From: AIPS 1994 Proceedings. Copyright © 1994, AAAI (www.aaai.org). All rights reserved.
Opportunities: A unifying frameworkfor planning and execution
Louise Pryor
School of ComputerScience
The University of Birmingham
Edgbaston
BirminghamBI5 2TY
UK
l.m.pryot~cs.bham
~tc.uk
Gregg Collins
TheInstitute for the LearningSciences
Northwestern
University
1890 MapleAvenue
EvanstonIL 60201
USA
collins@ils.nwu.edu
Abstract
Asuccessful agent in the real world mustboth plan
aheadandreact to the unexpected.Ideally, both processes should be carried out in a common
framework.
In this paper wedescribe su~a framework
basedon the
analysis of opportunities. Weargue that planningin
advancecan be viewedas a matter of anticipating opponanities, whilerespondingto the unexpectedshould
be seen as reacting to opportunitieswhenthey arise.
Wepresent an opportunistic planning agent, PARETO,
that operates in a simulatedrobot delivery world,and
implementsour approach.
1.1 Planning in an unpredictable world
Traditional AI planningsystems(Fikes and Nilsson 1971:
Sacerdoti 1977; Chapman
1987), knownas classical planners (Wilkins1988), haveeffectively decoupledplan constmctionand plan execution, operating on the assumption
that all neededinformationwill be freely availablein advance. Morespecifically, these systemsrely on three assumptionsaboutthe worldsin whichthey operate:
¯ Simplicity: it is possible to knoweverythingabout the
worldthat mightaffect the agent’sactions.
¯ Stasis: there will be no clumges
in the worldexceptthose
causedby the agent’sactions.
¯ Certainty:the agent’sactionshavedeterministicresults.
1. Introduction
Werethese assumptionsvalid, the world wouldhold no
Thereal worldis regular enoughto makeadvanceplanning surwisesfor the agent;henceplanscouldbe specif’w.din exworthwhile,yet unpredictableenoughto makeplanningto
act detail with no fear that an unexpectedoutcomewould
the last detail impossible.Anautonomous
agentmustthere- force subsequentrethinking.Of com~e,there are fewnatural
fore strike a balancebetweenplanningaheadandreacting to worldsin whichthe classical assumptionshold up. To take
changes.For example,a robot on a strange planet mustde- a simpleexample,consideran everydayhuman
activity like
terminewhichareas to explore,whereto take soil samples, preparingbreakfast. For mostpeople,this occursdayafter
whatroutes to take, andso on. Theright chokeswill depend day in the sameenvironment,at the sametime, using the
on details concerningthe terrain encountered,the atmo- sameingredients,andso on. It is an eventregular andpresphericconditions,andthe results of tests performed
on ear- dictable enoughto be described by a script (Schankand
lier samples--factors
that cannot,in general,be predictedin
Abelson1977). However,the apparent regularity of the
sufficient detail to allowfirm decisionsto be madein ad- breakfast worldis only an artifact of our loftily abstract
vance. Onthe other hand, undirectedwanderingmakeslittle
point of view;downat the level of detail at whichwemust
sense: enoughwill be knownin advanceto makesomedeci- treat the domainin orderto executea plan successfully,we
sions that will makea productivemissionmorelikely. The find a worldof wildandcapfi¢ionsunpredictability.Theplan
distinction is a matterof availableinformation:someof the
for makingbreakfast mayinvolve any numberof actions
informal/onthat wouldbe required to construct aa optimal suchas grasping,lifting, andpouringcontainersof milk,ceplan will not be availabkbeforephmexecutionbegins.
real, coffee, andso on; dishes andutensils mastbe manipuSince relevant informationmaybecomeavailable at any lated; appliancesmustbe operated;obstacles on the floor
point, evenwhilea plan is beingexecuted,an agentmustbe mustbe circumnavigated.
Toliterally makea completeplan
preparedto alter its plansto refle~newinformation.Ideally, in advance,it wouldbe necessaryto knowthe exact posithis replanningprocess will resembletbe wocessof plan- tion, orientation, andweightof everyrelevantobjectin the
ning in advanceas muchas possible, so that the same kitchen. It wouldbe necessaryto know,for example,the
knowledge
andprocessescen be applkd.In other we~ls,iris
preciseangleat whichthe boxof cornflakes shouldbe tilted
desirablethat planningandreplanmng
can be carried out in a to achieveoptimalflowinto the bowl,the properposition
commonframework. In this paper we propose such a
andaltitude of the milkcontainerduringpouringto miniframework,
basedon the notion of respondingto opportuni- mizesplashingcausedbythe flakes, andso on.
ties. Wediscuss the implementationof this approachin
Clearly.this is completelyunrealistic. Evenassuming
the
1 a system that notices and responds to opportuniPARETO,
theoretical possibility of gatheringsuchinformationin adties as it pursuesits goals(Pryor1994).
vance,the cost of acquiring,storing, andprocessingsucha
1 Plann~ and Acting in Realistic Envimnmeam
by Thinking about
Opportunities.Vilfredo Psreto (1848-1923)wasmIlslim eccaemist,so-
ciologht, and phiimepherbest knownfor the notion of Parelo opt~lity
mdthePartt°distributi°n’neither°fwhichbtmalindtisw°rk"
PRYOR
329
From: AIPS 1994 Proceedings. Copyright © 1994, AAAI (www.aaai.org). All rights reserved.
quantity of information about the world in general is prohibitive. Furthermore,given inaccuracies in sensors and the
interference of other agents, muchrelevant information is
likely to be unavailable in principle. In other words, the
breakfast world, like most natural environments, displays
the followingcharacteristics:
¯ Complexity: it is impossible to knoweverything.
¯ Dynamism:changes occur as a result of the actions of
other agents or of natural phenomena.
¯ Uncertainty: the agent cannot be sure what the results of
its actions will be.
Confrontedwith an environmentthat displays these characteristics, the classical planningparadigmbreaks down.There
are three key reasons for this. First, inaccurate information
maycause wrongdecisions to be madeduring the construction of a plan. For example, believing that there are clean
bowls in the cupboard, you might construct a plan that entails openingthe cupboarddoor. If you are wrongabout the
bowls, this step is unnecessary.The possibility of a faulty
plan implies that agents must monitor their plans during
executionand be preparedto recover fromfailures.
Second, information needed to make somedecisions may
not be available at the time a plan is chosen. For example,
you cannot accurately predict the movementsof your roommate, whichmeansthat there is potential for interference between your plan and your roommate’sactions. In order to
make an optimal decision on what path to take across the
kitchen, you must knowwhere your roommatewill be, or at
least knowthat she is out of the way. In general, decisions
that fall into this category should be deferred until enough
informationis available; since the infonnatiou will in many
cases not be available until after executionof the plan has
begun, the agent must prepared to interleave plan construction and plan execution.
Third, decisions mayarise that have not been foreseen in
the planning process. For example,the telephone might ring
while you are pouringa glass of orungejuice, forcing you to
decide whether to continue pouring or stop and unswerthe
phone. The agent mast recognizecircomstances under which
the need to maken,-t unforeseendecision arises, and must, if
necessary, be able to acquire the information neededto make
those decisions. The agent must be able to changeits plans
duringtheir executionto reflect anforeseensituations.
In sum, the inevitability of the unexpected means that
plans made in advance will require modification during
execution. Expendingeffort on the construction of elabcmae
and detailed plans is therefore often unpmdoctive.A more
effective apwmch
in the face of unlmxlictability is to expend
some effort on choosing simple plans, and to expend more
effort on adapting those plans as unforeseen circumstances
are encountered. This is the aPim3a~followed in PARETO.
1.2 Plan execution
The emphasisin the design of PARETO
is on recognizing the
need for and makingunforeseen decisions during plan execution. As an example of the kind of reasoning PARETO
is
meant to perform, suppose you happento see a sharp knife
as you a~ looking for a pair of scissors with whichto cut
330
POSTERS
string. In such a situation, unless there weresomeclear reason not to do so. you might well use the knife to achieve
your goal and abandonthe plan to find scissors. PARETO
recognizesand takes advanlageof such opportunities.
Instead of reasoning in detail about the interaction betweenplans for its various goals, as a classical planner
would, PARETO
constructs separate plans for each of its
goals and does not expandeffort attemptingto anticipate
tential interactions in advance.In addition, instead of expending a great deal of effort to gather all available information
at planning time, PARETO
depends on possibly faulty assumptionsabout the situation in whichit will execute its
plans. Obviously,the failure of these assumptionscan cause
PARETO’s
plans to fail. PARETO
is designedto react quickly
and flexibly to unexpectedcircumstances,rather than to minimizethe possibility of an uncertaintyarising.
2. Planning
and opportunities
Thecost of achievinga goal and the benefits of doingso can
vary wildly over the goal’s lifetime. For example,consider a
goal to buy gas for your car. Duringa rush to makean imporlant meeting, the cost of stopping for gas wouldbe very
high, since it is likely to makeyou late for the meeting.
The benefit--essentially the reduced probability of running
out of gas--is also high. but is likely outweighedby the
cost. During the meeting, the cost of pursuing the goal is
still higher, since walkingout of the meetingand driving off
to buy gas wouldbe a most undesirable course of action: the
benefit does not change. Onthe drive homeafter work, the
cost of buying gas is relatively low--assumingyou have no
urgent plans--while the benefit increases as the likelihood
of running out of gns becomesprogressively more acute. At
home,the cost of buyinggas is again high. as it is nowinconvenient to make a special journey, while the benefit
stays the same.Cost and benefit continue to vary over time.
dependingon the exact situation in whichyoufind yourself.
until youdecide to achievethe goal.
An effective planner must thus not only find workable
plans to achieveits goal, but must also, insofar as possible,
maximizethe benefit and minimizethe cost of doing so. In
short, it must wait until there is a good opportunity for
achieving a given goaLPlanning can be seen as the process
of predicting whenopportunities will arise and what form
they will take, and deciding in advanceto take advantageof
them. Althoughit is imwacticalto performdetailed predictions for all possible circumsUmces,
it is often possible to
performsimplifgd iavdk:tious. For example,it is routine to
plan whento refuel your car based on your knowledgeof the
amountremainingin the tank, the locations of gas stations,
and your travel plans over the next day or so.
2.1 Adapting the current plan
Aplan that is designedto achievea particular goal should be
revised whena predicted opportunity does not in fact arise,
or whena better opportunity comesalong that was not considereal during the planningprocess. Returningto an earlier
example,if youwere to receive a call on your car phoneinforming you that your meeting had been postponed, you
From: AIPS 1994 Proceedings. Copyright © 1994, AAAI (www.aaai.org). All rights reserved.
would be presented with an unexpected opportunity to buy
gas. In general an agent should notice such unexpectedopportunities and consider adopting new plans to take advantage of them. As far as possible, the agent should respond to
the unexpectedopportunity in the sameway as it wouldrespendto the oppcmmfity
had it been predicted in advance.
The paradigmof noticing and responding to opportunities
thus provides a unifying framework within which to approachboth the issue of planningand the issue of plan revision during execution. In this approach,an agent’s response
to an opportunity should be independent of whether it has
beenpredicted
ornot.Planning
inadvance
is:based on the
prediction
offurore
opportunities,
whileplanrevision
is
based
ontherecognition
ofcurrent
opportunities.
2.2 Switching between plans
PARETO’s
approach means that it is pursuing a number of
independent plans at any given time. The managementof
these diverse plans is thus a critical issue in PARETO’s
design. To managethem successfully, PARETO
must distinguish those plans that are being actively pursued from those
that arenot. In fact, it can generally be assumedthat only a
handful of the agent’s current plans will be pursued actively
at any given time. For example,consider a plan to follow a
recipe that says "soak the beans overnight." Clearly, pursuing this plan actively:once the beans are put in to soak--for
example, by sitting and watching them until morning-wouldbe a tremendouslyinefficient plan. Instead, the agent
should suspendthe executionof this plan, and turn its attention to the pursmtof other goals. At any one time, most of
an agent’s goals are suspended,tn order to manipulateplans
in this way, PARETO
must incoqxm~ general mechanisms
for deciding whenit should changefrom following the plan
for one goal to followingthe plan for another.
In general, a goodtime to attend to a particular goal is
whenthere is an oppornmi~for that goal, whether wedicted
or unpredicted. For example, you should return to the soaking beans whenyou are in a position to Wocee___d
with the
next step in the recipe. The existence of such an opportunity-to perform the next step in the preparation of the
beans---waspredicted in theconstruetion of the overall plan.
However,an unpredicted opportunity maysimilarly trigger a
change of attention from one goal to another. For example,
supposeyou have tried and failed to get hold of a friend on
the telephone to makesome arrangements
withher,
if you
see her in the supermarket, you maywell temporarily suspend your goal of {k)ing ymgr weekly shopping to pursue
your goal of making the arr~menm.
Thus, managementof plans is handled naturally within
PARETO’sparadigm of responding to opportunities.
Decisions about changing, plans, or reactivating suspended
plans, are based on the recognition of current opportunities,
while the plans themselvesare based on the prediction of future opportunities.
3, Pareto
In this section we describe PARETO,
a working system that
illustrates howthe frameworkdescribed in this paper is an
Fibre
I , PARETO’s w,orld
effective means, of combining
the execution ofplans with
apI~wiate responses to unexpectedsituations.
3.1 What PARETO does
PARETO
operates a simulated robot delivery truck. The simulator was built using TRUCKWORLD
(Firby and Hanks
1987; Hankset al. 1993). In a TRUCKWORLD
world, a robot
delivery truck travels between locations on a network of
roads, encountering and manipulating various objects as it
goes. In PARETO’s
world (described in detail in Pryor,
1994), there are several building sites whoseworkersuse the
truck to run delivery errands such as "fetch a hammer,"or
"fetch somethingto carry mytools in."
PARETO’s
world consists of a numberof locations linked
by roads along which the truck can travel (see figure 1).
Threeof the locations are building sites, one is the truck’s
base whereit can usually find feel, and the other locations
contain objects that the truck uses to fulfill its delivery
goals. Mostof these objects are used regularly by the construction workers whoseerrands the truck runs: hammers,
saws, ladders, paint, and so on. There are over 30 different
types of object in PARETO’s
world, of which20 are used for
deliveries. At any moment,there are typically well over 100
different objects at the various locations amtmdthe world.
PARETO’s
world is unpredictable: The truck has limited
perception, and can sense only those objects that are at its
current location, meaningthat from PARETO’s
perspective
the world is complex.It is also dynamic, since objects may
spontaneously change location, appear, or disappear.
Finally, the results of the truck’s actions are uncertain: it
maydrop objects that it is trying to grasp, and neither the
time taken to travel between locations nor the amountof
fuel used canbe predicted.
P~ receives defivery orders at unpredictable intervals
during its operation, with a typical run involving between
sevenand twelve separate deliveries. Plans that allow for every possible combination of goals would be far too complex; instead it uses a separate plan for each of its goals.
These plans are sketchy--they do not specify in detail every
action that should be performed. Instead, they specify the
overall strategy that should be used in terms of a few simple
steps, and PARETO
decides how each step should be performedas it executes the plans.
PRY’OR 331
From: AIPS 1994 Proceedings. Copyright © 1994, AAAI (www.aaai.org). All rights reserved.
PARETO’s
sketchy plans allow for someof the manycontingencies that mayarise (for example,they specify what to
do whenthe object being grasped is dropped) but manycircumstances cannot be f~seen in the plans. There are two
types of situation in which PARETO
must respond on the fly
to circumstances that it encounters. First, circumstances
may dicmte that PARETO
should switch its attention from
one goal to another. For example, PARETO
will not con-
Remove
tuk kern
out of fuel, but will instead concentrateon trying to refill its
fuel tank. 2 Second, an unforeseen opportunity
mayarise.
This mayentail either switching plans, or replacing a current plan with an alternative. For example, suppose the
truck has two delivery goals, one for something to carry
tools and one for something to cut twine. PARETO
maydecide to pursue the carry-tools goal fh-st’ and set off to the
warehousein which it expects to fund a box. If the truck
fnds a knife or a pair of scissors on the way, however,
PARETO
will temporarily switch its attention and pick up
the cutting tool. If PARETO
subsequendyencounters a bag
that wouldbe suitable to carry tools, it wouldabandonits
plan to find a box and instead pick up and deliver the bag.
PARETO
has an efficient mechanismfor spotting unpredic(ed opportunidessuch as these, and u’eats themin exacdy
the sameway as it does opportunities that have been predicted in its plans. The next sections explain PARETO’s
basic operationand characterization of plans and opporumities.
3.2 How PARETO works
3 plan execution system
PARETO
is based on Fuby’s RAPs
(F’trby 1987; 1989), and is deacrihed in (Pryor 1994).
PARETO
acquires a newgoal, it looks in its libraryof SAPs
(sketchy
plans)
foronedmtwillachieve
thegoalThesteps
in a RAPspecify subgoals that the system must achieve in
order to execute the plan successfully. PARETO
recursively
expands sketchy plans by choosing aplna for each subgoal.
Eventually, a subgoal will be achievable by performing a
simple action, and no further expansionis required.
When
PARI~TO
_ro~__~!_’ves
a goal, its lust action is to place a
task aimed at achieving that goal on the task agenda. The
task that is placed ou the task agendaconsists of the goal
and the RAPthat has been chosen to achieve it. PARETO°s
execution cycle is summarizedin figure 2, and consists of
the followingsteps:
¯ Cheosinga task from the agenda. The mkthat will be the
i.._
__ ~ task aoenm/
tinne tryingto makea deliverywhenthe truckis running
Figure2
PARETO’s
execution cycle
achieve its goal. For example,
a goalto find fuel might be
achieved by going to the location of a fuel drumthat
PARETO
knows about, by going to the base which is a
source location for fuel, or by wanderingaroundthe world
until a fuel drumis found. Eachsuch plan is a methodof
achieving the goal. PARETO
chooses one of the methods
of the task that is being l~rucessed, based on the state of
the worldat the time that the processing
takesplace.
¯ Addingnew tasks to the m~agenda for each of newgoals
created duringthe previousprocessingstep.
¯ Repmcessingthe original task when each of its submsks
has achievedits goal. The suecessful executionof the subasks is not enoughto guaranteethat the original task will
itself haveachieved
its goal,sincePAR~rO’s
worldis dy-
namicand sometime maypass betweenthe execution of a
task’s sublasks and the repeat lm3cessingof the original
rusk. For example,the truck mightsucceed
ingoing to the
location of a knownfuel drum, but the drummight have
meanwhiledisappeared, or the fuel in it been used by another agent. If the task has succeeded, it is removedfrom
the task agenda,else it ls Wecessedas descrihedabove.
To Wusimtethe execution cycle in action, consider what
happens whenPARETO
receives the two goals in the example above, to deliver somelhingto carry tools and something
to cut twine. As each goal is received, the deliver-object
RAPis chosenmachieve it and the relevant task is placed on
the task agenda. ’I~ deliver-obje~ RAPhas four steps:
PARETO
must find a su/table object, load it, travel to the
correct location, and unload the objecL After PARETO
has
processed
the deliver-object m~kforthecarry-tools goal,
a~ thuSfive tasks on the agendafor dmt gel: delivermost productive in ~ PAitLrrO’sgoals should he object, find-ohject, Ioad-payload-ol~ect,
truck4ravel-to,and
chosen. PARETO’s
task selector Ope~lt~ by looking for
unlond-at.Of these, the deliver-objecttaskis waitingfor the
opporl~mitlesto further its various goals. The ability to
other four to complete, and the Ioad-payload-obiect,
truckrecognize un~ oppommitles is a significunt
chan~
travel-to, and unload-at tasks are waiting for their
from the task selector used by Firby’s aAPssystem.
predecessm,sto complete. If PARETO
has no other goals, the
¯ Processingof the chosentask to fill in the demill of rite
next task to he chosenwill be the find-object task, which
incompletelyspecified plan that describes the task. A RAP will in turn be expanded
andits subgoalsplacedon ~ task
specifies all the different plans that might be used to
agmda.
If all goesaccordingto plan. all the sulxaskswill be
proc~ in ttwn and removedfrom the wsk alp~aKkuntil the
unload-at task has been achinved. Finally, PARETO
will
2 Aswell
as,,-,/very
Soals,
PAiteTO
has lm~envation
iio, I- (Schmk
md againprocessthe deliver-objecttask, fred that it has sucAbebon
1977)to mum
du~the mu:kdoenet nmouzd fuel mKIto keep c__~eded,
undremove
it fromtlz ruskaSenda.
m~etim mnoundinp.
31temive Aaion Padmlm
332 POSTEP~S
From: AIPS 1994 Proceedings. Copyright © 1994, AAAI (www.aaai.org). All rights reserved.
PARETO
thus characterizes opportunitiesas tasks on its
--> New order U~ER-5: sor~chtr~j to CARRY~ for ILS
--> New top level ~oal= ~LIV~-OB~CT:=[S:6]
task a~endathat are easyto achieve,h has an efficient mech--> New order ~-4: ~ng to CUT ".WINg for
~mismfor recognizing ~ties that uses a f’dtering pro--> ~ top level
~al:
~TV~-~ ::[7:7]
cess basedonthe functionalcharacteristicsof objectsin the
¯ .. ProcessL~g ta~-<rs~.~VER-CB3E~ ~-IU~S 113 (HI~-5>::
[6:6]
world (Pryor 1994). There are two waysin whichone
...
Prooessing
task~ (It~ff-~S
_ >=:[6:5]
PARETO’s
tasks maybe easily achievable,correspondingto
... ~he truck ~.s ~ff to £1nd a Jx~ ...
+~-~ Potential opp=~JJ~t¥ for
the twotypesof opportunitieswediscussedearlier:, either it
<DELIV~ (2~qff-TOC~
TI.S (lq~9>
has alreadysucceeded,
or it is readyfix processing(notwaitf~zn ITJ~I-20 ~AG)
ing for anyothers to be completed).Tasksthat havealready
~++ Po~enti~l
Opport~C¥
for ~ CPaSff-~OLS
_ >
fra. I~M-20 {Bl~;)
succeededmayindicate the presenceof an unexpected
oppor+4+ Potential
~.~r~,.rd.ty
for
tunity, whilea task that is readyfor processingmayindicate
<L(I%D-PA~
~ CRRRY-T(X)LS>
the presenceof a predictedopportunity--onethat has been
from I~-20 (BAG)
+++ Potentl~oRoortta~ty for
foreseenin a plan. In our example,the findingof a bagis
<DEIXV~-4:B~
~ ~ (1~1~-8>
unexpected:PAI~TO
had plannedto find abox.
frum IT~-I3(SCISSORS)
PARErO
thus characterizesopportunitiesin termsof easHas already
~ <~ND-Ci3J~
~-~(~S..
*** Taking tme~oecTJed ~.ta’~.ty:
[6:5]
ily achievabletasks on its task agenda.Expected
opportuni<FI1qD-G~ECTC~J~-~3XS.>
ties are associatedwithtasks that representthe nextstep in
... Pro~essir~ task- <lvDl:~]Bb~ Ciq~/-~q _ >: : [6:5]
... Pro~essin~task:
the plan for oneof its goals, andunexpectedopportunities
<IDAD-PAYI£1%D-CS..~C~
1"1vn4-20 C~R~-~(~S>::[6:8]
a’e associatedwithtasks that havealreadysucceeded.
... cbe cna:k loadsthebag ...
...
Task
~;
~PA~.J~D-cnJ~’T
~20 (~qRRY-~0(~S>::
[6:8]
4. Related work
*** Taking ~ ~I.l~k%~tiCy:
[7:7]
An
agent
should
revise
its plans whenit encountersan un<D~LIV~
~ T~R
U~-8>
"**
~ng ~oals
~ 6 Co 7
expectedopportumty,
but cannotafford to analyzeeverysit¯.. ProcessiJ~ task:
uation exhaustively.Thecentrality of opportunitiesto the
<[~IV~-Ci~
C~f~Z~ ~ Ue3~-8>: : [7:7]
executionof plansin an unwedictable
worldhas receivedlit... Prc~essir~.
task: <FI"I~)-CEk]~C~ ~ _ >: : [7:5]
¯..
Task
~:
tle arm,ionbyother researchersin the field.
<~I~D~ CIE-TR11~ => ~EH3J~P~-2 ITEM-13>::[7:5]
Mostcurrent resem~hon the problemof recognizingthe
¯.. Pr~esstng task:
<IQN>-PAYICN)-CB~CT
I~M-13CIE-TWI~>::[7:8]
needto makeunforeseendecisionsfails to addressthe issue
...the tzuc*loadsthe sc/ssors
...
of opportunismexplicitly 08resins and Drummond
1990;
¯.. Task w~w~___~
Fe~guson 1992; Lyons and Hendriks 1992; McDermott
<IEN)-P&~
I~(-13
~: : [7:8]
*** Changing~la from 7 no 6
1992).In general,this workrelies heavilyon projectingthe
... Processingtask: <~C~-~AVEL-TO
~LE-AVE>:: [6:5]
agent’s
current plans to detmninewhenreplanningwouldbe
... the rzu~c g~es off Co deliver
~hs haft...
desirable. Asprojection mayinvolve arbitrarily complex
reasoning, this approachfails to address the problemof
wx~3gnizing
the needto makedecisionsquicklyenoughthat
3.3 Opportunities in PARETO
the agentcan respondappropriatelyin a dynamic
world.
PARETO
chooseswhichtask to Ixocess next by considering
Theeurlicst workon opportunityrecognition, by Hayesthose tasks that are associated with opportunities. When Rothand Hayes-Roth
(1979), lookedat opportunism
in plan
PARETO
spots aft opportunity,it doesnot cousiderwbethef cons~, but did not consider plan execution.
or not the opportune
task is. the nextstep in the planfor the
Hammond
and his colleagues (Hammond
et al. 1993) pre.
currentgoal, Thus,it mayin effect ignorethe existingplan sent a methodof oppoNmfity
recognitionbasedon reoognizfor carryingout a ruskwhenan opportunityarises; if it suc- inK the features involvedin a goal’s achievement.
This apceedsin taking advantageof the oplmmmity,
the task andits
proachrelies on havingspecified the plan for the goal in
the previous plan are simply removedfrom the agenda. enoughdetail that the environmental
elementsinvolvedare
Furthermore, PARETO
is not co~i~ned by any notion of
already known.It does not allow aa agentto recognizeopthe"curre.t" _.m_~ in ~ ~x~m--~us. it can efportunities for goals that it has not yet decidedhowto
fectively switchplans wbetwv~
aa oppommity
arises.
achieve,anddoesnot allowthe recognitionof opportunities
In our example,PARW/’O
changesits plan for its current that require a diffe~nt metlxxiof achievement
fromthat in
goal by ¢k~idingto pick up tha.~ instead of continningto
the current plan. For example,Hammond’s
approachwould
lookfor a Uox.It also switchesits attention to anothorgoal not allow an agent to recognize the ~ity discussed
by picking up the ~’im~. Howeveg,the ~ switch is
above,in whichthe presenceof a bag allows an agent to
only temporary,as pickiagup the scissors does not achieve abandonits plan to find a box. This limited viewof oppof
the task of findingan objectthat cancarry tools, whichtask tunity recognitionwouldInvent the use of oppommities
as
remains on the agenda. After picking up the scissors
a framework
for combining
planningand responsiveness.
PARETO
returns to the c.arry-tools goal and goes off to
Theimpracticalityof unlimitedreplunninghas long bee.
deliver the bag. Figure 3 showsPARETO’s
output as it
recognizedas a serious problem
in the designof intelligent
~cognizesand rakes advantageof these oppommities.
agents. Thereale twoaspectsto the problem:the necessity
of limiting the amountof reasoningthat is performed,and
Fig,m~.. 3 ~.pARW£O
takes at~vauta~of twoopportunities
PRYOR
333
From: AIPS
Proceedings.
Copyright ©when
1994,this
AAAIlimited
(www.aaai.org).
All rightsBreslau,
reserved.J.
the 1994
necessity
of determining
reasoning
ImdM.Drummond.
1990. "Integrating Planningand
Reaction: A preliminary report." In Workingnotes of the
Spring
Symposium
onPlanning
inUncertain,
Unpredictable,
or
Changing
£nyirommen~,
Stanford,
CA,AAAI.
Chapman,
D. 1987. Planningfor ConjunctiveGoals.
Artificial
Intelligence 32 : 333-337.
Dem~T. and M. Boddy.1988. "Ananalysis of tlme-dependent
planning." In Proceedingsof the SeventhNational Conference
on Art~cial Intelligence, St Paul, MN,AAAI.
Ferguson,I. A. 1992. TouringMachines:An architecture for
dynamic, rational, mobile agents. ComputerLaboratory,
University of Cambridge.TechnicalReportNo. 273.
likes, R. E. and N. J. Nilsmm.1971. STRIPS:A newapproach
mthe application of theorem proving to problem solving.
Artificial Intelligence2 : 189-208.
Firby, R. J. 1987. "Aninvestigation into reactive planningin
complex domains." In Proceedings of the Sixth National
Conferenceon ArtOqcialIntelligence, Seattle, WA,AAAI.
Firby,
R. J.1989.Adaptive
execution
in complex
dynamic
worlds.
Department
of Computer
Science,
YaleUniversity.
Technical
Report
YALEUK~SD/RR
#672.
Firby, R. J. and S. Hanks. 1987. The simulator manual.
Departmentof ComputerScience, Yale University. Technical
Report YALEU~SD/RR
#563.
Hammond,
K., T. Converse, M. Marks, and C. Scifert. 1993.
Oppununiam
wAlearnin8. Machin~
Leorn/ng
10 : 279-309.
Hanks. S., M. E. Pollack, and P. R. Cohen. 1993.
"Benchmarks,testbeds, controlled experimentation, and the
design of agent architectures."A/Magazine,
Winte~1993.
Hanks, S. J. 1990. Projecting plans for uncertain worlds.
Departmentof ComputerScience, Yale University. Technical
ReportYALEUK~SD/RR
if056.
Haym-Roth,
B,1990."Dynamic
control
planning
in intelligentagents." In Workingnotes of the Spring Sympesiwnon
Planning in Uncertain, Unpedictable or Changing
F.nvironments,Stanford
University..
Htyes-Roth,
B.andF,Hayes-Roth.
1979.
A cognitive
model
of
plmmin
$. Conclusion
8. Cognitive Science 3 (4): 275-310.
Lynn, D. M. and A. J. Hendrila. 1992. "A practical approach
To be successful, an agent in the real world must both plan
to integrating reaction and deliberation." In Proceedingsof the
ahead and react to the unexpected. Ideally, both pmeeues First International Conferenceon Artificial Intelligence
should be carried out in a common
framework.In this paper
Planning Sy~n&Collage Park, Maryland, MorganKaufmann.
we have described such a frameworkbased on the analysis of
Maes,P. 1991. "Adaptiveaction selection." In Proceedingsof
opportunities, and a computer program, PARETO,
that imthe Thirteenth Annual Coherence of the Cogmtive Sconce
plements our approach.
Society, CI6cago,IL, LawronceErlbaumAssociates.
McDermott,
D. 1992.Transformational
planning
ofreactive
behavior. Yale University, Departmentof Computm
Science.
Acknowledgements
YALEUK~SD/RR
#941.
Mostof this workwas carried out while the first author was Pryor, L. 1994. "Opportunities and Planning in an
at the Institute for the Learning Sciences. This work was
Unpt~UctableWorkl." PhDdiu~ation, in preparation, The
supported in part by the AFOSRunder grant number
Institute
for the LearningSciences,Northwestern
University.
Secm’doti. E. 1977. A sir,crate for plan, and behavior. New
AFOSR-91-0341-DEF,and by DARPA,monitored by the
York: Ammicen
Elsevier.
AFOSR
under contract F496?J)-88-C.0058.The Institute for
Schsnk. R. C. and R. P. Abelson.1977. Scripts, plana, goals
the Learning Sciences was established in 1989 with the
and amm~,rmamd~r.
Hillsdele. NI: LawrenceErltmum
Associates.
support of Andersen Consulting, part of The Arthur
Wellmen,M. P. 198g. "Qualitative Pmbabilisfic Netwmks
for
Andersen WorldwideOq~mization. The Institute receives
Planning
under
Uncertainty."
In
Uncertainty
in
Artificial
additional support
fromAmerimch
andNorth WestWater, Intelligence, ed. J. F. Lemmer
and L. N. Kanal.2. Amsterdam:
Institute
Partners,
andfromIBM.
Elsevier.
Wellmm~
M. P. 1990.Pormslation
of Tradeoffs
in Planning
References
Uncertainly. LondomPitm~
Wilkiu, D. E. 1988. Practical
Plam6ng: F.xtendin|
the
Boddy, M. and T. Dean. 1989. "Solving time-dependent plmClassical
A!
Planning
Paradigm.
San
Marco,
CA:
Morgan
ning problems."In Proceedingsof the Eleventh International
Joint Conference
on Artificial Intelligence,
Detroit,
MI,AAAI. Kaufmmm.
should occur. Traditional AI ~bes have concenlrated on
limiting the amountof reasoning that is performed, by using either quantitativeapproximations
(Hanks1990) ~ qualitative techniques (Wellman1988; 1990). Anytime algorithms (Dean and Boddy1988; Beddyand Dean1989) address
the issues of designing reasoning algorithms that will pro.
duce an answerin limited time. Noneof this work addresses
the issue of whenthis reasoningshould be performed.
In Hayes-Roth’s GUARDIAN
system (Hayes-Roth 1990)
global control plans are used to direct the agent’s reasoning
towards important goals. These conlrol
plans
arechanged
by
global control decisions, whichappear to be trigget~ by the
receipt of sensory information. However,Hayes-Rothgives
no details of howthese decisions are Iriggered or the process
by which they are made. Presumably GUARDIAN’s
mechanism for makingthese decisions involves minimal reasoning, as they appear to occur rapidly whennecessary, but
there is no discussion of this aspecL GUARDIAN
thus limits
the amountof reasoning that need be done by focusing it on
a subset of the agent’s goals, but there is no clear answerto
the question of whensuch reasoning should be perfornw~
Maes(1991) describes a netwm’k-basedarchitecture with
parametersthat adjust the speed with whichthe agent reacts
to changes in its environment. If the environmentchanges
slowly, the agent can perform more reasoning before respending; in very unpredictable environments, the agent
must react with little or no reasoning. The parameters
changeonly in response to the unpredictability of the environment~and are not affected by the particular situation in
whichthe agent finds itself. The balance betweenacting and
reasoning changeson a global basis, and no attempt is made
mdirect the reasoningtowardsspecific goals.
334
POSTERS