From: AIPS 1994 Proceedings. Copyright © 1994, AAAI (www.aaai.org). All rights reserved. Quantitative Operator Selection for Planning Under Uncertaintyt Todd Michael ~ Manseli Departmentof ComputerScience The University of Melbonrne Parkville, Victoria 3052,Australi& mansell@modmel.mrl.dsto.gov.au. Abstract Thispaperdescribesthe best fLrst searchstrategyused by U-Plan(Mausell1993a), a planning systemthat constnt~ quantitatively ranked plans given an incompletedescription of an unce~environment.UPlanacceptsuncertainandincomplete information about the environment,characterisesit using a DempsterShaferinterval, andgeneratesa set of multiplepossible world sta~s. Plan cctmtruction takes place in an abstractionhierarchywherestrategic decisimm are made beforetactical decisions.Searchthroughthis abstraction hierarchyis guidedby a quantitativemeasure (expected fxdfilment) based on decision theory. This search strategyis best first withthe provisionto Ul~_~_teexpectedfulfdmentsandreviewpreviousdecisionsin the fight of planningdevelopments. U-Plangenerates multipleplans for multiplepossibleworlds,andwill attemptto use existingplansfor newwm’ld situatictm.A super-planis thenconstructed,basedonmerging the set of plansandappropriatelytimedknowledge acquisition operators, whichare used to decide betweenplan alternativesduringplanexecution. 1 Introduction to U-Plan Traditional planning systems describe the planning problem as the composingof a course of action that transforms the world from a given initial state to a desired goal state. This description makes two assumptions about the planning domain: complete and accurate informationabout the ~rld is available; and the environmentis static. Theseassumptionsensure that a constructedplan can be successfully executedin such an ideal world. However,such planners rarely produceplans that workin the real world, that by its nature is dynamic and its description is imprecise. To devise a useful plan given uncertain and/or incomplete information about a dynamic environment requires the removal of these assumptions. A major problemwhenplanning given incomplete and uncertain informationabout the environment is that it is not possible to construct oneinitial state that precisely tThis work was done under the CRCfor Intelligent Decision Systems, while the author was employedby DSTOAustralia. *Present address Materials Research Laboratory, PO Box 50, Ascot Vale, 3032, Melbourne, Australia. and nnambig~3~qly represents the world. U-Planuses a possibleworldsrepresentation,wherethe availableinitial informationis used to construct every possible initial state of theworld.Associatedwith each possible worldis a numericalmeasureof belief specifying the degree to whichthe evidencesupports each possible world as the onethat representsthe true state of the world. A hierarchical approach to planning is used as it significantly reduces thesearchspacebyfirst planning at abstract levels, and then expandingthese abstract plans into moredetailed plans. At the highestabstraction level strategic decisionsare made,whileat the lowestlevels of abstraction, ta~cal decisions about how best to implement the strategy, are made.U-Planutilises a set of (predefined) goal reduction operators that encodehow planninggoal is reducedby the operatoxasapplication. Whatresults is a planninghierarchytree wherethe goals are broken up into subgoals by the goal reduction operators.This allows us to first makethe strategic decisions, whichthen guidesall other decisions downto the tactical implementation of the subgoals. Ameasureof expectedfulfilment (section 5.1) is used whenselecting whichoperator to apply next. In support of hierarchical planning, eachpossible worldis describedat a numberof predefinedabstraction levels, resulting in decisions being madebased on a descriptionof the worldat a suitablelevel of detail. U-Planconstructs a plan for one possible world at a time, the first plan being constructed for the possible worldwith the greatest likelihoodof representingthe true world. Beforea plan is constructedfor the next possible world, the suitability of reapplyingan existing plan to this worldis assessed. Associatedwith each plan is the set of possible worldsit worksfor. If a plan partially worksfor anotherpossible world(e.g. the strategy works but someof the detail is different), then that part of the plan is used for this possible world, and planning continuesfromwherethe plan failed. Whena plan exists for every possible world, the operator order of all the plans is combinedto obtain a single planning tree that brancheswhenthe operator executionorder differs. At this point the ability to acquire additional knowledge is used. At each branch, a knowledgeacquisition operator can be inserted to gather current informationabout the state of the world and so determinewhichaction in the planningtree to carry out next. MANSELL 311 From: AIPS 1994 Proceedings. Copyright © 1994, AAAI (www.aaai.org). All rights reserved. Central to planningusing U-Planis: the set of states are representedat several abstractionlevels; the selection of reduction operators is not based on the modaltruth criterion, but dependenton a calculation of expected fulfilment; the system will plan to acquire additional knowledgewhenit is advantageous to do so, and an attempt is madeto apply an existing plan to morethan oneinitial state. 1.1 The Air Combat Domain U-Plan has been applied to a simplified air combat domain.The air combatdomainis dynamicand requires agents to act given uncertainand incompleteinformation. Generatinga plan for aircraft operating in such a domain requires the consideration of a large numberof plan strategies. Therole of an AI plannerin sucha domainis to constructa planthat best fulfils the stated goalsusing only available information, with the option to plan to acquireadditional informationif required. A simplified air combatdomainis introduced in this paper that considers two agents: the defender (whose objective is to defendhimselfand a designatedairspace) and the aggressor (who is invading the airspace controlled by the defender). Whilethe various actions availableto the agents in this domainare diverse, certain patterns in the moreabstract strategies can be identified. For examplesomeof the actions available to the defender are target monitoring, strategy selection, strategy implementation,and evaluation of attack strategy. To operate in such an environmentrequires a sophisticated planning and reasoningsystem. Thestrategies available to U-Planare centred around the selection of either a Beyond-Visual-Range Attack or a Visual-RangeAttack. Thedifferent wayssucha strategies can be carried out are numerous,each one involvesa uniquecourseof action. It is intendedthat U-Plangeneratea suitable courseof action for the defenderaircraft givenonly the information available to the aircraft at plan time. This plan should intend to acquire necessary information when appropriate,havea high likelihoodto success,be suitable to as manypossible worldsas feasible, and workfor the worldsthat are mostlikely to be true. Theplan produced by U-Planis intendedfor use in post missionanalysis of the defenderaircraft, not as a real time planningaid (as the systemis not runningin real time). $ State Representation Whenan incomplete model of the world is all that is available,a set of initial states canbe usedto describethe alternative environments.U-Planemploysa set of initial possiblestates (P-states) to describewhatmightbe true the world. AP-state, ps(a), is a completedescription one possible world using propositional statements. Each P-state is described hierarchically with n levels of abstraction, (ps(a)={tl(a) ... in(a)}) wheren is domain 312 POSTERS dependentand selected during knowledgeengineering. Thelevel ti(a ) is a completedescriptionof a worldat the ith level. Thehighestlevel of abstraction gives a coarse description of the state of the world. Thelowest level gives a detailed view of the world. Intermediate levels provide the description required to makea smooth transition betweenboth extremes. Associated with each P-state is a two valued quantitative measure(an evidential interval (Shafer 1976)) that characterises the weight of evidence that supportsthe P-state accuratelydescribesthe true state of the world. This information is used by U-Plan in determiningthe order in whichP-states are plannedfor, and the final executionorder of operators. Adetailed descriptionof howP-states are generatedandused can be found in t~ansel11993a). 3 Reduction Operator Planningoperators represent actions that the systemmay performin the given domain.Therole of an action is to changethe state of the world,the aimof an operatoris to represent howapplying that action will change the system’s view of the state of the world. U-Plan uses reduction operators to give alternative methodsfor achievingthe goal at a lowerlevel of abstraction, or at the tactical level it describesthe direct effects of an action on the P-state. Theseare SIPE-like operators (Wilkins 1988) where the closed world assumption is implemented,and hierarchical planningused. Each operator contains information about howand whenit should be applied (Mansell 1993a). Included this frameworkis the operator’s name,the abstraction level it operatesin, the necessarypreconditionsthat must be true of the P-state, andthe satisfiable preconditionsit will attempt to maketrue. The plot of the operator providesstep-by-step instructions on howto performthe action represented by the operator. This includes a description of the goal reduction operators that are appliedat the next level of abstraction, or at the lowest level of abstraction,howthe operatorchangesthe P-state. A function for calculating the probability of the reductionoperator succeedinggiventhe current P-state is also included. The availability of such a function is domainspecific and non-trivial to formulate. In the air combatexamplediscussed here the function is obtained empirically(basedon historical data). Theprobability succcssdoesnot providesufficient informationto select a reduction operator as it does not take into account the goals of the system.It is for this reasonthat associated with eachreductionoperatorlisted in the plot of a parent reduction operator is a measure of fulfilment, representing the degree to whichthe reduction operator achievesthe goalof the parent. U-Planuses a deductivecausal theozy (Wilkins1988), to deducethe context dependenteffects of applying a c’urca~ ~ [Boo] From: AIPS 1994 Proceedings. Copyright © 1994, AAAI (www.aaai.org). the All rights almsreserved. of the system,andthe [900,SO0] SVt_ATrA~ [,~Ol [2OOl [SOOl ~n~ZN_c~Br~L~01~ f200l [~o,~oJ~ol ARD [1000] / ATTA~ [ [700]--~_~9~.4D ]L~P Figure 1: The strategic portion of the absttaction hierarchy for the ~plified mr combatex~nple. reduction operator to a P-state. The effects that are deducedare considered to be side effects, wherethose that are introduceddirectly by the reductionoperator are the direct effects. Theuse of deducedeffects simplifies the description of the operators by removingthe needfor extensiveacid and delete lists. After the application of each reduction operator a set of triggers arc used to determineif the world has been changedin such a way that the deductive rules need he applied. If so, the deductivecausal theoryis usedto changethe P-state to be consistent with all the effects of an action. Theside effects of applyinganyreductionoperator are recordedin the planninghierarchytree. expansionof each layer in the tree describesa moredetailed level of abstraction in the plan wpresentedby the nodesin the tree. 5 Operator Selection Manyclassical planningsystemsnsc a state-based search strategy to solve planningproblems.To find a solution one applies operators to a state description until an expressiondescribingthe goal state is found. U-Planuses a quantitative measure,called expectedfulfilment, in an abstractionhierarchyto guidethe selectionof operators. A plan constitutes the successive application of reduction operator from~ highest level of abstraction (i.e. the goal function)downto the mosttactical detail. the air combatexample(fig. l) the goal of defending specifiedassets is accomplished by first choosingwhether to attack or turn awaythe aggressor, through specific strategies and manoeuvres available, downto the detailed implementation of a specific manoeuvre. The following sections outline howthe reduction operators are selected and implemented,and the process that is continuallyreviewingthese decisions. 5.1 Calculating Expected Fulfilment Goals in manydomainsdealing with uncertainty (for instance the air combat domain) are not precise requirements. Manygeneral goals can be fulfilled to various degrees by acl,.ieving alternative subgoals. However,not all subgoals are equally likely to be 4 The Abstraction Hierarchy achieved. We adopt an approach to planning by determininga course, of action that is likely to maximise U-plandoes not construct a state-based search tree, but the expected fulfilment of our goal. Consequently,our constructs a strategy hierarchy whichis a decision tree plans are not exhaustive. Theydo not elaborate all the like structure,, wherethe nodesin the hierarchyrepresent alternative actionsrequiredin all possibleworlds.Rather, a continuoustransition of actions fromthe strategic (at they specify alternative actions that are likely to the root node) to the tactical (at the leaf nodes). maximisothe expected fulfilment of our goal in the strategy hierarchy can be represented as an AND/OR possible worlds that are consistent with our partial search tree, the root noderepresentingthe strategic goal descriptionof the environment. of the system,andthe leaf nodesrepresentingthe tactical Expectedfulfilment is a quantitative measureused to details of howthe goal is to be achieved. Eachnode in rank the reduction operators whichachieve the goals of the tree is a subgoalnoderepresenting the current goal the active operator (i.e., the next reduction operator and P-state, and certain pairs of nodes are cormectedby chosen to be expanded) for selection purposes. For arcs representingthe application of a reductionoperator example,if Attack is the active operator who’sgoals we that producesthis subgoalnode. wish to achieve, the expectedfulfilment for BVRAttack For example, figure 1 shows part of the strategy and VRAttack(the operatorsthat achievethis subgoal)is hierarchyfor a simplified air combatdomain.Thegoal of calculatedandusedas a basis for the selection. defendingdesignatedassets can he achievedby either an Probability theory provides an effective methodfor Attack or Turn_Away.Thestrategy hierarchy showsthat choosing actions capable of producing consistently there are two reduction operators that can be applied to accurate choices. Information such as intelligence achieve anAttack, the BVR_Attack(or BeyondVisual reports, preferences, and raw data can be encodedand Range attack) and the VR_Attack (or Visual Range manipulated by a probabilistic inference engine to attack). Thenext level presents the manoeuvresapplied produce useful recommendations.Whereasprobabilities to achieve these specific attacks, whichare achievedby are usedto representthe likelihoodof events, fulfilments the implementationof tactical actions (not shown).The are used as a local measureof the degree to whichthe applicationof these operatorsrepresentsa clarification of consequentof the action achieves the intendedgoal and MANSELL 313 From: AIPS 1994 Proceedings. Copyright © 1994, AAAI (www.aaai.org). All rights reserved. the desirability of achievingthe goal using that action. Thetermfulfilment is usedto captureboth the essenceof utility scaled accordingto the desire to use a particular approach.Theterm utility is not used in this description to avoidconfusion,as there are subtle differencesin how they are usedand whatthey represent. The expected fulfilment is used as a measure of an action’s likelihood to produce the consequent that achievesthe agent’s goals. If weuse the measureF(c) represent the degree of fulfilment of consequentc, then the overall expectedfulfilmentassociatedwith action a is givenby: EF(a) = F(c)P(cla,e), (1) where, P(cla, e) is the probability of achieving consequence¢, conditioneduponselecting action a and observing evidence e. For example, to calculate the expectedfulfilmentof the Attackoperator in figure I, the probability of successfully executing an attack in the given P-state is multiplied by the degree of fulfilment obtainedby executingthe action (depictedin figure 1 as the first number in squarebracketsabovethe operator). Theexpectedfulfilment of action a, EF(a),is regarded as a gauge of the merit of action a. The expected fulfilment is used as a procedure for choosing among alternative (or competing) actions. Whengiven the choice betweentwo action (eg, Attack and Turn-Away) the selectionis basedon the actionthat yields the highest expected fulfilment (ie, EF(Attack)or EF(Turn-Away)). This result will dependon the description of the P-state whenthe selection is made.This process can be thought of as establishing a rank order in which one should attemptedto apply them. The semantics and justification of fulfilments are identicalto thoseoutlinedfor utility theory.Utility theory is not simplya convenientmathematicalformula, but is basedon studies of the psychologicalattitude towardrisk, choice, preference,and likelihood. Theessenceof utility theory, and therefore fulfilment, is captured by the axioms of utility Theory (Von Neumann & Morgenstern 1947). 5.2 Applying an Operator Oncethe reductionoperatorsthat achievethe goals of the parent operator have been ranked using the expected fulfilment calculation, they can be tested to determine their suitability to the P-state. U-Planwill attempt to applythe reductionoperatorsto the P-state in the order in which they were ranked (by their EF), until appropriateaction is found.If the necessarypreconditions of a reductionoperatorare true in the active P-state, then the reductionoperator is provisionallyselected, else the planfail for that operatoris applied(this usuallyinvolves backtracking). Whena reduction operator that satisfies the necessary preconditions has been found, the satisfiable preconditions are tested. If anyof these are not 314 POSTERS true, U-Plancan attempt to satisfy themusing reduction operators of equal or lower abstraction. If these preconditionsare not satisfied, the operator is rejected, andit’s planfail procedureis implemented. Onceboth sets of preconditionsof a reductionoperator can be shownto be true in the active P-state, the operator is accepted and its plot can be applied. The plot represents the effects the reduction operator has on the state of the world, andthe subgoalsthat maybe used to achieve this subgoal. Whenapplying the plot, the next level of the strategy hierarchyis exposed,andagain the subgoalwith the highestexpectedfulfilmentis selectedto be expanded next. The plot of operators ~presents actions at the lowestlevel of abstraction specifyhowthe P-state is physicallychangedbytheir application. This process of applyingoperators continuesuntil the next layer of the strategy hierarchyhas beenexposed.At this point, the earlier selection of specific actions are reviewedas describedbelow. 5.3 Reviewing Selected Operators When constructinga strategyhierarchyit is possiblethat, as a plan’s detail is filled out, it becomes less likely to succeed. This is partially becausethe initial strategic decisionsare basedon informationat a morecoarse level of abstraction. Asthe plan is expandedandmoretactical decisions aboutthe implementation of specific strategies are made, the expected fulfilment of specific plan branchesmaydecreases. Also, the EFcalculation for a higher level operator mayinclude a range of possible useful actions. However,as the plan becomesmore detailed, certain alternatives will be discarded,possibly resulting less that ideal actions beingselected, reducing the true e, xp~edfulfilmentof the plan branch. This makesit important to review earlier decisions while planning. After the application of a group of reduction operators U-Plan compares the e~ected fulfilment of the current subgoals,with those of previous subgoals, to determinesif the current subgoal remains favourable. U-Planuses an offset added to the current subgoal to take into account expected lower EF as increaseddetail is addedto the plan. Includingan offset is an iterative deepeningstrategy includedto avoid the problemof the system jumpingaround from branch to branchin the strategy hierarchy. Theoffset value will dependon the difference in abstraction level of the subgoals. Theiterative deepeningstrategy will not be discussionin the remainderof the paper, whichwill focus on the moregeneris process. To review these selections one must have a wayof updatingthe EFvalues calculated for the nodes in the abstraction hierarchy basedon the detail addedto that nodesplanningbranch.A set of updaterules are used to re-evaluate the fulfilment and probability of previously expanded operators given the most recent planning From: AIPS 1994 Proceedings. Copyright © 1994, AAAI (www.aaai.org). All rights reserved. developments.The update rules used dependon whether the nodeproducesan AND or ORbranch in the tree. In the case whenan operator is expandedproducingan ORbranchin the abstraction hierarchy, the updaterules used to determine the fulfilment and probability of a parentnodegivena set of possiblechildren are givenby: I MAX F(parent) = {F(child)]childreEF(child)} , (2) ]~I~o.o.,, [85o,81o,7oo1 /VR ATFACK - ( / SET_~.ARn~O / {iooo,o.9; [8001 VISUAL LOCK [sso.slo.6~ts] ACQUmE j ~ [~LOSE_D4 ~r_TARGET ~ {800~ "0/ / ’ ’ " AR LOCK pooo, o.80~\ gaoo.s.o} 0ooo.~71 {8o~o.81) \ I MAX e(~nO= {e(ch~ld)[childrenF~(child)L VntE ~ARM_WEAPON READY,m~{IO00,1.O} {zooo, o.9} ~mRE~N {ZO00,O.9} {zooo,o.o} Wherechildren is the set of children operatorsthat have been expandedor considered for expansionand have not beenrejoct~ Simplystated, rules 2 and3 tell us that the fulfilment andprobability values for the parent of an OR nodeis equalto the fulfilmentandprobabilityof the child nodewith the greatest expectedfulfilment that has not beenruled inapplicableto the P-state (i.e. dueto failure of preconditionsduring expansion).If the operator with greatest EFis not fromthe current subgoal, a changein the planningdirection occursand is recorded. The updating of the parent reduction operator at an ANDnode involves updating the probability and fulfilmentas follows: I MIN F(parent) = {F(child)lchildreF(Child)} , (4) P(parent) = I-I P(child). ffoo] (5) children In the AND case the rules and their justifications are less obvious(Mansell1994). The fulfilment of a parent operator is replaced by that of the child fromthe set of children with the lowest fulfilment. Normally, when confronted by an AND node, the children are given the samefulfilment as that of the parent. Thejustification being, that as the parent can ouly be achievedby a single sequence of actions, then this reduction is simply a refinement of the parent operator 0.e., these actions should wholly achieve the desired goal). Howeverthe situation mayarise when, at a lower level in the abstraction hierarchy, the fulfilment valuefor oneof the child operators mayitself be updated(by its subsequent descendants) to a new, lower valuer. Whensuch circumstancearise, the child operator is not fulfilled completely by its descendant, and consequently, the parent operator is no longer completelyfulfilled by its children. In the subset of the air combatexamplegiven in figure 2, a simple scenario for a Close_Inmanoeuvre operator is evaluated. Theexpectedfulfilment for this operator is calculated (the first numberin the squarebrackets above the operator, i.e., [850]) based on the fulfilment and probability values (shown in the braces below the Figure 2: Anexampleof the abstraction hieraroby for a Close In manoeuvre scenario that demon~ates the updating of Palfilments and probabilities. Progressive (fulfilment, probabilUy)and[EF]values are givenfor each operator. operator, i.e., 1000and0.85 respectively) obtainedfrom the operator. Onexpandingthe Close In operator, the next level of the plan is uncovered.This showsthat the Set_Bearing, Acquire_Targetand Fire_Readyoperator are to be applied. Thefulfilments and probabilities for these are calculated and shownin braces below the operators. As this is an AND operation, update rides 4 and5 are usedto updatethe fiflfilments andprobabilities for the parent, Close_In, operator (shownin the second set of braces belowthe operator, (1000,0.81)). These updatedvalues are used to calculate the updatedEFvalue for the Close_Inoperator(i.e., [810]). Figure 2 also includes an examplewhererules 2 and 3 are usedto updatethe fulfilmentandprobabilities for the parent of an ORnode. In this case the Acquire_Target can be achieved by either a I~sual_Lock or a Radar_Lock.The Visual_Lockis chosen as it has the higher EF of the two. However,the V~sual Lock has a m lower fulfilment than its parent and this is propagated back through the branch updating the Close_In EF to [648]. Asa result of propagatingthese valuesbackup the branch, the Side operator becomesfavourable over the expandedbranch Close_In. 5.4 Sensitivity of Expected Fulfilments The ~ fulfilment value is calculated by multiplyingthe probability of success of an operator by the degreeto whichthe operatorfulfils its intendedgoal. It is therefore useful to knowthe relative sensitivity (Karnavas,Sanchez,and Bahill 1993) of probability and fulfilment on the expected fulfilment function. The relative-sensitivity of a function F to the parameterct over the normaloperatingpoints is givenby: SF= %changein F °~’//F ~ ~nor = o~ar//a = ~[NOP FO’ MANSELL 315 From: AIPS 1994 Proceedings. Copyright © 1994, AAAI (www.aaai.org). All rights reserved. whereNOPand the subscripts 0 meansall functions and their parameters are evaluated over their normal operatingpoints. Given,the expectedfulfilment function in equation 1, the relative-sensitivities fimctions are computedas follows: NapEFoFo, $~ =--~ F0 P CO (8) Therelative-sensitivities of probability andfulfilment on the expected fulfilment are independent. This informationis used whendeterminingthe certainty with whichprobabilities and fulfiknents mustbe knowngiven a set of expected fulfilments. For example,given the choice betweentwo actions with close EFvalues, one can calculate the degreeof confidencein the fulfilmentvalues required of the actions, given the percentagedifference betweenthe actions probabilities. 6 Plan Reapplication U-Plan applies plan reapplication in an attempt to determineif a plan generatedfor one initial P-state can be adoptedfor anotherinitial P-state. Thedesiredresult being fewer plans than the numberof initial P-states. This is implementedby attempting to reapply plans generatedfor oneinitial P-stateto other initial P-state. Aplan is reapplicableif all the reductionoperatorsin the plan (that are not redundant)havetheir preconditions metunderthe newinitial P-state, andwhenapplied result in the goal state being achieved. If a plan, during re, application,fails dueto the unsuccessful applicationof an operator, that plan is not entirely discarded. U-Plan will attemptto use the part of the plan that wassuccessful and planning continues from the point where the plan failed. Thedesire is to construct plans with the sameor similar strategies by reusing, at least part of, the plan at the high level of abstraction. Whenmorethan one plan partially worksfor a newinitial P-state the best plan (Mansell1993a)is used. 7 Super-Plans Onceplans exists for all the P-states, with supportand plausibility abovesomethreshold, a single super-planis constructed. This is achievedby mergingthe set of plans constructedfor the set of initial P-states, that is applying identical operator sequencesand branchingat the point whereplans differ. At each branch in the super-plan a knowledgeacquisition operator is added, attaining the informationrequired to select whichaction in the superplan to apply next. Thecase mayarise whenthe required informationto differentiate betweenalternative branches is not available.In this case, the selectionis basedonthe 316 POSTERS degree of evidence supporting for each branch of the super-plan (see (Mansell1993a, Mansell1994) for more detail). This final step of producinga super-planis an importantpart of presentinguseful coarseof action that couldbe applied by the system. In most cases, U-Planproducesa super-plan in less time than it wouldtake a traditional planner to produce one plan for every possible world (see (Man~ll1993a, Mansell1993b)for moredetails on these results). 9 Conclusion U-Plan is a hierarchical planner that deals with information represented at a level of abstraction equivalentto the action being investigated. Outlinedin this paper is the quantitative best-first search method employed by U-Plan for operator selection in an abstraction hierarchy. As this process is a forward propagatingpartial decision tree, a methodfor reviewing previousdecisionsin the light moredetailed information is included. The update rides are presented in some detail, and an exampleof their operation presented. UPlan has provedto be a effective planningsystemin the air combatdomain(Mansell 1993a), and the expected fulfilment calculation a reliable formulafor operator selection. Acknowledgments I would like to thank Dr. GrahameSmith and Dr. Elizabeth Sonenbergfor their manyinsightful comments. References Karnavas,W.J., Sanchez,P. J., andBahill, A. T. 1993. Sensitivity Analysis of Continuousand Discrete Systems in the Timeand FrequencyDomainsIn IEEETrans. on Systems, Man,and Cybernetics, 23(2): 488-501. Mansell, T. M. 1993a. A Methodfor Planning Given Uncertainand IncompleteInformation. In Proceedingsof the Ninth Conference on Uncertainty in Artificial Intelligence, 350-358.WashingtonDC. Mansell, T. M. 1993b. Air CombatPlanning Using UPlan for Post MissionAnalysis. In Proceedingsof the First Australian and NewZealand Conference on Intelligent InformationSystems,644-648.Australia. Van Nenmann,J. and Morgenstern, O. 1947. Theory of Gamesand EconomicBehawour.2nd ed. Princeton, New Jersey: PrincetonUniversityPress. Mansell, T. M. 1994. Planning Given Uncertain and IncompleteInformation. Forthcoming. Shafer, G. A. 1976. A MathematicalTheoryof Evidence. Princeton,NewJersey: PrincetonUniversityPress. Wilkins, D. E. 1988. Practical Planning:Extendingthe Classical AI PlanningParadigm.Los Altos, California: MorganKaufmann.