newsletter of the european network of excellence in ai planning...

Newsletter of theEuropean Network

of Excellencein

AI Planning

Issue No.5

http://www.planet-noe.org

ISSN 1610-0204

The PLANET Newsletter 3

Issue No.5

EDITORIAL

Welcome to PLANET NEWS!

Welcome to Issue No 5 !

PLANET’s Industrial Information Days provide a fo-rum where successful industrial applications of Plan-ning and Scheduling technology are presented andpromising future exploitation is discussed. Represen-tatives from industry and academia use these eventsto promote the transfer of the technology and to pushmutual exchange and co-operation for the benefit ofboth.The recent information day took place in Rome inNovember. You’ll find a broad selection of contribu-tions from industrial speakers in the first section ofthis issue.Another major event was the Second PLANET In-ternational Summer School on AI Planning heldin Halkidiki, Greece in September. Students ofthe school were encouraged to present their Ph.D.projects in a separate poster session. This session wasvery successful and showed a variety of interestingaspects and new approaches. A report on the schooland several extended abstracts of the poster presenta-tions are included.The PLANFORM project aimed at the developmentof an Open Environment for Building Planners. Itwas carried out jointly at the universities of Hudder-

sfield, Salford, and Durham in the United Kingdom.You’ll find a final report on this project at page 38.The second PLANET Gap-bridging Seminar washeld in co-location with the UK PLANSIG meetingin November in Delft. A report on this seminar isprovided by Tim Grant.Announcements, job offers and information on forth-coming events, in particular those to be held in con-junction with ICAPS in June 2003 in Trento, and aninvitation to participate in the Supply Chain TradingCompetition 2003 can be found in the final section.Wishing you a successful and Happy New Year 2003,

Susanne BiundoBernd Schattenberg

Editors:

Susanne Biundo Network Coordinator, Dept. ofArtificial Intelligence, University of Ulm, Germany,[email protected]

Bernd Schattenberg Network Administrator,Dept. of Artificial Intelligence, University of Ulm,Germany, [email protected]

�

4 The PLANET Newsletter

Table of Contents

Editorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Industrial Information Day RomeS. de Givry, L. Jeannin, F. Josset, J. Mattioli,N. Museux, and P. SaveantThe THALES constraint programming frameworkfor hard and soft real-time applications . . . . . . . . . . . 5M. SanseverinoCOMPETE a Common Platform for ExtendedProject Management . . . . . . . . . . . . . . . . . . . . . . . . . . . 8J.E. SpraggIn Defence of Reactive Scheduling . . . . . . . . . . . . . 14B. DrabbleAdvanced Scheduling and Optimisation: Cutting theCosts of Manufacturing. . . . . . . . . . . . . . . . . . . . . . . .17

2nd PLANET Summer SchoolL. Compagna, G. Cortellessa, A. Farinelli, and N. Po-licella2nd International PLANET Summer School . . . . . 22A. Farinelli, G. Grisetti, L. Iocchi, D. Nardi, andR. RosatiGeneration and Execution of Partially Correct Plansin Dynamic Environments . . . . . . . . . . . . . . . . . . . . . 25F. McNeill, A. Bundy, M. SchorlemmerDynamic Ontology Refinement . . . . . . . . . . . . . . . . .27G. Cortellessa, N. Policella, A. Cesta, and A. OddiMEXAR: Integrated AI Technologies to SupportMARS EXPRESS Mission Planning . . . . . . . . . . . . . 29M. Baioletti, A. Milani, and V. PoggioniRDPPlan: an Extension of DPPlan for Planning withInterval Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31N.E. RichardsonExtending Operator Induction to Provide FullMethod Sets for Hierarchical Planning Domains . 33

O. Sapena and E. OnaindıaThe SimPlanner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

A. PolleresAnswer Set Planning with DLV

�

. . . . . . . . . . . . . . .36

Project ReportT.L. McCluskey, M. Fox, and R. AylettPlanform: An Open Environment for Building Plan-ners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

Gap-bridging SeminarT. GrantSecond PLANET Gap-Bridging Seminar . . . . . . . . 46

Announcements and Network NewsICAPS 2003. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .51

3rd PLANET International Summer School . . . . . . 52

ICAPS 2003 – Doctoral Consortium . . . . . . . . . . . . 52

ICAPS 2003 – Workshop Program. . . . . . . . . . . . . .53

ICAPS 2003 – Tutorials . . . . . . . . . . . . . . . . . . . . . . . 55

R. ArunachalamInvitation to Participate in TAC’03 - A Supply ChainTrading Competition . . . . . . . . . . . . . . . . . . . . . . . . . . 59

Postdoctoral and Doctoral Positions at Australian Na-tional ICT Center . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

InformationThe Members of PLANET . . . . . . . . . . . . . . . . . . . . 61

�



ARTICLE

The THALES Constraint Programming Framework for Hard andSoft Real-Time Applications

Authors: S. de Givry, L. Jeannin, F. Josset, J. Mattioli, N. Museux, and P. Saveant

This position paper presents the constraint tech-nology that has been developed at THALES since1997 for introducing Constraint Programming (CP)in THALES operational systems1 2. These systemsinvolve combinatorial optimization problems such asplanning and scheduling problems that can be ex-pressed with finite-domain variables and constraints.Typical examples of THALES systems concern su-pervision, for weapon allocation, radar configuration,weapon deployment and aircraft sequencing. Allthese systems are subject to specific requirementscoming from the operational constraints of embed-ded real-time systems and from the strategic contextof Defense applications:

1. The system involves several functions/tasks suchas situation assessment, resource management, vi-sualization, etc.; each task is periodical and theperiod can be much shorter than a second;

2. There is a memory space limit (a few megabytes);

3. The system has to be supported for a long time,typically over 20 years for Defense applications,including several retrofitting (functional and plat-form evolutions);

4. The system can be reused and modified for build-ing a specific system for a new client (productline);

5. The development of the system must be made andmastered in house for reasons of confidentialityand market protection.

The CP paradigm partially meets these requirements.A constraint model has modularity properties, i.e.adding/removing a constraint is easy, which enablesan incremental development process, reducing the

development time and effort. CP solvers provide effi-cient algorithms through the use of global constraints.The declarative nature of CP enables the programmerto focus on the application requirements rather thanon debugging low-level programming errors. Vali-dated CP models can be reused in a product line ap-proach.Unfortunately, off-the-shelf CP solvers do not pro-vide any guarantee on time and space usage. Theclassical backtracking search algorithm used in CPdoes not take into account any time contract. Re-cently an effort was made to provide better search al-gorithms in CP solvers, for instance in [1, 11, 14], butwithout any explicit time contract. Our aim is to ex-tend CP solver with new search features that wouldkeep the same nice software engineering propertiesas for modeling. This led to develop a high-levellanguage for designing search algorithms. This ap-proach allows proposing a set of search primitives ontop of the real-time finite-domain constraint solverEclair c

�

[13]. The resulting search algorithms arebased on partial search methods and take into accountthe time contract explicitly. Such algorithms can takeadvantage better of platform evolutions.

Eclair offers time and space guarantees. Deadlinesare guaranteed by the operating system alarm andEclair is able to restore a coherent state after an in-terruption in order to deliver a valid solution, or justa partial solution (when not all variables are instan-tiated). The memory allocation for the constraintsis static: a global constraint model is built once andonly parts of the model are made active and used ata given cyclical call. The memory consumed duringthe search is limited by using only restricted depth-first search or restricted best-first search.

1This work was partially funded by the EOLE project[7].2A former version of this paper appeared in [5].


Partial search methods are anytime algorithms [17]based on tree search methods having better qualityprofiles than the classical backtracking search algo-rithm. The main idea is to apply some arbitrary limitson the nodes visited in the tree search3 , depending onthe behavior of the heuristics and on the remainingcomputation time. We distinguish four approaches:the iterative weakening methods (e.g. [8]), the real-time search methods (e.g. [10]), the iterative sam-pling methods (e.g. [6]) and the interleaving methods(e.g. [12]). These methods use one or several searchschemes 4. The practical complexity of the searchcan be increasing, self-adjusting, or stable. In [3], wepropose the notion of parameterized search appliedto one search scheme. The parameters of the searchlimits are given explicitly. We can tune the degreeof incompleteness of the search by varying the valuesof the parameters. A tuning policy indicates the rele-vant values of the parameters for different time con-tracts. In [4], we integrate the parameterized searchapproach into a hybridization scheme to express par-tial search based on several search schemes. The hy-bridization scheme is a sequence or an interleavingof parameterized searches. The searches can coop-erate by exchanging solutions. A time-sharing pol-icy specifies how to distribute the time contract to thesearches.Our constraint optimization framework is calledToOLS c

�

(Templates Of On-Line Search). A searchalgorithm is expressed in ToOLS as the conjunctionof four distinct components:

� A set of heuristics to rank every choice;

� A set of primitives to express a search schemeindependent of any time limit; it is composedby predefined choice points and combinations ofchoice points as in the OPL language [9];

� A set of primitives to express the search limits thatdepend on the current node, the current path orthe current sub-tree; the resulting parameterized

search algorithm controls the size of the exploredsearch tree defined by one search scheme;

� A temporal strategy defined by a hybridizationscheme, i.e. a cooperation of several parameter-ized searches, dealing with time allocation andselecting the tuning strategy of the parameters(static tuning, iterative tuning or adaptive tuning).

A template of search defines an abstract componentof a search algorithm that can be reused to speed upthe development process of customized partial searchalgorithms. This framework makes it easier to trynew combinations of search limits and new temporalstrategies.Experiments on the weapon allocation problem showthat partial search algorithms significantly improvethe solution quality compared to a traditional ap-proach [3] and also demonstrates the gain in devel-opment time of new customized search algorithms.The code is clearer and more concise when using thesearch primitives. As the main result, our CP frame-work has been integrated in an operational on-boardhard real-time system of THALES.The hybridization scheme is a way to define spe-cific local search methods, such as large neighbor-hood search based on a sequence of partial searchesin different neighborhoods [15, 16]. Pure local searchmethods could also be introduced in our frameworkas a black-box used by the hybridization scheme. Thetemporal control could be enhanced by an on-linelearning mechanism, using the fact that similar prob-lems are repeatedly solved in a real-time system. [2]gave the base for this mechanism.

Bibliography

[1] Beldiceanu, N., Bourreau, E., Simonis, H., andRivreau, D. Introduction de metaheuristiquesdans CHIP. In Proc. of MIC-98., 1998.

[2] Crawford, Lara S., Fromherz, Markus P.J.,Guettier, Christophe, Shang, Yi. A Framework

3This description of partial search is compatible with the depth-first search principle. In [14], partial search methods are based onthe order of node exploration, which is memory consuming.

4A search scheme is a procedure that describes a search tree. For example, a combination of choice points.



for On-line Adaptive Control of Problem Solv-ing. In Proc. of CP-2001 workshop on On-Line combinatorial problem solving and Con-straint Programming, Paphos, Cyprus, Decem-ber 2001.

[3] de Givry, Simon, Saveant, Pierre, Jourdan,Jean. Optimization combinatoire en tempslimite: Depth first branch and bound adaptatif.In Proc. of JFPLC-99, pages 161-178, Lyon,France, 1999 (in french).

[4] de Givry, S., Hamadi, Y., Mattioli, J., Lemaıtre,M., Verfaillie, G., Aggoun, A., Gouachi, I.,Benoist, T., Bourreau, E., Laburthe, F., David,P., Loudni, S., Bourgault, S. Towards an on-lineoptimization framework. In Proc. of CP-2001workshop on On-Line combinatorial problemsolving and Constraint Programming, Paphos,Cyprus, December 2001. http://www.lcr.

thomson-csf.com/projects/www_eole/

workshop/olcp01-eole.ps

[5] de Givry, S., Gerard, P., Jeannin, L., Mattioli, J.,Museux, N., Saveant, P. A constraint optimiza-tion framework for real-time applications. InProc. of AIPS 2002 workshop on On-line Plan-ning and Scheduling, Toulouse, France, April2002.

[6] Gomes, C., Selman, B., Kautz, H. BoostingCombinatorial Search Through Randomization.In Proc. of the 15th National Conference on Ar-tificial Intelligence (AAAI-98), pages 431–437,Madison, WI, USA, 1998.

[7] RNRT EOLE project. http://www.lcr.

thomson-csf.com/projects/www_eole

[8] Harvey, William D., Ginsberg, Matthew L. Lim-ited discrepancy search. In Proc. of IJCAI-95,pages 607-613, Montreal, Canada, 1995.

[9] Van Hentenryck, P. OPL: The Optimiza-tion Programming Language. The MITPress,Cambridge,Mass., 1999.

[10] Korf, Richard E. Real-time heuristic search. Ar-tificial Intelligence, 42:189-211, 1990.

[11] Laburthe, Francois. SaLSA: a language forsearch algorithms. In Proc. of CP-98, pages310-324, Pisa, Italy, October 26-30 1998.

[12] Meseguer, P. Interleaved depth-first search. InProc. of IJCAI-97, Nagoya, Japan, pp. 1382-1387, 1997.

[13] PLATON Team. Eclair reference manual,Version 6.0. Technical Report Platon-01.16,THALES Research and Technology, Orsay,France, 2001.

[14] Perron, L. Search Procedures and Parallelismin Constraint Programming. In Proc. of CP-99,pages 346-360, Alexandria, Virginia, 1999.

[15] Pesant, G. and Gendreau, M. A Constraint Pro-gramming Framework for Local Search Meth-ods. Journal of Heuristics, 1999.

[16] Shaw, P. Using constraint programming andlocal search methods to solve vehicle routingproblems. In Proc. of CP-98, pp. 417-431, Pisa,Italy, 1998.

[17] Zilberstein, Shlomo. Using Anytime Algorithmsin Intelligent Systems. AI Magazine, 17(3):73-83, 1996.

Author Information

Simon de Givry INRA / Biometrics and Artifi-cial Intelligence, Chemin de Borde Rouge, BP27,31326 Castanet-Tolosan cedex France, [email protected]

Laurent Jeannin, Francois-Xavier Josset,Juliette Mattioli, Nicolas Museux, PierreSaveant THALES Research & Technology /PLATON Center, Domaine de Corbeville 91404Orsay cedex France, [email protected]

�


ARTICLE

COMPETE a Common Platform for Extended Project Management

Author: M. Sanseverino

Abstract

COMPETE (Common Platform for the ExTendedEnterprise) is an ESPRIT project which started in1999 and ended at the end of 2001. It is the re-sult of the cooperation among different Europeanpartners: Centro Ricerche Fiat, TXT e-solutions,Cap Gemini & Ernst & Young, BAE Systems andMagneti Marelli, University Federico II of Naples.Its main objective is to integrate methods and toolsto support process/project management in an ex-tended enterprise. In this environment processknow-how (how to do things) and available com-petences play a central role to shorten product de-velopment time and to support rapid decisions andflexibility in process actuation.

COMPETE has developed a software platformwhich integrates process modeling methods, com-petences management, project management andworkflow through a common data model shared byall company departments.

This paper , starting from the description of currentcompanies needs, describes COMPETE approachand functionalities and gives an idea of its generalarchitecture.

Keywords: Project Management, KnowledgeManagement, Competencies, Workflow, BusinessProcesses, Modeling, Simulation.

Introduction

The competitive arena in the Product Developmentscenario is marked today by rapidly evolving tech-nologies, dynamic, sophisticated and global markets,requiring high and customized performances at lowcost. This trend has put a tremendous pressure on thedesign process which must supply a stream of prod-ucts with high value/cost ratios at a rate fast enough towithstand technology obsolescence and demand va-garies.The main goal of COMPETE is to provide IT based

methodologies and tools to help companies struc-tured as Extended Enterprises to cope with the chal-lenges of globalization, deregulation and contract-ing life cycles, combining fast decision making andflexibility to change. Such methodologies embraceboth product function, market, life cycle analysis, to-gether with organisational and individual competen-cies identification and evaluation. The IT tools to beintegrated in a distributed Extended Enterprise archi-tecture (called BMA Business Modelling and Actua-tion ) are: a Business Process Definition toolset (withmodelling and simulation capabilities), a Competen-cies Identification and Evaluation environment, a Hu-man Resources Planner and Scheduler, a CommercialProject Management tool (OPENPLAN) and a Work-flow Management tool.Three main areas are targeted: the product, the pro-cess, and the competencies of human resources (p3

paradigm: product, people, process).On the product side, the project is concerned withthe early conceptual phase where the links betweenfunctional specifications and customer’s value areidentified and exploited to achieve maximum value atminimum cost, ensuring minimal environmental im-pact as well.

Figure 1: Scenario and objectives of COMPETE



On the people side, the project is concerned witha system for managing human resource competen-cies in the short and medium term. Given that pro-cess quality and competencies are correlated, an idealcompetencies profile is associated to each activityand candidates for the activity are selected who bettermatch this profile. To support this procedure, compe-tencies are routinely evaluated, and compared withmedium term requirements to orientate human re-sources policies. Specific methodologies have beenimplemented to find organizational competencies ansevaluate individualsOn the process side, the project is concerned with thegeneric product development process (PDP) from theearly product concept to the detailed design. To slashcosts and development times and to improve quality,enterprises are resorting to concurrency, BPR reengi-neering, standardisation, process control, knowledgedistribution, outsourcing and extended enterprising.Stand alone tools have appeared to support this trend.The project provides a platform which supports thesenew approaches, by integrating the existing tools,based on a sound logical model.Process modelling is crucial to achieve process ef-fectiveness. Actual processes are assembled fromstandard templates, which consolidate the best prac-

tices. Processes are released for execution after de-tailed simulation and planning, which balance effec-tiveness and competencies level. The unique com-bination of a Modeller, a Wf Management Tool anda Project Management tool caters for tight processcontrol, reactiveness to unforeseen events, high qual-ity and reduced processing times. The full potentialof this layout can be obtained by proper interfacingwith a PDM system. The systematic exploitation ofthe existing interfacing standards and use of a Webbased architecture makes the platform ideally suitedto support extended enterprise arrangements.

COMPETE Functions

The main functions supported by COMPETE platformare:

1. To store the operational know-how by meansof process modelling and to support the choiseamong different project alternatives estimatingcosts and times by means of simulation.

2. To support the automatic translation of a processmodel into operational projects and into workflowflow-charts.

Figure 2: Functional diagram of COMPETE Platform


3. To support planning according to competenciesrequired by process activities and automaticallyassigning resources according to workload in amultiproject environment.

4. To support progress control integrated withproject management.

5. To integrate different company departments, suchas human resources department, design, productplanning, project management, working teams, ona common platform and using a common datamodel.

6. To manage in a unique environment planning andprogress control of distributed teams in an ex-tended enterprise.

The platform integrates a project management tool(OPENPLAN) and commercial workflow manage-ment tools with new developments according to man-ufacturing companies requirements and has the fol-lowing advantages:

1. On the process side

� A clear structuring of operational know-howfor porpuse of documenting, distributing andconsolidating knowledge

� A poweful decisional support, based on costand time indicators, obtained by means of sim-ulation of different project alternatives

2. On the competencies management side

� Availability of a catalog of competencies be-longing to the organisations and to the in-dividuals of distributed teams, in a commondatabase and using a common glossary

� Easy evaluation of competency needs� Support to the definition of careers and re-

sources outsourcing� Support to training plan definition� Efficient resource allocation and conflict re-

duction� Improvement of product quality by means of a

better work share among working teams

3. On the project management side

� A higher reactiveness to emerging problems,by means of a deeper and structured knowl-edge of possible alternatives and available re-sources.

COMPETE Operational Flow

The operational flow supported by COMPETE projectstarts from process modeling, transforms process intooperational projects, identifying optimal choices interms of time and costs, supports activity planningand resources assignment according to a matching al-gorithm which compares the need of competenciesof process activities with the availability of compe-tencies in human resources according to their work-load. In the end COMPETE supports project execu-tions integrating project teams executed steps withthe project manager.

Process Modelling and Simulation The firststep is process modeling. At this point it is importantto explain which is the difference between a processand a project. A process describes a set of activi-ties and takes into account possible alternatives re-lated to sudden events which change the pre-definedflow. Unsatisfactory results of certain activities ornew information coming from the market may re-quire some changes or reworking or outsourcing. Aprocess model foresees these alternatives and sup-ports a preventive analysis of the different solutions.It is possible to describe these alternatives with mod-eling formalisms, COMPETE supports IDEF, whichcan be used by a simulation tool which estimatestimes and costs of the different solutions. This de-cision support is very important to react more rapidly



to problems during project execution. A process cangenerate many projects each one representing the ac-tivity flow of one alternative.COMPETE supports then the automatic translation ofIDEF models into Gantt diagrams used by a projectmanager tool.The modeler can start from scratch or use pre-definedprocess templates that are available in the platform.New processes can be created as combination of ex-isting ones. IDEF formalism is used to describe pro-cesses, sub-processes, activities and links. Each ac-tivity includes the description of its inputs, outputs,required competencies and duration. By means ofsimulation it is possible to choose the best project,according to times, costs and competences availabil-ity, comparing the different gantt charts.

Planning and Resource Scheduling

Once the best project solution has been chosen, hu-man resources will be allocated.

COMPETE has developed a matching algorithm that,according to the competencies required by each ac-tivity, selects the proper candidates comparing theirindividual competencies with the required ones. Ina second step the system compares required tim-ings with individual workloads on a multi-project ba-sis and proceeds in building working teams. A redtraffic-light informs the project manager the miss-ing of candidates for the required time interval andthe necessity of acquisition, outsourcing or train-ing. Manual adjustments are obviously possible, con-sidering that human resource scheduling need to bemuch more flexible than machine scheduling. At thispurpose an efficiency factor has been introduced toconsider the fact that human resources with the samecompetencies can have speed and quality very differ-ent.

Human resources assignment are then exported in aproject management tool which will represent graph-ically, gantt and workload.

Figure 4: IDEF modeling of process and competences


Progress Control

Once the planning and the resource scheduling arecompleted, the working teams are ready to start work-ing. COMPETE supports in this phase the aligningbetween the operative management and the progresscontrol. The activity flow, firs modeled and thenplanned is automatically imported in a workflowmanager tool which supports activity dispatching.The activity manager will insert in the system infor-mation related to activity status and progress and theproject manager tool will be automatically updatedaccording to these data.

Competencies Management

COMPETE stresses the importance of the relation be-tween competencies management and process qual-ity. A competence is a combination of knowledge,capacity to apply it and to make it applied. It is es-sential for a company to model its competencies and

to relate them to human resources. It is essential in anorganization to compare its competencies with pro-cess and project needs.COMPETE has developed a set of tools and a databaseto create a competence tree, to describe the organiza-tion, to relate competencies, organization and humanresources and to evaluate human resources accord-ing to their competencies. It is very important for acompany to build its competence tree in order to un-derstand which competencies are available and whichone need to be acquired in order to face technologicalevolution.

COMPETE Architecture

The improved prototypes developed in COMPETE arecurrently being engineered and packaged into a prod-uct suite composed by two main tools: SKILLPLAN(Skill & Process Planner) and P-CON (Process Actu-ation Control & Progress Review)

Figure 5: COMPETE Architecture



SKILLPLAN supports the process planning and op-timisation, by means of process modelling, specifica-tion of competencies’ requirements, simulation andwhat-if analysis. Process simulation caters for al-ternative flows of activities and recycling probabili-ties. The plan is then turned into operation: activitiesare assigned to resources according to required com-petencies, and scheduled. SKILLPLAN architectureis based on the XML integration of an IDEF-basedmodeller, which describes the process, a simulatoraimed to disambiguate it analysing all the relatedpossible project alternatives, a scheduler which usesan innovative algorithm to find the best resources,according to the required competencies and to thetime schedule. A commercial Project Manager hasbeen integrated: this is Open Plan (integrated into thesuite) which supports multi-projects facilities.The main and innovative features of SKILLPLANare: 1- the possibility to perform what-if analysis inorder to find the best projects to implement a certainprocess, 2- the introduction of a competency basedHR planning & Scheduling approach.P-CON aims at managing events, notifications andworkflow (dispatching the activities among the actorsinvolved) along the project lifecycle. It is tightly in-tegrated through XML interfaces to SKILLPLAN inorder to react as fast as possible to problems arisingduring project actuation.The formalisms used for description of processes andprojects , such as IDEF, gantt, flowcharts have beencompletely integrated and can be imported and ex-ported through XML interfaces.The architecture (Fig. 5) is applicable in an extendedenterprise environment. The communication amongthe distributed tools of COMPETE platform is basedon an asynchronous channel for XML models trans-mission and on a distributed database for competen-cies storage.

Conclusion

In conclusion a stream of new organizational ap-proaches have been introduced in the recent years tostreamline Product Development Process , to achieve

better quality and trim the costs. Concurrent engi-neering, businesses process engineering, benchmark-ing, knowledge management, process standardiza-tion, Extended Enterprises have been added to moretraditional automation tools such CAD, CAD/CAMand CAE. Stand alone tools have appeared to supportthese new practices. COMPETE brings together thesenew approaches in an integrated platform, which en-sures as much as possible the communication amongtools of different make. The benefits of the system, itscapability to coordinate concurrent activities in dis-tributed environment, are such as to significantly im-pact on quality, time-to-market and costs in modernextended enterprise.

Bibliography

[1] G. Morra, M. Sanseverino Un progetto Glob-ale: il caso Fiat, Sistemi e Impresa , September,1996.

[2] G. Zollo, A. Cannavacciuolo, G. Capaldo, A.Ventre, A. Volpe The Organisational EvaluationProcess, Fuzzy Economic Review, 1, no.1 (1996)3-30.

[3] M.Argilli, S.Gusmeroli, G. Secco Suardo, N.Matino, M. Sanseverino COMPETE : a com-mon platform for the extended enterprise 2000,e-Business and e-Work 2000 , October, 2000.

[4] G. Michellone, G.Zollo Competencies man-agement in knowledge based firms, Interna-tiona Journal of Technology management, vol20, 2000.

[5] G. Zollo, L. Iandoli, F. Borrelli, A Iuliano, M.Sanseverino Applications of soft computing tech-niques to Product Design, International Semi-nar on Intelligent Computation in ManufacturingCIRP June, 2000.

Author Information

Marialuisa Sanseverino FIAT Research Center,[email protected]

�


ARTICLE

In Defence of Reactive Scheduling

Author: J.E. Spragg

Introduction

Reactive Scheduling is the Cinderella of schedulingtechniques. Reactive Scheduling drudges away in thework place, delivering feasible solutions to highlycomplex and dynamic resource allocation problemsbut is seldom invited to conferences to show its fineryto the princes of the scheduling and planning commu-nity. Their attention is distracted by yet another algo-rithm that promises optimum results on some staticbenchmark problem. Indeed, a recent PLANET spon-sored workshop on on-line planning and scheduling[1] did not even include reactive scheduling as oneof its topics of interest until the second call for pa-pers. This is strange, on-line scheduling domains arethe most sensitive to sudden environmental change.The reality is that in stochastic environments the onlyoption open to a scheduler is to react to change. Tech-niques for anticipating change are limited and seldomprovide the functionality needed to manage resourceallocations that are continually being impacted bydisruptions of one form or another.

The Need for Reactive Schedulers

The requirement for reactive scheduling systems ismore widespread in the industrial world than most re-searchers from the academic world realise. For exam-ple, the production system used for garment manu-facture in the UK is called a Progressive Bundle Line(PBL) [2, 3, 4]. It is a flow line manufacturing systemin which work stations, comprising of machinist andsewing machine pairs, are arranged in a configurationso that the flow of work from one work station pro-vides work for the next work station in the productionsequence. The scheduling objective is to maintain aline balance so that the work in progress buffers be-tween each work station never overflow, causing stor-age and quality problems, or empty, causing machin-

ists to sit idle at their machines.There are examples of line balancing algorithms inthe scheduling and optimisation literature. The ma-jority of these techniques view line balancing as astatic optimisation problem and are wholly inade-quate for tackling the management of a PBL. The factof the matter is, in the real world, a well balancedline soon becomes unbalanced. Machines break-down, machinists go absent or start working below, orabove, standard performance, managers decide thatpriority job batches are no longer important and thatlow priority batches are important, quality controllersdecide that rework is required on some batches, etc.In clothing factories that employ PBL production sys-tems the most important person is the line supervisorwho continuously monitors, analyses, and revises theflow of work through the work stations. How wellshe (it is usually a she) does this depends upon herexperience and training.Providing reactive scheduling capabilities to a real-time automated scheduler involves mimicking the re-sponsibilities of the human supervisor employed tomanage a PBL. An ideal reactive scheduling frame-work employs an event driven multi agent approachthat applies monitoring, analysis, revision, and opti-misation tools in real-time to an executing scheduleto maintain its feasibility and quality over time.Reactive scheduling techniques can be reinforced bybuilding robust schedules or providing probabilisticmodels of system behaviour. While nobody doubtsthe importance of these techniques; if a machine isto breakdown it would be useful to have an indicatorof which machine and when; there is still the prob-lem of what the system can do about it. The responseto a machine breakdown for example is often contextsensitive and depends upon the current opportunitiesfor resource reallocation within the current schedule.Robust schedules limit the impact of a disturbanceon a schedule but again the system still has to reason



about strategies for bringing the solution back to fea-sibility. At the end of the day, the system will needto react by relaxing due dates or reallocating activi-ties to alternative resources or bumping lower priorityoperations.

Mixed Initiative Scheduling – and beyond

The OZONE Project [5] at the Intelligent Coordina-tion and Logistics Laboratory, Carnegie Mellon Uni-versity, back in the mid-90s of last century, recog-nised that current scheduling systems do not effec-tively support user tasks and requirements and donot support the iterative, evolving process of prob-lem understanding, requirements determination, con-flict resolution and solution refinement that is inher-ent in large, multi criteria problem solving. OZONE

tackled these issues by enabling flexible collabora-tive problem solving between user and system, sup-porting reconfiguration of system functionality to ac-commodate new environments and scheduling objec-tives. The OZONE approach to scheduling recognisedthat scheduling is an incremental process of ‘gettingthe constraints right’ in which human users alwayshave the big-picture decision-making expertise andknowledge to contribute but are unable to effectivelycope with the complexity of detailed solution devel-opment.The challenge for the new generation of real-time,on-line, schedulers is to encode the strategic knowl-edge that humans provide within an autonomousscheduling framework that can respond immediatelyand intelligently to change. When analysing ascheduling solution, both human and software agents,ask the same two questions:

� Where are the processing bottlenecks?� Where are the scheduling opportunities?

A mixed initiative scheduling tool supports user ex-perimentation with graphical displays and statisti-cal summaries. An autonomous scheduler can keepinternal representations of analysis results and pro-vide revision tools that are activated by internal statechanges to the control architecture. An autonomous

scheduler needs to evaluate the ‘correctness’ of its re-visions by comparing the results against ‘objective’criteria given by a multi-criteria cost function. Theinterpretation of such evaluations are often problem-atic and seldom objective. Techniques developed bydecision theorists, such as criteria voting [6], for re-solving conflicts in multi criteria decision makingsuggest promising alternatives to single valued mea-sures of ‘optimality’ for self-governing systems.

Reactive Problem Solvers – Schedulersfor the new era

The recent developments in scheduling technologyhave been partly driven by improvements in moni-toring technology. Communication technology hasmade reactive scheduling more important by enablingschedulers to keep up to date with environmentalchange and allowing them to respond to disruptionsby resolving conflicts in impacted parts of the so-lution. Gone are the days in manufacturing wherejobs were ‘chased’ through the factory by a progresschaser who spent their day sweet talking departmen-tal foreman and climbing over work inventories look-ing for job numbers. The advent of mechanicallyreadable bar codes means that work can now be con-tinuously monitored through a production process.Mobile telephony has had a major impact on lo-gistic scheduling, allowing huge savings by allow-ing field workers to communicate with headquar-ters in real-time. The British Telecom mobile work-force scheduling problem [7] addressed by a.p.solveis highly dependent upon mobile telephony. It is aproblem characterised by:

� A varying workload with one hour response times.� A diverse mixture of operational procedures that

range from those with hard start-time constraintsto those with highly relaxable start-time con-straints.

� The processing times of scheduling activities canrange from a few minutes to several days and aresubject to uncertainty.

� Resource scarcities in rare skills.


� Scheduling activities can be complicated by de-pendency relationships between activities.

� The environment in which mobile workers operateis subject to uncertainty and change. Calculatingtravel times is problematic and subject to trafficconditions.

British Telecom has a workforce of 20,000 tech-nicians nation-wide with several hundred thousandtasks to be scheduled and executed every day.In a resource constrained environment, with a smalltime budget to resolve conflicts and improve solutionquality, the system is forced to seek out spare capac-ity by searching a space of reallocation moves to findalternative resources and/or start times for activitiesin conflict. The most promising kinds of algorithmsfor this kind of constraint based search are those likeGinsberg’s and McAllester’s partial-order backtrack-ing [8] that combine aspects of both systematic andnon-systematic search. However, in practical envi-ronments partial-order backtracking needs to be sup-ported by texture analysis that indicates the aggregatedemand for alternative resources by providing mea-sures of contention and reliance [9].Given the industrial need for reactive scheduling sys-tems, why is the scheduling community obsessedwith static optimisation problems? In practice,scheduling problems require a ‘solution’ to be morethan the mere implementation of an algorithm forsolving a particular constraint satisfaction, or con-strained optimisation problem. Constructing sched-ules, in practical environments is an extended, iter-ated process that typically involves resolving con-flicts between competing schedule users and schedul-ing tools. In most applied domains, schedules needto be maintained over time through reactive revisionthat refine an initial, or current, solution, by adapt-ing it to changing environmental conditions and userpreferences.

Bibliography

[1] AIPS 2002 Workshop on On-line Planning andScheduling, Toulouse, France, April 24th, 2002.

[2] Fozzard, G., Spragg, J., and Tyler, D. Simula-tion of Flow Lines in Clothing Manufacture, Part1: model construction, Int.J. of Clothing Scienceand Technology, Vol 8, No. 4, pp 17-27, 1996.

[3] Fozzard, G., Spragg, J., and Tyler, D. Simulationof Flow Lines in Clothing Manufacture, Part 2:credibility issues and experimentation, Int.J. ofClothing Science and Technology, Vol 8, No. 5,pp 42-50, 1996.

[4] Spragg, J.E., Fozzard, G. and Tyler, D.J. FLEAS:a flowline environment for automated supervi-sion, The Int.J. of Manufacturing TechnologyManagement, Vol. 10, Number 6, pp 322-327,1999.

[5] Smith, S., Lassila, O., and Becker, M. Con-figurable, Mixed Initiative Systems for Planningand Scheduling, Advanced Planning Technology,Technological Achievements of the ARPA/RomeLaboratory Planning Initiative, Edited AustinTate, AAAI Press, pp 235-241, 1996.

[6] Croce, F.E., Tsoukias, A., and Moraitis, P. Whyis it Difficult to Make Decisions Under MultipleCriteria, AIPS-02, PLANET Sponsored Work-shop, Planning and Scheduling with MultipleCriteria, pp 41-45, 2002.

[7] Lesaint, D., Azarmi, N., Laithwaite, R., andWalker, P. Engineering Dynamic Scheduler forWork Manager, BT Technology Journal, Vol. 16.No. 3 July, pp 16-29, 1996.

[8] Ginsberg, M.L., and McAllester, D.A. GSAT andDynamic Backtracking, Knowledge Representa-tion and Reasoning Conference, 1994.

[9] Beck, J.C., Davenport, A.J., Davis, E.D., andFox, S. The ODO Project: Towards a UnifiedBasis for Constraint-Directed Scheduling, J. ofScheduling, 1, 89-125 (1998).

Author Information

John E. Spragg a.p.solve, (www.apsolve.com), Sirius House, Adastral Park, Martlesham, Ip-swich, Suffolk, IP5 3RE, United Kingdom, [email protected]

�



ARTICLE

Advanced Scheduling and Optimisation: Cutting the Costs ofManufacturing

Author: B. Drabble

Abstract

The aim of this article is to describe several in-telligent scheduling techniques that have been ap-plied to problems in manufacturing and assembly.The techniques have been evaluated in domainsincluding aircraft wing manufacturing, fiber opticcable manufacturing, submarine construction andCD manufacturing. The article will describe twoparticular scheduling techniques: schedule packingand squeaky wheel optimization. The techniqueshave resulted in a number of major improvementsincluding a reduction in make-span of up to 50%,an improvement in throughput of 40%, a reduc-tion in costs of 20% and the ability to tackle prob-lems up to 20 times larger. The article providesan overview of these techniques and describes twocase studies from Boeing and Electric Boat. Thearticle concludes with a description of the impactthese techniques have had in each of these orga-nizations and provides pointers to potential futureimprovements.

Introduction

Scheduling is the problem of assigning a set of tasksto a set of resources subject to a set of constraints.Examples of scheduling constraints include deadlines(e.g., job i must be completed by time t), resource ca-pacities (e.g., there are only four drills), precedenceconstraints on the order of tasks (e.g., a piece mustbe sanded before it is painted), and priorities on tasks(e.g., finish job j as soon as possible while meetingthe other deadlines). In addition the task assignmentsmust also meet a number of optimization criteria, e.g.minimize makespan, minimize set up times, maxi-mize work in progress. Examples of scheduling do-mains include classical job-shop scheduling, manu-facturing scheduling, and transportation scheduling.This article describes two generic scheduling tech-

niques that were developed and applied to problemsin aircraft manufacturing and submarine assembly. Ineach case the results obtained are currently the bestin the world. These techniques are now being aug-mented with additional functionality to tackle prob-lems involving ship repair/overhaul and mission plan-ning for the USAF. The techniques can be dividedinto two main areas:

� The use of non-systematic techniques such as“squeaky wheel” optimization [4] (SWO) andschedule packing (also known as doubleback opti-mization) [2] to solve problems that arise in man-ufacturing scheduling.

� The use of combined systematic and non-systematic techniques, such as limited discrep-ancy search (LDS) [3] with schedule packing, andsqueaky wheel with operations research methods.

Scheduling Technologies

This section provides an overview of the schedulingtechnologies that have been deployed in a number ofreal applications.

Limited Discrepancy Search (LDS) andHeuristics

In scheduling problems of any size it is unlikely thatalways using a merely good heuristic will get you re-ally close to an optimal schedule. A merely goodheuristic will be incorrect some of the time. Asthe complexity of scheduling problems increases, thenumber of decisions guided by the heuristic also in-creases. The more decisions made, the more likely itis that some of them are going to be incorrect. Howdoes LDS help address this problem? LDS is a system-atic method for disregarding the recommendation of


a heuristic a limited number of times (thus the nameLDS) when generating a schedule. With LDS, sched-ules are generated repeatedly, each of them followingthe heuristic for all decisions except one. The deci-sion at which the heuristic is ignored is different ineach schedule. If the heuristic leads to only one in-correct decision, then using LDS1 (the fastest formof LDS) will lead to a perfect schedule. Even if theheuristic leads to more than one incorrect decision(which is usually the case) then LDS1 will likely leadto a better solution then always following the heuris-tic.

Schedule Packing

Schedule Packing, also known as doubleback opti-mization involves “sloshing” a candidate schedule,repeatedly, right and left within a scheduling window.This has a remarkable impact on the length of mostschedules. The Doubleback process is analogous tofilling a box with blocks and then shaking the box.Shaking the box will almost always result in a denserpacking of blocks. Likewise, in schedules, Double-back almost always results in a denser packing oftasks in a schedule. The denser packing allows fortasks with few preceding activities to find appropri-ate holes in the schedule in which to be placed. Theschedule packing algorithm is appropriate for prob-lems with large numbers of precedences, e.g. assem-bly tasks, manufacturing, overhaul, etc.

Squeaky Wheel Optimization

The insight behind SWO is that in any real worldproblem it is impossible to capture all associated con-straints, e.g. context information. SWO uses a priorityqueue to determine the order in which tasks should bereleased to a greedy scheduling algorithm. The prior-ity queue is determined by how difficult the task is todeal with that is, i.e. higher the task is in the queuethe harder it is to find a good resource assignment.On each iteration of the algorithm, SWO quickly cre-ates a schedule and then examines it to identify theparts that were handled badly, for example, the task

was completed too late or assigned to an unsuitableagent. Any task that “squeaks” is promoted up thepriority queue, with the distance it is promoted deter-mined by the extent of the problem. The new prior-ity queue is then used to generate another schedulethat is analyzed for problems. This process contin-ues until no significant improvement in the scheduleis noted over several iterations or a predefined limit isreached i.e. cycle count or elapsed time. SWO is ex-tremely fast with each cycle of generate, analyze, andre-prioritize taking less than a second, even for largeproblems, e.g. 2500 tasks and 200 resources over afive day period.

Application Domains

This section describes two example domains thathave been tackled using either schedule packing orsqueaky wheel optimization. In each example a de-scription of the domain will provided together witha description of the impact the technology has made.In addition to the domains described here the tech-niques have also been applied to submarine assem-bly, aircraft mission scheduling [5] and CD man-ufacturing. Full details can be found via pointerwww.otsys.com/scheduling.html

Aircraft Assembly

The original aircraft assembly problems tackled byschedule packing were provided via a research groupat McDonnell Douglas. These were real schedulingproblems and were made available via the net to en-courage academic researchers to demonstrate the ap-plicability of their techniques. These particular prob-lems are relevant to large scale assembly and are in-stances of problems known as resource constrainedproject scheduling (RCPS). CIRL developed a sched-uler that produces the best known results on this prob-lem. CIRL has also converted several other bench-mark problems to the same format and solved themsuccessfully. The basic aircraft assembly problemhas the following features:



Figure 1: Example of aircraft assembly

Zone resources : A zone is an area of the aircraftin which work can be done. Zone resources spec-ify the maximum number of people that can workin that zone at the same time.

Labor resources : These specify how many la-borers are available with a particular skill set.

Shifts : The availability of labor resources variesover time, with more being available during theday.

Tasks : Tasks have a specified duration and set ofzone and labor resources that are needed to per-form the task.

Precedences : These specify which tasks must becompleted before other tasks can begin.

The program uses limited discrepancy search (LDS)and schedule packing (also known as doubleback op-timization) to generate solutions. LDS and sched-ule packing can be used in isolation or in combina-tion with each other with the best results producedusing LDS with schedule packing or schedule pack-ing on its own. When used in combination, multipleseed schedules generated with LDS are fed to sched-ule packing.In figure 1 the top part indicates resource usage overtime. There is one horizontal bar for each of the 17

resource types. The darker areas indicates a resourceis fully utilized, and the lighter indicates it is unused.The boxes in the bottom part of the figures representthe tasks. The width of a box represents the dura-tion of a task, and the height is an indicator of howmany resources a task requires. The PERT 1 sched-ule for this problem ends after 37 days, 2 hours and58 minutes, and is a loose lower bound on the mini-mum length any potential solution. The current bestschedule produced by McDonnell Douglas (now theBoeing Company) is just over 42 days and the bestschedule pack schedule is just over 39 days. Eachday of production removed from the schedule savesthe company approximately $600,000 and thus theschedule is able to save approximately $3.2 millionper sub-assembly. On Time Systems 2 has extendedthis scheduler to handle more general aircraft manu-facturing problems. These new constraints and fea-tures include:

� multiple sub-assemblies could be introduced intothe line at fixed rates, e.g. “every five days” or atarbitrary points, e.g. “40 wings over the next 10days”.

� Once a wing assembly was assigned to a bay itmust return to the same bay for further processingsteps.

1The PERT schedule is generated by starting tasks as early as possible and ignoring all resource conflicts2On Time Systems is a technology startup company that was developed to commercialize the optimization technology being de-

veloped at CIRL and from other groups around the world.


� The assembly lines have a number of robots thatmust be taken down for routine maintenance.

Submarine Assembly

Existing scheduling systems in use at shipyards to-day, such as ARTEMIS, SAP, or PRIMIVERA, relyon “makespan minimization” techniques to developschedules. Specifically, they try to schedule tasksas early as possible, subject to constraints and re-source availability profiles. Manual intervention istypically required to deal with overloaded resourcesand other difficulties, and the process of schedulingthe construction of a single ship can take months.This approach relies on the conventional wisdom thatit can never be wrong to get work done early, andthe assumption that a short schedule is likely to beefficient, since otherwise the inefficiently utilized re-sources could be loaded up to get more done earlier.OTS) research has shown that this conventional wis-dom is, in fact, misleading. In mass production en-vironments, makespan minimization often is a use-ful approach, since other jobs can help smooth outthe resource loading artifacts the process induces. Inshipyards, where it is not uncommon to have one, orat most a few, projects in process at any one time,makespan turns out to be a poor stand-in for the ship-yards’ more complex goals.OTS has developed a radically new approach to shipconstruction scheduling, one that addresses ship-building’s unique needs directly. The resulting sys-tem, ARGOS, is capable of scheduling multiple years’production across a whole yard in hours, instead ofmonths, without need for human intervention. Theresulting schedules typically exhibit a 10-20% reduc-tion in construction labor costs when compared withthose in use in the yards today. Conversely, in sit-uations where throughput is limited by the availablemanpower pool, ARGOS makes it possible to progress10 to 20% more work through the yard. All of thesesavings are achieved without changing the fundamen-tal production process or shipyard facility in any way.Table 1 shows the expected savings (over the cur-

rent schedule) for a single hull and Table 2 showsthe expected savings for the entire yard over the next5 years. Our estimates are that, if ARGOS were ap-plied to all new Navy construction, annual savingscould be expected to be between $200M and $500M.Numbers for re-fit and repair are more difficult to ob-tain, but the percentage savings (10-20%) appear tobe comparable.

iteration Time Savings1 2 min 8.4% $13.0M7 10 min 11.4% $17.7M20 34 min 11.8% $18.2MUltimate 24 hrs 15.5% $24.0M

Table 1: Expected Savings for a Single Hull

iteration Time Savings1 24 min 7.8% $49M7 1 hour 10.2% $65M20 4 hours 10.7% $68MUltimate 4 days 11.5% $73.0M

Table 2: Expected Savings Over the Entire Yard

Figure 2 provides a comparison of the resource uti-lization and manpower curves for the current sched-ule 3(top graph) and the one developed by ARGOS

(bottom graph) . The black line represents the amountof manpower required on a particular day. The redand blue lines represent the actual manpower avail-able in the shipyard schedule and ARGOS schedulerespectively. Deviation from the black line resultsin additional costs due to overtime, under-time orincreasing/reducing the total work-force. The AR-GOS schedule has a much smoother profile requir-ing fewer changes in manpower levels and providesthe ability to recover much more easily from unex-pected changes in project deadlines. Figure 3 pro-vides a comparison between the manpower needed inthe current schedule (red line) and the manpower re-quired by the ARGOS schedule for a single hull. Thisshows a smooth resource ramp “up and down” forthe ARGOS schedule and far less perturbation in theresource levels in the yard.

3This is the schedule the yard is currently building to.



Figure 2: Manpower and Resource UtilizationComparison

Figure 3: Single Hull Resource UtilizationComparison

Summary

Search based scheduling technologies have maturedto the point that they are now capable of solvinglarge real world problems and provide users with highquality solutions. Table 3 provides a summary of theimprovement in problem size and complexity that canbe handled by current techniques. In addition to be-ing able to solve complex real world problems it isalso interesting to note that the technology transferpath for these methods is short, which makes themeasily accessible to industrial users.

Tasks Resources Type Feasible1993 64 6 Job Shop No1996 570 17 RCPS Barely1999 1000s Dozens RCPS Yes2001 10,000s Hundreds RCPS Yes2002 millions Hundreds RCPS Yes

Table 3: Improvement in Problem Size andComplexity

Bibliography

[1] Aarup, M. and Arentoft, M.M. and Parrod, Y. andStokes, I. and Vadon, H. and Stader, J., Optimum-AIV: A Knowledge-Based Planning and SchedulingSystem for Spacecraft AIV, Intelligent Scheduling,(eds. Zweben, M. and Fox, M.S) Morgan Kaufmann,Inc, 1994, pp451-469.

[2] Crawford, J., An Approach to Resource ConstrainedProject Scheduling, in proceedings of the ArtificialIntelligence and Manufacturing Research PlanningWorkshop, June 1996, AAAI SIGMAN.

[3] Harvey, W. and Ginsberg, M., Limited Discrep-ancy Search, in the proceedings of Fourteenth Inter-national Joint Conference on Artificial Intelligence(IJCAI-95), July 1995, IJCAI Inc.

[4] D. E. Joslin and D. P. Clements, Squeakywheel Op-timization, in the proceedings of the Fifteenth Na-tional Conference on Artificial Intelligence, July,1998, AAAI Press, Menlo Park, CA.

[5] Drabble, B., Task Decomposition Support to Reac-tive Scheduling, in the proceedings of the 5th Eu-ropean Conference on Planning (ECP-99), Durham,September, 1999, Springer Verlag Press, New York,NY.

Author Information

Brian Drabble On Time Systems, Inc, 1850 Mill-race Drive, Suite 1, Eugene, OR 97403-1269, USA,[email protected]

�


REPORT

2nd International PLANET Summer School

Author: L. Compagna, G. Cortellessa, A. Farinelli, and N. Policella

After the great success of the first edition (Cyprus,September 2000), the 2nd International SummerSchool on AI Planning was held in Halkidiki, Greeceon September 16-22, 2002.This event was one of a number of activities orga-nized by PLANET (the European Network of Excel-lence in AI planning), whose main aim is to promoteknowledge exchange among students and researcherswho are new to the field with a view to foster inter-action and international collaborations. More thanfifty PhD students and researchers from academia andindustry attended the School, which offered coursesheld by top-level, internationally-renown speakers.The courses were divided into eight distinct parts,which covered most of the current “hot” topics in AIplanning:

Planning under Uncertainty with MarkovDecision Processes

Craig BoutilierMarkov Decision Processes (MDPs) are a widelyused computational method for solving sequential de-cision problems involving uncertainty. This courseprovided a brief introduction to MDPs, and focusedon the methods of representation and solution that arestrictly related to AI planning. A greater emphasiswas put on approaches that reduce the computationaleffort for solving MDPs through the adoption of tech-niques developed in the AI community.

Plan-based Control of Autonomous Robots

Michael BeetzGiven the increasing interest in autonomous roboticapplications and in improving the performancesof current systems (MARTHA, XAVIER, RHINO,MINERVA, REMOTE AGENT), this course aimedto provide the attendees with a broad overview of

the issues involved in the plan-based control of au-tonomous robots. Michael presented the computa-tional principles and basic software architectures thatenable robots to perform complex and diverse task indynamic environments. The need for an integratedapproach among plan representation, reasoning, exe-cution and learning was strongly underlined.

Planning history and Overview

Susanne BiundoThese lectures offered a comprehensive introductionto the field of AI Planning. They reviewed existingplanning methods (classical, heuristic search, hier-archical and hybrid, and logic-based), describing indetail their domain representations and introducingpresent and future application areas. In addition, Su-sanne also presented developments towards a system-atic combination and integration of different planningmethods, as well as the integration and use of tech-niques from related fields of research. The coursewas concluded by a useful historical overview of theAI Planning field.

Scheduling and Planning

Brian DrabbleBrian presented an overview of recent developmentsin intelligent scheduling and optimization and theways in which these systems and algorithms can beintegrated with planners to develop a comprehensiveapproach to the planning and scheduling problem.Traditionally, scheduling and planning were viewedas separate research areas. However, this is a simplis-tic view, as many decisions in planning have a directimpact on scheduling, and viceversa. The overviewprovided details of several scheduling techniques anddescribed many of their properties.



Planning and resources

Philippe LaboriePhilippe’s lectures began with the introduction of theconcept of resource – any substance or (set of) ob-ject(s) whose cost or available quantity induce someconstraint on the operations that use it – and by show-ing application domains for planning with resources.The course then proceeded with a review of the stateof the art of planning with resources, and presentedbasic tools and planning techniques. Finally, Philippealso described in detail one of the most promising ap-proach for dealing with resources in planning, whichrelies on the application of constraint-based tech-niques in partial-order or hierarchical task network(HTN) planning.

Planners Performance Evaluation

Maria FoxThe goal of this course was to analyze and dis-cuss methods for the evaluation of planning systems.Planners can be evaluated in a number of differ-

ent ways: by analysis of their formal properties, byempirical comparison with other similar systems, interms of the time/quality trade-offs they make, andso on. This course outlined the stages of a scientificapproach to evaluation of a data set. Moreover thecourse introduced some statistical tests that can beused to determine the significance of a feature of adata set.

Planning and Execution

Martha PollackThese lectures dealt with the interesting problem ofplanning and execution applied mainly to SimpleTemporal Problems. The speaker highlighted the factthat classical planning makes strong simplifying as-sumptions on the world model. In particular, in realenvironments the beliefs and goals at the base ofthe original planning problem are subject to changes,and hence the solution plan might loose consistency.Martha provided an helpful introduction to this topic,dwelling on the following techniques: (a) interleav-ing planning and execution; (b) monitoring plan ex-


ecution and inferring execution status; (c) recover-ing from failure by replanning; (d) maintaining thetemporal consistency of plans during execution (plandispatch); (e) managing commitments and updatingplans in response to changes in the environment.

Planning and the Web

Craig KnoblockThis course provided an interesting overview of thetechniques involved in the context of gathering andintegrating information from the web. These topicsinclude query planning for information gathering, in-teractive planning using constraint propagation, andefficient execution of information gathering plans.Craig pointed out how information gathering fromthe web can be tackled as a planning problem, as itcan be viewed as a process aimed at formulating ascheme or program for the accomplishment of somegoal.

Poster Session

The School also included a poster session, duringwhich attendees were able to present the recent devel-opments of their work. This allowed further interac-tion between students, researchers, lecturers and or-ganizers and enabled the authors (more than fifteen)to receive useful feedback and suggestions.

Social Activities

The local organizers put together a varied programmeof social and cultural activities. These activities com-prised an exciting excursion to the caves of Petralo-nia, a visit to the historical city of Thessaloniki, anda short trip to Nea Skioni, a traditional fishing vil-lage. This allowed the participants to get in touchwith the ancient history, local cuisine, origins, musicand dances of this fascinating country.

Conclusions

The School was a great success, providing not onlyan invaluable and up-to-date review of the recent

developments and techniques in AI Planning andScheduling, but also a friendly and informal en-vironment in which interaction was facilitated andresearch collaborations fostered. Many challeng-ing questions were raised during the lectures, stim-ulating discussion and promoting knowledge ex-change amongst all of the participants. Coursematerials, long abstract of the posters and lotsof funny photos can be found at the web pageof the school http://cswww.essex.ac.uk/PLANET/summer-school-02/.

Acknowledgements

The realization of this event – sponsored by PLANET– required the coordinated effort of several people,including Susanne Biundo (University of Ulm, Ger-many), Enrico Giunchiglia (University of Genoa,Italy), Sam Steel (University of Essex, UK), Ioan-nis Refanidis (University of Macedonia, Greece), andIoannis Vlahavas (Aristotele University of Thessa-loniki, Greece). We would also like to thank MaxGaragnani for his help in the production of this doc-ument.

Author Information

Luca Compagna DIST - University of Genoa,Italy, [email protected]

Gabriella Cortellessa ISTC-CNR [PST], Plan-ning and Scheduling Team, Institute for CognitiveScience and Technology, National Research Councilof Italy, Rome, Italy, [email protected]

Alessandro Farinelli DIS, Dipartimento di In-formatica e Sistemistica, Universita’ degli Studi diRoma “La Sapienza”, Rome, Italy, [email protected]

Nicola Policella DIS, Dipartimento di Informat-ica e Sistemistica, Universita’ degli Studi di Roma“La Sapienza”, Rome, Italy, [email protected]

�



ARTICLE

Generation and Execution of Partially Correct Plans in DynamicEnvironments

Authors: A. Farinelli, G. Grisetti, L. Iocchi, D. Nardi, and R. Rosati

Introduction

In this work we present the recent developments ofthe approach to the design of Cognitive Robots (i.e.robots whose actions are driven by a formally devel-oped theory of action), that are capable of performingtasks in a coordinated way. The logic of actions thatwe adopt is an epistemic dynamic logic, where it ispossible to derive acyclic branching plans (branchescorresponding to sensing actions), including primi-tive parallel actions.In the present work, we consider an extended notionof plan by admitting a simple class of cycles that arisefrom the attempt to recover from the failure statesoriginated by sensing actions. The proposed exten-sion allows us to address the problem of generatingplans that handle a form of synchronization based onthe recognition of specific situations through sensingactions, including forms of coordination required ina multi-robot scenario.

System Architecture

In this section we recall the layered hybrid archi-tecture used for our cognitive mobile robots (seealso [4]) displayed in Fig. 1, that has been imple-mented on several different kinds of robotic plat-forms, namely Sony AIBOs, Pioneer, and home-made wheeled robots.The deliberative level is formed by three main com-ponents: The Plan Execution Module that is executedon-line during the accomplishment of the robot’s taskand is responsible for executing a plan by coordinat-ing the primitive actions of a single robot. The Co-ordination Module that is responsible for assigningtasks to the robots in the team according to the currentsituation. The Plan Generation Module, that is exe-cuted off-line before the beginning of the robot’s mis-

sion, and generates a set of plans to deal with somespecific situations.

Actions Primitive

Generation Plan

CoordinationModule

Plan Execution

Perception WorldModel

KBLibrary

PlanDeliberative Level

Off-line

ConditionsHigh-level

ActuatorsSensors

Operative Level

On-line

Deliberative Level

Figure 1: Layered architecture for our robots

Plan Representation and Generation

In order to address the problem of synchronize dif-ferent robotic platforms using sensing action we useda particular notion of plan that we called PartiallyStrong Plan. This notion of plan is equivalent to thedefinition of strong cyclic plans given in [3]. A Par-tially Strong Plan is a plan that if terminate leads toa goal state. Namely, a partially strong plan is a planthat is not guaranteed to terminate: termination actu-ally depends on the outcome of the sensing actionsin the plan. However, if such a plan terminates, thenit always leads to the goal. The generation of plansis based on the use of a modal non monotonic de-scription logic

��[2]. As illustrated in [5],

the set of models of an��

knowledge base formalizing a dynamic system can be represented bymeans of a unique transition graph, called first-orderextension (FOE) of , which represents all the possi-ble evolutions of the dynamic system. For instance,Fig. 2 displays some examples of portions of FOEs.


X2X1

X0

BA

X2X1

X0 X0

senseA (T) senseA (F)

senseA (F)


senseA (F)

senseA (T)

senseA (F)senseA (T)

GOAL

senseA (T)

GOAL

senseA (F)

b)a) c)


senseA (T)

Figure 2: Plan structure for cyclic sensing actions

The plan generation module selects a portion of theFOE of the KB containing only those actions that arenecessary to achieve a goal starting from a given ini-tial situation. In fact, conditional plans can in princi-ple be generated in two steps (see in [5] for details).First, the FOE of the knowledge base is generated;this FOE can be seen as an action graph representingall possible plans starting from the initial state. Then,such a graph is visited, building a term (the cyclicconditional plan) representing a graph in which: (i)sensing actions generate branches; (ii) each branchleads either to a state satisfying the goal or to a previ-ous state of the plan. However, the current implemen-tation does not build the entire FOE before searchingfor the plan, but it builds the FOE starting from theinitial state with a breadth-first technique until a goalstate is reached. In case of sensing actions all thebranches are required to reach a goal state. In thisway it is possible to extract a minimal plan (with re-spect to the number of actions to be executed).

Implementation

We provided an implementation of our approach us-ing a simulator. The situation we experimented asan example for explaining our plan generation andexecution mechanism, is the dynamic exchange of

the role of goal keeper in the Sony Legged Leagueand the application of the two-defender rule. Thesituation presented is a typical situation in the SonyLegged League matches in which the goal keeper(robot number 1) is moving away from its own goaland is approaching the ball to push it away, while an-other robot (robot number 2) is far away from the balland it cannot help the goal keeper immediately. Inthis situation, it is more convenient for the team thatrobot 1 takes the role of attacker pushing the ball to-wards the opposite goal, while robot 2 goes back todefend its own goal acting as a goal keeper. However,in performing this role exchange the two robots mustcomply with the two defenders rule, and thus robot2 can enter the goal area only after robot 1 has leftit. The problem of complying with the two-defenderrule is solved by generating a plan in which one robot,before entering the goal area, must check that it is free(i.e. the other robot has left the area). This is achievedby adding in the knowledge base of the robot thespecification of a sensing action SenseFreeArea thatis used for verifying if the goal area is not occupiedby any robot of the team. Even though the simulationcannot provide a precise characterization of all theaspects that influence the performance of the robot inthe real environment, it can provide useful feedback



to the design of the coordination and plan executionmodules for actual robots. Through this simulator wehave verified the intended behaviour of the robots ineach of the roles in different scenarios.

Conclusions

As compared with previous work in CognitiveRobotics this is a novel attempt to generate plans thatinclude cycles in a partially known environment. Ascompared with the work on classical planning thereis a close relationship with the work in [1], whereonly conditional plans (tree-structured) are gener-ated. We are currently addressing the application ofmodel checking techniques, as done in [1], also in oursetting. Moreover, we are extending our analysis togenerate plans with cycles that are more general thanthe ones presented in this paper.

Bibliography

[1] P. Bertoli, A. Cimatti, M. Roveri, and P. Traverso.Planning in nondeterministic domains under par-tial observability via symbolic model checking.In Proc. of the 17th Int. Joint Conf. on ArtificialIntelligence (IJCAI 2001), 2001.

[2] Francesco M. Donini, Daniele Nardi, and Ric-cardo Rosati. Description logics of minimalknowledge and negation as failure. ACM Trans.on Computational Logic, 3(2):1–49, 2002.

[3] Fausto Giunchiglia and Paolo Traverso. Planningas model checking. In Proc. of the 5th Eur. Conf.on Planning (ECP’99), 1999.

[4] Luca Iocchi. Design and Development of Cog-nitive Robots. PhD thesis, Univ. ”La Sapienza”,Roma, Italy, 1999.

[5] Luca Iocchi, Daniele Nardi, and Riccardo Rosati.Planning with sensing, concurrency, and exoge-nous events: logical framework and implementa-tion. In Proceedings of the Seventh InternationalConference on Principles of Knowledge Repre-sentation and Reasoning (KR’2000), pages 678–689, 2000.

Author Information

Alessandro Farinelli, Giorgio Grisetti, LucaIocchi, Daniele Nardi, and Riccardo RosatiDipartimento di Informatica e Sistemistica, Univer-sita “La Sapienza”, Roma, Italy

�

ARTICLE

Dynamic Ontology Refinement

Authors: F. McNeill, A. Bundy, and M. Schorlemmer

Introduction

Creating a plan that is guaranteed to be executable ina certain domain depends not only on forming planscorrectly but also on having a perfect understandingof that domain. Unfortunately, developing this un-derstanding and representing it fully is possible onlyin small, static domains. In more complex environ-ments, plans may be based on incomplete or incorrectinformation and hence may be unexecutable. Interac-tion with the environment through attempted execu-tion of the plan leads to an enriched and fuller un-

derstanding of the environment but also leads to planfailure. This failure could be due to part of the theoryitself, such a missing axiom in a rule, or due to theunderlying ontology, such as a predicate with an in-correct arity. We therefore propose to devise a systemthat can dynamically incorporate this new knowledgeinto the theory as the plan is being executed.

Our system will be based around a central plan-implementation agent, who will control all the othercomponents of the system. This agent will firstlysend the theory, together with the goal, to the planner,which will then return a plan annotated with a justifi-


cation for each plan step. The plan-implementationagent will then use the plan justification, togetherwith further questioning of these agents, to discoverwhy this failure occurred and which part of the theoryis at fault. This problem point is then passed to the re-finement system, which will then correct the problemand replace it with the correction in the original the-ory. This refinement process will not only allow usto achieve a goal that would otherwise be unreach-able, but will also leave us with a domain theory thatis richer and more accurate.

Forming and Executing the Plan

We need to find a plan for achieving the goal and toannotate this with a justification for each step. Thereason this justification is required is that failure inthe plan execution can be immediately linked to aproblem in a particular part of the theory; namely,that part of the theory that was used to justify this planstep. We propose to use a state of the art planner suchas FF so that our system is capable of producing longand complex plans if necessary. However, using sucha planner will not provide us with information aboutthe inference rules and justifications behind each planstep. Therefore we intend to build a plan deconstuc-tor that will take the plan produced by FF and, usingthe theory, pseudo-execute it, at each stage annotat-ing the theory with the inference rule that was usedand the preconditions of that rule, together with theirjustifications.Once the plan-implementation agent has received theannotated plan, it will attempt to execute it by inter-action with other agents. For example, if the actionrequired in a plan step is to buy a visa, it will needto locate an embassy agent who is capable of fulfill-ing this requirement. It is through these interactionsthat any fault in the theory will come to light. Theseagents will have their own internal theories about howto perform these actions, and these may not match theplan-implementation agent’s theory.For example, the plan-implementation agent mayhave a visa represented in his theory as a simple, 0-

arity predicate, whereas the embassy agent may haveit represented as a 1-arity predicate visa(country)which takes a country as an argument. Or perhaps theembassy agent has a more complicated rule for buy-ing visas, which involves an extra precondition thatthe plan-implementation agent’s rule doesn’t have,for example, that an official invitation is required. Insuch cases, the actions will not be executed and theplan will fail. The plan-implementation agent canthen question the embassy agent further, using thejustification of the plan step, to find out exactly whichpart of the theory caused the failure and why.

Refinement Techniques

To create rules for specialising a theory or ontology,we looked first at rules for abstraction; that is, remov-ing detail from a theory. We inverted these rules toform anti-abstractions which can thus be used to adddetail to a theory:

� Predicate anti-abstraction - A single predicate isdivided into some number of predicates

� Domain anti-abstraction - Constants and func-tion symbols are divided up into different cases

� Propositional anti-abstraction - The arity of apredicate is increased

� Precondition anti-abstraction - Preconditionsare added to rules.

The refinement system will then select the appropri-ate type of refinement from those available to it, us-ing the information gleaned earlier about the problempoint. Once the refinement has been performed, it re-places the original in the theory, which can then bepresented to the planner.

Author Information

Fiona McNeill, Alan Bundy, and Marco Schor-lemmer Edinburgh University, UK, [email protected]

�



ARTICLE

MEXAR: Integrated AI Technologies to Support MARS EXPRESS

Mission Planning

Authors: G. Cortellessa, N. Policella, A. Cesta, and A. Oddi

Introduction

Space exploration missions require coordination of asignificant amount of activities. State of the art in-telligent planning and scheduling (P&S) technologycould potentially be of great help in supporting such acoordination. This work aim at showing an exampleof this technology in a support system for a specificmission scheduling problem related to the ESA pro-gram called MARS-EXPRESS [3].

MARS-EXPRESS is a space probe that will belaunched during 2003 and after six months will beorbiting around Mars for the following two years andmore. During the operational phase around Mars ateam of people, the Mission Planners, will be respon-sible for the on board operations of MARS-EXPRESS.They receive input from different teams of scientistsand cooperate with different specialists for variousspecific tasks (e.g., Flight Dynamics (FD) experts).Any single operation of a payload, named POR (Pay-load Operation Request), is decided well in advancethrough a negotiation phase among the different ac-tors involved in the process (e.g., scientists, missionplanners, FD experts).

The result of our study is a system called MEXAR thataddresses the memory dumping problem in MARS-EXPRESS. The specific problem that is addressedis defined as MEX-MDP and is described in detailin [7]. MEXAR is an interactive support systemthat may help mission planners in deciding policiesfor downlinking data to Earth during the temporalvisibility windows. The tool uses constraint-basedtechniques for representing the basic problem to besolved, namely the segmentation of on-board mem-ory in data packets during the visibility toward Earth.

The paragraph below introduces the two solvingalgorithms which have been developed: a greedy

heuristic and a local search procedure [7, contains acomplete explanation of these algorithms]. The fol-lowing section describes an important aspect of thiswork, the interactive functionalities developed to sup-port the user in his work [1, for more details].

The Packet Sequencing Algorithm

Scheduling problems such as MEX-MDP can be seenas a special types of Constraint Satisfaction Prob-lems (CSP) [6]. An instance of a CSP involves aset of variables � �� , a domain��

for each variable and a set of constraints �� s.t. � �� ! � � "� � �# � � ,which define feasible combinations of domain val-ues. A solution is an assignment of domain valuesto all variables which is consistent with all imposedconstraints.The CSP formalization of the MEX-MDP problem isbased on a partition of the temporal horizon $ �% & �('*) in a set of contiguous windows + �,��- � �% .0/ � . ��)21 .3/ � & �*45��-768� 9 . 6�:;� � . 6 )<1>=?�@ � � �BAC� . ��D $*� , such that 4FE�HG� - � ��$ . We con-sider a set of decision variables I �KJ 6 that represent theamount of data to be dumped from packet store LNM �in the window -O6 . The MEX-MDP contains differentkinds of constraints: (a) Given the characteristics ofthe packet stores the data must be downlink accordingto a FIFO philosophy; (b) the amount of data for eachpacket store does not exceed its capacity; (c) a finiteamount of data can be dumped in each transmissionwindow (finite transmission rate).All the proposed algorithms work over two levels ofabstraction: (1) the planning level, where the wholeset of decision variables are instantiated taking intoaccount the different constraints; (2) the schedulinglevel, where a sequence of memory dump operations


is generated over the communication links respectingthe constraints imposed over all the windows - 6 .In order to find an optimal solution we choose to re-alize a heuristic optimization strategy based on localsearch which is able to improve an initial solutiongiven as an input: Tabu search [4, 5]. The tabu meta-heuristic is founded on the notion of a move. A moveis a function which transforms one solution into an-other. For any solution

�, a subset of moves applied

to�

is computed. The result is the neighborhood of�

. The algorithm proceeds selecting, at each step,the best solution in the neighborhood, with respect toan objective function, till a fixed number of steps aremade without finding better solutions.

Figure 1: MEXAR layout

In MEX-MDP the move consists in bringing forwardsome data (for example data contained in observa-tions with the smallest volume of data) and delayingother ones; this should improve in many cases the ob-jective function (mean turn over time).

Mexar Interactive Functionalities

The MEXAR functionalities that are designed for theusers are summarized in Figure 2. As expected theproblem solving activity is central in the system. Thisfunctionality is guaranteed by the automated servicescentered on the constraint-satisfaction methodology(CSP) described above.In the figure 2 we show the flow of control duringthe use of the functionalities. It is possible to iden-

tify an activity that aims generically at defining a sin-gle problem. At present it simply consists of loadinga problem description from a file, it can be also bereplaced by a more complex incremental functional-ity that could be well coupled with the CSP model-ing used. The definition of a problem is followed byits solution according to the different algorithms pro-duced in our work. A different functionality allowsto refine the current problem. This activity consistsin deleting some of the Payload Operation Requests(PORs) from the associated timelines. This can beuseful to experiment different loads on the resourcesin specific intervals of the solution. This functionalityintroduces a cycle among these activities that couldbring the user to incrementally refine new MEX-MDP

problems. As shown we group these functionalitiesin an interaction layout called Problem Analyzer (seeFigure 2).Once a problem to solve is defined we can attack itwith different solution methods. Figure 1 shows anexample of a solved instance of MEX-MDP as it ispresented to the user.

Figure 2: MEXAR Interactive Environments

The availability of a portfolio of problem solvingprocedures has suggested the idea of involving moredeeply the user in the solution process. This has beenpursued by creating an environment in which it ispossible to save different solutions and, in addition,the user can guide search on how to improve the solu-tions applying different algorithms. We call this pro-cess solution space exploration. This aspect is strictlyconnected to the availability of evaluation metrics onthe solutions as discussed in [2]. The idea behind thesolution explorer is the one that the user can generatean initial solution, save it, try to improve it by lo-cal search, save the results, try to improve it by local



search with different tuning parameters and so on. Inthis way, it is possible to generate paths in the searchspace. The user can restore one of the previous so-lutions and try to improve it with a local search withdifferent parameters, etc. In this way he generates atree of solutions. This procedure can be repeated fordifferent starting points generating, in this way, a setof trees. Using at the same time the evaluation ca-pability on a single solution and its own experiencehe can generate different solution series, all of themsaved, and, at the end, choose the best candidate forexecution.

Bibliography

[1] Cesta, A.; Cortellessa, G.; Oddi, A.; and Poli-cella, N. Interaction Services for Mission Plan-ning in MARS EXPRESS. In Proceedings of the3rd International NASA Workshop on Planningand Scheduling for Space, 2002.

[2] Cesta, A.; Oddi, A.; Cortellessa, G.; and Poli-cella, N. Using AI Techniques to Solve MEX-MDP. Technical Report MEXAR-TR-02-08,ISTC-CNR - Planning and Scheduling Team,Rome, Italy, 2002.

[3] ESA. European Space Agency, MARS

EXPRESS Web Site. http://sci.esa.int/

marsexpress/, 2002.

[4] Glover, F. Tabu Search – Part I. ORSA Journalof Computing 1:190–206, 1989.

[5] Glover, F. Tabu Search – Part II. ORSA Journalof Computing 2:4–32, 1990.

[6] Montanari, U. Networks of Constraints: Fun-damental Properties and Applications to Pic-ture Processing. Information Sciences 7:95–132,1974

[7] Oddi, A.; Cesta, A.; Policella, N.; and Cortel-lessa, G. Scheduling Downlink Operations inMARS EXPRESS. In Proceedings of the 3rdInternational NASA Workshop on Planning andScheduling for Space, 2002.

Author Information

Gabriella Cortellessa � , Nicola Policella ,Amedeo Cesta, and Angelo Oddi ISTC-CNR, Italian National Research Council, Viale Marx15, I-00137 Rome, Italy, {corte,policella,cesta,oddi}@ip.rm.cnr.it1 is also Ph.D. student in Cognitive Psychology atUniversity of Rome “La Sapienza”2 is also Ph.D. student in Computer Science at Uni-versity of Rome “La Sapienza”

�

ARTICLE

RDPPlan: an Extension of DPPlan for Planning with IntervalResources

Authors: M. Baioletti, A. Milani, and V. Poggioni

RDPPLan is a model for planning with quantitativeresources. It is based on DPPPlan [1], a plannerwhich uses a non directional search algorithm on theplanning graph.

Most models of planning with resources, like [3], [4],[5], [6], [7] and [8], assume that an exact value canmodel the continuous quantities describing, in thereal world, a given resource. In other words these

models cannot deal with more realistic situations inwhich quantities are not known exactly. The RDP-Plan model allows one to manage domains where pre-conditions and effects over quantitative resources canbe specified by intervals of values; in addition mixedlogical/quantitative and pure numerical goals can bespecified.

Instead of initializing a resource with a unique real


value, we allow for the specification of a real inter-val �� as the range of the initial value of the resource� . The planner operates in an underspecified domainin which the value of some resource is not exactlyknown, but it is bound to be in an interval. Let us sup-pose that we do not know exactly how much gasolineis in the tank of our car: we just know for sure that thereal amount is between 5 and 10 liters. Similarly it ispossible to have underspecified effects of some oper-ators: the value which is added (subtracted), multi-plied (divided) or assigned to the current value of aresource is not exactly known, but only a lower andan upper bound is specified. Let us imagine that wedo not know which is the exact consumption of thecar in the previous example: all we know is that thecar travels from 10 to 15 kilometers per liter.We define two intervals associated to each resource� and each time level

.: �� and

� �� that respectivlyrepresent the Realized Interval and the Desired Inter-val of resource � at time-step

.. �� is initialized with

� � and it is updated, at each step, by the effects onthe resource � of the actions already inserted in theplan;

� �� , instead, contains all the admissible valuesfor the resource � that allow the execution of all theactions selected at time level

..

In this model the necessary and sufficient conditionfor a plan to be a solution of a given planning prob-lem is that for each resource � and time level

.,

�� . In other terms, a solution plan mustsolve every possible problem that is allowed by theconstraints specified in the initial state and in the ef-fects description.We use the same concept of simultaneous executabil-ity as expressed in [2] and in [6]. We say thatthe actions � �� E are simultaneously exe-cutable if, for every permutation of � J �HG � J�� J E , eachaction is executable in the order defined by the per-mutation and the effect over resources is always thesame. As a straightforward consequence, an assign-ment on resource � is not simultaneously executablewith any action changing � and an additive opera-tor (increase, decrease) is not simultaneously exe-cutable with any multiplicative operator (scale–up,scale–down). Moreover we have defined a provably

correct method that can allow us to calculate the De-sired Interval for the simultaneity execution of A ac-tions that have preconditions and/or effects on thesame resource � .In order to achieve the goals over resources, wehave defined strategies to solve ”pure numerical prob-lems“, i.e. with goals only on resources. Such strate-gies are combined with the ones for solving logicalgoals, by evaluating the difficulties of resources andlogical goals and selecting the most difficult goal tosolve.When we work with resources as intervals, we haveto handle with real intervals whose width in gen-eral grows, except when an assignment is performed.Moreover note that if the width of the Realized In-terval is large, it is more difficult that the solutioncondition �� "� �� holds.The action choice criterium takes into account thedistance between the corresponding Desired and Re-alized intervals and their width. In particular wehave defined two preference functions, one for inter-val widths, and one for distance between intervals,and the algorithm chooses the action that maximizesa linear combination of these functions.Investigations and experiments are planned in orderto develop more accurate heuristics and strategieswhich take resources into account. Moreover, in or-der to provide a meaningful evaluation, it will alsobe required the development of a set of significantbenchmarks for planning domains with interval re-sources. Finally it is worth investigating further ex-tensions to the resource model more accurate with re-spect to the uncertainty in the real world e.g. intervalswith given probability distribution over resource val-ues and fuzzy quantities.

Bibliography

[1] Baioletti M., Marcugini S. and Milani A. DP-Plan: an Algorithm for Fast Solution Extractionfrom a Planning Graph, In Proc. of AIPS-2000

[2] Fox M. and Long D., PDDL 2.1: An Extensionto PDDL for Expressing Temporal Planning Do-mains, 2002



[3] Koehler J. Planning under resource constraints.In Proc. of ECAI-1998

[4] Chien S. et al., ASPEN Automated Planningand Scheduling for Space Mission Operation, InProc. of SpaceOps 2000

[5] Wolfman S. and Weld D., The LPSAT Engine andits Application to Resource Planning, In Proc. ofIJCAI 1995

[6] Rintanen J. and Jungholt H. Numeric State Vari-ables in Constraint-Based Planning, In Proc. ofECP-1999

[7] Currie K. and Tate A., O-Plan: The open plan-ning architecture, In Artificial Intelligence 52,49-86, 1991

[8] Laborie P. and Ghallab M. Planning withsharable resource constraints, In Proc. of IJCAI1995

Author Information

Marco Baioletti Dip. Met. Quantitativi, Univer-sita di Siena, Italy

Alfredo Milani Dip. Mat. e Informatica, Univer-sita di Perugia, Italy

Valentina Poggioni Dip. Informatica e Aut.,Universita di Roma3, Italy

�

ARTICLE

Extending Operator Induction to Provide Full Method Sets forHierarchical Planning Domains

Author: N.E. Richardson

Abstract Modelling a world for efficient HTN plan-ning by hand is a lengthy and time consuming pro-cess. Our knowledge acquisition tool, GIPO (Graph-ical User Interface for Planning with Objects), offersGUI abstraction allowing the user to concentrate onthe model rather than a language syntax. Using GIPOwe aim to provide automatic operator induction forplanning domains with an hierarchical sort structureso that complex hierarchical operators can be con-structed.

At present we can induce operators for flat domains.Using GIPO we interactively construct a sequence ofactions to arrive at some goal state. Then we inputa partial domain (with no operators) and using op-maker, the induction tool in GIPO, we obtain a setof operators to complete the constructed task. Theseoperators may not be accurate at this stage.

We can also induce operator sequences to constructlow level methods (hierarchical operators for HTNdomains). The name method implies that these aredifferent recipies for achieving similar or relatedtasks and as such they often repeat actions or ac-

tion sequences. Issues arising from using repetitionin the sequences are that the same operators are in-duced more than once and, as it is desirable to haveonly one operator per action, we can use repetition togeneralise the operator. The debate is then to find away of revising the operator each time a new versionis induced or employ a learning and revision process.

We propose two systems to compare induced oper-ator specifications and generalise those descriptions.Comparing operators will give us a category of differ-ences. Some differences are allowable in the presentsystem. For example an operator can be inducedwithout a conditional clause and if the same opera-tor is used later in the sequence with a condition thenthe conditional clause is added to the original oper-ator. We would want to be able to merge operatorsthat have different names but are otherwise identical.Some operators will be more general than others butwe recognise that over-generalisation can mean thatthe operator is not sufficiently expressive.

Operators describe objects’ transitions from one setof substates prior to the action represented, to another


set of substates after the action. In generalising an op-erator the aim is to make it applicable to a wider setof objects but this can be overdone. A difficulty ingeneralising operators is to limit extent that the oper-ator’s expressiveness is compromised.Generalising operators and revising the operators inthe system as new ones are generated will allow usto constuct the higher levels of the method hierarchywhich are built from other methods as well as primi-tive operators. When we have these systems in place

it will be possible to have full method induction inGIPO for hierarchical domains.

Author Information

N.E. Richardson School of Computing and En-gineering The University of Huddersfield, Hudder-sfield HD1 3DH, UK, [email protected]

�

ARTICLE

The SimPlanner

Authors: O. Sapena and E. Onaindıa

Introduction

Off-line planning generates a complete plan beforeany action starts its execution [3]. This forces tomake some assumptions that are not possible in realenvironments like, for example, that actions are un-interruptable, that their effects are deterministic, thatthe planner has complete knowledge of the world orthat the world only changes through the executionof actions. On the other hand, on-line planning al-lows to start execution while the planner continuesworking in order to improve the overall planning andexecution time. Nowadays there are only some fewapproaches for planning in dynamic environmentsand/or with incomplete information [2]:

� Conditional planning: this approach tries to con-sider the possible contingencies that can occur inthe world.

� Parallel planning and execution: this approachseparates the planning process from the execution.

� Interleaving planning and execution: this ap-proach allows quick and effective responses tochanges in the environment.

SimPlanner is an integrated tool for planning and ex-ecution, and it is based on this latter approach. Thistool is thought to be used in real environments such

as the intelligent control of robots. However, it hasinitially been implemented as a simulator in order tocheck its behavior without having to integrate it indifferent several domains. SimPlanner is made up ofthree components: an on-line planner, a monitoringmodule and a replanner.

The on-line planner

The on-line planner is responsible for generating, inan incremental way, a plan to achieve the goals. Assoon as the planner calculates the first action, the plancan start its execution. From this moment on, theplanning and execution processes keep on working inparallel while no unexpected event is detected; other-wise, the execution must wait for the planner to makethe necessary modifications in the plan.The planner is based on a depth-first search, with noprovision for backup. The planning decisions (in-ferred actions) are consequently irrevocable. Theplanning algorithm uses heuristic functions to com-pute an approximate plan (AP) for each goal indepen-dently. Then, a conflict-checking mechanism detectsconflicts among actions in the approximate plans andselects the action from the AP that minimizes thenumber of conflicts. The selected action is insertedat the end of the plan, and the current state is updatedthrough its execution. This algorithm is iteratively



executed until all top-level goals are achieved [4].

The monitoring module

Monitoring is the process of observing the world andtrying to find discrepancies between the physical real-ity and the beliefs of the planner [1]. Basically, thereexists two types of plan execution monitoring [2]: ac-tion monitoring checks the validity of the action pre-conditions before it starts its execution and also thatits effects have taken place as expected. The environ-ment monitoring tries to capture information from theexternal world that conditions the remaining planningprocess. Monitoring is, therefore, domain-dependent.Since SimPlanner is being used at the moment as asimulator, this information is input to the system bythe user. The user is who decides what informationthe robot receives and which unexpected events thatoccur in the world are communicated.

The replanner

When an unexpected event is detected, the calculatedplan is checked in order to assure it is still valid [1]. Ifthis is the case, the execution simply continues. Oth-erwise the replanning module is invoked. The replan-ner tries to reuse as much of the calculated plan aspossible without losing the quality of the final plan.It uses a heuristic function to find out which is thebest reachable state through the actions in the origi-nal plan. If there are many reusable actions, the plan-ning process will save a lot of computation time. Inthe worst case, a new plan will be computed from

scratch. The replanner overhead is very small so itis worth trying to reuse the plan rather than planningfrom scratch [5].

Bibliography

[1] G. De Giacomo and R. Reiter, ‘Execution mon-itoring of high-level robot programs’, Princi-ples of Knowledge Representation and Reason-ing, 453–465, (1998).

[2] K.Z. Haigh and M. Veloso, ‘Interleaving plan-ning and robot execution for asynchronous userrequests’, Papers from the AAAI Spring Sympo-sium, 35–44, (1996).

[3] M.E. Pollack and J.F. Horty, ‘There’s more tolife than making plans: Plan management in dy-namic, multi-agent environments’, AI Magazine,20(4), 71–84, (1999).

[4] O Sapena and E. Onaindia, ‘Domain-independent on-line planning for STRIPSdomains’, Proceedings of Iberamia-02, LNAI,2527, 825–834, (2002).

[5] O Sapena and E. Onaindia, ‘Execution, moni-toring and replanning in dynamic environments’,Workshop on On-Line Planning and Scheduling,AIPS-02, (2002).

Author Information

Oscar Sapena, Eva Onaindıa Departamentode Sistemas Informaticos y Computacion, Univer-sidad Politecnica de Valencia, Spain, {osapena,onaindia}@dsic.upv.es

�


ARTICLE

Answer Set Planning with DLV�

Author: A. Polleres

Introduction

The knowledge based planning system DLV�

imple-ments answer set planning on top of the DLV system[1]. It is developed at TU Wien and supports thedeclarative language

�[3] and its extension

��[4].

The language��

is syntactically similar to the actionlanguage

�[6], but semantically closer to answer set

programming (by including default negation, for ex-ample).

� �offers the following distinguishing fea-

tures:

- Handling of incomplete knowledge: for a fluent � ,neither � nor its opposite �� need to be known inany state.

- Nondeterministic effects: actions may have multi-ple possible outcomes.

- Optimistic and secure (conformant) planning:construction of a “credulous” plan or a “sceptical”plan, which works in all cases.

- Parallel actions: More than one action may be ex-ecuted simultaneously.

- Optimal cost planning: In� �

, one can assign anarbitrary cost function to each action, where thetotal costs of the plan are minimized.

An operational prototype of DLV�

as well as sam-ple encodings of planning domains in the system areavailable at http://www.dlvsystem.com/K/.

Underlying Concepts

Action language� �

In� �

transitions aredescribed in a declarative way by means ofcausation rules, e.g. caused at( � ) aftertravel( !�� ), at( ). Furthermore, the lan-guage offers constructs to express executability andnon-executability of actions, ramifications, non-determinism and allows to assign costs to actions.Planning problems in this language are transformedto a logic program which is then evaluated under theanswer set semantics.

Answer Set Programming Answer Set pro-grams are logic programs in a syntax similar to Pro-log which allow for negation as failure in rule bod-ies and disjunction in rule heads. In our approach,the minimal models of these programs under the socalled Answer Set Semantics [5] correspond one-to-one to the plans of the resp. planning problemspecified in

� �. This view of models representing

plans can be partly compared to planning using SAT-Solvers.

System architecture

The architecture of DLV�

is outlined in Figure 1. Theinput of the system consists of domain descriptions(DLV

�

files) and optional static background knowl-edge specified by a logic program.

Control FlowData Flow

Plan Printer

Controller

Datalog Parser

Parser

Plan Generator

DLV Core

DLV

Core

Plan Checker

input

KnowledgeBackground

Figure 1: DLV

System Architecture



The Controller first invokes the Plan Generator,which translates the planning problem at hand intoa suitable program in the core language of DLV (dis-junctive logic programs under the answer set seman-tics). Then, the DLV kernel is invoked to solve thecorresponding problem. The resulting answer sets (ifany exist) are fed back to the Controller, which ex-tracts the solutions to the original problem from theseanswer sets and transforms them back to the originalplanning domain.If the user specified that secure/conformant planningshould be performed, the Controller then invokes thePlan Checker which verifies by another evaluation ofa logic program whether this plan is in fact also se-cure/conformant.Finally, the solutions found by the Generator (and op-tionally verified by the Checker) are translated backinto user output and printed.Details about the transformations mentioned abovecan be found in [2]. Performance and experimentalresults are reported in [2, 4].

Bibliography

[1] T. Eiter, W. Faber, N. Leone, and G. Pfeifer.Declarative Problem-Solving Using the DLVSystem. Logic-Based Artificial Intelligence, pp.79–103. Kluwer, 2000.

[2] T. Eiter, W. Faber, N. Leone, G. Pfeifer, andA. Polleres. A Logic Programming Approach toKnowledge-State Planning, II: the DLV

�

System.

Technical Report, December 2001. To appear inArtificial Intelligence.

[3] T. Eiter, W. Faber, N. Leone, G. Pfeifer, andA. Polleres. A Logic Programming Approach toKnowledge-State Planning: Semantics and Com-plexity. Technical Report, December 2001. Toappear in ACM Transactions on ComputationalLogic.

[4] T. Eiter, W. Faber, N. Leone, G. Pfeifer, andA. Polleres. Answer Set Planning under ActionCosts. Technical Report, October 2002.

[5] M. Gelfond and V. Lifschitz. Classical Negationin Logic Programs and Disjunctive Databases.New Generation Computing, 9:365–385, 1991.

[6] E. Giunchiglia and V. Lifschitz. An Action Lan-guage Based on Causal Explanation: PreliminaryReport. In AAAI ’98, pp. 623–630, 1998.

Author Information

Axel Polleres Institut fur Informationssysteme,TU Wien, A-1040 Wien, Austria, [email protected], [email protected] work was supported by FWF (Austrian Sci-ence Funds) under the projects P14781 and Z29-INF.Furthermore, the author thanks the OGAI (AustrianAssociation for Artificial Intelligence) and the OCG(Austrian Computer Society) for sponsoring travelcosts to the PLANET Summer School 2002

�


ARTICLE

Planform: An Open Environment for Building Planners

Authors: T.L. McCluskey, M. Fox, and R. Aylett

Project Overview

The PLANFORM project aimed to develop a high levelresearch platform for the systematic construction ofplanner domain models and automatically configuredplanning algorithms.Our objectives were, briefly:

� To assemble tools within an open environment forthe acquisition and modelling of planning domainmodels;

� To create languages for modelling of planning do-mains and to specify characteristics of plannersleading to the configuration of purpose-built plan-ning engines;

� To create a tool which synthesises a domain modeland a planning engine into a planning application;

� To evaluate the approach using realistic problemdomains.

During the project, the emphasis changed slightly sothat some of objectives developed in a different direc-

tion. Rather than abstractly defining planning algo-rithms, we decided to create a library of algorithmsand use domain analysis technology to design andconfigure a planning application. We perceived theknowledge acquisition bottleneck to be a significantproblem for AI Planning and concentrated more re-source than planned on this area.We have made substantial progress towards our ob-jectives. We have developed an environment enablingthe acquisitition and modelling of planning applica-tions and the configuration of planning engines suitedto those applications. Although time limitations pre-vented us from completing all aspects of the final in-tegration and evaluation we achieved our main objec-tives and laid a strong foundation for development ofthe project.The project consisted of a model engineering phaseand a planner engineering phase. Huddersfieldand Salford collaborated closely on modelling andknowledge acquisition issues, whilst Huddersfieldand Durham collaborated on the development of partsof the domain modelling tools. Durham worked onthe planner engineering phase of the project.

Application

Domain

Tools

Static and Dynamic Analysis

Planning

Heuristics

Planners

ToolAcquisitionKnowledge

General

Generic

Configuration

Application

Efficient

Model Engineering Phase Planner Engieenring Phase

Repres-Internal

entation ofDomain

Figure 1: Architectural Breakdown of Planform



Research at the University of Huddersfield

The Huddersfield contribution emphasised tools cre-ation and integration aspects of the project. Tools foracquisition, validation and domain modelling weredeveloped, in collaboration with colleagues at bothSalford and Durham. As part of this effort the Hud-dersfield team researched and developed a GUI-basedenvironment called GIPO1. In addition, as the initia-tors of PLANFORM, Huddersfield performed a cen-tral administrative function, producing the main ex-ternal website and also an internal website with ad-ditional resources such as Project Meeting minutes.NB: in the text below ‘the PLANFORM website’ refersto http://scom.hud.ac.uk/planform.

The first phase of the Huddersfield contribution wasconcerned with applications encoding and languagedevelopment. During this phase three application do-mains were used to explore the adequacy of the mod-elling languages �� and �� [20, 19]. �!��is a hierarchical version of �!�� enabling the mod-elling of planning domains as hierarchical task de-compositions. Salford and Huddersfield collaboratedon this effort with Salford working on the modellingof some existing domains using the two languages.Huddersfield tackled the problem of encoding the air-craft landing scheduling problem supplied by MarkWatson of the National Air Traffic Services. The con-straints (e.g. separation times for each type of air-craft) were encoded in �!�� [18] and �!�� . �!��was found adequate for all three domains, althoughthe encoding did demonstrate some required changes,and overall the pressing need for tool support.

The standard languages for communicating planningdomain descriptions are the PDDL variants [1]. In or-der to be able to experiment with our domain mod-els using exisiting planning technology we there-fore created tools to map between �� and PDDL.These tools continued to be developed throughout theproject allowing planners, and other tools that receiveinput in PDDL form, to be integrated into our environ-

ment. There are some interesting technical aspects ofthe mapping discussed in [27].Language manuals for �� and �!�� were main-tained during the project [12] and an online help fa-cility was constructed (see the PLANFORM website).Drawing on previous development work (e.g. [19]),we assembled tools that automate the syntactic andsemantic analysis of �!�� domain models. Analy-ses include ensuring that invariant properties of themodel are maintained and that syntactic rules are ob-served.The second phase of the project was concerned withdevelopment of the GUI and tools environment. Thefocus was on building and integrating knowledge ac-quisition and modelling tools for AI planning into anopen environment. The GUI and some of the toolsdescribed above were built in Java. Others were im-plemented in Prolog and it was necessary to integratethese via Sicstus Prolog’s JASPER interface.The JavaCup parser generator method was used torepresent the syntax rules of �!�� and to generatea parser for the language. This formed the input toolin GIPO (Graphical Interface for Planning with Ob-jects). The knowledge acquisition part of the toolwas structured using the method outlined in earlierwork [20]. The methods direct the user to define ob-jects, object sorts, relations and properties, classesand constraints on object situations, problems in theform of task specifications, and finally operators builtfrom these components. The design of the interfacewas based on the need to minimise the use of syntax,and use object rather than predicate centred ideas.Once users have entered all parts of a domain modelthey can utilise modelling tools to remove bugs andexperiment with the encoding. We created andadapted the following tools.

� plan stepper: This allows the user to pick actionschemas and apply them to a state, until a desiredgoal is reached. It is useful for identifying errorsin operators and operators sets.

� internal planning engines: this allows our ownin-house planning engines to be connected up to

1http://helios.hud.ac.uk/planform/gipo


GIPO. Sample tasks can be executed and the re-sulting solutions displayed.

� interface for external planning engines: This al-lows external planning engines to be ‘bolted’ intothe environment. The planner needs to be able toinput domain models in PDDL (from GIPO), andoutput solution in a prescribed format. Again, theresulting solutions are displayed through GIPO.

� a random task generator: This inputs the currentdomain model and randomly generates tasks to beused with a planner.

� an animator: After a domain model has been en-tered, and the planning engine has solved a taskwithin that model. the animator can be used totrack the transitions of each of the objects whichstarted in the initial state.

In the third phase, integration and evaluation, thetools outlined above were integrated into GIPO [28].The software was released and demonstrated atECP’01, and again at AIPS’02. It is available onUnix, Linux and Windows platforms from the PLAN-FORM website. As an initial indication of GIPO’simpact, the Huddersfield website recorded 147 exter-nal downloads of the system in the period November2001 - March 2002.The environment has been tested using a range ofcommon domains (details are in the resource sectionof the website). Further, GIPO was used as a teach-ing tool in a second year introductory course in Artfi-cial Intelligence. GIPO alleviates many user interfaceproblems by adopting an object modelling approachwhich seems natural to non-expert users. To amelio-rate the use of GIPO by non specialists the followingissues were explored:

� The use of an inductive approach to capturing op-erators2 . The opmaker tool was created whichoutputs a set of operators given a partial domainmodel and an example solution sequence [22, 21].

� The use of generic types to suggest planning de-sign patterns. These ideas were developed as

part of the Durham PLANFORM project, and Hud-dersfield is working in collaboration with col-leagues at Durham to integrate design patternsinto GIPO [26].

The final phase of our work has been to extend the in-ternal language and the surrounding tools from �!��(version 1) to �!�� (version 2). �!�� extends�� in two major ways: HTN operators can beused, and sort constraints can be put on each levelof the sort hierarchy meaning that objects of a prim-itive sort inherit all the constraints up the hierarchy.This modelling approach is being tested using a large‘Translog’ domain imported from a transport logisticdomain constructed by members of the University ofMaryland.The priorities for future development of the Hudder-sfield contribution are:

� The development of a suite of planning designpatterns and their integration with the GIPO tool;

� The evolution of the L�� .� L L � � tool into a gen-eral mixed-initiative plan authoring tool;

� Integration of the operator induction techniqueswith a plan authoring interface so that operatorspecifications can be induced and refined interac-tively.

� The development of the �!�� representation lan-guage to be on the expressive level of PDDL2.1(temporal representations), thereby enabling GIPO

to support the modelling of temporal planning do-mains.

Research at the University of Salford

Part of the set of objectives for the PLANFORM

project was to make AI planning technology accessi-ble to non-experts. In pursuit of this objective, workat Salford was based on the idea of the KnowledgeLevel [25] and Models of Expertise [6] as articulatedover many years in the KBS community, in which theproblem level is modelled separately from the design

2This work was primarily work undertaken in conjunction with a PhD student



level. Research in KBS technology has shown thatsupport for domain experts is feasible if it is based onthe generic task concept [7], and much earlier workhas been carried out round the generic task of diag-nosis [24]. Oddly, little of this has been applied toAI Planning. The basic idea is that a generic taskincorporates a skeletal model at the knowledge-levelwhich can then be used to direct a computer-basedknowledge acquisition process with a domain expert.Thus, to support domain experts, it was seen as nec-essary to build a knowledge-level tool as part of thePLANFORM environment, incorporating generic taskcomponents for planning, and supporting knowledge-level construction of the planning domain rather thanforcing the use of the domain design language, �!�� ,used internally. However such a tool would nec-essarily link to the �� GIPO tools being devel-oped at Huddersfield (see Figure 1) and thus output�!�� , so it was therefore vital to understand howplanning problems would be represented in �!�� . Astrong link with Huddersfield was built through mod-elling activity in which the objective was to under-stand this mapping and derive some constraints onthe knowledge-level interface, which was oriented to-wards users with no AI planning knowledge but withexpertise in a particular planning domain. A numberof domains were modelled including the multi-robotDrumstore from earlier work at Salford.Planning ontologies were identified as a key founda-tion for such a knowledge-level tool, and the meansby which the skeleton model mentioned above couldbe embodied. A survey of work in the field was car-ried out and the possibility of incorporating an exist-ing ontology such as CYC( http://www.cyc.com/cyc-2-1/cover.html.) was investigated but limited time didnot allow its use.The third component researched as the basis for aknowledge-level tool was the requirements of the do-main expert, and here an accessible domain was for-mulated (EVENTUS – arranging a weekend outing fora visiting researcher) and a knowledge acquisition ex-ercise was carried out with four people. The exer-cise was repeated with the robot Drumstore problem.Data was analysed for concept coverage and for in-

terface design issues, looking at the process a humanexpert goes through in conceptualising the domain.The details of this process are described in [2].The KA tool was constructed in Java, and in its firstversion supported passive model construction by theexpert with support from a small hand-coded ontol-ogy for the domains already investigated. In its sec-ond version, an active question-driven process wasadded based on the key planning concepts of Task,Agent and Object [5].The methodology embodies two successiveextraction-refinement processes: protocol to problemspecification; and problem-specification to conceptu-alisation. A part of the KOD (Knowledge OrientedDesign) method [30] was applied to obtain an accu-rate process for knowledge acquisition and to buildthe conceptual model through a set of examples andscenarios.The output of the KA-Tool is �!�� , which can thenbe loaded into the GIPO tool created at Huddersfield.By the formal end of the project, it was possible togenerate the world model in �� , and since then ithas become possible to generate planning operators,seen by most people as a key problem in formulatinga planning domain description. In the 26 months ofthe project, it was not possible to carry out any exten-sive evaluation programme, but it is proposed to carryon with the work for a limited period informally withthis as the key task.The modelling exercise enabled the development ofstrong links with Huddersfield, where frequent (al-most weekly) visits were made at some points. Atwo-week visit to Durham was also organised tostrengthen understanding of the requirements of thegenerative planning back-end.Further development of the KA-Tool is still being car-ried out in house, with the generation of Planning Op-erators now possible. If sufficient resources can befound, the next steps would include:

� Incorporation of a large ontology, such as CYC,into the KA tool;

� Integration of Durham’s generic types into this on-tology;


� Full integration of the KA tool with GIPO.

Two publications were generated jointly with theHuddersfield team (Simpson et al 2001a, 2001b)and two more by the Salford team on their own [4]and [2]. One further paper is under review [3] and ajournal article is in preparation.

Research at the University of Durham

The Durham contribution to PLANFORM focussed onthe development of planner configuration technology.The objective was to develop techniques by which,given a domain model elicited from a user, plannercomponents suited to the domain could be automat-ically identified and configured into a purpose-builtplanning system. Our approach has been to maintaina library of planner components, including heuristics,specialised solution strategies and problem-specificcontrol rules, and to access these by means of pattern-matching techniques once patterns have been identi-fied in the domain model.The techniques used to identify domain patterns arebased on static domain analysis algorithms devel-oped at Durham prior to the start of the PLANFORM

project [8]. The objective in the project was to extendthese algorithms to enable the recognition of generictypes and associated patterns of behaviour in plan-ning domains, and to associate these generic patternswith special purpose solution strategies [14, 13, 16].Briefly, a generic type is a collection of types shar-ing some fundamental behaviour. For example, thegeneric type of mobility contains all types of objectsthat are capable of movement while the generic typeof construction contains all types that are capable ofbeing combined into compounds (and subsequentlyrecovered by destruction of the compound).When a generic type is present its associated patternsof behaviour are present and these can be used bothto assist a domain designer in refining the model andto suggest appropriate solution approaches.The analysis techniques developed at Durham canidentify certain key generic types and associated pat-terns. For example, the pattern associated with mo-

bility comprises the mobile types, their maps and thepredicate that defines locatedness of a mobile objecton its map. A specific problem associated with mo-bility is route-planning, and a special purpose solu-tion strategy suited to this problem can be to exploittravelling saleman heuristics. We were able to auto-mate the configuration of planners with specialisedroute-planning capabilities enabling route-planningsub-problems to be handled using specialised ap-proaches instead of by search [9, 13, 10].The following software was developed at Durham aspart of the PLANFORM project:

� Versions 4 and 5 of the STAN planning system.STAN [9, 10] performs generic type analysis andconfigures a special purpose planner suited to theassociated generic patterns;

� Extensions to TIM [8, 15] to recognise a range ofgeneric types in a domain model;

� The TIM API providing access to the generic typeanalyses performed by the TIM system, enablingtheir exploitation by other planning systems. TheAPI is being exploited by other researchers in theinternational community;

� The OODL domain modelling language, support-ing the construction of domains around generictypes, and associated domain modelling toolDRAUGHTSMAN. These were developed by anMSc student and contributed to our collaborationwith Huddersfield on the GIPO tool.

In the scope of the project a number of other generictypes have been identified (for example, construction,resource-allocation, portability and others) and asso-ciated with specialised solution strategies. The con-figuration problem becomes complex when severalgeneric patterns co-occur and their solution strate-gies must be integrated, and we have not completedthe work required to support arbitrarily complex pat-terns of integration. We have, however, categorisedthe forms of integration that need to be handled andmade progress with configurations based on severalof these categories [17, 11, 9].



The library contains stored parametrized solutionstrategies (such as travelling salesman solvers, multi-processor scheduling heuristics etc) appropriate forsub-problems that commonly arise in planning do-mains. We do not try to guarantee complete cover-age of all such sub-problems – the configuration sys-tem defaults to search if no generic types or suitablelibrary components can be found. At present the li-brary contains only one solution strategy per iden-tified generic type, so extraction of a suitable stat-egy is simple. In general the extraction problem ismore difficult because there may be different formsin which generic patterns arise and these might needto be matched in some intelligent way against the li-brary. We have not explored this issue in the scopeof PLANFORM. A recent extension we have made tothe library is the addition of generalised control ruleswhich can be selected and instantiated to fit a specificproblem domain [23].Most recently our work in these areas has consideredthe use of generic patterns as a basis for the devel-opment of planning design patterns. Using these, adomain construction tool can prompt the user for thecomponents of generic patterns in a way that makesit simple for the user to enter that information. Initialwork on a tool capable of supporting this idea wasdone by an MSc student [29] who was temporarilyemployed on the PLANFORM project at Durham. Thework was continued in collaboration with the Hud-dersfield site which has focussed on the developmentof tool support [26].The planner configuration approach has been testedby entering a hybrid planning system, STAN ver-sion 4, into the international planning competitionin 2000. STAN 4 can automatically detect mobil-ity and resource-dependence patterns in planning do-mains and can extract route-planning and resource-allocation strategies from its library. Selected strate-gies are integrated, by means of a simple constraint-based interface, to a forward-search-based plan-ner [10]. STAN 4 was one of the prize-winning sys-tems, selected for the promise it showed in utilizingnovel approaches to solving complex planning prob-lems. Its plan quality was generally superior to that

produced by the other competing systems.Work remains to be done on increasing the sophisti-cation of the integration techniques that support co-ordination of different specialised solution strategieswithin the overall framework. An important aspectof making the configuration tools available to thegeneral planning community is to provide a cleanmeans of access to the library, pattern recognitiontechniques and planner interface. A priority for fu-ture development is to supply an API to the suite oftools we have, which other planning systems couldexploit. Although we made progress with the designand implementation of such an API we did not com-plete its implementation and it remains a topic for fu-ture work.

Overall Conclusions

The PLANFORM project set out to construct anopen environment for planner development, bring-ing knowledge acquisition tools, domain constructiontools, modelling languages and planner configurationcomponents into an integrated organisation makingplanning accessible to the non-specialist.The domain construction tool developed at Hudders-field produces PDDL domain descriptions providinga simple connection to the planner configuration ar-chitecture developed at Durham. The knowledge ac-quisition tools developed at Salford assist a user inconfronting the task of domain construction throughthe GIPO tool. At this stage it is possible for a naiveuser to follow the entire process of modelling, andplanning with, a specific problem domain without re-quiring detailed knowledge of any internal represen-tation language ( �!�� or PDDL). The level of ab-straction at which such a user can work within theenvironment will be further raised when the imple-mentation of design patterns, as a guiding principlein the modelling process, is completed.More time and work is necessary to evaluate the en-vironment in terms of how successful it is at makingplanning accessible to the non-expert. There are sev-eral key aspects of the environment that require eval-uation:


� The extent to which the object-oriented approachto modelling ameliorates the modelling task forthe planning non specialist;

� The extent to which design patterns can furtherameliorate this effort;

� The extent to which the proposed knowledge ac-quisition techniques can capture the modelling in-tentions of the user, and how transparent the envi-ronment can make the process of iterating over themodelling task until effective capture is achieved.

The last of these concerns the issue of how to pro-vide useful feedback to the user when the modellingprocess fails to result in a consistent model (due tomissing, or conflicting, information). Without a con-sistent domain model the plan configuration tools cando nothing useful and it is therefore desirable that theuser be able to develop a correct domain model in-crementally. We believe this is one of the most in-teresting technical challenges facing the PLANFORM

project at this stage.

Bibliography

[1] AIPS-98 Planning Competition Committee.PDDL - The Planning Domain Definition Lan-guage. Technical Report CVC TR-98-003/DCSTR-1165, Yale Center for Computational Vi-sion and Control, 1998.

[2] R. S. Aylett and C. Doniat. The PLANFORM-KA Tool. In Proceedings of the 20th UKPlanning and Scheduling Workshop (PLANSIG-2001), Edinburgh, Scotland, 2001.

[3] R. S. Aylett and C. Doniat. KA Tool and domainconstruction for AI planning applications (sub-mitted to ES2002). Technical report, Universityof Salford, 2002.

[4] R. S. Aylett and C. Doniat. Supporting the Do-main Expert in Planning Domain Construction. In Proceedings of the AIPS’02 Workshop onKnowledge Engineering Tools and Techniquesfor AI Planning, 2002.

[5] R. S. Aylett and S. Jones. Planner and Domain:Domain Configuration for a Task Planner. Int.Journal of Expert Systems, 9(2):279–318, 1996.

[6] J. Breuker and B. Wielinga. Models of Exper-tise in Knowledge Acquisition. In Topics in Ex-pert Systems Design: Methodologies and Tools(eds: G. Guida and C. Tasso), 1989.

[7] B. Chandrasekaran and T. R. Johnson. GenericTasks and Task Structures: History, Critiqueand New Directions. In Second Generation Ex-pert Systems (eds: J. M. David, J. P. Krivine andR. Simmons), 1993.

[8] M. Fox and D. Long. The Automatic Inferenceof State Invariants in TIM. Journal of AI Re-search, 9:367–421, 1998.

[9] M. Fox and D. Long. Hybrid STAN: Identify-ing and Managing Combinatorial OptimizationSub-problems in Planning. In Proceedings ofthe International Joint Conference on AI, 2001.

[10] M. Fox and D. Long. STAN 4: A Hybrid Plan-ning Strategy based on Sub-Problem Abstrac-tion. AI Magazine, AAAI Press, 22(4), 2001.

[11] M. Fox and D. Long. Integrating Multiple Sub-solving Strategies into a Hybrid Planning Sys-tem (submitted to the Journal of AI Research).Technical report, Department of Computer Sci-ence, University of Durham, 2002.

[12] D. Liu and T. L. McCluskey. The OCL Lan-guage Manual, Version 1.2. Technical report,Department of Computing and MathematicalSciences, University of Huddersfield , 2000.

[13] D. Long and M. Fox. Extracting Route-Planning: First Steps in Automatic Problem De-composition. In AIPS Workshop on Analyzingand Exploiting Domain Knowledge, 2000.

[14] D. Long and M. Fox. Recognising and exploit-ing generic types in planning domains. In Pro-ceedings of the International Conference on AIPlanning and Scheduling, 2000.



[15] D. Long and M. Fox. Recognizing and Exploit-ing Generic Types in Planning Domains). InProceedings of the Fourth International Con-ference on AI Planning and Scheduling (AIPS-2000), 2000.

[16] D. Long and M. Fox. Multi-Processor Schedul-ing Problems in Planning. In Proceedings of theInternational Conference on AI, 2001.

[17] D. Long, M. Fox, and M. Hamdi. Reformula-tion in Planning. In Proceedings of the Sym-posium on Abstraction, Reformulation and Ap-proximation, 2002.

[18] T. L. McCluskey. Planform: Draft Progress Re-port. Technical Report, School of Computingand Maths, Univ of Huddersfiled, 2001.

[19] T. L. McCluskey and D. E. Kitchin. A Tool-Supported Approach to Engineering HTN Plan-ning Models. In Proceedings of 10th IEEE In-ternational Conference on Tools with ArtificialIntelligence, 1998.

[20] T. L. McCluskey and J. M. Porteous. Engineer-ing and Compiling Planning Domain Models toPromote Validity and Effici ency. Artificial In-telligence, 95:1–65, 1997.

[21] T. L. McCluskey and N. E. Richardson. The in-duction of operator descriptions from examplesand structural domain knowledge. In Proceed-ings of the 20th Workshop of the UK Planningand Scheduling Special Interest Group, pages181–192, 2001.

[22] T. L. McCluskey, N. E. Richardson, and R. M.Simpson. An Interactive Method for Induc-ing Operator Descriptions. In The Sixth Inter-national Conference on Artificial IntelligencePlanning Systems, 2002.

[23] L. C. J. Murray. Reuse of Control Knowledgein Planning Domains. In AIPS Workshop onKnowledge Engineering Tools and Techniquesfor AIP Planning, 2002.

[24] M. Musen. Modern Architectures for IntelligentSystems: Reusable Ontologies and Problem-Solving Methods. In AMIA Annual Sympo-sium), 1998.

[25] A. Newell. The Knowledge Level. ArtificialIntelligence, 1982.

[26] R. M. Simpson, T. L. McCluskey, MariaFox, and Derek Long. Generic Types asDesign Patterns for Planning Domain Spec-ifications. In Proceedings of the AIPS’02Workshop on Knowledge EngineeringTools and Techniques for AI Planning,http://scom.hud.ac.uk/planet/aips02kett/, 2002.

[27] R. M. Simpson, T. L. McCluskey, D. Liu, andD. E. Kitchin. Knowledge Representation inPlanning: A PDDL to �!�� Translation. InProceedings of the 12th International Sympo-sium on Methodologies for Intelligent Systems,2000.

[28] R. M. Simpson, T. L. McCluskey, W. Zhao,R. S. Aylett, and C. Doniat. GIPO: An In-tegrated Graphical Tool to support KnowledgeEngineering in AI Planning. In Proceedingsof the 6th European Conference on Planning,2001.

[29] M. C. Tully. Object-Oriented Domain Mod-elling for Planning. Technical report, De-partment of Computer Science, University ofDurham, 2001.

[30] C. Vogel. Le Genie Cognitif. 1988.

Author Information

T.L. McCluskey University of Huddersfield, UK

Maria Fox University of Durham, UK

Ruth Aylett University of Salford, UKThis work was supported by EPSRC projectsGRM67421/01, GRM70568/01 and GRM68237/01,held at the three participating sites.

�


REPORT

Second PLANET Gap-Bridging Seminar (GBS-2)

Author: T. Grant

The second PLANET Gap-Bridging Seminar (GBS-2) took place in Delft, The Netherlands, on 21November 2002. GBS-2 was held in conjunc-tion with PlanSIG2002, the 21st workshop of theUK AI Planning & Scheduling Special InterestGroup. GBS-2 built on the success of the firstGap-Bridging Seminar, held in Edinburgh, Scot-land, in December 2001 (see PLANET News issue3, pages 30-31 and http://cswww.essex.ac.uk/conferences/planet/GBS-1/).The motivation for Gap-Bridging Seminars is the ob-servation that industry and academia both work onplanning and scheduling, but they do not work to-gether as well as one would hope. They have differ-ent goals. Industry must sell, academia must publish,and there is no time to talk to each other. PLANET’saim is to give representatives of both realms the op-portunity and the time to exchange views.In the Gap-Bridging Seminars, practitioners from in-dustry talk about their work – the techniques, the

problems, customer needs, what can and cannot bedone. This sets research in planning and schedul-ing in a wider context. Most of the speakers stayedfor the whole of PlanSIG2002, so that attendeescould have more opportunities to exchange viewswith them.In GBS-1 most of the speakers were from companiesthat developed and applied AI-based software prod-ucts for planning and scheduling. By contrast, theGBS-2 speakers were chosen to present a wide va-riety of real-world planning and scheduling applica-tions. The speakers and their titles were:

Alessandro Donati European Space OperationsCentre (ESOC), Darmstadt, Germany: “SpaceMission Operations Planning and Scheduling:Past, present and future”.

Henk Hesselink and Ron Seljee Dutch AerospaceLaboratory (NLR), Amsterdam, Netherlands: “AIPlanning, Waiting for results?”

Figure 1: GBS-2 venue in Delft.



Figure 2: Air traffic management problem in Europe.

Frank Oxener Atos Origin Nederland b.v., Nether-lands: “Work Force Planning: A logical next stepafter ERP”

Yossi Rissin and Roman Bartak VisOpt b.v.:“When Theory crashed into Reality”.

In addition, Tim Grant introduced GBS-2 andPLANET.The audience consisted of 42 participants, mostlyfrom the academic community (see Figure 1). UKand the Netherlands each provided a third of thepeople attending. The remaining third was dividedover Spain, Italy, Germany, France, Belgium, and theCzech Republic. Academic attendees were primarilypost-graduate students.Alessandro Donati identified the need for three com-munities – users, innovators and implementers – towork more closely with one another. He explainedthat ESOC is the part of the European Space Agency(ESA) responsible for launching and operating scien-tific spacecraft. A very recent survey at ESOC hadshown that the planning and scheduling processesdiffered radically between space missions. The rea-sons why had yet to be established. A priori, it would

seem to be more efficient to have a standard processand standard tools for all missions. He welcomed theinvolvement of the research community (the innova-tors) and PLANET, and mentioned the opportunitiesfor post-graduate students to work in ESA.

A lively discussion ensued, with researchers call-ing for information to be made available on com-plex, real-world planning and scheduling domainsand problems as case studies. This discussion re-sulted in a lunchtime splinter meeting on the sec-ond day between Alessandro and representatives ofthe innovator and implementer communities on howPLANET and ESOC could co-operate to documentone or more case studies, to mutual benefit.

Henk Hesselink talked about NLR’s experience inproviding software support for planning and schedul-ing in civil and military aircraft operations. TheNLR is a non-profit organisation founded in 1919to provide technical and scientific contributions toaerospace organisations in the private and public sec-tors. Turnover is EUR 70 million annually, split 65%civil and 35% military. The ratio between devel-opment and operations is 40:60, and between aero-nautical and space applications is 85:15. Facilities


available to the NLR include large wind-tunnels, twoaircraft, various simulators, and computing environ-ments (including a supercomputer). NLR co-operateswith Amsterdam Schiphol airport, Dutch industry,and Eurocontrol.Henk focussed on the forthcoming air traffic manage-ment (ATM) problem in Europe (see Figure 2). Thisproblem is not recognised by the controllers – the po-tential end-users – because ATM is divided into sub-problems. Many actors are involved, and each air-port is different, yet has features in common. Cur-rently, ATM controllers do not plan, but work on a“first-heard, first-served” basis. They are conserva-tive, with air safety being uppermost in their mind,followed by the need for fairness in allocating re-sources. They have neither the time nor the interestto talk to researchers. Henk argued that gap-bridginghad to start with the foundations of the bridge: fi-nance and user interest. He advocated the use of pro-totypes to make users aware of their problems andpotential solutions.Frank Oxener explained how the complexity of re-cent law-making in the Netherlands had given rise tocommercial opportunities for planning and schedul-ing applications in work-force management (see Fig-ure 3). Several Dutch companies were offering soft-ware products and associated services.One law required employers to give shift-workers 36hours of rest in any 7-day period or 60 hours rest inany 9-day period. Once every 5 weeks they had tobe allowed at least 32 hours continuous rest. An-other law provided for irregularity payments of 20%of salary to workers working Monday to Friday be-tween 06:00 and 08:00, providing they started before07:00. On Saturdays the payment was 40%, and onSundays 100%. These laws were a challenge to soft-ware systems for recording actual hours worked andplanning rest periods and irregularity payments. Heoutlined his experiences with applying a variety ofcommercially-available software packages.Frank elicited audience participation by asking hislisteners to write down a question on a sheet of paper.He then answered each question. Sample questionsincluded:

� Can the tools easily be adapted to the working-hour laws in other countries?

� What is the key feature in work-planning softwarefor user acceptability?

� What type of techniques are used to solve theproblems you have described?

� How does a different culture really affect a plan-ning process?

� How might a planning solution be overridden bylabour union objections?

� How long does it take to implement a system fromfirst requirement to operational use?

� Are your solutions sufficiently flexible to dealquickly with changes in the law?

� What is coming after workflow management?Why?

Roman Bartak focussed on the human factors issues.Human behaviour is inconsistent and readily affectedby mood, environment and psychological pressure. Itcan only be modelled statistically. Plant personneland planners are motivated by pride, their position inthe organisation, and future job security. Pride makesit difficult for users to admit mistakes, problems andweaknesses. They protect their position by being niceto superiors, gaining professional respect, and tryingto serve many masters at the same time. They protecttheir job by withholding knowledge. Internal politicsand power-plays are key factors in decision-making.

Figure 3: Right people, right place?(acknowledgements to Parallax b.v.)



Figure 4: VisOpt’s visual modelling language.

These factors and issues make academic research ir-relevant. One scheduling expert told him: “I havenever seen a Job Shop Scheduling problem in prac-tice”. From an academic point of view, the ideal fac-tory is one that is totally automated, populated withrobots and Automated Guided Vehicles (AGVs). Al-ternatively, it might be a new factory that has not yetbeen put into operation. There is then no previous“know-how”, rules and procedures, bad habits, andday-to-day reality to confront theory.

Roman contrasted the views of planners and aca-demics, looking at the “Not Invented Here” syn-drome. Summarising the lessons he had learned atVisOpt b.v., he advocated the development of a visualmodelling language as a way of improving communi-cation between the two communities (see Figure 4).

PLANET provided sponsorship for 14 post-graduatestudents from Spain, Germany, Netherlands, UK,Greece, and Italy to attend GBS-2. In return, the stu-dents wrote a short report on what they had learnedfrom GBS-2 and PlanSIG2002. Key quotes included:

� “The gap turned out to be bigger than I thought”.

� “The workshop gave me the chance to get ac-quainted to people whose work I had been readingin papers for quite some time”.

� “Industry does not always need the best possibleplan – it needs one that is good enough”.

� “An attempt should be made from both sides toapproach one another”.

� “Managers and planners need to be convinced tochange to new planning methods”.

� “Industry (NASA, ESOC, NLR, etc) can helpacademia by supplying real planning domains,problems, and case studies”.

� “[There are] four key items: visual modelling,planning as a step-by-step process, scheduling asa process to reason on resource constraints, andthe need for generating good enough plans in areasonable time”.

� “There should be cooperation between [human]planner and computer”.


� “[It is] necessary for researchers to show the ben-efits that can be obtained”.

� “There is also a gap between different academiccommunities: Operations Research and AI”.

� “Employees do not see the global view, but onlythe bit they are working on”.

� “One challenging line of research is the validationof planning domains”.

� “Domains in industry are far less predictableand much more dynamic that those used by re-searchers”.

� “Both communities need more communication”.

� “[GBS-2] encouraged new and interesting discus-sions, thus exposing many unexplored areas of re-search, and providing good ideas for future work”.

� “People from the two sides have to come closerand cooperate since this will be of great profit forboth”.

In summary, both speakers and audience came awayfrom the second PLANET Gap-Bridging Seminarhaving learned more about the academic and indus-trial aspects of planning and scheduling. All atten-dees were most grateful to PLANET and the CEC formaking it possible to exchange such a diversity ofviews, and look forward with eagerness to the thirdGap-Bridging Seminar.

Author Information

Tim Grant Atos-Origin Nederland b.v.,Nieuwegein, The Netherlands, [email protected]

�



ANNOUNCEMENT

ICAPS 2003

�


ANNOUNCEMENT

3rd PLANET International Summer School 2003

The International Summer School on AI Planning isa great opportunity for Ph.D. students and young re-searchers to be exposed to introductory and advancedcourses on various aspects of Artificial IntelligencePlanning, and to spend time and discuss research di-rections with their colleagues and with the teachers,foremost researchers in the field.The first edition of the school was held in Cyprus inSeptember 2000, while the second school was heldin Halkidiki, Greece, in September 2002. The thirdedition of the school will be held on the mountainsof Trentino, in the northern part of Italy, in June2003. The school will be colocated with InternationalConference on Automated Planning and Scheduling(ICAPS’03).The Program Chairs of the school are Daniel Borrajo

Millan (Universidad Carlos III de Madrid, Spain),and Alessandro Cimatti (ITC-irst, Italy). ITC-irstis also responsible for the local organization of theevent, with a team composed by Piergiorgio Bertoli,Mark Carman, Alessandro Cimatti and AlessandroTuccio.Additional, up-to-date information will be availableat the official web site of the school:

http:

//sra.itc.it/planet/summer-school-03/

�

CALL FOR PARTICIPATION

ICAPS 2003 – Doctoral Consortium

ICAPS-2003 invites PhD students to apply for theDoctoral Consortium, which will provide an oppor-tunity for a group of students to discuss and exploretheir research interests and career objectives with es-tablished researchers in Planning and Scheduling.The aims of the Doctoral Consortium are the follow-ing:

� to provide a forum for students to present theircurrent research, and receive feedback from otherstudents and senior researchers;

� to promote contacts among PhD students workingin the same area;

� to support students with information and adviceon academic, research and industrial careers;

� to financially support students by covering theconference registration fee and by partially con-tributing to travel expenses.

Programme

The programme will consist of students’ presenta-tions on their current research interests. A voluntarymentoring programme will be organized to link stu-dents with like-minded researchers.

Submissions

We encourage submissions from Ph.D. students atany level, and from any topic area and methodol-ogy within Planning and Scheduling. On the basis ofthe submissions, the Organizing Committee will se-lect a group of students that will be invited to presenttheir work during the Doctoral Consortium, and alsoto present a poster at the ICAPS-2003 poster ses-sion. Students accepted for participation in the Doc-



toral Consortium will have free conference registra-tion and a fixed allowance for travel/housing. Thestudents’ abstracts will be made available on the weband included as part of the conference proceedings.

Applicants should submit an extended abstract of5 pages maximum by email to one of the Doc-toral Consortium chairs. The submission should bein AAAI style format (http://www.aaai.org/Publications/Author/macros-link.html)and sent either as a PostScript or as a PDF file. Itshould describe original, unpublished work, com-pleted or in progress, that is part of the doctoral workof the student. If an extended version of the paperis also submitted to the technical programme, pleaseindicate it in the submission. Double submission isacceptable, but if the paper is accepted for the tech-nical programme, the student will present the workonly in the technical programme sessions and notduring the Doctoral Consortium.

In addition, the dissertation advisor should send aletter of recommendation by e-mail to one of theDoctoral Consortium chairs. It should include theexpected date for thesis submission, and the moti-vation/expected benefit for the student to attend the

Doctoral Consortium. This letter can be sent in aseither a text or a PostScript or a PDF file.

Important Dates

March 31st: deadline for submitting papers andletters of support

April 18th: notification of acceptance to program

April 25th: camera ready copy of the papers to thechairs

Doctoral Programme Chairs

Jeremy Frank , NASA Ames Research Center,[email protected]

Susanne Biundo , University of Ulm, [email protected]

Additional information is available at the ICAPSWebsite

http://icaps03.itc.it

This event will be sponsored by PLANET and Nasa.

�

ANNOUNCEMENT

ICAPS 2003 – Workshop Program

The ICAPS-03 workshop program (June 9-10, be-fore the main program), has been set. There willbe five workshops, covering a broad range of top-ics ranging from the current and future state of thePlanning Competition, to issues arising as planningand scheduling are applied to ever-more-complex do-mains, to specific application areas.The ICAPS-03 workshops:

Planning and Web Services

Web services are revolutionizing the way industryand government operate. Web services both pro-

vide information (e.g., available flights) and changethe world (e.g., buying a flight ticket). As theWeb evolves into the Semantic Web, the myriad ofavailable services are being described declaratively.Machine-understandable descriptions enable the au-tomatic discovery, use, and composition of web ser-vices.

With increased interest in the web services paradigm,composition of web services has become of primaryimportance. Several languages for describing webservices and their composition are currently being de-fined and seek to become standards. From a plan-ning perspective, the web services can be seen as op-


erators, specific web services compositions as plans,and automatic web service composition as a form ofplanning. This workshop will provide planning re-searchers with a forum for presenting planning resultsrelevant to web services, identify new challenges, andlead the development of the critically important fieldof web services.

Jose Luis Ambite

Workshop on Plan Execution

Much work in the planning community has focussedprimarily upon developing efficient ways of gen-erating plans that are not actually executed. Thiswas reflected in the AIPS 2002 Planning Competi-tion which measured the efficiency and optimality ofplans that were generated, but not executed. As ex-ecution may not result in the intended outcome thesequence of actions that is eventually executed maynot be as valuable as that in the original plan, whichleads to the question as to whether it is necessary togenerate plans that are near optimal. Researchers inthe planning community are increasingly concernedwith executing plans and the design of systems inwhich planning and execution are continually inter-leaved and actively managed.This workshop is intended for researchers who haveinterests in plan execution in a variety of domainssuch as robotics, space applications, informationgathering and other areas.

Alex Coddington

Workshop on PDDL

PDDL, originally developed by Drew McDermottand the 1998 planning competition committee, wasinspired by the need to encourage the empirical com-parison of planning systems and the exchange ofplanning benchmarks within the community. Its de-velopment improved the communication of researchresults and triggered an explosion in performance,expressivity and robustness of planning systems.PDDL has become a de facto standard language fordescribing planning domains, not only for the com-

petition but more widely, as it offers an opportunityto carry out empirical evaluation of planning systemson a growing collection of generally adopted standardbenchmark domains. The emergence of a languagestandard will have an impact on the entire field, in-fluencing what is seen as central and what peripheralin the development of planning systems. The adop-tion of PDDL in this role is itself an issue for debate:perhaps a completely different modelling languageis called for. We believe that it is therefore impor-tant to provide a forum in which the community cangive feedback and present their ideas to the languagedesigners, and in which the language designers candiscuss their ideas for maintaining and extending, oreven replacing the language.

Sylvie Thiebeaux

Planning under Uncertainty and IncompleteInformation

Controlling intelligent agents in complex real-worldenvironments poses requirements that are not ad-dressed in classical AI planning. Often it is not suffi-cient to find a sequence of actions leading to a givengoal, since the initial state may not be known withprecision, and action effects cannot be predicted withcertainty.In the past few years, there has been a growing in-terest in more general planning techniques, able totackle the problems of uncertainty, nondeterminism,and incompleteness of information. Several researchworks have proposed more expressive domain mod-els and description languages (e.g., allowing for ac-tions with multiple transitions, possibly with differ-ent probabilities, and with costs), and more complexmodels of execution (e.g., dealing with informationgathering at run-time). New planning techniques andalgorithms have been developed to operate on suchextended models and to produce plans which achievethe goals despite the uncertainty and incompletenessof information. The goal of this workshop is to bringtogether people working in different areas of plan-ning under uncertain and incomplete information.

Marco Pistore



The Planning Competition: Impact,Organization, Evaluation, Benchmarks

The planning competition series undoubtedly has hada huge impact on the field of AI planning, includingsuch aspects as growing standardization of complexplanning domain description languages, dramaticallyimproved scalability of existing approaches, and agrowing database of commonly used benchmark ex-amples. It is therefore important to provide an op-portunity for discussing topics related to the compe-tition. The workshop aims at doing just this. We wantto collect together panels on topics such as the role ofthe competition in and for the field, organizational as-

pects of the competition, (competition) results evalu-ation, and benchmarking issues. Technical presenta-tions are planned on topics related to the competitionsuch as language alternatives, and methods of empir-ical evaluation.

Joerg Hoffmann and Stefan Edelkamp

Submissions to all workshops are due by 31 March,2003, with notification and submission of camara-ready versions by the end of April.Additional detail about the workshops, as well as fur-ther information about ICAPS-03 in general, is alsoavailable on the conference website:

http://icaps03.itc.it/

�

ANNOUNCEMENT

ICAPS 2003 – Tutorials

Provisional Schedule

Monday, June 9morning afternoon

Timed Automata for Planning and SchedulingOded Maler (Verimag)

Tuesday, June 10morning afternoon

Practical Approaches to Handling Uncertainty inPlanning and SchedulingJ. Christopher Beck (University College Cork)and Thierry Vidal (ENIT)

Resource-Bounded and Time-Critical Reasoning

Lloyd Greenwald (Drexel University) andShlomo Zilberstein (University of Massahusetts)

Model Checking – A Hands-On IntroductionAlessandro Cimatti, Marco Pistore and MarcoRoveri (ITC- IRST)

ICAPS’03 Tutorial chairs: Anthony Barrett, NASA Jet Propulsion LaboratoryJussi Rintanen, Albert-Ludwigs-Universitat Freiburg


Timed Automata for Planning andScheduling

In this tutorial we propose the model of timed au-tomata, originating from the verification of real-timesystems, as a model for posing and solving time-dependent planning and scheduling problems. Webelieve that in the same sense as automata are usedas the major vehicle for verification of systems wherethe model of time is qualitative, timed automata canbe the center of a a unifying mathematical modelingframework for quantitative time, having the follow-ing attractive features:

1. It is sufficiently expressive to describe the essen-tial aspects of time-dependent real-life problemsin a variety of application domains.

2. It provides for models with well-defined and cleardynamic semantics.

3. These models are amenable to computer-aided de-sign methods such as simulation, testing, verifica-tion and automatic synthesis of (optimal) sched-ules and plans.

4. These methods are currently supported by toolsof various levels of maturity, that treat the spe-cific computational problems of time-related rea-soning.

Oded Maler was born in 1957 in Haifa, Israel. Heobtained his B.A. in Computer Science from theTechnion, Haifa in 1979 and his M.Sc. in Manage-ment Science from the University of Tel-Aviv at1984. In 1989 he finished his Ph.D. thesis (FiniteAutomata: Infinite Behavior, Learnability and De-composition ), under the supervision of A. Pnueli inthe department of Applied Mathematics and Com-puter Science, Weizmann Institute, Rehovot. Aftertwo years of post-doc at IRISA, Rennes, he movedto Grenoble at 1992 and obtained a research position(CR1) at the CNRS (French National Center of Sci-entific Research) in 1994. He has been promoted to“research director” (DR2) in 2001. Dr. Maler’s re-search is centered around the theory of automata and

its various extensions, most notably timed automata,hybrid automata and their application to control, em-bedded systems, scheduling and circuit timing analy-sis.

Practical Approaches to HandlingUncertainty in Planning and Scheduling

This tutorial presents techniques for dealing with thefact that the execution of plans and schedules in thereal world cannot assume a static environment: theworld changes non-deterministically during problemsolving and execution. We present techniques fromthe Artificial Intelligence and Operations Researchliterature for handling uncertainty in planning andscheduling with emphasis on practical techniques.Such techniques include reactive, on-line schedulingand planning, and proactive, off-line techniques thatbuild solutions that can cope with uncertain events,as well as intermediate approaches between these ex-tremes.J. Christopher Beck received a PhD in Artificial In-telligence in 1999 from the University of Toronto un-der the supervision of Mark S. Fox. From 1994 to1999 he was the project manager of the IntelligentScheduling Research Group at the Enterprise Inte-gration Laboratory at University of Toronto. Thefocus of his research was measurements of problemstructure as a basis for scheduling heuristics within aconstraint-based scheduling framework. From 1999until 2002 he was a software developer and SeniorScientist on the Scheduler team at ILOG, SA in Gen-tilly, France. As of June, 2002, he moved to the posi-tion of Staff Scientist at the Cork Constraint Com-putation Centre, University College Cork. His re-search interests focus on problem structure, hybridalgorithms, search in constraint-directed scheduling,and in the extension of constraint modeling and solv-ing capabilities to incorporate aspects of real-worldscheduling such as uncertainty, dynamic arrival of ac-tivities, and robustness.Thierry Vidal received a PhD in Artificial Intelli-gence in 1995 from the University of Toulouse underthe supervision of Malik Ghallab. He had worked



in the Robotics and AI team of the LAAS-CNRSin Toulouse, France, working on temporal constraintprocessing in temporal planning (the IxTeT system)and in task scheduling, with a special focus on un-certain durations. In 1996-97 he was a guest re-searcher in Erik Sandewall’s team at the Departmentof Computer Science of the University of Linkoping,Sweden, where he conducted basic research work inthe area of on-line decision making through contin-gent plan recognition and reactive controller synthe-sis. From 1997 he is assistant professor at ENITin Tarbes, France, working in the Automated Pro-duction team of the Production Engineering Labora-tory, with external collaborations with Helene Fargier(Possibilistic Reasoning team, IRIT, Toulouse), PaulMorris (NASA Ames Research Center, California,USA), and Ioannis Tsamardinos and Martha Pol-lack (University of Pittsburgh, USA). His current re-search interests are uncertain constraint reasoning inplanning, scheduling and resource allocation, multi-agent approaches to scheduling, reactivity, condi-tional planning and robust scheduling.

Resource-Bounded and Time-CriticalReasoning

A central problem in artificial intelligence is how todevelop computational models that allow decision-support systems or autonomous agents to react to asituation after performing the right amount of de-liberation. Frequently, the complexity of problemsolving makes it beneficial to use approximate so-lutions rather than try to compute the optimal an-swer. This issue arises in a wide range of applica-tion domains including medical trauma management,Bayesian inference, sequence alignment, graphicsrendering, web page prefetching, autonomous spaceexploration, real-time avionics, and robot navigation.This tutorial explores the theory and practice of build-ing intelligent systems that reason explicitly aboutemploying limited computational resources to gen-erate timely solutions to difficult combinatorial op-timization, planning and scheduling problems. Solu-tion techniques go beyond simple greedy or reactive

algorithms to achieve high-quality solutions whilemeeting both hard and soft real-time deadlines. Wewill explore over fifteen years of progress in this area,covering historical perspectives, state-of-the-art solu-tion techniques, and current and future challenges.Participants should be familiar with introductory arti-ficial intelligence, algorithm design and analysis, andintroductory probability and statistics.Lloyd Greenwald is an Assistant Professor of Com-puter Science and Director of the Intelligent Time-Critical Systems Lab at Drexel University. He re-ceived his Ph.D. in Computer Science from BrownUniversity. His research interests include time-critical planning and scheduling, mobile robotics,machine learning, ad hoc and sensor networks, andmedical decision making.Shlomo Zilberstein is an Associate Professor ofComputer Science and Director of the Resource-Bounded Reasoning Lab (http://anytime.cs.umass.edu) at the University of Mas-sachusetts, Amherst. He received his Ph.D. in Com-puter Science from the University of California,Berkeley. His research interests include approximatereasoning, decision theory, heuristic search, planningand scheduling, and resource-bounded reasoning.

Model Checking – A Hands-OnIntroduction

Model Checking is a formal technique for the verifi-cation of designs of concurrent systems. It is basedon the representation of the system being analyzed asa (finite state) transition systems (e.g. Kripke mod-els), while the requirements are typically expressedin temporal logics. A system satisfies a given prop-erty amounts to checking if the corresponding tem-poral formula is true in the Kripke model. Modelchecking is very effective in pinpointing design errorsthat are extremely hard to detect by means of test-ing, and is therefore being applied in the industrialdevelopment of reactive systems, hardware designs,and communication protocols. Furthermore, modelchecking techniques and tools are gaining interest


in several fields of Artificial Intelligence (e.g. Plan-ning, Multi-agent systems, and Model-based Diagno-sis) and Engineering (e.g. Requirement Verification,Safety Analysis). Of particular interest is SymbolicModel Checking, which makes it is possible to ana-lyze extremely large finite-state systems by means ofsymbolic representation techniques (e.g. Binary De-cision Diagrams, propositional satisfiability).Alessandro Cimatti is the leader of the formal meth-ods group within the Automated Reasoning Sys-tems division (SRA) at ITC-IRST (http://www.irst.itc.it). The activities carried out by thegroup include basic research, the development ofthe NuSMV (http://nusmv.irst.itc.it)model checker, and technology transfer in industrialprojects in the areas of safety critical applications(e.g., railways, avionics, aerospace, industrial plantcontrollers).Alessandro Cimatti has participated in and led severalindustrial projects aimed at the use of formal methodsfor the development and verification of safety criticalsystems and embedded controllers. Some examplesare the validation of Interlocking Systems, the devel-opment of Rail Traffic Management Systems, the de-sign of tools for on-board verification, and the verifi-cation of safety-critical communication protocols.Alessandro Cimatti is the leader of the develop-ment of NuSMV. His main research interests includethe development of advanced model checking tech-niques, and the application of model checking for thesynthesis of reactive controllers and test cases. Hehas also contributed to the research in theorem prov-ing, formal languages for the specification of multi-agent systems, planning and robotics.Marco Pistore is Associate Professor at the Depart-ment of Information and Communication Technolo-

gies of the University of Trento (http://www.dit.unitn.it/) and Research Consultant atITC-IRST (http://www.irst.itc.it). Hisresearch interests are in formal methods and in theapplication of formal methods to planning and to syn-thesis of controllers. Marco Pistore has been the re-sponsible of the development of the NuSMV checker.He is also working to the Formal Tropos project, aim-ing at the development of a formal language and offormal analysis techniques for the verification of re-quirements specifications. Marco Pistore has alsoparticipated to research and industrial projects on theapplication of formal methods to the design and ver-ification of safety-critical systems and of embeddedcontrollers.Marco Roveri received a PhD in Computer Sci-ence in 2002 from the University of Milano in col-laboration with ITC-Irst under the supervision ofA. Cimatti. His PhD thesis “Planning in Non-Deterministic Domains via Symbolic Model Check-ing” was awarded by the Italian Association for Ar-tificial Intelligence (AI*IA) the best price for PhDthesis in artificial intelligence in Italy. He his in thesteering committee of the NuSMV symbolic modelchecker, the first state-of-the-art open source sym-bolic model checker. Since 2001 he is working atITC-Irst in the Automated Reasoning Systems divi-sion. From 1997 until 2002 he was collaboratingwith ITC-Irst on topics related to Artificial Intelli-gence Planning and Formal Verification. His researchinterest are: integration of formal verification tech-niques along the whole software development pro-cess, planning in non-deterministic domains underdifferent assumptions on run-time observability usingsymbolic model checking techniques and symbolicmodel checking.

�



ANNOUNCEMENT

Invitation to Participate in TAC’03 - A Supply Chain TradingCompetition

Authors: R. Arunachalam, N. Sadeh, E. Aurell, J. Eriksson, N. Finne, and S. Janson

Supply chain management is concerned with plan-ning and coordinating the activities of organizationsacross the supply chain, from raw material procure-ment to finished goods delivery. In today’s globaleconomy, effective supply chain management is vitalto the competitiveness of manufacturing enterprisesas it directly impacts their ability to meet changingmarket demands in a timely and cost effective man-ner. With annual worldwide supply chain transac-tions in the trillions of dollars, the potential impactof performance improvements is tremendous. Whiletoday’s supply chains are essentially static, relyingon long-term relationships among key trading part-ners, more flexible and dynamic practices offer theprospect of better matches between suppliers and cus-tomers as market conditions change. Adoption ofsuch practices has however proven elusive, due to thecomplexity of many supply chain relationships andthe difficulty in effectively supporting more dynamictrading practices. TAC-03 was designed to capturemany of the challenges involved in supporting dy-namic supply chain practices, while keeping the rulesof the game simple enough to entice a large numberof competitors to submit entries. The game has beendesigned jointly by a team of researchers from thee-Supply Chain Management Lab at Carnegie Mel-lon University and the Swedish Institute of ComputerScience (SICS).

Specifically, TAC-03 features rounds where eight PCassembly agents compete for customer orders and forprocurement of a variety of components. Customersissue requests for quotes and select from quotes sub-mitted by the PC assemblers, based on delivery datesand prices. The assembly agents are limited by thecapacity of their assembly lines and have to procurecomponents from a set of possible suppliers. Thegame distinguishes between four types of compo-nents: CPUs, Motherboards, Memory Units and Disk

drives. It features a variety of components of eachtype (e.g. different CPUs, different motherboards,etc.). Customer demand comes in the form of re-quests for quotes for different types of PCs, each re-quiring a different combination of components.The PC assembly agents compete over a relativelylong period of time during which customer demandand availability of supplies varies according to prede-fined stochastic distributions. The aim of each com-petitor agent (PC assembly agent) is to maximize itsprofit, by (1) competing with other agents for valu-able customer orders and profitable supplier commit-ments, and (2) managing the assembly of products tomeet its existing customer delivery commitments.

The game is representative of a broad range of supplychain situations. It is challenging in that it requiresagents to concurrently compete in multiple markets(markets for different components on the supply sideand markets for different products on the customerside) with interdependencies and incomplete infor-mation. It allows agents to strategize (e.g. specializ-ing in particular types of products, stocking up com-ponents that are in low supply). To succeed, agentswill have to demonstrate their ability to react to vari-ations in customer demand and availability of sup-


plies, as well as adapt to the strategies adopted byother competing agents.We would like to invite you to consider submittingan entry to the competition. This is a unique oppor-tunity to develop and evaluate supply chain tradingtechnology in a competitive environment. Entrantswill also be invited to submit articles in an upcomingbook and will benefit from the publicity associatedwith the event in the form of press coverage and thepublication of an AI magazine article discussing thecompetition.A detailed game description, including the rulesof TAC03, will be published on the TAC website(http://www.sics.se/tac/) by late November2003. The TAC 03 game server and some simple PC

assembly agents will be available for practice gameson the TAC website by February 1, 2003. This willenable prospective agent designers to test and fine-tune their designs by playing practice games. Thecompetition itself will be played in a format sim-ilar to earlier TAC games with eight agents com-peting in each round. Qualification rounds will beheld in May 2003 with the finals slated to take placeat IJCAI-03 in Acapulco in August. Stay tuned on(http://www.sics.se/tac/) for more informa-tion.Raghu Arunachalam and Norman Sadeh(CMU),Erik Aurell, Joakim Eriksson, Niclas Finne,and Sverker Janson (SICS)

�

JOB OPENING

Postdoctoral and Doctoral Positions at Australian National ICTCenter

The Australian National ICT Center (NICTA) is anew research institute jointly set up by the Univer-sity of New South Wales (UNSW) and the AustralianNational University (ANU) with respective nodes inSydney and Canberra. It is funded by the Aus-tralian Federal and State Governments, in partnershipwith the two universities and industry. NICTA willhost top-ranked international researchers and gradu-ate programs, and will cover major areas in comput-ing, systems and telecommunications.The NICTA home page is

http://www.nicta.com.au

In it there are links to the existing programs (morewill be added in the future), and vacant positions.The activities of the Knowledge Representationand Reasoning (KRR) program (http://www.nicta.com.au/kr.html) are typified by the

content of papers appearing in, say, the KRR sectionof the IJCAI proceedings, with an additional focus onplanning and constraints. The program has the pre-existing multi-university Knowledge Systems Group(KSG) as its initial core, but now seeks to expand byrecruiting research personnel and graduate doctoralstudents. At the moment, post-doctoral fellows anddoctoral students are sought: for information on howto apply and conditions please follow the PositionsVacant link in the NICTA home page.

KRR welcomes preliminary inquiries about otherlevels of personnel.

Norman Foo, Maurice Pagnucco (UNSW), SylvieThiebaux (ANU)

Dr. Sylvie Thiebaux Research Fellow, RSISE,The Australian National University, Canberra ACT0200, Australiahttp://csl.anu.edu.au/˜[email protected]: +61 (2) 6125 8678, Fax: +61 (2) 6125 8651



INFORMATION

Member List for PLANET

Currently, PLANET has 58 nodes from 15 Europeancountries. Sites and contact persons are:

Austria- XIMES GmbH, Johannes Gartner,[email protected]

Belgium- Robonetics NV, Filip Verhaeghe,[email protected]

- Space Applications Services (SAS), RichardAked,[email protected]

Cyprus- University of Cyprus, Yannis Dimopoulos,[email protected]

Czech Republic- Charles University, Praha, Roman Bartak,[email protected]

France- COSYTEC S.A., Abderrahmane Aggoun,[email protected]

- ILOG S.A., Philippe Laborie,[email protected]

- Laboratoire d’ Analyse et d’ Architecture des Sys-temes (LAAS-CNRS), TCU Robot Planning, Ma-lik Ghallab,[email protected]

- Laboratoire d’ Informatique Marseille (LIM-CNRS), Camilla Schwind,[email protected]

- MASA Group, Emmanuel Chiva,[email protected]

- ONERA Systems Control and Flight Dynam-ics Department, TCU On-line Planning andScheduling, Gerard Verfaillie,[email protected]

- THOMSON-CSF, Simon De Givry,[email protected]

Germany- University of Ulm, Coordinating Node, Susanne

Biundo,[email protected]

- Aachen University of Technology, Gerhard Lake-meyer,[email protected]

- University of Bonn, Armin Cremers,[email protected]

- Bremer Institut fur Betriebstechnik und ange-wandte Arbeitswissenschaft (BIBA), Frithjof We-ber,[email protected]

- Darmstadt University of Technology, UlrichScholz,[email protected]

- German Research Center for Artificial Intelli-gence (DFKI), Markus Meyer,[email protected]

- University of Freiburg, Bernhard Nebel,[email protected]

- Fraunhofer - Autonomous intelligent Systems(AiS), Joachim Hertzberg,[email protected]


- Technical University of Munich, Michael Beetz,[email protected].

de

- Siemens AG, Wendelin Feiten,[email protected]

Greece- Aristotle University of Thessaloniki, Ioannis Re-

fanidis,[email protected]

- Foundation for Research and Technology - Hellas(ICS-FORTH), Dimitrios Plexousakis,[email protected]

- National Centre for Scientific Research”Demokritos”, Constantine Spyropoulos,[email protected]

- Technical University of Athens (ICCS), SpyrosTzafestas,[email protected]

- Technical University of Crete, ManolisKoubarakis,[email protected]

- University of Ioannina, Chrysostomos [email protected]

Hungary- Computer and Automation Research Institute

Hungarian Academy of Sciences (MTA SZTAKI),Laszlo Monostori,[email protected]

Italy- DEIS - University of Bologna, Paola Mello,[email protected]

- University of Brescia, Alfonso Gerevini,[email protected]

- DIST - University of Genoa, Enrico Giunchiglia,[email protected]

- Consiglio Nazionale delle Ricerche - Istituto diPsicologia (IP-CNR), TCU Aerospace Applica-tions, Amedeo Cesta,[email protected]

- University of Perugia, TCU Planning & Schedul-ing for the Web, Alfredo Milani,[email protected]

- Istituto per la Ricerca Scientifica e Tecnologia(IRST), Paolo Traverso,[email protected]

- University of Parma, Agostino Poggi,[email protected]

The Netherlands- Delft University of Technology, Cees Witteveen,[email protected]

- NLR – National Aerospace Laboratory, Henk Hes-selink,[email protected]

Portugal- Instituto Superior de Engenharia do Porto

ISEP/IPP, Joao Rocha,[email protected]

SLOVENIA- University of Maribor, Peter Kokol,[email protected]

Spain- iSOCO, Intelligent Software for the Networked

Economy, Antonio Reyes Moro,[email protected]

- Technical University of Catalonia, Lluıs Vila,[email protected]

- University of Granada, Luis Castillo,[email protected]

- University Carlos III of Madrid, TCU WorkflowManagement, Daniel Borrajo,[email protected]

- Universitat Politecnica de Catalunya, Institut deRobotica i Informatica Industrial, Tom Creemers,[email protected]

- Universidad Politecnica de Valencia, Eva Onain-dia,[email protected]



- Universitat Rovira i Virgili, Tarragona, MiguelAngel Garcia,[email protected]

Sweden- Linkoping University, Patrick Doherty,[email protected]

- Orebro University, Alessandro Saffiotti,[email protected]

United Kingdom- British Telecommunications, David Lesaint,[email protected]

- University of Durham, Julie Porteous,[email protected]

- University of Essex, Sam Steel,[email protected]

- University of Edinburgh, John Levine,[email protected]

- University of Huddersfield, TCU Knowledge En-gineering, Lee McCluskey,[email protected]

- University of Manchester, Nikolay Mehandjiev,[email protected]

- The Open University Walton Hall, MassimilianoGaragnani,[email protected]

- Salford University, TCU Intelligent Manufactur-ing, Ruth Aylett,[email protected]

- Troy Associates Ltd., Vince Long,[email protected]

Associated Members- Norman Sadeh, [email protected], Carnegie

Mellon University

- Peter Jarvis, [email protected], SRI

- Brian Drabble, [email protected],University of Oregon

- Sylvie Thiebaux, Sylvie.Thiebaux@anu.

edu.au, The Australian National University�

The network is open to new nodes at any time.

newsletter of the european network of excellence in ai planning...

Documents