Newsletter of the European Network of Excellence ... - Service - DFKI
Newsletter of the European Network of Excellence ... - Service - DFKI
Newsletter of the European Network of Excellence ... - Service - DFKI
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
<strong>Newsletter</strong> <strong>of</strong> <strong>the</strong><br />
<strong>European</strong> <strong>Network</strong><br />
http://www.planet-noe.org<br />
<strong>of</strong> <strong>Excellence</strong><br />
in<br />
AI Planning<br />
Issue No.5<br />
ISSN 1610-0204
PLANET News<br />
Issue No. 5, December 2002<br />
Copyright (C) 2002<br />
PLANET, <strong>the</strong> <strong>European</strong> <strong>Network</strong> <strong>of</strong><br />
<strong>Excellence</strong> in AI Planning<br />
Printed in Ulm, Germany<br />
ISSN 1610-0204
The PLANET <strong>Newsletter</strong> 3<br />
Welcome to PLANET NEWS!<br />
Welcome to Issue No 5 !<br />
PLANET’s Industrial Information Days provide a forum<br />
where successful industrial applications <strong>of</strong> Planning<br />
and Scheduling technology are presented and<br />
promising future exploitation is discussed. Representatives<br />
from industry and academia use <strong>the</strong>se events<br />
to promote <strong>the</strong> transfer <strong>of</strong> <strong>the</strong> technology and to push<br />
mutual exchange and co-operation for <strong>the</strong> benefit <strong>of</strong><br />
both.<br />
The recent information day took place in Rome in<br />
November. You’ll find a broad selection <strong>of</strong> contributions<br />
from industrial speakers in <strong>the</strong> first section <strong>of</strong><br />
this issue.<br />
Ano<strong>the</strong>r major event was <strong>the</strong> Second PLANET International<br />
Summer School on AI Planning held<br />
in Halkidiki, Greece in September. Students <strong>of</strong><br />
<strong>the</strong> school were encouraged to present <strong>the</strong>ir Ph.D.<br />
projects in a separate poster session. This session was<br />
very successful and showed a variety <strong>of</strong> interesting<br />
aspects and new approaches. A report on <strong>the</strong> school<br />
and several extended abstracts <strong>of</strong> <strong>the</strong> poster presentations<br />
are included.<br />
The PLANFORM project aimed at <strong>the</strong> development<br />
<strong>of</strong> an Open Environment for Building Planners. It<br />
was carried out jointly at <strong>the</strong> universities <strong>of</strong> Hudder-<br />
Issue No.5<br />
EDITORIAL<br />
sfield, Salford, and Durham in <strong>the</strong> United Kingdom.<br />
You’ll find a final report on this project at page 38.<br />
The second PLANET Gap-bridging Seminar was<br />
held in co-location with <strong>the</strong> UK PLANSIG meeting<br />
in November in Delft. A report on this seminar is<br />
provided by Tim Grant.<br />
Announcements, job <strong>of</strong>fers and information on forthcoming<br />
events, in particular those to be held in conjunction<br />
with ICAPS in June 2003 in Trento, and an<br />
invitation to participate in <strong>the</strong> Supply Chain Trading<br />
Competition 2003 can be found in <strong>the</strong> final section.<br />
Wishing you a successful and Happy New Year 2003,<br />
Editors:<br />
Susanne Biundo<br />
Bernd Schattenberg<br />
Susanne Biundo <strong>Network</strong> Coordinator, Dept. <strong>of</strong><br />
Artificial Intelligence, University <strong>of</strong> Ulm, Germany,<br />
biundo@informatik.uni-ulm.de<br />
Bernd Schattenberg <strong>Network</strong> Administrator,<br />
Dept. <strong>of</strong> Artificial Intelligence, University <strong>of</strong> Ulm,<br />
Germany, schatten@informatik.uni-ulm.<br />
de
4 The PLANET <strong>Newsletter</strong><br />
Table <strong>of</strong> Contents<br />
Editorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3<br />
Industrial Information Day Rome<br />
S. de Givry, L. Jeannin, F. Josset, J. Mattioli,<br />
N. Museux, and P. Savéant<br />
The THALES constraint programming framework<br />
for hard and s<strong>of</strong>t real-time applications . . . . . . . . . . . 5<br />
M. Sanseverino<br />
COMPETE a Common Platform for Extended<br />
Project Management . . . . . . . . . . . . . . . . . . . . . . . . . . . 8<br />
J.E. Spragg<br />
In Defence <strong>of</strong> Reactive Scheduling . . . . . . . . . . . . . 14<br />
B. Drabble<br />
Advanced Scheduling and Optimisation: Cutting <strong>the</strong><br />
Costs <strong>of</strong> Manufacturing. . . . . . . . . . . . . . . . . . . . . . . .17<br />
2nd PLANET Summer School<br />
L. Compagna, G. Cortellessa, A. Farinelli, and N. Policella<br />
2nd International PLANET Summer School . . . . . 22<br />
A. Farinelli, G. Grisetti, L. Iocchi, D. Nardi, and<br />
R. Rosati<br />
Generation and Execution <strong>of</strong> Partially Correct Plans<br />
in Dynamic Environments . . . . . . . . . . . . . . . . . . . . . 25<br />
F. McNeill, A. Bundy, M. Schorlemmer<br />
Dynamic Ontology Refinement. . . . . . . . . . . . . . . . .27<br />
G. Cortellessa, N. Policella, A. Cesta, and A. Oddi<br />
MEXAR: Integrated AI Technologies to Support<br />
MARS EXPRESS Mission Planning . . . . . . . . . . . . . 29<br />
M. Baioletti, A. Milani, and V. Poggioni<br />
RDPPlan: an Extension <strong>of</strong> DPPlan for Planning with<br />
Interval Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31<br />
N.E. Richardson<br />
Extending Operator Induction to Provide Full<br />
Method Sets for Hierarchical Planning Domains . 33<br />
http://www.planet-noe.org<br />
O. Sapena and E. Onaindía<br />
The SimPlanner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34<br />
A. Polleres<br />
Answer Set Planning with DLV<br />
. . . . . . . . . . . . . . .36<br />
Project Report<br />
T.L. McCluskey, M. Fox, and R. Aylett<br />
Planform: An Open Environment for Building Planners<br />
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38<br />
Gap-bridging Seminar<br />
T. Grant<br />
Second PLANET Gap-Bridging Seminar . . . . . . . . 46<br />
Announcements and <strong>Network</strong> News<br />
ICAPS 2003. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .51<br />
3rd PLANET International Summer School . . . . . . 52<br />
ICAPS 2003 – Doctoral Consortium . . . . . . . . . . . . 52<br />
ICAPS 2003 – Workshop Program. . . . . . . . . . . . . .53<br />
ICAPS 2003 – Tutorials . . . . . . . . . . . . . . . . . . . . . . . 55<br />
R. Arunachalam<br />
Invitation to Participate in TAC’03 - A Supply Chain<br />
Trading Competition . . . . . . . . . . . . . . . . . . . . . . . . . . 59<br />
Postdoctoral and Doctoral Positions at Australian National<br />
ICT Center . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60<br />
Information<br />
The Members <strong>of</strong> PLANET . . . . . . . . . . . . . . . . . . . . 61
The PLANET <strong>Newsletter</strong> 5<br />
ARTICLE<br />
The THALES Constraint Programming Framework for Hard and<br />
S<strong>of</strong>t Real-Time Applications<br />
Authors: S. de Givry, L. Jeannin, F. Josset, J. Mattioli, N. Museux, and P. Savéant<br />
This position paper presents <strong>the</strong> constraint technology<br />
that has been developed at THALES since<br />
1997 for introducing Constraint Programming (CP)<br />
in THALES operational systems 1 2 . These systems<br />
involve combinatorial optimization problems such as<br />
planning and scheduling problems that can be expressed<br />
with finite-domain variables and constraints.<br />
Typical examples <strong>of</strong> THALES systems concern supervision,<br />
for weapon allocation, radar configuration,<br />
weapon deployment and aircraft sequencing. All<br />
<strong>the</strong>se systems are subject to specific requirements<br />
coming from <strong>the</strong> operational constraints <strong>of</strong> embedded<br />
real-time systems and from <strong>the</strong> strategic context<br />
<strong>of</strong> Defense applications:<br />
1. The system involves several functions/tasks such<br />
as situation assessment, resource management, visualization,<br />
etc.; each task is periodical and <strong>the</strong><br />
period can be much shorter than a second;<br />
2. There is a memory space limit (a few megabytes);<br />
3. The system has to be supported for a long time,<br />
typically over 20 years for Defense applications,<br />
including several retr<strong>of</strong>itting (functional and platform<br />
evolutions);<br />
4. The system can be reused and modified for building<br />
a specific system for a new client (product<br />
line);<br />
5. The development <strong>of</strong> <strong>the</strong> system must be made and<br />
mastered in house for reasons <strong>of</strong> confidentiality<br />
and market protection.<br />
The CP paradigm partially meets <strong>the</strong>se requirements.<br />
A constraint model has modularity properties, i.e.<br />
adding/removing a constraint is easy, which enables<br />
an incremental development process, reducing <strong>the</strong><br />
1 This work was partially funded by <strong>the</strong> EOLE project[7].<br />
2 A former version <strong>of</strong> this paper appeared in [5].<br />
development time and effort. CP solvers provide efficient<br />
algorithms through <strong>the</strong> use <strong>of</strong> global constraints.<br />
The declarative nature <strong>of</strong> CP enables <strong>the</strong> programmer<br />
to focus on <strong>the</strong> application requirements ra<strong>the</strong>r than<br />
on debugging low-level programming errors. Validated<br />
CP models can be reused in a product line approach.<br />
Unfortunately, <strong>of</strong>f-<strong>the</strong>-shelf CP solvers do not provide<br />
any guarantee on time and space usage. The<br />
classical backtracking search algorithm used in CP<br />
does not take into account any time contract. Recently<br />
an effort was made to provide better search algorithms<br />
in CP solvers, for instance in [1, 11, 14], but<br />
without any explicit time contract. Our aim is to extend<br />
CP solver with new search features that would<br />
keep <strong>the</strong> same nice s<strong>of</strong>tware engineering properties<br />
as for modeling. This led to develop a high-level<br />
language for designing search algorithms. This approach<br />
allows proposing a set <strong>of</strong> search primitives on<br />
top <strong>of</strong> <strong>the</strong> real-time finite-domain constraint solver<br />
Eclair c<br />
[13]. The resulting search algorithms are<br />
based on partial search methods and take into account<br />
<strong>the</strong> time contract explicitly. Such algorithms can take<br />
advantage better <strong>of</strong> platform evolutions.<br />
Eclair <strong>of</strong>fers time and space guarantees. Deadlines<br />
are guaranteed by <strong>the</strong> operating system alarm and<br />
Eclair is able to restore a coherent state after an interruption<br />
in order to deliver a valid solution, or just<br />
a partial solution (when not all variables are instantiated).<br />
The memory allocation for <strong>the</strong> constraints<br />
is static: a global constraint model is built once and<br />
only parts <strong>of</strong> <strong>the</strong> model are made active and used at<br />
a given cyclical call. The memory consumed during<br />
<strong>the</strong> search is limited by using only restricted depthfirst<br />
search or restricted best-first search.
6 The PLANET <strong>Newsletter</strong><br />
Partial search methods are anytime algorithms [17]<br />
based on tree search methods having better quality<br />
pr<strong>of</strong>iles than <strong>the</strong> classical backtracking search algorithm.<br />
The main idea is to apply some arbitrary limits<br />
on <strong>the</strong> nodes visited in <strong>the</strong> tree search 3 , depending on<br />
<strong>the</strong> behavior <strong>of</strong> <strong>the</strong> heuristics and on <strong>the</strong> remaining<br />
computation time. We distinguish four approaches:<br />
<strong>the</strong> iterative weakening methods (e.g. [8]), <strong>the</strong> realtime<br />
search methods (e.g. [10]), <strong>the</strong> iterative sampling<br />
methods (e.g. [6]) and <strong>the</strong> interleaving methods<br />
(e.g. [12]). These methods use one or several search<br />
schemes 4 . The practical complexity <strong>of</strong> <strong>the</strong> search<br />
can be increasing, self-adjusting, or stable. In [3], we<br />
propose <strong>the</strong> notion <strong>of</strong> parameterized search applied<br />
to one search scheme. The parameters <strong>of</strong> <strong>the</strong> search<br />
limits are given explicitly. We can tune <strong>the</strong> degree<br />
<strong>of</strong> incompleteness <strong>of</strong> <strong>the</strong> search by varying <strong>the</strong> values<br />
<strong>of</strong> <strong>the</strong> parameters. A tuning policy indicates <strong>the</strong> relevant<br />
values <strong>of</strong> <strong>the</strong> parameters for different time contracts.<br />
In [4], we integrate <strong>the</strong> parameterized search<br />
approach into a hybridization scheme to express partial<br />
search based on several search schemes. The hybridization<br />
scheme is a sequence or an interleaving<br />
<strong>of</strong> parameterized searches. The searches can cooperate<br />
by exchanging solutions. A time-sharing policy<br />
specifies how to distribute <strong>the</strong> time contract to <strong>the</strong><br />
searches.<br />
Our constraint optimization framework is called<br />
ToOLS c (Templates Of On-Line Search). A search<br />
algorithm is expressed in ToOLS as <strong>the</strong> conjunction<br />
<strong>of</strong> four distinct components:<br />
A set <strong>of</strong> heuristics to rank every choice;<br />
A set <strong>of</strong> primitives to express a search scheme<br />
independent <strong>of</strong> any time limit; it is composed<br />
by predefined choice points and combinations <strong>of</strong><br />
choice points as in <strong>the</strong> OPL language [9];<br />
A set <strong>of</strong> primitives to express <strong>the</strong> search limits that<br />
depend on <strong>the</strong> current node, <strong>the</strong> current path or<br />
<strong>the</strong> current sub-tree; <strong>the</strong> resulting parameterized<br />
search algorithm controls <strong>the</strong> size <strong>of</strong> <strong>the</strong> explored<br />
search tree defined by one search scheme;<br />
A temporal strategy defined by a hybridization<br />
scheme, i.e. a cooperation <strong>of</strong> several parameterized<br />
searches, dealing with time allocation and<br />
selecting <strong>the</strong> tuning strategy <strong>of</strong> <strong>the</strong> parameters<br />
(static tuning, iterative tuning or adaptive tuning).<br />
A template <strong>of</strong> search defines an abstract component<br />
<strong>of</strong> a search algorithm that can be reused to speed up<br />
<strong>the</strong> development process <strong>of</strong> customized partial search<br />
algorithms. This framework makes it easier to try<br />
new combinations <strong>of</strong> search limits and new temporal<br />
strategies.<br />
Experiments on <strong>the</strong> weapon allocation problem show<br />
that partial search algorithms significantly improve<br />
<strong>the</strong> solution quality compared to a traditional approach<br />
[3] and also demonstrates <strong>the</strong> gain in development<br />
time <strong>of</strong> new customized search algorithms.<br />
The code is clearer and more concise when using <strong>the</strong><br />
search primitives. As <strong>the</strong> main result, our CP framework<br />
has been integrated in an operational on-board<br />
hard real-time system <strong>of</strong> THALES.<br />
The hybridization scheme is a way to define specific<br />
local search methods, such as large neighborhood<br />
search based on a sequence <strong>of</strong> partial searches<br />
in different neighborhoods [15, 16]. Pure local search<br />
methods could also be introduced in our framework<br />
as a black-box used by <strong>the</strong> hybridization scheme. The<br />
temporal control could be enhanced by an on-line<br />
learning mechanism, using <strong>the</strong> fact that similar problems<br />
are repeatedly solved in a real-time system. [2]<br />
gave <strong>the</strong> base for this mechanism.<br />
Bibliography<br />
[1] Beldiceanu, N., Bourreau, E., Simonis, H., and<br />
Rivreau, D. Introduction de métaheuristiques<br />
dans CHIP. In Proc. <strong>of</strong> MIC-98., 1998.<br />
[2] Crawford, Lara S., Fromherz, Markus P.J.,<br />
Guettier, Christophe, Shang, Yi. A Framework<br />
3<br />
This description <strong>of</strong> partial search is compatible with <strong>the</strong> depth-first search principle. In [14], partial search methods are based on<br />
<strong>the</strong> order <strong>of</strong> node exploration, which is memory consuming.<br />
4<br />
A search scheme is a procedure that describes a search tree. For example, a combination <strong>of</strong> choice points.<br />
http://www.planet-noe.org
The PLANET <strong>Newsletter</strong> 7<br />
for On-line Adaptive Control <strong>of</strong> Problem Solving.<br />
In Proc. <strong>of</strong> CP-2001 workshop on On-<br />
Line combinatorial problem solving and Constraint<br />
Programming, Paphos, Cyprus, December<br />
2001.<br />
[3] de Givry, Simon, Savéant, Pierre, Jourdan,<br />
Jean. Optimization combinatoire en temps<br />
limité: Depth first branch and bound adaptatif.<br />
In Proc. <strong>of</strong> JFPLC-99, pages 161-178, Lyon,<br />
France, 1999 (in french).<br />
[4] de Givry, S., Hamadi, Y., Mattioli, J., Lemaître,<br />
M., Verfaillie, G., Aggoun, A., Gouachi, I.,<br />
Benoist, T., Bourreau, E., Labur<strong>the</strong>, F., David,<br />
P., Loudni, S., Bourgault, S. Towards an on-line<br />
optimization framework. In Proc. <strong>of</strong> CP-2001<br />
workshop on On-Line combinatorial problem<br />
solving and Constraint Programming, Paphos,<br />
Cyprus, December 2001. http://www.lcr.<br />
thomson-csf.com/projects/www_eole/<br />
workshop/olcp01-eole.ps<br />
[5] de Givry, S., Gérard, P., Jeannin, L., Mattioli, J.,<br />
Museux, N., Savéant, P. A constraint optimization<br />
framework for real-time applications. In<br />
Proc. <strong>of</strong> AIPS 2002 workshop on On-line Planning<br />
and Scheduling, Toulouse, France, April<br />
2002.<br />
[6] Gomes, C., Selman, B., Kautz, H. Boosting<br />
Combinatorial Search Through Randomization.<br />
In Proc. <strong>of</strong> <strong>the</strong> 15th National Conference on Artificial<br />
Intelligence (AAAI-98), pages 431–437,<br />
Madison, WI, USA, 1998.<br />
[7] RNRT EOLE project. http://www.lcr.<br />
thomson-csf.com/projects/www_eole<br />
[8] Harvey, William D., Ginsberg, Mat<strong>the</strong>w L. Limited<br />
discrepancy search. In Proc. <strong>of</strong> IJCAI-95,<br />
pages 607-613, Montréal, Canada, 1995.<br />
[9] Van Hentenryck, P. OPL: The Optimization<br />
Programming Language. The MIT<br />
Press,Cambridge,Mass., 1999.<br />
[10] Korf, Richard E. Real-time heuristic search. Artificial<br />
Intelligence, 42:189-211, 1990.<br />
[11] Labur<strong>the</strong>, François. SaLSA: a language for<br />
search algorithms. In Proc. <strong>of</strong> CP-98, pages<br />
310-324, Pisa, Italy, October 26-30 1998.<br />
[12] Meseguer, P. Interleaved depth-first search. In<br />
Proc. <strong>of</strong> IJCAI-97, Nagoya, Japan, pp. 1382-<br />
1387, 1997.<br />
[13] PLATON Team. Eclair reference manual,<br />
Version 6.0. Technical Report Platon-01.16,<br />
THALES Research and Technology, Orsay,<br />
France, 2001.<br />
[14] Perron, L. Search Procedures and Parallelism<br />
in Constraint Programming. In Proc. <strong>of</strong> CP-99,<br />
pages 346-360, Alexandria, Virginia, 1999.<br />
[15] Pesant, G. and Gendreau, M. A Constraint Programming<br />
Framework for Local Search Methods.<br />
Journal <strong>of</strong> Heuristics, 1999.<br />
[16] Shaw, P. Using constraint programming and<br />
local search methods to solve vehicle routing<br />
problems. In Proc. <strong>of</strong> CP-98, pp. 417-431, Pisa,<br />
Italy, 1998.<br />
[17] Zilberstein, Shlomo. Using Anytime Algorithms<br />
in Intelligent Systems. AI Magazine, 17(3):73-<br />
83, 1996.<br />
Author Information<br />
Simon de Givry INRA / Biometrics and Artificial<br />
Intelligence, Chemin de Borde Rouge, BP27,<br />
31326 Castanet-Tolosan cedex France, degivry@<br />
toulouse.inra.fr<br />
Laurent Jeannin, François-Xavier Josset,<br />
Juliette Mattioli, Nicolas Museux, Pierre<br />
Savéant THALES Research & Technology /<br />
PLATON Center, Domaine de Corbeville 91404<br />
Orsay cedex France, laurent.jeannin@<br />
thalesgroup.com
8 The PLANET <strong>Newsletter</strong><br />
ARTICLE<br />
COMPETE a Common Platform for Extended Project Management<br />
Abstract<br />
COMPETE (Common Platform for <strong>the</strong> ExTended<br />
Enterprise) is an ESPRIT project which started in<br />
1999 and ended at <strong>the</strong> end <strong>of</strong> 2001. It is <strong>the</strong> result<br />
<strong>of</strong> <strong>the</strong> cooperation among different <strong>European</strong><br />
partners: Centro Ricerche Fiat, TXT e-solutions,<br />
Cap Gemini & Ernst & Young, BAE Systems and<br />
Magneti Marelli, University Federico II <strong>of</strong> Naples.<br />
Its main objective is to integrate methods and tools<br />
to support process/project management in an extended<br />
enterprise. In this environment process<br />
know-how (how to do things) and available competences<br />
play a central role to shorten product development<br />
time and to support rapid decisions and<br />
flexibility in process actuation.<br />
COMPETE has developed a s<strong>of</strong>tware platform<br />
which integrates process modeling methods, competences<br />
management, project management and<br />
workflow through a common data model shared by<br />
all company departments.<br />
This paper , starting from <strong>the</strong> description <strong>of</strong> current<br />
companies needs, describes COMPETE approach<br />
and functionalities and gives an idea <strong>of</strong> its general<br />
architecture.<br />
Keywords: Project Management, Knowledge<br />
Management, Competencies, Workflow, Business<br />
Processes, Modeling, Simulation.<br />
Introduction<br />
The competitive arena in <strong>the</strong> Product Development<br />
scenario is marked today by rapidly evolving technologies,<br />
dynamic, sophisticated and global markets,<br />
requiring high and customized performances at low<br />
cost. This trend has put a tremendous pressure on <strong>the</strong><br />
design process which must supply a stream <strong>of</strong> products<br />
with high value/cost ratios at a rate fast enough to<br />
withstand technology obsolescence and demand vagaries.<br />
The main goal <strong>of</strong> COMPETE is to provide IT based<br />
http://www.planet-noe.org<br />
Author: M. Sanseverino<br />
methodologies and tools to help companies structured<br />
as Extended Enterprises to cope with <strong>the</strong> challenges<br />
<strong>of</strong> globalization, deregulation and contracting<br />
life cycles, combining fast decision making and<br />
flexibility to change. Such methodologies embrace<br />
both product function, market, life cycle analysis, toge<strong>the</strong>r<br />
with organisational and individual competencies<br />
identification and evaluation. The IT tools to be<br />
integrated in a distributed Extended Enterprise architecture<br />
(called BMA Business Modelling and Actuation<br />
) are: a Business Process Definition toolset (with<br />
modelling and simulation capabilities), a Competencies<br />
Identification and Evaluation environment, a Human<br />
Resources Planner and Scheduler, a Commercial<br />
Project Management tool (OPENPLAN) and a Workflow<br />
Management tool.<br />
Three main areas are targeted: <strong>the</strong> product, <strong>the</strong> process,<br />
and <strong>the</strong> competencies <strong>of</strong> human resources (p 3<br />
paradigm: product, people, process).<br />
On <strong>the</strong> product side, <strong>the</strong> project is concerned with<br />
<strong>the</strong> early conceptual phase where <strong>the</strong> links between<br />
functional specifications and customer’s value are<br />
identified and exploited to achieve maximum value at<br />
minimum cost, ensuring minimal environmental impact<br />
as well.<br />
Figure 1: Scenario and objectives <strong>of</strong> COMPETE
The PLANET <strong>Newsletter</strong> 9<br />
On <strong>the</strong> people side, <strong>the</strong> project is concerned with<br />
a system for managing human resource competencies<br />
in <strong>the</strong> short and medium term. Given that process<br />
quality and competencies are correlated, an ideal<br />
competencies pr<strong>of</strong>ile is associated to each activity<br />
and candidates for <strong>the</strong> activity are selected who better<br />
match this pr<strong>of</strong>ile. To support this procedure, competencies<br />
are routinely evaluated, and compared with<br />
medium term requirements to orientate human resources<br />
policies. Specific methodologies have been<br />
implemented to find organizational competencies ans<br />
evaluate individuals<br />
On <strong>the</strong> process side, <strong>the</strong> project is concerned with <strong>the</strong><br />
generic product development process (PDP) from <strong>the</strong><br />
early product concept to <strong>the</strong> detailed design. To slash<br />
costs and development times and to improve quality,<br />
enterprises are resorting to concurrency, BPR reengineering,<br />
standardisation, process control, knowledge<br />
distribution, outsourcing and extended enterprising.<br />
Stand alone tools have appeared to support this trend.<br />
The project provides a platform which supports <strong>the</strong>se<br />
new approaches, by integrating <strong>the</strong> existing tools,<br />
based on a sound logical model.<br />
Process modelling is crucial to achieve process effectiveness.<br />
Actual processes are assembled from<br />
standard templates, which consolidate <strong>the</strong> best prac-<br />
tices. Processes are released for execution after detailed<br />
simulation and planning, which balance effectiveness<br />
and competencies level. The unique combination<br />
<strong>of</strong> a Modeller, a Wf Management Tool and<br />
a Project Management tool caters for tight process<br />
control, reactiveness to unforeseen events, high quality<br />
and reduced processing times. The full potential<br />
<strong>of</strong> this layout can be obtained by proper interfacing<br />
with a PDM system. The systematic exploitation <strong>of</strong><br />
<strong>the</strong> existing interfacing standards and use <strong>of</strong> a Web<br />
based architecture makes <strong>the</strong> platform ideally suited<br />
to support extended enterprise arrangements.<br />
COMPETE Functions<br />
Figure 2: Functional diagram <strong>of</strong> COMPETE Platform<br />
The main functions supported by COMPETE platform<br />
are:<br />
1. To store <strong>the</strong> operational know-how by means<br />
<strong>of</strong> process modelling and to support <strong>the</strong> choise<br />
among different project alternatives estimating<br />
costs and times by means <strong>of</strong> simulation.<br />
2. To support <strong>the</strong> automatic translation <strong>of</strong> a process<br />
model into operational projects and into workflow<br />
flow-charts.
10 The PLANET <strong>Newsletter</strong><br />
3. To support planning according to competencies<br />
required by process activities and automatically<br />
assigning resources according to workload in a<br />
multiproject environment.<br />
4. To support progress control integrated with<br />
project management.<br />
5. To integrate different company departments, such<br />
as human resources department, design, product<br />
planning, project management, working teams, on<br />
a common platform and using a common data<br />
model.<br />
6. To manage in a unique environment planning and<br />
progress control <strong>of</strong> distributed teams in an extended<br />
enterprise.<br />
The platform integrates a project management tool<br />
(OPENPLAN) and commercial workflow management<br />
tools with new developments according to manufacturing<br />
companies requirements and has <strong>the</strong> following<br />
advantages:<br />
1. On <strong>the</strong> process side<br />
A clear structuring <strong>of</strong> operational know-how<br />
for porpuse <strong>of</strong> documenting, distributing and<br />
consolidating knowledge<br />
A poweful decisional support, based on cost<br />
and time indicators, obtained by means <strong>of</strong> simulation<br />
<strong>of</strong> different project alternatives<br />
2. On <strong>the</strong> competencies management side<br />
Availability <strong>of</strong> a catalog <strong>of</strong> competencies belonging<br />
to <strong>the</strong> organisations and to <strong>the</strong> individuals<br />
<strong>of</strong> distributed teams, in a common<br />
database and using a common glossary<br />
Easy evaluation <strong>of</strong> competency needs<br />
Support to <strong>the</strong> definition <strong>of</strong> careers and resources<br />
outsourcing<br />
Support to training plan definition<br />
Efficient resource allocation and conflict reduction<br />
Improvement <strong>of</strong> product quality by means <strong>of</strong> a<br />
better work share among working teams<br />
http://www.planet-noe.org<br />
3. On <strong>the</strong> project management side<br />
A higher reactiveness to emerging problems,<br />
by means <strong>of</strong> a deeper and structured knowledge<br />
<strong>of</strong> possible alternatives and available resources.<br />
COMPETE Operational Flow<br />
The operational flow supported by COMPETE project<br />
starts from process modeling, transforms process into<br />
operational projects, identifying optimal choices in<br />
terms <strong>of</strong> time and costs, supports activity planning<br />
and resources assignment according to a matching algorithm<br />
which compares <strong>the</strong> need <strong>of</strong> competencies<br />
<strong>of</strong> process activities with <strong>the</strong> availability <strong>of</strong> competencies<br />
in human resources according to <strong>the</strong>ir workload.<br />
In <strong>the</strong> end COMPETE supports project executions<br />
integrating project teams executed steps with<br />
<strong>the</strong> project manager.<br />
Process Modelling and Simulation The first<br />
step is process modeling. At this point it is important<br />
to explain which is <strong>the</strong> difference between a process<br />
and a project. A process describes a set <strong>of</strong> activities<br />
and takes into account possible alternatives related<br />
to sudden events which change <strong>the</strong> pre-defined<br />
flow. Unsatisfactory results <strong>of</strong> certain activities or<br />
new information coming from <strong>the</strong> market may require<br />
some changes or reworking or outsourcing. A<br />
process model foresees <strong>the</strong>se alternatives and supports<br />
a preventive analysis <strong>of</strong> <strong>the</strong> different solutions.<br />
It is possible to describe <strong>the</strong>se alternatives with modeling<br />
formalisms, COMPETE supports IDEF, which<br />
can be used by a simulation tool which estimates<br />
times and costs <strong>of</strong> <strong>the</strong> different solutions. This decision<br />
support is very important to react more rapidly
The PLANET <strong>Newsletter</strong> 11<br />
to problems during project execution. A process can<br />
generate many projects each one representing <strong>the</strong> activity<br />
flow <strong>of</strong> one alternative.<br />
COMPETE supports <strong>the</strong>n <strong>the</strong> automatic translation <strong>of</strong><br />
IDEF models into Gantt diagrams used by a project<br />
manager tool.<br />
The modeler can start from scratch or use pre-defined<br />
process templates that are available in <strong>the</strong> platform.<br />
New processes can be created as combination <strong>of</strong> existing<br />
ones. IDEF formalism is used to describe processes,<br />
sub-processes, activities and links. Each activity<br />
includes <strong>the</strong> description <strong>of</strong> its inputs, outputs,<br />
required competencies and duration. By means <strong>of</strong><br />
simulation it is possible to choose <strong>the</strong> best project,<br />
according to times, costs and competences availability,<br />
comparing <strong>the</strong> different gantt charts.<br />
Planning and Resource Scheduling<br />
Once <strong>the</strong> best project solution has been chosen, human<br />
resources will be allocated.<br />
COMPETE has developed a matching algorithm that,<br />
according to <strong>the</strong> competencies required by each activity,<br />
selects <strong>the</strong> proper candidates comparing <strong>the</strong>ir<br />
individual competencies with <strong>the</strong> required ones. In<br />
a second step <strong>the</strong> system compares required timings<br />
with individual workloads on a multi-project basis<br />
and proceeds in building working teams. A red<br />
traffic-light informs <strong>the</strong> project manager <strong>the</strong> missing<br />
<strong>of</strong> candidates for <strong>the</strong> required time interval and<br />
<strong>the</strong> necessity <strong>of</strong> acquisition, outsourcing or training.<br />
Manual adjustments are obviously possible, considering<br />
that human resource scheduling need to be<br />
much more flexible than machine scheduling. At this<br />
purpose an efficiency factor has been introduced to<br />
consider <strong>the</strong> fact that human resources with <strong>the</strong> same<br />
competencies can have speed and quality very different.<br />
Human resources assignment are <strong>the</strong>n exported in a<br />
project management tool which will represent graphically,<br />
gantt and workload.<br />
Figure 4: IDEF modeling <strong>of</strong> process and competences
12 The PLANET <strong>Newsletter</strong><br />
Progress Control<br />
Once <strong>the</strong> planning and <strong>the</strong> resource scheduling are<br />
completed, <strong>the</strong> working teams are ready to start working.<br />
COMPETE supports in this phase <strong>the</strong> aligning<br />
between <strong>the</strong> operative management and <strong>the</strong> progress<br />
control. The activity flow, firs modeled and <strong>the</strong>n<br />
planned is automatically imported in a workflow<br />
manager tool which supports activity dispatching.<br />
The activity manager will insert in <strong>the</strong> system information<br />
related to activity status and progress and <strong>the</strong><br />
project manager tool will be automatically updated<br />
according to <strong>the</strong>se data.<br />
Competencies Management<br />
COMPETE stresses <strong>the</strong> importance <strong>of</strong> <strong>the</strong> relation between<br />
competencies management and process quality.<br />
A competence is a combination <strong>of</strong> knowledge,<br />
capacity to apply it and to make it applied. It is essential<br />
for a company to model its competencies and<br />
http://www.planet-noe.org<br />
to relate <strong>the</strong>m to human resources. It is essential in an<br />
organization to compare its competencies with process<br />
and project needs.<br />
COMPETE has developed a set <strong>of</strong> tools and a database<br />
to create a competence tree, to describe <strong>the</strong> organization,<br />
to relate competencies, organization and human<br />
resources and to evaluate human resources according<br />
to <strong>the</strong>ir competencies. It is very important for a<br />
company to build its competence tree in order to understand<br />
which competencies are available and which<br />
one need to be acquired in order to face technological<br />
evolution.<br />
COMPETE Architecture<br />
Figure 5: COMPETE Architecture<br />
The improved prototypes developed in COMPETE are<br />
currently being engineered and packaged into a product<br />
suite composed by two main tools: SKILLPLAN<br />
(Skill & Process Planner) and P-CON (Process Actuation<br />
Control & Progress Review)
The PLANET <strong>Newsletter</strong> 13<br />
SKILLPLAN supports <strong>the</strong> process planning and optimisation,<br />
by means <strong>of</strong> process modelling, specification<br />
<strong>of</strong> competencies’ requirements, simulation and<br />
what-if analysis. Process simulation caters for alternative<br />
flows <strong>of</strong> activities and recycling probabilities.<br />
The plan is <strong>the</strong>n turned into operation: activities<br />
are assigned to resources according to required competencies,<br />
and scheduled. SKILLPLAN architecture<br />
is based on <strong>the</strong> XML integration <strong>of</strong> an IDEF-based<br />
modeller, which describes <strong>the</strong> process, a simulator<br />
aimed to disambiguate it analysing all <strong>the</strong> related<br />
possible project alternatives, a scheduler which uses<br />
an innovative algorithm to find <strong>the</strong> best resources,<br />
according to <strong>the</strong> required competencies and to <strong>the</strong><br />
time schedule. A commercial Project Manager has<br />
been integrated: this is Open Plan (integrated into <strong>the</strong><br />
suite) which supports multi-projects facilities.<br />
The main and innovative features <strong>of</strong> SKILLPLAN<br />
are: 1- <strong>the</strong> possibility to perform what-if analysis in<br />
order to find <strong>the</strong> best projects to implement a certain<br />
process, 2- <strong>the</strong> introduction <strong>of</strong> a competency based<br />
HR planning & Scheduling approach.<br />
P-CON aims at managing events, notifications and<br />
workflow (dispatching <strong>the</strong> activities among <strong>the</strong> actors<br />
involved) along <strong>the</strong> project lifecycle. It is tightly integrated<br />
through XML interfaces to SKILLPLAN in<br />
order to react as fast as possible to problems arising<br />
during project actuation.<br />
The formalisms used for description <strong>of</strong> processes and<br />
projects , such as IDEF, gantt, flowcharts have been<br />
completely integrated and can be imported and exported<br />
through XML interfaces.<br />
The architecture (Fig. 5) is applicable in an extended<br />
enterprise environment. The communication among<br />
<strong>the</strong> distributed tools <strong>of</strong> COMPETE platform is based<br />
on an asynchronous channel for XML models transmission<br />
and on a distributed database for competencies<br />
storage.<br />
Conclusion<br />
In conclusion a stream <strong>of</strong> new organizational approaches<br />
have been introduced in <strong>the</strong> recent years to<br />
streamline Product Development Process , to achieve<br />
better quality and trim <strong>the</strong> costs. Concurrent engineering,<br />
businesses process engineering, benchmarking,<br />
knowledge management, process standardization,<br />
Extended Enterprises have been added to more<br />
traditional automation tools such CAD, CAD/CAM<br />
and CAE. Stand alone tools have appeared to support<br />
<strong>the</strong>se new practices. COMPETE brings toge<strong>the</strong>r <strong>the</strong>se<br />
new approaches in an integrated platform, which ensures<br />
as much as possible <strong>the</strong> communication among<br />
tools <strong>of</strong> different make. The benefits <strong>of</strong> <strong>the</strong> system, its<br />
capability to coordinate concurrent activities in distributed<br />
environment, are such as to significantly impact<br />
on quality, time-to-market and costs in modern<br />
extended enterprise.<br />
Bibliography<br />
[1] G. Morra, M. Sanseverino Un progetto Globale:<br />
il caso Fiat, Sistemi e Impresa , September,<br />
1996.<br />
[2] G. Zollo, A. Cannavacciuolo, G. Capaldo, A.<br />
Ventre, A. Volpe The Organisational Evaluation<br />
Process, Fuzzy Economic Review, 1, no.1 (1996)<br />
3-30.<br />
[3] M.Argilli, S.Gusmeroli, G. Secco Suardo, N.<br />
Matino, M. Sanseverino COMPETE : a common<br />
platform for <strong>the</strong> extended enterprise 2000,<br />
e-Business and e-Work 2000 , October, 2000.<br />
[4] G. Michellone, G.Zollo Competencies management<br />
in knowledge based firms, Internationa<br />
Journal <strong>of</strong> Technology management, vol<br />
20, 2000.<br />
[5] G. Zollo, L. Iandoli, F. Borrelli, A Iuliano, M.<br />
Sanseverino Applications <strong>of</strong> s<strong>of</strong>t computing techniques<br />
to Product Design, International Seminar<br />
on Intelligent Computation in Manufacturing<br />
CIRP June, 2000.<br />
Author Information<br />
Marialuisa Sanseverino FIAT Research Center,<br />
marialuisa.sanseverino@crf.it
14 The PLANET <strong>Newsletter</strong><br />
In Defence <strong>of</strong> Reactive Scheduling<br />
Introduction<br />
Reactive Scheduling is <strong>the</strong> Cinderella <strong>of</strong> scheduling<br />
techniques. Reactive Scheduling drudges away in <strong>the</strong><br />
work place, delivering feasible solutions to highly<br />
complex and dynamic resource allocation problems<br />
but is seldom invited to conferences to show its finery<br />
to <strong>the</strong> princes <strong>of</strong> <strong>the</strong> scheduling and planning community.<br />
Their attention is distracted by yet ano<strong>the</strong>r algorithm<br />
that promises optimum results on some static<br />
benchmark problem. Indeed, a recent PLANET sponsored<br />
workshop on on-line planning and scheduling<br />
[1] did not even include reactive scheduling as one<br />
<strong>of</strong> its topics <strong>of</strong> interest until <strong>the</strong> second call for papers.<br />
This is strange, on-line scheduling domains are<br />
<strong>the</strong> most sensitive to sudden environmental change.<br />
The reality is that in stochastic environments <strong>the</strong> only<br />
option open to a scheduler is to react to change. Techniques<br />
for anticipating change are limited and seldom<br />
provide <strong>the</strong> functionality needed to manage resource<br />
allocations that are continually being impacted by<br />
disruptions <strong>of</strong> one form or ano<strong>the</strong>r.<br />
The Need for Reactive Schedulers<br />
The requirement for reactive scheduling systems is<br />
more widespread in <strong>the</strong> industrial world than most researchers<br />
from <strong>the</strong> academic world realise. For example,<br />
<strong>the</strong> production system used for garment manufacture<br />
in <strong>the</strong> UK is called a Progressive Bundle Line<br />
(PBL) [2, 3, 4]. It is a flow line manufacturing system<br />
in which work stations, comprising <strong>of</strong> machinist and<br />
sewing machine pairs, are arranged in a configuration<br />
so that <strong>the</strong> flow <strong>of</strong> work from one work station provides<br />
work for <strong>the</strong> next work station in <strong>the</strong> production<br />
sequence. The scheduling objective is to maintain a<br />
line balance so that <strong>the</strong> work in progress buffers between<br />
each work station never overflow, causing storage<br />
and quality problems, or empty, causing machin-<br />
http://www.planet-noe.org<br />
ARTICLE<br />
Author: J.E. Spragg<br />
ists to sit idle at <strong>the</strong>ir machines.<br />
There are examples <strong>of</strong> line balancing algorithms in<br />
<strong>the</strong> scheduling and optimisation literature. The majority<br />
<strong>of</strong> <strong>the</strong>se techniques view line balancing as a<br />
static optimisation problem and are wholly inadequate<br />
for tackling <strong>the</strong> management <strong>of</strong> a PBL. The fact<br />
<strong>of</strong> <strong>the</strong> matter is, in <strong>the</strong> real world, a well balanced<br />
line soon becomes unbalanced. Machines breakdown,<br />
machinists go absent or start working below, or<br />
above, standard performance, managers decide that<br />
priority job batches are no longer important and that<br />
low priority batches are important, quality controllers<br />
decide that rework is required on some batches, etc.<br />
In clothing factories that employ PBL production systems<br />
<strong>the</strong> most important person is <strong>the</strong> line supervisor<br />
who continuously monitors, analyses, and revises <strong>the</strong><br />
flow <strong>of</strong> work through <strong>the</strong> work stations. How well<br />
she (it is usually a she) does this depends upon her<br />
experience and training.<br />
Providing reactive scheduling capabilities to a realtime<br />
automated scheduler involves mimicking <strong>the</strong> responsibilities<br />
<strong>of</strong> <strong>the</strong> human supervisor employed to<br />
manage a PBL. An ideal reactive scheduling framework<br />
employs an event driven multi agent approach<br />
that applies monitoring, analysis, revision, and optimisation<br />
tools in real-time to an executing schedule<br />
to maintain its feasibility and quality over time.<br />
Reactive scheduling techniques can be reinforced by<br />
building robust schedules or providing probabilistic<br />
models <strong>of</strong> system behaviour. While nobody doubts<br />
<strong>the</strong> importance <strong>of</strong> <strong>the</strong>se techniques; if a machine is<br />
to breakdown it would be useful to have an indicator<br />
<strong>of</strong> which machine and when; <strong>the</strong>re is still <strong>the</strong> problem<br />
<strong>of</strong> what <strong>the</strong> system can do about it. The response<br />
to a machine breakdown for example is <strong>of</strong>ten context<br />
sensitive and depends upon <strong>the</strong> current opportunities<br />
for resource reallocation within <strong>the</strong> current schedule.<br />
Robust schedules limit <strong>the</strong> impact <strong>of</strong> a disturbance<br />
on a schedule but again <strong>the</strong> system still has to reason
The PLANET <strong>Newsletter</strong> 15<br />
about strategies for bringing <strong>the</strong> solution back to feasibility.<br />
At <strong>the</strong> end <strong>of</strong> <strong>the</strong> day, <strong>the</strong> system will need<br />
to react by relaxing due dates or reallocating activities<br />
to alternative resources or bumping lower priority<br />
operations.<br />
Mixed Initiative Scheduling – and beyond<br />
The OZONE Project [5] at <strong>the</strong> Intelligent Coordination<br />
and Logistics Laboratory, Carnegie Mellon University,<br />
back in <strong>the</strong> mid-90s <strong>of</strong> last century, recognised<br />
that current scheduling systems do not effectively<br />
support user tasks and requirements and do<br />
not support <strong>the</strong> iterative, evolving process <strong>of</strong> problem<br />
understanding, requirements determination, conflict<br />
resolution and solution refinement that is inherent<br />
in large, multi criteria problem solving. OZONE<br />
tackled <strong>the</strong>se issues by enabling flexible collaborative<br />
problem solving between user and system, supporting<br />
reconfiguration <strong>of</strong> system functionality to accommodate<br />
new environments and scheduling objectives.<br />
The OZONE approach to scheduling recognised<br />
that scheduling is an incremental process <strong>of</strong> ‘getting<br />
<strong>the</strong> constraints right’ in which human users always<br />
have <strong>the</strong> big-picture decision-making expertise and<br />
knowledge to contribute but are unable to effectively<br />
cope with <strong>the</strong> complexity <strong>of</strong> detailed solution development.<br />
The challenge for <strong>the</strong> new generation <strong>of</strong> real-time,<br />
on-line, schedulers is to encode <strong>the</strong> strategic knowledge<br />
that humans provide within an autonomous<br />
scheduling framework that can respond immediately<br />
and intelligently to change. When analysing a<br />
scheduling solution, both human and s<strong>of</strong>tware agents,<br />
ask <strong>the</strong> same two questions:<br />
Where are <strong>the</strong> processing bottlenecks?<br />
Where are <strong>the</strong> scheduling opportunities?<br />
A mixed initiative scheduling tool supports user experimentation<br />
with graphical displays and statistical<br />
summaries. An autonomous scheduler can keep<br />
internal representations <strong>of</strong> analysis results and provide<br />
revision tools that are activated by internal state<br />
changes to <strong>the</strong> control architecture. An autonomous<br />
scheduler needs to evaluate <strong>the</strong> ‘correctness’ <strong>of</strong> its revisions<br />
by comparing <strong>the</strong> results against ‘objective’<br />
criteria given by a multi-criteria cost function. The<br />
interpretation <strong>of</strong> such evaluations are <strong>of</strong>ten problematic<br />
and seldom objective. Techniques developed by<br />
decision <strong>the</strong>orists, such as criteria voting [6], for resolving<br />
conflicts in multi criteria decision making<br />
suggest promising alternatives to single valued measures<br />
<strong>of</strong> ‘optimality’ for self-governing systems.<br />
Reactive Problem Solvers – Schedulers<br />
for <strong>the</strong> new era<br />
The recent developments in scheduling technology<br />
have been partly driven by improvements in monitoring<br />
technology. Communication technology has<br />
made reactive scheduling more important by enabling<br />
schedulers to keep up to date with environmental<br />
change and allowing <strong>the</strong>m to respond to disruptions<br />
by resolving conflicts in impacted parts <strong>of</strong> <strong>the</strong> solution.<br />
Gone are <strong>the</strong> days in manufacturing where<br />
jobs were ‘chased’ through <strong>the</strong> factory by a progress<br />
chaser who spent <strong>the</strong>ir day sweet talking departmental<br />
foreman and climbing over work inventories looking<br />
for job numbers. The advent <strong>of</strong> mechanically<br />
readable bar codes means that work can now be continuously<br />
monitored through a production process.<br />
Mobile telephony has had a major impact on logistic<br />
scheduling, allowing huge savings by allowing<br />
field workers to communicate with headquarters<br />
in real-time. The British Telecom mobile workforce<br />
scheduling problem [7] addressed by a.p.solve<br />
is highly dependent upon mobile telephony. It is a<br />
problem characterised by:<br />
A varying workload with one hour response times.<br />
A diverse mixture <strong>of</strong> operational procedures that<br />
range from those with hard start-time constraints<br />
to those with highly relaxable start-time constraints.<br />
The processing times <strong>of</strong> scheduling activities can<br />
range from a few minutes to several days and are<br />
subject to uncertainty.<br />
Resource scarcities in rare skills.
16 The PLANET <strong>Newsletter</strong><br />
Scheduling activities can be complicated by dependency<br />
relationships between activities.<br />
The environment in which mobile workers operate<br />
is subject to uncertainty and change. Calculating<br />
travel times is problematic and subject to traffic<br />
conditions.<br />
British Telecom has a workforce <strong>of</strong> 20,000 technicians<br />
nation-wide with several hundred thousand<br />
tasks to be scheduled and executed every day.<br />
In a resource constrained environment, with a small<br />
time budget to resolve conflicts and improve solution<br />
quality, <strong>the</strong> system is forced to seek out spare capacity<br />
by searching a space <strong>of</strong> reallocation moves to find<br />
alternative resources and/or start times for activities<br />
in conflict. The most promising kinds <strong>of</strong> algorithms<br />
for this kind <strong>of</strong> constraint based search are those like<br />
Ginsberg’s and McAllester’s partial-order backtracking<br />
[8] that combine aspects <strong>of</strong> both systematic and<br />
non-systematic search. However, in practical environments<br />
partial-order backtracking needs to be supported<br />
by texture analysis that indicates <strong>the</strong> aggregate<br />
demand for alternative resources by providing measures<br />
<strong>of</strong> contention and reliance [9].<br />
Given <strong>the</strong> industrial need for reactive scheduling systems,<br />
why is <strong>the</strong> scheduling community obsessed<br />
with static optimisation problems? In practice,<br />
scheduling problems require a ‘solution’ to be more<br />
than <strong>the</strong> mere implementation <strong>of</strong> an algorithm for<br />
solving a particular constraint satisfaction, or constrained<br />
optimisation problem. Constructing schedules,<br />
in practical environments is an extended, iterated<br />
process that typically involves resolving conflicts<br />
between competing schedule users and scheduling<br />
tools. In most applied domains, schedules need<br />
to be maintained over time through reactive revision<br />
that refine an initial, or current, solution, by adapting<br />
it to changing environmental conditions and user<br />
preferences.<br />
Bibliography<br />
[1] AIPS 2002 Workshop on On-line Planning and<br />
Scheduling, Toulouse, France, April 24th, 2002.<br />
http://www.planet-noe.org<br />
[2] Fozzard, G., Spragg, J., and Tyler, D. Simulation<br />
<strong>of</strong> Flow Lines in Clothing Manufacture, Part<br />
1: model construction, Int.J. <strong>of</strong> Clothing Science<br />
and Technology, Vol 8, No. 4, pp 17-27, 1996.<br />
[3] Fozzard, G., Spragg, J., and Tyler, D. Simulation<br />
<strong>of</strong> Flow Lines in Clothing Manufacture, Part 2:<br />
credibility issues and experimentation, Int.J. <strong>of</strong><br />
Clothing Science and Technology, Vol 8, No. 5,<br />
pp 42-50, 1996.<br />
[4] Spragg, J.E., Fozzard, G. and Tyler, D.J. FLEAS:<br />
a flowline environment for automated supervision,<br />
The Int.J. <strong>of</strong> Manufacturing Technology<br />
Management, Vol. 10, Number 6, pp 322-327,<br />
1999.<br />
[5] Smith, S., Lassila, O., and Becker, M. Configurable,<br />
Mixed Initiative Systems for Planning<br />
and Scheduling, Advanced Planning Technology,<br />
Technological Achievements <strong>of</strong> <strong>the</strong> ARPA/Rome<br />
Laboratory Planning Initiative, Edited Austin<br />
Tate, AAAI Press, pp 235-241, 1996.<br />
[6] Croce, F.E., Tsoukias, A., and Moraitis, P. Why<br />
is it Difficult to Make Decisions Under Multiple<br />
Criteria, AIPS-02, PLANET Sponsored Workshop,<br />
Planning and Scheduling with Multiple<br />
Criteria, pp 41-45, 2002.<br />
[7] Lesaint, D., Azarmi, N., Laithwaite, R., and<br />
Walker, P. Engineering Dynamic Scheduler for<br />
Work Manager, BT Technology Journal, Vol. 16.<br />
No. 3 July, pp 16-29, 1996.<br />
[8] Ginsberg, M.L., and McAllester, D.A. GSAT and<br />
Dynamic Backtracking, Knowledge Representation<br />
and Reasoning Conference, 1994.<br />
[9] Beck, J.C., Davenport, A.J., Davis, E.D., and<br />
Fox, S. The ODO Project: Towards a Unified<br />
Basis for Constraint-Directed Scheduling, J. <strong>of</strong><br />
Scheduling, 1, 89-125 (1998).<br />
Author Information<br />
John E. Spragg a.p.solve, (www.apsolve.<br />
com), Sirius House, Adastral Park, Martlesham, Ipswich,<br />
Suffolk, IP5 3RE, United Kingdom, john.<br />
spragg@bt.com
The PLANET <strong>Newsletter</strong> 17<br />
ARTICLE<br />
Advanced Scheduling and Optimisation: Cutting <strong>the</strong> Costs <strong>of</strong><br />
Manufacturing<br />
Abstract<br />
The aim <strong>of</strong> this article is to describe several intelligent<br />
scheduling techniques that have been applied<br />
to problems in manufacturing and assembly.<br />
The techniques have been evaluated in domains<br />
including aircraft wing manufacturing, fiber optic<br />
cable manufacturing, submarine construction and<br />
CD manufacturing. The article will describe two<br />
particular scheduling techniques: schedule packing<br />
and squeaky wheel optimization. The techniques<br />
have resulted in a number <strong>of</strong> major improvements<br />
including a reduction in make-span <strong>of</strong> up to 50%,<br />
an improvement in throughput <strong>of</strong> 40%, a reduction<br />
in costs <strong>of</strong> 20% and <strong>the</strong> ability to tackle problems<br />
up to 20 times larger. The article provides<br />
an overview <strong>of</strong> <strong>the</strong>se techniques and describes two<br />
case studies from Boeing and Electric Boat. The<br />
article concludes with a description <strong>of</strong> <strong>the</strong> impact<br />
<strong>the</strong>se techniques have had in each <strong>of</strong> <strong>the</strong>se organizations<br />
and provides pointers to potential future<br />
improvements.<br />
Introduction<br />
Scheduling is <strong>the</strong> problem <strong>of</strong> assigning a set <strong>of</strong> tasks<br />
to a set <strong>of</strong> resources subject to a set <strong>of</strong> constraints.<br />
Examples <strong>of</strong> scheduling constraints include deadlines<br />
(e.g., job i must be completed by time t), resource capacities<br />
(e.g., <strong>the</strong>re are only four drills), precedence<br />
constraints on <strong>the</strong> order <strong>of</strong> tasks (e.g., a piece must<br />
be sanded before it is painted), and priorities on tasks<br />
(e.g., finish job j as soon as possible while meeting<br />
<strong>the</strong> o<strong>the</strong>r deadlines). In addition <strong>the</strong> task assignments<br />
must also meet a number <strong>of</strong> optimization criteria, e.g.<br />
minimize makespan, minimize set up times, maximize<br />
work in progress. Examples <strong>of</strong> scheduling domains<br />
include classical job-shop scheduling, manufacturing<br />
scheduling, and transportation scheduling.<br />
This article describes two generic scheduling tech-<br />
Author: B. Drabble<br />
niques that were developed and applied to problems<br />
in aircraft manufacturing and submarine assembly. In<br />
each case <strong>the</strong> results obtained are currently <strong>the</strong> best<br />
in <strong>the</strong> world. These techniques are now being augmented<br />
with additional functionality to tackle problems<br />
involving ship repair/overhaul and mission planning<br />
for <strong>the</strong> USAF. The techniques can be divided<br />
into two main areas:<br />
The use <strong>of</strong> non-systematic techniques such as<br />
“squeaky wheel” optimization [4] (SWO) and<br />
schedule packing (also known as doubleback optimization)<br />
[2] to solve problems that arise in manufacturing<br />
scheduling.<br />
The use <strong>of</strong> combined systematic and nonsystematic<br />
techniques, such as limited discrepancy<br />
search (LDS) [3] with schedule packing, and<br />
squeaky wheel with operations research methods.<br />
Scheduling Technologies<br />
This section provides an overview <strong>of</strong> <strong>the</strong> scheduling<br />
technologies that have been deployed in a number <strong>of</strong><br />
real applications.<br />
Limited Discrepancy Search (LDS) and<br />
Heuristics<br />
In scheduling problems <strong>of</strong> any size it is unlikely that<br />
always using a merely good heuristic will get you really<br />
close to an optimal schedule. A merely good<br />
heuristic will be incorrect some <strong>of</strong> <strong>the</strong> time. As<br />
<strong>the</strong> complexity <strong>of</strong> scheduling problems increases, <strong>the</strong><br />
number <strong>of</strong> decisions guided by <strong>the</strong> heuristic also increases.<br />
The more decisions made, <strong>the</strong> more likely it<br />
is that some <strong>of</strong> <strong>the</strong>m are going to be incorrect. How<br />
does LDS help address this problem? LDS is a systematic<br />
method for disregarding <strong>the</strong> recommendation <strong>of</strong>
18 The PLANET <strong>Newsletter</strong><br />
a heuristic a limited number <strong>of</strong> times (thus <strong>the</strong> name<br />
LDS) when generating a schedule. With LDS, schedules<br />
are generated repeatedly, each <strong>of</strong> <strong>the</strong>m following<br />
<strong>the</strong> heuristic for all decisions except one. The decision<br />
at which <strong>the</strong> heuristic is ignored is different in<br />
each schedule. If <strong>the</strong> heuristic leads to only one incorrect<br />
decision, <strong>the</strong>n using LDS1 (<strong>the</strong> fastest form<br />
<strong>of</strong> LDS) will lead to a perfect schedule. Even if <strong>the</strong><br />
heuristic leads to more than one incorrect decision<br />
(which is usually <strong>the</strong> case) <strong>the</strong>n LDS1 will likely lead<br />
to a better solution <strong>the</strong>n always following <strong>the</strong> heuristic.<br />
Schedule Packing<br />
Schedule Packing, also known as doubleback optimization<br />
involves “sloshing” a candidate schedule,<br />
repeatedly, right and left within a scheduling window.<br />
This has a remarkable impact on <strong>the</strong> length <strong>of</strong> most<br />
schedules. The Doubleback process is analogous to<br />
filling a box with blocks and <strong>the</strong>n shaking <strong>the</strong> box.<br />
Shaking <strong>the</strong> box will almost always result in a denser<br />
packing <strong>of</strong> blocks. Likewise, in schedules, Doubleback<br />
almost always results in a denser packing <strong>of</strong><br />
tasks in a schedule. The denser packing allows for<br />
tasks with few preceding activities to find appropriate<br />
holes in <strong>the</strong> schedule in which to be placed. The<br />
schedule packing algorithm is appropriate for problems<br />
with large numbers <strong>of</strong> precedences, e.g. assembly<br />
tasks, manufacturing, overhaul, etc.<br />
Squeaky Wheel Optimization<br />
The insight behind SWO is that in any real world<br />
problem it is impossible to capture all associated constraints,<br />
e.g. context information. SWO uses a priority<br />
queue to determine <strong>the</strong> order in which tasks should be<br />
released to a greedy scheduling algorithm. The priority<br />
queue is determined by how difficult <strong>the</strong> task is to<br />
deal with that is, i.e. higher <strong>the</strong> task is in <strong>the</strong> queue<br />
<strong>the</strong> harder it is to find a good resource assignment.<br />
On each iteration <strong>of</strong> <strong>the</strong> algorithm, SWO quickly creates<br />
a schedule and <strong>the</strong>n examines it to identify <strong>the</strong><br />
parts that were handled badly, for example, <strong>the</strong> task<br />
http://www.planet-noe.org<br />
was completed too late or assigned to an unsuitable<br />
agent. Any task that “squeaks” is promoted up <strong>the</strong><br />
priority queue, with <strong>the</strong> distance it is promoted determined<br />
by <strong>the</strong> extent <strong>of</strong> <strong>the</strong> problem. The new priority<br />
queue is <strong>the</strong>n used to generate ano<strong>the</strong>r schedule<br />
that is analyzed for problems. This process continues<br />
until no significant improvement in <strong>the</strong> schedule<br />
is noted over several iterations or a predefined limit is<br />
reached i.e. cycle count or elapsed time. SWO is extremely<br />
fast with each cycle <strong>of</strong> generate, analyze, and<br />
re-prioritize taking less than a second, even for large<br />
problems, e.g. 2500 tasks and 200 resources over a<br />
five day period.<br />
Application Domains<br />
This section describes two example domains that<br />
have been tackled using ei<strong>the</strong>r schedule packing or<br />
squeaky wheel optimization. In each example a description<br />
<strong>of</strong> <strong>the</strong> domain will provided toge<strong>the</strong>r with<br />
a description <strong>of</strong> <strong>the</strong> impact <strong>the</strong> technology has made.<br />
In addition to <strong>the</strong> domains described here <strong>the</strong> techniques<br />
have also been applied to submarine assembly,<br />
aircraft mission scheduling [5] and CD manufacturing.<br />
Full details can be found via pointer<br />
www.otsys.com/scheduling.html<br />
Aircraft Assembly<br />
The original aircraft assembly problems tackled by<br />
schedule packing were provided via a research group<br />
at McDonnell Douglas. These were real scheduling<br />
problems and were made available via <strong>the</strong> net to encourage<br />
academic researchers to demonstrate <strong>the</strong> applicability<br />
<strong>of</strong> <strong>the</strong>ir techniques. These particular problems<br />
are relevant to large scale assembly and are instances<br />
<strong>of</strong> problems known as resource constrained<br />
project scheduling (RCPS). CIRL developed a scheduler<br />
that produces <strong>the</strong> best known results on this problem.<br />
CIRL has also converted several o<strong>the</strong>r benchmark<br />
problems to <strong>the</strong> same format and solved <strong>the</strong>m<br />
successfully. The basic aircraft assembly problem<br />
has <strong>the</strong> following features:
The PLANET <strong>Newsletter</strong> 19<br />
Zone resources : A zone is an area <strong>of</strong> <strong>the</strong> aircraft<br />
in which work can be done. Zone resources specify<br />
<strong>the</strong> maximum number <strong>of</strong> people that can work<br />
in that zone at <strong>the</strong> same time.<br />
Labor resources : These specify how many laborers<br />
are available with a particular skill set.<br />
Shifts : The availability <strong>of</strong> labor resources varies<br />
over time, with more being available during <strong>the</strong><br />
day.<br />
Tasks : Tasks have a specified duration and set <strong>of</strong><br />
zone and labor resources that are needed to perform<br />
<strong>the</strong> task.<br />
Precedences : These specify which tasks must be<br />
completed before o<strong>the</strong>r tasks can begin.<br />
The program uses limited discrepancy search (LDS)<br />
and schedule packing (also known as doubleback optimization)<br />
to generate solutions. LDS and schedule<br />
packing can be used in isolation or in combination<br />
with each o<strong>the</strong>r with <strong>the</strong> best results produced<br />
using LDS with schedule packing or schedule packing<br />
on its own. When used in combination, multiple<br />
seed schedules generated with LDS are fed to schedule<br />
packing.<br />
In figure 1 <strong>the</strong> top part indicates resource usage over<br />
time. There is one horizontal bar for each <strong>of</strong> <strong>the</strong> 17<br />
Figure 1: Example <strong>of</strong> aircraft assembly<br />
resource types. The darker areas indicates a resource<br />
is fully utilized, and <strong>the</strong> lighter indicates it is unused.<br />
The boxes in <strong>the</strong> bottom part <strong>of</strong> <strong>the</strong> figures represent<br />
<strong>the</strong> tasks. The width <strong>of</strong> a box represents <strong>the</strong> duration<br />
<strong>of</strong> a task, and <strong>the</strong> height is an indicator <strong>of</strong> how<br />
many resources a task requires. The PERT 1 schedule<br />
for this problem ends after 37 days, 2 hours and<br />
58 minutes, and is a loose lower bound on <strong>the</strong> minimum<br />
length any potential solution. The current best<br />
schedule produced by McDonnell Douglas (now <strong>the</strong><br />
Boeing Company) is just over 42 days and <strong>the</strong> best<br />
schedule pack schedule is just over 39 days. Each<br />
day <strong>of</strong> production removed from <strong>the</strong> schedule saves<br />
<strong>the</strong> company approximately $600,000 and thus <strong>the</strong><br />
schedule is able to save approximately $3.2 million<br />
per sub-assembly. On Time Systems 2 has extended<br />
this scheduler to handle more general aircraft manufacturing<br />
problems. These new constraints and features<br />
include:<br />
multiple sub-assemblies could be introduced into<br />
<strong>the</strong> line at fixed rates, e.g. “every five days” or at<br />
arbitrary points, e.g. “40 wings over <strong>the</strong> next 10<br />
days”.<br />
Once a wing assembly was assigned to a bay it<br />
must return to <strong>the</strong> same bay for fur<strong>the</strong>r processing<br />
steps.<br />
1 The PERT schedule is generated by starting tasks as early as possible and ignoring all resource conflicts<br />
2 On Time Systems is a technology startup company that was developed to commercialize <strong>the</strong> optimization technology being de-<br />
veloped at CIRL and from o<strong>the</strong>r groups around <strong>the</strong> world.
20 The PLANET <strong>Newsletter</strong><br />
The assembly lines have a number <strong>of</strong> robots that<br />
must be taken down for routine maintenance.<br />
Submarine Assembly<br />
Existing scheduling systems in use at shipyards today,<br />
such as ARTEMIS, SAP, or PRIMIVERA, rely<br />
on “makespan minimization” techniques to develop<br />
schedules. Specifically, <strong>the</strong>y try to schedule tasks<br />
as early as possible, subject to constraints and resource<br />
availability pr<strong>of</strong>iles. Manual intervention is<br />
typically required to deal with overloaded resources<br />
and o<strong>the</strong>r difficulties, and <strong>the</strong> process <strong>of</strong> scheduling<br />
<strong>the</strong> construction <strong>of</strong> a single ship can take months.<br />
This approach relies on <strong>the</strong> conventional wisdom that<br />
it can never be wrong to get work done early, and<br />
<strong>the</strong> assumption that a short schedule is likely to be<br />
efficient, since o<strong>the</strong>rwise <strong>the</strong> inefficiently utilized resources<br />
could be loaded up to get more done earlier.<br />
OTS) research has shown that this conventional wisdom<br />
is, in fact, misleading. In mass production environments,<br />
makespan minimization <strong>of</strong>ten is a useful<br />
approach, since o<strong>the</strong>r jobs can help smooth out<br />
<strong>the</strong> resource loading artifacts <strong>the</strong> process induces. In<br />
shipyards, where it is not uncommon to have one, or<br />
at most a few, projects in process at any one time,<br />
makespan turns out to be a poor stand-in for <strong>the</strong> shipyards’<br />
more complex goals.<br />
OTS has developed a radically new approach to ship<br />
construction scheduling, one that addresses shipbuilding’s<br />
unique needs directly. The resulting system,<br />
ARGOS, is capable <strong>of</strong> scheduling multiple years’<br />
production across a whole yard in hours, instead <strong>of</strong><br />
months, without need for human intervention. The<br />
resulting schedules typically exhibit a 10-20% reduction<br />
in construction labor costs when compared with<br />
those in use in <strong>the</strong> yards today. Conversely, in situations<br />
where throughput is limited by <strong>the</strong> available<br />
manpower pool, ARGOS makes it possible to progress<br />
10 to 20% more work through <strong>the</strong> yard. All <strong>of</strong> <strong>the</strong>se<br />
savings are achieved without changing <strong>the</strong> fundamental<br />
production process or shipyard facility in any way.<br />
Table 1 shows <strong>the</strong> expected savings (over <strong>the</strong> cur-<br />
3 This is <strong>the</strong> schedule <strong>the</strong> yard is currently building to.<br />
http://www.planet-noe.org<br />
rent schedule) for a single hull and Table 2 shows<br />
<strong>the</strong> expected savings for <strong>the</strong> entire yard over <strong>the</strong> next<br />
5 years. Our estimates are that, if ARGOS were applied<br />
to all new Navy construction, annual savings<br />
could be expected to be between $200M and $500M.<br />
Numbers for re-fit and repair are more difficult to obtain,<br />
but <strong>the</strong> percentage savings (10-20%) appear to<br />
be comparable.<br />
iteration Time Savings<br />
1 2 min 8.4% $13.0M<br />
7 10 min 11.4% $17.7M<br />
20 34 min 11.8% $18.2M<br />
Ultimate ˜24 hrs 15.5% $24.0M<br />
Table 1: Expected Savings for a Single Hull<br />
iteration Time Savings<br />
1 24 min 7.8% $49M<br />
7 1 hour 10.2% $65M<br />
20 4 hours 10.7% $68M<br />
Ultimate 4 days 11.5% $73.0M<br />
Table 2: Expected Savings Over <strong>the</strong> Entire Yard<br />
Figure 2 provides a comparison <strong>of</strong> <strong>the</strong> resource utilization<br />
and manpower curves for <strong>the</strong> current schedule<br />
3 (top graph) and <strong>the</strong> one developed by ARGOS<br />
(bottom graph) . The black line represents <strong>the</strong> amount<br />
<strong>of</strong> manpower required on a particular day. The red<br />
and blue lines represent <strong>the</strong> actual manpower available<br />
in <strong>the</strong> shipyard schedule and ARGOS schedule<br />
respectively. Deviation from <strong>the</strong> black line results<br />
in additional costs due to overtime, under-time or<br />
increasing/reducing <strong>the</strong> total work-force. The AR-<br />
GOS schedule has a much smoo<strong>the</strong>r pr<strong>of</strong>ile requiring<br />
fewer changes in manpower levels and provides<br />
<strong>the</strong> ability to recover much more easily from unexpected<br />
changes in project deadlines. Figure 3 provides<br />
a comparison between <strong>the</strong> manpower needed in<br />
<strong>the</strong> current schedule (red line) and <strong>the</strong> manpower required<br />
by <strong>the</strong> ARGOS schedule for a single hull. This<br />
shows a smooth resource ramp “up and down” for<br />
<strong>the</strong> ARGOS schedule and far less perturbation in <strong>the</strong><br />
resource levels in <strong>the</strong> yard.
The PLANET <strong>Newsletter</strong> 21<br />
Figure 2: Manpower and Resource Utilization<br />
Comparison<br />
Figure 3: Single Hull Resource Utilization<br />
Comparison<br />
Summary<br />
Search based scheduling technologies have matured<br />
to <strong>the</strong> point that <strong>the</strong>y are now capable <strong>of</strong> solving<br />
large real world problems and provide users with high<br />
quality solutions. Table 3 provides a summary <strong>of</strong> <strong>the</strong><br />
improvement in problem size and complexity that can<br />
be handled by current techniques. In addition to being<br />
able to solve complex real world problems it is<br />
also interesting to note that <strong>the</strong> technology transfer<br />
path for <strong>the</strong>se methods is short, which makes <strong>the</strong>m<br />
easily accessible to industrial users.<br />
Tasks Resources Type Feasible<br />
1993 64 6 Job Shop No<br />
1996 ˜570 17 RCPS Barely<br />
1999 1000s Dozens RCPS Yes<br />
2001 10,000s Hundreds RCPS Yes<br />
2002 millions Hundreds RCPS Yes<br />
Table 3: Improvement in Problem Size and<br />
Complexity<br />
Bibliography<br />
[1] Aarup, M. and Arent<strong>of</strong>t, M.M. and Parrod, Y. and<br />
Stokes, I. and Vadon, H. and Stader, J., Optimum-<br />
AIV: A Knowledge-Based Planning and Scheduling<br />
System for Spacecraft AIV, Intelligent Scheduling,<br />
(eds. Zweben, M. and Fox, M.S) Morgan Kaufmann,<br />
Inc, 1994, pp451-469.<br />
[2] Crawford, J., An Approach to Resource Constrained<br />
Project Scheduling, in proceedings <strong>of</strong> <strong>the</strong> Artificial<br />
Intelligence and Manufacturing Research Planning<br />
Workshop, June 1996, AAAI SIGMAN.<br />
[3] Harvey, W. and Ginsberg, M., Limited Discrepancy<br />
Search, in <strong>the</strong> proceedings <strong>of</strong> Fourteenth International<br />
Joint Conference on Artificial Intelligence<br />
(IJCAI-95), July 1995, IJCAI Inc.<br />
[4] D. E. Joslin and D. P. Clements, Squeakywheel Optimization,<br />
in <strong>the</strong> proceedings <strong>of</strong> <strong>the</strong> Fifteenth National<br />
Conference on Artificial Intelligence, July,<br />
1998, AAAI Press, Menlo Park, CA.<br />
[5] Drabble, B., Task Decomposition Support to Reactive<br />
Scheduling, in <strong>the</strong> proceedings <strong>of</strong> <strong>the</strong> 5th <strong>European</strong><br />
Conference on Planning (ECP-99), Durham,<br />
September, 1999, Springer Verlag Press, New York,<br />
NY.<br />
Author Information<br />
Brian Drabble On Time Systems, Inc, 1850 Millrace<br />
Drive, Suite 1, Eugene, OR 97403-1269, USA,<br />
drabble@otsys.com
22 The PLANET <strong>Newsletter</strong><br />
2nd International PLANET Summer School<br />
After <strong>the</strong> great success <strong>of</strong> <strong>the</strong> first edition (Cyprus,<br />
September 2000), <strong>the</strong> 2nd International Summer<br />
School on AI Planning was held in Halkidiki, Greece<br />
on September 16-22, 2002.<br />
This event was one <strong>of</strong> a number <strong>of</strong> activities organized<br />
by PLANET (<strong>the</strong> <strong>European</strong> <strong>Network</strong> <strong>of</strong> <strong>Excellence</strong><br />
in AI planning), whose main aim is to promote<br />
knowledge exchange among students and researchers<br />
who are new to <strong>the</strong> field with a view to foster interaction<br />
and international collaborations. More than<br />
fifty PhD students and researchers from academia and<br />
industry attended <strong>the</strong> School, which <strong>of</strong>fered courses<br />
held by top-level, internationally-renown speakers.<br />
The courses were divided into eight distinct parts,<br />
which covered most <strong>of</strong> <strong>the</strong> current “hot” topics in AI<br />
planning:<br />
Planning under Uncertainty with Markov<br />
Decision Processes<br />
Craig Boutilier<br />
Markov Decision Processes (MDPs) are a widely<br />
used computational method for solving sequential decision<br />
problems involving uncertainty. This course<br />
provided a brief introduction to MDPs, and focused<br />
on <strong>the</strong> methods <strong>of</strong> representation and solution that are<br />
strictly related to AI planning. A greater emphasis<br />
was put on approaches that reduce <strong>the</strong> computational<br />
effort for solving MDPs through <strong>the</strong> adoption <strong>of</strong> techniques<br />
developed in <strong>the</strong> AI community.<br />
Plan-based Control <strong>of</strong> Autonomous Robots<br />
Michael Beetz<br />
Given <strong>the</strong> increasing interest in autonomous robotic<br />
applications and in improving <strong>the</strong> performances<br />
<strong>of</strong> current systems (MARTHA, XAVIER, RHINO,<br />
MINERVA, REMOTE AGENT), this course aimed<br />
to provide <strong>the</strong> attendees with a broad overview <strong>of</strong><br />
http://www.planet-noe.org<br />
REPORT<br />
Author: L. Compagna, G. Cortellessa, A. Farinelli, and N. Policella<br />
<strong>the</strong> issues involved in <strong>the</strong> plan-based control <strong>of</strong> autonomous<br />
robots. Michael presented <strong>the</strong> computational<br />
principles and basic s<strong>of</strong>tware architectures that<br />
enable robots to perform complex and diverse task in<br />
dynamic environments. The need for an integrated<br />
approach among plan representation, reasoning, execution<br />
and learning was strongly underlined.<br />
Planning history and Overview<br />
Susanne Biundo<br />
These lectures <strong>of</strong>fered a comprehensive introduction<br />
to <strong>the</strong> field <strong>of</strong> AI Planning. They reviewed existing<br />
planning methods (classical, heuristic search, hierarchical<br />
and hybrid, and logic-based), describing in<br />
detail <strong>the</strong>ir domain representations and introducing<br />
present and future application areas. In addition, Susanne<br />
also presented developments towards a systematic<br />
combination and integration <strong>of</strong> different planning<br />
methods, as well as <strong>the</strong> integration and use <strong>of</strong> techniques<br />
from related fields <strong>of</strong> research. The course<br />
was concluded by a useful historical overview <strong>of</strong> <strong>the</strong><br />
AI Planning field.<br />
Scheduling and Planning<br />
Brian Drabble<br />
Brian presented an overview <strong>of</strong> recent developments<br />
in intelligent scheduling and optimization and <strong>the</strong><br />
ways in which <strong>the</strong>se systems and algorithms can be<br />
integrated with planners to develop a comprehensive<br />
approach to <strong>the</strong> planning and scheduling problem.<br />
Traditionally, scheduling and planning were viewed<br />
as separate research areas. However, this is a simplistic<br />
view, as many decisions in planning have a direct<br />
impact on scheduling, and viceversa. The overview<br />
provided details <strong>of</strong> several scheduling techniques and<br />
described many <strong>of</strong> <strong>the</strong>ir properties.
The PLANET <strong>Newsletter</strong> 23<br />
Planning and resources<br />
Philippe Laborie<br />
Philippe’s lectures began with <strong>the</strong> introduction <strong>of</strong> <strong>the</strong><br />
concept <strong>of</strong> resource – any substance or (set <strong>of</strong>) object(s)<br />
whose cost or available quantity induce some<br />
constraint on <strong>the</strong> operations that use it – and by showing<br />
application domains for planning with resources.<br />
The course <strong>the</strong>n proceeded with a review <strong>of</strong> <strong>the</strong> state<br />
<strong>of</strong> <strong>the</strong> art <strong>of</strong> planning with resources, and presented<br />
basic tools and planning techniques. Finally, Philippe<br />
also described in detail one <strong>of</strong> <strong>the</strong> most promising approach<br />
for dealing with resources in planning, which<br />
relies on <strong>the</strong> application <strong>of</strong> constraint-based techniques<br />
in partial-order or hierarchical task network<br />
(HTN) planning.<br />
Planners Performance Evaluation<br />
Maria Fox<br />
The goal <strong>of</strong> this course was to analyze and discuss<br />
methods for <strong>the</strong> evaluation <strong>of</strong> planning systems.<br />
Planners can be evaluated in a number <strong>of</strong> differ-<br />
ent ways: by analysis <strong>of</strong> <strong>the</strong>ir formal properties, by<br />
empirical comparison with o<strong>the</strong>r similar systems, in<br />
terms <strong>of</strong> <strong>the</strong> time/quality trade-<strong>of</strong>fs <strong>the</strong>y make, and<br />
so on. This course outlined <strong>the</strong> stages <strong>of</strong> a scientific<br />
approach to evaluation <strong>of</strong> a data set. Moreover <strong>the</strong><br />
course introduced some statistical tests that can be<br />
used to determine <strong>the</strong> significance <strong>of</strong> a feature <strong>of</strong> a<br />
data set.<br />
Planning and Execution<br />
Martha Pollack<br />
These lectures dealt with <strong>the</strong> interesting problem <strong>of</strong><br />
planning and execution applied mainly to Simple<br />
Temporal Problems. The speaker highlighted <strong>the</strong> fact<br />
that classical planning makes strong simplifying assumptions<br />
on <strong>the</strong> world model. In particular, in real<br />
environments <strong>the</strong> beliefs and goals at <strong>the</strong> base <strong>of</strong><br />
<strong>the</strong> original planning problem are subject to changes,<br />
and hence <strong>the</strong> solution plan might loose consistency.<br />
Martha provided an helpful introduction to this topic,<br />
dwelling on <strong>the</strong> following techniques: (a) interleaving<br />
planning and execution; (b) monitoring plan ex-
24 The PLANET <strong>Newsletter</strong><br />
ecution and inferring execution status; (c) recovering<br />
from failure by replanning; (d) maintaining <strong>the</strong><br />
temporal consistency <strong>of</strong> plans during execution (plan<br />
dispatch); (e) managing commitments and updating<br />
plans in response to changes in <strong>the</strong> environment.<br />
Planning and <strong>the</strong> Web<br />
Craig Knoblock<br />
This course provided an interesting overview <strong>of</strong> <strong>the</strong><br />
techniques involved in <strong>the</strong> context <strong>of</strong> ga<strong>the</strong>ring and<br />
integrating information from <strong>the</strong> web. These topics<br />
include query planning for information ga<strong>the</strong>ring, interactive<br />
planning using constraint propagation, and<br />
efficient execution <strong>of</strong> information ga<strong>the</strong>ring plans.<br />
Craig pointed out how information ga<strong>the</strong>ring from<br />
<strong>the</strong> web can be tackled as a planning problem, as it<br />
can be viewed as a process aimed at formulating a<br />
scheme or program for <strong>the</strong> accomplishment <strong>of</strong> some<br />
goal.<br />
Poster Session<br />
The School also included a poster session, during<br />
which attendees were able to present <strong>the</strong> recent developments<br />
<strong>of</strong> <strong>the</strong>ir work. This allowed fur<strong>the</strong>r interaction<br />
between students, researchers, lecturers and organizers<br />
and enabled <strong>the</strong> authors (more than fifteen)<br />
to receive useful feedback and suggestions.<br />
Social Activities<br />
The local organizers put toge<strong>the</strong>r a varied programme<br />
<strong>of</strong> social and cultural activities. These activities comprised<br />
an exciting excursion to <strong>the</strong> caves <strong>of</strong> Petralonia,<br />
a visit to <strong>the</strong> historical city <strong>of</strong> Thessaloniki, and<br />
a short trip to Nea Skioni, a traditional fishing village.<br />
This allowed <strong>the</strong> participants to get in touch<br />
with <strong>the</strong> ancient history, local cuisine, origins, music<br />
and dances <strong>of</strong> this fascinating country.<br />
Conclusions<br />
The School was a great success, providing not only<br />
an invaluable and up-to-date review <strong>of</strong> <strong>the</strong> recent<br />
http://www.planet-noe.org<br />
developments and techniques in AI Planning and<br />
Scheduling, but also a friendly and informal environment<br />
in which interaction was facilitated and<br />
research collaborations fostered. Many challenging<br />
questions were raised during <strong>the</strong> lectures, stimulating<br />
discussion and promoting knowledge exchange<br />
amongst all <strong>of</strong> <strong>the</strong> participants. Course<br />
materials, long abstract <strong>of</strong> <strong>the</strong> posters and lots<br />
<strong>of</strong> funny photos can be found at <strong>the</strong> web page<br />
<strong>of</strong> <strong>the</strong> school http://cswww.essex.ac.uk/<br />
PLANET/summer-school-02/.<br />
Acknowledgements<br />
The realization <strong>of</strong> this event – sponsored by PLANET<br />
– required <strong>the</strong> coordinated effort <strong>of</strong> several people,<br />
including Susanne Biundo (University <strong>of</strong> Ulm, Germany),<br />
Enrico Giunchiglia (University <strong>of</strong> Genoa,<br />
Italy), Sam Steel (University <strong>of</strong> Essex, UK), Ioannis<br />
Refanidis (University <strong>of</strong> Macedonia, Greece), and<br />
Ioannis Vlahavas (Aristotele University <strong>of</strong> Thessaloniki,<br />
Greece). We would also like to thank Max<br />
Garagnani for his help in <strong>the</strong> production <strong>of</strong> this document.<br />
Author Information<br />
Luca Compagna DIST - University <strong>of</strong> Genoa,<br />
Italy, compa@dist.unige.it<br />
Gabriella Cortellessa ISTC-CNR [PST], Planning<br />
and Scheduling Team, Institute for Cognitive<br />
Science and Technology, National Research Council<br />
<strong>of</strong> Italy, Rome, Italy, corte@ip.rm.cnr.it<br />
Alessandro Farinelli DIS, Dipartimento di Informatica<br />
e Sistemistica, Universita’ degli Studi di<br />
Roma “La Sapienza”, Rome, Italy, farinelli@<br />
dis.uniroma1.it<br />
Nicola Policella DIS, Dipartimento di Informatica<br />
e Sistemistica, Universita’ degli Studi di Roma<br />
“La Sapienza”, Rome, Italy, policella@dis.<br />
uniroma1.it
The PLANET <strong>Newsletter</strong> 25<br />
ARTICLE<br />
Generation and Execution <strong>of</strong> Partially Correct Plans in Dynamic<br />
Environments<br />
Introduction<br />
In this work we present <strong>the</strong> recent developments <strong>of</strong><br />
<strong>the</strong> approach to <strong>the</strong> design <strong>of</strong> Cognitive Robots (i.e.<br />
robots whose actions are driven by a formally developed<br />
<strong>the</strong>ory <strong>of</strong> action), that are capable <strong>of</strong> performing<br />
tasks in a coordinated way. The logic <strong>of</strong> actions that<br />
we adopt is an epistemic dynamic logic, where it is<br />
possible to derive acyclic branching plans (branches<br />
corresponding to sensing actions), including primitive<br />
parallel actions.<br />
In <strong>the</strong> present work, we consider an extended notion<br />
<strong>of</strong> plan by admitting a simple class <strong>of</strong> cycles that arise<br />
from <strong>the</strong> attempt to recover from <strong>the</strong> failure states<br />
originated by sensing actions. The proposed extension<br />
allows us to address <strong>the</strong> problem <strong>of</strong> generating<br />
plans that handle a form <strong>of</strong> synchronization based on<br />
<strong>the</strong> recognition <strong>of</strong> specific situations through sensing<br />
actions, including forms <strong>of</strong> coordination required in<br />
a multi-robot scenario.<br />
System Architecture<br />
In this section we recall <strong>the</strong> layered hybrid architecture<br />
used for our cognitive mobile robots (see<br />
also [4]) displayed in Fig. 1, that has been implemented<br />
on several different kinds <strong>of</strong> robotic platforms,<br />
namely Sony AIBOs, Pioneer, and homemade<br />
wheeled robots.<br />
The deliberative level is formed by three main components:<br />
The Plan Execution Module that is executed<br />
on-line during <strong>the</strong> accomplishment <strong>of</strong> <strong>the</strong> robot’s task<br />
and is responsible for executing a plan by coordinating<br />
<strong>the</strong> primitive actions <strong>of</strong> a single robot. The Coordination<br />
Module that is responsible for assigning<br />
tasks to <strong>the</strong> robots in <strong>the</strong> team according to <strong>the</strong> current<br />
situation. The Plan Generation Module, that is executed<br />
<strong>of</strong>f-line before <strong>the</strong> beginning <strong>of</strong> <strong>the</strong> robot’s mis-<br />
Authors: A. Farinelli, G. Grisetti, L. Iocchi, D. Nardi, and R. Rosati<br />
sion, and generates a set <strong>of</strong> plans to deal with some<br />
specific situations.<br />
Off-line<br />
Deliberative Level<br />
On-line<br />
Deliberative Level<br />
Operative Level<br />
Sensors<br />
KB<br />
Perception<br />
Plan<br />
Generation<br />
High-level<br />
Conditions<br />
Coordination<br />
Module<br />
World<br />
Model<br />
Plan<br />
Library<br />
Plan<br />
Execution<br />
Primitive<br />
Actions<br />
Figure 1: Layered architecture for our robots<br />
Plan Representation and Generation<br />
Actuators<br />
In order to address <strong>the</strong> problem <strong>of</strong> synchronize different<br />
robotic platforms using sensing action we used<br />
a particular notion <strong>of</strong> plan that we called Partially<br />
Strong Plan. This notion <strong>of</strong> plan is equivalent to <strong>the</strong><br />
definition <strong>of</strong> strong cyclic plans given in [3]. A Partially<br />
Strong Plan is a plan that if terminate leads to<br />
a goal state. Namely, a partially strong plan is a plan<br />
that is not guaranteed to terminate: termination actually<br />
depends on <strong>the</strong> outcome <strong>of</strong> <strong>the</strong> sensing actions<br />
in <strong>the</strong> plan. However, if such a plan terminates, <strong>the</strong>n<br />
it always leads to <strong>the</strong> goal. The generation <strong>of</strong> plans<br />
is based on <strong>the</strong> use <strong>of</strong> a modal non monotonic de-<br />
scription logic ¢¡¤£¦¥¨§�©<br />
[2]. As illustrated in [5],<br />
<strong>the</strong> set <strong>of</strong> models <strong>of</strong> an �¡¤£¦¥ §�© knowledge base �<br />
formalizing a dynamic system can be represented by<br />
means <strong>of</strong> a unique transition graph, called first-order<br />
extension (FOE) <strong>of</strong> � , which represents all <strong>the</strong> possible<br />
evolutions <strong>of</strong> <strong>the</strong> dynamic system. For instance,<br />
Fig. 2 displays some examples <strong>of</strong> portions <strong>of</strong> FOEs.
26 The PLANET <strong>Newsletter</strong><br />
senseA (T)<br />
X1<br />
X0<br />
senseA (T) senseA (F)<br />
A<br />
senseA (T)<br />
senseA (F)<br />
X2<br />
B<br />
senseA (F)<br />
senseA (T)<br />
senseA (T)<br />
X1<br />
GOAL<br />
X0 X0<br />
senseA (T)<br />
senseA (F)<br />
senseA (F)<br />
X2<br />
senseA (F)<br />
senseA (T)<br />
a) b)<br />
c)<br />
The plan generation module selects a portion <strong>of</strong> <strong>the</strong><br />
FOE <strong>of</strong> <strong>the</strong> KB containing only those actions that are<br />
necessary to achieve a goal starting from a given initial<br />
situation. In fact, conditional plans can in principle<br />
be generated in two steps (see in [5] for details).<br />
First, <strong>the</strong> FOE <strong>of</strong> <strong>the</strong> knowledge base is generated;<br />
this FOE can be seen as an action graph representing<br />
all possible plans starting from <strong>the</strong> initial state. Then,<br />
such a graph is visited, building a term (<strong>the</strong> cyclic<br />
conditional plan) representing a graph in which: (i)<br />
sensing actions generate branches; (ii) each branch<br />
leads ei<strong>the</strong>r to a state satisfying <strong>the</strong> goal or to a previous<br />
state <strong>of</strong> <strong>the</strong> plan. However, <strong>the</strong> current implementation<br />
does not build <strong>the</strong> entire FOE before searching<br />
for <strong>the</strong> plan, but it builds <strong>the</strong> FOE starting from <strong>the</strong><br />
initial state with a breadth-first technique until a goal<br />
state is reached. In case <strong>of</strong> sensing actions all <strong>the</strong><br />
branches are required to reach a goal state. In this<br />
way it is possible to extract a minimal plan (with respect<br />
to <strong>the</strong> number <strong>of</strong> actions to be executed).<br />
Implementation<br />
We provided an implementation <strong>of</strong> our approach using<br />
a simulator. The situation we experimented as<br />
an example for explaining our plan generation and<br />
execution mechanism, is <strong>the</strong> dynamic exchange <strong>of</strong><br />
http://www.planet-noe.org<br />
Figure 2: Plan structure for cyclic sensing actions<br />
GOAL<br />
senseA (F)<br />
<strong>the</strong> role <strong>of</strong> goal keeper in <strong>the</strong> Sony Legged League<br />
and <strong>the</strong> application <strong>of</strong> <strong>the</strong> two-defender rule. The<br />
situation presented is a typical situation in <strong>the</strong> Sony<br />
Legged League matches in which <strong>the</strong> goal keeper<br />
(robot number 1) is moving away from its own goal<br />
and is approaching <strong>the</strong> ball to push it away, while ano<strong>the</strong>r<br />
robot (robot number 2) is far away from <strong>the</strong> ball<br />
and it cannot help <strong>the</strong> goal keeper immediately. In<br />
this situation, it is more convenient for <strong>the</strong> team that<br />
robot 1 takes <strong>the</strong> role <strong>of</strong> attacker pushing <strong>the</strong> ball towards<br />
<strong>the</strong> opposite goal, while robot 2 goes back to<br />
defend its own goal acting as a goal keeper. However,<br />
in performing this role exchange <strong>the</strong> two robots must<br />
comply with <strong>the</strong> two defenders rule, and thus robot<br />
2 can enter <strong>the</strong> goal area only after robot 1 has left<br />
it. The problem <strong>of</strong> complying with <strong>the</strong> two-defender<br />
rule is solved by generating a plan in which one robot,<br />
before entering <strong>the</strong> goal area, must check that it is free<br />
(i.e. <strong>the</strong> o<strong>the</strong>r robot has left <strong>the</strong> area). This is achieved<br />
by adding in <strong>the</strong> knowledge base <strong>of</strong> <strong>the</strong> robot <strong>the</strong><br />
specification <strong>of</strong> a sensing action SenseFreeArea that<br />
is used for verifying if <strong>the</strong> goal area is not occupied<br />
by any robot <strong>of</strong> <strong>the</strong> team. Even though <strong>the</strong> simulation<br />
cannot provide a precise characterization <strong>of</strong> all <strong>the</strong><br />
aspects that influence <strong>the</strong> performance <strong>of</strong> <strong>the</strong> robot in<br />
<strong>the</strong> real environment, it can provide useful feedback
The PLANET <strong>Newsletter</strong> 27<br />
to <strong>the</strong> design <strong>of</strong> <strong>the</strong> coordination and plan execution<br />
modules for actual robots. Through this simulator we<br />
have verified <strong>the</strong> intended behaviour <strong>of</strong> <strong>the</strong> robots in<br />
each <strong>of</strong> <strong>the</strong> roles in different scenarios.<br />
Conclusions<br />
As compared with previous work in Cognitive<br />
Robotics this is a novel attempt to generate plans that<br />
include cycles in a partially known environment. As<br />
compared with <strong>the</strong> work on classical planning <strong>the</strong>re<br />
is a close relationship with <strong>the</strong> work in [1], where<br />
only conditional plans (tree-structured) are generated.<br />
We are currently addressing <strong>the</strong> application <strong>of</strong><br />
model checking techniques, as done in [1], also in our<br />
setting. Moreover, we are extending our analysis to<br />
generate plans with cycles that are more general than<br />
<strong>the</strong> ones presented in this paper.<br />
Bibliography<br />
[1] P. Bertoli, A. Cimatti, M. Roveri, and P. Traverso.<br />
Planning in nondeterministic domains under partial<br />
observability via symbolic model checking.<br />
In Proc. <strong>of</strong> <strong>the</strong> 17th Int. Joint Conf. on Artificial<br />
Intelligence (IJCAI 2001), 2001.<br />
Dynamic Ontology Refinement<br />
Introduction<br />
Creating a plan that is guaranteed to be executable in<br />
a certain domain depends not only on forming plans<br />
correctly but also on having a perfect understanding<br />
<strong>of</strong> that domain. Unfortunately, developing this understanding<br />
and representing it fully is possible only<br />
in small, static domains. In more complex environments,<br />
plans may be based on incomplete or incorrect<br />
information and hence may be unexecutable. Interaction<br />
with <strong>the</strong> environment through attempted execution<br />
<strong>of</strong> <strong>the</strong> plan leads to an enriched and fuller un-<br />
[2] Francesco M. Donini, Daniele Nardi, and Riccardo<br />
Rosati. Description logics <strong>of</strong> minimal<br />
knowledge and negation as failure. ACM Trans.<br />
on Computational Logic, 3(2):1–49, 2002.<br />
[3] Fausto Giunchiglia and Paolo Traverso. Planning<br />
as model checking. In Proc. <strong>of</strong> <strong>the</strong> 5th Eur. Conf.<br />
on Planning (ECP’99), 1999.<br />
[4] Luca Iocchi. Design and Development <strong>of</strong> Cognitive<br />
Robots. PhD <strong>the</strong>sis, Univ. ”La Sapienza”,<br />
Roma, Italy, 1999.<br />
[5] Luca Iocchi, Daniele Nardi, and Riccardo Rosati.<br />
Planning with sensing, concurrency, and exogenous<br />
events: logical framework and implementation.<br />
In Proceedings <strong>of</strong> <strong>the</strong> Seventh International<br />
Conference on Principles <strong>of</strong> Knowledge Representation<br />
and Reasoning (KR’2000), pages 678–<br />
689, 2000.<br />
Author Information<br />
Alessandro Farinelli, Giorgio Grisetti, Luca<br />
Iocchi, Daniele Nardi, and Riccardo Rosati<br />
Dipartimento di Informatica e Sistemistica, Università<br />
“La Sapienza”, Roma, Italy<br />
ARTICLE<br />
Authors: F. McNeill, A. Bundy, and M. Schorlemmer<br />
derstanding <strong>of</strong> <strong>the</strong> environment but also leads to plan<br />
failure. This failure could be due to part <strong>of</strong> <strong>the</strong> <strong>the</strong>ory<br />
itself, such a missing axiom in a rule, or due to <strong>the</strong><br />
underlying ontology, such as a predicate with an incorrect<br />
arity. We <strong>the</strong>refore propose to devise a system<br />
that can dynamically incorporate this new knowledge<br />
into <strong>the</strong> <strong>the</strong>ory as <strong>the</strong> plan is being executed.<br />
Our system will be based around a central planimplementation<br />
agent, who will control all <strong>the</strong> o<strong>the</strong>r<br />
components <strong>of</strong> <strong>the</strong> system. This agent will firstly<br />
send <strong>the</strong> <strong>the</strong>ory, toge<strong>the</strong>r with <strong>the</strong> goal, to <strong>the</strong> planner,<br />
which will <strong>the</strong>n return a plan annotated with a justifi-
28 The PLANET <strong>Newsletter</strong><br />
cation for each plan step. The plan-implementation<br />
agent will <strong>the</strong>n use <strong>the</strong> plan justification, toge<strong>the</strong>r<br />
with fur<strong>the</strong>r questioning <strong>of</strong> <strong>the</strong>se agents, to discover<br />
why this failure occurred and which part <strong>of</strong> <strong>the</strong> <strong>the</strong>ory<br />
is at fault. This problem point is <strong>the</strong>n passed to <strong>the</strong> refinement<br />
system, which will <strong>the</strong>n correct <strong>the</strong> problem<br />
and replace it with <strong>the</strong> correction in <strong>the</strong> original <strong>the</strong>ory.<br />
This refinement process will not only allow us<br />
to achieve a goal that would o<strong>the</strong>rwise be unreachable,<br />
but will also leave us with a domain <strong>the</strong>ory that<br />
is richer and more accurate.<br />
Forming and Executing <strong>the</strong> Plan<br />
We need to find a plan for achieving <strong>the</strong> goal and to<br />
annotate this with a justification for each step. The<br />
reason this justification is required is that failure in<br />
<strong>the</strong> plan execution can be immediately linked to a<br />
problem in a particular part <strong>of</strong> <strong>the</strong> <strong>the</strong>ory; namely,<br />
that part <strong>of</strong> <strong>the</strong> <strong>the</strong>ory that was used to justify this plan<br />
step. We propose to use a state <strong>of</strong> <strong>the</strong> art planner such<br />
as FF so that our system is capable <strong>of</strong> producing long<br />
and complex plans if necessary. However, using such<br />
a planner will not provide us with information about<br />
<strong>the</strong> inference rules and justifications behind each plan<br />
step. Therefore we intend to build a plan deconstuctor<br />
that will take <strong>the</strong> plan produced by FF and, using<br />
<strong>the</strong> <strong>the</strong>ory, pseudo-execute it, at each stage annotating<br />
<strong>the</strong> <strong>the</strong>ory with <strong>the</strong> inference rule that was used<br />
and <strong>the</strong> preconditions <strong>of</strong> that rule, toge<strong>the</strong>r with <strong>the</strong>ir<br />
justifications.<br />
Once <strong>the</strong> plan-implementation agent has received <strong>the</strong><br />
annotated plan, it will attempt to execute it by interaction<br />
with o<strong>the</strong>r agents. For example, if <strong>the</strong> action<br />
required in a plan step is to buy a visa, it will need<br />
to locate an embassy agent who is capable <strong>of</strong> fulfilling<br />
this requirement. It is through <strong>the</strong>se interactions<br />
that any fault in <strong>the</strong> <strong>the</strong>ory will come to light. These<br />
agents will have <strong>the</strong>ir own internal <strong>the</strong>ories about how<br />
to perform <strong>the</strong>se actions, and <strong>the</strong>se may not match <strong>the</strong><br />
plan-implementation agent’s <strong>the</strong>ory.<br />
For example, <strong>the</strong> plan-implementation agent may<br />
have a visa represented in his <strong>the</strong>ory as a simple, 0-<br />
http://www.planet-noe.org<br />
arity predicate, whereas <strong>the</strong> embassy agent may have<br />
it represented as a 1-arity predicate visa(country)<br />
which takes a country as an argument. Or perhaps <strong>the</strong><br />
embassy agent has a more complicated rule for buying<br />
visas, which involves an extra precondition that<br />
<strong>the</strong> plan-implementation agent’s rule doesn’t have,<br />
for example, that an <strong>of</strong>ficial invitation is required. In<br />
such cases, <strong>the</strong> actions will not be executed and <strong>the</strong><br />
plan will fail. The plan-implementation agent can<br />
<strong>the</strong>n question <strong>the</strong> embassy agent fur<strong>the</strong>r, using <strong>the</strong><br />
justification <strong>of</strong> <strong>the</strong> plan step, to find out exactly which<br />
part <strong>of</strong> <strong>the</strong> <strong>the</strong>ory caused <strong>the</strong> failure and why.<br />
Refinement Techniques<br />
To create rules for specialising a <strong>the</strong>ory or ontology,<br />
we looked first at rules for abstraction; that is, removing<br />
detail from a <strong>the</strong>ory. We inverted <strong>the</strong>se rules to<br />
form anti-abstractions which can thus be used to add<br />
detail to a <strong>the</strong>ory:<br />
Predicate anti-abstraction - A single predicate is<br />
divided into some number <strong>of</strong> predicates<br />
Domain anti-abstraction - Constants and function<br />
symbols are divided up into different cases<br />
Propositional anti-abstraction - The arity <strong>of</strong> a<br />
predicate is increased<br />
Precondition anti-abstraction - Preconditions<br />
are added to rules.<br />
The refinement system will <strong>the</strong>n select <strong>the</strong> appropriate<br />
type <strong>of</strong> refinement from those available to it, using<br />
<strong>the</strong> information gleaned earlier about <strong>the</strong> problem<br />
point. Once <strong>the</strong> refinement has been performed, it replaces<br />
<strong>the</strong> original in <strong>the</strong> <strong>the</strong>ory, which can <strong>the</strong>n be<br />
presented to <strong>the</strong> planner.<br />
Author Information<br />
Fiona McNeill, Alan Bundy, and Marco Schorlemmer<br />
Edinburgh University, UK, fionam@<br />
dai.ed.ac.uk
The PLANET <strong>Newsletter</strong> 29<br />
ARTICLE<br />
MEXAR: Integrated AI Technologies to Support MARS EXPRESS<br />
Mission Planning<br />
Introduction<br />
Space exploration missions require coordination <strong>of</strong> a<br />
significant amount <strong>of</strong> activities. State <strong>of</strong> <strong>the</strong> art intelligent<br />
planning and scheduling (P&S) technology<br />
could potentially be <strong>of</strong> great help in supporting such a<br />
coordination. This work aim at showing an example<br />
<strong>of</strong> this technology in a support system for a specific<br />
mission scheduling problem related to <strong>the</strong> ESA program<br />
called MARS-EXPRESS [3].<br />
MARS-EXPRESS is a space probe that will be<br />
launched during 2003 and after six months will be<br />
orbiting around Mars for <strong>the</strong> following two years and<br />
more. During <strong>the</strong> operational phase around Mars a<br />
team <strong>of</strong> people, <strong>the</strong> Mission Planners, will be responsible<br />
for <strong>the</strong> on board operations <strong>of</strong> MARS-EXPRESS.<br />
They receive input from different teams <strong>of</strong> scientists<br />
and cooperate with different specialists for various<br />
specific tasks (e.g., Flight Dynamics (FD) experts).<br />
Any single operation <strong>of</strong> a payload, named POR (Payload<br />
Operation Request), is decided well in advance<br />
through a negotiation phase among <strong>the</strong> different actors<br />
involved in <strong>the</strong> process (e.g., scientists, mission<br />
planners, FD experts).<br />
The result <strong>of</strong> our study is a system called MEXAR that<br />
addresses <strong>the</strong> memory dumping problem in MARS-<br />
EXPRESS. The specific problem that is addressed<br />
is defined as MEX-MDP and is described in detail<br />
in [7]. MEXAR is an interactive support system<br />
that may help mission planners in deciding policies<br />
for downlinking data to Earth during <strong>the</strong> temporal<br />
visibility windows. The tool uses constraint-based<br />
techniques for representing <strong>the</strong> basic problem to be<br />
solved, namely <strong>the</strong> segmentation <strong>of</strong> on-board memory<br />
in data packets during <strong>the</strong> visibility toward Earth.<br />
The paragraph below introduces <strong>the</strong> two solving<br />
algorithms which have been developed: a greedy<br />
Authors: G. Cortellessa, N. Policella, A. Cesta, and A. Oddi<br />
heuristic and a local search procedure [7, contains a<br />
complete explanation <strong>of</strong> <strong>the</strong>se algorithms]. The following<br />
section describes an important aspect <strong>of</strong> this<br />
work, <strong>the</strong> interactive functionalities developed to support<br />
<strong>the</strong> user in his work [1, for more details].<br />
The Packet Sequencing Algorithm<br />
Scheduling problems such as MEX-MDP can be seen<br />
as a special types <strong>of</strong> Constraint Satisfaction Problems<br />
(CSP) [6]. An instance <strong>of</strong> a CSP involves a<br />
set <strong>of</strong> variables ¡£¢¤ ¦¥¤§¨ �©�§������¤§¨ ���� , a domain<br />
���<br />
for each variable and a set <strong>of</strong> ��¡<br />
constraints<br />
, �<br />
����� � �<br />
¥�� ©����������<br />
¢���¥¤§���©�§�������§����¤� s.t. �<br />
which define feasible combinations <strong>of</strong> domain values.<br />
A solution is an assignment <strong>of</strong> domain values<br />
to all variables which is consistent with all imposed<br />
constraints.<br />
The CSP formalization <strong>of</strong> <strong>the</strong> MEX-MDP problem is<br />
based on a partition <strong>of</strong> � ¡<br />
� �<br />
<strong>the</strong> temporal horizon<br />
in a set <strong>of</strong> � ¡�¢¤� ¥ ¡<br />
� ��� §���� � ��� contiguous �<br />
� windows �<br />
����¢¤����¡ � �¤��¥�§ ��������¡<br />
�<br />
§ � ���<br />
¥¨��� ¡<br />
, such ������<br />
��������§<br />
�<br />
� ¡��<br />
that . We<br />
���<br />
��� consider<br />
a set <strong>of</strong> � � decision variables that<br />
¥<br />
represent <strong>the</strong><br />
amount <strong>of</strong> data to be dumped ���<br />
from packet store<br />
in <strong>the</strong> window ��� . The MEX-MDP contains different<br />
kinds <strong>of</strong> constraints: (a) Given <strong>the</strong> characteristics <strong>of</strong><br />
<strong>the</strong> packet stores <strong>the</strong> data must be downlink according<br />
to a FIFO philosophy; (b) <strong>the</strong> amount <strong>of</strong> data for each<br />
packet store does not exceed its capacity; (c) a finite<br />
amount <strong>of</strong> data can be dumped in each transmission<br />
window (finite transmission rate).<br />
All <strong>the</strong> proposed algorithms work over two levels <strong>of</strong><br />
abstraction: (1) <strong>the</strong> planning level, where <strong>the</strong> whole<br />
set <strong>of</strong> decision variables are instantiated taking into<br />
account <strong>the</strong> different constraints; (2) <strong>the</strong> scheduling<br />
level, where a sequence <strong>of</strong> memory dump operations<br />
�
30 The PLANET <strong>Newsletter</strong><br />
is generated over <strong>the</strong> communication links respecting<br />
<strong>the</strong> constraints imposed over all <strong>the</strong> windows � � .<br />
In order to find an optimal solution we choose to realize<br />
a heuristic optimization strategy based on local<br />
search which is able to improve an initial solution<br />
given as an input: Tabu search [4, 5]. The tabu metaheuristic<br />
is founded on <strong>the</strong> notion <strong>of</strong> a move. A move<br />
is a function which transforms one solution into ano<strong>the</strong>r.<br />
For any solution ¡<br />
, a subset <strong>of</strong> moves applied<br />
to ¡<br />
is computed. The result is <strong>the</strong> neighborhood <strong>of</strong><br />
¡<br />
. The algorithm proceeds selecting, at each step,<br />
<strong>the</strong> best solution in <strong>the</strong> neighborhood, with respect to<br />
an objective function, till a fixed number <strong>of</strong> steps are<br />
made without finding better solutions.<br />
Figure 1: MEXAR layout<br />
In MEX-MDP <strong>the</strong> move consists in bringing forward<br />
some data (for example data contained in observations<br />
with <strong>the</strong> smallest volume <strong>of</strong> data) and delaying<br />
o<strong>the</strong>r ones; this should improve in many cases <strong>the</strong> objective<br />
function (mean turn over time).<br />
Mexar Interactive Functionalities<br />
The MEXAR functionalities that are designed for <strong>the</strong><br />
users are summarized in Figure 2. As expected <strong>the</strong><br />
problem solving activity is central in <strong>the</strong> system. This<br />
functionality is guaranteed by <strong>the</strong> automated services<br />
centered on <strong>the</strong> constraint-satisfaction methodology<br />
(CSP) described above.<br />
In <strong>the</strong> figure 2 we show <strong>the</strong> flow <strong>of</strong> control during<br />
<strong>the</strong> use <strong>of</strong> <strong>the</strong> functionalities. It is possible to iden-<br />
http://www.planet-noe.org<br />
tify an activity that aims generically at defining a single<br />
problem. At present it simply consists <strong>of</strong> loading<br />
a problem description from a file, it can be also be<br />
replaced by a more complex incremental functionality<br />
that could be well coupled with <strong>the</strong> CSP modeling<br />
used. The definition <strong>of</strong> a problem is followed by<br />
its solution according to <strong>the</strong> different algorithms produced<br />
in our work. A different functionality allows<br />
to refine <strong>the</strong> current problem. This activity consists<br />
in deleting some <strong>of</strong> <strong>the</strong> Payload Operation Requests<br />
(PORs) from <strong>the</strong> associated timelines. This can be<br />
useful to experiment different loads on <strong>the</strong> resources<br />
in specific intervals <strong>of</strong> <strong>the</strong> solution. This functionality<br />
introduces a cycle among <strong>the</strong>se activities that could<br />
bring <strong>the</strong> user to incrementally refine new MEX-MDP<br />
problems. As shown we group <strong>the</strong>se functionalities<br />
in an interaction layout called Problem Analyzer (see<br />
Figure 2).<br />
Once a problem to solve is defined we can attack it<br />
with different solution methods. Figure 1 shows an<br />
example <strong>of</strong> a solved instance <strong>of</strong> MEX-MDP as it is<br />
presented to <strong>the</strong> user.<br />
Figure 2: MEXAR Interactive Environments<br />
The availability <strong>of</strong> a portfolio <strong>of</strong> problem solving<br />
procedures has suggested <strong>the</strong> idea <strong>of</strong> involving more<br />
deeply <strong>the</strong> user in <strong>the</strong> solution process. This has been<br />
pursued by creating an environment in which it is<br />
possible to save different solutions and, in addition,<br />
<strong>the</strong> user can guide search on how to improve <strong>the</strong> solutions<br />
applying different algorithms. We call this process<br />
solution space exploration. This aspect is strictly<br />
connected to <strong>the</strong> availability <strong>of</strong> evaluation metrics on<br />
<strong>the</strong> solutions as discussed in [2]. The idea behind <strong>the</strong><br />
solution explorer is <strong>the</strong> one that <strong>the</strong> user can generate<br />
an initial solution, save it, try to improve it by local<br />
search, save <strong>the</strong> results, try to improve it by local
The PLANET <strong>Newsletter</strong> 31<br />
search with different tuning parameters and so on. In<br />
this way, it is possible to generate paths in <strong>the</strong> search<br />
space. The user can restore one <strong>of</strong> <strong>the</strong> previous solutions<br />
and try to improve it with a local search with<br />
different parameters, etc. In this way he generates a<br />
tree <strong>of</strong> solutions. This procedure can be repeated for<br />
different starting points generating, in this way, a set<br />
<strong>of</strong> trees. Using at <strong>the</strong> same time <strong>the</strong> evaluation capability<br />
on a single solution and its own experience<br />
he can generate different solution series, all <strong>of</strong> <strong>the</strong>m<br />
saved, and, at <strong>the</strong> end, choose <strong>the</strong> best candidate for<br />
execution.<br />
Bibliography<br />
[1] Cesta, A.; Cortellessa, G.; Oddi, A.; and Policella,<br />
N. Interaction <strong>Service</strong>s for Mission Planning<br />
in MARS EXPRESS. In Proceedings <strong>of</strong> <strong>the</strong><br />
3rd International NASA Workshop on Planning<br />
and Scheduling for Space, 2002.<br />
[2] Cesta, A.; Oddi, A.; Cortellessa, G.; and Policella,<br />
N. Using AI Techniques to Solve MEX-<br />
MDP. Technical Report MEXAR-TR-02-08,<br />
ISTC-CNR - Planning and Scheduling Team,<br />
Rome, Italy, 2002.<br />
[3] ESA. <strong>European</strong> Space Agency, MARS<br />
EXPRESS Web Site. http://sci.esa.int/<br />
marsexpress/, 2002.<br />
[4] Glover, F. Tabu Search – Part I. ORSA Journal<br />
<strong>of</strong> Computing 1:190–206, 1989.<br />
[5] Glover, F. Tabu Search – Part II. ORSA Journal<br />
<strong>of</strong> Computing 2:4–32, 1990.<br />
[6] Montanari, U. <strong>Network</strong>s <strong>of</strong> Constraints: Fundamental<br />
Properties and Applications to Picture<br />
Processing. Information Sciences 7:95–132,<br />
1974<br />
[7] Oddi, A.; Cesta, A.; Policella, N.; and Cortellessa,<br />
G. Scheduling Downlink Operations in<br />
MARS EXPRESS. In Proceedings <strong>of</strong> <strong>the</strong> 3rd<br />
International NASA Workshop on Planning and<br />
Scheduling for Space, 2002.<br />
Author Information<br />
Gabriella Cortellessa ¥ , Nicola Policella © ,<br />
Amedeo Cesta, and Angelo Oddi ISTC-<br />
CNR, Italian National Research Council, Viale Marx<br />
15, I-00137 Rome, Italy, {corte,policella,<br />
cesta,oddi}@ip.rm.cnr.it<br />
1 is also Ph.D. student in Cognitive Psychology at<br />
University <strong>of</strong> Rome “La Sapienza”<br />
2 is also Ph.D. student in Computer Science at University<br />
<strong>of</strong> Rome “La Sapienza”<br />
RDPPlan: an Extension <strong>of</strong> DPPlan for Planning with Interval<br />
Resources<br />
RDPPLan is a model for planning with quantitative<br />
resources. It is based on DPPPlan [1], a planner<br />
which uses a non directional search algorithm on <strong>the</strong><br />
planning graph.<br />
Most models <strong>of</strong> planning with resources, like [3], [4],<br />
[5], [6], [7] and [8], assume that an exact value can<br />
model <strong>the</strong> continuous quantities describing, in <strong>the</strong><br />
real world, a given resource. In o<strong>the</strong>r words <strong>the</strong>se<br />
ARTICLE<br />
Authors: M. Baioletti, A. Milani, and V. Poggioni<br />
models cannot deal with more realistic situations in<br />
which quantities are not known exactly. The RDP-<br />
Plan model allows one to manage domains where preconditions<br />
and effects over quantitative resources can<br />
be specified by intervals <strong>of</strong> values; in addition mixed<br />
logical/quantitative and pure numerical goals can be<br />
specified.<br />
Instead <strong>of</strong> initializing a resource with a unique real
32 The PLANET <strong>Newsletter</strong><br />
¢¡<br />
£<br />
£ � �<br />
¤¥¡§¦ ¡¨¦<br />
value, we allow for <strong>the</strong> specification <strong>of</strong> a real interval<br />
as <strong>the</strong> range <strong>of</strong> <strong>the</strong> initial value <strong>of</strong> <strong>the</strong> resource<br />
. The planner operates in an underspecified domain<br />
in which <strong>the</strong> value <strong>of</strong> some resource is not exactly<br />
known, but it is bound to be in an interval. Let us suppose<br />
that we do not know exactly how much gasoline<br />
is in <strong>the</strong> tank <strong>of</strong> our car: we just know for sure that <strong>the</strong><br />
real amount is between 5 and 10 liters. Similarly it is<br />
possible to have underspecified effects <strong>of</strong> some operators:<br />
<strong>the</strong> value which is added (subtracted), multiplied<br />
(divided) or assigned to <strong>the</strong> current value <strong>of</strong> a<br />
resource is not exactly known, but only a lower and<br />
an upper bound is specified. Let us imagine that we<br />
do not know which is <strong>the</strong> exact consumption <strong>of</strong> <strong>the</strong><br />
car in <strong>the</strong> previous example: all we know is that <strong>the</strong><br />
car travels from 10 to 15 kilometers per liter.<br />
We define two intervals associated to each resource<br />
and each time level : and that respectivly<br />
represent <strong>the</strong> Realized Interval and <strong>the</strong> Desired Interval<br />
<strong>of</strong> resource £ at time-step � ¤©¡§¦<br />
¡<br />
. is initialized with<br />
and it is updated, at each step, by <strong>the</strong> effects on<br />
<strong>the</strong> resource £ <strong>of</strong> <strong>the</strong> actions already inserted in <strong>the</strong><br />
plan; �<br />
¡¨¦ , instead, contains all <strong>the</strong> admissible values<br />
for <strong>the</strong> resource £<br />
that allow <strong>the</strong> execution <strong>of</strong> all <strong>the</strong><br />
actions selected at time level � .<br />
In this model <strong>the</strong> necessary and sufficient condition<br />
for a plan to be a solution <strong>of</strong> a given planning problem<br />
is that for each resource £<br />
and time level � �<br />
¡¨¦ ¤�¡§¦��<br />
¥ � ©<br />
�<br />
§��<br />
�<br />
§������<br />
���<br />
§�� �<br />
������� �<br />
¥ � �<br />
,<br />
. In o<strong>the</strong>r terms, a solution plan must<br />
solve every possible problem that is allowed by <strong>the</strong><br />
constraints specified in <strong>the</strong> initial state and in <strong>the</strong> effects<br />
description.<br />
We use <strong>the</strong> same concept <strong>of</strong> simultaneous executability<br />
as expressed in [2] and in [6]. We say that<br />
<strong>the</strong> actions are simultaneously executable<br />
if, for every permutation <strong>of</strong> , each<br />
action is executable in <strong>the</strong> order defined by <strong>the</strong> permutation<br />
and <strong>the</strong> effect over resources is always <strong>the</strong><br />
same. As a straightforward consequence, an assignment<br />
on resource £<br />
is not simultaneously executable<br />
with any action changing £ �<br />
correct method that can allow us to calculate <strong>the</strong> Desired<br />
Interval for <strong>the</strong> simultaneity execution <strong>of</strong> actions<br />
that have preconditions and/or effects on <strong>the</strong><br />
same resource<br />
and an additive operator<br />
(increase, decrease) is not simultaneously executable<br />
with any multiplicative operator (scale–up,<br />
scale–down). Moreover we have defined a provably<br />
£ ���<br />
¡§¦ ¤�¡§¦<br />
.<br />
In order to achieve <strong>the</strong> goals over resources, we<br />
have defined strategies to solve ”pure numerical problems“,<br />
i.e. with goals only on resources. Such strategies<br />
are combined with <strong>the</strong> ones for solving logical<br />
goals, by evaluating <strong>the</strong> difficulties <strong>of</strong> resources and<br />
logical goals and selecting <strong>the</strong> most difficult goal to<br />
solve.<br />
When we work with resources as intervals, we have<br />
to handle with real intervals whose width in general<br />
grows, except when an assignment is performed.<br />
Moreover note that if <strong>the</strong> width <strong>of</strong> <strong>the</strong> Realized Interval<br />
is large, it is more difficult that <strong>the</strong> solution<br />
condition holds.<br />
The action choice criterium takes into account <strong>the</strong><br />
distance between <strong>the</strong> corresponding Desired and Realized<br />
intervals and <strong>the</strong>ir width. In particular we<br />
have defined two preference functions, one for interval<br />
widths, and one for distance between intervals,<br />
and <strong>the</strong> algorithm chooses <strong>the</strong> action that maximizes<br />
a linear combination <strong>of</strong> <strong>the</strong>se functions.<br />
Investigations and experiments are planned in order<br />
to develop more accurate heuristics and strategies<br />
which take resources into account. Moreover, in order<br />
to provide a meaningful evaluation, it will also<br />
be required <strong>the</strong> development <strong>of</strong> a set <strong>of</strong> significant<br />
benchmarks for planning domains with interval resources.<br />
Finally it is worth investigating fur<strong>the</strong>r extensions<br />
to <strong>the</strong> resource model more accurate with respect<br />
to <strong>the</strong> uncertainty in <strong>the</strong> real world e.g. intervals<br />
with given probability distribution over resource values<br />
and fuzzy quantities.<br />
Bibliography<br />
[1] Baioletti M., Marcugini S. and Milani A. DP-<br />
Plan: an Algorithm for Fast Solution Extraction<br />
from a Planning Graph, In Proc. <strong>of</strong> AIPS-2000<br />
[2] Fox M. and Long D., PDDL 2.1: An Extension<br />
to PDDL for Expressing Temporal Planning Domains,<br />
2002<br />
http://www.planet-noe.org
The PLANET <strong>Newsletter</strong> 33<br />
[3] Koehler J. Planning under resource constraints.<br />
In Proc. <strong>of</strong> ECAI-1998<br />
[4] Chien S. et al., ASPEN Automated Planning<br />
and Scheduling for Space Mission Operation, In<br />
Proc. <strong>of</strong> SpaceOps 2000<br />
[5] Wolfman S. and Weld D., The LPSAT Engine and<br />
its Application to Resource Planning, In Proc. <strong>of</strong><br />
IJCAI 1995<br />
[6] Rintanen J. and Jungholt H. Numeric State Variables<br />
in Constraint-Based Planning, In Proc. <strong>of</strong><br />
ECP-1999<br />
[7] Currie K. and Tate A., O-Plan: The open planning<br />
architecture, In Artificial Intelligence 52,<br />
49-86, 1991<br />
[8] Laborie P. and Ghallab M. Planning with<br />
sharable resource constraints, In Proc. <strong>of</strong> IJCAI<br />
1995<br />
Author Information<br />
Marco Baioletti Dip. Met. Quantitativi, Università<br />
di Siena, Italy<br />
Alfredo Milani Dip. Mat. e Informatica, Università<br />
di Perugia, Italy<br />
Valentina Poggioni Dip. Informatica e Aut.,<br />
Università di Roma3, Italy<br />
ARTICLE<br />
Extending Operator Induction to Provide Full Method Sets for<br />
Hierarchical Planning Domains<br />
Abstract Modelling a world for efficient HTN planning<br />
by hand is a lengthy and time consuming process.<br />
Our knowledge acquisition tool, GIPO (Graphical<br />
User Interface for Planning with Objects), <strong>of</strong>fers<br />
GUI abstraction allowing <strong>the</strong> user to concentrate on<br />
<strong>the</strong> model ra<strong>the</strong>r than a language syntax. Using GIPO<br />
we aim to provide automatic operator induction for<br />
planning domains with an hierarchical sort structure<br />
so that complex hierarchical operators can be constructed.<br />
At present we can induce operators for flat domains.<br />
Using GIPO we interactively construct a sequence <strong>of</strong><br />
actions to arrive at some goal state. Then we input<br />
a partial domain (with no operators) and using opmaker,<br />
<strong>the</strong> induction tool in GIPO, we obtain a set<br />
<strong>of</strong> operators to complete <strong>the</strong> constructed task. These<br />
operators may not be accurate at this stage.<br />
We can also induce operator sequences to construct<br />
low level methods (hierarchical operators for HTN<br />
domains). The name method implies that <strong>the</strong>se are<br />
different recipies for achieving similar or related<br />
tasks and as such <strong>the</strong>y <strong>of</strong>ten repeat actions or ac-<br />
Author: N.E. Richardson<br />
tion sequences. Issues arising from using repetition<br />
in <strong>the</strong> sequences are that <strong>the</strong> same operators are induced<br />
more than once and, as it is desirable to have<br />
only one operator per action, we can use repetition to<br />
generalise <strong>the</strong> operator. The debate is <strong>the</strong>n to find a<br />
way <strong>of</strong> revising <strong>the</strong> operator each time a new version<br />
is induced or employ a learning and revision process.<br />
We propose two systems to compare induced operator<br />
specifications and generalise those descriptions.<br />
Comparing operators will give us a category <strong>of</strong> differences.<br />
Some differences are allowable in <strong>the</strong> present<br />
system. For example an operator can be induced<br />
without a conditional clause and if <strong>the</strong> same operator<br />
is used later in <strong>the</strong> sequence with a condition <strong>the</strong>n<br />
<strong>the</strong> conditional clause is added to <strong>the</strong> original operator.<br />
We would want to be able to merge operators<br />
that have different names but are o<strong>the</strong>rwise identical.<br />
Some operators will be more general than o<strong>the</strong>rs but<br />
we recognise that over-generalisation can mean that<br />
<strong>the</strong> operator is not sufficiently expressive.<br />
Operators describe objects’ transitions from one set<br />
<strong>of</strong> substates prior to <strong>the</strong> action represented, to ano<strong>the</strong>r
34 The PLANET <strong>Newsletter</strong><br />
set <strong>of</strong> substates after <strong>the</strong> action. In generalising an operator<br />
<strong>the</strong> aim is to make it applicable to a wider set<br />
<strong>of</strong> objects but this can be overdone. A difficulty in<br />
generalising operators is to limit extent that <strong>the</strong> operator’s<br />
expressiveness is compromised.<br />
Generalising operators and revising <strong>the</strong> operators in<br />
<strong>the</strong> system as new ones are generated will allow us<br />
to constuct <strong>the</strong> higher levels <strong>of</strong> <strong>the</strong> method hierarchy<br />
which are built from o<strong>the</strong>r methods as well as primitive<br />
operators. When we have <strong>the</strong>se systems in place<br />
The SimPlanner<br />
Introduction<br />
Off-line planning generates a complete plan before<br />
any action starts its execution [3]. This forces to<br />
make some assumptions that are not possible in real<br />
environments like, for example, that actions are uninterruptable,<br />
that <strong>the</strong>ir effects are deterministic, that<br />
<strong>the</strong> planner has complete knowledge <strong>of</strong> <strong>the</strong> world or<br />
that <strong>the</strong> world only changes through <strong>the</strong> execution<br />
<strong>of</strong> actions. On <strong>the</strong> o<strong>the</strong>r hand, on-line planning allows<br />
to start execution while <strong>the</strong> planner continues<br />
working in order to improve <strong>the</strong> overall planning and<br />
execution time. Nowadays <strong>the</strong>re are only some few<br />
approaches for planning in dynamic environments<br />
and/or with incomplete information [2]:<br />
Conditional planning: this approach tries to consider<br />
<strong>the</strong> possible contingencies that can occur in<br />
<strong>the</strong> world.<br />
Parallel planning and execution: this approach<br />
separates <strong>the</strong> planning process from <strong>the</strong> execution.<br />
Interleaving planning and execution: this approach<br />
allows quick and effective responses to<br />
changes in <strong>the</strong> environment.<br />
SimPlanner is an integrated tool for planning and execution,<br />
and it is based on this latter approach. This<br />
tool is thought to be used in real environments such<br />
http://www.planet-noe.org<br />
it will be possible to have full method induction in<br />
GIPO for hierarchical domains.<br />
Author Information<br />
N.E. Richardson School <strong>of</strong> Computing and Engineering<br />
The University <strong>of</strong> Huddersfield, Huddersfield<br />
HD1 3DH, UK, n.e.richardson@hud.<br />
ac.uk<br />
ARTICLE<br />
Authors: O. Sapena and E. Onaindía<br />
as <strong>the</strong> intelligent control <strong>of</strong> robots. However, it has<br />
initially been implemented as a simulator in order to<br />
check its behavior without having to integrate it in<br />
different several domains. SimPlanner is made up <strong>of</strong><br />
three components: an on-line planner, a monitoring<br />
module and a replanner.<br />
The on-line planner<br />
The on-line planner is responsible for generating, in<br />
an incremental way, a plan to achieve <strong>the</strong> goals. As<br />
soon as <strong>the</strong> planner calculates <strong>the</strong> first action, <strong>the</strong> plan<br />
can start its execution. From this moment on, <strong>the</strong><br />
planning and execution processes keep on working in<br />
parallel while no unexpected event is detected; o<strong>the</strong>rwise,<br />
<strong>the</strong> execution must wait for <strong>the</strong> planner to make<br />
<strong>the</strong> necessary modifications in <strong>the</strong> plan.<br />
The planner is based on a depth-first search, with no<br />
provision for backup. The planning decisions (inferred<br />
actions) are consequently irrevocable. The<br />
planning algorithm uses heuristic functions to compute<br />
an approximate plan (AP) for each goal independently.<br />
Then, a conflict-checking mechanism detects<br />
conflicts among actions in <strong>the</strong> approximate plans and<br />
selects <strong>the</strong> action from <strong>the</strong> AP that minimizes <strong>the</strong><br />
number <strong>of</strong> conflicts. The selected action is inserted<br />
at <strong>the</strong> end <strong>of</strong> <strong>the</strong> plan, and <strong>the</strong> current state is updated<br />
through its execution. This algorithm is iteratively
The PLANET <strong>Newsletter</strong> 35<br />
executed until all top-level goals are achieved [4].<br />
The monitoring module<br />
Monitoring is <strong>the</strong> process <strong>of</strong> observing <strong>the</strong> world and<br />
trying to find discrepancies between <strong>the</strong> physical reality<br />
and <strong>the</strong> beliefs <strong>of</strong> <strong>the</strong> planner [1]. Basically, <strong>the</strong>re<br />
exists two types <strong>of</strong> plan execution monitoring [2]: action<br />
monitoring checks <strong>the</strong> validity <strong>of</strong> <strong>the</strong> action preconditions<br />
before it starts its execution and also that<br />
its effects have taken place as expected. The environment<br />
monitoring tries to capture information from <strong>the</strong><br />
external world that conditions <strong>the</strong> remaining planning<br />
process. Monitoring is, <strong>the</strong>refore, domain-dependent.<br />
Since SimPlanner is being used at <strong>the</strong> moment as a<br />
simulator, this information is input to <strong>the</strong> system by<br />
<strong>the</strong> user. The user is who decides what information<br />
<strong>the</strong> robot receives and which unexpected events that<br />
occur in <strong>the</strong> world are communicated.<br />
The replanner<br />
When an unexpected event is detected, <strong>the</strong> calculated<br />
plan is checked in order to assure it is still valid [1]. If<br />
this is <strong>the</strong> case, <strong>the</strong> execution simply continues. O<strong>the</strong>rwise<br />
<strong>the</strong> replanning module is invoked. The replanner<br />
tries to reuse as much <strong>of</strong> <strong>the</strong> calculated plan as<br />
possible without losing <strong>the</strong> quality <strong>of</strong> <strong>the</strong> final plan.<br />
It uses a heuristic function to find out which is <strong>the</strong><br />
best reachable state through <strong>the</strong> actions in <strong>the</strong> original<br />
plan. If <strong>the</strong>re are many reusable actions, <strong>the</strong> planning<br />
process will save a lot <strong>of</strong> computation time. In<br />
<strong>the</strong> worst case, a new plan will be computed from<br />
scratch. The replanner overhead is very small so it<br />
is worth trying to reuse <strong>the</strong> plan ra<strong>the</strong>r than planning<br />
from scratch [5].<br />
Bibliography<br />
[1] G. De Giacomo and R. Reiter, ‘Execution monitoring<br />
<strong>of</strong> high-level robot programs’, Principles<br />
<strong>of</strong> Knowledge Representation and Reasoning,<br />
453–465, (1998).<br />
[2] K.Z. Haigh and M. Veloso, ‘Interleaving planning<br />
and robot execution for asynchronous user<br />
requests’, Papers from <strong>the</strong> AAAI Spring Symposium,<br />
35–44, (1996).<br />
[3] M.E. Pollack and J.F. Horty, ‘There’s more to<br />
life than making plans: Plan management in dynamic,<br />
multi-agent environments’, AI Magazine,<br />
20(4), 71–84, (1999).<br />
[4] O Sapena and E. Onaindia, ‘Domainindependent<br />
on-line planning for STRIPS<br />
domains’, Proceedings <strong>of</strong> Iberamia-02, LNAI,<br />
2527, 825–834, (2002).<br />
[5] O Sapena and E. Onaindia, ‘Execution, monitoring<br />
and replanning in dynamic environments’,<br />
Workshop on On-Line Planning and Scheduling,<br />
AIPS-02, (2002).<br />
Author Information<br />
Oscar Sapena, Eva Onaindía Departamento<br />
de Sistemas Informáticos y Computación, Universidad<br />
Politécnica de Valencia, Spain, {osapena,<br />
onaindia}@dsic.upv.es
36 The PLANET <strong>Newsletter</strong><br />
Answer Set Planning with DLV<br />
Introduction<br />
The knowledge based planning system DLV implements<br />
answer set planning on top <strong>of</strong> <strong>the</strong> DLV system<br />
[1]. It is developed at TU Wien and supports <strong>the</strong><br />
declarative language ¥<br />
[3] and its extension ¥¢¡ [4].<br />
The language ¥£¡<br />
is syntactically similar to <strong>the</strong> action<br />
language £<br />
¡ ¥<br />
[6], but semantically closer to answer set<br />
programming (by including default negation, for example).tures:<br />
<strong>of</strong>fers <strong>the</strong> following distinguishing fea-<br />
- Handling <strong>of</strong> incomplete knowledge: for a fluent ¤ ,<br />
nei<strong>the</strong>r ¤ nor its opposite ¥¦¤ need to be known in<br />
any state.<br />
- Nondeterministic effects: actions may have multiple<br />
possible outcomes.<br />
- Optimistic and secure (conformant) planning:<br />
construction <strong>of</strong> a “credulous” plan or a “sceptical”<br />
plan, which works in all cases.<br />
- Parallel actions: More than one action may be executed<br />
simultaneously.<br />
- Optimal cost planning: In ¥ ¡ , one can assign an<br />
arbitrary cost function to each action, where <strong>the</strong><br />
total costs <strong>of</strong> <strong>the</strong> plan are minimized.<br />
An operational prototype <strong>of</strong> DLV as well as sample<br />
encodings <strong>of</strong> planning domains in <strong>the</strong> system are<br />
available at http://www.dlvsystem.com/K/.<br />
©<br />
input<br />
Background<br />
Knowledge<br />
Control Flow<br />
Data Flow<br />
http://www.planet-noe.org<br />
© Parser<br />
Datalog Parser<br />
Underlying Concepts<br />
ARTICLE<br />
Author: A. Polleres<br />
¥ ¡ ¡ ¥<br />
at(§<br />
at(� travel(��§¨§<br />
Action language In transitions are<br />
described in a declarative way by means <strong>of</strong><br />
causation rules, e.g. caused ) after<br />
), ). Fur<strong>the</strong>rmore, <strong>the</strong> language<br />
<strong>of</strong>fers constructs to express executability and<br />
non-executability <strong>of</strong> actions, ramifications, nondeterminism<br />
and allows to assign costs to actions.<br />
Planning problems in this language are transformed<br />
to a logic program which is <strong>the</strong>n evaluated under <strong>the</strong><br />
answer set semantics.<br />
Answer Set Programming Answer Set programs<br />
are logic programs in a syntax similar to Prolog<br />
which allow for negation as failure in rule bodies<br />
and disjunction in rule heads. In our approach,<br />
<strong>the</strong> minimal models <strong>of</strong> <strong>the</strong>se programs under <strong>the</strong> so<br />
called Answer Set Semantics [5] correspond oneto-one<br />
to <strong>the</strong> plans <strong>of</strong> <strong>the</strong> resp. planning problem<br />
specified in ¥ ¡ . This view <strong>of</strong> models representing<br />
plans can be partly compared to planning using SAT-<br />
Solvers.<br />
System architecture<br />
The architecture <strong>of</strong> DLV is outlined in Figure 1. The<br />
input <strong>of</strong> <strong>the</strong> system consists <strong>of</strong> domain descriptions<br />
(DLV files) and optional static background knowledge<br />
specified by a logic program.<br />
Plan Generator<br />
DLV © Core<br />
DLV Core<br />
Figure 1: DLV © System Architecture<br />
Plan Checker<br />
Controller<br />
Plan Printer
The PLANET <strong>Newsletter</strong> 37<br />
The Controller first invokes <strong>the</strong> Plan Generator,<br />
which translates <strong>the</strong> planning problem at hand into<br />
a suitable program in <strong>the</strong> core language <strong>of</strong> DLV (disjunctive<br />
logic programs under <strong>the</strong> answer set semantics).<br />
Then, <strong>the</strong> DLV kernel is invoked to solve <strong>the</strong><br />
corresponding problem. The resulting answer sets (if<br />
any exist) are fed back to <strong>the</strong> Controller, which extracts<br />
<strong>the</strong> solutions to <strong>the</strong> original problem from <strong>the</strong>se<br />
answer sets and transforms <strong>the</strong>m back to <strong>the</strong> original<br />
planning domain.<br />
If <strong>the</strong> user specified that secure/conformant planning<br />
should be performed, <strong>the</strong> Controller <strong>the</strong>n invokes <strong>the</strong><br />
Plan Checker which verifies by ano<strong>the</strong>r evaluation <strong>of</strong><br />
a logic program whe<strong>the</strong>r this plan is in fact also secure/conformant.<br />
Finally, <strong>the</strong> solutions found by <strong>the</strong> Generator (and optionally<br />
verified by <strong>the</strong> Checker) are translated back<br />
into user output and printed.<br />
Details about <strong>the</strong> transformations mentioned above<br />
can be found in [2]. Performance and experimental<br />
results are reported in [2, 4].<br />
Bibliography<br />
[1] T. Eiter, W. Faber, N. Leone, and G. Pfeifer.<br />
Declarative Problem-Solving Using <strong>the</strong> DLV<br />
System. Logic-Based Artificial Intelligence, pp.<br />
79–103. Kluwer, 2000.<br />
[2] T. Eiter, W. Faber, N. Leone, G. Pfeifer, and<br />
A. Polleres. A Logic Programming Approach to<br />
Knowledge-State Planning, II: <strong>the</strong> DLV System.<br />
Technical Report, December 2001. To appear in<br />
Artificial Intelligence.<br />
[3] T. Eiter, W. Faber, N. Leone, G. Pfeifer, and<br />
A. Polleres. A Logic Programming Approach to<br />
Knowledge-State Planning: Semantics and Complexity.<br />
Technical Report, December 2001. To<br />
appear in ACM Transactions on Computational<br />
Logic.<br />
[4] T. Eiter, W. Faber, N. Leone, G. Pfeifer, and<br />
A. Polleres. Answer Set Planning under Action<br />
Costs. Technical Report, October 2002.<br />
[5] M. Gelfond and V. Lifschitz. Classical Negation<br />
in Logic Programs and Disjunctive Databases.<br />
New Generation Computing, 9:365–385, 1991.<br />
[6] E. Giunchiglia and V. Lifschitz. An Action Language<br />
Based on Causal Explanation: Preliminary<br />
Report. In AAAI ’98, pp. 623–630, 1998.<br />
Author Information<br />
Axel Polleres Institut für Informationssysteme,<br />
TU Wien, A-1040 Wien, Austria, axel@kr.<br />
tuwien.ac.at, pfeifer@dbai.tuwien.<br />
ac.at<br />
This work was supported by FWF (Austrian Science<br />
Funds) under <strong>the</strong> projects P14781 and Z29-INF.<br />
Fur<strong>the</strong>rmore, <strong>the</strong> author thanks <strong>the</strong> ÖGAI (Austrian<br />
Association for Artificial Intelligence) and <strong>the</strong> OCG<br />
(Austrian Computer Society) for sponsoring travel<br />
costs to <strong>the</strong> PLANET Summer School 2002
38 The PLANET <strong>Newsletter</strong><br />
Planform: An Open Environment for Building Planners<br />
Project Overview<br />
The PLANFORM project aimed to develop a high level<br />
research platform for <strong>the</strong> systematic construction <strong>of</strong><br />
planner domain models and automatically configured<br />
planning algorithms.<br />
Our objectives were, briefly:<br />
To assemble tools within an open environment for<br />
<strong>the</strong> acquisition and modelling <strong>of</strong> planning domain<br />
models;<br />
To create languages for modelling <strong>of</strong> planning domains<br />
and to specify characteristics <strong>of</strong> planners<br />
leading to <strong>the</strong> configuration <strong>of</strong> purpose-built planning<br />
engines;<br />
To create a tool which syn<strong>the</strong>sises a domain model<br />
and a planning engine into a planning application;<br />
To evaluate <strong>the</strong> approach using realistic problem<br />
domains.<br />
During <strong>the</strong> project, <strong>the</strong> emphasis changed slightly so<br />
that some <strong>of</strong> objectives developed in a different direc-<br />
Application<br />
Domain<br />
Knowledge<br />
Acquisition<br />
Tools<br />
http://www.planet-noe.org<br />
Internal<br />
Representation<br />
<strong>of</strong><br />
Domain<br />
Static and Dynamic Analysis<br />
ARTICLE<br />
Authors: T.L. McCluskey, M. Fox, and R. Aylett<br />
tion. Ra<strong>the</strong>r than abstractly defining planning algorithms,<br />
we decided to create a library <strong>of</strong> algorithms<br />
and use domain analysis technology to design and<br />
configure a planning application. We perceived <strong>the</strong><br />
knowledge acquisition bottleneck to be a significant<br />
problem for AI Planning and concentrated more resource<br />
than planned on this area.<br />
We have made substantial progress towards our objectives.<br />
We have developed an environment enabling<br />
<strong>the</strong> acquisitition and modelling <strong>of</strong> planning applications<br />
and <strong>the</strong> configuration <strong>of</strong> planning engines suited<br />
to those applications. Although time limitations prevented<br />
us from completing all aspects <strong>of</strong> <strong>the</strong> final integration<br />
and evaluation we achieved our main objectives<br />
and laid a strong foundation for development <strong>of</strong><br />
<strong>the</strong> project.<br />
The project consisted <strong>of</strong> a model engineering phase<br />
and a planner engineering phase. Huddersfield<br />
and Salford collaborated closely on modelling and<br />
knowledge acquisition issues, whilst Huddersfield<br />
and Durham collaborated on <strong>the</strong> development <strong>of</strong> parts<br />
<strong>of</strong> <strong>the</strong> domain modelling tools. Durham worked on<br />
<strong>the</strong> planner engineering phase <strong>of</strong> <strong>the</strong> project.<br />
Configuration<br />
Tool<br />
General<br />
Planners<br />
Generic<br />
Heuristics<br />
Model Engineering Phase Planner Engieenring Phase<br />
Figure 1: Architectural Breakdown <strong>of</strong> Planform<br />
Efficient<br />
Planning<br />
Application
The PLANET <strong>Newsletter</strong> 39<br />
Research at <strong>the</strong> University <strong>of</strong> Huddersfield<br />
The Huddersfield contribution emphasised tools creation<br />
and integration aspects <strong>of</strong> <strong>the</strong> project. Tools for<br />
acquisition, validation and domain modelling were<br />
developed, in collaboration with colleagues at both<br />
Salford and Durham. As part <strong>of</strong> this effort <strong>the</strong> Huddersfield<br />
team researched and developed a GUI-based<br />
environment called GIPO 1 . In addition, as <strong>the</strong> initiators<br />
<strong>of</strong> PLANFORM, Huddersfield performed a central<br />
administrative function, producing <strong>the</strong> main external<br />
website and also an internal website with additional<br />
resources such as Project Meeting minutes.<br />
NB: in <strong>the</strong> text below ‘<strong>the</strong> PLANFORM website’ refers<br />
to http://scom.hud.ac.uk/planform.<br />
The first phase <strong>of</strong> <strong>the</strong> Huddersfield contribution was<br />
concerned with applications encoding and language<br />
development. During this phase three application domains<br />
were used to explore <strong>the</strong> adequacy <strong>of</strong> <strong>the</strong> modelling<br />
languages ��¢¡ and ��¢¡¤£ [20, 19]. ��¥¡¦£<br />
is a hierarchical version <strong>of</strong> ��¢¡ enabling <strong>the</strong> modelling<br />
<strong>of</strong> planning domains as hierarchical task decompositions.<br />
Salford and Huddersfield collaborated<br />
on this effort with Salford working on <strong>the</strong> modelling<br />
<strong>of</strong> some existing domains using <strong>the</strong> two languages.<br />
Huddersfield tackled <strong>the</strong> problem <strong>of</strong> encoding <strong>the</strong> aircraft<br />
landing scheduling problem supplied by Mark<br />
Watson <strong>of</strong> <strong>the</strong> National Air Traffic <strong>Service</strong>s. The constraints<br />
(e.g. separation times for each type <strong>of</strong> aircraft)<br />
were encoded in ��¢¡ [18] and ��¢¡§£ . ��¢¡¦£<br />
was found adequate for all three domains, although<br />
<strong>the</strong> encoding did demonstrate some required changes,<br />
and overall <strong>the</strong> pressing need for tool support.<br />
The standard languages for communicating planning<br />
domain descriptions are <strong>the</strong> PDDL variants [1]. In order<br />
to be able to experiment with our domain models<br />
using exisiting planning technology we <strong>the</strong>refore<br />
created tools to map between ��¢¡ and PDDL.<br />
These tools continued to be developed throughout <strong>the</strong><br />
project allowing planners, and o<strong>the</strong>r tools that receive<br />
input in PDDL form, to be integrated into our environ-<br />
1 http://helios.hud.ac.uk/planform/gipo<br />
ment. There are some interesting technical aspects <strong>of</strong><br />
<strong>the</strong> mapping discussed in [27].<br />
Language manuals for ��¢¡ and ��¥¡¤£ were maintained<br />
during <strong>the</strong> project [12] and an online help facility<br />
was constructed (see <strong>the</strong> PLANFORM website).<br />
Drawing on previous development work (e.g. [19]),<br />
we assembled tools that automate <strong>the</strong> syntactic and<br />
semantic analysis <strong>of</strong> ��¥¡ domain models. Analyses<br />
include ensuring that invariant properties <strong>of</strong> <strong>the</strong><br />
model are maintained and that syntactic rules are observed.<br />
The second phase <strong>of</strong> <strong>the</strong> project was concerned with<br />
development <strong>of</strong> <strong>the</strong> GUI and tools environment. The<br />
focus was on building and integrating knowledge acquisition<br />
and modelling tools for AI planning into an<br />
open environment. The GUI and some <strong>of</strong> <strong>the</strong> tools<br />
described above were built in Java. O<strong>the</strong>rs were implemented<br />
in Prolog and it was necessary to integrate<br />
<strong>the</strong>se via Sicstus Prolog’s JASPER interface.<br />
The JavaCup parser generator method was used to<br />
represent <strong>the</strong> syntax rules <strong>of</strong> ��¢¡ and to generate<br />
a parser for <strong>the</strong> language. This formed <strong>the</strong> input tool<br />
in GIPO (Graphical Interface for Planning with Objects).<br />
The knowledge acquisition part <strong>of</strong> <strong>the</strong> tool<br />
was structured using <strong>the</strong> method outlined in earlier<br />
work [20]. The methods direct <strong>the</strong> user to define objects,<br />
object sorts, relations and properties, classes<br />
and constraints on object situations, problems in <strong>the</strong><br />
form <strong>of</strong> task specifications, and finally operators built<br />
from <strong>the</strong>se components. The design <strong>of</strong> <strong>the</strong> interface<br />
was based on <strong>the</strong> need to minimise <strong>the</strong> use <strong>of</strong> syntax,<br />
and use object ra<strong>the</strong>r than predicate centred ideas.<br />
Once users have entered all parts <strong>of</strong> a domain model<br />
<strong>the</strong>y can utilise modelling tools to remove bugs and<br />
experiment with <strong>the</strong> encoding. We created and<br />
adapted <strong>the</strong> following tools.<br />
plan stepper: This allows <strong>the</strong> user to pick action<br />
schemas and apply <strong>the</strong>m to a state, until a desired<br />
goal is reached. It is useful for identifying errors<br />
in operators and operators sets.<br />
internal planning engines: this allows our own<br />
in-house planning engines to be connected up to
40 The PLANET <strong>Newsletter</strong><br />
GIPO. Sample tasks can be executed and <strong>the</strong> resulting<br />
solutions displayed.<br />
interface for external planning engines: This allows<br />
external planning engines to be ‘bolted’ into<br />
<strong>the</strong> environment. The planner needs to be able to<br />
input domain models in PDDL (from GIPO), and<br />
output solution in a prescribed format. Again, <strong>the</strong><br />
resulting solutions are displayed through GIPO.<br />
a random task generator: This inputs <strong>the</strong> current<br />
domain model and randomly generates tasks to be<br />
used with a planner.<br />
an animator: After a domain model has been entered,<br />
and <strong>the</strong> planning engine has solved a task<br />
within that model. <strong>the</strong> animator can be used to<br />
track <strong>the</strong> transitions <strong>of</strong> each <strong>of</strong> <strong>the</strong> objects which<br />
started in <strong>the</strong> initial state.<br />
In <strong>the</strong> third phase, integration and evaluation, <strong>the</strong><br />
tools outlined above were integrated into GIPO [28].<br />
The s<strong>of</strong>tware was released and demonstrated at<br />
ECP’01, and again at AIPS’02. It is available on<br />
Unix, Linux and Windows platforms from <strong>the</strong> PLAN-<br />
FORM website. As an initial indication <strong>of</strong> GIPO’s<br />
impact, <strong>the</strong> Huddersfield website recorded 147 external<br />
downloads <strong>of</strong> <strong>the</strong> system in <strong>the</strong> period November<br />
2001 - March 2002.<br />
The environment has been tested using a range <strong>of</strong><br />
common domains (details are in <strong>the</strong> resource section<br />
<strong>of</strong> <strong>the</strong> website). Fur<strong>the</strong>r, GIPO was used as a teaching<br />
tool in a second year introductory course in Artficial<br />
Intelligence. GIPO alleviates many user interface<br />
problems by adopting an object modelling approach<br />
which seems natural to non-expert users. To ameliorate<br />
<strong>the</strong> use <strong>of</strong> GIPO by non specialists <strong>the</strong> following<br />
issues were explored:<br />
The use <strong>of</strong> an inductive approach to capturing operators<br />
2 . The opmaker tool was created which<br />
outputs a set <strong>of</strong> operators given a partial domain<br />
model and an example solution sequence [22, 21].<br />
The use <strong>of</strong> generic types to suggest planning design<br />
patterns. These ideas were developed as<br />
2 This work was primarily work undertaken in conjunction with a PhD student<br />
http://www.planet-noe.org<br />
part <strong>of</strong> <strong>the</strong> Durham PLANFORM project, and Huddersfield<br />
is working in collaboration with colleagues<br />
at Durham to integrate design patterns<br />
into GIPO [26].<br />
The final phase <strong>of</strong> our work has been to extend <strong>the</strong> internal<br />
language and <strong>the</strong> surrounding tools from ��¥¡<br />
(version 1) to ��¥¡¦£ (version 2). ��¥¡¦£ extends<br />
��¢¡ in two major ways: HTN operators can be<br />
used, and sort constraints can be put on each level<br />
<strong>of</strong> <strong>the</strong> sort hierarchy meaning that objects <strong>of</strong> a primitive<br />
sort inherit all <strong>the</strong> constraints up <strong>the</strong> hierarchy.<br />
This modelling approach is being tested using a large<br />
‘Translog’ domain imported from a transport logistic<br />
domain constructed by members <strong>of</strong> <strong>the</strong> University <strong>of</strong><br />
Maryland.<br />
The priorities for future development <strong>of</strong> <strong>the</strong> Huddersfield<br />
contribution are:<br />
The development <strong>of</strong> a suite <strong>of</strong> planning design<br />
patterns and <strong>the</strong>ir integration with <strong>the</strong> GIPO tool;<br />
�©¨ ¨<br />
The �¡ £¢¥¤§¦<br />
£<br />
evolution <strong>of</strong> <strong>the</strong> tool into a general<br />
mixed-initiative plan authoring tool;<br />
Integration <strong>of</strong> <strong>the</strong> operator induction techniques<br />
with a plan authoring interface so that operator<br />
specifications can be induced and refined interactively.<br />
The development <strong>of</strong> <strong>the</strong> ��¥¡ representation language<br />
to be on <strong>the</strong> expressive level <strong>of</strong> PDDL2.1<br />
(temporal representations), <strong>the</strong>reby enabling GIPO<br />
to support <strong>the</strong> modelling <strong>of</strong> temporal planning domains.<br />
Research at <strong>the</strong> University <strong>of</strong> Salford<br />
Part <strong>of</strong> <strong>the</strong> set <strong>of</strong> objectives for <strong>the</strong> PLANFORM<br />
project was to make AI planning technology accessible<br />
to non-experts. In pursuit <strong>of</strong> this objective, work<br />
at Salford was based on <strong>the</strong> idea <strong>of</strong> <strong>the</strong> Knowledge<br />
Level [25] and Models <strong>of</strong> Expertise [6] as articulated<br />
over many years in <strong>the</strong> KBS community, in which <strong>the</strong><br />
problem level is modelled separately from <strong>the</strong> design<br />
� �
The PLANET <strong>Newsletter</strong> 41<br />
level. Research in KBS technology has shown that<br />
support for domain experts is feasible if it is based on<br />
<strong>the</strong> generic task concept [7], and much earlier work<br />
has been carried out round <strong>the</strong> generic task <strong>of</strong> diagnosis<br />
[24]. Oddly, little <strong>of</strong> this has been applied to<br />
AI Planning. The basic idea is that a generic task<br />
incorporates a skeletal model at <strong>the</strong> knowledge-level<br />
which can <strong>the</strong>n be used to direct a computer-based<br />
knowledge acquisition process with a domain expert.<br />
Thus, to support domain experts, it was seen as necessary<br />
to build a knowledge-level tool as part <strong>of</strong> <strong>the</strong><br />
PLANFORM environment, incorporating generic task<br />
components for planning, and supporting knowledgelevel<br />
construction <strong>of</strong> <strong>the</strong> planning domain ra<strong>the</strong>r than<br />
forcing <strong>the</strong> use <strong>of</strong> <strong>the</strong> domain design language, ��¥¡ ,<br />
used internally. However such a tool would necessarily<br />
link to <strong>the</strong> ��¢¡ GIPO tools being developed<br />
at Huddersfield (see Figure 1) and thus output<br />
��¥¡ , so it was <strong>the</strong>refore vital to understand how<br />
planning problems would be represented in ��¥¡ . A<br />
strong link with Huddersfield was built through modelling<br />
activity in which <strong>the</strong> objective was to understand<br />
this mapping and derive some constraints on<br />
<strong>the</strong> knowledge-level interface, which was oriented towards<br />
users with no AI planning knowledge but with<br />
expertise in a particular planning domain. A number<br />
<strong>of</strong> domains were modelled including <strong>the</strong> multi-robot<br />
Drumstore from earlier work at Salford.<br />
Planning ontologies were identified as a key foundation<br />
for such a knowledge-level tool, and <strong>the</strong> means<br />
by which <strong>the</strong> skeleton model mentioned above could<br />
be embodied. A survey <strong>of</strong> work in <strong>the</strong> field was carried<br />
out and <strong>the</strong> possibility <strong>of</strong> incorporating an existing<br />
ontology such as CYC( http://www.cyc.com/cyc-<br />
2-1/cover.html.) was investigated but limited time did<br />
not allow its use.<br />
The third component researched as <strong>the</strong> basis for a<br />
knowledge-level tool was <strong>the</strong> requirements <strong>of</strong> <strong>the</strong> domain<br />
expert, and here an accessible domain was formulated<br />
(EVENTUS – arranging a weekend outing for<br />
a visiting researcher) and a knowledge acquisition exercise<br />
was carried out with four people. The exercise<br />
was repeated with <strong>the</strong> robot Drumstore problem.<br />
Data was analysed for concept coverage and for in-<br />
terface design issues, looking at <strong>the</strong> process a human<br />
expert goes through in conceptualising <strong>the</strong> domain.<br />
The details <strong>of</strong> this process are described in [2].<br />
The KA tool was constructed in Java, and in its first<br />
version supported passive model construction by <strong>the</strong><br />
expert with support from a small hand-coded ontology<br />
for <strong>the</strong> domains already investigated. In its second<br />
version, an active question-driven process was<br />
added based on <strong>the</strong> key planning concepts <strong>of</strong> Task,<br />
Agent and Object [5].<br />
The methodology embodies two successive<br />
extraction-refinement processes: protocol to problem<br />
specification; and problem-specification to conceptualisation.<br />
A part <strong>of</strong> <strong>the</strong> KOD (Knowledge Oriented<br />
Design) method [30] was applied to obtain an accurate<br />
process for knowledge acquisition and to build<br />
<strong>the</strong> conceptual model through a set <strong>of</strong> examples and<br />
scenarios.<br />
The output <strong>of</strong> <strong>the</strong> KA-Tool is ��¢¡ , which can <strong>the</strong>n<br />
be loaded into <strong>the</strong> GIPO tool created at Huddersfield.<br />
By <strong>the</strong> formal end <strong>of</strong> <strong>the</strong> project, it was possible to<br />
generate <strong>the</strong> world model in ��¢¡ , and since <strong>the</strong>n it<br />
has become possible to generate planning operators,<br />
seen by most people as a key problem in formulating<br />
a planning domain description. In <strong>the</strong> 26 months <strong>of</strong><br />
<strong>the</strong> project, it was not possible to carry out any extensive<br />
evaluation programme, but it is proposed to carry<br />
on with <strong>the</strong> work for a limited period informally with<br />
this as <strong>the</strong> key task.<br />
The modelling exercise enabled <strong>the</strong> development <strong>of</strong><br />
strong links with Huddersfield, where frequent (almost<br />
weekly) visits were made at some points. A<br />
two-week visit to Durham was also organised to<br />
streng<strong>the</strong>n understanding <strong>of</strong> <strong>the</strong> requirements <strong>of</strong> <strong>the</strong><br />
generative planning back-end.<br />
Fur<strong>the</strong>r development <strong>of</strong> <strong>the</strong> KA-Tool is still being carried<br />
out in house, with <strong>the</strong> generation <strong>of</strong> Planning Operators<br />
now possible. If sufficient resources can be<br />
found, <strong>the</strong> next steps would include:<br />
Incorporation <strong>of</strong> a large ontology, such as CYC,<br />
into <strong>the</strong> KA tool;<br />
Integration <strong>of</strong> Durham’s generic types into this ontology;
42 The PLANET <strong>Newsletter</strong><br />
Full integration <strong>of</strong> <strong>the</strong> KA tool with GIPO.<br />
Two publications were generated jointly with <strong>the</strong><br />
Huddersfield team (Simpson et al 2001a, 2001b)<br />
and two more by <strong>the</strong> Salford team on <strong>the</strong>ir own [4]<br />
and [2]. One fur<strong>the</strong>r paper is under review [3] and a<br />
journal article is in preparation.<br />
Research at <strong>the</strong> University <strong>of</strong> Durham<br />
The Durham contribution to PLANFORM focussed on<br />
<strong>the</strong> development <strong>of</strong> planner configuration technology.<br />
The objective was to develop techniques by which,<br />
given a domain model elicited from a user, planner<br />
components suited to <strong>the</strong> domain could be automatically<br />
identified and configured into a purpose-built<br />
planning system. Our approach has been to maintain<br />
a library <strong>of</strong> planner components, including heuristics,<br />
specialised solution strategies and problem-specific<br />
control rules, and to access <strong>the</strong>se by means <strong>of</strong> patternmatching<br />
techniques once patterns have been identified<br />
in <strong>the</strong> domain model.<br />
The techniques used to identify domain patterns are<br />
based on static domain analysis algorithms developed<br />
at Durham prior to <strong>the</strong> start <strong>of</strong> <strong>the</strong> PLANFORM<br />
project [8]. The objective in <strong>the</strong> project was to extend<br />
<strong>the</strong>se algorithms to enable <strong>the</strong> recognition <strong>of</strong> generic<br />
types and associated patterns <strong>of</strong> behaviour in planning<br />
domains, and to associate <strong>the</strong>se generic patterns<br />
with special purpose solution strategies [14, 13, 16].<br />
Briefly, a generic type is a collection <strong>of</strong> types sharing<br />
some fundamental behaviour. For example, <strong>the</strong><br />
generic type <strong>of</strong> mobility contains all types <strong>of</strong> objects<br />
that are capable <strong>of</strong> movement while <strong>the</strong> generic type<br />
<strong>of</strong> construction contains all types that are capable <strong>of</strong><br />
being combined into compounds (and subsequently<br />
recovered by destruction <strong>of</strong> <strong>the</strong> compound).<br />
When a generic type is present its associated patterns<br />
<strong>of</strong> behaviour are present and <strong>the</strong>se can be used both<br />
to assist a domain designer in refining <strong>the</strong> model and<br />
to suggest appropriate solution approaches.<br />
The analysis techniques developed at Durham can<br />
identify certain key generic types and associated patterns.<br />
For example, <strong>the</strong> pattern associated with mo-<br />
http://www.planet-noe.org<br />
bility comprises <strong>the</strong> mobile types, <strong>the</strong>ir maps and <strong>the</strong><br />
predicate that defines locatedness <strong>of</strong> a mobile object<br />
on its map. A specific problem associated with mobility<br />
is route-planning, and a special purpose solution<br />
strategy suited to this problem can be to exploit<br />
travelling saleman heuristics. We were able to automate<br />
<strong>the</strong> configuration <strong>of</strong> planners with specialised<br />
route-planning capabilities enabling route-planning<br />
sub-problems to be handled using specialised approaches<br />
instead <strong>of</strong> by search [9, 13, 10].<br />
The following s<strong>of</strong>tware was developed at Durham as<br />
part <strong>of</strong> <strong>the</strong> PLANFORM project:<br />
Versions 4 and 5 <strong>of</strong> <strong>the</strong> STAN planning system.<br />
STAN [9, 10] performs generic type analysis and<br />
configures a special purpose planner suited to <strong>the</strong><br />
associated generic patterns;<br />
Extensions to TIM [8, 15] to recognise a range <strong>of</strong><br />
generic types in a domain model;<br />
The TIM API providing access to <strong>the</strong> generic type<br />
analyses performed by <strong>the</strong> TIM system, enabling<br />
<strong>the</strong>ir exploitation by o<strong>the</strong>r planning systems. The<br />
API is being exploited by o<strong>the</strong>r researchers in <strong>the</strong><br />
international community;<br />
The OODL domain modelling language, supporting<br />
<strong>the</strong> construction <strong>of</strong> domains around generic<br />
types, and associated domain modelling tool<br />
DRAUGHTSMAN. These were developed by an<br />
MSc student and contributed to our collaboration<br />
with Huddersfield on <strong>the</strong> GIPO tool.<br />
In <strong>the</strong> scope <strong>of</strong> <strong>the</strong> project a number <strong>of</strong> o<strong>the</strong>r generic<br />
types have been identified (for example, construction,<br />
resource-allocation, portability and o<strong>the</strong>rs) and associated<br />
with specialised solution strategies. The configuration<br />
problem becomes complex when several<br />
generic patterns co-occur and <strong>the</strong>ir solution strategies<br />
must be integrated, and we have not completed<br />
<strong>the</strong> work required to support arbitrarily complex patterns<br />
<strong>of</strong> integration. We have, however, categorised<br />
<strong>the</strong> forms <strong>of</strong> integration that need to be handled and<br />
made progress with configurations based on several<br />
<strong>of</strong> <strong>the</strong>se categories [17, 11, 9].
The PLANET <strong>Newsletter</strong> 43<br />
The library contains stored parametrized solution<br />
strategies (such as travelling salesman solvers, multiprocessor<br />
scheduling heuristics etc) appropriate for<br />
sub-problems that commonly arise in planning domains.<br />
We do not try to guarantee complete coverage<br />
<strong>of</strong> all such sub-problems – <strong>the</strong> configuration system<br />
defaults to search if no generic types or suitable<br />
library components can be found. At present <strong>the</strong> library<br />
contains only one solution strategy per identified<br />
generic type, so extraction <strong>of</strong> a suitable stategy<br />
is simple. In general <strong>the</strong> extraction problem is<br />
more difficult because <strong>the</strong>re may be different forms<br />
in which generic patterns arise and <strong>the</strong>se might need<br />
to be matched in some intelligent way against <strong>the</strong> library.<br />
We have not explored this issue in <strong>the</strong> scope<br />
<strong>of</strong> PLANFORM. A recent extension we have made to<br />
<strong>the</strong> library is <strong>the</strong> addition <strong>of</strong> generalised control rules<br />
which can be selected and instantiated to fit a specific<br />
problem domain [23].<br />
Most recently our work in <strong>the</strong>se areas has considered<br />
<strong>the</strong> use <strong>of</strong> generic patterns as a basis for <strong>the</strong> development<br />
<strong>of</strong> planning design patterns. Using <strong>the</strong>se, a<br />
domain construction tool can prompt <strong>the</strong> user for <strong>the</strong><br />
components <strong>of</strong> generic patterns in a way that makes<br />
it simple for <strong>the</strong> user to enter that information. Initial<br />
work on a tool capable <strong>of</strong> supporting this idea was<br />
done by an MSc student [29] who was temporarily<br />
employed on <strong>the</strong> PLANFORM project at Durham. The<br />
work was continued in collaboration with <strong>the</strong> Huddersfield<br />
site which has focussed on <strong>the</strong> development<br />
<strong>of</strong> tool support [26].<br />
The planner configuration approach has been tested<br />
by entering a hybrid planning system, STAN version<br />
4, into <strong>the</strong> international planning competition<br />
in 2000. STAN 4 can automatically detect mobility<br />
and resource-dependence patterns in planning domains<br />
and can extract route-planning and resourceallocation<br />
strategies from its library. Selected strategies<br />
are integrated, by means <strong>of</strong> a simple constraintbased<br />
interface, to a forward-search-based planner<br />
[10]. STAN 4 was one <strong>of</strong> <strong>the</strong> prize-winning systems,<br />
selected for <strong>the</strong> promise it showed in utilizing<br />
novel approaches to solving complex planning problems.<br />
Its plan quality was generally superior to that<br />
produced by <strong>the</strong> o<strong>the</strong>r competing systems.<br />
Work remains to be done on increasing <strong>the</strong> sophistication<br />
<strong>of</strong> <strong>the</strong> integration techniques that support coordination<br />
<strong>of</strong> different specialised solution strategies<br />
within <strong>the</strong> overall framework. An important aspect<br />
<strong>of</strong> making <strong>the</strong> configuration tools available to <strong>the</strong><br />
general planning community is to provide a clean<br />
means <strong>of</strong> access to <strong>the</strong> library, pattern recognition<br />
techniques and planner interface. A priority for future<br />
development is to supply an API to <strong>the</strong> suite <strong>of</strong><br />
tools we have, which o<strong>the</strong>r planning systems could<br />
exploit. Although we made progress with <strong>the</strong> design<br />
and implementation <strong>of</strong> such an API we did not complete<br />
its implementation and it remains a topic for future<br />
work.<br />
Overall Conclusions<br />
The PLANFORM project set out to construct an<br />
open environment for planner development, bringing<br />
knowledge acquisition tools, domain construction<br />
tools, modelling languages and planner configuration<br />
components into an integrated organisation making<br />
planning accessible to <strong>the</strong> non-specialist.<br />
The domain construction tool developed at Huddersfield<br />
produces PDDL domain descriptions providing<br />
a simple connection to <strong>the</strong> planner configuration architecture<br />
developed at Durham. The knowledge acquisition<br />
tools developed at Salford assist a user in<br />
confronting <strong>the</strong> task <strong>of</strong> domain construction through<br />
<strong>the</strong> GIPO tool. At this stage it is possible for a naive<br />
user to follow <strong>the</strong> entire process <strong>of</strong> modelling, and<br />
planning with, a specific problem domain without requiring<br />
detailed knowledge <strong>of</strong> any internal representation<br />
language ( ��¥¡ or PDDL). The level <strong>of</strong> abstraction<br />
at which such a user can work within <strong>the</strong><br />
environment will be fur<strong>the</strong>r raised when <strong>the</strong> implementation<br />
<strong>of</strong> design patterns, as a guiding principle<br />
in <strong>the</strong> modelling process, is completed.<br />
More time and work is necessary to evaluate <strong>the</strong> environment<br />
in terms <strong>of</strong> how successful it is at making<br />
planning accessible to <strong>the</strong> non-expert. There are several<br />
key aspects <strong>of</strong> <strong>the</strong> environment that require evaluation:
44 The PLANET <strong>Newsletter</strong><br />
The extent to which <strong>the</strong> object-oriented approach<br />
to modelling ameliorates <strong>the</strong> modelling task for<br />
<strong>the</strong> planning non specialist;<br />
The extent to which design patterns can fur<strong>the</strong>r<br />
ameliorate this effort;<br />
The extent to which <strong>the</strong> proposed knowledge acquisition<br />
techniques can capture <strong>the</strong> modelling intentions<br />
<strong>of</strong> <strong>the</strong> user, and how transparent <strong>the</strong> environment<br />
can make <strong>the</strong> process <strong>of</strong> iterating over <strong>the</strong><br />
modelling task until effective capture is achieved.<br />
The last <strong>of</strong> <strong>the</strong>se concerns <strong>the</strong> issue <strong>of</strong> how to provide<br />
useful feedback to <strong>the</strong> user when <strong>the</strong> modelling<br />
process fails to result in a consistent model (due to<br />
missing, or conflicting, information). Without a consistent<br />
domain model <strong>the</strong> plan configuration tools can<br />
do nothing useful and it is <strong>the</strong>refore desirable that <strong>the</strong><br />
user be able to develop a correct domain model incrementally.<br />
We believe this is one <strong>of</strong> <strong>the</strong> most interesting<br />
technical challenges facing <strong>the</strong> PLANFORM<br />
project at this stage.<br />
Bibliography<br />
[1] AIPS-98 Planning Competition Committee.<br />
PDDL - The Planning Domain Definition Language.<br />
Technical Report CVC TR-98-003/DCS<br />
TR-1165, Yale Center for Computational Vision<br />
and Control, 1998.<br />
[2] R. S. Aylett and C. Doniat. The PLANFORM-<br />
KA Tool. In Proceedings <strong>of</strong> <strong>the</strong> 20th UK<br />
Planning and Scheduling Workshop (PLANSIG-<br />
2001), Edinburgh, Scotland, 2001.<br />
[3] R. S. Aylett and C. Doniat. KA Tool and domain<br />
construction for AI planning applications (submitted<br />
to ES2002). Technical report, University<br />
<strong>of</strong> Salford, 2002.<br />
[4] R. S. Aylett and C. Doniat. Supporting <strong>the</strong> Domain<br />
Expert in Planning Domain Construction<br />
. In Proceedings <strong>of</strong> <strong>the</strong> AIPS’02 Workshop on<br />
Knowledge Engineering Tools and Techniques<br />
for AI Planning, 2002.<br />
http://www.planet-noe.org<br />
[5] R. S. Aylett and S. Jones. Planner and Domain:<br />
Domain Configuration for a Task Planner. Int.<br />
Journal <strong>of</strong> Expert Systems, 9(2):279–318, 1996.<br />
[6] J. Breuker and B. Wielinga. Models <strong>of</strong> Expertise<br />
in Knowledge Acquisition. In Topics in Expert<br />
Systems Design: Methodologies and Tools<br />
(eds: G. Guida and C. Tasso), 1989.<br />
[7] B. Chandrasekaran and T. R. Johnson. Generic<br />
Tasks and Task Structures: History, Critique<br />
and New Directions. In Second Generation Expert<br />
Systems (eds: J. M. David, J. P. Krivine and<br />
R. Simmons), 1993.<br />
[8] M. Fox and D. Long. The Automatic Inference<br />
<strong>of</strong> State Invariants in TIM. Journal <strong>of</strong> AI Research,<br />
9:367–421, 1998.<br />
[9] M. Fox and D. Long. Hybrid STAN: Identifying<br />
and Managing Combinatorial Optimization<br />
Sub-problems in Planning. In Proceedings <strong>of</strong><br />
<strong>the</strong> International Joint Conference on AI, 2001.<br />
[10] M. Fox and D. Long. STAN 4: A Hybrid Planning<br />
Strategy based on Sub-Problem Abstraction.<br />
AI Magazine, AAAI Press, 22(4), 2001.<br />
[11] M. Fox and D. Long. Integrating Multiple Subsolving<br />
Strategies into a Hybrid Planning System<br />
(submitted to <strong>the</strong> Journal <strong>of</strong> AI Research).<br />
Technical report, Department <strong>of</strong> Computer Science,<br />
University <strong>of</strong> Durham, 2002.<br />
[12] D. Liu and T. L. McCluskey. The OCL Language<br />
Manual, Version 1.2. Technical report,<br />
Department <strong>of</strong> Computing and Ma<strong>the</strong>matical<br />
Sciences, University <strong>of</strong> Huddersfield , 2000.<br />
[13] D. Long and M. Fox. Extracting Route-<br />
Planning: First Steps in Automatic Problem Decomposition.<br />
In AIPS Workshop on Analyzing<br />
and Exploiting Domain Knowledge, 2000.<br />
[14] D. Long and M. Fox. Recognising and exploiting<br />
generic types in planning domains. In Proceedings<br />
<strong>of</strong> <strong>the</strong> International Conference on AI<br />
Planning and Scheduling, 2000.
The PLANET <strong>Newsletter</strong> 45<br />
[15] D. Long and M. Fox. Recognizing and Exploiting<br />
Generic Types in Planning Domains). In<br />
Proceedings <strong>of</strong> <strong>the</strong> Fourth International Conference<br />
on AI Planning and Scheduling (AIPS-<br />
2000), 2000.<br />
[16] D. Long and M. Fox. Multi-Processor Scheduling<br />
Problems in Planning. In Proceedings <strong>of</strong> <strong>the</strong><br />
International Conference on AI, 2001.<br />
[17] D. Long, M. Fox, and M. Hamdi. Reformulation<br />
in Planning. In Proceedings <strong>of</strong> <strong>the</strong> Symposium<br />
on Abstraction, Reformulation and Approximation,<br />
2002.<br />
[18] T. L. McCluskey. Planform: Draft Progress Report.<br />
Technical Report, School <strong>of</strong> Computing<br />
and Maths, Univ <strong>of</strong> Huddersfiled, 2001.<br />
[19] T. L. McCluskey and D. E. Kitchin. A Tool-<br />
Supported Approach to Engineering HTN Planning<br />
Models. In Proceedings <strong>of</strong> 10th IEEE International<br />
Conference on Tools with Artificial<br />
Intelligence, 1998.<br />
[20] T. L. McCluskey and J. M. Porteous. Engineering<br />
and Compiling Planning Domain Models to<br />
Promote Validity and Effici ency. Artificial Intelligence,<br />
95:1–65, 1997.<br />
[21] T. L. McCluskey and N. E. Richardson. The induction<br />
<strong>of</strong> operator descriptions from examples<br />
and structural domain knowledge. In Proceedings<br />
<strong>of</strong> <strong>the</strong> 20th Workshop <strong>of</strong> <strong>the</strong> UK Planning<br />
and Scheduling Special Interest Group, pages<br />
181–192, 2001.<br />
[22] T. L. McCluskey, N. E. Richardson, and R. M.<br />
Simpson. An Interactive Method for Inducing<br />
Operator Descriptions. In The Sixth International<br />
Conference on Artificial Intelligence<br />
Planning Systems, 2002.<br />
[23] L. C. J. Murray. Reuse <strong>of</strong> Control Knowledge<br />
in Planning Domains. In AIPS Workshop on<br />
Knowledge Engineering Tools and Techniques<br />
for AIP Planning, 2002.<br />
[24] M. Musen. Modern Architectures for Intelligent<br />
Systems: Reusable Ontologies and Problem-<br />
Solving Methods. In AMIA Annual Symposium),<br />
1998.<br />
[25] A. Newell. The Knowledge Level. Artificial<br />
Intelligence, 1982.<br />
[26] R. M. Simpson, T. L. McCluskey, Maria<br />
Fox, and Derek Long. Generic Types as<br />
Design Patterns for Planning Domain Specifications.<br />
In Proceedings <strong>of</strong> <strong>the</strong> AIPS’02<br />
Workshop on Knowledge Engineering<br />
Tools and Techniques for AI Planning,<br />
http://scom.hud.ac.uk/planet/aips02kett/, 2002.<br />
[27] R. M. Simpson, T. L. McCluskey, D. Liu, and<br />
D. E. Kitchin. Knowledge Representation in<br />
Planning: A PDDL to ��¥¡¤£ Translation. In<br />
Proceedings <strong>of</strong> <strong>the</strong> 12th International Symposium<br />
on Methodologies for Intelligent Systems,<br />
2000.<br />
[28] R. M. Simpson, T. L. McCluskey, W. Zhao,<br />
R. S. Aylett, and C. Doniat. GIPO: An Integrated<br />
Graphical Tool to support Knowledge<br />
Engineering in AI Planning. In Proceedings<br />
<strong>of</strong> <strong>the</strong> 6th <strong>European</strong> Conference on Planning,<br />
2001.<br />
[29] M. C. Tully. Object-Oriented Domain Modelling<br />
for Planning. Technical report, Department<br />
<strong>of</strong> Computer Science, University <strong>of</strong><br />
Durham, 2001.<br />
[30] C. Vogel. Le Genie Cognitif. 1988.<br />
Author Information<br />
T.L. McCluskey University <strong>of</strong> Huddersfield, UK<br />
Maria Fox University <strong>of</strong> Durham, UK<br />
Ruth Aylett University <strong>of</strong> Salford, UK<br />
This work was supported by EPSRC projects<br />
GRM67421/01, GRM70568/01 and GRM68237/01,<br />
held at <strong>the</strong> three participating sites.
46 The PLANET <strong>Newsletter</strong><br />
Second PLANET Gap-Bridging Seminar (GBS-2)<br />
The second PLANET Gap-Bridging Seminar (GBS-<br />
2) took place in Delft, The Ne<strong>the</strong>rlands, on 21<br />
November 2002. GBS-2 was held in conjunction<br />
with PlanSIG2002, <strong>the</strong> 21st workshop <strong>of</strong> <strong>the</strong><br />
UK AI Planning & Scheduling Special Interest<br />
Group. GBS-2 built on <strong>the</strong> success <strong>of</strong> <strong>the</strong> first<br />
Gap-Bridging Seminar, held in Edinburgh, Scotland,<br />
in December 2001 (see PLANET News issue<br />
3, pages 30-31 and http://cswww.essex.ac.<br />
uk/conferences/planet/GBS-1/).<br />
The motivation for Gap-Bridging Seminars is <strong>the</strong> observation<br />
that industry and academia both work on<br />
planning and scheduling, but <strong>the</strong>y do not work toge<strong>the</strong>r<br />
as well as one would hope. They have different<br />
goals. Industry must sell, academia must publish,<br />
and <strong>the</strong>re is no time to talk to each o<strong>the</strong>r. PLANET’s<br />
aim is to give representatives <strong>of</strong> both realms <strong>the</strong> opportunity<br />
and <strong>the</strong> time to exchange views.<br />
In <strong>the</strong> Gap-Bridging Seminars, practitioners from industry<br />
talk about <strong>the</strong>ir work – <strong>the</strong> techniques, <strong>the</strong><br />
http://www.planet-noe.org<br />
Figure 1: GBS-2 venue in Delft.<br />
REPORT<br />
Author: T. Grant<br />
problems, customer needs, what can and cannot be<br />
done. This sets research in planning and scheduling<br />
in a wider context. Most <strong>of</strong> <strong>the</strong> speakers stayed<br />
for <strong>the</strong> whole <strong>of</strong> PlanSIG2002, so that attendees<br />
could have more opportunities to exchange views<br />
with <strong>the</strong>m.<br />
In GBS-1 most <strong>of</strong> <strong>the</strong> speakers were from companies<br />
that developed and applied AI-based s<strong>of</strong>tware products<br />
for planning and scheduling. By contrast, <strong>the</strong><br />
GBS-2 speakers were chosen to present a wide variety<br />
<strong>of</strong> real-world planning and scheduling applications.<br />
The speakers and <strong>the</strong>ir titles were:<br />
Alessandro Donati <strong>European</strong> Space Operations<br />
Centre (ESOC), Darmstadt, Germany: “Space<br />
Mission Operations Planning and Scheduling:<br />
Past, present and future”.<br />
Henk Hesselink and Ron Seljée Dutch Aerospace<br />
Laboratory (NLR), Amsterdam, Ne<strong>the</strong>rlands: “AI<br />
Planning, Waiting for results?”
The PLANET <strong>Newsletter</strong> 47<br />
Frank Oxener Atos Origin Nederland b.v., Ne<strong>the</strong>rlands:<br />
“Work Force Planning: A logical next step<br />
after ERP”<br />
Yossi Rissin and Roman Bartak VisOpt b.v.:<br />
“When Theory crashed into Reality”.<br />
In addition, Tim Grant introduced GBS-2 and<br />
PLANET.<br />
The audience consisted <strong>of</strong> 42 participants, mostly<br />
from <strong>the</strong> academic community (see Figure 1). UK<br />
and <strong>the</strong> Ne<strong>the</strong>rlands each provided a third <strong>of</strong> <strong>the</strong><br />
people attending. The remaining third was divided<br />
over Spain, Italy, Germany, France, Belgium, and <strong>the</strong><br />
Czech Republic. Academic attendees were primarily<br />
post-graduate students.<br />
Alessandro Donati identified <strong>the</strong> need for three communities<br />
– users, innovators and implementers – to<br />
work more closely with one ano<strong>the</strong>r. He explained<br />
that ESOC is <strong>the</strong> part <strong>of</strong> <strong>the</strong> <strong>European</strong> Space Agency<br />
(ESA) responsible for launching and operating scientific<br />
spacecraft. A very recent survey at ESOC had<br />
shown that <strong>the</strong> planning and scheduling processes<br />
differed radically between space missions. The reasons<br />
why had yet to be established. A priori, it would<br />
Figure 2: Air traffic management problem in Europe.<br />
seem to be more efficient to have a standard process<br />
and standard tools for all missions. He welcomed <strong>the</strong><br />
involvement <strong>of</strong> <strong>the</strong> research community (<strong>the</strong> innovators)<br />
and PLANET, and mentioned <strong>the</strong> opportunities<br />
for post-graduate students to work in ESA.<br />
A lively discussion ensued, with researchers calling<br />
for information to be made available on complex,<br />
real-world planning and scheduling domains<br />
and problems as case studies. This discussion resulted<br />
in a lunchtime splinter meeting on <strong>the</strong> second<br />
day between Alessandro and representatives <strong>of</strong><br />
<strong>the</strong> innovator and implementer communities on how<br />
PLANET and ESOC could co-operate to document<br />
one or more case studies, to mutual benefit.<br />
Henk Hesselink talked about NLR’s experience in<br />
providing s<strong>of</strong>tware support for planning and scheduling<br />
in civil and military aircraft operations. The<br />
NLR is a non-pr<strong>of</strong>it organisation founded in 1919<br />
to provide technical and scientific contributions to<br />
aerospace organisations in <strong>the</strong> private and public sectors.<br />
Turnover is EUR 70 million annually, split 65%<br />
civil and 35% military. The ratio between development<br />
and operations is 40:60, and between aeronautical<br />
and space applications is 85:15. Facilities
48 The PLANET <strong>Newsletter</strong><br />
available to <strong>the</strong> NLR include large wind-tunnels, two<br />
aircraft, various simulators, and computing environments<br />
(including a supercomputer). NLR co-operates<br />
with Amsterdam Schiphol airport, Dutch industry,<br />
and Eurocontrol.<br />
Henk focussed on <strong>the</strong> forthcoming air traffic management<br />
(ATM) problem in Europe (see Figure 2). This<br />
problem is not recognised by <strong>the</strong> controllers – <strong>the</strong> potential<br />
end-users – because ATM is divided into subproblems.<br />
Many actors are involved, and each airport<br />
is different, yet has features in common. Currently,<br />
ATM controllers do not plan, but work on a<br />
“first-heard, first-served” basis. They are conservative,<br />
with air safety being uppermost in <strong>the</strong>ir mind,<br />
followed by <strong>the</strong> need for fairness in allocating resources.<br />
They have nei<strong>the</strong>r <strong>the</strong> time nor <strong>the</strong> interest<br />
to talk to researchers. Henk argued that gap-bridging<br />
had to start with <strong>the</strong> foundations <strong>of</strong> <strong>the</strong> bridge: finance<br />
and user interest. He advocated <strong>the</strong> use <strong>of</strong> prototypes<br />
to make users aware <strong>of</strong> <strong>the</strong>ir problems and<br />
potential solutions.<br />
Frank Oxener explained how <strong>the</strong> complexity <strong>of</strong> recent<br />
law-making in <strong>the</strong> Ne<strong>the</strong>rlands had given rise to<br />
commercial opportunities for planning and scheduling<br />
applications in work-force management (see Figure<br />
3). Several Dutch companies were <strong>of</strong>fering s<strong>of</strong>tware<br />
products and associated services.<br />
One law required employers to give shift-workers 36<br />
hours <strong>of</strong> rest in any 7-day period or 60 hours rest in<br />
any 9-day period. Once every 5 weeks <strong>the</strong>y had to<br />
be allowed at least 32 hours continuous rest. Ano<strong>the</strong>r<br />
law provided for irregularity payments <strong>of</strong> 20%<br />
<strong>of</strong> salary to workers working Monday to Friday between<br />
06:00 and 08:00, providing <strong>the</strong>y started before<br />
07:00. On Saturdays <strong>the</strong> payment was 40%, and on<br />
Sundays 100%. These laws were a challenge to s<strong>of</strong>tware<br />
systems for recording actual hours worked and<br />
planning rest periods and irregularity payments. He<br />
outlined his experiences with applying a variety <strong>of</strong><br />
commercially-available s<strong>of</strong>tware packages.<br />
Frank elicited audience participation by asking his<br />
listeners to write down a question on a sheet <strong>of</strong> paper.<br />
He <strong>the</strong>n answered each question. Sample questions<br />
included:<br />
http://www.planet-noe.org<br />
Can <strong>the</strong> tools easily be adapted to <strong>the</strong> workinghour<br />
laws in o<strong>the</strong>r countries?<br />
What is <strong>the</strong> key feature in work-planning s<strong>of</strong>tware<br />
for user acceptability?<br />
What type <strong>of</strong> techniques are used to solve <strong>the</strong><br />
problems you have described?<br />
How does a different culture really affect a planning<br />
process?<br />
How might a planning solution be overridden by<br />
labour union objections?<br />
How long does it take to implement a system from<br />
first requirement to operational use?<br />
Are your solutions sufficiently flexible to deal<br />
quickly with changes in <strong>the</strong> law?<br />
What is coming after workflow management?<br />
Why?<br />
Roman Bartak focussed on <strong>the</strong> human factors issues.<br />
Human behaviour is inconsistent and readily affected<br />
by mood, environment and psychological pressure. It<br />
can only be modelled statistically. Plant personnel<br />
and planners are motivated by pride, <strong>the</strong>ir position in<br />
<strong>the</strong> organisation, and future job security. Pride makes<br />
it difficult for users to admit mistakes, problems and<br />
weaknesses. They protect <strong>the</strong>ir position by being nice<br />
to superiors, gaining pr<strong>of</strong>essional respect, and trying<br />
to serve many masters at <strong>the</strong> same time. They protect<br />
<strong>the</strong>ir job by withholding knowledge. Internal politics<br />
and power-plays are key factors in decision-making.<br />
Figure 3: Right people, right place?<br />
(acknowledgements to Parallax b.v.)
The PLANET <strong>Newsletter</strong> 49<br />
These factors and issues make academic research irrelevant.<br />
One scheduling expert told him: “I have<br />
never seen a Job Shop Scheduling problem in practice”.<br />
From an academic point <strong>of</strong> view, <strong>the</strong> ideal factory<br />
is one that is totally automated, populated with<br />
robots and Automated Guided Vehicles (AGVs). Alternatively,<br />
it might be a new factory that has not yet<br />
been put into operation. There is <strong>the</strong>n no previous<br />
“know-how”, rules and procedures, bad habits, and<br />
day-to-day reality to confront <strong>the</strong>ory.<br />
Roman contrasted <strong>the</strong> views <strong>of</strong> planners and academics,<br />
looking at <strong>the</strong> “Not Invented Here” syndrome.<br />
Summarising <strong>the</strong> lessons he had learned at<br />
VisOpt b.v., he advocated <strong>the</strong> development <strong>of</strong> a visual<br />
modelling language as a way <strong>of</strong> improving communication<br />
between <strong>the</strong> two communities (see Figure 4).<br />
PLANET provided sponsorship for 14 post-graduate<br />
students from Spain, Germany, Ne<strong>the</strong>rlands, UK,<br />
Greece, and Italy to attend GBS-2. In return, <strong>the</strong> students<br />
wrote a short report on what <strong>the</strong>y had learned<br />
from GBS-2 and PlanSIG2002. Key quotes included:<br />
“The gap turned out to be bigger than I thought”.<br />
Figure 4: VisOpt’s visual modelling language.<br />
“The workshop gave me <strong>the</strong> chance to get acquainted<br />
to people whose work I had been reading<br />
in papers for quite some time”.<br />
“Industry does not always need <strong>the</strong> best possible<br />
plan – it needs one that is good enough”.<br />
“An attempt should be made from both sides to<br />
approach one ano<strong>the</strong>r”.<br />
“Managers and planners need to be convinced to<br />
change to new planning methods”.<br />
“Industry (NASA, ESOC, NLR, etc) can help<br />
academia by supplying real planning domains,<br />
problems, and case studies”.<br />
“[There are] four key items: visual modelling,<br />
planning as a step-by-step process, scheduling as<br />
a process to reason on resource constraints, and<br />
<strong>the</strong> need for generating good enough plans in a<br />
reasonable time”.<br />
“There should be cooperation between [human]<br />
planner and computer”.
50 The PLANET <strong>Newsletter</strong><br />
“[It is] necessary for researchers to show <strong>the</strong> benefits<br />
that can be obtained”.<br />
“There is also a gap between different academic<br />
communities: Operations Research and AI”.<br />
“Employees do not see <strong>the</strong> global view, but only<br />
<strong>the</strong> bit <strong>the</strong>y are working on”.<br />
“One challenging line <strong>of</strong> research is <strong>the</strong> validation<br />
<strong>of</strong> planning domains”.<br />
“Domains in industry are far less predictable<br />
and much more dynamic that those used by researchers”.<br />
“Both communities need more communication”.<br />
“[GBS-2] encouraged new and interesting discussions,<br />
thus exposing many unexplored areas <strong>of</strong> research,<br />
and providing good ideas for future work”.<br />
http://www.planet-noe.org<br />
“People from <strong>the</strong> two sides have to come closer<br />
and cooperate since this will be <strong>of</strong> great pr<strong>of</strong>it for<br />
both”.<br />
In summary, both speakers and audience came away<br />
from <strong>the</strong> second PLANET Gap-Bridging Seminar<br />
having learned more about <strong>the</strong> academic and industrial<br />
aspects <strong>of</strong> planning and scheduling. All attendees<br />
were most grateful to PLANET and <strong>the</strong> CEC for<br />
making it possible to exchange such a diversity <strong>of</strong><br />
views, and look forward with eagerness to <strong>the</strong> third<br />
Gap-Bridging Seminar.<br />
Author Information<br />
Tim Grant Atos-Origin Nederland b.v.,<br />
Nieuwegein, The Ne<strong>the</strong>rlands, Tim.Grant@<br />
atosorigin.com
The PLANET <strong>Newsletter</strong> 51<br />
ICAPS 2003<br />
ANNOUNCEMENT
52 The PLANET <strong>Newsletter</strong><br />
3rd PLANET International Summer School 2003<br />
The International Summer School on AI Planning is<br />
a great opportunity for Ph.D. students and young researchers<br />
to be exposed to introductory and advanced<br />
courses on various aspects <strong>of</strong> Artificial Intelligence<br />
Planning, and to spend time and discuss research directions<br />
with <strong>the</strong>ir colleagues and with <strong>the</strong> teachers,<br />
foremost researchers in <strong>the</strong> field.<br />
The first edition <strong>of</strong> <strong>the</strong> school was held in Cyprus in<br />
September 2000, while <strong>the</strong> second school was held<br />
in Halkidiki, Greece, in September 2002. The third<br />
edition <strong>of</strong> <strong>the</strong> school will be held on <strong>the</strong> mountains<br />
<strong>of</strong> Trentino, in <strong>the</strong> nor<strong>the</strong>rn part <strong>of</strong> Italy, in June<br />
2003. The school will be colocated with International<br />
Conference on Automated Planning and Scheduling<br />
(ICAPS’03).<br />
The Program Chairs <strong>of</strong> <strong>the</strong> school are Daniel Borrajo<br />
ICAPS 2003 – Doctoral Consortium<br />
ICAPS-2003 invites PhD students to apply for <strong>the</strong><br />
Doctoral Consortium, which will provide an opportunity<br />
for a group <strong>of</strong> students to discuss and explore<br />
<strong>the</strong>ir research interests and career objectives with established<br />
researchers in Planning and Scheduling.<br />
The aims <strong>of</strong> <strong>the</strong> Doctoral Consortium are <strong>the</strong> following:<br />
to provide a forum for students to present <strong>the</strong>ir<br />
current research, and receive feedback from o<strong>the</strong>r<br />
students and senior researchers;<br />
to promote contacts among PhD students working<br />
in <strong>the</strong> same area;<br />
to support students with information and advice<br />
on academic, research and industrial careers;<br />
to financially support students by covering <strong>the</strong><br />
conference registration fee and by partially contributing<br />
to travel expenses.<br />
http://www.planet-noe.org<br />
ANNOUNCEMENT<br />
Millan (Universidad Carlos III de Madrid, Spain),<br />
and Alessandro Cimatti (ITC-irst, Italy). ITC-irst<br />
is also responsible for <strong>the</strong> local organization <strong>of</strong> <strong>the</strong><br />
event, with a team composed by Piergiorgio Bertoli,<br />
Mark Carman, Alessandro Cimatti and Alessandro<br />
Tuccio.<br />
Additional, up-to-date information will be available<br />
at <strong>the</strong> <strong>of</strong>ficial web site <strong>of</strong> <strong>the</strong> school:<br />
http:<br />
//sra.itc.it/planet/summer-school-03/<br />
Programme<br />
CALL FOR PARTICIPATION<br />
The programme will consist <strong>of</strong> students’ presentations<br />
on <strong>the</strong>ir current research interests. A voluntary<br />
mentoring programme will be organized to link students<br />
with like-minded researchers.<br />
Submissions<br />
We encourage submissions from Ph.D. students at<br />
any level, and from any topic area and methodology<br />
within Planning and Scheduling. On <strong>the</strong> basis <strong>of</strong><br />
<strong>the</strong> submissions, <strong>the</strong> Organizing Committee will select<br />
a group <strong>of</strong> students that will be invited to present<br />
<strong>the</strong>ir work during <strong>the</strong> Doctoral Consortium, and also<br />
to present a poster at <strong>the</strong> ICAPS-2003 poster session.<br />
Students accepted for participation in <strong>the</strong> Doc
The PLANET <strong>Newsletter</strong> 53<br />
toral Consortium will have free conference registration<br />
and a fixed allowance for travel/housing. The<br />
students’ abstracts will be made available on <strong>the</strong> web<br />
and included as part <strong>of</strong> <strong>the</strong> conference proceedings.<br />
Applicants should submit an extended abstract <strong>of</strong><br />
5 pages maximum by email to one <strong>of</strong> <strong>the</strong> Doctoral<br />
Consortium chairs. The submission should be<br />
in AAAI style format (http://www.aaai.org/<br />
Publications/Author/macros-link.html)<br />
and sent ei<strong>the</strong>r as a PostScript or as a PDF file. It<br />
should describe original, unpublished work, completed<br />
or in progress, that is part <strong>of</strong> <strong>the</strong> doctoral work<br />
<strong>of</strong> <strong>the</strong> student. If an extended version <strong>of</strong> <strong>the</strong> paper<br />
is also submitted to <strong>the</strong> technical programme, please<br />
indicate it in <strong>the</strong> submission. Double submission is<br />
acceptable, but if <strong>the</strong> paper is accepted for <strong>the</strong> technical<br />
programme, <strong>the</strong> student will present <strong>the</strong> work<br />
only in <strong>the</strong> technical programme sessions and not<br />
during <strong>the</strong> Doctoral Consortium.<br />
In addition, <strong>the</strong> dissertation advisor should send a<br />
letter <strong>of</strong> recommendation by e-mail to one <strong>of</strong> <strong>the</strong><br />
Doctoral Consortium chairs. It should include <strong>the</strong><br />
expected date for <strong>the</strong>sis submission, and <strong>the</strong> motivation/expected<br />
benefit for <strong>the</strong> student to attend <strong>the</strong><br />
ICAPS 2003 – Workshop Program<br />
The ICAPS-03 workshop program (June 9-10, before<br />
<strong>the</strong> main program), has been set. There will<br />
be five workshops, covering a broad range <strong>of</strong> topics<br />
ranging from <strong>the</strong> current and future state <strong>of</strong> <strong>the</strong><br />
Planning Competition, to issues arising as planning<br />
and scheduling are applied to ever-more-complex domains,<br />
to specific application areas.<br />
The ICAPS-03 workshops:<br />
Planning and Web <strong>Service</strong>s<br />
Web services are revolutionizing <strong>the</strong> way industry<br />
and government operate. Web services both pro-<br />
Doctoral Consortium. This letter can be sent in as<br />
ei<strong>the</strong>r a text or a PostScript or a PDF file.<br />
Important Dates<br />
March 31st: deadline for submitting papers and<br />
letters <strong>of</strong> support<br />
April 18th: notification <strong>of</strong> acceptance to program<br />
April 25th: camera ready copy <strong>of</strong> <strong>the</strong> papers to <strong>the</strong><br />
chairs<br />
Doctoral Programme Chairs<br />
Jeremy Frank , NASA Ames Research Center,<br />
frank@email.arc.nasa.gov<br />
Susanne Biundo , University <strong>of</strong> Ulm, susanne.<br />
biundo@informatik.uni-ulm.de<br />
Additional information is available at <strong>the</strong> ICAPS<br />
Website<br />
http://icaps03.itc.it<br />
This event will be sponsored by PLANET and Nasa.<br />
ANNOUNCEMENT<br />
vide information (e.g., available flights) and change<br />
<strong>the</strong> world (e.g., buying a flight ticket). As <strong>the</strong><br />
Web evolves into <strong>the</strong> Semantic Web, <strong>the</strong> myriad <strong>of</strong><br />
available services are being described declaratively.<br />
Machine-understandable descriptions enable <strong>the</strong> automatic<br />
discovery, use, and composition <strong>of</strong> web services.<br />
With increased interest in <strong>the</strong> web services paradigm,<br />
composition <strong>of</strong> web services has become <strong>of</strong> primary<br />
importance. Several languages for describing web<br />
services and <strong>the</strong>ir composition are currently being defined<br />
and seek to become standards. From a planning<br />
perspective, <strong>the</strong> web services can be seen as op-
54 The PLANET <strong>Newsletter</strong><br />
erators, specific web services compositions as plans,<br />
and automatic web service composition as a form <strong>of</strong><br />
planning. This workshop will provide planning researchers<br />
with a forum for presenting planning results<br />
relevant to web services, identify new challenges, and<br />
lead <strong>the</strong> development <strong>of</strong> <strong>the</strong> critically important field<br />
<strong>of</strong> web services.<br />
Jose Luis Ambite<br />
Workshop on Plan Execution<br />
Much work in <strong>the</strong> planning community has focussed<br />
primarily upon developing efficient ways <strong>of</strong> generating<br />
plans that are not actually executed. This<br />
was reflected in <strong>the</strong> AIPS 2002 Planning Competition<br />
which measured <strong>the</strong> efficiency and optimality <strong>of</strong><br />
plans that were generated, but not executed. As execution<br />
may not result in <strong>the</strong> intended outcome <strong>the</strong><br />
sequence <strong>of</strong> actions that is eventually executed may<br />
not be as valuable as that in <strong>the</strong> original plan, which<br />
leads to <strong>the</strong> question as to whe<strong>the</strong>r it is necessary to<br />
generate plans that are near optimal. Researchers in<br />
<strong>the</strong> planning community are increasingly concerned<br />
with executing plans and <strong>the</strong> design <strong>of</strong> systems in<br />
which planning and execution are continually interleaved<br />
and actively managed.<br />
This workshop is intended for researchers who have<br />
interests in plan execution in a variety <strong>of</strong> domains<br />
such as robotics, space applications, information<br />
ga<strong>the</strong>ring and o<strong>the</strong>r areas.<br />
Alex Coddington<br />
Workshop on PDDL<br />
PDDL, originally developed by Drew McDermott<br />
and <strong>the</strong> 1998 planning competition committee, was<br />
inspired by <strong>the</strong> need to encourage <strong>the</strong> empirical comparison<br />
<strong>of</strong> planning systems and <strong>the</strong> exchange <strong>of</strong><br />
planning benchmarks within <strong>the</strong> community. Its development<br />
improved <strong>the</strong> communication <strong>of</strong> research<br />
results and triggered an explosion in performance,<br />
expressivity and robustness <strong>of</strong> planning systems.<br />
PDDL has become a de facto standard language for<br />
describing planning domains, not only for <strong>the</strong> com-<br />
http://www.planet-noe.org<br />
petition but more widely, as it <strong>of</strong>fers an opportunity<br />
to carry out empirical evaluation <strong>of</strong> planning systems<br />
on a growing collection <strong>of</strong> generally adopted standard<br />
benchmark domains. The emergence <strong>of</strong> a language<br />
standard will have an impact on <strong>the</strong> entire field, influencing<br />
what is seen as central and what peripheral<br />
in <strong>the</strong> development <strong>of</strong> planning systems. The adoption<br />
<strong>of</strong> PDDL in this role is itself an issue for debate:<br />
perhaps a completely different modelling language<br />
is called for. We believe that it is <strong>the</strong>refore important<br />
to provide a forum in which <strong>the</strong> community can<br />
give feedback and present <strong>the</strong>ir ideas to <strong>the</strong> language<br />
designers, and in which <strong>the</strong> language designers can<br />
discuss <strong>the</strong>ir ideas for maintaining and extending, or<br />
even replacing <strong>the</strong> language.<br />
Sylvie Thiebeaux<br />
Planning under Uncertainty and Incomplete<br />
Information<br />
Controlling intelligent agents in complex real-world<br />
environments poses requirements that are not addressed<br />
in classical AI planning. Often it is not sufficient<br />
to find a sequence <strong>of</strong> actions leading to a given<br />
goal, since <strong>the</strong> initial state may not be known with<br />
precision, and action effects cannot be predicted with<br />
certainty.<br />
In <strong>the</strong> past few years, <strong>the</strong>re has been a growing interest<br />
in more general planning techniques, able to<br />
tackle <strong>the</strong> problems <strong>of</strong> uncertainty, nondeterminism,<br />
and incompleteness <strong>of</strong> information. Several research<br />
works have proposed more expressive domain models<br />
and description languages (e.g., allowing for actions<br />
with multiple transitions, possibly with different<br />
probabilities, and with costs), and more complex<br />
models <strong>of</strong> execution (e.g., dealing with information<br />
ga<strong>the</strong>ring at run-time). New planning techniques and<br />
algorithms have been developed to operate on such<br />
extended models and to produce plans which achieve<br />
<strong>the</strong> goals despite <strong>the</strong> uncertainty and incompleteness<br />
<strong>of</strong> information. The goal <strong>of</strong> this workshop is to bring<br />
toge<strong>the</strong>r people working in different areas <strong>of</strong> planning<br />
under uncertain and incomplete information.<br />
Marco Pistore
The PLANET <strong>Newsletter</strong> 55<br />
The Planning Competition: Impact,<br />
Organization, Evaluation, Benchmarks<br />
The planning competition series undoubtedly has had<br />
a huge impact on <strong>the</strong> field <strong>of</strong> AI planning, including<br />
such aspects as growing standardization <strong>of</strong> complex<br />
planning domain description languages, dramatically<br />
improved scalability <strong>of</strong> existing approaches, and a<br />
growing database <strong>of</strong> commonly used benchmark examples.<br />
It is <strong>the</strong>refore important to provide an opportunity<br />
for discussing topics related to <strong>the</strong> competition.<br />
The workshop aims at doing just this. We want<br />
to collect toge<strong>the</strong>r panels on topics such as <strong>the</strong> role <strong>of</strong><br />
<strong>the</strong> competition in and for <strong>the</strong> field, organizational as-<br />
ICAPS 2003 – Tutorials<br />
Provisional Schedule<br />
Monday, June 9<br />
pects <strong>of</strong> <strong>the</strong> competition, (competition) results evaluation,<br />
and benchmarking issues. Technical presentations<br />
are planned on topics related to <strong>the</strong> competition<br />
such as language alternatives, and methods <strong>of</strong> empirical<br />
evaluation.<br />
Joerg H<strong>of</strong>fmann and Stefan Edelkamp<br />
Submissions to all workshops are due by 31 March,<br />
2003, with notification and submission <strong>of</strong> camaraready<br />
versions by <strong>the</strong> end <strong>of</strong> April.<br />
Additional detail about <strong>the</strong> workshops, as well as fur<strong>the</strong>r<br />
information about ICAPS-03 in general, is also<br />
available on <strong>the</strong> conference website:<br />
http://icaps03.itc.it/<br />
morning afternoon<br />
Timed Automata for Planning and Scheduling<br />
Oded Maler (Verimag)<br />
Tuesday, June 10<br />
morning afternoon<br />
Practical Approaches to Handling Uncertainty in<br />
Planning and Scheduling<br />
J. Christopher Beck (University College Cork)<br />
and Thierry Vidal (ENIT)<br />
Model Checking – A Hands-On Introduction<br />
Alessandro Cimatti, Marco Pistore and Marco<br />
Roveri (ITC- IRST)<br />
ANNOUNCEMENT<br />
Resource-Bounded and Time-Critical Reasoning<br />
Lloyd Greenwald (Drexel University) and<br />
Shlomo Zilberstein (University <strong>of</strong> Massahusetts)<br />
ICAPS’03 Tutorial chairs: Anthony Barrett, NASA Jet Propulsion Laboratory<br />
Jussi Rintanen, Albert-Ludwigs-Universität Freiburg
56 The PLANET <strong>Newsletter</strong><br />
Timed Automata for Planning and<br />
Scheduling<br />
In this tutorial we propose <strong>the</strong> model <strong>of</strong> timed automata,<br />
originating from <strong>the</strong> verification <strong>of</strong> real-time<br />
systems, as a model for posing and solving timedependent<br />
planning and scheduling problems. We<br />
believe that in <strong>the</strong> same sense as automata are used<br />
as <strong>the</strong> major vehicle for verification <strong>of</strong> systems where<br />
<strong>the</strong> model <strong>of</strong> time is qualitative, timed automata can<br />
be <strong>the</strong> center <strong>of</strong> a a unifying ma<strong>the</strong>matical modeling<br />
framework for quantitative time, having <strong>the</strong> following<br />
attractive features:<br />
1. It is sufficiently expressive to describe <strong>the</strong> essential<br />
aspects <strong>of</strong> time-dependent real-life problems<br />
in a variety <strong>of</strong> application domains.<br />
2. It provides for models with well-defined and clear<br />
dynamic semantics.<br />
3. These models are amenable to computer-aided design<br />
methods such as simulation, testing, verification<br />
and automatic syn<strong>the</strong>sis <strong>of</strong> (optimal) schedules<br />
and plans.<br />
4. These methods are currently supported by tools<br />
<strong>of</strong> various levels <strong>of</strong> maturity, that treat <strong>the</strong> specific<br />
computational problems <strong>of</strong> time-related reasoning.<br />
Oded Maler was born in 1957 in Haifa, Israel. He<br />
obtained his B.A. in Computer Science from <strong>the</strong><br />
Technion, Haifa in 1979 and his M.Sc. in Management<br />
Science from <strong>the</strong> University <strong>of</strong> Tel-Aviv at<br />
1984. In 1989 he finished his Ph.D. <strong>the</strong>sis (Finite<br />
Automata: Infinite Behavior, Learnability and Decomposition<br />
), under <strong>the</strong> supervision <strong>of</strong> A. Pnueli in<br />
<strong>the</strong> department <strong>of</strong> Applied Ma<strong>the</strong>matics and Computer<br />
Science, Weizmann Institute, Rehovot. After<br />
two years <strong>of</strong> post-doc at IRISA, Rennes, he moved<br />
to Grenoble at 1992 and obtained a research position<br />
(CR1) at <strong>the</strong> CNRS (French National Center <strong>of</strong> Scientific<br />
Research) in 1994. He has been promoted to<br />
“research director” (DR2) in 2001. Dr. Maler’s research<br />
is centered around <strong>the</strong> <strong>the</strong>ory <strong>of</strong> automata and<br />
http://www.planet-noe.org<br />
its various extensions, most notably timed automata,<br />
hybrid automata and <strong>the</strong>ir application to control, embedded<br />
systems, scheduling and circuit timing analysis.<br />
Practical Approaches to Handling<br />
Uncertainty in Planning and Scheduling<br />
This tutorial presents techniques for dealing with <strong>the</strong><br />
fact that <strong>the</strong> execution <strong>of</strong> plans and schedules in <strong>the</strong><br />
real world cannot assume a static environment: <strong>the</strong><br />
world changes non-deterministically during problem<br />
solving and execution. We present techniques from<br />
<strong>the</strong> Artificial Intelligence and Operations Research<br />
literature for handling uncertainty in planning and<br />
scheduling with emphasis on practical techniques.<br />
Such techniques include reactive, on-line scheduling<br />
and planning, and proactive, <strong>of</strong>f-line techniques that<br />
build solutions that can cope with uncertain events,<br />
as well as intermediate approaches between <strong>the</strong>se extremes.<br />
J. Christopher Beck received a PhD in Artificial Intelligence<br />
in 1999 from <strong>the</strong> University <strong>of</strong> Toronto under<br />
<strong>the</strong> supervision <strong>of</strong> Mark S. Fox. From 1994 to<br />
1999 he was <strong>the</strong> project manager <strong>of</strong> <strong>the</strong> Intelligent<br />
Scheduling Research Group at <strong>the</strong> Enterprise Integration<br />
Laboratory at University <strong>of</strong> Toronto. The<br />
focus <strong>of</strong> his research was measurements <strong>of</strong> problem<br />
structure as a basis for scheduling heuristics within a<br />
constraint-based scheduling framework. From 1999<br />
until 2002 he was a s<strong>of</strong>tware developer and Senior<br />
Scientist on <strong>the</strong> Scheduler team at ILOG, SA in Gentilly,<br />
France. As <strong>of</strong> June, 2002, he moved to <strong>the</strong> position<br />
<strong>of</strong> Staff Scientist at <strong>the</strong> Cork Constraint Computation<br />
Centre, University College Cork. His research<br />
interests focus on problem structure, hybrid<br />
algorithms, search in constraint-directed scheduling,<br />
and in <strong>the</strong> extension <strong>of</strong> constraint modeling and solving<br />
capabilities to incorporate aspects <strong>of</strong> real-world<br />
scheduling such as uncertainty, dynamic arrival <strong>of</strong> activities,<br />
and robustness.<br />
Thierry Vidal received a PhD in Artificial Intelligence<br />
in 1995 from <strong>the</strong> University <strong>of</strong> Toulouse under<br />
<strong>the</strong> supervision <strong>of</strong> Malik Ghallab. He had worked
The PLANET <strong>Newsletter</strong> 57<br />
in <strong>the</strong> Robotics and AI team <strong>of</strong> <strong>the</strong> LAAS-CNRS<br />
in Toulouse, France, working on temporal constraint<br />
processing in temporal planning (<strong>the</strong> IxTeT system)<br />
and in task scheduling, with a special focus on uncertain<br />
durations. In 1996-97 he was a guest researcher<br />
in Erik Sandewall’s team at <strong>the</strong> Department<br />
<strong>of</strong> Computer Science <strong>of</strong> <strong>the</strong> University <strong>of</strong> Linköping,<br />
Sweden, where he conducted basic research work in<br />
<strong>the</strong> area <strong>of</strong> on-line decision making through contingent<br />
plan recognition and reactive controller syn<strong>the</strong>sis.<br />
From 1997 he is assistant pr<strong>of</strong>essor at ENIT<br />
in Tarbes, France, working in <strong>the</strong> Automated Production<br />
team <strong>of</strong> <strong>the</strong> Production Engineering Laboratory,<br />
with external collaborations with Hélène Fargier<br />
(Possibilistic Reasoning team, IRIT, Toulouse), Paul<br />
Morris (NASA Ames Research Center, California,<br />
USA), and Ioannis Tsamardinos and Martha Pollack<br />
(University <strong>of</strong> Pittsburgh, USA). His current research<br />
interests are uncertain constraint reasoning in<br />
planning, scheduling and resource allocation, multiagent<br />
approaches to scheduling, reactivity, conditional<br />
planning and robust scheduling.<br />
Resource-Bounded and Time-Critical<br />
Reasoning<br />
A central problem in artificial intelligence is how to<br />
develop computational models that allow decisionsupport<br />
systems or autonomous agents to react to a<br />
situation after performing <strong>the</strong> right amount <strong>of</strong> deliberation.<br />
Frequently, <strong>the</strong> complexity <strong>of</strong> problem<br />
solving makes it beneficial to use approximate solutions<br />
ra<strong>the</strong>r than try to compute <strong>the</strong> optimal answer.<br />
This issue arises in a wide range <strong>of</strong> application<br />
domains including medical trauma management,<br />
Bayesian inference, sequence alignment, graphics<br />
rendering, web page prefetching, autonomous space<br />
exploration, real-time avionics, and robot navigation.<br />
This tutorial explores <strong>the</strong> <strong>the</strong>ory and practice <strong>of</strong> building<br />
intelligent systems that reason explicitly about<br />
employing limited computational resources to generate<br />
timely solutions to difficult combinatorial optimization,<br />
planning and scheduling problems. Solution<br />
techniques go beyond simple greedy or reactive<br />
algorithms to achieve high-quality solutions while<br />
meeting both hard and s<strong>of</strong>t real-time deadlines. We<br />
will explore over fifteen years <strong>of</strong> progress in this area,<br />
covering historical perspectives, state-<strong>of</strong>-<strong>the</strong>-art solution<br />
techniques, and current and future challenges.<br />
Participants should be familiar with introductory artificial<br />
intelligence, algorithm design and analysis, and<br />
introductory probability and statistics.<br />
Lloyd Greenwald is an Assistant Pr<strong>of</strong>essor <strong>of</strong> Computer<br />
Science and Director <strong>of</strong> <strong>the</strong> Intelligent Time-<br />
Critical Systems Lab at Drexel University. He received<br />
his Ph.D. in Computer Science from Brown<br />
University. His research interests include timecritical<br />
planning and scheduling, mobile robotics,<br />
machine learning, ad hoc and sensor networks, and<br />
medical decision making.<br />
Shlomo Zilberstein is an Associate Pr<strong>of</strong>essor <strong>of</strong><br />
Computer Science and Director <strong>of</strong> <strong>the</strong> Resource-<br />
Bounded Reasoning Lab (http://anytime.<br />
cs.umass.edu) at <strong>the</strong> University <strong>of</strong> Massachusetts,<br />
Amherst. He received his Ph.D. in Computer<br />
Science from <strong>the</strong> University <strong>of</strong> California,<br />
Berkeley. His research interests include approximate<br />
reasoning, decision <strong>the</strong>ory, heuristic search, planning<br />
and scheduling, and resource-bounded reasoning.<br />
Model Checking – A Hands-On<br />
Introduction<br />
Model Checking is a formal technique for <strong>the</strong> verification<br />
<strong>of</strong> designs <strong>of</strong> concurrent systems. It is based<br />
on <strong>the</strong> representation <strong>of</strong> <strong>the</strong> system being analyzed as<br />
a (finite state) transition systems (e.g. Kripke models),<br />
while <strong>the</strong> requirements are typically expressed<br />
in temporal logics. A system satisfies a given property<br />
amounts to checking if <strong>the</strong> corresponding temporal<br />
formula is true in <strong>the</strong> Kripke model. Model<br />
checking is very effective in pinpointing design errors<br />
that are extremely hard to detect by means <strong>of</strong> testing,<br />
and is <strong>the</strong>refore being applied in <strong>the</strong> industrial<br />
development <strong>of</strong> reactive systems, hardware designs,<br />
and communication protocols. Fur<strong>the</strong>rmore, model<br />
checking techniques and tools are gaining interest
58 The PLANET <strong>Newsletter</strong><br />
in several fields <strong>of</strong> Artificial Intelligence (e.g. Planning,<br />
Multi-agent systems, and Model-based Diagnosis)<br />
and Engineering (e.g. Requirement Verification,<br />
Safety Analysis). Of particular interest is Symbolic<br />
Model Checking, which makes it is possible to analyze<br />
extremely large finite-state systems by means <strong>of</strong><br />
symbolic representation techniques (e.g. Binary Decision<br />
Diagrams, propositional satisfiability).<br />
Alessandro Cimatti is <strong>the</strong> leader <strong>of</strong> <strong>the</strong> formal methods<br />
group within <strong>the</strong> Automated Reasoning Systems<br />
division (SRA) at ITC-IRST (http://www.<br />
irst.itc.it). The activities carried out by <strong>the</strong><br />
group include basic research, <strong>the</strong> development <strong>of</strong><br />
<strong>the</strong> NuSMV (http://nusmv.irst.itc.it)<br />
model checker, and technology transfer in industrial<br />
projects in <strong>the</strong> areas <strong>of</strong> safety critical applications<br />
(e.g., railways, avionics, aerospace, industrial plant<br />
controllers).<br />
Alessandro Cimatti has participated in and led several<br />
industrial projects aimed at <strong>the</strong> use <strong>of</strong> formal methods<br />
for <strong>the</strong> development and verification <strong>of</strong> safety critical<br />
systems and embedded controllers. Some examples<br />
are <strong>the</strong> validation <strong>of</strong> Interlocking Systems, <strong>the</strong> development<br />
<strong>of</strong> Rail Traffic Management Systems, <strong>the</strong> design<br />
<strong>of</strong> tools for on-board verification, and <strong>the</strong> verification<br />
<strong>of</strong> safety-critical communication protocols.<br />
Alessandro Cimatti is <strong>the</strong> leader <strong>of</strong> <strong>the</strong> development<br />
<strong>of</strong> NuSMV. His main research interests include<br />
<strong>the</strong> development <strong>of</strong> advanced model checking techniques,<br />
and <strong>the</strong> application <strong>of</strong> model checking for <strong>the</strong><br />
syn<strong>the</strong>sis <strong>of</strong> reactive controllers and test cases. He<br />
has also contributed to <strong>the</strong> research in <strong>the</strong>orem proving,<br />
formal languages for <strong>the</strong> specification <strong>of</strong> multiagent<br />
systems, planning and robotics.<br />
Marco Pistore is Associate Pr<strong>of</strong>essor at <strong>the</strong> Department<br />
<strong>of</strong> Information and Communication Technolo-<br />
http://www.planet-noe.org<br />
gies <strong>of</strong> <strong>the</strong> University <strong>of</strong> Trento (http://www.<br />
dit.unitn.it/) and Research Consultant at<br />
ITC-IRST (http://www.irst.itc.it). His<br />
research interests are in formal methods and in <strong>the</strong><br />
application <strong>of</strong> formal methods to planning and to syn<strong>the</strong>sis<br />
<strong>of</strong> controllers. Marco Pistore has been <strong>the</strong> responsible<br />
<strong>of</strong> <strong>the</strong> development <strong>of</strong> <strong>the</strong> NuSMV checker.<br />
He is also working to <strong>the</strong> Formal Tropos project, aiming<br />
at <strong>the</strong> development <strong>of</strong> a formal language and <strong>of</strong><br />
formal analysis techniques for <strong>the</strong> verification <strong>of</strong> requirements<br />
specifications. Marco Pistore has also<br />
participated to research and industrial projects on <strong>the</strong><br />
application <strong>of</strong> formal methods to <strong>the</strong> design and verification<br />
<strong>of</strong> safety-critical systems and <strong>of</strong> embedded<br />
controllers.<br />
Marco Roveri received a PhD in Computer Science<br />
in 2002 from <strong>the</strong> University <strong>of</strong> Milano in collaboration<br />
with ITC-Irst under <strong>the</strong> supervision <strong>of</strong><br />
A. Cimatti. His PhD <strong>the</strong>sis “Planning in Non-<br />
Deterministic Domains via Symbolic Model Checking”<br />
was awarded by <strong>the</strong> Italian Association for Artificial<br />
Intelligence (AI*IA) <strong>the</strong> best price for PhD<br />
<strong>the</strong>sis in artificial intelligence in Italy. He his in <strong>the</strong><br />
steering committee <strong>of</strong> <strong>the</strong> NuSMV symbolic model<br />
checker, <strong>the</strong> first state-<strong>of</strong>-<strong>the</strong>-art open source symbolic<br />
model checker. Since 2001 he is working at<br />
ITC-Irst in <strong>the</strong> Automated Reasoning Systems division.<br />
From 1997 until 2002 he was collaborating<br />
with ITC-Irst on topics related to Artificial Intelligence<br />
Planning and Formal Verification. His research<br />
interest are: integration <strong>of</strong> formal verification techniques<br />
along <strong>the</strong> whole s<strong>of</strong>tware development process,<br />
planning in non-deterministic domains under<br />
different assumptions on run-time observability using<br />
symbolic model checking techniques and symbolic<br />
model checking.
The PLANET <strong>Newsletter</strong> 59<br />
ANNOUNCEMENT<br />
Invitation to Participate in TAC’03 - A Supply Chain Trading<br />
Competition<br />
Authors: R. Arunachalam, N. Sadeh, E. Aurell, J. Eriksson, N. Finne, and S. Janson<br />
Supply chain management is concerned with planning<br />
and coordinating <strong>the</strong> activities <strong>of</strong> organizations<br />
across <strong>the</strong> supply chain, from raw material procurement<br />
to finished goods delivery. In today’s global<br />
economy, effective supply chain management is vital<br />
to <strong>the</strong> competitiveness <strong>of</strong> manufacturing enterprises<br />
as it directly impacts <strong>the</strong>ir ability to meet changing<br />
market demands in a timely and cost effective manner.<br />
With annual worldwide supply chain transactions<br />
in <strong>the</strong> trillions <strong>of</strong> dollars, <strong>the</strong> potential impact<br />
<strong>of</strong> performance improvements is tremendous. While<br />
today’s supply chains are essentially static, relying<br />
on long-term relationships among key trading partners,<br />
more flexible and dynamic practices <strong>of</strong>fer <strong>the</strong><br />
prospect <strong>of</strong> better matches between suppliers and customers<br />
as market conditions change. Adoption <strong>of</strong><br />
such practices has however proven elusive, due to <strong>the</strong><br />
complexity <strong>of</strong> many supply chain relationships and<br />
<strong>the</strong> difficulty in effectively supporting more dynamic<br />
trading practices. TAC-03 was designed to capture<br />
many <strong>of</strong> <strong>the</strong> challenges involved in supporting dynamic<br />
supply chain practices, while keeping <strong>the</strong> rules<br />
<strong>of</strong> <strong>the</strong> game simple enough to entice a large number<br />
<strong>of</strong> competitors to submit entries. The game has been<br />
designed jointly by a team <strong>of</strong> researchers from <strong>the</strong><br />
e-Supply Chain Management Lab at Carnegie Mellon<br />
University and <strong>the</strong> Swedish Institute <strong>of</strong> Computer<br />
Science (SICS).<br />
Specifically, TAC-03 features rounds where eight PC<br />
assembly agents compete for customer orders and for<br />
procurement <strong>of</strong> a variety <strong>of</strong> components. Customers<br />
issue requests for quotes and select from quotes submitted<br />
by <strong>the</strong> PC assemblers, based on delivery dates<br />
and prices. The assembly agents are limited by <strong>the</strong><br />
capacity <strong>of</strong> <strong>the</strong>ir assembly lines and have to procure<br />
components from a set <strong>of</strong> possible suppliers. The<br />
game distinguishes between four types <strong>of</strong> components:<br />
CPUs, Mo<strong>the</strong>rboards, Memory Units and Disk<br />
drives. It features a variety <strong>of</strong> components <strong>of</strong> each<br />
type (e.g. different CPUs, different mo<strong>the</strong>rboards,<br />
etc.). Customer demand comes in <strong>the</strong> form <strong>of</strong> requests<br />
for quotes for different types <strong>of</strong> PCs, each requiring<br />
a different combination <strong>of</strong> components.<br />
The PC assembly agents compete over a relatively<br />
long period <strong>of</strong> time during which customer demand<br />
and availability <strong>of</strong> supplies varies according to predefined<br />
stochastic distributions. The aim <strong>of</strong> each competitor<br />
agent (PC assembly agent) is to maximize its<br />
pr<strong>of</strong>it, by (1) competing with o<strong>the</strong>r agents for valuable<br />
customer orders and pr<strong>of</strong>itable supplier commitments,<br />
and (2) managing <strong>the</strong> assembly <strong>of</strong> products to<br />
meet its existing customer delivery commitments.<br />
The game is representative <strong>of</strong> a broad range <strong>of</strong> supply<br />
chain situations. It is challenging in that it requires<br />
agents to concurrently compete in multiple markets<br />
(markets for different components on <strong>the</strong> supply side<br />
and markets for different products on <strong>the</strong> customer<br />
side) with interdependencies and incomplete information.<br />
It allows agents to strategize (e.g. specializing<br />
in particular types <strong>of</strong> products, stocking up components<br />
that are in low supply). To succeed, agents<br />
will have to demonstrate <strong>the</strong>ir ability to react to variations<br />
in customer demand and availability <strong>of</strong> sup-
60 The PLANET <strong>Newsletter</strong><br />
plies, as well as adapt to <strong>the</strong> strategies adopted by<br />
o<strong>the</strong>r competing agents.<br />
We would like to invite you to consider submitting<br />
an entry to <strong>the</strong> competition. This is a unique opportunity<br />
to develop and evaluate supply chain trading<br />
technology in a competitive environment. Entrants<br />
will also be invited to submit articles in an upcoming<br />
book and will benefit from <strong>the</strong> publicity associated<br />
with <strong>the</strong> event in <strong>the</strong> form <strong>of</strong> press coverage and <strong>the</strong><br />
publication <strong>of</strong> an AI magazine article discussing <strong>the</strong><br />
competition.<br />
A detailed game description, including <strong>the</strong> rules<br />
<strong>of</strong> TAC03, will be published on <strong>the</strong> TAC website<br />
(http://www.sics.se/tac/) by late November<br />
2003. The TAC 03 game server and some simple PC<br />
assembly agents will be available for practice games<br />
on <strong>the</strong> TAC website by February 1, 2003. This will<br />
enable prospective agent designers to test and finetune<br />
<strong>the</strong>ir designs by playing practice games. The<br />
competition itself will be played in a format similar<br />
to earlier TAC games with eight agents competing<br />
in each round. Qualification rounds will be<br />
held in May 2003 with <strong>the</strong> finals slated to take place<br />
at IJCAI-03 in Acapulco in August. Stay tuned on<br />
(http://www.sics.se/tac/) for more information.<br />
Raghu Arunachalam and Norman Sadeh<br />
(CMU),<br />
Erik Aurell, Joakim Eriksson, Niclas Finne,<br />
and Sverker Janson (SICS)<br />
JOB OPENING<br />
Postdoctoral and Doctoral Positions at Australian National ICT<br />
Center<br />
The Australian National ICT Center (NICTA) is a<br />
new research institute jointly set up by <strong>the</strong> University<br />
<strong>of</strong> New South Wales (UNSW) and <strong>the</strong> Australian<br />
National University (ANU) with respective nodes in<br />
Sydney and Canberra. It is funded by <strong>the</strong> Australian<br />
Federal and State Governments, in partnership<br />
with <strong>the</strong> two universities and industry. NICTA will<br />
host top-ranked international researchers and graduate<br />
programs, and will cover major areas in computing,<br />
systems and telecommunications.<br />
The NICTA home page is<br />
http://www.nicta.com.au<br />
In it <strong>the</strong>re are links to <strong>the</strong> existing programs (more<br />
will be added in <strong>the</strong> future), and vacant positions.<br />
The activities <strong>of</strong> <strong>the</strong> Knowledge Representation<br />
and Reasoning (KRR) program (http://www.<br />
nicta.com.au/kr.html) are typified by <strong>the</strong><br />
http://www.planet-noe.org<br />
content <strong>of</strong> papers appearing in, say, <strong>the</strong> KRR section<br />
<strong>of</strong> <strong>the</strong> IJCAI proceedings, with an additional focus on<br />
planning and constraints. The program has <strong>the</strong> preexisting<br />
multi-university Knowledge Systems Group<br />
(KSG) as its initial core, but now seeks to expand by<br />
recruiting research personnel and graduate doctoral<br />
students. At <strong>the</strong> moment, post-doctoral fellows and<br />
doctoral students are sought: for information on how<br />
to apply and conditions please follow <strong>the</strong> Positions<br />
Vacant link in <strong>the</strong> NICTA home page.<br />
KRR welcomes preliminary inquiries about o<strong>the</strong>r<br />
levels <strong>of</strong> personnel.<br />
Norman Foo, Maurice Pagnucco (UNSW), Sylvie<br />
Thiebaux (ANU)<br />
Dr. Sylvie Thiebaux Research Fellow, RSISE,<br />
The Australian National University, Canberra ACT<br />
0200, Australia<br />
http://csl.anu.edu.au/˜thiebaux<br />
sylvie.thiebaux@anu.edu.au<br />
Tel: +61 (2) 6125 8678, Fax: +61 (2) 6125 8651
The PLANET <strong>Newsletter</strong> 61<br />
Member List for PLANET<br />
Currently, PLANET has 58 nodes from 15 <strong>European</strong><br />
countries. Sites and contact persons are:<br />
Austria<br />
- XIMES GmbH, Johannes Gärtner,<br />
gaertner@ximes.com<br />
Belgium<br />
- Robonetics NV, Filip Verhaeghe,<br />
filip.verhaeghe@roboentics.com<br />
- Space Applications <strong>Service</strong>s (SAS), Richard<br />
Aked,<br />
ra@sas.be<br />
Cyprus<br />
- University <strong>of</strong> Cyprus, Yannis Dimopoulos,<br />
yannis@cs.ucy.ac.cy<br />
Czech Republic<br />
- Charles University, Praha, Roman Barták,<br />
bartak@kti.mff.cuni.cz<br />
France<br />
- COSYTEC S.A., Abderrahmane Aggoun,<br />
abderrahmane.aggoun@cosytec.com<br />
- ILOG S.A., Philippe Laborie,<br />
laborie@ilog.fr<br />
- Laboratoire d’ Analyse et d’ Architecture des Systemes<br />
(LAAS-CNRS), TCU Robot Planning, Malik<br />
Ghallab,<br />
malik@laas.fr<br />
- Laboratoire d’ Informatique Marseille (LIM-<br />
CNRS), Camilla Schwind,<br />
schwind@lim.univ-mrs.fr<br />
- MASA Group, Emmanuel Chiva,<br />
emmanuel.chiva@masagroup.net<br />
- ONERA Systems Control and Flight Dynamics<br />
Department, TCU On-line Planning and<br />
Scheduling, Gérard Verfaillie,<br />
Gerard.Verfaillie@cert.fr<br />
INFORMATION<br />
- THOMSON-CSF, Simon De Givry,<br />
simon.degivry@thalesgroup.com<br />
Germany<br />
- University <strong>of</strong> Ulm, Coordinating Node, Susanne<br />
Biundo,<br />
biundo@informatik.uni-ulm.de<br />
- Aachen University <strong>of</strong> Technology, Gerhard Lakemeyer,<br />
gerhard@cs.rwth-aachen.de<br />
- University <strong>of</strong> Bonn, Armin Cremers,<br />
abc@informatik.uni-bonn.de<br />
- Bremer Institut für Betriebstechnik und angewandte<br />
Arbeitswissenschaft (BIBA), Frithj<strong>of</strong> Weber,<br />
web@biba.uni-bremen.de<br />
- Darmstadt University <strong>of</strong> Technology, Ulrich<br />
Scholz,<br />
scholz@informatik.tu-darmstadt.de<br />
- German Research Center for Artificial Intelligence<br />
(<strong>DFKI</strong>), Markus Meyer,<br />
meyer@dfki.de<br />
- University <strong>of</strong> Freiburg, Bernhard Nebel,<br />
nebel@informatik.uni-freiburg.de<br />
- Fraunh<strong>of</strong>er - Autonomous intelligent Systems<br />
(AiS), Joachim Hertzberg,<br />
hertzberg@ais.fraunh<strong>of</strong>er.de
62 The PLANET <strong>Newsletter</strong><br />
- Technical University <strong>of</strong> Munich, Michael Beetz,<br />
Michael.Beetz@informatik.tu-muenchen.<br />
de<br />
- Siemens AG, Wendelin Feiten,<br />
wendelin.feiten@mchp.siemens.de<br />
Greece<br />
- Aristotle University <strong>of</strong> Thessaloniki, Ioannis Refanidis,<br />
yrefanid@csd.auth.gr<br />
- Foundation for Research and Technology - Hellas<br />
(ICS-FORTH), Dimitrios Plexousakis,<br />
dp@csi.forth.gr<br />
- National Centre for Scientific Research<br />
”Demokritos”, Constantine Spyropoulos,<br />
costass@iit.demokritos.gr<br />
- Technical University <strong>of</strong> A<strong>the</strong>ns (ICCS), Spyros<br />
Tzafestas,<br />
tzafesta@s<strong>of</strong>tlab.ece.ntua.gr<br />
- Technical University <strong>of</strong> Crete, Manolis<br />
Koubarakis,<br />
manolis@ced.tuc.gr<br />
- University <strong>of</strong> Ioannina, Chrysostomos Stylios<br />
stylios@cs.uoi.gr<br />
Hungary<br />
- Computer and Automation Research Institute<br />
Hungarian Academy <strong>of</strong> Sciences (MTA SZTAKI),<br />
László Monostori,<br />
laszlo.monostori@sztaki.hu<br />
Italy<br />
- DEIS - University <strong>of</strong> Bologna, Paola Mello,<br />
pmello@deis.unibo.it<br />
- University <strong>of</strong> Brescia, Alfonso Gerevini,<br />
gerevini@ing.unibs.it<br />
- DIST - University <strong>of</strong> Genoa, Enrico Giunchiglia,<br />
Enrico@dist.unige.it<br />
- Consiglio Nazionale delle Ricerche - Istituto di<br />
Psicologia (IP-CNR), TCU Aerospace Applications,<br />
Amedeo Cesta,<br />
cesta@ip.rm.cnr.it<br />
http://www.planet-noe.org<br />
- University <strong>of</strong> Perugia, TCU Planning & Scheduling<br />
for <strong>the</strong> Web, Alfredo Milani,<br />
milani@dipmat.unipg.it<br />
- Istituto per la Ricerca Scientifica e Tecnologia<br />
(IRST), Paolo Traverso,<br />
traverso@irst.itc.it<br />
- University <strong>of</strong> Parma, Agostino Poggi,<br />
poggi@ce.unipr.it<br />
The Ne<strong>the</strong>rlands<br />
- Delft University <strong>of</strong> Technology, Cees Witteveen,<br />
witt@cs.tudelft.nl<br />
- NLR – National Aerospace Laboratory, Henk Hesselink,<br />
hessel@nlr.nl<br />
Portugal<br />
- Instituto Superior de Engenharia do Porto<br />
ISEP/IPP, João Rocha,<br />
jrocha@ipp.pt<br />
SLOVENIA<br />
- University <strong>of</strong> Maribor, Peter Kokol,<br />
kokol@uni-mb.si<br />
Spain<br />
- iSOCO, Intelligent S<strong>of</strong>tware for <strong>the</strong> <strong>Network</strong>ed<br />
Economy, Antonio Reyes Moro,<br />
toni@isoco.com<br />
- Technical University <strong>of</strong> Catalonia, Lluís Vila,<br />
vila@lsi.upc.es<br />
- University <strong>of</strong> Granada, Luis Castillo,<br />
L.Castillo@decsai.ugr.es<br />
- University Carlos III <strong>of</strong> Madrid, TCU Workflow<br />
Management, Daniel Borrajo,<br />
dborrajo@ia.uc3m.es<br />
- Universitat Politècnica de Catalunya, Institut de<br />
Robòtica i Informàtica Industrial, Tom Creemers,<br />
creemers@iri.upc.es<br />
- Universidad Politecnica de Valencia, Eva Onaindia,<br />
onaindia@dsic.upv.es
The PLANET <strong>Newsletter</strong> 63<br />
- Universitat Rovira i Virgili, Tarragona, Miguel<br />
Angel Garcia,<br />
magarcia@etse.urv.es<br />
Sweden<br />
- Linköping University, Patrick Doherty,<br />
patdo@ida.liu.se<br />
- Örebro University, Alessandro Saffiotti,<br />
alessandro.saffiotti@aass.oru.se<br />
United Kingdom<br />
- British Telecommunications, David Lesaint,<br />
david.lesaint@bte.bt.com<br />
- University <strong>of</strong> Durham, Julie Porteous,<br />
j.m.porteous@durham.ac.uk<br />
- University <strong>of</strong> Essex, Sam Steel,<br />
sam@essex.ac.uk<br />
- University <strong>of</strong> Edinburgh, John Levine,<br />
johnl@aiai.ed.ac.uk<br />
- University <strong>of</strong> Huddersfield, TCU Knowledge Engineering,<br />
Lee McCluskey,<br />
t.l.mccluskey@zeus.hud.ac.uk<br />
- University <strong>of</strong> Manchester, Nikolay Mehandjiev,<br />
Nikolay.Mehandjiev@co.umist.ac.uk<br />
- The Open University Walton Hall, Massimiliano<br />
Garagnani,<br />
M.Garagnani@open.ac.uk<br />
- Salford University, TCU Intelligent Manufacturing,<br />
Ruth Aylett,<br />
R.S.Aylett@iti.salford.ac.uk<br />
- Troy Associates Ltd., Vince Long,<br />
vlong@troyassoc.com<br />
Associated Members<br />
- Norman Sadeh, sadeh@cs.cmu.edu, Carnegie<br />
Mellon University<br />
- Peter Jarvis, Jarvis@ai.sri.com, SRI<br />
- Brian Drabble, drabble@cirl.uoregon.edu,<br />
University <strong>of</strong> Oregon<br />
- Sylvie Thiebaux, Sylvie.Thiebaux@anu.<br />
edu.au, The Australian National University<br />
The network is open to new nodes at any time.