From: AAAI Technical Report SS-94-06. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. TowardA General Dynamic Decision Modeling Language: An Integrated FrameworkFor Planning Under Uncertainty Tze-YunLeong Clinical Decision MakingGroup MITLaboratory for ComputerScience Cambridge, MA leong@lcs.mit.edu can generatemanyoptimalcoursesof actions or plans, conditional onall possibleevolutionsof the eventsor states in the environment. Abstract Decisionmakingis often complicatedby the dynamismand uncertainty of the information involved. This paper presents a newmethodology for formulating,analyzing, and solving dynamic decision problems.Theproposedformalismintegrates multi-disciplinary approaches; it also generalizes current dynamicdecision modeling frameworks.The ontology, graphical presentation, and mathematicalrepresentation of the decisionfactors are distinguished;modularextensions to the scope of the admissible decision problemsand systematic improvements to the efficiencyof their solutionsare supported. 2 Dynamic Decision Modeling A dynamicdecision model,e.g., dynamicinfluence diagram [Tatmanand Shachter, 1990], Markovcycle tree [Hollenberg,1984], or stochastic tree [Hazen,1992], is basedon a graphicalmodelinglanguagefor visualizing the relevant variablesin a dynamic decisionproblem.In general, sucha modelconsists of the followingsix components, the first five of whichconstitute a conventionaldecision model: ¯ Aset of decisionnodeslisting the alternative actions that the decisionmakercan take; 1 Introduction ¯ Aset of chancenodesoutliningthe possibleoutcomesor happeningsthat the decision makerhas Decisionmakingin our daily lives is often complicatedby no control over; the dynamism and uncertaintyof the informationinvolved. ¯ Asingle or a set of valuefunctionscapturingthe In medicine,for instance, a common problemis to choose desirability of eachoutcome or action; a courseof optimalIreatmentsfor a patient whosephysical ¯ A set ofprobabilistic dependencies depicting how conditions mayvary in time. Dynamic decision modelsare the outcomes of each chance node depend on othframeworks for modelingand solving such dynamicdecier outcomes or actions; sion problems.Theseframeworks are basedon structural ¯ Aset of informationaldependenciesindicating and semanticalextensions of conventionaldecision modthe informationavailable whenthe decision makels, e.g., decision trees andinfluencediagrams,with the er makesa decision; and mathematical definitionsof finite-state stochasticprocess¯ Anunderlying stochastic process governingthe es. evolutionin time for the abovefive components. This paper presents a newmethodology for dynamicdeciIn existing frameworks,the mathematicalproperties of a sion modeling, called DYNAMOL (for DYNAmic decision dynamicdecision modelare defined with respect to its MOdelingLanguage). The DYNAMOL design integrates approachesand techniquesin AI, decisiontheory, and con- graphicalstructures; solvingsucha modelinvolvesdirectly the graphicalstructures. 1Iol theory;it supportsformulation,analysis, andsolution manipulating of dynamicdecision models. DYNAMOL generalizes existRecently, [Leong,1993] has identified semi-Markov deciing dynamicdecision modelingframeworks; it differs from sion processes as a common theoretical basis of dynamic previousapproachesby distinguishingthe decision ontolo- decision modeling. DYNAMOL attempts to integrate the gy, the graphical presentation, and the mathematical graphicalcapabilities of the existing frameworks, and the representation of dynamicdecision problems.The soluconcisepropertiesandvaried solutionsof the mathematical tions to the models formulated in DYNAMOL, called formulations. policies or decisionrules, are guidelinesfor choosingthe optimalactions over time to achievesomegoals in a given 3 The DYNAMOLDesign environment.In the AI planningvocabulary,a dynamicdeTheDYNAMOL languagehas three distinct but integrated cision modelcorrespondsto a planningproblem;a policy 155 components: a dynamicdecision grammar,a graphical presentation convention, and a formal mathematical representation. Thedecision grammar allowsspecification of the decision factors and cons~aints. Thepresentation convention,in the tradition of graphicaldecision models, enables visualization of the decision factors and constraints. The mathematicalrepresentation provides a conciseformulationof the decisionproblem;it admitsvarious solution methods, depending on the different properties of the formalmodel. 3.1 DynamicDecision Grammar A dynamicdecision model formulated in DYNAMOL has the followingcomponents:1) the time-horizon, denoting the time-frame for the decisionproblem;2) a set of valuefunctions, denotingthe evaluationcriteria of the decision problem;3) a set of states, denotingthe possibleconditions that would affect the valuefunctions;4) a set of actions,denoting the alternative choicesat each state; 5) a set of events, denoting occurrencesin the environment,conditional on the actions selected, that might affect the evolutionof the states; 6) a set ofprobabilisticparameters, denotingthe transition characteristics, conditionalon the actions, among the states andevents;7) a set of declaratory constraints,suchas the valid durationof particular states, actions,events,andprobabilisticparameters; and8) a set of strategic constraints, suchas the numberof applicability for certain actions,the validordering of a subsetof actions, andthe time spanbetweenapplicability of twosuccessive actions. The dynamicdecision grammarfor DYNAMOL is an abstract grammar. Following the convention in [Meyer,1990], this grammar contains the followingcomponents: ¯ Afinite set of namesof constructs; ¯ Afinite set of productions,eachassociatedwith a construct. Anexampleof a production that defines the construct "model"is as follows: Model.->name:Identifier; contexts: Context-list; definitions: Definition-list; constraints: Constraint-list; solution: Optimal-policy Eachconstructdescribesthe structure of a set of objects, called the specimensof the construct. Theconstructis the (syntactic) type of its specimens. In the aboveexample, object with the type Modelhas five parts: name(with type Identifier),contexts(withtypeContext-list).... etc. Theconstruets/types appearingon the fight-handside of the above definition are similarly definedby different productions.A set of primitiveconsmacts/types areassumed, e.g, String, Cumulative DistributionFunction, etc. Therearesome fixed waysin whichthe structureof a construct can be specified. Theaboveexampleshowsan "aggregate" production,i.e., the constructhasspecimens comprising of a fixed number of components. Othertypes of productions include: "Choice"productions, e.g., the timedurationin the decisionhorizoncanbe finite or infinite: "time-duration --> $+ u {0} I oo "List" productions,e.g. the state-space of the decision problemconsists of oneor morestates: State-space ---> State+ The DYNAMOL grammardefines the structure of a dynamic decisionmodelin termsof its components; the smJctures of these components are recursively defined in a similar manner.Thegrammar specifies the informationrequired to build a model.Onthe other hand, there can be manyways to manipulatesuch information.In other words, the abstract grammar can support different interface implementations.For example,an object of type Statespacemaybe specified in 3 different ways: ¯ Text command interface: Type"state-space state1 state-2 state-3" to command-prompt. ¯ Graphicalinterface: Drawthree state-nodes labeled"state-l,’, "state-2,""state-3"in the display window. ¯ Routineor codeinterface: Type"(define-statespace ’state-1 ’state-2 ’state-3)" to the Common Lisp prompt. Thegrammar defines the syntax of the language. Thesemanticsof the languageis definedby the correspondence fromthe constructs to the underlyingmathematicalrepresentation of a semi-Markov decision process. 3.2 GraphicalPresentation Convention The graphical presentation conventionin DYNAMOL prescribes howthe decisionfactors andconstraintsexpressible in the grammar are displayed.It is againdefinedin termsof the correspondencefromthe grammarand graphical constructs to the underlyingmathematical representationof the model.Followingthe conventionin decision modeling,for instance, an action-variableis denotedby a rectangle, an event-variable an oval, etc. In DYNAMOL, however,both declaratoryandstrategic constraintsare also visualizable. Themodelcan be presentedin part or in wholeas desired. Moreover,all decision components can be presentedin different perspectives. Figure1, shownlater in this paper, depicts twodifferent perspectivesof a simpledynamicde- 156 cision modelfor a medical example. 3.3 Semi-MarkovDecision Process A dynamicdecision modelspecified in the decision grammar in DYNAMOL is automatically compiled or interpreted into a semi-Markovdecision process. The modelspecification maybe more general than the semi-Markovdecision process definition wouldallow. For instance, the events and their correspondingconditional probabilities are compiled into the transition probabilities in the state space of the semi-Markov process Formally, a semi-Markovdecision process is characterized by the following components [Howard, 1971] [Heyman and Sobel, 1984]: ¯ A time index set ~, ¯ A decision or control process denoted by a set of random variables {D(t);te T}, where D (t)~ A = { 1, 2 ..... k} is the decision made at timet; and ¯ A semi-Markovreward process denoted by a set of random variables {S (t) ;t ~ T], where S(t) e S = {0, 1,2 .... } is the state of the process at timet, with: 1. an embeddedMarkovchain denoted by a set of random variables {Sm;m> 0} ; such that Sm = S (Tm), where T1 < T2 < T3 <... are the randomvariables denoting the successive epochs (i.e., instants of time) at which the process makestransitions; probabilities sets of transition 2. k {P!.a);i>O,j>O, 1 <a<k} among the states of the eml~lded chain, such that for any given decision a ~ A, and states i, j ~ S: p.C ) q = P {Sin+ 1 =Jl Sm=i, Dm= a } = P{S(Tm+ l) =jIS(Tm) =i,D(Tm) = whichalso satisfies the Markovianproperty: q = P{Sm+l =jlSm= i, Dm=a } = P[Sm+1 =jlSm=i, Sm_l=h ..... Dm = a} 3.k sets of holding times {x~.a);i>0,j>0, 1 <a<k} amongthe states of the embe~ldedchain, whichare randora numbers with corresponding distributions {H~a) (n);i>O,j>O, l<a<k}, such that for any giveh decision a 6 A, and states i, j E S: (a) Hij (n) = P{Tm+I-Tm <nlSm=i, Sm+1 =j, Dm=a } 4.ksetsofrewardsoryields {r~a)t (l);i>O, 1 <a<k} associated with the states of the embeddedchain, such that for any given decision a ~ A, r~a)" (l) is the value achievable in state i e S of the chain over time interval (l, l+l). 3.4 Solution Methods A dynamic decision problem can be expressed as the dynamic programming equation, or Bellman optimality equation, of a semi-Markovdecision process. For a decision horizon of duration n time units, with a discount factor 9, the optimal value achievablein any state i, Vi , gwenan initial value Vi (0), is [Howard,1971]: (n,9) co = maxaI~p(a)~n J (a) + h -ij ra= ! t r a) (l) (m) + n ~P(ija’ ~ (a, ( m)x J ra=l F~.~m-I Ll~=O~lr~a) (l) + ~mv; (n-m, n>O;i,j>O;1 ; 11 <a<k (EO 1) The first addend in EQ 1 indicates the expected value achievableif the next transition out of state i occurs after time duration n, and the second addend indicates the expected value achievable if the next transition occurs before that time duration. This formulation assumesthe samevalue-function or reward structure r[a) (.)in each state i conditional on an action a, independentof the next state j. The most direct solution method, called value iteration, is to solve the optimality equation shownin EQ1. The solution to such an equation is an optimal policy, i.e., a sequenceof decisions over time (whichcould be one single decision repeated indefinitely) that maximizes Vstar ? (N), the optimal expected value or reward for starting state start, at time t = 0, or for decision horizon/Q, wheren = N - t is the remaining duration for the decision problem. For infinite horizon problems,EQ1 simplifies into: and 157 unchanged; onlythe valid entities are displayedfor particular timeinstancesor durations. v7 (~) = maxa[~Pij (a) ~ (a) Ehij (m) -J i,j~ 0;1 <a<k Thesenewconstructsdirectly correspondto a generalclass of semi-Markov decision processes with dynamicstatespace andaction-space.Thevalue-iteration solution methods as described in EQ1 and EQ2 remain applicable to these problems. m= 1 (EQ2) Mostsolution techniques for existing dynamicdecision modelingframeworksare based on the value-iteration method. [Shachter and Peot, 1992] has also employed probabilistic inferencetechniquesfor solving dynamic influence diagrams. DYNAMOL, however,wouldadmit other moreefficient solutions methodsfor semi-Markov decision processes.Forinstance, policyiteration, adaptiveaggregation, or linear programmingmaybe applicable in the DYNAMOL frameworkif certain assumptionsor conditions are met.Theseconditionsincludestationary policies, constant discountfactors, homogeneous state-space, etc. 4-2 Non-Homogeneous Transition Probabilities In manyreal-life examples,the transition probabilitymaybe time-dependent.For example,the morbidityrate of a patient maydependon his age; a 70 year-old personis more likely to havea heart attack than a 40 year-old one. Such time-dependentor non-homogeneous transition probabilities have been incorporated into DYNAMOL. Again extendingthe decision grammar andgraphical presentation are straightforward. Extendingthe underlyingmathematical representation leads to redefining the transition probabilityin Section4.3 as a functionof time. Thevalueiteration methodin EQ1 has to be modifiedaccordingly. Byseparating the modeling(with the decision grammar) and the solution (with the mathematicalrepresentation) tasks, therefore, different solution techniquescan be employed. Moreover,employinga newsolution technique doesnot involveanychangeto the languageitself; all solution techniques reference only the mathematical representationof a model. 4.3 Limited Memory Transitionprobabilities governthe destinationsof transitions. As comparedto a Markovprocess, a Semi-Markov processsupportsa seconddimensionof uncertainty: duration in a state. A semi-Markov process is only "semi-" Markovbecausethere is a limited memory about the time since entry into anyparticular state. In somecases, limited memory about previousstates or actions are importantin a 4 SupportingLanguageExtension dynamicdecision problem. For example, having had a heart attack beforewouldrenderhavinga secondheart atTheinitial version of DYNAMOL contains only a basic set can of languageconstructs; manyof the constraints mentioned tack morelikely duringsurgery. Suchlimited memory be incorporated into a semi-Markov or Markov process by in Section3.1 are not includedin the initial design. DYtechniques such as state-augmentation. NAMOL supports modular extensions to the language constructs. This section sketches the approachesto adIn DYNAMOL, a newset of syntactic constructs can be indressing someof these issues; someof these possible troduced to specify such limited memory.A newset of extensionshavealreadybeenincorporatedinto the current correspondence rules will then be incorporatedto perform version of DYNAMOL. automatic state-augmentationand calculate the correspondingprobabilistic parameters.Wedo not yet knowif 4-1 Static vs. DynamicSpaces this processcan be fully automated.Theresulting matheOnlystatic state-spaceandstatic action-spaceare currently matical representation, however,should be a well-formed allowedin the DYNAMOL decision grammar.Static statesemi-Markov decision process; the solution methodswill spaceindicatesthat the samestates are valid throughoutthe remainapplicable. decision horizon; static action-spaceindicates that the sameset of actionsare applicablein eachstate, overtime. 4.4 Strategic Constraints Theoretically, wecan modelchangesin the state-space or Strategicconstraintsare closelyrelated to the limitedmemaction-space by providing some"dummy" states and "no- ory capabilities. Someexamplesof these constraints are" op" actions. In practice, however,suchextraneousentities action-number constraints, whereone or moreactions canmaycompromise modelclarity andsolution efficiency. not be applied for morethan a specific numberof times; The DYNAMOL decision grammarcan be easily extended and action-orderconstraints, whereoneaction mustalways to incorporate dynamicstate-space and action-space. New or never follow another action. Thereare two methodsto productionsfor the corresponding constructscan be written incorporatesuch constraints into a dynamicdecision modto incorporatethe valid time or durationfor eachstate and el. Thefirst methodis by augmentingthe state-space to action. Thegraphicalpresentationrules can remainmostly keeptrack of the numberand/or order of the actions ap158 plied; this wouldusually result in a very complexstatespace, compromisingthe clarity of the model. The second methodis to let the solution moduleworry about keeping track of the constraints. In either case, new constructs can be introduced in DYNAMOL to express the constraints. If the’ state- PTCA,CABG}.For ease of exposition, assume that each state s E S is a function of a set of binary state attribute or health outcome variables 0 = {Status, It,tl, Restonosis } , e.g., "Well":=(Status.= alive, MI = absent, Restenosis = absent), "MI":=(Status alive, MI = present, Restenosis = absent), etc. augmentation method is adopted,a newset of correspon- A partial modelspecification of the exampleproblemin the dencerules need to be added as described earlier. If the dynamicdecision grammaris as follows’: solution methodis responsible for the constraints, these factors only need to be accessible. ($time-horizon:nature discrete :duration60:unit month) ($actionmodieal-treatment :durationall) 5 A Prototype Implementation i$statewell:duration all) ($staterestenosis :duration all) ¯ ,. A prototype implementation of DYNAMOL is currently un($transition :fromwell:to dead :transition-probability 0.1 derway. The system is implemented in Lucid Common :duration30) Lisp on a Sun Spare Station, with the GARNET graphics ($transition :fromwell:to dead :transition-probability 0.2 package.It includes a graphical user interface that allows ;duration(3160]) interactive modelspecification. The specification can be ($transition :fromwell:to restenosis) strictly text-based, or aided by the graphical presentations ($event hight-fat-diet:outcomes (true false):predwell of the available decision variables or constraints. Onlya :succrestenosis :condl-probability <distribution>) subset of the dynamicdecision grammaris included in the ($event stress:outcomes (true false) :predwell :succ first version of this implementation,and the solution method supported is value iteration. Future agenda on the ($inlluence :fromstress:to MI:cond-probability <distribution> :duration all) project include improving the implementation, extending and refining the dynamic decision grammar, identifying ($value-function <function> :stateall) meta-level dynamicdecision problemtypes, searching for ($discount-factor 0.8:duration all) moreeffective solution methods,and investigating relevant ($action-number-constraint CABG :number 3) issues in supporting automated generation of dynamicdecision models. The current domainfor examiningall these Figure 1 showstwo graphical perspectivesof someof the issues is medical decision makingin general. decisionfactors in this model. 6 An Example Theproblemis to determinethe relative efficacies of different treatments for chronic stable angina (chest pain), the major manifestation of chronic ischemic heart disease (CIHD). The alternatives considered are medical treatments, percutaneous transluminal angioplasty (PTCA),and coronary artery bypass graft (CABG).The manifestations of CIHDare progressive; if the angina worsensafter a treat° ment, for instance, subsequent actions will be considered. Evenafter successful treatment, restenosis, i.e., renewed occlusion of the coronary arteries, mayoccur. Hencethe decisions must be madein sequence. The treatment efficacies in lowering mortality decline as time progresses; the treatment complications worsenas time progresses. A major complication for PTCAis perioperative myocardial infarction (MI), or heart attack, which would render emergencyCABG necessary. The effectiveness of the different treatments is evaluated with respect to qualityadjusted life expectancy(QALE). In this problem,assumethat the states S = {"Well", "Restenosis", "MI", "MI+Restenosis", "Dead"} represent the possible physical conditions or health outcomesof a patient, given any particular treatment a E A = {MedRx, In termsof the mathematical representation,the set of actions is A = {MedRx, PTCA, CABG}. The semiMarkov,or in this case, Markovrewardprocess,with time indexset T ~_{0, 1, 2 .... } , is definedby: 1) the embedded Markov chain with state-space S = {"Well", "Restenosis", "MI", "MI+Restenosis","Dead"} as illusmated in Figure1 a; 2) threesets of transition probabilities among the states in S, corresponding to the actions in A; 3) constant holding times with distributions H~/a)(n) = 1 (n - 1), where1 (n - 1) is a step function at time n = 1 (in anyunit); and4) three sets of rewards, correspondingto the amountof QALE achievableper unit timein eachstate in S, with respectto the actionsin A. 7 Related Work This work is based on and extends existing dynamicdecision modeling frameworks such as dynamic influence diagrams, Markovcycle trees, and stochastic trees. It also integrates manyideas in control theory for the mathematical representation and solution of semi-Markovdecision t. Theseare approximations to the graphicalinterfacecommands;the plaintext command interfaceis not complete. 159 ..,...~.,...,..,..,..,..~.. .~...~.~.~.~.,..,.: .,... ~:‘‘~‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘~‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘~‘‘‘~‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘‘~! ~‘ ".:’ ~:~i~ii~’ ...... .~.~.~’i’i’~.~.~.;J ..... ~i~i~i~i~i~i~:,~’,~’,~:,~’,~i~:,:..~;~’,~i~’~:~;~’,’:’3 ENTITIES-LIFESPAN ~ All REfiT ION S-LIFESPAN = All Figure ] Twographical perspectivesof a dynamic decisionmodel.Figure 1 (a) depicts the Markovstate transition diagram.Thelinks representpossibletransitions fromonestate to another;givenanyaction; a transition-functionis usuallyassociatedwith eachlink. Figure1 (b) depictsprobabilistic andinformationaldependencies, in the convention of influencediagrams, among the decisionfactors at timeunit n. processes.Thenotionof abstractionthat lies in the core of ticular, the declaratory and strategic constraints are the methodology design is a fundamentalconcept in Com- explicitly incorporatedinto the graphicalstructure of the model;the probabilistic and temporalparametersare exputer Science,andperhapsmanyother disciplines. plicitly encodedfor each time slice or period considered. [Egar and Musen,1993] has examinedthe grammarapThe dynamicdecision grammarin DYNAMOL, on the other proach to specifying decision model variables and hand,supportsabstract statementsaboutthe decisionsituconstraints. This grammar is basedon the graphicalstrucation, e.g., statements about the validity duration of ture of decisionmodel,andhas not beenextendedto handle particular states, statementsaboutthe orderingconstraints dynamicdecision models. [Deanetai., 1993a], [Dean on different subsetsof actions, etc. Theseabstract stateet al., 1993b],andother related efforts havedevisedmeth- ments are analogous to the macro constructs in otis based on Markovdecision processes for planningin conventionalprogramming languages. Byfocusing on the stochastic domains;they apply the mathematicalformula- decision problemontologyinstead of the decision model tion directly, anddo not addressthe ontologyof general components, DYNAMOL would provide a more concise dynamicdecision problems.Muchcan be learned fromthis and yet moreexpressive platform for supporting model line of work, however,both in navigating throughlarge construction. state spacesfor findingoptimalpolicies, andin identifying Theadvantagesof the graphicalnature of existing dynamic meta-levelproblemtypes for devisingsolution strategies. The graphical representation of semi-Markov processes decision modelinglanguagesare preservedand extendedin Thegraphical presentation conventionrenders has been explored in [Berzuini etal., 1989] and [Dean DYNAMOL. and constraint examinablegraphiet al., 1992];theyfocusonlyon the single-perspective pre- every modelcomponent sentationof relevantdecisionvariablesin a belief network cally. Thedifferent perspectivesin the presentationwould furthercontributeto the visualizationeaseandclarity. or influence diagramformalism. Theoretically, semi-Markov decision processes can approximatemoststochastic processesby state augmentation Theresulting state-space, however, Modelspecification in DYNAMOL is expressedin a higher or other mechanisms. may be too complex for direct manipulationor visualizalevel languagethan that in existing dynamic decision modtion. On the other hand, efficient solution methodsmaynot eling frameworks. In frameworks such as dynamic exist for more general stochastic models.Bydistinguishing influencediagramsor Markercycle trees, mostparameters the specification grammar and the underlyingmathematical of the modelneedto be explicitlyspecifiedin detail. In par8 Discussion 160 model,DYNAMOL aims to preserve the clarity and expressivenessof the modelstructure, while minimizing the loss of information.This wouldcontribute towardthe ease and effectivenessof modelanalysis. Acknowledgments Egar, J.W. and Musen, M.A. (1993). Graph-grammar assistance for automatedgenerationof influencediagrams. In Heckerman,D. and Mamdami, A., editors, Uncertainty in Artificial Intelligence: Proceedingsof the Ninth Conference, pages 235-242, San Mateo, CA. Morgan Kaufmann. Hazen,G. B. (1992). Stochastictrees: Anewtechniquefor temporal medical decision modeling. MedicalDecision Making,12:163-178. I wouldlike to thank Peter Szolovits for advising this project, and MikeWellmanand Steve Pauker for many helpful discussions. This research was supportedby the National Institutes of Health Grant No. R01LM04493 Heyman, D. P. and Sobel, M. J. (1984). Stochastic Models fromthe National Library of Medicine,and by tile USAF in Operations Research: Stochastic Optimization, RomeLaboratory and DARPA under contract F30602-91- volume2. McGraw-Hill. C-0018. Hollenberg, J.P. (1984). Markovcycle trees: A new representation for complexMarkovprocesses. Medical References Decision Making,4(4). Abstract from the Sixth Annual Berzuini, C., Bellazzi, R., and Quaglini, S. (1989). Meetingof the Society for MedicalDecisionMaking. Temporalreasoningwith probabilities. In Proceedingsof Howard,R.A. (1971). DynamicProbabilistic Systems, the Fifth Workshop on Uncertainty in Artificial volume1 &2. John Wileyand Sons, NewYork. Intelligence, pages14-21. Associationfor Uncertaintyin Leong, T.-Y. (1993). Dynamicdecision modeling Artificial Intelligence. medicine: A critique of existing formalisms. In Dean,T., Kaelbling,L. P., Kirman,J., andNicholson,A. Proceedings of the Seventeenth Annual Symposiumon (1993a). Deliberation scheduling for time-critical ComputerApplications in MedicalCare, pages 478-484. sequential decision making. In Heckerman, D. and IEEE. Mamdami,A., editors, Uncertainty in Artificial Intelligence: Proceedingsof the Ninth Conference,pages Meyer, B. (1990). Introduction to the Theory of Programming Languages.Prentice Hall. 309-316, San Mateo, CA. MorganKaufmann. Dean,T., Kaelbling,L. P., Kirman,J., and Nicholson,A. Shachter, R. D. and Peot, M. A. (1992). Decisionmaking using probabilistic inference methods. In Dubois, D., (1993b).Planningwith deadlinesin stochastic domains. Wellman, M.P., D’Ambrosio, B.D., and Smets, P., Proceedings of the Eleventh National Conference on editors, Uncertainty in Artificial Intelligence:Proceedings Artificial Intelligence, pages574-579. of the Eighth Conference, pages 276-283. Morgan Dean, T., Kriman, J., and K,~azawa, K. (1992). Kaufmann. Probabilistic networkrepresentations of continuous-time stochastic processes for applications in planning and Tatman, J.A. and Shachter, R.D. (1990). Dynamic programming and influence diagrams. IEEETransactions control. In Proceedings of the First International on Systems, Man,and Cybernetics, 20(2):365-379. Conferenceon AI PlanningSystems. 161