionArhiteture

Here,theAgentViewerinterfaeommuniateswithX-BDIthroughsoketsbysendingtheinputfromthe

environmentin whih theagentisembedded andreeivingtheresultoftheagent'sdeliberation. Throughthe

AgentViewer theuseranalsospeifytheagentintermsofitsdesires,ationsandinitialbeliefs. OneX-BDI

reeivestheagentspeiation,itommuniateswiththeplanningmodule throughoperatingsystemlesand

theProlog/C++interfae. Theplanner is responsiblefor generating aset ofintentionsfor theagent. When

theagentdeliberates,itonvertssubsetsoftheagent'sdesiredpropertiesintopropositionalplanningproblems

andinvokestheplanningalgorithm tosolvethese problemsuntil eitheraplan thatsolvesthehighestpriority

desiresisfound,orthealgorithmdeterminesthat itisnotpossibleto solveanyoneoftheseproblems.

4. A BDI Prodution Cell. Inthis work weuseaBDI agentinorder tomodelaprodution ellasa

asestudy, andasameansto verifythevalidityof thearhiteturedesribed in Setion3. Inpartiular, the

rationalutilisationofequipmentinindustrialfailitiesisaomplexproblem,espeiallyshedulingitsuse. This

problemisompliatedwhenthefailityproduesmultipleomponenttypes,whereeahtyperequiresasubset

of the equipment available. Inour test senario, the proposed prodution ell[46℄, illustrated in Figure 4.1,

isomposed ofseven devies: afeed belt,adeposit belt andfour proessing units upon whih omponentsare

movedtobeproessed.

Componentsenterthe produtionellforproessingthroughthefeed belt and,one proessed by allthe

appropriateproessingunits,theyareremovedfromtheellthroughthedepositbelt. Everyproessingunit is

responsiblefor performingadierentkindof operationontheomponentbeingproessed,and anhold only

oneomponentatagivenmoment. Eahomponentintroduedintotheellanbeproessedbyoneormore

(9)

0000

1111

_Processing

Unit 3

Belt

Deposit

Feed

Belt

Processing

Unit 2

Processing

Unit 4

Processing

L1

Unit 1

Fig.4.1. ABDIProdutionCell.

X-BDI,whih should shedulethe workof theprodutionellin relationto itsbeliefs,desiresand intentions,

re-shedulingwheneversomehangein thesystemours.

TherststepinmodellinganyproblemusingaSTRIPS-likeformalismisthehoieoftheprediatesused

torepresenttheproblem's objet-typesanditsstates. Wedene thefollowingprediatesrepresentingobjets

intheell:

•

omponent(C)denotesthatCisaomponent tobeproessed;

•

proUnit(P)denotesthatPisaproessingunit,whihisalsoadevie;

•

devie(D)denotesthat Dis adevie;

•

feedBeltrepresentsthefeedbelt;

•

depositBeltrepresentsthedeposit belt.

Similarly,wehavethefollowingprediatesrepresentingsystemstates:

•

over(C,D)denotesthat omponentCisoverdevieD;

•

empty(P)denotesthatproessingunitPisempty,i.e. hasnoomponentoverit;

•

proessed(C,P)denotesthatomponentChasalreadybeenproessedbyproessingunitP;

•

finished(C)denotesthatomponentChasalreadybeenproessedbyallappropriateproessingunits andhasbeenremovedfromtheprodutionell;

Next,wedenetheationstheagentisapableofperformingintheontextoftheproposedproblem,these

aresummarisedin Table4.1. Informally,ationproess(C,P)representstheproessingthataproessingunit

PperformsonaomponentCoverit;onsume(C)representstheremovalofomponentCfrom theprodution

ell through the deposit belt; and move(C,D1,D2)represents the motion of omponent C from devie D1 to

devieD2.

Table4.1

Ationspeiationfortheprodutionellagent.

Ation Preonditions Eets

proess(C,P) proUnit(P) proessed(C,P)

omponent(C)

over(C,P)

onsume(C) omponent(C)

¬

over(C,depositBelt) over(C,depositBelt) empty(depositBelt)

finished(C)

move(C,D1,D2) over(C,D1) over(C, D2)

empty(D2)

¬

over(C,D1)

omponent(C)

¬

empty(D2)

devie(D1) empty(D1)

(10)

Theproessingrequirementsofomponentsandtheirprioritiesaremodelledthroughdesires. Thus,wean

modelanagent,pCell,whihneedstoproessomponentomp1byproessingunitsproUnit1,proUnit2and

proUnit3assoonasthisomponentisinsertedintotheprodutionellusingthespeiationofListing1 2

.

Listing1

Speiationofdesiresrelated toproessing omp1.

des(pCell,finished(omp1),Tf,[0.7℄)

if bel(pCell, omponent(omp1)),

bel(pCell, proessed(omp1,proUnit1)),

bel(pCell, -finished(omp1)).

des(pCell,proessed(omp1,proUnit1),Tf,[0.6℄)

bel(pCell, -proessed(omp1,proUnit1)).

Similarly, weanmodel theagent'sneed to proessomponentblo2by proessingunit proUnit3and

proUnit4byaddingto theagentspeiationthedesiresofListing2.

Listing 2

Speiationofdesiresrelated toproessing omp2.

des(pCell,finished(omp2),Tf,[0.6℄)

bel(pCell, -finished(omp2)).

Finally, wemodel the agent'sstati knowledge regardingthe problem domain, in partiular theobjet's

lassesandtheinitialworld-statewiththebeliefsspeiedinListing3.

Thearrivalof anewomponentin theprodutionellissignalledbythesensorsthroughtheinlusion of

omponent(omp1)andover(omp1,feedBelt)intheagent'sbeliefdatabase,ativatingtheagent's

reonsid-erationproess. Given thedesire'spre-onditionspreviouslydened,only thedesiresrelated tothe following

propertiesbeomeeligible:

(11)

Listing3

Domainknowledgefortheprodutionell.

bel(pCell, proUnit(proUnit1)).

bel(pCell, devie(proUnit1)).

bel(pCell, devie(depositBelt)).

bel(pCell, devie(feedBelt)).

bel(pCell, empty(proUnit1)).

bel(pCell, empty(depositBelt)).

•

proessed(omp1,proUnit1);

•

These desiresarethen analysedby theproess ofseleting andidatedesires. Inthis proess, theagent's

eligibledesiresandbeliefsareusedinthereationofplanningproblemsthataresenttoGraphplanforresolution.

Theresultofthisproessingisaplanthatsatisesalltheeligibledesires,withthefollowingsteps:

1. move(omp1,feedBelt,proUnit2)

2. proess(omp1,proUnit2)

3. move(omp1,proUnit2,proUnit1)

TheexisteneofthisplanindiatestoX-BDIthatthespeiedsetofeligibledesiresispossible,thusturning

the previousset of desires into andidate desires, whih generate primary intentionsrepresenting theagent's

ommitment. Next, relative intentionsare generated using the steps in the reently reatedplan, with one

intentionforeah stepoftheplan. Theseleadtheagenttoperform theappropriateations. Onetheations

are exeuted, the andidate desiresfrom the previousdeliberation are satised. Moreover,the pre-ondition

of thedesireto aomplish finished(omp1)beomestrue, reativatingthe agent'sdeliberative proessand

generatingthefollowingplan:

1. move(omp1,proUnit3,depositBelt)

2. onsume(omp1)

Onemore,thisplanbringsaboutsomeintentionsand,eventually,leadstheagenttoat. Now,supposethat

duringtheagent'soperation,anewomponentintheprodutionellarrives. Ifthisourredimmediatelyafter

thedeliberationthatreatedtherstplan,itwouldbesignaledbytheagent'ssensorsthroughtheinlusion of

omponent(omp2)andover(omp2,feedBelt)inthebeliefsdatabase,whihwouldmodifytheeligibledesires

hosenintheseonddeliberationyleto:

•

finished(omp1);

•

ThesedesiresbeomeandidatedesiresbeauseGraphplanisapableofgeneratingaplanthatsatisesall

thedesires. Thenewplanis:

(12)

8. onsume(omp2)

Thestepsofthisplanthusgeneraterelativeintentions,eventuallyleadingtheagenttotheexeutionofits

ations.

5. Conlusions. Inthis paper, wehavedisussed therelationship betweenpropositionalplanning

algo-rithmsandmeans-endreasoninginBDIagents. Totesttheviabilityofusingpropositionalplannerstoperform

means-ends reasoning in a BDI arhiteture, we have desribed a modiation to the X-BDI agent model.

Throughout this modiation, new denitions of desires and intentions were reated in order for the agent

model to maintainthetheoretial propertiespresentin itsoriginal version,espeiallyregardingthedenition

ofdesiresandintentionsimpossibility. Moreover,itwasneessaryto deneamappingbetweenthestrutural

omponentsofaBDIagentandpropositionalplanningproblems. Theresultofimplementingthese denitions

in a prototype an be seen in the ase study of Setion 4, whih represents a problem that the means-end

reasoningproessoftheoriginalX-BDIouldnotsolve.

Consideringthat mostimplementationsofBDI agentsuseaplanlibraryformeans-end reasoningin order

to bypassthe inherentomplexity of performing planningat runtime, X-BDIoers aninnovativewayof

im-plementingmoreexibleagentsthroughitsfullydelarativespeiation. However,itsplanningmehanismis

notablyineient. Forexample,theasestudy desribedinSetion4wasnottratablein theoriginalX-BDI

planningproess. Thus, the main ontributionof ourwork onsistsin addressingthis limitation throughthe

denitionofamappingfromBDImeans-endreasoningtofastplanningalgorithms. Moreover,suhanapproah

enablestheagentarhiteturetobeextendedwithanypropositionalplanningalgorithmthat usesaformalism

ompatible withtheproposedmapping,thusallowinganagentto usemorepowerfulplannersastheybeome

available,ortousemoresuitableplanningstrategiesfordierentproblemlasses.

Otherapproahestoperformingruntimeplanninghavealsobeenproposed,themostnotablereentonesby

Sardinaetal.[36℄andWalzaketal.[41℄. SardinaproposesthetightintegrationoftheJACKagentframework

[8℄ with the SHOP hierarhialplanner [29℄. This approah relies on new onstruts added to an otherwise

proeduralagentrepresentationandtakesadvantageofthesimilarityofhierarhial tasknetwork (HTN)

plan-ning to BDI reasoning. The work of Walzak proposes the integration of JADEX [32℄ike℄What is JADEX?

to a ustomisedknowledge-basedplanner operating in parallel to agent exeution, using a similar proess of

agent-stateonversiontoworkofMeneguzzietal. [26,47℄,aswellastheonepresentedinthis paper.

Someramiationsofthisworkareforeseenasfuture work,inpartiular,theinorporationofthevarious

Graphplanimprovements,aswellastheondutionoftestsusingotherpropositionalplanningalgorithms,SAT

beinganexample. Itislearthatotheragentarhiteturesanbenetfromtheusageofplanningomponents

toallowagentstoopewithunforeseeneventsatruntime,asdemonstratedbyreenteortsinplanningagents

[36,41℄. Therefore,investigatinghowtointegrateplanningapabilitiestoAgentSpeak-basedagentsouldreate

agentsthatantakeadvantageofboththefastresponseofpre-ompiledplansandtheexibilityofbeingable

toplanat runtimetoopewithunforeseensituations.

Aknowledgments. Wewishto aknowledgeX,YandZforthesupportin thiswork.

REFERENCES

[1℄ J.J.AlferesandL.M.Pereira,ReasoningwithLogiProgramming,SpringerVerlag,1996.

[2℄ S. Biundo, K. L. Myers, and K. Rajan, eds., Proeedings of the Fifteenth International Conferene on Automated

PlanningandSheduling(ICAPS2005),June5-102005,Monterey,California,USA,AAAI,2005.

[3℄ A.L.BlumandM.L.Furst,Fastplanningthroughplanninggraphanalysis,ArtiialIntelligene,90(1997),pp.281300.

[4℄ R.H.Bordini,M.Dastani,J.Dix,andA.E.Fallah-Seghrouhni,Multi-AgentProgramming: Languages,Platforms

andAppliations,vol.15ofMultiagentSystems,ArtiialSoieties,andSimulatedOrganizations,Springer,2005.

[5℄ R. H. Bordini, M. Fisher, C. Pardavila, and M. Wooldridge, Model hekingAgentSpeak, inProeedingsof the

2ndInternationalConfereneonAutonomousAgentsandMultiagentSystems(AAMAS-03),Melbourne,Australia,July

2003,ACMPress,pp.409416.

[6℄ M.E.Bratman,Intention,PlansandPratialReason,HarvardUniversityPress,Cambridge,MA,1987.

(13)

[8℄ P. Busetta,R. Rönnquist,A.Hodgson,andA.Luas,Jakintelligentagentsomponents forintelligentagentsin

java. AgentLinkNewsletter,January1999.Whitepaper,http://www.agent-softwa re. om. au

[9℄ T.Bylander,Theomputationalomplexityof propositional STRIPSplanning,ArtiialIntelligene,69(1994),pp.165

204.

[10℄ D.Chapman,Planningforonjuntivegoals,ArtiialIntelligene,32(1987),pp.333377.

[11℄ P.R.CohenandH.J.Levesque,Intention ishoiewithommitment,ArtiialIntelligene,42(1990),pp.213261.

[12℄ J.S.Cox,E.H.Durfee,andT.Bartold,Adistributedframeworkforsolvingthemultiagentplanoordinationproblem,

inAAMAS'05:ProeedingsofthefourthinternationaljointonfereneonAutonomousagentsandmultiagentsystems,

NewYork,NY,USA,2005,ACMPress,pp.821827.

[13℄ W.V.derHoekandM.Wooldridge,Towardsalogiofrationalageny,LogiJournaloftheIGPL,11(2003),pp.133

157.

[14℄ M.d'InvernoandM.Luk,EngineeringAgentSpeak(L):Aformalomputationalmodel,JournalofLogiandComputation,

8(1998),pp.233260.

[15℄ R. Fikes andN.Nilsson,STRIPS: Anewapproahtothe appliation oftheorem provingtoproblem solving,Artiial

Intelligene,2(1971),pp.189208.

[16℄ M.Georgeff,B.Pell,M.E.Pollak,M.Tambe,andM.Wooldridge,Thebelief-desire-intentionmodelofageny,

inProeedingsofthe5thInternationalWorkshoponIntelligentAgentsV:AgentTheories,Arhitetures,andLanguages

(ATAL-98),J.Müller,M.P.Singh,andA.S.Rao,eds.,vol.1555,Springer-Verlag: Heidelberg,Germany,1999,pp.110.

[17℄ J.HoffmannandB.Nebel,TheFFplanningsystem: Fastplan generationthroughheuristisearh,JournalofArtiial

IntelligeneResearh(JAIR),14(2001),pp.253302.

[18℄ F. F. Ingrand, M. P. Georgeff, and A. S. Rao, An arhiteture forreal-time reasoning and system ontrol, IEEE

Expert,Knowledge-BasedDiagnosisinProessEngineering,7(1992),pp.3344.

[19℄ N.R.Jennings,Onagent-basedsoftwareengineering,ArtiialIntelligene,117(2000),pp.277296.

[20℄ H.KautzandB.Selman,Planningassatisability,inProeedingsoftheTenthEuropeanConfereneonArtiial

Intel-ligene,Chihester,UK,1992,Wiley,pp.359363.

[21℄ J.Köhler,Solving omplexplanningtasks through extrationof subproblems, inProeedingsofthe Fourth International

Conferene on Artiial Intelligene PlanningSystems, R.Simmons, M. Veloso, and S.Smith, eds., Pittsburg, 1998,

AAAIPress,pp.6269.

[22℄ J.Köhler,B.Nebel,J.Hoffmann,andY.Dimopoulos,ExtendingplanninggraphstoanADLsubset,inProeedings

of the 4thEuropeanConfereneon Planning, S.Steel, ed.,vol. 1348 ofLeture Notes inComputerSiene,Springer

Verlag,Germany,1997,pp.273285.

[23℄ R.A.KowalskiandM.J.Sergot,Alogi-basedalulusofevents,NewGenerationComputing,4(1986),pp.6795.

[24℄ D. Long and M. Fox, Eient implementation of theplan graph in STAN,Journalof Artiial Intelligene Researh

(JAIR),10(1999),pp.87115.

[25℄ D.LongandM.FoxAutomatisynthesisanduseofgeneritypesinplanning,inProeedingsoftheFifthInternational

ConfereneonArtiialIntelligenePlanningSystems,Brekenridge,Colorado,2000,AAAIPress,pp.196205.

[26℄ F.R.Meneguzzi,A.F.Zorzo,andM.D.C.Móra,PropositionalplanninginBDIagents,inProeedingsofthe2004

ACMSymposiumonAppliedComputing,Niosia,Cyprus,2004,ACMPress,pp.5863.

[27℄ M. d. C. Móra, J. G. P. Lopes, R. M. Viari, and H. Coelho, BDImodels and systems: Bridging thegap., in

IntelligentAgentsV,AgentTheories,Arhitetures,andLanguages,FifthInternationalWorkshop,ATAL'98,vol.1555

ofLNCS,Springer,Paris,Frane,1999,pp.1127.

[28℄ J.P.Müller,Thedesignofintelligentagents:Alayeredapproah,inTheDesignofIntelligentAgents:ALayeredApproah,

vol.1177ofLeture NotesinComputerSiene,SpringerVerlag,Germany,1996.

[29℄ D. S. Nau, Y. Cao, A. Lotem, and H. Muñoz-Avila, Shop: Simple hierarhial ordered planner., in IJCAI, 1999,

pp.968975.

[30℄ B.Nebel,Ontheompilabilityandexpressivepowerofpropositional planningformalisms,JournalofArtiialIntelligene

Researh(JAIR),12(2000),pp.271315.

[31℄ N.NideandS.Takata,DedutionsystemsforBDIlogisusingsequentalulus,inProeedingsoftheFirstInternational

JointConfereneonAutonomousAgentsandMultiagentSystems,ACMPress,2002,pp.928935.

[32℄ A.Pokahr, L. Braubah, and W. Lamersdorf,Jadex: Implementing a bdi-infrastruture forjade agents, EXPin

searhofinnovation(SpeialIssueonJADE),3(2003),pp.7685.

[33℄ A.S.Rao,AgentSpeak(L):BDIagentsspeakoutinalogialomputablelanguage,inProeedingsoftheSeventhEuropean

WorkshoponModellingAutonomousAgentsinaMulti-AgentWorld,W.V.deVeldeandJ.W.Perram,eds.,vol.1038

ofLNCS,Springer,Eindhoven,TheNetherlands,1996,pp.4255.

[34℄ A. S. Rao and M. P. Georgeff, Formal models and deision proedures for multi-agent systems, Teh. Report 61,

AustralianArtiialIntelligeneInstitute,171LaTrobeStreet,Melbourne,Australia,1995. TehnialNote.

[35℄ S.J.RusselandP.Norvig,ArtiialIntelligene: AModernApproah,PrentieHall,NewJersey,1994.

[36℄ S. Sardina, L. de Silva, and L. Padgham,Hierarhial Planning in BDI Agent Programming Languages: AFormal

Approah,inAAMAS'06:ProeedingsoftheFifthInternationalJointConfereneonAutonomousAgentsandMultiagent

Systems,Hakodate,Japan,2006,ACMPress,pp.10011008.

[37℄ M.ShutandM.Wooldridge,Theontrolofreasoninginresoure-bounded agents,TheKnowledgeEngineeringReview,

16(2001).

[38℄ D.E.SmithandD.S.Weld,Temporalplanningwithmutualexlusionreasoning,inProeedingsofthe17thInternational

JointConfereneonArtiialIntelligene(IJCAI),1999,pp.326337.

[39℄ J.R. Surdu, J. M. D. Hill, andU. W. Pooh, Antiipatory planning support system, inWSC '00: Proeedingsof

the 32ndonfereneon Wintersimulation,SanDiego,CA,USA,2000,SoietyforComputerSimulationInternational,

(14)

[40℄ B.vanRiemsdijk,M.Dastani,F. Dignum,andJ.-J.C.Meyer,Dynamisofdelarativegoalsinagentprogramming.,

inLNCS,vol.3476,2004,pp.118.

[41℄ A. Walzak, L. Braubah, A. Pokahr, and W. Lamersdorf, AugmentingBDI Agentswith Deliberative Planning

Tehniques,inTheFifthInternationalWorkshoponProgrammingMultiagentSystems(PROMAS-2006),2006.

[42℄ P.Wegner,Whyinterationismorepowerfulthanalgorithms,CommuniationsoftheACM,40(1997),pp.8091.

[43℄ D.S.Weld,ReentAdvanesinAIPlanning,AIMagazine,20(1999),pp.93123.

[44℄ M.Wooldridge,Reasoning aboutRationalAgents,TheMITPress,2000.

[45℄ M. Wooldridge, TheComputational Complexity of Agent Design Problems, inProeedings of the Fourth International

ConfereneonMulti-AgentSystems(ICMAS2000),E.Durfee,ed.,IEEEPress,2000,pp.341348.

[46℄ A.F.Zorzo,L.A.Cassol,A.L.Nodari,L.A.Oliveira,andL.R.Morais,Shedulingsafety-ritialsystemsusing

apartial orderplanningalgorithm, inAdvanesinLogi,ArtiialIntelligene andRobotis,SãoPaulo,Brazil,2002,

IOSPress,pp.3340.

[47℄ A.F.ZorzoandF.R.Meneguzzi,Anagentmodelforfault-tolerantsystems,inSAC'05:Proeedingsofthe2005ACM

symposiumonAppliedomputing,NewYork,NY,USA,2005,ACMPress,pp.6065.

Editedby: MarinPaprzyki,NiranjanSuri

Reeived: Otober1,2006