Adding P e e ences and Mo al Values in an
Agen -Based Simula ion amewo k o
High-Pe o mance Compu ing
Da id Ma in Gu ie ez3, Ja ie V´azquez-Salceda1, Se gio Al a ez-Napagao12,
and Dmi y Gna yshak2
1Uni e si a Poli `ecnica de Ca alunya-Ba celonaTECH, Edi ici Omega.
C/ Jo di Gi ona 1-3, E - 08034 Ba celona
{[email p o ec ed], [email p o ec ed]}
2Ba celona Supe compu e Cen e (BSC). C/ Jo di Gi ona 1-3, E - 08034 Ba celona
{[email p o ec ed]}
3PwC Spain. A . Diagonal, 640, E - 08017 Ba celona
{[email p o ec ed]}
Abs ac Agen -Based Simula ion is a sui able app oach used now-a-
days o simula e and analyze complex socie al en i onmen s and sce-
na ios. Cu en Agen -Based Simula ion amewo ks ei he scale qui e
well in compu a ion bu implemen e y simple easoning mechanisms,
o employ complex easoning sys ems a he expense o scalabili y. In
his pape we p esen ou wo k o ex end an agen -based HPC pla o m,
enabling goal-d i en agen s wi h HTN planning capabili ies o scale and
un pa allelly. Ou ex ension includes p e e ences o e hei objec i es,
p e e ences o e hei plans, ac ions, and mo al alues. We show he
exp esi eness o he ex ended pla o m wi h a sample scena io.
Keywo ds: agen -based simula ion, goals, p e e ences, alues
1 In oduc ion
Agen -Based Simula ion (ABS) is a compu a ional app oach o simula ing he
ac i i ies and in e ac ions o au onomous agen s in o de o be e unde s and
how a sys em beha es. Fu he mo e, hey allow o he simula ion o complex
en i onmen s whe e pe cep ion, decision-making p ocesses and ac ions ca ied
ou a e dispe sed among se e al s akeholde s o agen s. The pu pose o ABS is
he e o e o ob ain explana o y insigh in o he beha io o a g oup o agen s
which sha e a common en i onmen . ABS can be applied o many ields such
as biology, social sciences, ecology, economics, policy-making, e c. Speci ically,
ABS can be used o analyze he social ela ionships be ween agen s by means
o no ms, mo al alues, and social con en ions, hei adhe ence o hose no ms
and alues, how hey a ec and limi hei ac ions, and how hey may change
o e ime as he agen s in e ac wi h each o he and hei en i onmen .
Many ABS amewo ks ha e been buil ocusing on la ge simula ions o be
un in High-Pe o mance Compu ing (HPC) pla o ms. In cu en HPC-Based
ABS app oaches (such as Repas [24], NETLOGO[21], and MASON[14]) models
may be ele a ed o and examined a genuinely la ge scales a he expense o ha -
ing agen s wi h limi ed easoning capabili ies and/o limi ed in e ac ion among
hem, some imes e en educing agen s o me e ule-based o unc ional inpu - o-
ou pu ans o me s. An opposi e app oach a e Mul i-Agen amewo ks (such
as Jadex[5], 2APL[8], BDI4Jade[15] o GOAL[10]) ha o e cogni i e agen s
wi h mo e powe ul p ac ical easoning capabili ies, bu a he expense o ha -
ing e y limi ed scalabili y. Many o he app oaches in li e a u e o e di e en
le els o easoning and scalabili y ([1] and [18] p o ide an in e es ing compa a i e
analysis on many o hem, showing he easoning le el s. scalabili y ade-o ).
In [9], Gna yshak e al. p esen a cus om Py hon-based BDI-agen simula ion
amewo k capable o bo h hos ing agen s imbued wi h mo e powe ul p ac ical
easoning capaci y and unning simula ions wi h la ge numbe s o hese agen s.
Scalabili y is ackled in his amewo k by pa allelizing ia PyCOMPSs[20] he
easoning cycle o goal-o ien ed agen s. In his pape we add ess he issue o
u he enhancing his amewo k by gi ing agen s he capabili y o deal wi h
p e e ences o e hei objec i es, p e e ences o e he ac ions hey ake in o de
o accomplish hose objec i es and (mo al) alues, as a i s s ep owa ds a pow-
e ul agen -based mic o-simula ion amewo k o analyse he impac o social
alues, no ms and con en ions in la ge popula ions. In his wo k we also aim o
explo e how a we can go wi hou using numbe s in ou p e e ence mechanisms.
Gene ally, humans do no eason using ha d numbe s bu in quali a i e e ms.
Howe e , all s a e-o - he-a app oaches we ha e analysed [6,17,23,7,22] end up
adding ha d numbe s and/o ad-hoc nume ical o mulae o hei selec ion s a -
egy. So we aim o explo e how no using numbe s limi s he exp essi eness o ou
sys em, how se e e his limi a ion is, and d aw some conclusions as o whe he
i is accep able o use numbe s o a ain a desi able le el o complex easoning.
This pape is s uc u ed as ollows: in §2 we b ie ly desc ibe he p e ious
wo ks we used as e e ence; in §3 we desc ibe he concep ual model and how
we added goals, p e e ences o e goals, p e e ences o e plans and ac ions, and
suppo o he exp ession o mo al alues; in §4 we show how ou addi ions o
he model wo k in a sample scena io; and in §5 we conclude by discussing some
limi a ions o he cu en model and ex ensions o be explo ed as u u e wo k.
2 Rela ed Wo k
Ou model o goals has been inspi ed by wo agen amewo ks wi h wo king
implemen a ions: GOAP and BDI4JADE.
GOAP [16] is he AI c ea ed o he enemies o he ideo game F.E.A.R,
mainly o malized by Je O kin. In GOAP, goals a e ep esen ed by speci ying
adesi ed s a e o he wo ld ha agen s s i e o achie e. This desi ed s a e
is desc ibed using he same s uc u e used o he cu en s a e o he wo ld, an
agen ’s belie s, ac ions’ e ec s, e c. Agen s can ha e many independen goals, bu
can only pu sue one a he same ime. In o de o plan, an agen mus ha e a se
o a ailable ac ions, a se o belie s abou he wo ld and senso s o pe iodically
upda e hose belie s, and a se o goals. Each goal has a cu en p io i y, and he
agen will choose o plan o he goal wi h he highes cu en p io i y. GOAP
uses nume ic p io i ies (i.e., a quan i a i e ela ion a he han quali a i e). A*
is used o plan wi h a heu is ic minimizing he weigh ed numbe o ac ions used
o each he desi ed s a e., i.e., minimize he sum o cos s o he ac ions in he
plan. We bo ow such goals de ined as desi ed wo ld s a es (see §3.1).
Ing id Nunes’s BDI4JADE [15] pla o m p o ides a BDI laye on op o JADE
[2]. I uses he same s uc u e as O kin’s GOAP o ep esen goals (desi ed
s a e o he wo ld). I suppo s he decla a ion o di e en ypes o goals: ‘belie
goals’ (goals ha deal wi h s a es o he wo ld desc ibed by boolean a iables),
‘belie se alue goals’ (same as be o e, bu a iables a e con inuous o ha e mo e
han wo possible alues), ‘composi e goals’ (used o ep esen goals composed
o subgoals which ha e o be achie ed sequen ially o in pa allel), e c. I also
di e en ia es be ween desi es (non-commi ed goals) and in en ions (commi ed
goals). Plans a e an o de ed se o ac ions and a e execu ed o achie e a speci ic
goal. In BDI4JADE agen s do no ha e a se o ac ions ha hey can use o
build plans, bu a he , hey ha e a lib a y o plans ha he agen s can choose
om. Each plan in he lib a y has some applicabili y condi ions (equi alen
o ac ions’ p econdi ions) ha a e used in he plan selec ion p ocess. We ge
inspi a ion om BDI4JADE on i s plan selec ion s a egy.
Ou main inspi a ion o he modelling o p e e ences o e goals comes om
CP-ne s[4]. Al hough ou ac ual implemen a ion is de ini ely no an implemen a-
ion o a CP-ne , he main inspi a ions we ha e d awn om hem is o es ablish
one de aul and many condi ional p eo de ela ionships o e goals, and building
a g aph o bo h isualize hem and in e p e hem. We also analysed Dignum e
al. app oach in [7] o model alues ( o adap i o model p e e ences o e goals),
bu upon close inspec ion, we decided no o ollow his app oach since i uses
nume ical alues and in his wo k we aim o a mo e quali a i e app oach.
In he case o p e e ences o e plans, we d ew a g ea deal o inspi a ion
om Visse ’s wo k in [22]. I in oduces he concep s o goals’ p ope ies, which
we use ex ensi ely in ou modeling o p io i ies o e plans. We also make use
o hei mechanism o p ope y p opaga ion in ou implemen a ion. We should
no e ha ou implemen a ion is simple han hei s. Fo ins ance, he pape
de ines bo h p ope ies o goals (disc e e alues ha a p ope y can ake) and
esou ces o goals (nume ical alues and in e als ha ep esen how much o a
esou ce -e.g., money, ood- is being consumed by a goal o a sub-goal), bu we
chose o simpli y he app oach and add only disc e e p ope ies, as we wan o
explo e a quali a i e, scala - ee p e e ence app oach.
3 Concep ual model
Amul i-agen sys em Mis de ined as he uple M={E, A+,C} whe e E
is an en i onmen , in which he agen s eside, ha hey can pe cei e, ga he
in o ma ion om, and ac on; A+is a non-emp y se o agen s; Cis a con olle ,
de ined as he uple C={I, inAcs}whe e Iis he inbox o all he agen s’
ou going messages (suppo ing agen communica ion), and inAcs is he se o
all he ac ions o be exe cised on he en i onmen ( egula ing how agen s access
and ac upon i ).
An agen is de ined as Ai={ID, msgQs, ou Acs, Bh, B,G, gc,Pc,MP,Pg,Pp}
whe e:
–ID ={AgID, AgDesc}is Ai’s iden i y da a:
•AgID is he unique iden i ie o Ai
•AgDesc is an a bi a y desc ip ion o Ai
–msgQs ={I,O} is he se o Ai’s message queues
• I ={. . . , msgi, . . . }is he Inbox, he se o messages sen o Ai
• O ={. . . , msgi, . . . }is he Ou box, he se o messages sen by Ai
•msgi={AgIDs, AgID , pe o ma i e, con en , p io i y}is a message
sen om agen wi h ID =AgIDs o he agen wi h ID =AgID , wi h
he co esponding (FIPA-like) pe o ma i e ype, con en , and p io i y.
–ou Acs is he se o ex e nal ac ions o be execu ed on he en i onmen . I
is composed o uples o he o m: {sende ID, ae}, whe e ID is he sende ’s
ID, and aeis he ac ion ha is being sen .
–Bh ={RG,P}is Ai’s ole beha io
•RG is he se o ole goals associa ed wi h he Bh which Aiis enac ing
•Pis he se o plans Passocia ed wi h he Bh
–Bis he se o Ai’s belie s. I uses he same wo ld s a e s uc u e as E
–Gis he se o Ai’s own goals (see §3.1).
–gc∈(G∪RG) is he cu en commi ed goal (see §3.1).
–Pc={. . . , abi, . . . }is Ai’s cu en plan, which is an o de ed se o ac ion
blocks. Each ac ion block abi={...,aij, . . . }is an o de ed se o ac ions
(each aij is an ac ion). The e a e h ee ypes o ac ions: in e nal ac ions
(ac ions ha a e execu ed by he agen in o de o change hei belie s),
ex e nal ac ions (ac ions ha a e sen by he agen o he con olle in
o de o be execu ed on he en i onmen o al e i ), message ac ions
(ac ions ha a e used o gene a e messages in ended o o he agen s)
–MP is he me aplanne , a lib a y o plans o each goal (see §3.2).
–Pgis he se o p e e ences o e goals (see §3.3).
–Ppis he se o p e e ences o e plans (see §3.4).
Ou concep ual model ex ends he one p esen ed in [9]. Ou ex ensions a e
desc ibed in he ollowing sec ions.
3.1 Adding goal s uc u e
We ex end he concep ual model in [9] by p o iding a o mal model o goals:
wha hey a e, how hey a e de ined, and how hey a e ela ed wi h plans. We
ha e chosen o model goals as desi ed s a es o he wo ld ha agen s s i e o
achie e. I is equi alen o he concep o desi es in BDI. A goal is he e o e
de ined by a collec ion o subse s o he a iables ha desc ibe a s a e o he wo ld
(i s condi ions), and an asse ion o hei desi ed alue(s). These condi ions a e
exp essions such as ‘cash==10’ o ‘speed>=50’ o mean ha ha ing exac ly 10
uni s o cash and ha main aining a speed o 50 o abo e a e pa o he desi ed
s a e o he wo ld, espec i ely. Each subse desc ibes a conjunc ion o a iables
ha desc ibe a desi ed s a e o he wo ld and, in o de o a goal o be conside ed
achie ed, i is equi ed ha all he a iables o a leas one o hese subse ha e
he desi ed alues in he eyes o he agen (i s belie s).
We o mally de ine he s uc u e o a se o goals Gas an uno de ed se
o he o m G={g1, g2, . . . , gn}whe e each giis an indi idual goal among
he many goals an agen has. A goal is de ined as gi={name, desc , C, s a us}
whe e name is a unique iden i ie o he goal, desc is an op ional ex desc ibing
he goal, Cis he se o condi ions o e he s a e o he wo ld o he goal o be
conside ed achie ed, and s a us is a boolean alue ha is T ue i and only i he
condi ions Ca e sa is ied acco ding o he agen ’s cu en belie s B.
Ase o condi ions o e he s a e o he wo ld is de ined as uno de ed
collec ions o asse ions o e he s a e o he wo ld ( he en i onmen ) o he o m
C={a1, a2, . . . , an}whe e ai={n1⋆ 1, n2⋆ 2, . . . , nm⋆ m}is a conjunc ion
o s a emen s o e he alues o a iables o he agen ’s belie s, de ined by ni,
which is he unique name o a a iable o he agen ’s belie s; ⋆, which is a bina y
ope a o ({=,=, >, ≥, <, ≤}); and i, which is he alue o in e es ha is being
asse ed o ni.
The agen possesses he capabili ies o check whe he o no an indi idual
goal has been achie ed acco ding o i s belie s: check goal(gi,B) ou pu s T ue
i , acco ding o he agen ’s belie s, he condi ions o he goal ha e been me ,
and alse o he wise. Ou agen s a e allowed o ha e mul iple goals (own goals
Gand ole goals RG), bu a e es ic ed o pu suing only one a a ime. This
commi men o a goal ha is in ended o be pu sued (gcin he agen uple)
is equi alen o he concep o in en ion in BDI. Agen s ha e he capabili y
o e-conside which goal hey wan o pu sue, and may change he goal hey
a e commi ed o e en i hey ha e no achie ed i , depending on hei cu en
belie s and he s a e o he wo ld hey pe cei e.
3.2 Adding a lib a y o plans
We also ex end [9] o enable speci ying di e en plans o each goal, and o pick
di e en plans o a commi ed goal wi h an elemen ha will ac as a lib a y
o plans, The implemen a ion o he means-ends easone o he pla o m is
a Hie a chical Task Ne wo k (HTN) planne [13]. The HTN is a ee composed
o h ee ypes o nodes: (i) P imi i e Tasks, (ii) Me hods, and (iii) Compound
Tasks. The oo o he HTN is an abs ac compound ask (e.g., o de ood).
Figu e 1 p o ides an example. Ou agen s ha e a lib a y o p ede ined HTN
plans ha he agen can pick om, and hese plans will be ela ed o goals
by means o he s uc u e o he me aplanne , which is he MP elemen o
he agen uple. Fo mally, i can be iewed as MP :G−→ P∗, a ma ching
ela ionship om goals owa ds plans, whe e Pis he se o plans Passocia ed
wi h goal giand P∗is used o indica e ha i can ou pu uples o plans o
a bi a y ca dinali y (meaning one speci ic goal may ha e, o ins ance, h ee
plans associa ed o i , while a di e en goal migh ha e i e, o wo). We need
also o add applicabili y condi ions o plans: P={C, ab1, . . . abn}, whe e Cis
he se o condi ions o e he s a e o he wo ld (see §3.1) ha de e mine a plan
o be applicable, and each abiis an ac ion block.
O he no ewo hy aspec s o he me aplanne a e ha i inco po a es ap-
p op ia e unc ions o plan selec ion. The e o e, i will no simply ac as a
lib a y/collec ion o plans, bu i will also pe o m pa o he easoning. This
easoning includes bo h checking which o he associa ed plans a e a ailable o
applica ion, as well as o de ing hem based on he p e e ences4. Fo he i s unc-
ionali y, he me aplanne ea u es a ge a ailable plans(gi,B) unc ion which,
aking in o accoun he cu en belie s o he agen , i ou pu s a subse o he se
o plans associa ed wi h he goal, con aining only all plans ha a e applicable.
Fo he second unc ionali y, he me aplanne has a pick plans(gi,B, p e sP)
unc ion, whe e p e sPa e he agen ’s p e e ences o e plans, ha will pick
he plan ha is mo e adequa e o he cu en si ua ion acco ding o he agen ’s
p e e ences and belie s, om among all he applicable plans.
3.3 Adding p e e ences o e goals
The nex ex ension we in oduce in he model a e p e e ences o e goals. As
we explained in §2 we d ew inspi a ion om CP-ne s and condi ional p e e ence
o mulas, o some ex en , bu we simpli ied he app oach in o de o be able o
wo k wi hou scala s, ha is, ha ing a ully quali a i e app oach o speci ying
p e e ences o e goals.
To de ine p e e ences o e a se o goals, he app oach we ha e aken is o
es ablish a s ic pa ial o de ela ion be ween hem o indica e which goals mus
be pu sued be o e ying o achie e o he goals. These bina y ela ions be ween
goals a e e lexi e, ansi i e and assyme ic. To model he con ex -dependen
na u e o p e e ences, we allow he decla a ion o condi ional p e e ences, which
a e also a s ic p eo de ela ion o e goals, bu hey only apply when hei
igge condi ions a e me . A nice p ope y o s ic p eo de s is ha hey ha e
always a unique di ec acyclic g aph (DAG) associa ed o hem.
In o de o encode p e e ences o e goals in ou agen s, we ha e added
he ollowing elemen , Pg(which s ands o “P e e ences o e goals”) o he
agen uple. We de ine i as Pg={dGP, cGP1, cGP2, . . . , cGPn}, whe e dGP a e
he de aul p e e ences o e goals ( hey apply unde ‘no mal’ ci cums ances),
and cGPia e condi ional p e e ences o e goals ( hey ha e some igge se o
condi ions Cio e he s a e o he wo ld as de ined in §3.1).
The dGP and each cGPia e de ined as a DAG ha co esponds di ec ly
o a s ic pa ial o de ela ionship be ween goals, and he only di e ence
be ween hem is ha he dGP is he one ac i e by de aul (does no need any
condi ions o be me ), while he a ious cGPibecome ac i e and eplace dGP
i some associa ed condi ions a e ue.
Once all he s ic p eo de ela ions ha e been es ablished, we deduce hei
associa ed DAGs. F om hose DAGs, we compu e a alid opological o de ing o
4We desc ibe how we model p e e ences o e plans in §3.4
each, and hese o de s a e he ones in which goals will be pu sued by he agen s
(by choosing he i s non-achie ed goal in he opological o de ing), e.g.:
–We ha e one agen A′which has he goals G0={g0, g1, g2}.g0is a goal o
idy he agen ’s bed oom, g1is a goal o idy he agen ’s ki chen, and g2is
a goal o s o e clo hes ha a e hanging ou o d y in he open.
–I we deno e “goal imus be achie ed be o e goal j” as gi→gj, he de aul
p e e ences o e goals o agen A′a e {g0→g2, g1→g2}, ha is, be o e
s o ing he clo hes ha a e ou side, A′mus ha e cleaned bo h his bed oom
and his ki chen. No ice how bo h g0and g1mus be accomplished be o e
ocusing on g2, bu he e is no es ablished o de be ween g0and g1, as i is
a s ic pa ial o de . A alid opological o de ing migh be: g0,g1,g2, bu
also g1,g0,g2. By de aul , A′will pu sue his goals in ei he o hose o de s.
–The se o condi ional p e e ences o e goals o agen A′is {g2→g0, g1→
g0}wi h he associa ed igge condi ions ha he a iable ‘ aining’ mus
be T ue. I i is aining, he agen ’s op p io i y goal will be o collec he
clo hes, hen cleaning hei ki chen o bed oom, in no speci ic o de . The e-
o e, he momen i s a s o ain, A′will swi ch o any o he opological
o de ings ha can be gi en o his se ( o ins ance, g2,g1,g0)5.
3.4 Adding p e e ences o e plans and ac ions
By adding p e e ences o e goals we p o ide agen s wi h he capaci y o choose
wha o pu sue. Bu we also need o p o ide hem wi h means o ha e p e e ences
o e how o achie e wha hey a e pu suing. We humans ha e p e e ences no
only o e wha goals we wan o achie e, bu also o e how we wan o achie e
hem, and hese p e e ences may be con ex -dependen . Some people migh p e-
e o d i e o hei wo kplace, while some o he s would a he walk he e. Bu
he p e e ence on walking may change in he case he wea he is e y cold o
ainy, hen p e e ing o commu e o wo k by a combina ion o anspo a ion
modes. In o de o encode p e e ences o e plans and ac ions in ou agen s,
we ha e added elemen Pp(which s ands o “P e e ences o e plans”) o he
agen uple. We de ine i as Pp={gP1, gP2, . . . , gPn}.We deno e he p e e -
ences o e plans o each goal giby gPi={dP P, cP P1, cP P2, . . . , cPPn}, whe e
dP P a e he de aul p e e ences o e plans o goal gi(unde ‘no mal’ ci cum-
s ances), and cP Pia e condi ional p e e ences o e plans o goal gi( hey ha e
some igge se o condi ions Cio e he s a e o he wo ld).
Ap ope y o a goal is he name o a a iable o in e es ha a goal
has he capaci y o al e . Said a iable does no necessa ily ha e o be he
name o a a iable in he se o belie s o an agen . I is simply some hing no e-
wo hy ha achie ing a goal has he capaci y o gi e a speci ic se o alues.
Fo example, i a goal is o ‘o de dinne ’, some o he p ope ies migh be
‘ ege a ian’ and ‘cuisine’, and hei possibles alues migh be {T ue, F alse}
and {‘F ench’, ‘I alian’, ‘Spanish’, ‘Tu kish’}, espec i ely. In ou model each
5In case o con lic s be ween p e e ences, he de aul beha iou is o choose by o de
o decla a ion in he HTN. This can be o e iden by he designe . Re e o §5.
goal, plan, subplan, and ac ion may ha e a se o p ope ies P S, o he o m
P S ={p op1, p op2, . . . , p opn}, and each p ope y p opiis o he o m p opi=
{ 1, 2, . . . , n}whe e: p opiis he unique name/iden i ie o he p ope y, and
iis one o he possible alues ha he p ope y can ake. These alues can be
boolean, nume ic, e c., depending on he na u e o he p ope y i sel . The se
o alues ha make up each p ope y a e used o indica e possible alues he
p ope y can ake. All p ope ies can ha e he special None alue inside he se
o hei possible alues. The p esence o his alue in a p ope y o a plan o
subplan indica es ha said plan o subplan can be achie ed h ough one o mo e
ac ions ha do no use o al e he p ope y in ques ion a all.
P opaga ion o p ope ies consis s in sending he p ope ies ‘upwa ds’
om he mos conc e e ac ions, up o he oo goal, passing h ough e e y sub-
plan and subgoal in he way. The ull desc ip ion o he me hod is p o ided in
[22]. Gi en wo sequen ial ac ions ha ha e he same pa en , he pa en ’s se o
p ope ies will be he esul o compu ing he union be ween he wo child en’s
p ope ies. Each child will no ha e di e en possible alues o he same p op-
e ies, since hey a e sequen ial ac ions, and i would no make sense o design
a plan in which child ac ion no. 1 se s ‘cuisine’=‘Spanish’ only o he child ac-
ion no. 2 o se he cuisine o be ‘F ench’. The e o e, he p ope ies o he wo
(sequen ial) child en will always be di e en , and he esul ing p ope ies o he
pa en node will simply be he joining o he child en’s se s o p ope ies, and i
is i ial o see ha his p ocess applies o nsequen ial child en ac ions.
Gi en wo al e na i e ac ions ha ha e he same pa en , he pa en ’s se
o p ope ies will be he esul o me ging he p ope ies o he child en in he
ollowing manne : i bo h child en se di e en alues o he same p ope y hen,
o he a he , he alues o he p ope y will be he union o he alues ha
he child en had (e.g., i child no. 1 had ‘cuisine’=‘Spanish’ and child no. 2 had
‘cuisine’=‘F ench’, he pa en ask will ha e ‘cuisine’={‘Spanish’, ‘F ench’} o
indica e ha i ha node is chosen, we will limi he possible alues o ‘cuisine’
o hose wo alues). I ei he child has a p ope y ha he o he does no , he
pa en will simply ake he same p ope ies o he child ha has i , and will add
he special alue None, o indica e ha i ha node is chosen, he e is a pa h o
he plan ha accomplishes he goal wi hou e e gi ing a alue o ha p ope y.
Figu e 1 p o ides an example o p ope y p opaga ion. I shows he se o
plans associa ed o a goal o o de ing dinne . The e a e h ee possible op ions:
a plan o o de bu ge s, a plan o o de ala el, and a plan o o de pizza. We
assume ha he e is only a local bu ge , a local ala el, and bo h a local pizza
es au an and a big company ha makes pizza. O he assump ions ha we ake
a e ha all bu ge s and pizzas a e non- egan, and ha all ala els a e ege a ian.
The designe only needs o decla e p ope ies on he ac ions. Then, as a esul
o he p ope y p opaga ion p ocess, all e ices ha e hei own se o p ope ies
ha ha e p opaga ed upwa ds, om he lea es (ac ions). No ice how, in gene al,
all p ope ies ha e p opaga ed owa ds he upwa d nodes. Howe e , mos o hese
p opaga ions ha e been e y simple ones: om single child o pa en , al hough
he e a e wo cases wo h men ioning. The i s one is he p opaga ion om he
Figu e 1. P ope y p opaga ion on an HTN plan assoca ed o he o de dinne goal
subplans o o de local pizza and o de om big pizza company. No ice how hei
p ope ies a e he same in all ields excep o he ‘local’ ield, wi h one holding i
as T ue, and he o he as False. Howe e , hese wo al e na i e subplans sha e a
common pa en , and when hei p ope ies a e p opaga ed o i , hey a e me ged
in he way we desc ibed ea lie : he pa en has i s p ope y ‘local’ wi h all o
i s child en alues, o ep esen ha , i ha subgoal (o i s pa en subplan) is
picked, hen we can s ill o de om ei he a local es au an o a big chain. The
o he no e-wo hy example is he p opaga ion o p ope ies o he oo node,
whe e all op ions ha e been compiled in i s p ope ies.
3.5 Selec ion o plans and ac ions using p ope ies
We will now b ie ly desc ibe he p ocess o choosing a plan aking p e e ences
in o accoun . An assump ion we make h oughou his whole example is ha
all plans a e a ailable, ha is, ou choices a e no es ic ed by he en i onmen
in any way, shape, o o m. Gi en a conc e e goal gi(o de dinne ) an agen
has a se o p e e ences o e he plans o achie e gi. We can de ine his se
as gPi={dP P, cP P1, cP P2}, whe e dP P is he de aul se o p e e ences, and
cP P1, cP P2a e condi ional se s o p e e ences. We assume ha we ha e he
ollowing p e e ences o e how o achie e he goal o o de dinne (see Figu e 1):
Figu e 5. Agen Bob has used his p e e ed means o anspo o when i ains and
his “local business” alues o choose he local pizza op ion.
Pe haps he bigges limi a ion in ou decla a ion o p e e ences o e goals and
plans is ha hey a e absolu e, and his s ems om he ac ha we aimed o
no use numbe s in ou model. The e o e, we canno exp ess hings like ‘I p e e
his a li le mo e han ha ’, o ‘I p e e ha a lo mo e han his’ ha could be
used o sol e con lic s (such as Bob’s con lic s be ween he plan p e e ence and
his alues). Visse ’s e al app oach [22] p o ides a mo e complex s uc u e ha
allows hei agen s o ha e mo e complex p e e ences (e.g., agen s can eason
abou quan i ies, quan i y op imiza ion, limi a ion by quan i y, e c.). Also, hei
agen s a e able o au oma ically ex ac p ope ies o goals by looking a he
ac ions, and hen de i e he ele an p ope ies o he goals. Ou model elies on
he designe ca e ully lis ing (wi hin he scena io desc ip ion ile) he p ope ies
and he p e e ences in he igh o de . In u u e wo k we will explo e mo e
lexible and exp essi e ways o sol e his (wi h no nume ical alues, i possible).
One ela ed issue we plan o in es iga e u he is ela ed o wha o do
when he igge condi ions o non-de aul p e e ences o e goals o e lap (e.g.,
i is snowing and a medical eme gency occu s), especially in he case hey de ine
di e en p eo de s. Ou cu en app oach is o pick he i s goal p eo de (by
decla a ion o de in he scena io ile), and o allow he designe o implemen an
ad-hoc, mo e complex solu ion, i hei scena io equi es so. I would be be e
o modi y ou model o allow o a na i e way o handle his issue.
Finally, ou encoding o mo al alues also o ally elies on he designe ca e-
ully lis ing which ac ions ha e wha mo al implica ions and, while his is good
om an exp essi eness poin o iew (i allows us o decla e mo al ela i ism
as di e en agen s ha ing di e en mo al con ic ions) and con ex -dependen
mo ali y ( he same ac ion ca ied ou unde di e en ci cums ances ha ing di -
e en mo al implica ions), i is a e y exhaus i e and daun ing ask. I would
be good o ha e he sys em pa ly au oma ed, pe haps employing some ma ch-
ing be ween he pu pose o an ac ion and a alue- ee s uc u e oo ed in a
well- ounded model o alues (such as Schwa z’s[19], which is used in [7,12,11]).
Acknowledgemen s
This wo k has been pa ially suppo ed by EU Ho izon 2020 P ojec S ai wAI
(g an ag eemen No. 101017142).
Re e ences
1. Aba , S., Theodo opoulos, G.K., Lema inie , P., O’Ha e, G.M.: Agen based mod-
elling and simula ion ools: A e iew o he s a e-o -a so wa e. Compu e Science
Re iew 24, 13–33 (2017)
2. Belli emine, F., Poggi, A., Rimassa, G.: JADE - A FIPA-complian agen ame-
wo k, pp. 97–108. The P ac ical Applica ion Company L d. (1999), h p://jade.
csel .i /pape s/PAAM.pd
3. Bo dini, R.H., H¨ubne , J.F., Woold idge, M.: P og amming Mul i-Agen Sys ems
in Agen Speak Using Jason (Wiley Se ies in Agen Technology). John Wiley &
Sons, Inc., Hoboken, NJ, USA (2007)
4. Bou ilie , C., B a man, R.I., Domshlak, C., Hoos, H.H., Poole, D.: CP-ne s: A
Tool o Rep esen ing and Reasoning wi h Condi ional Ce e is Pa ibus P e e ence
S a emen s. Jou nal o A i icial In elligence Resea ch 21, 135–191 (2004)
5. B aubach, L., Pokah , A., Lame sdo , W.: Jadex: A BDI easoning engine. In:
Mul i-agen p og amming: Languages, pla o ms and applica ions. pp. 149–174
(2005)
6. Casali, A., Godo, L., Sie a, C.: A g aded bdi agen model o ep esen
and eason abou p e e ences. A i icial In elligence 175(7), 1468–1478 (2011).
h ps://doi.o g/h ps://doi.o g/10.1016/j.a in .2010.12.006
7. C ane ield, S., Winiko , M., Dignum, V., Dignum, F.: No Pizza o You: Value-
based Plan Selec ion in BDI Agen s. In: P oceedings o he 26 h In e na ional Join
Con e ence on A i icial In elligence. p. 178–184. IJCAI’17, AAAI P ess (2017)
8. Das ani, M.: 2APL: A p ac ical agen p og amming language. Au onomous Agen s
and Mul i-Agen Sys ems 16, 214–248 (2008)
9. Gna yshak, D., Oli a-Felipe, L., ´
Al a ez Napagao, S., Padge , J., V´azquez-Salceda,
J., Ga cia-Gasulla, D., Co ´es, U.: Towa ds a goal-o ien ed agen -based simula ion
amewo k o high-pe o mance compu ing. In: A i icial In elligence Resea ch and
De elopmen : P oceedings o he 22nd In e na ional Con e ence o he Ca alan
Associa ion o A i icial In elligence. pp. 329–338. IOS P ess (2019)
10. Hind iks, K.V., Robe i, T.: GOAL as a planning o malism. In: Lec u e No es in
A i icial In elligence. ol. 5774, pp. 29–40 (2009)
11. Kammle , C., Dignum, F., Wije mans, N.: U ilizing he ull po en ial o no ms o
he agen ’s decision p ocess. In: P oceedings o he Social Con e ence 2022, Milan,
12-16 Sep embe (2022)
12. K uelen, K., de B uin, B., Gho bani, A., Mellema, R., Kammle , C., Vanh´ee, L.,
Dignum, V., Dignum, F.: How Cul u e In luences he Managemen o a Pandemic:
A Simula ion o he COVID-19 C isis. Jou nal o A i icial Socie ies and Social
Simula ion 25(3), 1013–1020 (2022)
13. Ku luhan, E., Hendle , J., Nau, D.S.: HTN Planning: Complexi y and Exp essi i y.
In: P oceedings o he AAAI Con e ence on A i icial In elligence, 12. pp. 1123–
1128 (1994)
14. Luke, S., Cio i-Re illa, C., Panai , L., Sulli an, K., Balan, G.C.: MASON: A Mul-
iagen Simula ion En i onmen . SIMULATION: T ansac ions o The Socie y o
Modeling and Simula ion In e na ional 81(7) pp. 517–527 (2005)
15. Nunes, I., Lucena, C., Luck, M.: BDI4JADE: a BDI laye on op o JADE. In:
P oceedings o he 9 h Wo kshop on P og amming Mul iagen Sys ems. pp. 88–
103 (2012)
16. O kin, J.: Th ee S a es and a Plan: The A.I. o F.E.A.R. Game De elope s Con-
e ence 2006 (2006), h ps://alumni.media.mi .edu/~jo kin/gdc2006 _o kin
_je _ ea .pd
17. Padgham, L., Singh, D.: Si ua ional p e e ences o bdi plans. 12 h In e na ional
Con e ence on Au onomous Agen s and Mul iagen Sys ems 2013, AAMAS 2013
2, 1013–1020 (2013)
18. Rousse , A., He mann, B., Lang, C., Philippe, L.: A su ey on pa allel and dis-
ibu ed mul i-agen sys ems o high pe o mance compu ing. Compu e Science
Re iew 22, 27–46 (2016)
19. Schwa z, S.H.: An o e iew o he schwa z heo y o basic alues. Online Readings
in Psychology and Cul u e 2(12 2012)
20. Tejedo , E., Bece a, Y., Aloma , G., Que al , A., Badia, R.M., To es, J., Co es,
T., Laba a, J.: Pycompss: Pa allel compu a ional wo k lows in py hon. In e na-
ional Jou nal o High Pe o mance Compu ing Applica ions 31, 66–82 (1 2017)
21. Tisue, S., Wilensky, U.: Ne Logo: A Simple En i onmen o Modeling Complexi y.
In: In e na ional Con e ence on Complex Sys ems, Vol. 21. pp. 16–21 (2004)
22. Visse , S., Thanga ajah, J., Ha land, J., Dignum, F.: P e e ence-based easoning
in BDI agen sys ems. Au onomous Agen s and Mul i-Agen Sys ems 30, 291–330
(3 2016)
23. Winiko , M., Sido enko, G., Dignum, V., Dignum, F.: Why bad co ee? explain-
ing bdi agen beha iou wi h aluings. A i icial In elligence 300, 103554 (2021).
h ps://doi.o g/h ps://doi.o g/10.1016/j.a in .2021.103554
24. Zia, K., Riene , A., Fa ahi, K., Fe scha, A.: A New Oppo uni y o U ban E ac-
ua ion Analysis: Ve y La ge Scale Simula ions o Social Agen Sys ems in Repas
HPC. In: 2012 ACM/IEEE/SCS 26 h Wo kshop on P inciples o Ad anced and
Dis ibu ed Simula ion. pp. 233–242 (2012). h ps://doi.o g/10.1109/PADS.2012.4
