From: AAAI Technical Report SS-92-01. Compilation copyright © 1992, AAAI (www.aaai.org). All rights reserved. Predit: a TemporalPredictive Framework for Scheduling Systems E.Paolucci, E.Patriarca, M.Sem,G.Gini Politecnico di Milano Piazza Leonardoda Vinci 32 20133 Milano - Italy Paolucci@ipmelLpolimi.it Abstract Schedulingcan be formalizedas a ConstraintSatisfaction Problem(CSP). Withinthis frameworkactivities belongingto a plan are interconnected via temporal constraints that accountfor slack amongthem.Temporal representationmustincludemethodsfor constraints propagationand provide a logic for symbolicand numerical deductions. In this paper we describe a support frameworkfor opportunisticreasoningin constraint directed scheduling. In order to focusthe attention of an incremental scheduler on critical problemaspects, somediscrete temporalindexesare presented.Theyare also useful for the predictionof the degreeof resourcescontention. Thepredictive methodexpressed through our indexes can be seen as a Knowledge Sourcefor an opportunistic schedulerwith a blackboardarchitecture. 1. Formalization of scheduling problem and strategies for its solution Schedulingcan be formalizedas a ConstraintSatisfaction Problem(CSP)[Kengand Yun, 1989]. This approachis concernedwith the assignmentof values to variablessubjectto a set of constraints. In scheduling variablesare constituted by activities start times and fromresources allocation; for this reason wehaveto deal explicitely with twotypes of constraints:temporal relations amongtasks andresourcescapacity[Fox, 86]. In our approachweassumethat wehavea set of plans to be scheduled,wherea plan is defined as a partial orderingof activities. Eachactivity mayrequire oneor more resources and for each of them there can be alternative choices. Besideresources capacity can be used contemporarly by different tasks; for the sake of simplicity wewill assumeall resources with unary capacity. Weare concernedwith the issue of howit’s possible to focus the attention of an incrementalscheduleron the mostcritical schedulingchoicesin order to evaluate whichare the mostcritical points, whichdecisionsseem to be the mostpromisingin reducingsearch complexity andimprovingquality of resulting schedule. Ourstrategy is to identify the most"solvable"aspects of the problemthroughthe evaluationof the degreeof interaction existing among activities belongingto different orders. Theaimis to reducethe number of steps requiredto obtain a solution. Thenecessity to overcome the limits of partial decomposition approach,such as order-basedand resourcebased decompositions, has led us towards an event-basedperspective whit chronologically-grouped information. This basic searchstrategy is realized throughmost-constrainedandleast-impactpolicies. Everystep is divided into twoparts: first the rnost-constrainedpoliey selects dynamicallyon whichagent must be focused scheduling attention; then, the least-impactpolicy choosesfor that agent a value whoseimpact on the rest of the non-scheduled agentsis as small as possibile. Thegoal is the identification of critical activitiesthat heavilyrely on the possessionof highly contendedtemporalintervals or resourcesbecauseof intra-order andinter-order interactions (look-aheadstrategy). This two policies neednumericindexes which,analyzing the particular structure of a problem,are able to measurethe interaction among activities andresources in terms of variable looseness and value goodness [Sadehand Fox, 88]. Scheduling is an NP-hard problem and methodsrequiredfor its solution mustface this complexity.In our research wedecidedto focusour attention on contribution in schedulingcomingfromAI, andparticularly on opportunistic reasoning[Hayes-Roth,79]. 1. This workhas been partially supportedby CNRPRF Project and Maurogrants 140 Variable looseness is the measureof howconstrained is a resource or an activity; value goodnessmeasures whichvariable value, among all the feasible ones, gives the least impact (i.e. a sort of maximum slack) availability of feasible (and good)valuesfor non-scheduledvariables. Weidentified somenumericindexesthat contain informationrequiredto realize an event-basedpolicy: these indexesare usefulfor different reasons: the>, makepossibleto point out critical resources andactivities; [] they identify "islandof certainty" that will be a part of problemsolution; [] an internal bound(INT), whichrepresents the minimuntime interval whichmust separate the endof the first lapse fromthe beginof the second of tworelated lapses; [] an external bound0EXT),whichrepresents the maximum time interval fromthe begin of the first lapse to the endof the second. n they give informationaboutactivities start times that havethe least impacton non-scheduled activities. This behaviouris a sort of "oppommisticreasoning" [Hayes-Roth,79]: this term has been used to characterize a problem-solvingprocess wherereasoning is consistently directed towardsthose actions that appear mostpromisingfor solving a problem. Throughthese two parameters it’s possible to model any temporalrelation in a schedulingproblem.Theyare simplerthan thirteen Allen’sprimitiverelations; moreover, INTand EXTimprovegreatly the efficiency of numerictemporalreasoning, that is instead a limit in Allen’sprimitive. Ourpredictive approach,used together with an opportunistic reasoning,is also usefulto detect unsatisfiable CSPsas soon as possible, simply by analyzing the indexes wedefined. In this sense the system can be viewedas a Knowledge Source in a blackboardarchitecture, whichassumesresponsibility for preventive analysisof activities interactionsandfor the detection of prospectivebottlenecks. 3. The Predit indexes Temporal relation constraints are usedto describepartial orderingsamong activities as providedby the process planningstep. 2. The predictive approach: basic assumptions Themaingoal of our research wasto provide a simple but completeinference mechanism to support scheduling, workingin a discrete time domain.This mechanism is basedon someindexesand is designedto performan a-priori guidancefor search in schedulingdomain.We kept a particular attention on the efficiencyandon the speed of such a mechanism,becausewerealized that such propertiesare necessaryin schedulingsystemsfor real applicativeenvironments like, for instance, manufacturing ones. For this reasonwedecideto considera discrete representationof time instead of a continous one. Wehaveto schedulea set of activities (A1,A2 ..... AN). Let Ik be the timeinterval associatedwithAk.Activities are connectedby a set of temporalrelation constraints, thereby forming a TCG.Weview TCGsas undirected graphs. AnArc in a TCGindicates the presence of a temporalrelation betweentwo intervals representedby the coupleINT-EXT (Fig. 1). Wewill refer to the graphdefinedby these constraints, for a given CSP,as the CSP’sTemporalConstraint Graphs O’CG). Ai I I ETi STi Aj Ourindexesare basedon the constraints analysis (and on the propagation of the temporal ones) and on particular representationof existingtime relations. In terms of constraints analysis wedifferentiate betweenrestrictions and preferences[Fox, 86]. Temporal preferencesare representedthroughutility functions definedon activity start timesthat mapspossiblevalues onto utility levels rangingfrom0 to 1. Moreover in our analysis we consider the existence of intra-order (among activities belongingto an order) andinter-order constraints (among activities belongingto different orders). Themodeladoptatedin representing time and in reasoningabout temporalrelation is basedon the concept of lapse, that is definedas the periodof time associated withan activity. In a temporalaxis a lapseis represented by two temporalparameters,namelystart time and end time. Relationsbetweendifferent lapses are expressed by twoparameters[Paolucci, 90]: 141 Figure1 Additionallythere are capacityconstraints limiting the use of eachresourceto onlyoneactivity at a time. The next examplepresents a simple case of a TCGcomposedof two orders. In order to providea predictivesupportfor opportunistic schedulersoperating in a discrete time domainwe have consideredinteractions amongactivities caused by temporalrelations. Thefirst issue wefacedwasto detect as soonas possible duringthe schedulingthe possiblearising of conflicts dueto interactions among activities. For tiffs issue wedefinedan index called D.ggr.~(CD),whichmeasuresthe howtight is the link existing betweentwo generic activities Ai and Aj connectedby a temporalrelation constraint (representedby INT-EXT) in a contraint graph. To sumup, the CDindex detects (following the mostconstrainedpolicy)the mostcritical activities withrespeet to intra-order temporalrelations (expressedby INTand EXT)and to temporal windows(expressed activity start andendtime). The second index, called Preferential Start Time (PST), is a local measureof value goodnessand globally, a measure of variableloosenessfor activities start times. It helps in choosingamongall admissiblestart [1] Di + Dj +/Nr ~ ~j - sri times the one that minimizesfuture conflicts. It is calculated betweeneachpair of activities connectedin [2] Di + Dj + INT ~ the TCG (i.e. AiandAj) andit dependson the start time of the first activity(i.e. sti). Thefirst inequalityverifies that time interval composed of activities durationsandInternal Bound is includedin The main goal of PSTindex was to introduce some maximum temporal windowallowed by Aj latest end estimationrule for activity start timesin orderto identtimeandAiearliest start time. ify the least impactvalues arising from intra-order interactions. Thesecondinequalitycontrols that the sametime interval doesn’tviolate ExternalBoundtemporalconstraint. PSI’ indexis computed for everyactivity start time sti These inequalities lead to define the CDformula evaluatedbetweenearliest start time (STi), or value througha multiplicationof their members: allowedby INT-EXT, andlatest start time (ETi-Di), valueallowedby INT-EXT, increasingsti with the fixed timeunit. 2 ( Di* +( Di+ PSTis expressedby the ratio: ETi- INT) S~ ) [31 CDij- EXT Temporalrelations betweenintervals maysimply be expressedby using potential inequalities associated with the boundsof intervals suchas : O <CO~i < l [4] PSTij( sti ) Dk = Akduration INT= internal boundbetween Ai and Aj "intii( sti ) EXT- Di - Dj - INT o <_ PST O(st~)<_1 STk= Akstart time EXT= external boundbetweenAi and Aj ETk= Ak end time where: = intij(s~ ) = relative internal bound The Constraint Degreeis calculated on the notion of slack betweentwoactivities tied by temporallinks. = EXT-Di-Dj-INT = maximum slack between the activities = CDij = 1 meansthat Ai allows no slack to Aj (mostconstrained) Thenumeratoris calculatedfor sti valuesfromEarliest Start Timeto the maximum allowed by temporal constraints, increasingeachtimesti witha chosentimeunit. It maybe also consideredas "actual" slack betweenthe twoactivites corresponding to sti value. = CDij= 0 meansthat Ai allows maximum slack to Aj. The CDcomputational algorithm considers all connectedactivities fromthe beginningto the endof the graph.Therefore,for endingactivities weset CDindex to zero (endingactivities are not constrained,with temporal relation, withanyother activity in the graph). The validity of CDindex is preservedby a previous optimizationprocedurein orderto adjust activities temporal windows cutting out start time values that can never be involved in CDcomputation(the same is madefor other indexes). 142 Therefore, the denominatormaybe viewedas the maximum slack betweenthe two activities. Thecloser is PSTijvalueto one, the greater is the slack betweenAi andAj. Therefore,for a generic activity PSTmeasures for eachadmissiblestart time its goodnessandlikelihoodto minimizefuture schedulingconflicts. To computeactivity Indlvidnal Demandfor resources, wehave combinedthe value goodnessof every start time (expressedby PST)with the activities durations. Moreover,as assumedin [Sadeh-Fox,88], an activity Aican use a resourceRjif Ai is active at timet andAi usesRj at timet to fulfill its resourcerequirement. Fromeach PSTgraph we achieve an Individual Demandgraph (whosevalues are expressed by ID index) for each activity, expandingPSTvalues with a lapse equalto the activity duration andaddingall valuesin function of time. Weobtain a histogramrepresenting activity resource demand in function of time. Individual Demandvalues are combinedto measure resource Ag~egateDemand(AD), always in function of time. ADshowswhenresource competitionis particularly high andwhichare activities that heavilyrely in the possessionof these resources. ADvaluesmustbe tightly evaluatedin function of time becausetemporal consa’aint propagationdoesn’t allow for any resource preference(as explainedbefore, weassumeall resources with unary capacity). Therefore, ADindex can estimate the amountof contention for each resource overtemporalaxis but only as a functionof start time. Moreover,it’s easy to improvethis approachrepresenting, for example,resourcespreferenceswith utility functionsand propagatingthese resourcesreservations through the TCGgraph. AD Resource R3 AD 10 - 10 4" "4 2- -2 0 -- , 10 , 20 , 30 , 40 Time , rio , 70 ¯ 00 0 60 Figure3 Psi a7 Order 2 Pit. j 0.3. ’ ’ ’0;2:1 ..’ ...................................... O;2~IF’ O.~, ................................................ 0,25¯ "" ~ 0,2t 0,21 0,16 0.16 0.100’2 ,’ Figure2 showsa simpleexamplein order to illustrate our graphic results concerningtemporaldiscrete indexes presented above. 0.1" 0,05- 10 12 1.4 16 10 20 22 24 23 20 -qO Time Figure4 Next results are concernedwith the reasonablesteps that an OpportunisticSchedulershouldachievein order to producethe final Ganttchart. Order 2 lpo A1 0,25 A2 0,2 A3 0,2 A4 0.182 A5 0 Table1: CDvaluesfor order 1 activities A6 0,25 Figure2 Thetemporalconstraint associated at each linker is equalfor all coupleof activity andit is expressedby INT=0and EXT--40.However,these values maybe optimizedas described before. Start Timeand EndTime are expressed by numbers aboveeach activity andthe sameis madefor requested resources.For the sake of simplicity,in this example we havenot introducedpreferentialstart times(so activity start timesare equallypreferred). 143 A7 0,25 A8 0 Table2: CDvaluesfor order2 activities Amongthe aggregate demands,the most highly contendedresourceis R3(fig. 3), required by A2,A3,A4, A7;the nextactivities wewill focusour attention onare A4, A3and A7(becauseA2has an alternative in R1). Takinga look at the CDindexesof order 1 andorder 2, A7appearsto be the mostconstrainedactivity because of its highest CDvalue. Now,A7PSTgraph (fig. 4) presents a maximum for t=20 and scheduling A7with st=20 wecan assign the resourceR1to the activity A2 at the samestart time. predictive supportin an opportunisticschedulingsystem. The same considerations based on temporal indexes evaluationallowthe identification ofotheractivitypreferential start timesleadingto the Ganttchart presented belowin fig. 5. Weimplementedthis model in a MS-DOS environment with a particular attention towardsspeedperformances. Ourexperimentsindicate that our approachis successful in supportingopporumisticscheduling.This system is very efficient (it takes fewsecondsto calculate indexesin non-trivial real problems). Ourmodelseemsto be highly appropriatefor problems wherethe costs of backtracking is highbecauseit’s able to point out scheduling decisions that will minimize intra-orderandinter-orderconflicts. It increasessignificantly the performances of an opportunisticscheduler, makingit possible to introduce such a tools in real applications. GANTT CHART order I o 5 to ,s ~o 2~ so 35 40 ,5 ~o 55 GANTT CHART order 2 6n iS5 ?0 Moreoverthe policies used by Predit to control the solution search(mustconstrainedandleast impact)can be used also in dynamicmanufacturingenvironments. Weare developing our researchin this sense, also trying to support reactive scheduling and to managemultiagent productioncontrol systems. [--rf---] 0 5 I0 15 20 I~5 30 35 40 4 l~ 50 61~ 60 65 70 ’rtme Figure5 Thequality of a scheduleis basedon the capability of the schedulerto satisfy a set of performance measures. References [Fox, 1896], FoxM., Observationon the role of constraints in problemsolving, ProceedingsSixth Canadian Conferenceon Al, Montreal,1986. Moreover,a satisfiable schedule is always a com[Hayes-Roth, 79], Hayes-Roth B. et al., Modeling planning as an incrementalopportunisticproblem,Proceedpromisebetweenthe attempt to meet performancerequiredandthe necessityto respect all its constraints: ings 3rd lJCAl, Tokio, 1979. schedulequality mirrors this trade-off. Eachset of [Kengand Yun,89], Keng.N. and YunD., A planning organizationalconstraints has its effects on final proscheduling methodologyfor the constrained resource duetion schedulesand, followingthe CSPformulation, if wechangethe constraints the solution will change problem", ProceedingsIJCA11989,pp. 998-1002. tOO. [Paolucci,90], Paolucciet. al., "CRONOS-III: requirements for a knowledge-based scheduling tool covering In order to improveschedulequality, our research is a broad class of production environments",in Expert focusingon the evaluationof whichimpactmighthave an unexpectedevent on the resulting solution. PREDIT systemsin engineering,G.Gottloband W.Nejdl(eds.), approach through theevaluation ofdiscrete temporal Springer-Verlag,1990. indexes produces relatively accurate early predictions [Sadehand Fox, 88], SadehN. and FoxM., Preference ofactivities behaviour assoon asPREDIT receives their propagationin temporal capacity constraint graphs, changes andaslongasconstraints remain constant Technical report CMU-CS-88-193. during indexes computation. Theability toreact to changes thatoccur indynamic environments providing a feasible solution ina sufficiently short time isvery important cxpecially inmanufacturing scheduling domain. 4. Concluding remarks Theapproachwepresentedin this paperconstitutes the basis for integrating an event-basedmechanism and a 144