Local Search for Planning and Scheduling

Withtheincreasingdeploymentofplanningandschedulingsystems,developers oftenhavetodealwithverylargesearchspaces,real-timeperformancedemands, anddynamicenvironments. Completere?nementmethodsdonotscalewell,- kinglocalsearchmethodstheonlypracticalalternative. Adynamicenvironment alsopromotestheapplicationoflocalsearch,thesearchheuristicsnotnormally beinga?ectedbymodi?cationsofthesearchspace. Furthermore,localsearchis wellsuitedforanytimerequirementsbecausetheoptimizationgoalisimproved iteratively. Suchadvantagesareo?setbytheincompletenessofmostlocalsearch methods,whichmakesitimpossibletoprovetheinconsistencyoroptimalityof thesolutionsgenerated. Popularlocalsearchapproachesincludeevolutionary- gorithms,simulatedannealing,tabusearch,min-con?icts,GSAT,andWalksat. The?rstarticleinthisbook aninvitedcontributionbyStefanVoß givesan overviewofthesemethods. ThebookisbasedonthecontributionstotheWorkshoponLocalSearchfor Planning&Scheduling,heldonAugust21,2000atthe14thEuropeanCon- renceonArti?cialIntelligence(ECAI2000)inBerlin,Germany. Theworkshop broughttogetherresearchersfromtheplanningandschedulingcommunitiesto explorethesetopicswithrespecttolocalsearchprocedures. Aftertheworkshop, asecondreviewprocessresultedinthecontributionstothepresentvolume. Voß soverviewisfollowedbytwoarticles,byHamiezandHaoandGerevini andSerina,onspeci?c classical combinatorialsearchproblems. Thearticleby HamiezandHaoaddressestheproblemofsports-leaguescheduling,presenting results achieved by a tabu search method based on a neighborhood of value swaps. GereviniandSerina sarticleaddressesthetopicthatdominatestherest ofthebook:actionplanning. Itbuildsontheirpreviousworkonlocalsearch onplanninggraphs,presentinganewsearchguidanceheuristicwithdynamic parametertuning. Thenextsetofarticlesdealwithplanningsystemsthatareabletoinc- porateresourcereasoning. The?rstarticle,ofwhichIamtheauthor,makesit clearwhyconventionalplanningsystemscannotproperlyhandleplanningwith resourcesandgivesanoverviewoftheconstraint-basedExcaliburagent spl- ningsystem,whichdoesnothavetheserestrictions. Thenextthreearticlesare aboutNASAJPL sASPEN/CASPERsystem. The?rstone byChien,Knight, andRabideau focusesonthereplanningcapabilitiesoflocalsearchmethods, presentingtwoempiricalstudiesinwhichacontinuousplanningprocessclearly outperformsarestartstrategy. Thenextarticle,byEngelhardtandChien,shows howlearningcanbeusedtospeedupthesearchforaplan. Thegoalisto?nda setofsearchheuristicsthatguidethesearchaswellaspossible. Thelastarticle inthisblock byKnight,Rabideau,andChien proposesanddemonstrates, a technique for aggregating single search moves so that distant states can be reachedmoreeasily. VI Preface Thelastthreearticlesinthisbookaddresstopicsthatarenotdirectlyrelated tolocalsearch,butthedescribedmethodsmakeverylocaldecisionsduringthe search. RefanidisandVlahavasdescribeextensionstotheGRTplanner,e. g. ,a hill-climbingstrategyforactionselection. Theextensionsresultinmuchbetter performancethanwiththeoriginalGRTplanner. Thesecondarticle byO- india, Sebastia, and Marzal presents a planning algorithm that successively re?nes a start graph by di?erent phases, e. g. , a phase to guarantee comp- teness. Inthelastarticle,HiraishiandMizoguchipresentasearchmethodfor constructingaroutemap. Constraintswithrespecttomemoryandtimecanbe incorporatedintothesearchprocess. Iwishtoexpressmygratitudetothemembersoftheprogramcommittee, whoactedasreviewersfortheworkshopandthisvolume. Iwouldalsoliketo thank all those who helped to make this workshop a success including, of course,theparticipantsandtheauthorsofpapersinthisvolume. June2001 AlexanderNareyek WorkshopChair ProgramCommittee EmileH. L. Aarts PhilipsResearch Jos eLuisAmbite Univ. ofSouthernCalifornia BlaiBonet UniversityofCalifornia RonenI. Brafman Ben-GurionUniversity SteveChien NASAJPL AndrewDavenport IBMT. J. Watson AlfonsoGerevini Universit`adiBrescia HolgerH. Hoos Univ. ofBritishColumbia AlexanderNareyek GMDFIRST AngeloOddi IP-CNR Mar ?aC. Ri? Univ. T ec. Fed. SantaMar

Includes supplementary material: sn.pub/extras

Klappentext

Thenextsetofarticlesdealwithplanningsystemsthatareabletoinc- porateresourcereasoning. The?rstarticle,ofwhichIamtheauthor,makesit clearwhyconventionalplanningsystemscannotproperlyhandleplanningwith resourcesandgivesanoverviewoftheconstraint-basedExcaliburagent spl- ningsystem,whichdoesnothavetheserestrictions. Thenextthreearticlesare aboutNASAJPL sASPEN/CASPERsystem. The?rstone byChien,Knight, andRabideaüfocusesonthereplanningcapabilitiesoflocalsearchmethods, presentingtwoempiricalstudiesinwhichacontinuousplanningprocessclearly outperformsarestartstrategy. Thenextarticle,byEngelhardtandChien,shows howlearningcanbeusedtospeedupthesearchforaplan. Thegoalisto?nda setofsearchheuristicsthatguidethesearchaswellaspossible. Thelastarticle inthisblock byKnight,Rabideau,andChien proposesanddemonstrates, a technique for aggregating single search moves so that distant states can be reachedmoreeasily. VI Preface Thelastthreearticlesinthisbookaddresstopicsthatarenotdirectlyrelated tolocalsearch,butthedescribedmethodsmakeverylocaldecisionsduringthe search. RefanidisandVlahavasdescribeextensionstotheGRTplanner,e. g. ,a hill-climbingstrategyforactionselection. Theextensionsresultinmuchbetter performancethanwiththeoriginalGRTplanner. Thesecondarticle byO- india, Sebastia, and Marzal presents a planning algorithm that successively re?nes a start graph by di?erent phases, e. g. , a phase to guarantee comp- teness. Inthelastarticle,HiraishiandMizoguchipresentasearchmethodfor constructingaroutemap. Constraintswithrespecttomemoryandtimecanbe incorporatedintothesearchprocess. Iwishtoexpressmygratitudetothemembersoftheprogramcommittee, whoactedasreviewersfortheworkshopandthisvolume. Iwouldalsoliketo thank all those who helped to make this workshop a success including, of course,theparticipantsandtheauthorsofpapersinthisvolume. June2001 AlexanderNareyek WorkshopChair ProgramCommittee EmileH. L. Aarts PhilipsResearch Jos eLuisAmbite Univ. ofSouthernCalifornia BlaiBonet Universityof

Inhalt
Invited Paper.- Meta-heuristics: The State of the Art.- Combinatorial Optimization.- Solving the Sports League Scheduling Problem with Tabu Search.- Lagrange Multipliers for Local Search on Planning Graphs.- Planning with Resources.- Beyond the Plan-Length Criterion.- An Empirical Evaluation of the Effectiveness of Local Search for Replanning.- Board-Laying Techniques Improve Local Search in Mixed Planning and Scheduling.- Empirical Evaluation of Local Search Methods for Adapting Planning Policies in a Stochastic Environment.- Related Approaches.- The GRT Planner: New Results.- Incremental Local Search for Planning Problems.- Map Drawing Based on a Resource-Constrained Search for a Navigation System.