Contents

Welcome . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Committees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Sponsors and Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Overview of Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Overview of Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Opening Session . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Plenary and Semi-Plenary Sessions . . . . . . . . . . . . . . . . . . . . . 10
Parallel Sessions: Overview . . . . . . . . . . . . . . . . . . . . . . . . 19
Parallel Sessions: Rooms and Clusters . . . . . . . . . . . . . . . . . . . 20
Parallel Sessions: Abstracts . . . . . . . . . . . . . . . . . . . . . . . . . 31
Index of Keywords . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Speakers, Coauthors and Chairpersons . . . . . . . . . . . . . . . . . . . 176
List of Exhibitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

ISMP2003: Program and Abstracts

(on-line version of printed book)
2

The Logo and the Petersen Graph

The ISMP2003 logo, which is used as From 1859 to 1871 Petersen taught at a contributions. His contributions to graph
cover for this book and additionally prestigious private high school in Copen- theory, inspired by Sylvester, on the fac-
used as symbol in our posters and on hagen. (The headmaster Bohr had a torization of regular graphs, are his best
the homepage etc., is composed of two gifted son Christian, who became pro- known work. He was known for his clar-
main components: the bridge connect- fessor of physiology at the University of ity of exposition, problem-solving skills,
ing Denmark and Sweden, and the Pe- Copenhagen. He in turn had two gifted originality, vigour and sense of humour.
tersen graph. The bridge symbolizes in- sons, Niels Bohr — Nobel prize winner He was bubbling and witty in debate, al-
ternational collaboration as well as engi- in physics — and Harald Bohr — profes- ways ready to brake a lance with any-
neering and mathematical programming sor in mathematics at the University of body, in unfailing trust of the soundness
skills. The Petersen graph is due to the Copenhagen.) of his own arguments. He made a point
Danish mathematician Julius Petersen; out of not reading the writings of other
the following is a short biography writ- mathematicians, not to have his own way
ten by Bjarne Toft, University of Southern of thinking misled or destroyed, some-
Denmark. times with embarrassing consequences.
He is quoted to have said: ’I started to
Julius Petersen (1839–1910) was born in read the proof, but it was eight pages
the Danish town of Sorø on Zealand. His long, so I preferred to do it myself’.
father was a dyer. In 1854 his parents
were no longer able to pay for his school- Petersen’s mind for numbers can be illus-
ing, so he became an apprentice in an trated by the following account, told by
uncle’s grocery store in Jutland. The one of his high school students. To en-
uncle died however when Petersen had tertain the class, Petersen let each pupil
been with him for a year, leaving Pe- write two three digit numbers on the
tersen enough money to return to school. blackboard. He then for a minute or two
In 1857 he began to study engineering
at the Polytechnic School in Copenhagen While teaching at Bohr’s school, Pe- looked at the list of more than 30 num-
(now the Technical University of Den- tersen married, and he and his wife Laura bers. With his back to the blackboard, he
mark), later deciding to concentrate on had three children. He obtained a math- could now repeat the whole list of num-
mathematics. When his inheritance ran ematics degree from the University of bers, forward and backward, and he was
out he had to teach to make a living, Copenhagen in 1866 and a doctorate in able immediately to tell each number’s
and he started a lifelong occupation writ- 1871. In 1871 the Technical University place in the list.
ing textbooks, the first in 1859 on loga- expanded and changed from a two year
rithms. One of his books, ’Methods and to a one year schedule, and Petersen was In 1908 Petersen suffered a stroke. When
Theories for the Solution of Geometri- appointed as its professor of mathemat- his friends pitied his deteriorating condi-
cal Problems’, were translated into eight ics, a position he held for 16 years. In tion he answered them: ’When through-
languages, with the English version last 1887 he changed to a professorship at the out life you have obtained honour and
reprinted in the 1960’ies by Chelsea (in University of Copenhagen. money for enjoying yourself — what
the collection ’String Figures and other more can you ask for ?’. Petersen’s death
Monographs’) and the French version as Petersen worked in a wide range of ar- in 1910 was front page news. A newspa-
recently as 1990, more than a century af- eas, including algebra, analysis, cryptog- per of the time described him as a Hans
ter the first publication date 1866. Pe- raphy, geometry, mechanics, mathemat- Christian Andersen of science — a child
tersen’s books are short, succinct and el- ical economics and number theory. In of the people who made good in the aca-
egant, still a pleasure to read. many of these fields he made pioneering demic world.
 3

Welcome to ISMP2003

On behalf of the Technical University of Denmark, the University of Copenhagen, and

the Mathematical Programming Society we welcome you to ISMP2003, the 18th Interna-
tional Symposium on Mathematical Programming. The symposium is arranged jointly by
the two universities and is held at the campus of the Technical University in Lyngby just
North of Copenhagen.
The conference program includes 17 invited lectures and 750 contributed talks of which
approximately half are included in sessions organized by the participants themselves. The
topics span the complete universe of mathematical programming from theory over mod-
els, modeling tools and solution methods to applications.
We are indebted to a number of persons, companies and institutions for their support of
the symposium. We gratefully acknowledge the contributions from our sponsors, which
made it possible to waive the conference fee for more than 100 participants and to provide
partial support for accommodation expenses for a number of the participating graduate
students.
We are sure that the symposium will initiate and enhance collaboration between mathe-
matical programmers from all over the world. Also, the city of Copenhagen welcomes
ISMP2003, and we encourage all of you to explore the many attractions of the city.
We wish you all an enjoyable ISMP2003 in Copenhagen.

Jens Clausen, Technical University of Denmark

Jørgen Tind, University of Copenhagen
4

Committees

Program Committee

Jørgen Tind (chair), University of Copenhagen; Jens Clausen (cochair), Technical Univer-
sity of Denmark; Rainer Burkard, Technische Universität, Graz; Martin Grötschel, ZIB,
Berlin; Michael J. Todd, Cornell University; Stephen J. Wright, University of Wisconsin.
Local Organizing Committee

Jens Clausen (chair), Technical University of Denmark; Jørgen Tind (cochair), University
of Copenhagen; Hans Bruun Nielsen, Technical University of Denmark; David Pisinger,
University of Copenhagen; Martin Zachariasen, University of Copenhagen.
Nordic Committee

Kim Allan Andersen, University of Aarhus; Erling D. Andersen, EKA Consulting ApS;
Jørgen Bang Jensen, University of Southern Denmark; Claus Carøe, Nordea; Arne Drud,
ARKI Consulting A/S; Anders Forsgren, Royal Institute of Technology, Sweden; Gud-
mund Frandsen, University of Aarhus; Kaj Holmberg, Linköping University; Kurt Jørn-
sten, Norwegian School of Economics and Business Administration; Niklas Kohl, Car-
men Systems AB; Pekka Korhonen, Helsinki School of Economics; Jakob Krarup, Uni-
versity of Copenhagen; Allan Larsen, Technical University of Denmark; Jesper Larsen,
Technical University of Denmark; Per S. Laursen, PriceWaterhouseCoopers; Christos
Levcopolos, Lund Institute of Technology; Per Olov Lindberg, Linköping University;
Kaj Madsen, Technical University of Denmark; Morten Gjølbye Madsen, Carl Bro; Oli
B. G. Madsen, Technical University of Denmark; Kaisa Miettinen, University of Jy-
vaskyla; Teemu Pennanen, Helsinki School of Economics; Niels Chr. Petersen, Univer-
sity of Southern Denmark; Michael Patriksson, Chalmers University of Technology; Sven
Skyum, University of Aarhus; Trond Steihaug, University of Bergen; Anette Vainer, Carl
Bro; Stein W. Wallace, Molde University College; Pawel Winter, University of Copen-
hagen.
 5

Sponsors and Support

We gratefully acknowledge the generous support by our sponsors and supporters.

Sponsors

SAS, Scandinavian Airlines System

Carmen Systems
ARKI Consulting & Development A/S
Springer-Verlag
IBM Research
ILOG
Dash Optimization

Support

Mica Fonden
Thomas B. Thriges Fond
6

Overview of Events

Registration and Get-Together

Sunday, August 17, 17:00 – 20:00, Sportshal, Building 101. Registration is also open on
Monday, August 18, 8:00 – 9:00, Sportshal, Building 101.

Opening Session

Monday, August 18, 9:00 – 10:45, Sportshal, Building 101. Welcome addresses, awarding
of prizes and music by “The Nordic Five”. See full program on page 9.

Plenary Sessions

See program on page 10.

Parallel Sessions

See program on page 19.

Conference Dinner

Tuesday, August 19, Base Camp Restaurant, Halvtolv Bygning 148, Holmen. Boat de-
parture from Nyhavn at 18:00, dinner at 19:00.

Reception

Wednesday, August 20, 19:00 – 21:00. Two parallel receptions have been arranged: One
at the City Hall, Rådhuspladsen, and the other at the University Celebration Hall, Frue
Plads. Your reception ticket indicates in which reception you have been invited to partic-
ipate.

Social Program and Activities

A number of different trips are planned by VanHauen Conferences & Incentives. Also,
some social activities — such as sport activities — have been arranged. For more infor-
mation, please contact VanHauen at the registration desk.
 7

Overview of Program

Monday Tuesday Wednesday Thursday Friday

Opening Parallel Parallel Parallel Parallel
9:00 – 10:30 Sessions Sessions Sessions Sessions
Session TU1 WE1 TH1 FR1
10:30 – 10:45
10:45 – 11:00 Coffee Break
11:00 – 11:15 Plenary Plenary Plenary Plenary
11:15 – 12:00 Plenary
Session Session Session Session
12:00 – 12:15 Session

12:15 – 13:15 Lunch

Parallel Parallel Parallel Parallel

13:15 – 14:45 Sessions Sessions Sessions Sessions
MO2 TU2 WE2 TH2

14:45 – 15:15 Coffee Break

Semi- Semi- Semi- Semi-
15:15 – 16:15 Plenary Plenary Plenary Plenary
Sessions Sessions Sessions Sessions
16:15 – 16:30 Break Break
Parallel Parallel
16:30 – 18:00 Sessions Sessions
MO3 TH3

Dinner Reception

Boat 19:00–21:00
departure
at 18:00
8
 9

Opening Session

Monday, August 18, 9:00 – 10:45, Sportshal, Building 101.

Chair: Jens Clausen, Local Organizing Committee Chair.
Call to Order
Jørgen Tind, Program Committee Chair.
Welcome Address
Lars Pallesen, Rector, DTU.
Chairman’s Message
Robert Bixby, Chairman of the Mathematical Programming Society.
Prizes
Dantzig Prize — for original research having a major impact on mathematical pro-
gramming.
Lagrange Prize — for outstanding works in the area of continuous optimization.
Fulkerson Prize — for outstanding papers in discrete mathematics.
Beale-Orchard-Hays Prize — for excellence in computational mathematical program-
ming.
Tucker Prize — for an outstanding paper by a student.
Music
The Nordic Five: A modern musical performance of legends and tunes from Scandinavia.
10 Plenary and Semi-Plenary Sessions: Monday

Plenary and Semi-Plenary Sessions

Monday, 11:15 – 12:15, Sportshal, Building 101 (Plenary)

In pursuit of the traveling salesman
W ILLIAM J. C OOK
Georgia Institute of Technology
Coauthors: David Applegate, Robert Bixby and Vašek Chvátal

Chair: George L. Nemhauser

Dantzig, Fulkerson, and Johnson (1954) introduced the cutting-plane method as a means
of attacking the traveling salesman problem; this method has been applied to broad classes
of problems in combinatorial optimization and integer programming. In this paper we
discuss an implementation of Dantzig et al.’s method that is suitable for TSP instances
having 1,000,000 or more cities. Our aim is to use the study of the TSP as a step towards
understanding the applicability and limits of the general cutting-plane method in large-
scale applications.

Monday, 15:15 – 16:15, Sportshal, Building 101 (Semi-Plenary)

The smoothed analysis of algorithms
DANIEL A. S PIELMAN
Massachusetts Institute of Technology
Coauthor: Shang-Hua Teng

Chair: Stephen J. Wright

We perform a smoothed analysis of a termination phase for linear programming algo-

rithms. By combining this analysis with the smoothed analysis of Renegar’s condition
number by Dunagan, Spielman and Teng we show that the smoothed complexity of
interior-point algorithms for linear programming is O m3 log m
σ  . In contrast, the best
known bound on the worst-case complexity of linear programming is O m 3 L  , where L
could be as large as m. We include an introduction to smoothed analysis and a tutorial on
proof techniques that have been useful in smoothed analyses.
 11

Monday, 15:15 – 16:15, Aud 81, Building 116 (Semi-Plenary)

Applications of convex optimization in signal processing and digital communication
Z HI -Q UAN L UO
University of Minnesota
Chair: Hans Bruun Nielsen

In the last two decades, the mathematical programming community has witnessed some
spectacular advances in interior point methods and robust optimization. These advances
have recently started to significantly impact various fields of applied sciences and en-
gineering where computational efficiency is essential. This paper focuses on two such
fields: digital signal processing and communication.
In the past, the widely used optimization methods in both fields had been the gradi-
ent descent or least squares methods, both of which are known to suffer from the usual
headaches of stepsize selection, algorithm initialization and local minima. With the re-
cent advances in conic and robust optimization, the opportunity is ripe to use the newly
developed interior point optimization techniques and highly efficient software tools to
help advance the fields of signal processing and digital communication. This paper sur-
veys recent successes of applying interior point and robust optimization to solve some
core problems in these two fields.
The successful applications considered in this paper include adaptive filtering, robust
beamforming, design and analysis of multi-user communication system, channel equal-
ization, decoding and detection. Throughout, our emphasis is on how to exploit the hid-
den convexity, convex reformulation of semi-infinite constraints, analysis of convergence,
complexity and performance, as well as efficient practical implementation.

Monday, 15:15 – 16:15, Aud 42, Building 302 (Semi-Plenary)

Online algorithms: a survey
S USANNE A LBERS
Institut für Informatik, Albert-Ludwigs-Universität Freiburg
Chair: Martin Zachariasen

During the last 15 years online algorithms have received considerable research interest.
In this survey we give an introduction to the competitive analysis of online algorithms
and present important results. We study interesting application areas and identify open
problems.
12 Plenary and Semi-Plenary Sessions: Tuesday

Tuesday, 11:00 – 12:00, Sportshal, Building 101 (Plenary)

Recent advances in the solution of quadratic assignment problems
K URT M. A NSTREICHER
Department of Management Sciences, University of Iowa
Chair: Jens Clausen

The quadratic assignment problem (QAP) is notoriously difficult for exact solution meth-
ods. In the past few years a number of long-open QAPs, including those posed by Stein-
berg (1961), Nugent et al. (1968) and Krarup (1972) were solved to optimality for the
first time. The solution of these problems has utilized both new algorithms and novel
computing structures. We describe these developments, as well as recent work which is
likely to result in the solution of even more difficult instances.

Tuesday, 15:15 – 16:15, Sportshal, Building 101 (Semi-Plenary)

Approximation schemes for NP-hard geometric optimization problems: a survey
S ANJEEV A RORA
Computer Science Department, Princeton University
Chair: David Pisinger

Traveling Salesman, Steiner Tree, and many other famous geometric optimization prob-
lems are NP-hard. Since we do not expect to design efficient algorithms that solve these
problems optimally, researchers have tried to design approximation algorithms, which can
compute a provably near-optimal solution in polynomial time.
We survey such algorithms, in particular a new technique developed over the past few
years that allows us to design approximation schemes for many of these problems. For
any fixed constant c 0, the algorithm can compute a solution whose cost is at most
1 c  times the optimum. (The running time is polynomial for every fixed c 0, and

in many cases is even nearly linear.) We describe how these schemes are designed, and
survey the status of a large number of problems.
 13

Tuesday, 15:15 – 16:15, Aud 81, Building 116 (Semi-Plenary)

How to compute the volume in high dimension?
M IKLÓS S IMONOVITS
Alfréd Rényi Mathematical Institute, Budapest
Chair: Kim A. Andersen

In some areas of theoretical computer science we feel that randomized algorithms are
better and in some others we can prove that they are more efficient than the deterministic
ones.
Approximating the volume of a convex n-dimensional body, given by an oracle is one
of the areas where this difference can be proved. In general, if we use a deterministic algo-
rithm to approximate the volume, it requires exponentially many oracle questions in terms
of n as n ∞. Dyer, Frieze and Kannan gave a randomized polynomial approximation
algorithm for the volume of a convex body K
 n , given by a membership oracle.
The DKF algorithm was improved in a sequence of papers. The area is full of deep
and interesting problems and results. This paper is an introduction to this field and also a
survey.

Tuesday, 15:15 – 16:15, Aud 42, Building 302 (Semi-Plenary)

Variational conditions with smooth constraints: structure and analysis
S TEPHEN M. ROBINSON
University of Wisconsin-Madison
Chair: Kaj Madsen

This is an expository paper about the analysis of variational conditions over sets defined in
finite-dimensional spaces by fairly smooth functions satisfying a constraint qualification.
The primary focus is on results that can provide quantitative and computable sensitivity
information for particular instances of the problems under study, and our objective is to
give a personal view of the state of current knowledge in this area and of gaps in that
knowledge that require future work. The writing style is informal, in keeping with the
objective of focusing the reader’s attention on the basic concepts and the relationships
between them, rather than on details of the particular results themselves.
14 Plenary and Semi-Plenary Sessions: Wednesday

Wednesday, 11:00 – 12:00, Sportshal, Building 101 (Plenary)

Necessary conditions and feedback design in optimal and stabilizing control
F RANCIS C LARKE
University of Lyon 1
Chair: Anders Forsgren

We present necessary conditions of optimality for a general control problem formulated

in terms of a differential inclusion. These conditions unify and substantially extend pre-
vious results in the literature, and also incorporate a ‘stratified’ feature of a novel nature.
When specialized to the calculus of variations, the results yield necessary conditions and
regularity theorems that go significantly beyond the previous standard. They also give rise
to new and stronger maximum principles of Pontryagin type. The principal focus here is
the development of structural criteria on the problem which guarantee a priori that the
necessary conditions apply to any local minimizer.

Wednesday, 15:15 – 16:15, Sportshal, Building 101 (Semi-Plenary)

Progress on perfect graphs
ROBIN T HOMAS
Georgia Institute of Technology
Coauthors: Maria Chudnovsky, Neil Robertson and P. D. Seymour

Chair: Rainer Burkard

A graph is perfect if for every induced subgraph, the chromatic number is equal to the
maximum size of a complete subgraph. The class of perfect graphs is important for several
reasons. For instance, many problems of interest in practice but intractable in general can
be solved efficiently when restricted to the class of perfect graphs. Also, the question of
when a certain class of linear programs always have an integer solution can be answered
in terms of perfection of an associated graph.
In the first part of the paper we survey the main aspects of perfect graphs and their
relevance. In the second part we outline our recent proof of the Strong Perfect Graph
Conjecture of Berge from 1961, the following: a graph is perfect if and only if it has no
induced subgraph isomorphic to an odd cycle of length at least five, or the complement of
such an odd cycle.
 15

Wednesday, 15:15 – 16:15, Aud 81, Building 116 (Semi-Plenary)

First- and second-order methods for semidefinite programming
R ENATO D. C. M ONTEIRO
Georgia Institute of Technology
Chair: Michael J. Todd

In this paper, we survey the most recent methods that have been developed for the solution
of semidefinite programs. We first concentrate on the methods that have been primarily
motivated by the interior point (IP) algorithms for linear programming, putting special
emphasis in the class of primal-dual path-following algorithms. We also survey methods
that have been developed for solving large-scale SDP problems. These include first-order
nonlinear programming (NLP) methods and more specialized path-following IP methods
which use the (preconditioned) conjugate gradient or residual scheme to compute the
Newton direction and the notion of matrix completion to exploit data sparsity.

Wednesday, 15:15 – 16:15, Aud 42, Building 302 (Semi-Plenary)

Stochastic programming with integer variables
RÜDIGER S CHULTZ
University Duisburg-Essen
Chair: Stein W. Wallace

Including integer variables into traditional stochastic linear programs has considerable
implications for structural analysis and algorithm design. Starting from mean-risk ap-
proaches with different risk measures we identify corresponding two- and multi-stage
stochastic integer programs that are large-scale block-structured mixed-integer linear pro-
grams if the underlying probability distributions are discrete. We highlight the role of
mixed-integer value functions for structure and stability of stochastic integer programs.
When applied to the block structures in stochastic integer programming, well known al-
gorithmic principles such as branch-and-bound, Lagrangian relaxation, or cutting plane
methods open up new directions of research. We review existing results in the field and
indicate departure points for their extension.
16 Plenary and Semi-Plenary Sessions: Thursday

Thursday, 11:00 – 12:00, Sportshal, Building 101 (Plenary)

Strong formulations for mixed integer programs: valid inequalities and extended
formulations
L AURENCE A. W OLSEY
Université Catholique de Louvain, Belgium
Chair: Jørgen Tind

We examine progress over the last fifteen years in finding strong valid inequalities and
tight extended formulations for simple mixed integer sets lying both on the “easy" and
“hard" sides of the complexity frontier. Most progress has been made in studying sets
arising from knapsack and single node flow sets, and a variety of sets motivated by dif-
ferent lot-sizing models. We conclude by citing briefly some of the more intriguing new
avenues of research.

Thursday, 15:15 – 16:15, Sportshal, Building 101 (Semi-Plenary)

Capacities of quantum channels and how to find them
P ETER W. S HOR
AT&T Labs – Research
Chair: Kaj Holmberg

We survey what is known about the information transmitting capacities of quantum chan-
nels, and give a proposal for how to calculate some of these capacities using linear pro-
gramming.
 17

Thursday, 15:15 – 16:15, Aud 81, Building 116 (Semi-Plenary)

The mathematics of eigenvalue optimization
A DRIAN S. L EWIS
Simon Fraser University
Chair: Trond Steihaug

Optimization problems involving the eigenvalues of symmetric and nonsymmetric ma-

trices present a fascinating mathematical challenge. Such problems arise often in the-
ory and practice, particularly in engineering design, and are amenable to a rich blend
of classical mathematical techniques and contemporary optimization theory. This essay
presents a personal choice of some central mathematical ideas, outlined for the broad op-
timization community. I discuss the convex analysis of spectral functions and invariant
matrix norms, touching briefly on semidefinite representability, and then outlining two
broader algebraic viewpoints based on hyperbolic polynomials and Lie algebra. Analo-
gous nonconvex notions lead into eigenvalue perturbation theory. The last third of the
article concerns stability, for polynomials, matrices, and associated dynamical systems,
ending with a section on robustness. The powerful and elegant language of nonsmooth
analysis appears throughout, as a unifying narrative thread.

Thursday, 15:15 – 16:15, Aud 42, Building 302 (Semi-Plenary)

Toward fully automated fragments of graph theory
S IEMION FAJTLOWICZ
University of Houston
Chair: Martin Grötschel

A recent review of all claims concerning automated conjectures refers to six programs.
Since then, two more such claims were made. One of them concerns AGX and the other -
Graffiti.PC. Graffiti is the first computer program whose conjectures inspired mathemati-
cians to write papers about them. The program made a number of appealing conjectures
that inspired about eighty papers, but even more significant is that for the first time, math-
ematicians worked on ideas invented by a computer program.
A paper of the author, published at about the same time, addressed aspects of conjecture-
making programs in the context of the Turing test and Penrose’s claim that digital com-
puters cannot have human mathematical intuitions; it is self-evident that attributing to a
program conjecture-making abilities automatically precludes the possibility of confusing
such a program with a program written for the purpose of verification, confirmation or
testing of human-invented conjectures. Still, for the author, the main goal of automa-
tion of the conjecture-making processes is to understand what makes a good conjecture;
everything else is a secondary consideration. This goal is germane to Putnam’s idea of
quasi-empirical mathematics, in which proofs deliberately play a secondary role to obser-
vations.
18 Plenary and Semi-Plenary Sessions: Friday

The same experiment was performed with a chemistry version of the program. Minute-
man is an offshoot of Graffiti for making conjectures about fullerenes and recently about
benzenoids. As a consequence of two conjectures of Minuteman, the author suggested
that the stable fullerenes tend to be good expanders and later that they tend to minimize
their independence numbers. We know now, thanks to extensive computation of Larson
using Fullgen, that for classical fullerenes the independence number is a better statistical
predictor of stability than any of the previously proposed criteria.

Friday, 11:00 – 12:00, Sportshal, Building 101 (Plenary)

Optimization in forestry
M IKAEL RÖNNQVIST
Linköping University and the Forestry Research Institute of Sweden
Chair: P. O. Lindberg

Optimization models and methods have been used extensively in the forest industry. In
this paper we describe the general wood-flow in forestry and a variety of planning prob-
lems. These cover planning periods from a fraction of a second to more than one hundred
years. The problems are modelled using linear, integer and nonlinear models. Solution
methods used depend on the required solution time and include for example dynamic pro-
gramming, LP methods, branch & bound methods, heuristics and column generation. The
importance of modelling and qualitative information is also discussed.
 19

Parallel Sessions: Overview

A parallel session consists of three presentations, and has a duration of 90 minutes. In total
10 time blocks have been scheduled for parallel sessions from Monday to Friday. There
are at most three blocks of parallel sessions each day. They take place at the following
times:
1: 9:00 – 10:30
2: 13:15 – 14:45
3: 16:30 – 18:00

On each day from Monday to Wednesday there are two blocks of parallel sessions, on
Thursday three blocks and on Friday one block of parallel sessions. Please consult page 7
for details; the blocks are coded according to their weekday and time: MO2, MO3, TU1,
TU2 etc.
Up to 26 parallel sessions are taking place simultaneously in each time block. The distri-
bution of clusters (or topics) for all parallel sessions is indicated on the next page.
The rooms are all located in buildings of the 3rd quadrant of DTU. These buildings are
numbered 3xx. The location of these buildings is given on the map on the inside of the
back cover of this book. Within each building, the room names are indicated, e.g., as
302/41 and 308/T1.
20 Parallel Sessions: Overview

Parallel Sessions: Rooms and Clusters

Room Monday Tuesday Wednesday Thursday Friday

2 3 1 2 1 2 1 2 3 1
302/41
302/42 COMBINATORIAL OPTIMIZATION

302/44
302/45
302/49 INTEGER AND MIXED INTEGER PROGRAMMING

306/31 CLP

306/32
306/33 NONLINEAR PROGRAMMING

306/34
306/35 SOFTWARE

306/36 NONSMOOTH PAR NONSMOOTH OPTIMIZATION

306/37 CONVEX PROGRAMMING

306/38 PAR GLOBAL OPTIMIZATION

308/11 AIRLINE OPT. GENERALIZED CONVEXITY / MONOTONICITY

308/12 STOCHASTIC PROGRAMMING

308/13 TUCKER SEMI

308/T1 FINANCE AND ECONOMICS DYNAMIC PROGRAMMING

308/T2 COMPLEMENTARY AND VARIATIONAL INEQUALITIES

308/T3 MODELING LANG . AND SYSTEMS ON - LINE OPT. TELECOMMUNICATION NETWORK DESIGN

341/21 INTERIOR POINT ALGORITHMS

341/22 LOGISTICS AND TRANSPORTATION

341/23 BIO APPROXIMATION ALGORITHMS

321/053 PRODUCTION PAR PRODUCTION AND SCHEDULING

321/033 NETWORKS MULTICRITERIA OPTIMIZATION

321/133 GRAPHS AND MATROIDS LINEAR PROGRAMMING

305/205 GAME THEORY

BIO : BIOINFORMATICS AND OPTIMIZATION

CLP : CONSTRAINT LOGIC PROGRAMMING

PAR : PARALLEL COMPUTING

SEMI : SEMI - INFINITE AND INFINITE DIMENSIONAL PROGRAMMING

 21

Room
Page Parallel Sessions: Monday 13:15 – 14:45 (MO2)
302/41 Submodular functions and discrete convexity I, organizers : Satoru Fujishige and Kazuo Murota, COMBINATORIAL OPTIMIZATION

chair : Satoru Fujishige

p.31 K AZUO M UROTA, Conjugacy relationship between M- A KIYOSHI S HIOURA, Minimization algorithms for M- H ARIHAR NARAYANAN, Polyhedrally tight functions and
convex and L-convex functions in continuous variables convex functions convexity
302/42 Discrete and continuous optimization in VLSI design, organizers : Bernhard Korte and Jens Vygen, COMBINATORIAL OPTIMIZATION

chair : Jens Vygen

p.31 S TEPHAN H ELD, Slack balance algorithms M ARKUS S TRUZYNA, Matrix decompositions and itera- C HRISTIAN S ZEGEDY, Gate and wire sizing using La-
tive solutions of very large quadratic programs grangian relaxation
302/44 Combinatorial optimization I, chair : Bill Cunningham COMBINATORIAL OPTIMIZATION
p.32 A DRIAN Z YMOLKA, Polyhedral investigations of stable G IANPAOLO O RIOLO, Solving stable set problems on su- B ILL C UNNINGHAM, The polytope of even permutations
multi-sets perclasses of interval graphs
302/45 Branch-and-cut approaches, chair : Serigne Gueye COMBINATORIAL OPTIMIZATION
p.32 TAMAS K IS, Polyhedral results and a branch-and-cut al- G EORG K LIEWER, A branch-and-cut system for capaci- S ERIGNE G UEYE, A branch-and-cut algorithm for the
gorithm for project scheduling with variable intensity ac- tated network design graph bipartitioning problem
tivities
302/49 Integer programming column generation I, organizer/chair : Marco Lübbecke INTEGER AND MIXED INTEGER PROGRAMMING
p.33 M IKKEL M ÜHLDORFF S IGURD, Stabilized column gen- E DUARDO U CHOA, Integer program reformulation for F RANCOIS VANDERBECK, Automated Dantzig-Wolfe re-
eration for set partitioning robust branch-and-cut-and-price formulation or how to exploit simultaneaously original
formulation and column generation re-formulation
306/31 Integer programming, organizer/chair : Gabor Pataki INTEGER AND MIXED INTEGER PROGRAMMING
p.34 JANNY L EUNG, Min-up/min-down polytopes DANIEL B IENSTOCK, Subset-algebra lift operators for BALA K RISHNAMOORTHY, The complexity of branch-
0/1 integer programming and-bound
306/32 Nonlinear programming, chair : Stanislav Zakovic NONLINEAR PROGRAMMING
p.34 M ARK A BRAMSON, Applying a mixed variable filter pat- S UNYOUNG K IM, Numerical stability of path tracing in S TANISLAV Z AKOVIC, An interior point algorithm for
tern search algorithm to thermal insulation system design polyhedral homotopy continuation methods worst-case optimisation

306/33 Nonlinear systems, chair : Per Åke Wedin NONLINEAR PROGRAMMING

p.34 M ICHEL B IERLAIRE, A new class of robust methods to H IDEFUMI K AWASAKI, Conjugate sets for a nonlinear P ER Å KE W EDIN, Primal and dual optimization prob-
solve noisy systems of nonlinear equations programming problem lems in tikhonov regularization
306/34 New trust region algorithms, organizer/chair : Jorge Nocedal NONLINEAR PROGRAMMING
p.35 R ICHARD WALTZ, An active-set trust-region algorithm M ICHAEL P OWELL , A new algorithm for unconstrained S TEPHEN W RIGHT, A feasible trust-region algorithm for
for nonlinear optimization minimization without derivatives nonlinear optimization
306/35 Recent developments in filter methods I, organizers : Michael Ulbrich, Stefan Ulbrich and Luís N. Vicente, NONLINEAR PROGRAMMING

chair : Michael Ulbrich

p.35 C HARLES AUDET, A filter pattern search method for non- DAVID S HANNO, Filter methods and NLP’s with un- R ENATA S ILVA, On a primal-dual interior-point filter
smooth constrained optimization bounded multipliers method for nonlinear programming
306/36 Recent advances in nonsmooth optimization I, organizers : Jean-Louis Goffin and Jean-Philippe Vial, NONSMOOTH OPTIMIZATION

chair : Jean-Philippe Vial

p.36 J EAN -P HILIPPE V IAL, Proximal ACCPM, a cutting plane Y URII N ESTEROV, Smooth minimization of non-smooth C ORALIE T RIADOU, A bundle method for convex opti-
method for column generation and Lagrangian relaxation: functions mization: implementation and illustrations
application to the p-median problem
306/37 Convex optimization methods, chair : Jean-Baptiste Hiriart-Urruty CONVEX PROGRAMMING
p.36 M ATIAS C OURDURIER, Coupling general penalty W IESLAWA O BUCHOWSKA, Conditions for boundedness J EAN -BAPTISTE H IRIART-U RRUTY, Solving some ma-
schemes for convex programming with the steepest de- and the existence results for quasi-convex programmes trix nearness problems via convex optimization
scent and the proximal point algorithm
306/38 Distributed solution of LP and MIP problems, chair : Richard Van Slyke PARALLEL COMPUTING
p.37 L UCIA D RUMMOND, Towards a grid enabled branch-and- Y UJI S HINANO, Parallelization of a MIP solver in the R ICHARD VAN S LYKE, Distributed computing and the
bound algorithm PUBB2 framework simplex method
308/11 Airline scheduling I, organizer/chair : Stefan Karisch AIRLINE OPTIMIZATION APPLICATIONS
p.38 S ERGEI C HEBALOV, Robust airline crew pairing opti- C URT H JORRING, Solving crew augmented pairing prob- L ENNART B ENGTSSON, Challenges in railway crew pair-
mization lems using column generation ing
308/12 Applications of stochastic programming, organizer : COSP (Chefi Triki), chair : Chefi Triki STOCHASTIC PROGRAMMING
p.38 L AUREANO F. E SCUDERO, FRC-S3, a fix-and-relax co- M ARTE F ODSTAD, Weekly supply chain coordination C HEFI T RIKI, Optimal capacity allocation in multi-
ordination scheme for stochastic sequencing and schedul- with stochastic demand auction electricity markets under uncertainty
ing
308/13 Stability in stochastic programming, chair : Rene Henrion STOCHASTIC PROGRAMMING
p.39 M. E BRU A NGUN, Response surface merthodology with I OANA P OPESCU, Robust mean-covariance solutions for R ENE H ENRION, Hoelder and Lipschitz stability of solu-
stochastic constraints for expensive simulation stochastic optimization tion sets in programs with probabilistic constraints
308/T1 Model building, chair : H.Paul Williams FINANCE AND ECONOMICS
p.39 L UDMILA KOSHLAI, An econometric model of state bud- M IKHALEVICH M IKHAIL, Optimization model of H.PAUL W ILLIAMS , The allocation of shared fixed costs:
get planned technological-structural changes fairness versus efficiency
308/T2 Semidefinite linear complementarity problem, organizers : T. Parthasarathy and Srinivasa R. Mohan, COMPLEMENTARITY AND VARIATIONAL INEQUALITIES

chair : Srinivasa R. Mohan

p.40 M ADHUR M ALIK, Geometrical notions of p and related M UDDAPPA G OWDA, P-properties of linear transforma- S RINIVASA R. M OHAN, Semidefinite complementary
properties in semidefinite linear complementarity prob- tions on Euclidean Jordan algebras cones and faces in SDLCP
lems
308/T3 Modeling languages and systems I, organizer/chair : Robert Fourer MODELING LANGUAGES AND SYSTEMS
p.40 A LKIS VAZACOPOULOS, Solving large scale supply J EAN -P IERRE G OUX, Modeling, optimization and simu- L EONARDO L OPES, A modeling language for stochastic
chain optimization problems lation of stochastic dynamic systems programming
341/21 Computational conic programming, organizers : Mituhiro Fukuda and Masakazu Kojima, INTERIOR POINT ALGORITHMS

chair : Mituhiro Fukuda

p.41 Y IN Z HANG, A non-polyhedral primal active-set ap- M ASAKAZU M URAMATSU, A pivoting structure in M ICHAL KOCVARA, Solving nonconvex SDP problems
proach to SDP second-order cone programming of structural optimization by PENNON
341/22 Facility location, chair : Francisco Saldanha-da-Gama LOGISTICS AND TRANSPORTATION
p.42 H IROAKI M OHRI, Lagrangian approach for hybrid prob- C ARMEN P LESCHIUTSCHNIG, Inverse 1-median prob- F RANCISCO S ALDANHA - DA -G AMA, Location models:
lem of facility location and network design lems in the plane a reformulation by discretization
341/23 Transportation, chair : Uwe Zimmermann LOGISTICS AND TRANSPORTATION
p.42 F RANK C RITTIN, A new approach to solve large-scale M ARTIN J OBORN, Solving large scale integer multicom- U WE Z IMMERMANN, Optimal shunting
systems of nonlinear equations modity network flow problems in one second
321/053 Planning in medical applications, chair : Rembert Reemtsen PRODUCTION AND SCHEDULING
p.43 KONRAD E NGEL, Fast simultaneous angle, wedge, and ATSUKO I KEGAMI , An algorithm for scheduling doctors R EMBERT R EEMTSEN, The solution of nonlinear prob-
beam intensity optimization in inverse radiotherapy plan- for night duty lems of intensity modulated radiation therapy planning by
ning a Lagrangian barrier-penalty algorithm
321/033 Integer programming for design problems, chair : Corinne Feremans NETWORKS
p.43 A DELAIDE C ERVEIRA, Truss topology design with bi- B ERNARD F ORTZ, The 2-edge connected subgraph prob- C ORINNE F EREMANS , The generalized subgraph prob-
nary variables lem with bounded rings lem: complexity, approximability and polyhedra
321/133 Graphs and matroids, chair : Geir Dahl GRAPHS AND MATROIDS
p.44 D OMINGOS M OREIRA C ARDOSO, Graphs with convex- A NNEGRET K ATRIN WAGLER, Perfectness is an elusive G EIR DAHL, Matrices of zeros and ones with given line
QP stability number graph property sums and a zero block
22 Parallel Sessions: Overview

Room
Page Parallel Sessions: Monday 16:30 – 18:00 (MO3)
302/41 Submodular functions and discrete convexity II, organizers : Satoru Fujishige and Kazuo Murota, COMBINATORIAL OPTIMIZATION

chair : Kazuo Murota

p.45 S ATORU I WATA, A fully combinatorial algorithm for sub- B IANCA S PILLE , A Gallai-Edmonds-type structure theo- S ATOKO M ORIGUCHI , Capacity scaling algorithm for
modular function minimization rem for path-matchings scalable M-convex submodular flow problems
302/42 Network flows in VLSI layout, organizers : Bernhard Korte and Jens Vygen, chair : Bernhard Korte COMBINATORIAL OPTIMIZATION
p.45 U LRICH B RENNER, Network flow algorithms in VLSI D IRK M UELLER, Fast flow augmentation in bit-pattern- J ENS V YGEN, Multicommodity flow with congestion
placement represented grid graphs costs
302/44 Combinatorial optimization II, chair : Pasquale Avella COMBINATORIAL OPTIMIZATION
p.45 A NDREI H ORBACH, On the facets and the diameter of the JAVIER L EONARDO M ARENCO, Chromatic scheduling PASQUALE AVELLA, Computational study of large-scale
k-cycle polytope polytopes coming from the bandwidth allocation problem p-median problems
in point-to-multipoint radio access systems
302/45 Set partitioning applications, chair : David M. Ryan COMBINATORIAL OPTIMIZATION
p.46 J UN I MAIZUMI , A column generation approach for dis- K EISUKE H OTTA, Political redistricting: a case study in DAVID M. RYAN, Improved solution techniques for
crete lotsizing and scheduling problem on identical paral- japan multi-period area-based forest harvest scheduling prob-
lel machines lems
302/49 Integer programming column generation II, organizer : Marco Lübbecke, chair : Ralf Borndörfer INTEGER AND MIXED INTEGER PROGRAMMING
p.46 R ICARDO F UKASAWA, Robust branch-and-cut-and-price L UIS M IGUEL TORRES, Online vehicle dispatching for I VANA L JUBIC, Combining column generation with
for the capacitated minimum spanning tree problem service units memetic algorithms
306/31 Integer and mixed integer programming I, chair : Pedro Martins INTEGER AND MIXED INTEGER PROGRAMMING
p.47 ROBERT W EISMANTEL , Reformulation techniques based H IROO S AITO, A polyhedral approach to hub location P EDRO M ARTINS, Capacitated minimum spanning tree
on group relaxations problems problem: revisiting hop-indexed models
306/32 Selection and design, chair : Jens Starke NONLINEAR PROGRAMMING
p.48 DAG H AUGLAND, Local search methods for the subset L AURA N ICOLETA H EINRICH -L ITAN, Computing the J ENS S TARKE, Surface design for heterogeneous catalytic
selection problem with minimum unit norm center of area of a convex polygon reactions
306/33 Noisy problems, chair : Antanas Zilinskas NONLINEAR PROGRAMMING
p.48 M ICHAEL KOKKOLARAS, Analytical target cascading in WALTRAUD H UYER, Stable noisy optimization by branch A NTANAS Z ILINSKAS, One dimensional global opti-
design optimization of hierarchical multilevel systems un- and fit mization in the presence of noise
der uncertainty
306/34 Optimization applications in biology and chemistry, organizer/chair : Todd Munson NONLINEAR PROGRAMMING
p.49 TODD M UNSON, Computing transition states M ICHAEL WAGNER, Design of protein folding potentials BASTIAAN J. B RAAMS, Semidefinite programming for
using mathematical programming electronic structure calculations
306/35 Recent developments in filter methods II, organizers : Michael Ulbrich, Stefan Ulbrich and Luís N. Vicente, NONLINEAR PROGRAMMING

chair : Luís N. Vicente

p.49 S TEFAN U LBRICH, On the superlinear local convergence A NDREAS WAECHTER, An interior-point filter line- E LIZABETH K ARAS, On the convergence of a filter algo-
of a filter-SQP method search method for large-scale nonlinear programming rithm with independent feasibility and optimality phases

306/36 Recent advances in non-smooth optimization II, organizers : Jean-Philippe Vial and Jean-Louis Goffin, NONSMOOTH OPTIMIZATION

chair : Jean-Philippe Vial

p.50 C LOVIS C. G ONZAGA, Using the central path for com- F RÉDÉRIC BABONNEAU, Nonlinear multicommodity S AMIR E LHEDHLI , ACCPM-based branch-and-bound
puting the analytic center of apolytope flow and traffic equilibrium methods for integer programming
306/37 Semidefinite and conic optimization, chair : Karthik Natarajan CONVEX PROGRAMMING
p.50 L AURA W YNTER, Nested optimization for experimental Y URIY Z INCHENKO, The dual cone for the derivative of K ARTHIK NATARAJAN, Optimal mean-variance upper
inference problems the non-negative orthant bounds on the expectation of highest order statistics
306/38 Parallel methods for combinatorial and global optimization, organizer/chair : Jonathan Eckstein PARALLEL COMPUTING
p.50 D IEGO K LABJAN, Parallel constrained shortest path J EFF L INDEROTH, Computational grid implementations T ED R ALPHS, A framework for scalable parallel tree
of branch-and-bound methods for nonconvex quadratic search
programming
308/11 Airline scheduling II, organizer/chair : Stefan Karisch AIRLINE OPTIMIZATION APPLICATIONS
p.51 M AREK O ZANA, Large scale airline crew rostering prob- O LIVIER DU M ERLE, A proximal analytic center cutting C LAUDE P H . M EDARD, Airline crew scheduling: from
lems in practice plane method to solve the Lagrangian relaxation of the planning to operations
aircraft rotation problem
308/12 Real-life applications in stochastic programming, chair : Sin C. Ho STOCHASTIC PROGRAMMING
p.51 N INA L INN U LSTEIN, A stochastic programming model J ÖRGEN B LOMVALL, Optimization based derivative pric- S IN C. H O, Designing routing zones for VRP with
for strategic capacity planning in the metal-working in- ing in incomplete and imperfect markets stochastic demands
dustry
308/13 Stochastic integer programs and network interdiction, organizer : COSP (Maarten H. van der Vlerk), STOCHASTIC PROGRAMMING

chair : Stein W. Wallace

p.52 M AARTEN H. VAN DER V LERK, On six-pack recourse: J OYCE Y EN, A stochastic programming approach to the K EVIN WOOD, A stochastic network-interdiction prob-
convex approximations for a simple mixed-integer re- generalized assignment problem lem: maximizing the probability of sufficient delay
course model
308/T1 Finance and economics, chair : Alan King FINANCE AND ECONOMICS
p.52 S. R AGHAVAN, CAMBO: combinatorial auction using V LADIMIR K RIVONOZHKO, Parametric optimization A LAN K ING, Calibration of options prices to market data
matrix bids with order methods for visualization of frontier in DEA and produc- using stochastic programming
tivity analysis of Russian banks
308/T2 Complementarity and variational inequalities, chair : Dolf Talman COMPLEMENTARITY AND VARIATIONAL INEQUALITIES
p.53 O LGA P INYAGINA, Application of the d-gap function ap- D OLF TALMAN, Perfection and stability of stationary
proach to geheral equlibrium problems in Banach spaces points with applications to noncooperative games

308/T3 Modeling languages and systems II, organizer/chair : Robert Fourer MODELING LANGUAGES AND SYSTEMS
p.53 K ENNETH H OLMSTRÖM, Large-scale optimization with J OHANNES B ISSCHOP, The open outer approximation DAVID G AY, Update on the AMPL/solver interface li-
TOMLAB / MATLAB MINLP solver in AIMMS brary
341/21 Primal-dual algorithms, chair : Mohamed El ghami INTERIOR POINT ALGORITHMS
p.54 A NA PAULA A IRES B ORGES T EIXEIRA, An extension of A NDRE T ITS, Primal-dual interior-point algorithms for M OHAMED E L GHAMI, Primal-dual interior-point meth-
a predictor-corrector method variant from linear program- indefinite quadratic programming ods based on a new class of barrier functions
ming to semidefinite programming
341/22 Facility location problems, chair : Jakob Krarup LOGISTICS AND TRANSPORTATION
p.54 A SGEIR TOMASGARD, Facility location under eco- H ELENE G UNNARSSON, Terminal location integrated JAKOB K RARUP, The p q-active facility location prob-
nomics of scale with ship routing lem
341/23 Vehicle routing, chair : Jens Lysgaard LOGISTICS AND TRANSPORTATION
p.55 PAULA L ORENA Z ABALA, A two-commodity flow ap- V ICKY M AK, New valid inequalities and an exact algo- J ENS LYSGAARD, Exact solution of the CVRP for a given
proach to vehicle routing problem rithm for the ATSP-TW set of seed customers: a branch-and-bound approach
321/053 Sequencing and scheduling, chair : Jacques Desrosiers PRODUCTION AND SCHEDULING
p.55 T HOMAS E PPING, Theoretical and practical aspects of G ÜNTER ROTE, Pursuit-evasion with imprecise target lo- JACQUES D ESROSIERS, A proximal trust-region algo-
color sequencing cation rithm for stabilized column generation
321/033 Constrained paths, chair : Scott Provan NETWORKS
p.55 N JÅL F OLDNES, On hop-constrained walk polytopes N ICOLAS S TIER, Selfish routing in capacitated networks S COTT P ROVAN, A primal-dual algorithm for finding
maximum capacity paths with recourse
321/133 Matroids and antimatroids, chair : Abdón Sanchez-Arroyo GRAPHS AND MATROIDS
p.56 M ASAHIRO H ACHIMORI , Rooted circuits of convex ge- M ASATAKA NAKAMURA, The max-flow min-cut theo- A BDÓN S ANCHEZ -A RROYO, On independence systems
ometries from affine point configurations rem for stems of rooted circuits of affine convex geome- and Rado-Hall theorems
tries
 23

Room
Page Parallel Sessions: Tuesday 9:00 – 10:30 (TU1)
302/41 Submodular functions and discrete convexity III, organizers : Satoru Fujishige and Kazuo Murota, COMBINATORIAL OPTIMIZATION

chair : Satoru Fujishige

p.57 A KIHISA TAMURA, A general two-sided matching mar- FABIO TARDELLA, Substitutes and complements in lin- YOSHIO O KAMOTO, Matroid representation of clique
ket with discrete concave utility functions ear and nonlinear programming complexes
302/42 Steiner trees, organizer/chair : Gunnar W. Klau COMBINATORIAL OPTIMIZATION
p.57 S IAVASH VAHDATI DANESHMAND, Practical algorithms B OTING YANG, A lower bound and test problem genera- J OÃO T ELHADA, On the multi-weighted Steiner tree
for the Steiner problem tor for the group Steiner minimal tree problem problem: a reformulation by intersection
302/44 Combinatorial optimization III, chair : Klaus Michael Wenger COMBINATORIAL OPTIMIZATION
p.58 A DAM L ETCHFORD, Fast algorithms for planar graphs, D IRK T HEIS, Fast computation of a minimum t-cut in an K LAUS M ICHAEL W ENGER, Maximally violated mod-k
Part I: cycles and cuts undirected graph cuts and the capacitated vehicle routing problem
302/45 Machine scheduling, chair : Jose Rafael Correa COMBINATORIAL OPTIMIZATION
p.58 M ARCUS O SWALD, A branch-and-cut approach for cou- G ENRIKH L EVIN, Mathematical models and methods for J OSE R AFAEL C ORREA, Single machine scheduling with
pled task problems balancing production lines with parallel blocks of opera- precedence constraints
tions at workstations
302/49 Integer programming column generation III, organizer : Marco Lübbecke, chair : David M. Ryan INTEGER AND MIXED INTEGER PROGRAMMING
p.59 M ICHELE M ONACI, Models and algorithms for staff R ALF B ORNDÖRFER , Solving duty scheduling problems M ARCUS P OGGI DE A RAGAO, Robust branch-and-cut-
scheduling problems in public transit and-price for the capacitated vehicle routing problem
306/31 Integer and mixed integer programming II, chair : Robert Nauss INTEGER AND MIXED INTEGER PROGRAMMING
p.59 E UGENE Z AK, Simplifying branch-and-price algorithms PABLO A. R EY, Eliminating redundant solutions of com- ROBERT NAUSS, A computational study of elastic and
through lexicographic ordering binatorial integer programs classical generalized assignment problems
306/32 Optimization of engineering systems governed by simulations/PDE, part 1, NONLINEAR PROGRAMMING

organizer/chair : Natalia Alexandrov

p.60 M ARTIN B ENDSOE, Optimization with multiple scales I VAN O LIVEIRA, Reduced-basis techniques for rapid reli- E KATERINA KOSTINA, Robust optimal design of experi-
using wavelet based reduced order models able optimization of systems described by parametric par- ments for DAE and PDE
tial differential equations
306/33 Augmented Lagrangian methods I, chair : Otero Rolando Gárciga NONLINEAR PROGRAMMING
p.61 F REDERIC D ELBOS, Application of an SQP augmented J EAN C HARLES G ILBERT , Finite time identification of OTERO ROLANDO G ÁRCIGA, A strongly convergent
Lagrangian method to a large-scale problem in 3d reflec- active constraints with an augmented Lagrangian algo- augmented Lagrangian method
tion tomography rithm
306/34 Advances in algorithms and software for nonlinear programming, organizer/chair : Stephen Wright NONLINEAR PROGRAMMING
p.61 H ANDE B ENSON, LOQO: an interior-point method for J ORGE N OCEDAL, Improving to nonlinear interior meth- M ICHAEL PAUL F RIEDLANDER, An LCL algorithm for
nonconvex nonlinear programming ods constrained optimization
306/35 Software session, Dash, chair : James Tebboth OTHER
p.62
306/36 Implementation, chair : Andreas Grothey PARALLEL COMPUTING
p.62 K AZUHIRO KOBAYASHI , Parallel implementation of Z ORANA L UZANIN, Newton-like method for special A NDREAS G ROTHEY, Object-oriented parallel interior
metaheuristics for combinatorial optimization problems class of nonlinear systems point solver for structured nonlinear programs
on global computing systems
306/37 Decomposition in convex programming, chair : Kaj Holmberg CONVEX PROGRAMMING

p.62 Y URI L EVIN, A decomposition approach to the mini- Á NGEL M ARÍN, Solving system design using a type of K AJ H OLMBERG, Mean value cross decomposition for
mization of a separable convex function on a convex poly- convex submodels: computational experience nonlinear convex problems
hedron generated by a parallel-consecutive structure
306/38 Global optimization I, chair : Nikolaos Sahinidis GLOBAL OPTIMIZATION
p.63 M IRJAM D UER, Solving fractional programs with J UN - YA G OTOH, Minimal ellipsoid circumscribing a N IKOLAOS S AHINIDIS, Global optimization in informat-
stochastic algorithms polytope defined by a system of linear inequalities ics problems in biology, chemistry, and physics
308/11 Disruption management, organizer/chair : Niklas Kohl AIRLINE OPTIMIZATION APPLICATIONS
p.63 J OHN -PAUL C LARKE, Degradable airline schedules R AJAN BATTA, A branch-and-price approach for opera- DARREN D EJUN H ANG, Fast aircraft schedule recovery
tional aircraft maintenance routing with crew and passenger constraints
308/12 Computational stochastic integer programming, organizer/chair : Andrew Schaefer STOCHASTIC PROGRAMMING
p.64 TAKAYUKI S HIINA, Lagrangian relaxation method for A NDREW S CHAEFER , SPAR: stochastic programming NAN KONG, Two-stage integer programming with
price-based unit commitment problem with adversarial research stochastic integer right-hand sides
308/13 Sampling and control in stochastic programming, chair : Andreas Eichhorn STOCHASTIC PROGRAMMING
p.64 M ATTI KOIVU, Integration quadratures in discretization S EIICHI I WAMOTO, Nondeterministic dynamic program- A NDREAS E ICHHORN, Stochastic integer programming:
of stochastic programs ming limit theorems and confidence intervals
308/T1 Modeling under uncertainty, chair : Ken Kortanek FINANCE AND ECONOMICS
p.65 VADIM A RKIN, Stochastic model of optimal investor be- W LODZIMIERZ O GRYCZAK, On extending the LP com- K EN KORTANEK, How serious are Wets’ et al “serious
havior in real sector putable risk measures to account downside risk zeros curves”?
308/T2 Regularization methods, organizer/chair : Rainer Tichatschke COMPLEMENTARITY AND VARIATIONAL INEQUALITIES
p.66 DAN BUTNARIU, Convergence and stability of a regular- A LFREDO N OEL I USEM, A proximal point method for T IM VOETMANN, Interior point methods for variational
ization method for maximal monotone inclusions convex vector-valued optimization inequalities with weak regularisation
308/T3 Modeling languages and systems III, organizer/chair : Robert Fourer MODELING LANGUAGES AND SYSTEMS
p.66 PATRICK VALENTE , XML representation of mathemati- ROBERT F OURER, Analyzing submissions to the NEOS K AISA M IETTINEN, Interactive, synchronous nimbus
cal programming models for distributed optimisation ap- server for the purpose of recommending solvers method for multiobjective optimization
plications
341/21 Complexity of interior point methods, chair : Robert Freund INTERIOR POINT ALGORITHMS
p.67 S IMING H UANG, Expected number of iterations of inte- P ETRA H UHN, Average complexity of interior point ROBERT F REUND, Comparison of complexity of ipms
rior point algorithms for linear programming methods: the expected number of steps and the ellipsoid method for conic feasibility and opti-
mization
341/22 Applications of optimization in pricing, organizer/chair : Georgia Perakis LOGISTICS AND TRANSPORTATION
p.67 D IMITRIS B ERTSIMAS, A robust optimization approach K ALYAN TALLURI , New math programming models for G EORGIA P ERAKIS, A variational inequality model for
to supply chain management revenue management dynamic pricing under competition
341/23 Vehicle routing with time windows, chair : Oli B.G. Madsen LOGISTICS AND TRANSPORTATION
p.68 M YRNA PALMGREN, Solving truck scheduling problems S TEFAN RØPKE, A general heuristic for vehicle routing O LI B.G. M ADSEN, Vehicle routing with time windows
with branch-and-price problems
321/053 Lot sizing, chair : Hamish Waterer PRODUCTION AND SCHEDULING
p.68 S ERGEI C HUBANOV, A FPTAS for a generalized single- H ALDUN S URAL, A Lagrangean relaxation based H AMISH WATERER , Big bucket lot sizing problems with
item lot-sizing problem branch-and-bound heuristic for lot sizing with setup times changeover times

321/033 Network flow problems, chair : Dorit Hochbaum NETWORKS

p.69 K ATHARINA L ANGKAU, Flows over time with flow- P HILIPPE M AHEY, Solving multicommodity flow prob- D ORIT H OCHBAUM, A faster algorithm for bipartite
dependent transit times lems with separable piecewise convex costs matching and for the maximum flow problem on closure
graphs
321/133 Combinatorial algorithms, organizer/chair : Kathie Cameron GRAPHS AND MATROIDS
p.69 K ATHIE C AMERON, Finding list partitions of graphs J EAN F ONLUPT, Chromatic characterization of biclique JACK E DMONDS, Revisiting oracle methods for well-
cover described polytopes
305/205 Game programming, chair : M. Aránzazu Estévez-Fernández GAME THEORY
p.70 N IELS O LIEMAN, Stability of coalitions in a Nash-cartel A NATOLY A NTIPIN, Extra-proximal method for comput- M. A RÁNZAZU E STÉVEZ -F ERNÁNDEZ, On properties
game: a probability analysis ing Nash equilibria for two-person nonzero-sum game of several refinements of optimal solutions in linear pro-
gramming
24 Parallel Sessions: Overview

Room
Page Parallel Sessions: Tuesday 13:15 – 14:45 (TU2)
302/41 Flows over time and fluid networks, organizer/chair : Martin Skutella COMBINATORIAL OPTIMIZATION
p.71 M ARTIN S KUTELLA, Multicommodity flows over time: L ISA F LEISCHER, Quickest flows over time JAY S ETHURAMAN, Approximately optimal control of
efficient algorithms and complexity fluid networks
302/42 Prize-collecting Steiner trees, organizer/chair : Gunnar W. Klau COMBINATORIAL OPTIMIZATION
p.71 A LEXANDRE C UNHA, A relax and cut algorithm for the C RISTINA F ERNANDES, Primal-dual algorithms for G UNNAR W. K LAU, The fractional prize-collecting
prize collecting Steiner problem in graphs prize-collecting Steiner tree Steiner tree problem on trees
302/44 Combinatorial optimization IV, chair : Francesco Maffioli COMBINATORIAL OPTIMIZATION
p.72 U RIEL G. ROTHBLUM , Vertices, edge-directions and G AUTIER S TAUFFER , On the non-rank facets of the stable F RANCESCO M AFFIOLI, On the problem of finding min-
combinatorial optimization set polytope of claw-free and circulant graphs imum fundamental cycle bases
302/45 Traveling salesman I, chair : Genevieve Benoit COMBINATORIAL OPTIMIZATION
p.72 K EVIN C HEUNG, On the subtour-elimination polytope S YLVIA B OYD, Finding the exact integrality gap for small G ENEVIEVE B ENOIT, Finding violated cut constraints for
travelling salesman problems the STSP using a decomposition approach
302/49 Integer programming column generation IV, organizer : Marco Lübbecke, chair : Jacques Desrosiers INTEGER AND MIXED INTEGER PROGRAMMING
p.73 N IMA G OLBAHARAN, Routing of snowploughs- a col- DAVID S CHINDL , Graph-coloring and linear program- A RISTIDE M INGOZZI , An exact algorithm for period and
umn generation approach ming multi-depot VRP
306/31 Integer and mixed integer programming III, chair : Katalin Meszaros INTEGER AND MIXED INTEGER PROGRAMMING
p.73 A NDREW M ASON, Branch-and-cut for cyclic staff I GOR VASIL’ EV, Computational experience with a cut- K ATALIN M ESZAROS, A fuzzy logic application as a con-
scheduling ting plane algorithm for the university timetabling prob- tribution to solving the bus crew scheduling problem
lem
306/32 Optimization of engineering systems governed by simulations/PDE, part 2, NONLINEAR PROGRAMMING

organizer/chair : Natalia Alexandrov

p.74 J OERG G ABLONSKY, Effective parallel optimization of A NTHONY G IUNTA, Surrogate-based optimization under E RIC N IELSEN, Aerodynamic design optimization using
expensive functions uncertainty: formulations and applications the Navier-Stokes equations
306/33 Augmented Lagrangian methods II, chair : Giampaolo Liuzzi NONLINEAR PROGRAMMING
p.74 E LIJAH P OLAK, An algorithm for generalized semi- E VA KOMAROMI , Entropy-like Lagrangian method as a G IAMPAOLO L IUZZI, Use of a truncated Newton direc-
infinite optimization problems based on augmented La- unifying approach for solving convex programs tion in an augmented Lagrangian framework
grangians
306/34 Stability of stationary points, organizer/chair : Bernd Kummer NONLINEAR PROGRAMMING
p.75 B ERND K UMMER, Strong Lipschitz stability of stationary D IETHARD K LATTE, Strong Lipschitz stability of station- M IKHAIL S OLODOV, Newton methods for optimization
solutions in variational analysis under MFCQ: Part I the ary solutions in variational analysis under MFCQ, Part II: problems without constraint qualifications
general condition application to nonlinear optimization
306/35 Software session, ILOG, organizer/chair : Irvin Lustig OTHER
p.75
306/36 Recent advances in nonsmooth optimization III, NONSMOOTH OPTIMIZATION

organizers : John Mitchell, Jean-Louis Goffin and Jean-Philippe Vial, chair : John Mitchell
p.76 C ESAR B ELTRAN, The face simplex method in the cut- J OHN M ITCHELL , Properties of a cutting plane method FATMA G ZARA, Cut and column generation based on an-
ting plane framework for semidefinite programming alytic centers for large scale integer programming
306/37 Second-order cone programming, chair : Michael Todd CONVEX PROGRAMMING
p.76 Y U X IA, An algorithm for perturbed second-order cone PAUL T SENG, Approximation algorithms for conic pro- M ICHAEL TODD, Distance weighted discrimination: ap-
programs with application to the Steiner minimal tree gram with extreme ray constraints plying SOCP to pattern recognition
problem
306/38 Global optimization problems with special structure, organizer/chair : Marco Locatelli GLOBAL OPTIMIZATION
p.77 YOSHITSUGU YAMAMOTO, Global optimization method M ARCO L OCATELLI , Undominated D.C. decompositions B ERNARDETTA A DDIS, Smoothing local searches: a new
for solving the minimum maximal flow problem of quadratic functions and applications to branch-and- approach to global optimization
bound approaches
308/11 Generalized convexity I, organizer/chair : Laura Martein GENERALIZED CONVEXITY / MONOTONICITY
p.77 L AURA M ARTEIN, Pseudomonotonicity and pseudocon- I MMANUEL M. B OMZE, From Kato’s theorem to a new D IEHARD E RNST PALLASCHKE, Pairs of compact con-
vexity under the Charnes-Cooper transformation class of generalized convex functions? vex sets
308/12 Asset-liability management, chair : Fredrik Altenstedt STOCHASTIC PROGRAMMING
p.78 RONALD H OCHREITER, Application of dependent dis- P ETRI H ILLI, A stochastic programming model for asset F REDRIK A LTENSTEDT , A comparison between stochas-
crete 2-coupling with reduction to financial management and liability management of a finnish pension company tic programming and parametrized policies for asset lia-
bility management
308/13 Tucker prize, chair : Rainer Burkard OTHER
p.78
308/T1 Equilibrium models, chair : Jorge Rivera FINANCE AND ECONOMICS
p.78 P EDRO DANIEL JARA M ORONI, A sequential land use M ONIQUE F LORENZANO, General equilibrium analysis J ORGE R IVERA, The second welfare theorem with public
equilibrium model with endogenous incomes in ordered topological vector spaces goods in general economies
308/T2 Developments in proximal algorithms, organizer/chair : Jonathan Eckstein COMPLEMENTARITY AND VARIATIONAL INEQUALITIES
p.79 R AINER T ICHATSCHKE , Interior proximal methods for PAULO J OSE DA S ILVA E S ILVA, Double regularization B ENAR S VAITER , Separation-based proximal splitting
non-paramonotone variational inequalities proximal methods for complementarity methods
308/T3 Modeling languages and systems IV, chair : Bob Daniel MODELING LANGUAGES AND SYSTEMS
p.79 A RNOLD N EUMAIER , Testing global optimization soft- B OB DANIEL, Xpress-mosel: a modular environment for
ware modelling and solving optimization problems
341/21 Second order cone programming, chair : Farid Alizadeh INTERIOR POINT ALGORITHMS
p.79 C HEK B ENG C HUA, The primal-dual second-order cone G ONGYUN Z HAO, On treating second order cone prob- FARID A LIZADEH , The q method for optimization over
approximations algorithm lem as a special case of semidefinite problem symmetric cones
341/22 Design in communication and allocation, chair : Chung-Piaw Teo LOGISTICS AND TRANSPORTATION
p.80 P ETER B ROSTROM , Internet protocol network design and M USTAFA C. P INAR, Design of virtual private networks C HUNG -P IAW T EO, Berth allocation planning optmiza-
routing under aggregate traffic uncertainty tion in container terminal
341/23 Optimization in medicine I, organizer/chair : Panos M. Pardalos BIOINFORMATICS AND OPTIMIZATION
p.81 O LVI M ANGASARIAN, Survival-time classification of A RIELA S OFER, Optimal treatment planning for radiofre- E VA L EE, Beam geometry and intensity map optimization
breast cancer patients and chemotherapy quency ablation of liver tumors in IMRT via mixed integer programming
321/053 Parallel computation in optimization: semidefinite programming and related issues, PARALLEL COMPUTING

organizers : Mituhiro Fukuda and Masakazu Kojima, chair : Mituhiro Fukuda

p.81 K ATSUKI F UJISAWA, High performance grid computing M AKOTO YAMASHITA , Parallel computation for K AZUHIDE NAKATA, Parallel semidefinite programming
for mathematical programming semidefinite programming algorithm using matrix completion
321/033 Shortest path problems, chair : Stefan Krause NETWORKS
p.82 I RINA D UMITRESCU, A Lagrangean relaxation approach A NDREAS B LEY, The minimum congestion shortest path S TEFAN K RAUSE, Increasing distances in graphs by
to solving the resource constrained shortest path problem routing problem deleting minimum edge sets
321/133 Graphs and discovery, organizer/chair : Pierre Hansen GRAPHS AND MATROIDS
p.82 D RAGAN S TEVANOVIC , Computer-assisted research in G ILLES C APOROSSI, What forms do interesting conjec- P IERRE H ANSEN, How far should, is and could
graph theory tures have in graph theory? conjecture-making be automated in graph theory?
305/205 Cooperative games, chair : Sjur Didrik Flåm GAME THEORY
p.83 E LENA S ÁIZ P ÉREZ, Determination of stable structures H ERBERT H AMERS, Characterizing convexity of games S JUR D IDRIK F LÅM, Extremal convolution and games
in a multiple coalition game using marginal vectors
 25

Room
Page Parallel Sessions: Wednesday 9:00 – 10:30 (WE1)
302/41 Domination analysis of combinatorial optimization algorithms, organizer/chair : Gregory Gutin COMBINATORIAL OPTIMIZATION
p.84 G REGORY G UTIN, Introduction to domination analysis J OERGEN BANG -J ENSEN, When the greedy algorithm A NDREAS S. S CHULZ, Approximate local search
fails
302/42 Combinatorial optimization V, chair : Marek Libura COMBINATORIAL OPTIMIZATION
p.84 S ANTOSH K ABADI, A strongly polynomial algorithm for L UCAS L ÉTOCART , Minimal multicut and maximal inte- M AREK L IBURA, Adjustment problem for binary con-
integer version of the multipath flow network synthesis ger multiflow in rings strained linear programming problems
problem
302/44 Knapsack problems, organizer/chair : David Pisinger COMBINATORIAL OPTIMIZATION
p.85 DAVID P ISINGER , A survey of upper bounds and exact U LRICH P FERSCHY, A survey of stochastic aspects of H ANS K ELLERER, A survey of approximation algorithms
algorithms for the quadratic knapsack problem knapsack problems for knapsack-type problems
302/45 Traveling salesman II, chair : Nicholas Pearson COMBINATORIAL OPTIMIZATION
p.85 M ICHAL S TERN, On the optimal clustering TSP path and J OHN R AFFENSPERGER , Solving the TSP with N ICHOLAS P EARSON, Fast algorithms for planar graphs,
stars problems decomposition-based pricing Part II: the traveling salesman problem
302/49 Integer programming, organizer/chair : Daniel Bienstock INTEGER AND MIXED INTEGER PROGRAMMING
p.86 M ICHAEL P ERREGAARD, A practical implementation of G EORGE N EMHAUSER , A branch-and-cut algorithm for S ANJEEB DASH, On the use of Gomory’s cyclic group
lift-and-project cuts nonconvex quadratic programming polyhedra in cutting-plane generation
306/31 Integer and mixed integer programming IV, chair : Agostinho Agra INTEGER AND MIXED INTEGER PROGRAMMING
p.86 D UAN L I, Exact solution to separable integer program- D ORIN P REDA, Discrete time LQR optimal control prob- AGOSTINHO AGRA, MIP cuts based on knapsacks with 2
ming lem with logical constraints integer variables
306/32 Optimization of engineering systems governed by simulations/PDE, part 3, NONLINEAR PROGRAMMING

organizer/chair : Natalia Alexandrov

p.87 E MILIO F. C AMPANA, High-fidelity solvers in the design M ARTIN B ERGGREN, Boundary shape optimization for NATALIA A LEXANDROV, Problem formulation and mod-
optimization of ships fluids and waves eling in simulation-based design
306/33 Least square problems, chair : Jerry Eriksson NONLINEAR PROGRAMMING
p.87 T HOMAS V IKLANDS, On global minimization of I RINEL D RAGAN, The least square values and the Shap- J ERRY E RIKSSON, The superiority of using regulariza-
procrustes-penrose regression problems ley value tion over trust-region for ill-conditioned nonlinear least
squares
306/34 Large scale nonlinear programming I, organizers : Sven Leyffer and Richard Waltz, chair : Richard Waltz NONLINEAR PROGRAMMING
p.88 D OMINIQUE O RBAN, An interior-point L1-penalty R ICHARD B YRD, Convergence and constraint activity in ROGER F LETCHER, An f nonmonotonic filter algorithm
method for nonlinear optimization a successive LP algorithm for NLP
306/35 Software session, COIN-OR, organizer/chair : Ted Ralphs OTHER
p.88
306/36 Cutting plane methods for conic optimization problems, NONSMOOTH OPTIMIZATION

organizers : John Mitchell, Jean-Louis Goffin and Jean-Philippe Vial, chair : John Mitchell
p.88 K ARTIK K RISHNAN, Cutting plane methods for semidef- VASILE BASESCU, Analytic center cutting plane method M OHAMMAD R. O SKOOROUCHI , Second-order cone
inite programming in conic programming cuts vs. semidefinite cuts in the analytic center cutting
plane method
306/37 Complexity in convex programming, chair : Stefan M Stefanov CONVEX PROGRAMMING
p.89 F ERNANDO O RDONEZ, Condition number complexity S TEFAN M S TEFANOV, Convex separable optimization
for non-conic convex optimization problems with bounded variables
306/38 Bilevel programming, chair : Anton Evgrafov GLOBAL OPTIMIZATION
p.89 M ARCIA FAMPA, A bilevel model and solution algorithm S OPHIE D EWEZ, Heuristics for toll setting problem A NTON E VGRAFOV, Smoothing by relaxing the equilib-
for the stochastic taxation problem rium conditions in topology optimization problems for
Truss structures in contact
308/11 Generalized convexity II, organizer : Laura Martein, chair : Riccardo Cambini GENERALIZED CONVEXITY / MONOTONICITY
p.90 E VGENY B ELOUSOV, Discontinuity properties of convex R ICCARDO C AMBINI , Mixed type duality for multiobjec-
polynomial mappings tive optimization problems with set constraints
308/12 Modeling and computation issues in stochastic programming, organizer/chair : Gus Gassmann STOCHASTIC PROGRAMMING
p.90 G US G ASSMANN, SMPS and SMPSReader: an input for- C HANDRA P OOJARI, Scalability and implementation is- S HANE DYE, Subtree decomposition for multistage
mat and user routines for stochastic programs sues in stochastic programming algorithms stochastic programs
308/13 Stochastic programming applications in the power and gas industries, STOCHASTIC PROGRAMMING

organizer : COSP (Markus Westphalen), chair : Markus Westphalen

p.91 F RODE RØMO, Value chain management in the liberal- M ARKUS W ESTPHALEN, A two-stage stochastic MILP
ized natural gas market for natural gas transmission optimization with transient
flow
308/T1 Portfolio planning, chair : Gautam Mitra FINANCE AND ECONOMICS
p.91 D IANA ROMAN, A review of risk measures with applica- E RANDA C ELA, Fin4cast: from data mining to optimal G AUTAM M ITRA, Algorithms for the solution of large-
tion in financial portfolio analysis portfolio selection scale quadratic programming (QP) and quadratic mixed
integer programming (qmip) models
308/T2 Complementarity principle in games, organizer/chair : Samir Kumar Neogy COMPLEMENTARITY AND VARIATIONAL INEQUALITIES
p.92 S AMIR K UMAR N EOGY, Complementarity principle and A RUP K UMAR DAS, Some classes of matrices in linear U WE S CHAEFER, The linear complementarity problem
some structured stochastic games complementarity problem and matrix games with interval data
308/T3 Deterministic on-line optimization, chair : Joan Boyar ON - LINE OPTIMIZATION
p.92 W INFRIED H OCHSTÄTTLER , On-line matching on a line PATRIK F LISBERG, On-line bleaching control at pulp J OAN B OYAR, New paging results using the relative worst
mills order ratio
341/21 Methods in semidefinite programming, chair : Henry Wolkowicz INTERIOR POINT ALGORITHMS
p.93 M ITUHIRO F UKUDA, Preprocessing sparse semidefinite PAULO ROBERTO O LIVEIRA, New scaling algorithms H ENRY WOLKOWICZ , Robust search directions for large
programs by the conversion method and a hybrid bundle method for SDP programming sparse semidefinite programming (SDP)
341/22 Practical routing problems, chair : K. Ganesh LOGISTICS AND TRANSPORTATION
p.93 TOBIAS A NDERSSON, Decision support for real-time A RMIN F UEGENSCHUH, Simultaneous optimization of K. G ANESH, Optimization of vehicle routing using evolu-
ambulance planning and control school starting times and public bus services tionary algorithms for contraceptive logistics in a supply
chain
341/23 Optimization in medicine II, organizer/chair : Panos M. Pardalos BIOINFORMATICS AND OPTIMIZATION
p.94 M ICHAEL F ERRIS, Fractionation in radiation treatment WANPRACHA C HAOVALITWONGSE , Optimization tech- E DWIN ROMEIJN, A column generation approach to aper-
planning niques in seizure prediction ture modulation in radiation therapy treatment planning
321/053 Real-life scheduling, chair : Matthew Berge PRODUCTION AND SCHEDULING
p.95 RYUHEI M IYASHIRO, Characterizing feasible pattern sets YOSHIKO I KEBE, Feasibility problems in sports schedul- M ATTHEW B ERGE, Airline schedule recovery - models
with a minimum number of breaks ing and algorithms for air traffic management concept analy-
sis
321/033 Efficiency analysis in multicriteria optimization III, chair : Henri Bonnel MULTICRITERIA OPTIMIZATION
p.95 E MILIO C ARRIZOSA, Inferring efficient weights from S TEFAN T IGAN, Efficiency and generalized concavity in H ENRI B ONNEL, Semivectorial bilevel optimization
pairwise comparison matrices stochastic multiobjective programming problem
321/133 Linear programming and Markov decision, organizer/chair : Yinyu Ye LINEAR PROGRAMMING
p.96 Y INYU Y E, A new complexity result on solving the M ICHAEL O’S ULLIVAN, Polynomial-time computation TAKASHI T SUCHIYA, A new iteration-complexity bound
Markov decision problem of optimal policies in Markov decision chains for the MTY predictor-corrector algorithm
305/205 Optimization in electricity markets, organizer/chair : Andy Philpott GAME THEORY
p.96 G OLBON Z AKERI, Estimation of market distribution A NDY P HILPOTT , Optimization models for hydro- E DWARD A NDERSON, Learning collusive strategies in
functions in electricity pool markets reservoir operations in electricity pools electricity markets
26 Parallel Sessions: Overview

Room
Page Parallel Sessions: Wednesday 13:15 – 14:45 (WE2)
302/41 Computational methods for graph-coloring I, organizer/chair : Michael Trick COMBINATORIAL OPTIMIZATION
p.98 K AZUNORI M IZUNO, A systematic approach to generate S HAHADAT H OSSAIN, Graph-coloring in the estimation M ICHAEL T RICK, Column generation approaches for
very hard 3COL instances of mathematical derivatives graph-coloring generalizations
302/42 Combinatorial optimization VI, chair : Tomomi Matsyu COMBINATORIAL OPTIMIZATION
p.98 TAMON S TEPHEN, The distribution of values in the F ILIPA D UARTE DE C ARVALHO, Exact disclosure pre- TOMOMI M ATSYU, Perfect sampling algorithm for two-
quadratic assignment problem vention in general and nonnegative statistical tables rowed contingency tables
302/44 Networks and communication, chair : Fulvio Piccinonno COMBINATORIAL OPTIMIZATION
p.98 M ARC C. S TEINBACH , Nonlinear optimization in gas N ELSON M ACULAN, New integer linear programming F ULVIO P ICCINONNO , A new map-like algorithm with
network operation formulations for the SRAP low memory requirements for decoding error correcting
codes
302/45 Postman optimization problems, chair : Julián Arturo Aráoz COMBINATORIAL OPTIMIZATION
p.99 F RANCISCO JAVIER Z ARAGOZA M ARTINEZ, Restricted D INO A HR, A branch-and-cut approach for the min-max J ULIÁN A RTURO A RÁOZ, Privatized rural postman prob-
postman problems on mixed graphs k-Chinese postman problem lems
302/49 Decomposition algorithms and dynamic cut generation, organizer/chair : Ted Ralphs INTEGER AND MIXED INTEGER PROGRAMMING
p.100 M ATTHEW G ALATI, Decomposition and dynamic cut A BILIO L UCENA, Optimal rectangular partitions L ASZLO L ADANYI, Solving lexicographic multiobjective
generation in integer programming MIPs with branch-cut-price
306/31 Computational methods for solving IPs I, organizer/chair : Eva Lee INTEGER AND MIXED INTEGER PROGRAMMING
p.100 L ISA E VANS, Generating cutting planes using master A NDREA L ODI, Lagrangean approaches for optimizing M ATHIEU VAN V YVE, Mixing two sets of mixed integer
cyclic group polyehdra over the semimetric polytope inequalities
306/32 Algorithms for degenerate nonlinear optimization, organizer/chair : Olga Brezhneva NONLINEAR PROGRAMMING
p.101 C ATARINA AVELINO, Behavior of NLP codes on a de- C HRISTIAN NAGEL, Quadratic convergence of a nons- O LGA B REZHNEVA, Methods for nonlinear program-
generate optimal control problem mooth Newton-type method for semidefinite programs ming without strict complementarity assumption
306/33 Approximation of gradients, chair : Benoit Hamelin NONLINEAR PROGRAMMING
p.101 S ERGE S HISHKIN, Computationally efficient approxima- R ENTSEN E NKHBAT, Analysis of response surface prob- B ENOIT H AMELIN, Automatic finite differences, an al-
tion of derivatives for robust optimization lems using quadratic programming ternative to automatic derivatives
306/34 Large scale nonlinear programming II, organizers : Sven Leyffer and Richard Waltz, chair : Richard Waltz NONLINEAR PROGRAMMING
p.102 J IE S UN, A robust primal-dual interior point algorithm for A NDERS F ORSGREN, On the solution of linear equations A NNICK S ARTENAER , On interior-point methods using a
nonlinear programs arising in interior methods for nonconvex optimization suitable decomposition for multistage nonlinear stochas-
tic programming
306/35 NLP software - state of the art I, organizer/chair : Arne Stolbjerg Drud NONLINEAR PROGRAMMING
p.102 A RNE S TOLBJERG D RUD, Detecting unboundedness in S TEVEN D IRKSE, Mathematical programs with equilib- M ICHAEL BUSSIECK, Global optimization with GAMS -
practical nonlinear models rium constraints: automatic reformulation and solution applications and performance
via constrained optimization
306/36 Non-convex problems, chair : Michael Overton NONSMOOTH OPTIMIZATION
p.103 WARREN H ARE, Partial smoothness and prox-regularity A LEXEI D EMYANOV, Nonlocal methods for solving M ICHAEL OVERTON, A robust gradient sampling algo-
in optimization some classes of nonsmooth and nonconvex optimization rithm for nonsmooth, nonconvex optimization
problems
306/37 Duality and optimality, chair : Michael L. Flegel CONVEX PROGRAMMING
p.103 A SUMAN O ZDAGLAR, Enhanced optimality conditions ROMAN P OLYAK, Lagrangean transformation in convex M ICHAEL L. F LEGEL , On the Guignard constraint quali-
and informative multipliers for convex programming optimization fication for mathematical programs with equilibrium con-
straints
306/38 Global optimization II, chair : Eligius Maria Theodorus Hendrix GLOBAL OPTIMIZATION
p.104 C HRISTODOULOS F LOUDAS, Deterministic global opti- H ERMANN S CHICHL , Global optimization in the coconut E LIGIUS M ARIA T HEODORUS H ENDRIX, Investigating
mization: a new class of improved convex underestima- project population algorithms; a spatial GA
tors for constrained NLPs
308/11 Generalized convexity III, organizer : Laura Martein, chair : Rita Pini GENERALIZED CONVEXITY / MONOTONICITY
p.105 R ITA P INI, Coercivity conditions and equilibrium prob- DANIELA M ARIAN, Generalized convexity on allures in A NNA M ARCHI, On the connectedness of mix semi-
lems the sense of elena popoviciu efficient frontier
308/12 Option prices and optimization, chair : Teemu Pennanen STOCHASTIC PROGRAMMING
p.105 A NA M ARGARIDA M ONTEIRO, Recovering risk-neutral M ARKKU K ALLIO, Real option valuation via stochastic T EEMU P ENNANEN, Arbitrage pricing of American con-
probabilities from option prices-a convex quadratic pro- optimization tingent claims in incomplete markets
gramming approach
308/13 Stochastic programming with risk measures, organizer/chair : Andrzej Ruszczynski STOCHASTIC PROGRAMMING
p.106 A NDRZEJ RUSZCZYNSKI , Risk aversion via stochastic DARINKA D ENTCHEVA, Risk aversion via stochastic W ERNER ROEMISCH, Risk functionals in stochastic pro-
dominance constraints (Part I) dominance constraints (Part II) gramming: stability and algorithmic issues
308/T1 Combinatorial auctions I, organizer/chair : Susan Powell FINANCE AND ECONOMICS
p.106 R ICHARD S TEINBERG, Pause: a general combinatorial A RNE A NDERSSON, Combinatorial auctions in practice R AKESH VOHRA, On ascending vickrey auctions for het-
auction procedure erogeneous objects
308/T2 Complementarity methods for simulating nonsmooth mechanical systems, organizer/chair : Mihai Anitescu COMPLEMENTARITY AND VARIATIONAL INEQUALITIES
p.106 M IHAI A NITESCU, Optimization problems with comple- J ONG -S HI PANG, Differential variational inequalities X INWEI L IU, A robust sequential quadratic programming
mentarity constraints in mechanics method for mathematical programs with linear comple-
mentarity constraints
308/T3 On-line optimization algorithms, chair : Tjark Vredeveld ON - LINE OPTIMIZATION
p.107 D IANA P OENSGEN, Minimizing the maximum flow time R ENE S ITTERS , The generalized 2-server problem T JARK V REDEVELD, Smoothening helps: a probabilistic
in the online dial-a-ride problem analysis of the multi-level feedback algorithm
341/21 Interior point methods in linear programming, chair : Miguel Argaez INTERIOR POINT ALGORITHMS
p.107 H UA W EI, An iterative method for solving the search di- C ORALIA C ARTIS, On the convergence of interior point M IGUEL A RGAEZ, The role of the notion of quasicentral
rection of linear programming methods for linear programming path for linear programming
341/22 Applications of optimization in transportation, organizer/chair : Georgia Perakis LOGISTICS AND TRANSPORTATION
p.108 T HANASIS Z ILIASKOPOULOS, An inner approximation S OULAYMANE K ACHANI, A fluid dynamics model in S IRIPHONG L AWPHONGPANICH, On second-best toll
algorithm for the dynamic user equilibrium problem distribution systems and transportation pricing problems
341/23 Approximation algorithms for metric facility location and spanning tree, organizer/chair : Yinyu Ye APPROXIMATION ALGORITHMS
p.108 B O C HEN, On universal facility location problem A SAF L EVIN, An efficient polynomial time approxima- M ATTHIAS M UELLER -H ANNEMANN , Approximation of
tion scheme for the constrained minimum spanning tree rectilinear Steiner trees with length restrictions on obsta-
using matroid intersection cles
321/053 Flow shop scheduling, chair : Débora Pretti Ronconi PRODUCTION AND SCHEDULING
p.109 O ULAMARA A MMAR, Scheduling with simultaneous B ERTRAND L IN, A branch-and-bound algorithm for the D ÉBORA P RETTI RONCONI, A branch-and-bound algo-
task processing for two-machine no-wait flowshop prob- minimization of weighted tardiness in a two-machine rithm to minimize the makespan in a flowshop with block-
lem flowshop ing
321/033 Efficiency analysis in multicriteria optimization I, chair : Fernanda Maria Pereira Raupp MULTICRITERIA OPTIMIZATION
p.109 D OLORES ROMERO M ORALES, A biobjective method F RANK H EYDE, Duality in risk minimization and vector F ERNANDA M ARIA P EREIRA R AUPP, On minimization
for sample allocation in stratified sampling optimization over weakly efficient sets
321/133 Computational and algorithmic progress in cone programming, organizer/chair : Jos F. Sturm LINEAR PROGRAMMING
p.110 J OS F. S TURM, A new variant of the interior point method S AM BURER, Semidefinite programming in the space of R EHA T UTUNCU, New developments in SDPT3
with rank-1 updates partial positive semidefinite matrices
305/205 Game theory applications I, organizer/chair : Daniel Granot GAME THEORY
p.110 WALTER K ERN, On the complexity of the kernel M ARILDA S OTOMAYOR , Buying and selling strategies in DANIEL G RANOT, Chinese postman games on a class of
the assignment game Eulerian graphs
 27

Room
Page Parallel Sessions: Thursday 9:00 – 10:30 (TH1)
302/41 Computational methods for graph-coloring II, organizer/chair : Michael Trick COMBINATORIAL OPTIMIZATION
p.112 I GOR D JUKANOVIC , Graph-coloring heuristics based on R ICARDO C ORRÊA, Cliques, holes and lower bounds for I SABEL M ENDEZ -D IAZ , Application of mathematical
lovasz theta number the vertex coloring problem programming to graph-coloring
302/42 Combinatorial optimization VII, chair : Marco Lübbecke COMBINATORIAL OPTIMIZATION
p.112 B ETTINA K LINZ, A fast parametric assignment algorithm A LANTHA N EWMAN, On semidefinite relaxations for the M ARCO L ÜBBECKE , On the stabbing number of match-
with applications in max-algebra linear ordering problem ings, trees, and triangulations
302/44 Counting problems, chair : Takeaki Uno COMBINATORIAL OPTIMIZATION
p.113 YASUKO M ATSUI, An algorithm for enumerating all sub- S AMMANI DANWAWU A BDULLAHI , Counting the ver- TAKEAKI U NO, Fast algorithms for enumerating maximal
stitutions a rooted ordered tree into a term tree tices of certain classes of polyhedra cliques
302/45 Integer and mixed integer programming V, chair : Atsushi Kawamoto INTEGER AND MIXED INTEGER PROGRAMMING
p.113 C ARLO V ERCELLIS, On classes of mixed integer pro- DAVID WARME, So it’s facet-defining... big, fat, hairy ATSUSHI K AWAMOTO, Articulated mechanism designs:
gramming problems arising in discriminant analysis deal! topology and geometry variations
302/49 Alternative methods, chair : Chuangyin Dang INTEGER AND MIXED INTEGER PROGRAMMING
p.114 U TZ -U WE H AUS, An augmentation framework for inte- M ATTHIAS KOEPPE, The integral basis method and ex- C HUANGYIN DANG, A simplicial approach to integer
ger programming tensions programming
306/31 Computational methods for solving IPs II, organizers : Eva Lee and Andrew Miller, chair : Eva Lee INTEGER AND MIXED INTEGER PROGRAMMING
p.114 K EREM A KARTUNALI , Strong formulations and separa- QUENTIN L OUVEAUX, Lifting, superadditivity and sin- D ENISS K UMLANDER, Problems of optimization: an ex-
tion for multi-level lot-sizing problems gle node flow sets revisited act algorithm for finding a maximum clique optimized for
the dense graphs
306/32 Applications, chair : Katia Demaseure NONLINEAR PROGRAMMING
p.115 F RANK S TRAUSS, Shape optimization in rotordynamics O LEG M AKARENKOV, A food problem for a predator- K ATIA D EMASEURE , The progressive addition lens prob-
with eigenvalue and eigenvector constraints prey interaction model lem : an hybrid approach combining optimization and
perturbation techniques
306/33 PDE-constrained optimization (1), organizers : Michael Ulbrich and Stefan Ulbrich, chair : Stefan Ulbrich NONLINEAR PROGRAMMING
p.115 M ATTHIAS H EINKENSCHLOSS , Optimization of time- C ARL K ELLEY, Optimal design of groundwater remedia- M ICHAEL U LBRICH, Multilevel and preconditioning ap-
dependent partial differential equations tion systems proaches for inequality constrained optimization prob-
lems governed by PDEs
306/34 Large scale problems I, chair : Trond Steihaug NONLINEAR PROGRAMMING
p.116 C RISTINA V ILLALOBOS , On the development of a trust- J R . H ERNANDEZ JAIME, An inexact Newton trust region T ROND S TEIHAUG, When sparsity counts: optimal direct
region interior-point method for large-scale nonlinear pro- interior-point algorithm for nonlinear programming prob- Jacobian computation
grams lems
306/35 NLP software - the state of the art II, organizer/chair : Hans D. Mittelmann NONLINEAR PROGRAMMING
p.116 H ANS D. M ITTELMANN, The state of the art in software E RLING A NDERSEN, Functional versus conic optimiza- A RMIN P RUESSNER, Automated performance analysis in
for SDP&SOCP problems tion: what is the better the evaluation of nonlinear programming solvers
306/36 Bundle-based methods, chair : Marko M. Makela NONSMOOTH OPTIMIZATION
p.117 C LAUDIA A LEJANDRA S AGASTIZÁBAL , An infeasible M ARJO H AARALA, Large-scale nonsmooth optimiza- M ARKO M. M AKELA, A hybrid of simulated annealing
bundle method for nonsmooth convex constrained opti- tion: new limited memory bundle method and proximal bundle method for continuous global opti-
mization without a penalty function or a filter mization
306/37 Robust optimization, organizer/chair : Garud Lyengar CONVEX PROGRAMMING
p.117 G ARUD LYENGAR, Robust quadratically constrained A HARON B EN -TAL, Robust optimization approach to M ELVYN S IM, Robust discrete optimization and network
quadratic programming multi-stage uncertain decision problems flows
306/38 Global optimization III, chair : Ivo Nowak GLOBAL OPTIMIZATION
p.118 V INCENZO S CALZO, Relaxation of the semicontinuity M OHIT TAWARMALANI , Convex extensions and polyhe- I VO N OWAK, Rounding and partitioning heuristics for
for marginal functions dral basis solving nonconvex minlps
308/11 Generalized convexity IV, organizer : Laura Martein, chair : Nicolas Hadjisavvas GENERALIZED CONVEXITY / MONOTONICITY
p.118 N ICOLAS H ADJISAVVAS , Quasimonotone variational in- S ANGHO K UM, Generalized vector variational and quasi- A NGELO G UERRAGGIO, Well-posed vector variational
equalities variational inequalities with operator solutions inequalities
308/12 Dynamic stochastic programming: theory and applications, organizer/chair : Lisa Korf STOCHASTIC PROGRAMMING
p.119 L ISA KORF, Dynamic stochastic programming: models, L EONARD M AC L EAN, Risk control of dynamic invest- A LEXEI G AIVORONSKI , Planning of supply chains under
information structures, and duality ment models uncertainty: interworking of stochastic programming and
simulation
308/13 Stochastic programming: a panel discussion, organizer/chair : Gautam Mitra STOCHASTIC PROGRAMMING
p.119
308/T1 Combinatorial auctions II, organizer/chair : Susan Powell FINANCE AND ECONOMICS
p.119 M AKOTO YOKOO, Secure multi-agent dynamic program- T HOMAS E LENDNER, Winner determination in combina- S VEN DE V RIES, Branch-and-price and new testproblems
ming and its application to combinatorial auction torial auctions: Lagrangean heuristic for a generalization for spectrum auctions
of the WJISP
308/T2 Topics in mathematical programs with equilibrium constraints, COMPLEMENTARITY AND VARIATIONAL INEQUALITIES

organizers : Angel-Victor DeMiguel and Michael Paul Friedlander, chair : Michael Paul Friedlander
p.120 A NGEL -V ICTOR D E M IGUEL, A superlinearly conver- C HE -L IN S U, Algorithms for solving EPECs DANIEL R ALPH, Local properties of MPCC methods
gent interior-point method for MPECs
308/T3 Integer programming for the design of ring networks, chair : Rosemary T. Berger TELECOMMUNICATION NETWORK DESIGN
p.120 I RENE L OISEAU, A column generation approach for a H ERNAN A BELEDO, An exact algorithm for optimal rout- ROSEMARY T. B ERGER, Designing stacks of intercon-
network design problem ing and wavelength assignment in ring networks nected bidirectional sonet rings
341/21 Large scale problems, chair : Michael Alan Saunders INTERIOR POINT ALGORITHMS
p.121 S HANGHUA T ENG, Smoothed analysis of condition num- L ETICIA V ELAZQUEZ , A trust region inexact New- M ICHAEL A LAN S AUNDERS, Interior-point solution of
bers of linear programs ton interior-point method for solving large scale uncon- large-scale entropy maximization problems
strained minimization problems
341/22 Railways / public transportation, organizer/chair : Leo Kroon LOGISTICS AND TRANSPORTATION
p.121 L EO K ROON, Rolling stock circulations for passenger D ENNIS H UISMAN, Multiple-depot integrated vehicle R AMON L ENTINK, Shunting passenger train units
trains and crew scheduling
341/23 Algorithms for routing and partitioning, chair : Kimmo Tapio Nieminen APPROXIMATION ALGORITHMS
p.122 YOSHIKO WAKABAYASHI , Approximation algorithms P IERRE F OUILHOUX, The multi-layer constrained via K IMMO TAPIO N IEMINEN, A genetic algorithm for the
for partitioning a graph into connected subgraphs minimization problem: formulation andpolyhedral issues vehicle routing problem

321/053 Scheduling with complex constraints, chair : Chris Potts PRODUCTION AND SCHEDULING
p.122 Y UE W U, A time staged linear programming model for C LIFFORD S TEIN, Scheduling to simultaneously opti- C HRIS P OTTS, Rescheduling for new orders
multi-site aggregate production planning problems mize two objectives
321/033 Planning by multicriteria optimization, chair : Alexander Scherrer MULTICRITERIA OPTIMIZATION
p.123 K ATHRIN K LAMROTH, An MCDM approach to portfolio G ANNA S HYSHKANOVA, Optimization of debt manage- A LEXANDER S CHERRER , Intensity modulated radiation
optimization ment in uncertainty by algorithms on graph therapy - a multicriteria optimization problem
321/133 Aspects of linear programming, chair : Vincent Guigues LINEAR PROGRAMMING
p.124 F.JAVIER TOLEDO -M ELERO, Distance to ill-posedness E TIENNE L OUTE, Gaussian elimination as a computa- V INCENT G UIGUES, Application of robust counterpart
for linear optimization problems: distance to insolvabil- tional paradigm technique to production management
ity
28 Parallel Sessions: Overview

Room
Page Parallel Sessions: Thursday 13:15 – 14:45 (TH2)
302/41 Computational methods for graph-coloring III, organizer/chair : Michael Trick COMBINATORIAL OPTIMIZATION
p.125 A RIE M.C.A. KOSTER, Lightpath coloring in all-optical S TEVEN P RESTWICH, Hybrid local search on two multi-
networks colouring models
302/42 Combinatorial optimization VIII, chair : Sonoko Moriyama COMBINATORIAL OPTIMIZATION
p.125 T RULS F LATBERG, Reconstructing (0,1)-matrices using BALAZS KOTNYEK, A new class of matrices for half- S ONOKO M ORIYAMA, Shellability of simplicial com-
Lagrangean decomposition integrality plexes and optimization in acyclic orientations
302/44 Paths in graphs, chair : L. Sunil Chandran COMBINATORIAL OPTIMIZATION
p.125 E UGENE BARSKY, NP-hardness of some maximum DAVID H UYGENS, Two edge-disjoint hop-constrained L. S UNIL C HANDRAN, Isoperimetric inequalities and the
edge-disjoint paths problem paths and polyhedra width parameters of graphs
302/45 Generalized assignment, chair : Alberto Ceselli INTEGER AND MIXED INTEGER PROGRAMMING
p.126 J OHN W ILSON, An LP-based heuristic procedure for the A NDREAS D REXL, Cut-and-branch for the generalized A LBERTO C ESELLI , A branch-and-price algorithm for
generalized assignment problem with special ordered sets assignment problem the multilevel generalized assignment problem

302/49 Routing in networks, chair : Bruce Shepherd INTEGER AND MIXED INTEGER PROGRAMMING
p.126 P ER O LOV L INDBERG, A prize collecting connected sub- TOMMY T HOMADSEN, The quadratic selective travelling B RUCE S HEPHERD, Maximum multicommodity demand
graph problem salesman problem flows
306/31 Open-source software for mathematical programming, organizer/chair : Robin Lougee-Heimer INTEGER AND MIXED INTEGER PROGRAMMING
p.127 M ATTHEW S ALTZMAN, The COIN-OR open solver in- ROBIN L OUGEE -H EIMER, The COIN-OR initative: C LAUDE M ARTIN TADONKI, Prox-Accpm: a cutting
terface: a progress report open-source tools for mathematical programmers plane solver for convex optimization
306/32 Surrogate modelling for engineering optimization I, organizer/chair : Kaj Madsen NONLINEAR PROGRAMMING
p.127 A NDREW C ONN, On multivariate polynomial interpola- L UÍS N. V ICENTE , On multivariate polynomial interpo- JAN -E RIK K ÄCK, Constrained global optimization with
tion and its use in derivative-free optimization (parts I and lation and its use in derivative-free optimization (parts I radial basis functions
II) and II)
306/33 PDE-constrained optimization (2), organizers : Michael Ulbrich and Stefan Ulbrich, chair : Michael Ulbrich NONLINEAR PROGRAMMING
p.128 F REDI T ROELTZSCH, Optimal control of PDEs with M ICHAEL H INTERMÜLLER, Space mapping for optimal PAUL T. B OGGS, Software for rapid source location in
mixed control-state constraints control of partial differential equations chemical/biological defense
306/34 Large scale problems II, chair : Jacek Gondzio NONLINEAR PROGRAMMING
p.129 M EHDI L ACHIHEB, Approximation of optimum sets of M ASSIMO ROMA, Dynamic scaling based precondition- JACEK G ONDZIO, Parallel solution techniques in finan-
extreme points in a SQP integrated simple decomposition ing for truncated Newton methods in large scale uncon- cial planning problems
strained optization
306/35 Newton-like algorithms I, chair : Oleg Burdakov NONLINEAR PROGRAMMING
p.129 J ELENA N EDIC, Newton-like search directions J IAN -Z HONG Z HANG, Value function and error bounds O LEG BURDAKOV, On limited memory methods with
of a basic trust region model shape changing trust region
306/36 Covering problems, chair : Frits Spieksma NONSMOOTH OPTIMIZATION
p.130 M ICHAEL PATRIKSSON, Global optimality conditions for A DILSON E LIAS X AVIER, Optimum order p covering F RITS S PIEKSMA, An approximation algorithm for an in-
discrete and nonconvex optimiation, with applications to of plane domains by circles via hyperbolic smoothing terval stabbing problem
Lagrangian heuristics and column generation method
306/37 Conic programming and positive polynomials, organizer/chair : Javier Pena CONVEX PROGRAMMING
p.130 L UIS Z ULUAGA, A conic programming approach to gen- E TIENNE DE K LERK, Global optimization of rational J UAN V ERA, Using algebraic geometry in 0-1 program-
eralized Tchebycheff inequalities functions: a semidefinite programming approach ming
306/38 Global optimization IV, chair : Tibor Csendes GLOBAL OPTIMIZATION
p.131 JÁNOS F ÜLÖP, Global optimization methods for approx- YOUKANG FANG, Box region in which the local mini- T IBOR C SENDES, A new interval selection technique for
imation by consistent matrices mizer is global global optimization
308/11 Generalized convexity V, organizer : Laura Martein, chair : Piera Mazzoleni GENERALIZED CONVEXITY / MONOTONICITY
p.131 P IERA M AZZOLENI , Generalized convexity and depen- G ABOR K ASSAY, On generalized equilibrium points DAVIDE L A TORRE, On generalized convexity for set val-
dence structures in equilibrium problems ued maps
308/12 Risk issues in stochastic programming, organizer : COSP (Stephan Tiedemann), chair : Stephan Tiedemann STOCHASTIC PROGRAMMING
p.132 S HABBIR A HMED, Risk modeling in stochastic program- A NDREAS M AERKERT, Risk aversion in stochastic pro- S TEPHAN T IEDEMANN, Risk aversion in stochastic pro-
ming grams with mixed-integer recourse, Part I grams with mixed-integer recourse, Part II
308/T1 Dynamic programming and optimal control, chair : Julien Laurent-Varin DYNAMIC PROGRAMMING AND OPTIMAL CONTROL
p.132 M ORITZ D IEHL, Robust dynamic programming for poly- A NDREI D MITRUK, A relaxation theorem for nonlinear J ULIEN L AURENT-VARIN , An interior points algorithm
topic systems with piecewise linear cost control systems with sliding modes for optimal control problems
308/T2 Mathematical programs with equilibrium constraints, organizers : Sven Leyffer and Richard Waltz, COMPLEMENTARITY AND VARIATIONAL INEQUALITIES

chair : Sven Leyffer

p.133 L ORENZ B IEGLER , Solution and application of MPECs G ABRIEL L OPEZ -C ALVA, Interior point methods for S VEN L EYFFER, Solving complementarity constraints via
in process engineering MPECs nonlinear optimization
308/T3 Scheduling and routing in networks, chair : Hans-Florian Geerdes TELECOMMUNICATION NETWORK DESIGN
p.133 S YLVIE B ORNE, Design of survivable IP-over-optical H ANS -F LORIAN G EERDES, The potential of relaying
networks cellular wireless computer networks
341/21 Recent development in interior-point methods, organizer/chair : Jiming Peng INTERIOR POINT ALGORITHMS
p.134 K EES ROOS, What is special with the logarithmic barrier A KIKO YOSHISE, A homogeneous model for p0 and p TAMAS T ERLAKY, Self-regular proximity based dynamic
function in optimization? nonlinear complementarity problems large-update primal-dual ipms for linear optimization
341/22 Routing by mixed integer programming, chair : Jean-Francois Cordeau LOGISTICS AND TRANSPORTATION
p.134 J ENNY K ARLSSON, Modelling of road investments H ENRIK A NDERSSON, A ship routing problem in the J EAN -F RANCOIS C ORDEAU, A branch-and-cut approach
pulp industry for the dial-a-ride problem
341/23 Approximation and trees, chair : Abdel Lisser APPROXIMATION ALGORITHMS
p.135 V LADIMIR D EINEKO, The double-tree approximation for M ICHEL G OEMANS, Covering minimum spanning trees A BDEL L ISSER, Robust minimum spanning tree
metric TSP: is the best good enough? of induced subgraphs
321/053 Scheduling AND/OR constraints, chair : Valery Gordon PRODUCTION AND SCHEDULING
p.136 B ERIT J OHANNES, Scheduling with or-precedence con- VANESSA K AEAEB, Scheduling with AND/OR-networks VALERY G ORDON, Assignment of due dates in schedul-
straints ing partially ordered jobs
321/033 Convexity and generalized convexity in multicriteria optimization, chair : Andreas Loehne MULTICRITERIA OPTIMIZATION
p.136 K HADIJA K HAZAFI, Efficiency conditions and duality for L UIS M AURICIO G RAÑA D RUMMOND, A projected gra- A NDREAS L OEHNE, On convex functions with values in
multiobjective fractional variational programming prob- dient method for vector optimization semi-linear spaces
lems
321/133 Numerical approaches I, chair : Neelam Gupta LINEAR PROGRAMMING
p.136 G AUTAM A PPA, Solving DEA models by the Fourier G IDEON W EISS, A simplex algorithm for separated con- N EELAM G UPTA, Una: a simple numerical algorithm to
Motzkin elimination method tinuous linear programs solve linear constraints in real variables
 29

Room
Page Parallel Sessions: Thursday 16:30 – 18:00 (TH3)
302/41 Semidefinite liftings in combinatorial optimization, organizer/chair : Miguel Anjos COMBINATORIAL OPTIMIZATION
p.138 M IGUEL A NJOS, Solving the satisfiability problem using A NGELIKA W IEGELE , Semidefinite relaxations for M ONIQUE L AURENT, Semidefinite representations for fi-
semidefinite programming sparse max-cut problems nite varieties
302/42 Combinatorial optimization IX, chair : Safia Kedad-Sidhoum COMBINATORIAL OPTIMIZATION
p.138 J UAN J OSÉ S ALAZAR G ONZÁLEZ , Controlled rounding YOUNG -S OO M YUNG, The graph disconnectivity prob- S AFIA K EDAD -S IDHOUM , A new lower bound for the
and other new statistical disclosure limitation methods for lem: an improved algorithm and new models one machine problem with earliness andtardiness penal-
tabular data ties
302/44 Minimum trees, chair : Francisco Barahona COMBINATORIAL OPTIMIZATION
p.139 C RISTINA R EQUEJO, A 2-paths approach for odd- M ARCUS B RAZIL, Flexibility of Steiner minimum trees F RANCISCO BARAHONA, Network reinforcement
diameter-constrained minimum spanning and Steiner in uniform orientation metrics
trees
302/45 Industrial applications, chair : Ken McKinnon INTEGER AND MIXED INTEGER PROGRAMMING
p.139 J OHN K ARLOF, An integer programming solution to op- B JØRN N YGREEN, Optimal routing of oil and gas from K EN M C K INNON, Branch-and-price to solve a MILP
timally locating radioactive waste facilities wells to separators feed mill problem
302/49 Theory of 0/1 and integer programming, organizer/chair : Friedrich Eisenbrand INTEGER AND MIXED INTEGER PROGRAMMING
p.140 F RIEDRICH E ISENBRAND, Fast integer programming in VOLKER K AIBEL, The graph-density of random 0/1- E RNST A LTHAUS, Point containment in the integer hull
fixed dimension polytopes of a polygon
306/31 Constraint programming, chair : Irvin Lustig CONSTRAINT LOGIC PROGRAMMING
p.140 E RIC B OURREAU, Optimizing TV-break packages of- M IGUEL C ONSTANTINO, A combined mixed integer pro- I RVIN L USTIG, Scheduling the national football league
fered by satellite channels gramming and constraint programming approach to plan- with constraint programming
ning and scheduling in a paint company
306/32 Surrogate modelling for engineering optimization II, organizer/chair : Kaj Madsen NONLINEAR PROGRAMMING
p.140 H ANS B RUUN N IELSEN, Surrogate modelling by Kriging RODRIGUE O EUVRAY, A derivative-free algorithm based D ICK DEN H ERTOG, Gradient estimation using Lagrange
on radial basis functions (RBF) interpolation polynomials
306/33 PDE-constrained optimization (3), organizers : Michael Ulbrich and Stefan Ulbrich, chair : Stefan Ulbrich NONLINEAR PROGRAMMING
p.141 M ICHAEL H INZE, A generalized discretization concept ROLAND B ECKER, Adaptive finite elements for control C HRISTIAN Z ILLOBER , Very large scale nonlinear pro-
for control constrained optimal control problems and its of convection-diffusion equation gramming by scp methods
numerical realization
306/34 Large scale problems III, chair : Ernesto G. Birgin NONLINEAR PROGRAMMING
p.142 G IOVANNI FASANO, Issues on conjugate gradient-type M ARIA H ELENA C AUTIERO H ORTA JARDIM, An E RNESTO G. B IRGIN, Inexact spectral projected gradient
algorithms within truncated Newton methods, in large o n2 log n algorithm for projecting a vector on the inter- methods on convex sets

scale unconstrained optimization section of two hyperplanes and a box in rn

306/35 Newton-like algorithms II, chair : Mehiddin Al-Baali NONLINEAR PROGRAMMING
p.142 H ONGCHAO Z HANG, A nonmonotone line search tech- J EAN -P IERRE D USSAULT, High order Newton-penalty M EHIDDIN A L -BAALI , Quasi-Wolfe conditions for
nique and its application to unconstrained optimization algorithms Newton-like methods
306/36 Nonsmooth optimization, chair : Robert Mifflin NONSMOOTH OPTIMIZATION
p.143 A DIL BAGIROV, Max-min separability and its applica- M ATTHIAS K NOBLOCH, Large-scale Lagrangian decom- ROBERT M IFFLIN, Properties of functions with primal-
tions position problems solved by cutting plane methods dual gradient structure
306/37 Geometry and duality in convex optimization, organizer/chair : Gabor Pataki CONVEX PROGRAMMING
p.143 S HUZHONG Z HANG, The D-induced duality and its ap- JAVIER P ENA, On the block-structured distance to infea- G ABOR PATAKI, Bad semidefinite systems: they look all
plications sibility the same
306/38 Global optimization techniques via reformulations, smoothing approaches and DC programming, GLOBAL OPTIMIZATION

organizer/chair : Tao Pham Dinh

p.144 H OAI A N L E T HI, A difference of convex functions TAO P HAM D INH, DC programming and DCA ap- J IANMING S HI, D.C. and smooth optimization methods
programming approach to image restoration via Markov proaches to variational inequality problems for solving minimum maximal network flow problem
model
308/11 Generalized convexity VI, organizer : Laura Martein, chair : Matteo Rocca GENERALIZED CONVEXITY / MONOTONICITY
p.144 M ARIELLA ROMANIELLO, On a dual method for quasi- M ATTEO ROCCA, Minty variational inequalities and vec- G IOVANNI PAOLO C RESPI, Minty variational inequalities
variational inequalities tor optimization and well-posedness in scalar optimization
308/12 Scenario tree generation, organizer : COSP (David Morton), chair : David Morton STOCHASTIC PROGRAMMING
p.145 DAVID M ORTON, Monte Carlo scenario tree generation S TEIN W. WALLACE, Scenario generation in stochastic M ATTHIAS P ETER N OWAK, Generating scenario trees
for stochastic programming programming - what can we reasonably ask for? for multidimensional stochastic processes
308/13 Complexity in semi-infinite programming, chair : Klaus Meer SEMI - INFINITE AND INFINITE DIMENSIONAL PROGRAMMING
p.145 M ARCO A. L ÓPEZ -C ERDÁ, Distance to ill-posedness for O LIVER S TEIN, An outer approximation method for gen- K LAUS M EER, On the complexity of some problems in
linear semi-infinite inequality systems: applications to eral semi-infinite programming without discretization interval arithmetic
lipschitzian analysis
308/T1 Linear programming approaches to dynamic and stochastic optimization, organizer/chair : Jose Nino-Mora DYNAMIC PROGRAMMING AND OPTIMAL CONTROL
p.146 L ODEWIJK K ALLENBERG, (Parametric) linear program- M ARC U ETZ, Scheduling jobs with uncertain processing J OSE N INO -M ORA, Dynamic allocation indices for rest-
ming for Markovian control problems times on parallel machines less projects and queueing admission control: a polyhe-
dral approach
308/T2 Linear complementarity problems, chair : Vitalij Zhadan COMPLEMENTARITY AND VARIATIONAL INEQUALITIES
p.147 S IMON S CHURR, On polyhedral games, linear comple- D EFENG S UN, Nonsmooth matrix valued functions de- V ITALIJ Z HADAN, Finite Newton’s methods for LP and
mentarity problems, and interior point methods fined by singular values LCP
308/T3 Optimization problems in wireless telecommunications, organizer/chair : Federico Malucelli TELECOMMUNICATION NETWORK DESIGN
p.147 P ETER VÄRBRAND, Optimization of pilot power in 3G D I Y UAN, Scheduling of spatial time division multiple F EDERICO M ALUCELLI , Planning of a wireless LAN
mobile networks access in multi-hop radio networks
341/21 Advances in interior point methods, organizer/chair : Tamas Terlaky INTERIOR POINT ALGORITHMS
p.148 M ASAKAZU KOJIMA, Exploiting sparsity in sum of J IMING P ENG, Hybrid primal-dual interior-point methods G ORAN L ESAJA, Introducing interior-point methods into
squares of polynomials based on self-regular functions the first operations research course
341/22 Supply chain optimization, chair : Huifu Xu LOGISTICS AND TRANSPORTATION
p.148 DAVID B REDSTRÖM , Ship routing and scheduling in the C HIA -S HIN C HUNG, Inventory placement problem for a H UIFU X U, Optimal supply functions in electricity mar-
pulp mill industry serial supply chain with the satisficing objective kets with piecewise continuously differentiable profit
functions
341/23 Approximation algorithms for scheduling problems, chair : David Shmoys APPROXIMATION ALGORITHMS
p.149 V ITALY S TRUSEVICH, Approximation schemes for the F LAVIO K EIDI M IYAZAWA, Computational experience DAVID S HMOYS, Approximation algorithms for the joint
open shop scheduling problem with non-availability con- on approximation algorithms for scheduling unrelated replenishment problem
straints machines
321/053 Industrial scheduling problems, chair : Johannes Hatzl PRODUCTION AND SCHEDULING
p.149 J ENS BUCHHOLZ, Optimal machine scheduling in a ship- J ESPER H ANSEN, Industrialising the application of math- J OHANNES H ATZL, Makespan minimization for chemi-
yard ematical models and methods cal batch processes
321/033 Scalarized functions and functionals in multicriteria optimization, chair : Christiane Tammer MULTICRITERIA OPTIMIZATION
p.150 P ETRA W EIDNER, A method for calculating tradeoffs in M ATTHIAS E HRGOTT, Scalarizations in multiobjective C HRISTIANE TAMMER, A new density result in multicri-
multicriteria optimization combinatorial optimization teria optimization
321/133 Numerical approaches II, chair : Karl Heinz Borgwardt LINEAR PROGRAMMING
p.150 W ILLIAM H AGER, A sparse implementation of the LP I STVAN M AROS, An enhanced piecewise linear dual K ARL H EINZ B ORGWARDT , Average-case-behaviour of
dual active set algorithm phase-1 algorithm for the simplex method three convex-hull-algorithms under rotation-symmetry
30 Parallel Sessions: Overview

Room
Page Parallel Sessions: Friday 9:00 – 10:30 (FR1)
302/41 Packing, partitioning and trees, chair : Michal Penn COMBINATORIAL OPTIMIZATION
p.152 G UNTRAM S CHEITHAUER , Phi-functions for complex M IHALY C SABA M ARKOT, The densest packings of 28, M ICHAL P ENN, An approximation algorithm for the
2d-objects 29 and 30 congruent circles in the unit square - a reliable group Steiner network problem
optimality proof
302/42 Combinatorial optimization X, chair : Bjarne Toft COMBINATORIAL OPTIMIZATION
p.152 A LEXANDRE L AUGIER, K-edge-connected graphs and L EONIDAS P ITSOULIS, Bounds for greedy type algo- B JARNE TOFT, Scheduling without waiting periods
positive semidefiniteness rithms for the k-dimensional assignment problem
302/44 Maximum feasible subsystem problem, organizer/chair : Edoardo Amaldi COMBINATORIAL OPTIMIZATION
p.152 M ARC P FETSCH, The maximum feasible subsystem R APHAEL H AUSER, Generalized boundedness of the re- E DOARDO A MALDI, Randomized relaxation methods for
problem and related polyhedra laxation method with an explicit bound in terms of a con- the maximum feasible subsystem problem
dition number
302/45 Improved algorithms for mixed-integer programming, organizer/chair : John W. Chinneck INTEGER AND MIXED INTEGER PROGRAMMING
p.153 J ONATHAN E CKSTEIN, Pivot, cut, and dive: a class of E MILIE DANNA, Exploring relaxation induced neighbor- J OHN W. C HINNECK, Faster MIP solutions through bet-
heuristics for general MIP hoods to improve MIP solutions ter variable ordering
302/49 Integer programming, organizer/chair : Ismael de Farias INTEGER AND MIXED INTEGER PROGRAMMING
p.154 I SMAEL DE FARIAS, Semi-continuous cuts for mixed- A LPER ATAMTURK, A polyhedral approach to robust in- K ENT H ØJ A NDERSEN, Improving the performance of
integer programming teger programming mixed integer Gomory cuts
306/31 Hybrid methods, organizer/chair : John Hooker CONSTRAINT LOGIC PROGRAMMING
p.154 D IMITRIS M AGOS, Constraint and integer programming J OHN H OOKER, A relaxation of the cumulative constraint
for the orthogonal latin squares problem for resource-constrained scheduling
306/32 Surrogate modelling for engineering optimization III, organizer/chair : Kaj Madsen NONLINEAR PROGRAMMING
p.154 K AJ M ADSEN, Motivations and applications of the space J OHN BANDLER, Space mapping: engineering modeling F RANK P EDERSEN, Space mapping optimization using
mapping technique and optimization exploiting surrogates interpolating surrogates
306/33 Semidefinite problems, chair : Joachim Dahl NONLINEAR PROGRAMMING
p.155 H ECTOR R AMIREZ C ABRERA, A global algorithm for M ALICK J EROME, Covariance matrices calibration via J OACHIM DAHL, Gabor frames with optimal time and fre-
nonlinear semidefinite programming semidefinite least-squares quency localization
306/34 Large scale problems IV, chair : Gerardo Toraldo NONLINEAR PROGRAMMING
p.155 M ICHAEL S TINGL, PENNON - a code for large-scale J OAO -L AURO FACO ´, A parallel NPL code for optimal G ERARDO TORALDO, An hybrid algorithm for large-
nonconvex npl and SDP: algorithm, theory, numerial re- control problems with bounded states scale quadratic programs
sults
306/35 Semidefinite programming and nonnegative polynomials, organizer/chair : Etienne de Klerk NONLINEAR PROGRAMMING
p.156 D MITRII V. PASECHNIK, Optimization subject to a fixed Y VAN H ACHEZ, On the rank-reducibility of barrier func- PABLO PARRILO, Sum of squares decompositions for
number of quadratic constraints in polytime tions structured polynomials
306/36 Nonsmooth optimization II, chair : José Herskovits NONSMOOTH OPTIMIZATION
p.157 H RISTO S ENDOV, Twice differentiable spectral functions D OMINIQUE A ZÉ, Characterization of error bounds for J OSÉ H ERSKOVITS, A new feasible directions algorithm
lower semicontinuous functions for nonsmooth convex optimization
306/37 Large-scale semidefinite programming, organizer/chair : Madhu Nayakkankuppam CONVEX PROGRAMMING
p.157 K IM -C HUAN TOH, Solving large scale semidefinite pro- C HRISTOPH H ELMBERG, A conic bundle approach to lin- M ADHU NAYAKKANKUPPAM, Solving large-scale SDP’s
gramming via Krylov subspace methods ear programming over symmetric cones in parallel
306/38 Quadratic optimization, chair : Andreas Brieden GLOBAL OPTIMIZATION
p.158 J OHANNES J. DE N IJS, Solving nonconvex quadratic pro- C LAUDIO S ODINI, A solution algorithm for a class of box A NDREAS B RIEDEN, Clustering in agriculture by means
gramming problems with simple bound constraints constrained quadratic programming problem of quadratic optimization
308/11 Generalized convexity/monotonicity, chair : Evgeny G. Gol’shtein GENERALIZED CONVEXITY / MONOTONICITY
p.158 O RIZON F ERREIRA, Convex- and monotone- S USANA S CHEIMBERG, A projection-type method with E VGENY G. G OL’ SHTEIN, A method for solving mono-
transformable mathematical programming problems variational metric for solving a general variational in- tone variational inequalities
and a proximal-like point method equality problem
308/12 Sensitivity and stability, chair : Alexey Izmailov STOCHASTIC PROGRAMMING
p.159 V LADIMIR N ORKIN, Nonparametric kernel estimation A RAM A RUTYUNOV, Stability theory for systems of in- A LEXEY I ZMAILOV, Sensitivity analysis for abnormal
via nonstationary stochastic optimization eaqualities at abnormal points optimization problems
308/13 Semi-infinite and infinite dimensional programming, chair : Tatyana Stepanchuk SEMI - INFINITE AND INFINITE DIMENSIONAL PROGRAMMING
p.159 R ADHIA B ESSI, The obstacle problem for water tanks TATYANA S TEPANCHUK, Solution of the optimal set par-
titioning problem as problem of minimization the sets
function
308/T1 Large-scale dynamic programming - algorithms and approximation, organizer/chair : Marina Epelman DYNAMIC PROGRAMMING AND OPTIMAL CONTROL
p.159 I RWIN S CHOCHETMAN, Efficient average-cost optima M ATTHEW BAILEY, Solution acceleration through nested M ARINA E PELMAN, Aggregation in stochastic dynamic
for infinite horizon optimization aggregation programming
308/T2 Nonlinear complementarity problems, chair : Natasa Krejic COMPLEMENTARITY AND VARIATIONAL INEQUALITIES
p.160 J OAO PATRICIO, An investigation of a nonlinear obstacle S ANDRA AUGUSTA S ANTOS, Mixed nonlinear comple- NATASA K REJIC, On a quasi-Newton method for com-
plate optimization problem mentarity problems via nonlinear optimization: numer- plementarity problems
ical results on multi-rigid-body contact problems with
Coulomb friction
308/T3 Models and simulations for planning and control of UMTS (momentum), TELECOMMUNICATION NETWORK DESIGN

organizer/chair : Arie M.C.A. Koster

p.160 A NDREAS E ISENBLÄTTER, Challenges in UMTS radio ROLAND W ESSÄLY, Models for UMTS radio network A LEXANDER M ARTIN, Optimisation methods for UMTS
network planning planning radio network planning
341/21 Second-order cone optimization and extensions, organizer/chair : Erling Andersen INTERIOR POINT ALGORITHMS
p.161 D ON G OLDFARB, The simplex method for semidefinite F RANÇOIS G LINEUR , Second-order cone optimization J EROME O’N EAL, Uniform boundedness of a precondi-
and second-order cone programming with a single second-order cone tioned normal matrix used in interior point methods
341/22 Logistics and transportation, chair : Maria Daneva LOGISTICS AND TRANSPORTATION
p.161 K IN K EUNG L AI, A robust optimization model for empty M ARIA DANEVA, Conjugate direction Frank-Wolfe
container allocations methods with applications to the traffic assignment prob-
lem
341/23 Priority algorithms: models and approximation bounds for greedy-like algorithms, APPROXIMATION ALGORITHMS

organizer/chair : Joan Boyar

p.162 A LLAN B ORODIN, What is a greedy (approximation) al- RUSSELL I MPAGLIAZZO, Models of greedy algorithms K IM S. L ARSEN, Restricted models for priority algo-
grorithm for graph problems rithms on graphs
321/053 Scheduling with variable parameters, chair : Mohamed Haouari PRODUCTION AND SCHEDULING
p.163 NATALIA S HAKHLEVICH, Scheduling problems with M ING L IANG, Optimizing job sequening and tool re- M OHAMED H AOUARI, Optimal parallel machines
controllable processing times and preemption placement decisions scheduling with availability constraints
321/033 Efficiency analysis in multicriteria optimization II, chair : Irina Pospelova MULTICRITERIA OPTIMIZATION
p.163 C ÉSAR G UTIÉRREZ, On a new concept of proper approx- G IANCARLO B IGI, Second order optimality criteria in I RINA P OSPELOVA, Existence conditions of game solu-
imate Pareto efficiency nonsmooth vector optimization tion in multicriteria case
321/133 Mathematical programming techniques in data envelopment analysis, LINEAR PROGRAMMING

organizer/chair : Niels Christian Petersen

p.164 N IELS C HRISTIAN P ETERSEN, A new algorithm for O LE B ENT O LESEN, Identification and use of efficient TOR B ELTOV, How to bootstrap DEA efficiency scores
DEA faces and facets in DEA by using the slice model approach
 31

Parallel Sessions: Abstracts

MO2-302/41 first polynomial-time algorithm is pro-

COMBINATORIAL OPTIMIZATION We give sufficient conditions on the dual
posed by Shioura (1998), where it is generator structures of f g in order that h
Submodular functions and shown that a given vector can be effi- is integral when f g are integral.
discrete convexity I ciently separated from a minimizer of an Our results can be refined in terms of
organizers : Satoru Fujishige and M-convex function. Later, Moriguchi– what may be called ‘legal dual cone’ and
Kazuo Murota Murota–Shioura (2002) showed a prox- ‘legal dual generator’ structures.
chair : Satoru Fujishige imity theorem and proposed a scaling ap-
Conjugacy relationship between proach for M-convex function minimiza-
tion, although the algorithm runs in poly- MO2-302/42 COMBINATORIAL OPTIMIZATION

M-convex and L-convex func- nomial time only for a restricted class Discrete and continuous
tions in continuous variables of M-convex functions. The scaling ap- optimization in VLSI design
K AZUO M UROTA proach of Moriguchi et al. was polished organizers : Bernhard Korte and Jens
University of Tokyo to a polynomial-time algorithm applica- Vygen
ble to general M-convex functions by chair : Jens Vygen
coauthor: Akiyoshi Shioura Tamura (2002). Recently, Shioura (2003)
keywords: combinatorial optimization, proposed two scaling-based simple and Slack balance algorithms
convex function, matroid, submodular fast algorithms for the minimization of
function an M-convex function, both of which are S TEPHAN H ELD
The main theme of Discrete Convex fastest so far for M-convex function min- University of Bonn
Analysis is to extract key combinatorial imization. We survey this progress on keywords: combinatorial optimization,
structures in well-solved nonlinear com- the development of polynomial-time al- VLSI design, slack balancing
binatorial optimization problems, and M- gorithms for M-convex function mini- In the timing analysis of VLSI chips
convex and L-convex functions play the mization. slacks are defined as the differences be-
primary roles. The concepts of M- tween required and real arrival times.
convexity and L-convexity were origi- Polyhedrally tight functions and They act as indicators for the timing-
nally introduced by Murota (1996,1998) convexity criticality of a chip.
for functions on the integer lattice. These
Slack balance problems are a general-
concepts were extended recently to poly- H ARIHAR NARAYANAN
ization of the minimum ratio cycle prob-
hedral convex functions and quadratic EE Dept., IIT Bombay
lem. We want to show how slack bal-
functions in continuous variables by keywords: submodular function,
ance algorithms on digraphs can be used
Murota–Shioura (2000, 2001). To fully convexity, polyhedrally tight functions,
in VLSI-design optimization. The slack
cover the well-solved nonlinear combina- separation theorem
torial optimization problems, it is desir- A set function f : 2S
ℜ, is said to be
balance problem is defined as follows:
We are given a directed graph G with
able to further extend these concepts to polyhedrally tight (dually polyhedrally
arbitrary arc-costs and nonnegative arc-
more general convex functions. In this tight) iff in the set polyhedron (dual set
parameters. G must not have negative
talk, we consider a further extension of polyhedron) denoted by Pf (P f ) defined
cost cycles without positive parameters.
M- and L-convexity to general convex by   For any node potential the parameterized
functions, and show that the conjugacy x X  f X  X  S
relationship, under the Fenchel-Legendre    slack for an arc with positiv parameter is
transformation, between such M-convex x X 
f X  X  S  defined as the ratio between the reduced
costs and the arc parameter. We want to
and L-convex functions. every inequality can be satisfied as an
find a node potential such that the vector
equality. The collection of all dual cones
of these slacks is lexicographically max-
associated with the faces of Pf (Pg ) is de-
Minimization algorithms for M- noted ‘the dual cone structure of f (g)’ imum after sorting in nondecreasing or-
convex functions and the corresponding sets of generators
der.
The special case that all parameters
A KIYOSHI S HIOURA taken from the rows of the coefficient ma-
are either 0 or 1 is already known from
Tohoku University trix of the defining inequalities is called
clock schedule optimization.
‘the dual  generator structure of f (g)’. We
coauthor: Kazuo Murota extend f g  to convex and concave func- In this talk we focus on the general
keywords: combinatorial optimization, case and its applications. Fast Algo-
tions on ℜS by rithms for the special case can be applied
convexity, minimization, matroid

In this talk, we consider the minimiza- fcup c   maxx  Pf cT x to the general case losing the strongly
polynomial running time. We show how
tion of an M-convex function. It is  
gcap c  minx  Pg cT x 
Megiddos concept of optimizing ratio-
a fundamental problem concerning M- nal objective functions can be applied
convex functions, and various algorithms We then have to achieve strongly polynomial running
have been proposed so far. The lo- times.
cal minimality implies the global mini- Theorem If f is polyhedrally tight, g is
mality for M-convex functions. There- dually polyhedrally tight and f
g and
fore, a minimizer of an M-convex func- Pf and Pg have the same dual cone struc- Matrix decompositions and it-
tion can be found by a greedy algorithm, ture then fcup
gcap and there exists a erative solutions of very large
which may require exponential time. The modular function h s.t. f
h
g. quadratic programs
32 Parallel Sessions: Monday 13:15 – 14:45 (MO2)

M ARKUS S TRUZYNA Stable multi-sets represent a generaliza- If we consider a permutation on a set V of

University of Bonn tion of the well-known stable sets. A sta- n elements, as a subset of V  V , then the
keywords: quadratic programming, ble multi-set is an assignment of integers convex hull of all permutations is nicely
VLSI design, iterative methods to the vertices of a graph such that speci- described by the well-known theorem of
fied bounds on vertices and edges are not Birkhoff. Many years ago Alan Hoffman
Iterative methods like steepest descent exceeded. Stable sets are the special case asked what could be said about the con-
and conjugate gradients have been used in which all these bounds equal one. vex hull of even permutations. We de-
for decades for solving large sparse lin- For the stable multi-set polytope, sev- scribe classes of facet inducing inequal-
ear equations systems and quadratic pro- eral valid and facet-defining inequalities ities for this polytope, proving a conjec-
grams, often using various precondition- are known. In this talk, we report on ture of Brualdi and Liu that it cannot be
ing strategies as incomplete Cholesky de- further polyhedral investigations. We described as the solution set of polyno-
composition. present a characterization of the extreme mially many linear inequalities. We also
We consider geometric quadratic pro- points of the linear relaxation polytope. consider the difficulty of the membership
grams arising in VLSI placement, with For the so-called odd-valued odd cycle problem for the polytope, the subject of a
symmetric and positive definite matri- inequalities we specify conditions under second conjecture of Brualdi and Liu.
ces. For these instances we present an which they define facets. Moreover, a
efficient strategy to subdivide extremely polynomial time algorithm for their sep-
large matrices and reduce the number of aration is derived. Finally, we discuss the
MO2-302/45 COMBINATORIAL OPTIMIZATION

nonzero entries. This technique allows results of computational experiments. Branch-and-cut approaches
very fast approximative solutions – for chair : Serigne Gueye
example instances with 2.5 million vari- Solving stable set problems on
ables and over 7 million non-zero entries superclasses of interval graphs Polyhedral results and a
can be solved in less than a minute. This branch-and-cut algorithm for
subdivision strategy permits even paral- G IANPAOLO O RIOLO project scheduling with variable
lel implementation: the matrix is decom- Universita’ Tor Vergata intensity activities
posed into several diagonal blocks which coauthors: Carlo Mannino, Federico TAMAS K IS
can be computed simultaneously and in- Ricci Computer and Automation Research
dependently to obtain further speed-up. keywords: interval graphs, independent Institute
sets, boxicity of a graph, frequency keywords: project scheduling, resource
Gate and wire sizing using La- assignment constraints, integer programming,
grangian relaxation We introduce a graph invariant, called cutting planes
thinness, and show that a maximum  We consider a project scheduling prob-
C HRISTIAN S ZEGEDY weighted stable set on a graph G V E 
Institue for Discrete Mathematics, 
lem in which the total resource require-
University Bonn with thinness k may be found in O Vk  k - ments of the activities are known in ad-
time by dynamic programming, when a vance, but the intensity of each activity,
keywords: convexity, relaxation, certain representation is given. A graph and proportionally its current resource
computer aided design has thinness 1 if and only if it is an inter- usage may be varied over time. Each ac-
tivity has a time window and may require
The goal of gate and wire sizing in VLSI val graph.
We investigate some properties of the several resources. Activities are partially
design is to determine sizes for the tran-
thinness. We show that a graph with ordered by a precedence relation. The re-
sistors and wires of a chip subject to tim-
thinness k is the intersection graph of k- sources are renewable, continuously di-
ing constraints. Convex optimization via
visible, and have finite capacities. Each
the subgradient method and Lagrangian dimensional boxes and discuss connec-
resource has a limit above which its uti-
duality turned out to be an extremely ef- tions with some superclasses of interval
lization incurs a cost. One has to deter-
ficient method to solve this problem op- graphs.
We discuss two applications where mine the intensity of the activities over
timally. In this talk, we are going to
these results naturally apply: the fre- time while respecting all constraints and
present an overview of this method with
quency assignment problem (fap) in minimizing the total cost.
emphasis on practical tricks that allow for
We have shown that the above prob-
finding optimal solutions for large scale GSM networks and the single machine
lem is NP-hard in the strong sense. Hav-
instances with millions of variables very scheduling problem. We show that an ef-
ficient search in exponential neighbour- ing formalized as a mixed integer-linear
quickly.
hoods for the (fap) may be done in poly- program, we solved it by a branch-and-
nomial time by our dynamic program- cut algorithm using the polyhedral re-
MO2-302/44 COMBINATORIAL OPTIMIZATION
ming algorithm. This led us to improv- sults obtained. In fact, a relaxation
Combinatorial optimization I ing the best known solutions for several of MIP led to studying   the follow-
benchmark instances of a relevant test- ing
 polytope P  conv x z   n 
chair : Bill Cunningham
bed. 0 1
q  ∑ j  1 x j  1 0 x j az j j 
n

Polyhedral investigations of sta- 1    q 0 x j a j  q  1    n z j

ble multi-sets The polytope of even permuta- z j  1 j  1    q 1
 . We have ob-
tions tained a complete description of P in
A DRIAN Z YMOLKA terms of facet defining inequalities by an-
Konrad-Zuse-Zentrum fuer B ILL C UNNINGHAM alyzing the cuts of a capacitated network.
Informationstechnik Berlin (ZIB) University of Waterloo We have also designed an efficient sep-
coauthor: Arie M.C.A. Koster coauthor: Yaoguang Wang aration algorithm based on the matroid
keywords: combinatorics, cutting keywords: polyhedra, facets, graphs, greedy algorithm. Computational results
plane/facet, networks/graphs permutations show that our problem formulation and
 33

the generated cuts are really effective in M IKKEL M ÜHLDORFF S IGURD This fact is still little known outside the
solving even large size instances with Dept. of Computer Science, University "column generation community" and its
hundreds of activities. of Copenhagen consequences on integer programming
coauthor: David M. Ryan are just beginning to be explored. This
keywords: column generation, set work intends to be a detailed analysis of
A branch-and-cut system for ca- how to reformulate an integer program in
pacitated network design partitioning
order to build an efficient robust branch-
When applying the column generation and-cut-and-price. In particular, we pro-
G EORG K LIEWER method, one often experiences a seem- pose an alternative master problem that
University of Paderborn ingly random behaviour of the dual vari- can be quite advantageous in some situ-
coauthor: Larissa Timajev ables of the master problem. The dual ations. Another key issue addressed is
keywords: network design, variable values may assume extremely large (pos- how to avoid the pitfalls that arise from
fixing, branch-and-cut, transportation itive or negative) values and may change variable symmetries in the original for-
radically from one iteration to the next. mulations of many problems. We present
We present a Lagrangian relaxation This behaviour is often observed in the extensive computational experiments on
based branch-and-bound system for the beginning of the column generation pro- the capacitated minimum spanning tree,
capacitated network design problem. We cess and it reflects the poor informa- capacitated vehicle routing, and general-
use subgradient optimisation or bundle tion content of columns currently in the ized assignment problems. Remarkable
methods for the computation of lower LP basis. Since new columns are gen- results on benchmark instances from the
bounds and compare their performance in erated based on the value of the dual literature clearly attest the power of com-
the context of branch-and-bound. Sev- variables, this behaviour causes further bining cut and column generation.
eral variable fixing strategies (exact and strange columns to be generated.
heuristic) are used to reduce the algo- We present a simple method of sta-
rithm running time. Furthermore we bilizing the column generation method Automated Dantzig-Wolfe re-
use problem specific cuts and extend the which overcomes the randomness be- formulation or how to exploit
branch-and-bound algorithm to a branch- haviour of the dual variables. simultaneaously original formu-
Our lation and column generation
and-cut algorithm. The performance of method is a special case of the stabiliza-
the implemented system components is tion method proposed by du Merle et al. re-formulation
compared and evaluated on a large set of [1]. In considering stabilization in the F RANCOIS VANDERBECK
benchmark instances. context of set partitioning or set cover- MAB, Universite Bordeaux 1
ing we show that our method reduces the keywords: Lagrangian relaxation,
A branch-and-cut algorithm for number of simplex iterations and number Dantzig-Wolfe decomposition,
the graph bipartitioning prob- of columns generated by a factor 10. Fur- branch-and-price
lem thermore, we show that the stabilization A compact model representing a decom-
method eliminates the need for hot start- posable mixed integer program is intro-
S ERIGNE G UEYE ing the column generation with a good duced. From this representation, a com-
Laboratoire d’Informatique d’Avignon feasible solution. puter code can generate automatically
[1] Olivier du Merle, Daniel Vil- the column generation formulation of the
coauthor: Philippe Michelon leneuve, Jacques Desrosiers, Pierre problem as well as the formulation in its
keywords: branch-and-cut, Hansen: "Stabilized Column Genera- original variables (sometimes refered to
linearization, quadratic programming, tion", Disc. Math. 194:229-237, 1999. as the compact formulation). The pric-
polyhedral techniques ing sub-problem formulations can also be
In order to find exact solution of Integer program reformulation generated if one choose to use a MIP-
the graph bipartitioning problem, some for robust branch-and-cut-and- solver as oracle. The Lagrangian dual
branch-and-bound schemes, for which price formulation is readily available in case
the lower bound are derived with a one wants to solve the master problem
linear programming approach, may be E DUARDO U CHOA using a dual cutting plane approach such
used. Unfortunately, the linear formula- Universidade Federal Fluminense as accpm or the bundle method, or even a
tion, corresponding to the lower bound, sub-gradient algorithm.
coauthor: Marcus Poggi de Aragão
usually involves a huge number of vari- Based on this framework, we have
keywords: integer programming,
ables and constraints, making the resolu- developped a generic branch-and-price
column generation, cutting planes
tion very difficult. To solve the problem, code, named BaPCod, where the user in-
we propose a branch-and-cut, based on Since cut and column generation were es- put is limited to the problem represen-
an "economical" linear model, involving tablished as two of the most important tation in terms of its original variables
less variables than a "classical" lineariza- techniques in integer programming, re- and constraints, with no need to specify
tion of the problem. Numerical results searchers have looked for ways of com- the form of the master columns, their re-
are presented. bining them into a robust branch-and- duced cost and the like. Moreover, the
cut-and-price algorithm. Here, “robust” link between the different formulations
means that neither branching nor the ad- of the problem (in particular between
MO2-302/49 INTEGER AND MIXED INTEGER
PROGRAMMING dition of cuts should change the structure the formulation in the original variables
Integer programming column of the pricing subproblems. In the last and its column generation counterpart)
generation I few years, several researchers indepen- is exploited to allow an intelligent ini-
organizer/chair : Marco Lübbecke dently noted that cuts expressed in terms tialisation of the master program, to
of variables from a suitable original for- get stronger dual bound through pre-
Stabilized column generation mulation could be added to the master processing and variable fixing, to imple-
for set partitioning problem without disturbing the pricing. ment primal heuristics, and to recognise
34 Parallel Sessions: Monday 13:15 – 14:45 (MO2)

or enforce integrality through branching. We analyse the number of branch-and- Numerical stability of path trac-
bound nodes that are necessary to solve ing in polyhedral homotopy con-
an IP feasibilty problem in terms of a tinuation methods
condition number of the describing linear
MO2-306/31 INTEGER AND MIXED INTEGER
PROGRAMMING
system. We propose a method to reduce S UNYOUNG K IM
Integer programming this condition number, and give compu- Ewha W. University
organizer/chair : Gabor Pataki tational results on several classes of hard coauthor: Masakazu Kojima
IPs. Notably, we solve strongly corre- keywords: polynomial system,
Min-up/min-down polytopes lated knapsack problems with large coef- polyhedral homotopy cont., path tracing
ficients, that are amenable neither to LP The reliability of polyhedral homotopy
JANNY L EUNG based algorithms without precondition-
The Chinese University of Hong Kong continuation methods for solving a poly-
ing, nor to dynamic programming meth- nomial system becomes increasingly im-
coauthors: Jon Lee, Francois Margot ods. portant as the dimension of the poly-
keywords: mixed integer programming, nomial system increases. High powers
cutting planes, facets, power generation MO2-306/32 NONLINEAR PROGRAMMING
of the homotopy continuation parame-
In power generation and other produc- Nonlinear programming ter t and ill-conditioned Jacobian ma-
tion settings, technological constraints trices encountered in tracing of homo-
chair : Stanislav Zakovic
force restrictions on the number of time- topy paths affect the numerical stability.
periods that a machine must stay on once Applying a mixed variable fil- We present modified homotopy functions
activated, and stay off once deactivated. ter pattern search algorithm to with a new homotopy continuation pa-
We characterize the polyhedral structure thermal insulation system de- rameter s and various scaling strategies to
of a model representing these restric- sign enhance the numerical stability. Advan-
tions. We also describe a cutting-plane tages of employing the new homotopy
M ARK A BRAMSON parameter s are discussed. Numerical re-
method for solving integer programs in- Air Force Institute of Technology
volving such min-up and min-down times sults are included to illustrate improved
for machines. Finally, we demonstrate coauthors: Charles Audet, John Dennis performance of the presented techniques.
how the polytope of our study generalises keywords: nonlinear programming,
the well-known cross polytope (.i.e gen- categorical variables, pattern search,
eralised octahedron). filter ods An interior point algorithm for
This presentation describes the optimiza- worst-case optimisation
tion of a load-bearing thermal insula-
Subset-algebra lift operators for tion system. The optimization problem S TANISLAV Z AKOVIC
0/1 integer programming is represented as a mixed variable pro- Imperial College
gramming (MVP) problem with nonlin- coauthor: Berc Rustem
DANIEL B IENSTOCK ear constraints, in which the objective is keywords: minimax problems, multiple
Columbia University to minimize the power required to main- maxima
coauthor: Mark Zuckerberg tain one surface sufficiently cold. MVP In this paper an algorithm for the contin-
keywords: integer programming, problems are more general than mixed uous minimax problem with constraints
lift-and-project, Chvatal-Gomory cuts integer nonlinear programming (MINLP) is considered. The algorithm uses quasi–
problems in that the discrete variables are Newton search direction, conditional on
We present a generalization of the categorical; i.e., they must always take on
Sherali-Adams, Lovasz-Schrijver, Balas- approximate maximisers. The initial
values from a predefined enumerable set problem is transformed to an equivalent
Ceria-Cornuejols and Lasserre operators, or list. Thus, traditional approaches that
that lifts a 0/1 vector in n dimensions equality constrained problem, where the
use branch and bound techniques cannot logarithmic barrier function is used to en-
to one whose coordinates are indexed by be applied.
a polynomial-size set of logical expres- sure feasibility. Satisfaction of the equal-
The mixed variable generalized pat- ity constraints is enforced through incor-
sions on the initial n variables. The re- tern search (GPS)algorithm of Audet and
sulting relaxations are provable at least as porating an adoptive quadratic penalty
Dennis is extended to handle nonlinear function into the objective. Computa-
strong as those cited above, and in sev- constraints by incorporating a filter. Non-
eral important cases, they are exponen- tional results are included which illus-
linear constraints on stress, mass, and trate the efficient performance of the al-
tially stronger. One of our results is that thermal contraction are included to gen-
for any fixed tolerance ε, and any inte- gorithm
erate a more realistic feasible design.
ger r, there is a polynomial-time relax- Several computational experiments show
ation  of set-covering whose value is at MO2-306/33 NONLINEAR PROGRAMMING

substantial improvement in power re-

least 1 ε  times the value of the rank-r quired to maintain the system, as com- Nonlinear systems
Chvatal-Gomory closure. pared to the previous literature. The ad- chair : Per Åke Wedin
dition of the new constraints leads to
a very different design without signifi- A new class of robust methods to
The complexity of branch-and- solve noisy systems of nonlinear
bound cantly changing the power required. The
results strate that the new algorithm can equations
BALA K RISHNAMOORTHY be applied to a very broad class of opti- M ICHEL B IERLAIRE
University of North Carolina mization problems, for which no previ- EPFL
ous algorithm with provable convergence
coauthor: Gabor Pataki coauthor: Frank Crittin
results could be applied.
keywords: branch-and-bound, basis keywords: nonlinear equations, noisy
reduction, correlated knapsack problems problems, fixed point
 35

Many problems arising in physics, trans- a need to incorporate a special kind of M ICHAEL P OWELL
portations or economics are formulated constraints, here called regularizing con- University of Cambridge,
as the search of a fixed-point, or equiv- straints. We will consider the regular- keywords: derivative free optimization
alently as the resolution of a system of 
ization of the finite dimensional nonlin-
nonlinear equations. Unfortunately, in ear system of equations f x   0. To In trust region methods for unconstrained
many practical applications, the presence simplify, let us here only consider least optimization, a typical new vector of
of noise in the evaluation of the function squares regularization. We introduce two variables is generated by minimizing an
prevents from using standard algorithms. formulations of the regularization prob- approximation to the objective function,
Classical secant Newton methods, as the lem. First, one corresponding to a con- subject to a bound on the distance from
Broyden method, are disrupted by the strained least squares problem the best trial vector of variables so far.
noise since they try to interpolate the last We let the approximation be quadratic as
two iterates. Classical inexact Newton 1  1   usual, which has (n+1)(n+2)/2 degrees of
f x  2 s  t  L x xc 2 1  freedom, where n is the number of vari-

methods are useless since they rely on fi- 2 2

nite differences. ables. For moderate n it is expensive,
In this talk we present a new class of and the Tikhonov formulation however, to take up all this freedom by
interpolation to function values, which is
1  1  
methods building a linear model using a
done by several algorithms when deriva-
population of previous iterates in order to f x  2  λ L x xc 2 2  tives are not available. Therefore, 16

implicitly smooth the objective function. 2 2

We propose a least-square approach with months ago, the author began to investi-
an explicit control of the numerical sta- ∆ and/or λ have to be varied to find gate the use of only about 2n+1 interpo-
bility. a suitable solution of the  regularization  lation conditions, which provides some
problem. The points 12 f x  2 12 L x information about curvature. The free-

One instance of this class of methods

xc 2  corresponding to solutions of (1) lie dom in each new quadratic model is fixed

produces a generalization of the Broy-

den method. In that case, we proof  the
on  2 1decreasing
convex  L-curve. Points by minimizing the Frobenius norm of
λ 2 f x  2 L x xc 2  correspond- the change to the second derivative ma-
1

the local convergence of the correspond-

ing quasi-Newton method. We first vali- ing to the dual problem (2) lie on a con- trix of the model, this technique being
date the method on noise-free problems. cave increasing curve. These two curves analogous to the symmetric Broyden for-
It exhibits a robust behavior compared will be shown to be quite useful in practi- mula when derivatives are available. Nu-
to standard methods. We then present cal computations. This is true even if we merical experiments have provided good
numerical evidence of the efficiency of do not compute exact solutions of (1) and news and bad news. Recently, however, a
our new class of methods in presence of (2). It is also true if there exist several lo- matrix factorization has been found that
noisy nonlinear system of equations. cal minima of (1) and (2). halves the number of digits that are lost
through cancellation, and also a positive
semi-definiteness property is preserved
Conjugate sets for a nonlinear MO2-306/34 NONLINEAR PROGRAMMING
automatically, but the factorization had
programming problem New trust region algorithms not been tried in practice at the time of
H IDEFUMI K AWASAKI organizer/chair : Jorge Nocedal writing this abstract. An up-to-date ac-
Graduate school of mathematics, count of the research will be presented.
Kyushu University An active-set trust-region algo-
keywords: conjugate sets rithm for nonlinear optimiza- A feasible trust-region algo-
tion rithm for nonlinear optimiza-
The conjugate point is a global concept in
the calculus of variations. In variational tion
R ICHARD WALTZ
problems, the variable is not a vector in Northwestern University S TEPHEN W RIGHT
Rn but a function. So, a simple and natu- University of Wisconsin
coauthors: Richard Byrd, Nick Gould,
ral question arises. Is it possible to estab-
Jorge Nocedal keywords: nonlinear programming,
lish a conjugate points theory for a non-
keywords: trust region, active set, SLP, trust region method, sequential quadratic

linear programming problem: minimize
nonlinear optimization
f x  on x  Rn ? Recently, this problem progr.
was positively solved. In this talk, we in- We will describe an implementation of A sequential quadratic programming al-
troduce a conjugate set which is a gener- a SLP-EQP algorithm for nonlinear opti- gorithm for smooth constrained nonlin-
alization of the conjugate point for a non- mization. This algorithm relies on the so- ear optimization problems that maintains
linear programming problem. lution of linear programming and equal- feasibility of all iterates with respect to
ity constrained subproblems which we the constraints will be described. Global
Primal and dual optimization believe makes this approach better able and local convergence analysis will be
problems in tikhonov regular- to solve large-scale problems compared discussed. The usefulness of the ap-
ization with existing active-set methods. We will proach will be demonstrated by outlining
present numerical results to support this an application to nonlinear model predic-
P ER Å KE W EDIN claim and discuss algorithmic advances. tive control.
Department of Computing Science Particular focus will be paid to issues
coauthors: Jerry Eriksson, Thomas of warm-starts and approximate LP solu- MO2-306/35 NONLINEAR PROGRAMMING

Viklands tions for the subproblems. Recent developments in filter

keywords: primal-dual, regularization, methods I
ill-conditioning A new algorithm for uncon- organizers : Michael Ulbrich, Stefan
In order to solve an ill-conditioned prob- strained minimization without Ulbrich and Luís N. Vicente
lem with numerical optimization there is derivatives chair : Michael Ulbrich
36 Parallel Sessions: Monday 13:15 – 14:45 (MO2)

A filter pattern search method technique of Fletcher and Leyffer (1997) In this paper we propose a new ap-
for non-smooth constrained op- to the globalization of the primal-dual proach for constructing efficient schemes
timization interior-point algorithm, avoiding the use for non-smooth convex optimization. It
of merit functions and the updating of is based on a special smoothing tech-
C HARLES AUDET penalty parameters. nique, which can be applied to the func-
École Polytechnique de Montréal - The algorithm decomposes the tions with explicit max-structure. Our ap-
GERAD primal-dual step obtained from the per- proach can be considered as an alterna-
coauthor: John Dennis turbed first-order necessary conditions tive to the black-box minimization. From
keywords: nonlinear programming, into a normal and a tangential step, the viewpoint of efficiency estimates, we
filter algorithm, pattern search, whose sizes are controlled by a trust- manage to improve the traditional bounds
nonsmooth optimization region type parameter. Each entry in on the number of iterations of the gradi-
This talk will focus on defining and an- the filter is a pair of coordinates: one ent schemes from O ε12 to O 1ε , keep-


alyzing a filter pattern search method for resulting from feasibility and centrality,

ing basically the complexity of each iter-

general nonlinear programming without and associated with the normal step; the ation unchanged.
derivatives. Since it is a pattern search other resulting from optimality (comple-
method, this method is highly dependent mentarity and duality), and related with
on the set of directions it is given to work the tangential step. The method pos- A bundle method for convex op-
with. Linear inequalities are treated by an sesses global convergence to first-order timization: implementation and
adequate choice of directions, and non- critical points. illustrations
linear constraints are handled by a fil- We present numerical results for
ter. We present a hierarchy of optimal- large-scale problems and discuss exten- C ORALIE T RIADOU
ity results tied to local differentiability of sions of the original algorithm and of its CERMSEM (Paris 1) - Artelys
the problem and to the set of directions convergence theory. coauthors: Claude Lemaréchal, Jean
used by the algorithm. The smoothness Maeght, Arnaud Renaud
assumptions that we analyze vary from MO2-306/36 NONSMOOTH OPTIMIZATION
keywords: bundle methods, Lagrangian
lower-semicontinuity, Lipschitz continu- Recent advances in nonsmooth relaxation, combinatorial problems
ity up to strict and continuous differentia- optimization I
bility. We present the implementation of a bun-
organizers : Jean-Louis Goffin and dle method, oriented toward Lagrangian
Jean-Philippe Vial relaxation of combinatorial problems.
Filter methods and NLP’s with chair : Jean-Philippe Vial Problems to be solved have a non-
unbounded multipliers differentiable objective function, and
Proximal ACCPM, a cutting bound constraints on the variables. The
DAVID S HANNO plane method for column gen-
Rutgers University bundle method is a cutting-plane method,
eration and Lagrangian relax- in which the deviation of the next iterate
coauthors: Hande Benson, Arun Sen, ation: application to the p- from the reference point is penalized by a
Igor Griva, Robert J. Vanderbei median problem quadratic term. The next iterate is given
keywords: filter methods, unbounded by a quadratic programm, ours is a two
multipliers, interior point method J EAN -P HILIPPE V IAL level one in order to take into account the
University of Geneva
The talk discusses using a filter method bounds on the variables. A dynamic and
interior point code on problems with coauthor: Olivier du Merle a static management of the penalization’s
unbounded sets of Lagrange multipli- keywords: Lagrangian relaxation, coefficient will be compared.
ers. Theoretical and computational re- cutting plane methods, analytic center, The presented tests belong to differ-
sults will be given demonstrating that proximal methods ent categories : L1- and L2-norm mini-
constraint relaxation can lead to a very Proximal ACCPM is a variant of the an- mization, Lagrangian relaxations of com-
efficient algorithm. Implications for con- alytic center cutting plane method, in binatorial problems, such as traveling
vergence theory of filter methods will be which a proximal term is added to the salesman (Held & Karp) or p-median.
discussed. barrier function that defines the center. We present results for our bundle method
The present paper gives a detailed pre- Artelys Dualis on these problems, outlin-
On a primal-dual interior-point sentation of the method and of its imple- ing the peculiarity of each category.
filter method for nonlinear pro- mentation. Proximal ACCPM is used to
gramming solve the Lagrangian relaxation of the p- MO2-306/37 CONVEX PROGRAMMING

median problem on two sets of problem

R ENATA S ILVA instances. Problems of the same collec-
Convex optimization methods
Department of Mathematics - University tion are tentatively solved with the clas- chair : Jean-Baptiste Hiriart-Urruty
of Coimbra sical column generation scheme.
coauthors: Michael Ulbrich, Stefan
Coupling general penalty
Ulbrich, Luís N. Vicente
schemes for convex program-
keywords: interior point method, Smooth minimization of non- ming with the steepest descent
primal-dual, filter, global convergence smooth functions and the proximal point algo-
rithm
In this talk, we present the primal-dual Y URII N ESTEROV
interior-point filter method for nonlinear CORE/UCL M ATIAS C OURDURIER
programming developed by Ulbrich, Ul- keywords: non smooth optimization, University of Washington
brich, and Vicente (2000). The method structural optimization, gradient coauthor: R. Cominetti
is based on the application of the filter methods, complexity keywords: penalty methods
 37

The standard approach in penalty meth- Finally we show that our extension of the links and processors of the same cluster
ods is trying to trace the path generated Frank-Wolfe theorem immediately im- connected via high-speed links). Com-
by the solutions of penalized versions of plies continuity of the solution set de- ponents of this Grid can fail eventually.
the original problem. If the limit point fined by the considered system of (quasi- Because of these characteristics we
of these penalized solutions solves the )convex inequalities. propose a distributed branch-and-bound
original convex program, then by tracing algorithm that includes procedures of
them we obtain a solution to the problem. Solving some matrix nearness load balance and fault tolerance. Con-
This approach is useful because the pe- problems via convex optimiza- cerning the load balance problem, a dis-
nalized versions are easier to solve that tion tributed dynamic model was employed
the original constrained convex program, that favoured the balance among proces-
but we can be expending a lot of compu- J EAN -BAPTISTE sors concentrated geografically. In or-
tational resources in trying to trace this H IRIART-U RRUTY der to treat the fault tolerance problem,
penalized path. This might be unneces- Paul Sabatier University checkpoint protocols that take advantage
sary when there is a hole family of trajec- of the hierarquical structure of Grids are
coauthors: Marcel Mongeau,
tories converging to the original optimal also been developed.
Pawoumodom L. Takouda
set. Maybe in that case we can attain con- The procedures proposed are been
keywords: convexity, matrices,
vergence to a solution just by following applied on an already existing sequen-
algorithms
the ’general direction’ of this trajectories, tial branch-and-bound algorithm for the
by following the direction of the flow. In We intend solving the following prob- Steiner Problem in graphs. Excellent
this paper we study how in the convex lem : given a matrix, find its near- speed ups are obtained, allowing the res-
program case the dynamics raising from est matrix (in the usual Frobenius-Schur olution of a number of formerly open in-
the coupling of the penalty scheme with norm) in a convex set of matrices struc- stances from the SteinLib library in rea-
the steepest descent method gives such a tured as the intersection of a closed con- sonable times. The computational ex-
family of trajectories. And how by cou- vex cone with an affine subspace. In- periments are been run on a Grid with
pling the penalty schemes with an ’in- stances are: the set of doubly stochastic Globus-MPI.
exact proximal point’ we attain conver- matrices, the set of correlation matrices,
gence by following the general direction the so-called spectraplex,...Applications
of the flow. In the linear program case arise from: social choice theory (ag- Parallelization of a MIP solver
we also study the convergence of some gregation of preferences), risk manage- in the PUBB2 framework
naturally associated dual variables to an ment in portfolio optimization, eigen-
optimal dual solution. value optimization,... Among the algo- Y UJI S HINANO
rithmic procedures we might thing of us- Tokyo University of Agriculture and
ing(alternate projections, based on con- Technology
Conditions for boundedness and vex duality, fixed point approach, interior coauthors: Tetsuya Fujie, Yuusuke
the existence results for quasi- point methods), we focus aur attention on Kounoike
convex programmes an adaptation of Boyle-Dykstra alternat- keywords: parallel algorithm, mixed
W IESLAWA O BUCHOWSKA ing projection algorithm. We report some integer programming,
Chowan College numerical experiments (matrices of size branch-and-bound, software framework
up to some hundreds).
keywords: mathematics, convexity We introduce a simple method of paral-
We consider the problem of boundedness lelizing a MIP (Mixed Integer Program-
MO2-306/38 PARALLEL COMPUTING
ming) solver. This method is differ-
and the existence of an optimal solution
to the constrained optimization problem. Distributed solution of LP and ent from a standard implementation that
We present necessary and sufficient con- MIP problems constructs a parallel branch-and-cut al-
ditions for boundedness of a convex pro- chair : Richard Van Slyke gorithm from scratch (except using an
gram where the objective function and LP solver). The MIP solver we use is
the inequality constraints are either faith-
Towards a grid enabled branch- ILOG-CPLEX MIP Optimizer (Version
fully convex functions (satisfying some
and-bound algorithm 8.0), which is one of the most efficient
mild condition) or quasi-convex polyno- implementations of branch-and-cut algo-
L UCIA D RUMMOND rithms. The parallelization of the solver
mials. The conditions are provided in the Fluminense Federal University
form of an algorithm, terminating after is performed by using the software tool
a finite number of iterations, the imple- coauthors: Jose Viterbo, Clicia Stelling PUBB2 developed by the authors. We
mentation of which requires the identifi- Castro, Eduardo Uchoa report a part of our computational expe-
cation of implicit equalities in a homoge- keywords: distributed algorithms, load rience using up to 24 processors. In ad-
neous linear system. We also prove that balance, branch-and-bound, fault dition, we point out some problems that
the optimal solution set of the consid- tolerance should be resolved for a more efficient
ered problem is nonempty, which gener- This work presents a distributed branch- parallelization.
alizes attainability results proved very re- and-bound algorithm to be run on a wide-
cently by other authors for quasi-convex area-network system provided with an Distributed computing and the
quadratic objective function and con- infra-structure that allows to execute ap- simplex method
vex quadratic constraints and for convex plications that require much computa-
polynomial objective function and con- tional power, the so-called Grid. These R ICHARD VAN S LYKE
straints. The latter results are on the systems are usually composed of many Polytechnic University
other hand extensions of the well known heterogeneous computers connected in a coauthor: Gavriel Yarmish
Frank-Wolfe theorem for a quadratic ob- hierarquical fashion (clusters of different keywords: simplex method, distributed
jective function and linear constraints. institutions connected through low-speed computing
38 Parallel Sessions: Monday 13:15 – 14:45 (MO2)

Several proposals for parallel and dis- Carmen Systems In the talk, I will present how Carmen
tributed implementations of the simplex keywords: personnel scheduling, crew Systems has approached some of the dif-
method have been based on the standard, pairing, airline ficulties in the railway crew pairing prob-
rather than the revised, simplex method. lem. The case in study is Deutsche Bahn,
This is primarily because the standard The basic crew pairing problem is a well one of the largest transportation compa-
method is "embarrassingly parallel" in known OR problem with direct applica- nies worldwide.
the sense that it can be implemented with tion to airlines, and many published pa-
minimal communication between proces- pers. However, there are a number of MO2-308/12 STOCHASTIC PROGRAMMING

sors; this, in turn, allows substantial important variants to the pairing problem Applications of stochastic
scalability even when the communica- that are less studied. One of these is the programming
tion latency between processors is rela- pairing problem with crew augmentation, organizer : COSP (Chefi Triki)
tively high. On the other hand, it is well in which the number of crew to operate chair : Chefi Triki
known that the revised simplex method per flight is a decision variable.
is much more efficient least for serial The crew augmentation problem usu- FRC-S3, a fix-and-relax coor-
processors solving problems which are ally arises when planning pairings for dination scheme for stochastic
sparse, and/or have high aspect ratio (ra- long-haul flights. Depending on union sequencing and scheduling
tio of columns to rows). However, highly and safety rules, an airline may be able L AUREANO F. E SCUDERO
scalable revised methods don’t exist for to reduce rest times between flights by Universidad Miguel Hernández
using coarse grained, distributed process- augmenting the crew, i.e. adding an ex- coauthors: M. F. Clement, Antonio
ing. tra crew member. Even though there are Alonso, M. L. Gil, M. T. Ortuño
We exam the tradeoff with the stan- additional costs associated with the ex- keywords: sequencing and scheduling,
dard method between scalability and the tra crew member, it may still be advan- stochastic integer programming, fix and
inefficiency of not using the revised tageous to operate with additional crew, relax coordination
method, especially as a function of prob- due to the efficiency improvement from We present a framework for solving pure
lem density. To this end we have pro- the reduced rest times. 0-1 programs for scheduling and se-
grammed a new standard method which The crew augmentation problem can quencing problems with uncertainty in
can easily be executed in a distributed be approximated using a two step ap- the objective function coefficients, the
environment. We present and validate proach that involves solving two basic constraint matrix and the right-hand-side.
performance models for this approach on crew pairing problems. A difficulty with Typical elements are: limited availabil-
"real life" and synthetic problems using that approach is to estimate the cost of ity of the resources, multiperiod opera-
our implementation. the extra crew when solving the first ba- tions and precedence relationships. The
sic problem. In this talk we introduce a stochasticity of the problem is due to
MO2-308/11 AIRLINE OPTIMIZATION APPLICATIONS
one-shot approach that accurately mod- the operations’ execution cost and the re-
Airline scheduling I els the crew augmentation problem, and sources’ consumption by the operations
compare that with the two step approach and their availability along the time hori-
organizer/chair : Stefan Karisch on real world problems. zon. A multistage scenario analysis with
Robust airline crew pairing op- a mixture of full and simple recourse is
timization Challenges in railway crew pair- used. By considering a splitting variable
ing mathematical representation of the Deter-
S ERGEI C HEBALOV ministic Equivalent Model, we present a
University of Illinois at heuristic approach the so-called Fix-and-
Urbana-Champaign L ENNART B ENGTSSON
Carmen Systems Relax Coordination algorithmic frame-
coauthor: Diego Klabjan work to exploit the model’s structure and,
keywords: crew scheduling, airline keywords: personnel scheduling, crew specifically, the non-anticipativity con-
operations research, optimization pairing, railway straints for each scenario group in the
Due to the flight disruptions in opera- Automatic crew planning system are stochastic model. We consider a given
tions, the crew scheduling cost at the end nowadays standard within the airline in- strategy for partitioning the 0-1 variables
of a month is substantially higher than the dustry, and they are also finding applica- in clusters, such that the approach selec-
projected cost in planning. We present tions within some of the major European tively explores the active nodes of the
a model that yields more robust crew railway companies, such as Deutsche Branch-and-Fix (BF) tree for obtaining
schedules in planning. Besides the objec- Bahn and Swedish Railways. This talk (hopefully) good solutions for each sce-
tive of minimizing the cost, we introduce will focus on the crew pairing part of the nario related model. The algorithm uses
the objective of maximizing the number crew planning process, in which one or the twin node family concept. It is de-
of crew that can be swapping in opera- two-day pairings are composed from the signed for coordinating and reinforcing
tions. We present a solution methodology individual train legs. the node pruning, variables fixing and
for solving the resulting model. The pro- Compared to the airline crew pair- branching node and variable selection at
duced crew schedules are evaluated with ing problem, the railway network is very each scenario related BF tree.
a simulation. Several possible extensions dense, with many short legs within a
and directions will be discussed. limited geographic region. In addition, Weekly supply chain coordina-
the total number of legs per day often tion with stochastic demand
Solving crew augmented pairing exceeds 10,000, whereas a large airline M ARTE F ODSTAD
problems using column genera- might operate in the order of 1,000 flights SINTEF
tion per day. These two factors make the coauthors: Asgeir Tomasgard, Erik
railway crew pairing problem quite chal- Høeg, Peter Schütz
C URT H JORRING lenging to solve. keywords: supply chain optimization
 39

In this paper we present supply chain op- MO2-308/13 STOCHASTIC PROGRAMMING option-type. We explore applications in
timisation models for the food industry. robust portfolio management and multi-
The focus of the decision support mod- Stability in stochastic product pricing.
els is to maximize expected profit sub- programming
ject to uncertain demand and shortfall chair : Rene Henrion
Hoelder and Lipschitz stability
costs. This is done by co-ordination of Response surface merthodology of solution sets in programs with
production at a set of facilities and inven- with stochastic constraints for probabilistic constraints
tories in different regions and at differ- expensive simulation
ent levels in the supply chain. The mod- R ENE H ENRION
els describe decisions at the operational M. E BRU A NGUN Weierstrass Institute Berlin
level at a weekly resolution with a two Tilburg University coauthor: Werner Roemisch
month horizon. We consider the situa- coauthors: Gul Gurkan, Dick den keywords: probabilistic constraints,
tion with uncertain demand and present Hertog, Jack P. C. Kleijnen Lipschitz stability, stochastic
mathematical models based on stochastic keywords: simulation based programming
linear programming. We describe how to optimization, stochastic constraints,
generate scenarios based on quantile re- We consider a convex program with
metaheuristics, expensive simulation probabilistic constraints induced by an
gression and present results from a pilot
In this presentation, we investigate r-concave probability distribution. Usu-
test at the Norwegian meat Co-operative. simulation-based optimization problems ally, only limited information about
with a stochastic objective function and the underlying probability distribution
stochastic constraints, as well as deter- is available, and the problem is solved
Optimal capacity allocation in ministic box constraints. More specif- on the basis of suitable approximations.
multi-auction electricity mar- ically, we generalize Response Sur- Therefore, it is reasonable to study quan-
kets under uncertainty face Methodology (RSM) to account for titative stability of solution sets to such
stochastic constraints. This extension is programs under perturbations of the dis-
C HEFI T RIKI obtained through the generalization of tribution. First we provide conditions im-
University of Lecce the estimated steepest descent - used in plying Hoelder continuity of the solution
conventional RSM - using ideas from set w.r.t. the Kolmogorov distance of
keywords: competitive electricity interior point methods, especially affine probability distributions. Here, the per-
market, bidding strategies, stochastic scaling. The new search direction is scale turbed distributions are completely free,
model independent. This is especially impor- which allows to include the important
With deregulation any producer finds tant from the practitioners’ point of view, class of empirical approximations. Sec-
himself in front of the problem of choos- since it avoids some numerical compli- ondly, applying general stability results
ing how to allocate the available capacity cations and problems commonly encoun- by Bonnans/Shapiro, we derive upper
in order to maximize his profit. The mul- tered. Furthermore, we provide a heuris- Lipschitz continuity for solution sets un-
tiplicity of auctions in electricity markets tic that uses this search direction iter- der more restrictive conditions on the
and the non trivial constraints imposed atively. The heuristic is primarily in- original program and on the perturbed
by technical and bidding rules make the tended for simulation-based optimization distributions. The stability results are il-
problem of crucial applicability impor- problems in which each simulation run is lustrated by numerical tests showing the
tance and difficult to model and solve. very expensive and the simulation bud- different asymptotic behaviour of para-
Further difficulties are represented by the get is limited, so that one needs to reach metric and empirical estimates in a pro-
dynamic and stochastic natures that char- a neighborhood of the true optimum in gram with a probabilistic constraint un-
acterize the decision process. Mathemat- only a few simulation runs. Numerical der multivariate normal distribution.
ical programming under uncertainty ap- illustrations of the heuristic for a Monte
pears to be the most appropriate frame- Carlo example give encouraging results. MO2-308/T1 FINANCE AND ECONOMICS

work to capture all the aspects of this Model building

kind of problems. Robust mean-covariance solu- chair : H.Paul Williams
The formulation proposed is basi- tions for stochastic optimization
cally a unit commitment based model An econometric model of state
that covers a multi-period time horizon.
I OANA P OPESCU budget
Uncertainty related to the clearing infor- INSEAD
mation is incorporated in the model ex- keywords: robust optimization
L UDMILA KOSHLAI
Senior Researcher
plicitly by means of a set of scenarios We provide a method for deriving robust
incorporated in a multi-stage stochastic solutions to stochastic optimization prob- coauthor: M. Mikhalevich
program. The recourse stages correspond lems with general objective, based only keywords: budget modeling,
to the different bidding moments refer- on mean-covariance information about behavioural econometric models,
ring to the same time period, and not, the distribution underlying the random regression function, nondifferentiable
as traditionally assumed, the successive cost vector. For a general class of ob- optimization
time periods referring to the same auc- jective functions, we show that the ro- Accounting a rational budget is one of the
tion. bust optimization problem is equivalent vital issues of economic growth. Such
Under the assumption that the sup- to solving a certain deterministic para- characteristics of a transition economy as
plier is price-taking bidder, the result- metric quadratic program. Interesting re- the dominance of formerly state-owned
ing model is a multi-stage mixed-integer sults arise from comparing the robust so- property, strong monopoly influence and
stochastic optimization. We verified the lutions with those corresponding to en- prohibitive technology costs result in the
effectiveness of the generated bidding tropy maximizing distributions for vari- complexity of budgetary policy. Dy-
strategies by using Lingo. ous criteria, such as target, fractile and namic simulation model is considered for
40 Parallel Sessions: Monday 13:15 – 14:45 (MO2)

solving this problem. Account of bud- variables. Original numerical algorithms the semidefinite cone and compared with
get influences the prices and optimiza- based on method of nonsmooth and com- the algebraic generalizations mabe by
tion of taxes are the main its features. binatorial optimization are proposed for Gowda and others.
Accurate modelling of budget influence the mentioned problem solution. Real-
needs the development of econometric data numerical experiments were done P-properties of linear transfor-
behavioural models for main economic on the basis of 18-industry balances of mations on Euclidean Jordan al-
agents. Another problem of budget mod- Ukrainian economy. Several scenarios of gebras
elling is constructing an objective func- possible technological changes were cre-
tion for its optimization. Methods of or- ated. M UDDAPPA G OWDA
dinary regression can be applied for such University of Maryland, Baltimore
problems solution. Parameters of regres- The allocation of shared fixed County
sion function are estimated by this ap- costs: fairness versus efficiency keywords: P-property, Euclidean Jordan
proach as the solution of the system of algebra
inequalities. The methods of nondiffer- H.PAUL W ILLIAMS An n by n real matrix is said to be a P-
entiable optimization can be applied to London School of Economics matrix if all its principal minors are posi-
find a solution (or pseudosolution) to the keywords: fixed cost allocation, tive. It is well known that this property
system and thereby to determine the ob- fairness, integer programming duality, can be described in numerous equiva-
jective and behavioural functions. The game theory lent ways, such as: the non-sign-reversal
model of state budget is constructed for We consider the problem of sharing the property based on the componentwise
such known functions as the system of fixed costs of facilities among the users. product of vectors, the order P-property
finite-difference equations. The gradient This can be regarded as seeking ’dual val- based on the minimum and maximum of
methods for the optimal control problems ues’ for an Integer Programme. Alterna- vectors, the uniqueness property in the
can be used to find the "best" solution tive solutions are given and considered in standard linear complementarity prob-
in such model. The mentioned approach relation to criteria of both fairness and ef- lem, the (Lipschitzian) homeomorphism
is realized in computer decision-support ficiency. The obtaining of a ’fair’ solu- property of the (standard) normal map,
system. The demonstration version of it tion connects to concepts from coopera- etc. In this talk, we describe how these
will be presented. tive Game Theory. The problem is illus- concepts can be extended for a linear
trated by that of sharing the costs of cen- transformation defined on an Euclidean
Optimization model of planned tral computing provision among the dif- Jordan Algebra.
technological-structural ferent faculties in a large university.
changes Semidefinite complementary
MO2-308/T2 COMPLEMENTARITY AND VARIATIONAL
INEQUALITIES cones and faces in SDLCP
M IKHALEVICH M IKHAIL Semidefinite linear S RINIVASA R. M OHAN
Academy of Foreign Trade complementarity problem Indian Statistical Institute, Delhi Centre
coauthor: Ludmila Koshlai organizers : T. Parthasarathy and keywords: semidefinite LCP,
keywords: aggregated interindustry Srinivasa R. Mohan semidefinite cone, face, cone preserving
model, non linear problems, numerical chair : Srinivasa R. Mohan transformation
algorithms, combinatorial optimization Given a real symmetric matrix Q and a
Geometrical notions of p and re-
Monetary stabilization and reforms in fi- lated properties in semidefinite linear transformation L from the space
nancial sector for countries in transition linear complementarity prob- of all real symmetric matrices into it-
must be accompanied by radical adjust- lems self, we consider the semidefinite lin-
ment in industrial technological system. ear complementarity problem of finding
M ADHUR M ALIK
Decrease of production costs is a major 
a symmetric positive semidefinite matrix
goal of such technological changes. It Indian Statistical Institute, Delhi Centre X such that L X   Q is positive
  semidef-
can be achieved in a twofold way: first, keywords: semidefinite inite and the trace of X L X   Q   0 
to reduce inputs of commodities, materi- complementarity, face, complementary There are a number of applications for
als and other manufactured products per cone, subtransformations this problem and a study of such prob-
unit of output, and second, by gradually In this paper we generalize the notion lems is also relevant for problems of vari-
reducing the use of energy- and resource- of complementary cones to the semidef- ational inequalities over nonpolyhedral
inefficient technologies. Both these ap- inite setting. The semi-definite linear closed convex cones. We use the geome-
proaches reduce the technological coef- complementarity problem is a general- try of the semidefinite cone and the struc-
ficients of input-output matrix in aggre- ization of the well known linear comple- ture of its faces to study the properties
gated interindustry model. The optimiza- mentarity problem. It is a compelmen- of transformations that preserve the pos-
tion model for this purpose is consid- tarity problem, over the nonpolyhedral itive semidefinite cone and transforma-
ered in our paper. Let A be the ma- self dual closed convex cone of semidef- tions that include the Stein transforma-
trix of technological coefficient and q is inite matrices and is relevant for a study tion, studied earlier by Gowda and oth-
the vector of share of labour cost in the of the problem of variational inequalities. ers.
price of each industry product. The prob- It arises naturally in the unified formu-
lem is to determine what changes in el- lation of a pair of primal-dual semidef- MO2-308/T3 MODELING LANGUAGES AND SYSTEMS

ements of A and q will maximize the inite programming problems. Certain Modeling languages and
consumer income without triggering ad- concepts and results such as those relat- systems I
ditional cost-inflationary forces. Depen- ing to P-matrix available for linear com- organizer/chair : Robert Fourer
dently from assumptions about possible plementarity problems are then general-
changes, we obtain the non-linear prob- ized to semidefinite linear complemen- Solving large scale supply chain
lems with continuous or (and) discrete tarity problems using the geometry of optimization problems
 41

A LKIS VAZACOPOULOS simulations for risk analysis. Design and M ASAKAZU M URAMATSU
Dash Optimization, Inc. algorithmic principles will be addressed. The University of
coauthor: Alan Dormer Applications in the area of energy stock Electro-Communications
keywords: constrained programming, management, financial contract pricing keywords: second order cone
integer programming, supply chain and supply chain management will be programming, pivoting algorithm
planning, modeling system presented.
We give a framework for a pivotting al-
The paper will examine the results of a
three year EU funded R&D project in A modeling language for gorithm for second-order cone program-
stochastic programming ming. First, we discuss how to define ba-
large-scale integrated supply chain opti-
sic and nonbasic variables, and a dictio-
mization. The objective of the project L EONARDO L OPES nary using them, for second-order cone
was to use mixed-integer programming Northwestern University programming. Then a pivotting proce-
(MIP) and constraint programming (CP)
coauthor: Robert Fourer dure is proposed to exchange basic and
to address the combined planning and
keywords: stochastic programming, nonbasic variables. To pick up a variable
scheduling problem that arises in supply
modeling languages entering to the basis, we need to solve
chain planning and optimization. The
normal approach is to have two models - StAMPL a modeling tool specifically de- low-dimensional subproblems. When all
a planning model and a scheduling model signed for stochastic programming prob- the subproblems have zero optimal solu-
- with the former solved with MIP and lems with recourse. It exploits these tion, we terminate the algorithm. This
the latter solved with CP. The need to problems’ special structure, using a syn- is a try to extend the simplex method
use two models and two separate sys- tax that stresses aspects important from for linear and quadratic programming to
tems increases the complexity, reliabil- a modeling focus. As a result, StAMPL second-order cone programming.
ity and lifecycle cost of the system. It models emphasize the dynamic aspect
also requires some manual intervention of the decision model, with a notation Solving nonconvex SDP prob-
to iterate between the two systems which that focuses on stochastic programming- lems of structural optimization
is expensive and unreliable. The project specific concepts while taking advantage by PENNON
has two main results that will be dis- of features in the AMPL modeling lan-
cussed: firstly a unified achitecture (one guage to handle the algebraic compo- M ICHAL KOCVARA
system/one model) has been found that nents of the model. With this approach, Inst. of Applied Mathematics, University
works well for the planning and schedul- StAMPL can represent models in a clean of Erlangen
ing problem and secondly this architec- and scalable format, uncomplicated by
tire has been applied to several real- indexing over either periods or scenarios. coauthor: Michael Stingl
world problems. Further developments keywords: semidefinite programming,
are in progress to refine the technology structural optimization, nonlinear
programming
and make it applicable to the more gen- MO2-341/21 INTERIOR POINT ALGORITHMS

eral planning and scheduling problems Computational conic Semidefinite programming offers an el-
that arise in transportation and human re- programming egant way how to formulate optimiza-
sources, for example. tion problems with constraints on eigen-
organizers : Mituhiro Fukuda and
Masakazu Kojima values. We introduce new formula-
Modeling, optimization and chair : Mituhiro Fukuda tions of Truss Topology and Free Ma-
simulation of stochastic dy- terial Optimization problems with con-
namic systems A non-polyhedral primal active- straints on self-vibration and global sta-
set approach to SDP bility of the optimal structure. The re-
J EAN -P IERRE G OUX sulting semidefinite programming prob-
Artelys Y IN Z HANG
Rice University lems will be solved by optimization code
coauthors: Nicolas Bonnard, Olivier PENNON.
Teytaud, Arnaud Renaud coauthors: Kartik Krishnan, Gabor
The problems with constraints on
keywords: stochastic optimization, Pataki
self-vibration lead to large-scale linear
dynamic programming, linear keywords: active-set approach,
SDP, while the problems with global sta-
programming, risk management semidefinite programming
bility control (that are of more practical
The management of hydro reservoirs, We present a non-polyhedral primal importance) result in large-scale noncon-
natural gas storages, financial assets port- active-set approach to semidefinite pro- vex SDP.
folios or cement reserves requires to gramming which exploits the low rank The SDP problems, both the linear
make a sequence of decisions balancing of optimal primal extreme point solu- and the nonlinear one, are large-scale
present gains and expected future gains. tions. The goal is to find a proper su- and extremely difficult to solve. The
Complex dynamic management strate- perset of the range space of an optimal algorithm used in PENNON is a gen-
gies that satisfy demand while maxim- primal extreme-point solution. The algo- eralized version of the Augmented La-
imizing expected profit and minimizing rithm generates a sequence of primal fea- grangian method, originally introduced
risk are necessary. sible iterares with non-increasing objec- for convex nonlinear programming prob-
Artelys DynOpt is a C++ model tive values. Under a nondegeneracy as- lems. The efficiency of this approach is
building and solving environment that fa- sumption, the objective values are strictly based on a special choice of the penalty
cilitates the description and the analysis decreasing. We will also discuss conver- function for matrix inequalities. This
of such systems. Artelys DynOpt imple- gence issues for this approach. special choice affects the complexity of
ments powerful stochastic dynamic pro- the algorithm, in particular the complex-
gramming algorithms to compute robust A pivoting structure in second- ity of Hessian assembling, which is the
management strategies and Monte-Carlo order cone programming bottleneck of all SDP codes working with
42 Parallel Sessions: Monday 13:15 – 14:45 (MO2)

second-order information. Unlike codes weights have to meet upper and lower Newton methods to solve these kinds of
based on interior-point techniques, PEN- bounds. We show that in space the in- problems.
NON can be directly generalized for non- verse 1-median problem can be solved In this talk we present a matrix-
linear nonconvex SDP. in O(n log n) time, provided that the free iterative algorithm designed to solve
Number of practical examples will il- distances are measured in rectangular or large scale problems without resorting to
lustrate the importance of the eigenvalue maximum norm. finite differences, which could be a de-
constraints and the efficiency of PEN- cisive property in the presence of noise,
NON. Location models: a reformula- and without imposing particular struc-
tion by discretization tures on the problems or the Jacobian.
MO2-341/22 LOGISTICS AND TRANSPORTATION
Practically this method allows us to
Facility location F RANCISCO solve problems, coming from the trans-
chair : Francisco Saldanha-da-Gama S ALDANHA - DA -G AMA portation field, with more than 120’000
University of Lisbon / Operational variables. Computational experiments
Lagrangian approach for hy- Research Centre on standard problems show that this al-
brid problem of facility location coauthor: Luis Gouveia gorithm outperforms classical large-scale
and network design keywords: discrete location, extended quasi-Newton methods in terms of effi-
H IROAKI M OHRI formulations ciency and robustness. Moreover, its nu-
Waseda University merical performances are equivalent to
The usual location models contain a set the performances of the Newton-Krylov
coauthors: Naoto Katayama, Ming Zhe of binary variables Yi indicating whether methods currently considered as one of
Chen or not, a facility is built at location i. In the best methods to solve high-dimension
keywords: network application, this talk we discuss models using binary nonlinear systems of equations.
network design, facility location, variables Ziq indicating whether a facil-
Lagrangian relaxation ity is built at location i AND is serving
We introduce a hybrid problem of facility q clients. We relate formulations involv-
location and network design. Our prob- ing the new set of variables with previ- Solving large scale integer mul-
lem is a kind of network design problem ous known formulations. We also present ticommodity network flow prob-
combined with a facility location prob- and discuss a set of inequalities using the lems in one second
lem. This problem includes the network new set of variables and generated by the
design problem considering fixed arc cost Chvatal-Gomory rounding method. We M ARTIN J OBORN
and routing cost as a special case. Also, also discuss variations of the problem Carmen Consulting
the problem has many applications espe- with nonlinear costs at the facilities (as
cially in planning telecommunication and these are easily modelled by the new vari- keywords: railway transportation,
computer networks. In such cases, ca- ables). Computational results taken from empty distribution, multicommodity
bles are arcs and facilities, which are au- instances with up to 100 client nodes and network flow, decomposition
tomatic switching machines as such, are 20 facility nodes are presented for several
special nodes on a connected graph. This variations of the location problem. In an application of real time fleet allo-
problem is proved to be a NP hard prob- cation, an integer multicommodity net-
lem in terms of Computing Complexity MO2-341/23 LOGISTICS AND TRANSPORTATION work flow problem has to be solved re-
Theory. That is why it is difficult to solve peatedly and a solution is needed very
Transportation quickly. The problem is decomposed ge-
the large size problem. We show that a
solution is reached by using a mixed in- chair : Uwe Zimmermann ographically by studying the structure of
teger programming formulation and effi- A new approach to solve large- the transportation network. In an on-line
cient algorithms with a focus on the La- scale systems of nonlinear equa- setting, a fast heuristic checks the feasi-
grangian relaxation approach. If time tions bility of demands. In the heuristic, the
permits, we will show some numerical problem is decomposed by commodity,
examples. F RANK C RITTIN by regarding the flow of all commodi-
IMA-EPFL ties but one as fixed. A continuous al-
coauthor: Michel Bierlaire gorithm provides near optimal solutions.
Inverse 1-median problems in The system and algorithms are designed
the plane keywords: nonlinear equations, large
scale problems, transportation for the real time setting, and has very fast
C ARMEN P LESCHIUTSCHNIG response time and handles a continuous
Graz University of Technology / One of the most basic numerical prob- flow in updated input data. The applica-
Department of Mathematics B lems encountered in computational sci- tion is implemented at the rail freight car-
ence is to find the solution of a system rier Green Cargo, Sweden.
coauthors: Rainer Burkard, Jian-Zhong of nonlinear equations, or equivalently to
Zhang find a fixed point of a given mapping.
keywords: inverse optimization, Due to the advances in computer sciences Optimal shunting
1-median, location problem these problems involve an increasingly
The inverse 1-median problem consists in large number of variables, as for example
changing weights of existing facilities at in transportation or economics. More- U WE Z IMMERMANN
minimum cost such that a pre-specified over it is not uncommon that the Jaco- Dept. Mathematical Optimization. TU
location becomes the 1-median under the bian matrix of the system is unavailable Braunschweig
new weights. The cost is assumed to be or, despite progresses in automatic dif-
proportional to the increase or decrease ferentiation, costly to compute motivat- keywords: combinatorial optimization,
of the corresponding weight. All new ing the use of quasi-Newton or inexact mixed integer programming, logistics
and transportation
 43

Cost for shunting are rather expensive. be assigned to every night duty. Also, convergence. The outcomes of the algo-
How to reduce these efforts to the abso- to balance the workload among the doc- rithm are demonstrated by clinical radia-
lute necessary amount? Nice mathemati- tors, the number of night duties that a tion plans.
cal models lead to various easy and hard doctor works on Saturday, Sunday, and
combinatorial optimization problems. In week nights must also be considered, as MO2-321/033 NETWORKS

view of its applications, shunting occurs well as the number in total. Certain work
as one part of complex railway problems patterns that may adversely affect doc- Integer programming for
in public and industrial transportation. tors’ health, such as consecutive night design problems
We report about models, solved and open duties, are also not permitted. The con- chair : Corinne Feremans
mathematical problems as well as experi- straints of this problem are of block- Truss topology design with bi-
ence in related practical projects. angular structure. This structure consists nary variables
of diagonal blocks of constraints for a
MO2-321/053 PRODUCTION AND SCHEDULING specific doctor that can be dealt with in- A DELAIDE C ERVEIRA
dependently without a set of linking con- UTAD(Trás-os -Montes e Alto Douro
Planning in medical straints. We deal with this problem as a University)-Lisbon’s University
applications variant of the nurse scheduling problem Operational Research centre
chair : Rembert Reemtsen and we solve it by an algorithm based coauthor: Fernando Bastos
Fast simultaneous angle, wedge, on the ‘Subproblem-centric approach’ keywords: semidefinite programming,
and beam intensity optimization for nurse scheduling. The subproblem- Truss topology design, duality
in inverse radiotherapy plan- centric approach improves the objective
function by repeatedly solving subprob- In this talk, we consider a truss topol-
ning ogy design problem, where all the exis-
lems, where the objective function in-
cludes minimizing violations for the link- tent bars have the same cross-sectional
KONRAD E NGEL area. The goal is to find the stiffest
Universität Rostock/Fachbereich ing constraints. This algorithm obtains
good schedules efficiently. truss subject to the equilibrium equation,
Mathematik
a constraint defining the maximum vol-
coauthor: Eckhard Tabbert ume (weight) and integrality constraints
keywords: treatment planning, radiation The solution of nonlinear prob- on the variables. We can formulate the
therapy optimization, beam orientation, lems of intensity modulated ra- problem as a semidefinite problem with
IMRT diation therapy planning by a binary variables. We consider and solve,
We present a new very fast radiotherapy Lagrangian barrier-penalty al- using a freely available SDP software,
planning algorithm which determines ap- gorithm several semidefinite relaxations of that
proximatively optimal gantry and table problem. We use also another approach
angles, kinds of wedges, leaf positions R EMBERT R EEMTSEN based on lagrangean decomposition.
and intensities simultaneously in a global Brandenburg Technical University
way. Other parameters are optimized lo- Cottbus The 2-edge connected subgraph
cally, i.e. independently of each other. coauthor: M. Alber problem with bounded rings
The algorithm uses an elaborated field keywords: radiation therapy planning,
management and field reduction. Beam large scale optimization, nonlinear B ERNARD F ORTZ
intensities are determined via a variant IAG, Université Catholique de Louvain
optimization, barrier-penalty algorithm
of a projected Newton method of Bert- coauthors: Pierre Pesneau, Ali Ridha
sekas. The objective function is a stan- We discuss the solution of large-scale Mahjoub, S. Thomas McCormick
dard piecewise quadratic penalty func- nonlinear optimization problems arising keywords: network design,
tion, but it is built with efficient upper from a biological model for intensity branch-and-cut
bounds which are calculated during the modulated radiation therapy planning in
optimization process. Instead of pencil cancer treatment. The model developed We consider the problem of determin-
beams basic leaf positions are included. elsewhere leads to convex and noncon- ing a 2-edge connected subgraph of min-
The algorithm is implemented in the new vex constrained optimization problems imum cost such that each edge belongs
beam modelling and dose optimization with several thousands of variables, for to a cycle of using at most K edges. We
modul HoMo OptiS. which gradients of the involved func- investigate this problem from a polyhe-
tions are available, but computation of dral point of view. We first give an in-
Hessians is numerically too costly. It teger programming formulation for the
An algorithm for scheduling is suggested to solve the optimization problem using the natural set of vari-
doctors for night duty problems by a modification of an al- ables, and we discuss several classes of
ATSUKO I KEGAMI gorithm by Conn, Gould, and Toint, valid inequalities for the associated poly-
Seikei University which for convex problems and special tope. We characterize the dimension of
choices of parameters coincides essen- that polytope and describe conditions for
keywords: staff scheduling, nurse tially with the nonlinear rescaling method these inequalities to be facet defining.
scheduling, night duty, subproblem for the modified-logarithmic-quadratic Moreover, we study separation routines
centric approach barrier-penalty function by Polyak. In for some of these inequalities. Numeri-
Every month, doctors usually work sev- particular, the subproblems in the algo- cal results obtained with a Branch&Cut
eral night duties in addition to their usual rithm are solved by a conjugate gradi- algorithm are presented.
day duties. Here we deal with the prob- ent method, since the favorable eigen-
lem of scheduling doctors to night duties value structure of the Hessian of the La- The generalized subgraph prob-
at hospitals in Japan. The appropriate grangian at a solution of such problems lem: complexity, approximabil-
number of doctors and skill level must indicates a sufficiently fast (local) rate of ity and polyhedra
44 Parallel Sessions: Monday 13:15 – 14:45 (MO2)

C ORINNE F EREMANS Graphs with convex-QP stabil- A graph property is called elusive (or eva-
Universiteit Maastricht, Faculty of ity number sive) if every algorithm for testing this
Economics and Business Administration, property has to read in the worst case
Department of Quantitative Economics D OMINGOS M OREIRA n
2 entries of the adjacency matrix of
C ARDOSO


coauthors: Martine Labbé, Adam the given graph. Several graph proper-
Departamento de Matemática, ties have been shown to be elusive, e.g.
Letchford, Juan José Salazar González
Universidade de Aveiro planarity (Best et al 1974), k-colorability
keywords: network design
This paper is concerned with a prob- keywords: graph theory, stability (Bollobas 1978), 2-connectivity (Triesch
lem on networks which we call the Gen- number, programming involving graphs 1982), or the membership in any mi-
eralized Subgraph Problem (GSP). The Graphs with convex-QP stability number nor closed family (Chakrabarti, Khot, Shi
GSP is defined on an undirected graph are graphs for which the stability num- 2002). A famous conjecture of Karp
where the vertex set is partitioned into ber is equal to the optimal value of a (1973) says that every non-trivial mono-
clusters. The task is to find a sub- convex quadratic program. It is known tone graph property is elusive. We prove
graph which touches at most one ver- that there is an infinite number of graphs that a non-monotone but hereditary graph
tex in each cluster so as to maximize with convex-QP stability number. For in- property is elusive: perfectness.
the sum of vertex and edge weights. stance, a connected graph G with at least
The GSP is a relaxation of several im- one edge which is not a star neither a tri-
portant problems of a ‘generalized’ type angle has a perfect matching if and only if
and, interestingly, has strong connections its line graph, L G  has convex-QP sta- Matrices of zeros and ones with
with various other well-known combina- bility number. On the other hand, if G is a given line sums and a zero block
torial problems, such as the matching, connected graph   with an even number of
max-flow / min-cut, uncapacitated facil- edges then L L G   has convex-QP sta-
ity location, stable set, quadratic semi- bility number. In this presentation we in- G EIR DAHL
assignment, vertex cover and max-cut troduce several combinatorial and spec- University of Oslo
problems. tral characterizations of graphs for which coauthor: Richard A. Brualdi
In this paper, we examine the GSP it is possible to recognize, in polynomial- keywords:  0 1  -matrices, Gale-Ryser
from a theoretical viewpoint. We show time, if they have (or not) convex-QP sta- algorithm, structure matrix, majorization
that the GSP is strongly NP-hard, but bility number. Furthermore, some hered-
solvable in polynomial time in several itary classes of graphs for which such
special cases. We also give several ap- polynomial-time recognition can be done We study the existence of  0 1  -matrices
proximation results. Finally, we exam- are presented. with given line sums and a fixed zero
ine two 0-1 integer programming formu-
block. An algorithm is given to construct
lations and derive new classes of valid
and facet-inducing inequalities that could
Perfectness is an elusive graph such a matrix which is based on three
property applications of the well-known  Gale-
Ryser algorithm for constructing 0 1  -
be useful to develop a cutting plane ap-
proach for the exact or heuristic resolu- A NNEGRET K ATRIN WAGLER matrices with given line sums. A charac-
tion of the problem. Konrad-Zuse-Institute for Information terization in terms of a certain “structure
Technology matrix” is proved. Further properties of
MO2-321/133 GRAPHS AND MATROIDS

coauthor: Stefan Hougardy this structure matrix are also established,

Graphs and matroids keywords: perfect graphs, elusiveness, and its rank is determined and interpreted
chair : Geir Dahl graph property testing combinatorially.
 45

MO3-302/41 COMBINATORIAL OPTIMIZATION and the M-convex submodular flow prob- We describe a very fast method for find-
lem is one of the most general frame- ing augmenting paths in grid graphs, a
Submodular functions and works of efficiently solvable combinato- problem motivated by the capacity es-
discrete convexity II rial optimization problems. It includes timation in the global routing of VLSI
organizers : Satoru Fujishige and the minimum cost flow and the submod- chips. Nodes in the given graph, as well
Kazuo Murota ular flow problems as its special cases. as edges with positive residual capac-
chair : Kazuo Murota In this talk, we first devise a successive ity, are encoded in arrays of bit vectors.
A fully combinatorial algorithm shortest path algorithm for the M-convex Given these bit vectors, we show how
for submodular function mini- submodular flow problem. We then pro- to find an augmenting path by perform-
mization pose an efficient algorithm based on a ing a series of logical and bit shift op-
capacity scaling framework for the scal- erations on them. Using this method as
S ATORU I WATA able M-convex submodular flow prob-  the core of an almost-maximum (multi-
University of Tokyo lem. Here an M-convex  function f x  is commodity) flow algorithm, we achieve
keywords: submodular function, said to be scalable if f ax  b  is also M- running times reduced by several orders
combinatorial algorithm, strongly convex for any positive integer a and any of magnitude on the real world instances
polynomial algorithm integer vector b. which arise from the capacity estimation
problem mentioned above, compared to
Combinatorial strongly polynomial al- well-known maximum flow algorithms
gorithms have been developed for min- MO3-302/42 COMBINATORIAL OPTIMIZATION

for general graphs, like the one due to

imizing submodular functions indepen- Network flows in VLSI layout Goldberg and Tarjan.
dently by Schrijver and by Iwata, Fleis- organizers : Bernhard Korte and Jens
cher, and Fujishige (IFF). These combi- Vygen
natorial algorithms perform multiplica- chair : Bernhard Korte Multicommodity flow with con-
tions and divisions despite the definition gestion costs
of submodular functions does not involve
Network flow algorithms in
such numerical operations. In this talk
VLSI placement J ENS V YGEN
we present a fully combinatorial variant U LRICH B RENNER University of Bonn
of the IFF algorithm using only addi- Research Institute for Discrete keywords: VLSI design, global routing,
tions, subtractions, comparisons, and or- Mathematics, University of Bonn multicommodity flow
acle calls for the function values. The
keywords: flow algorithms, VLSI Many optimization problems in prac-
algorithm can be applied to submodu-
design tice can be modelled by multicommod-
lar functions over an arbitrary totally or-
dered additive group. In VLSI placement, millions of compo- ity flows. In contrast to the standard
nents have to be placed disjointly in a model, it is often more expensive to use
given area. In its simplest case, this task congested edges. One obvious example
A Gallai-Edmonds-type struc- can
ture theorem for path- lem beandformulated as an assignment prob-
it is natural to apply network
is routing traffic flows with congestion-
dependent travel times, another one -
matchings flow algorithms to it. which we concentrate on - is global rout-
B IANCA S PILLE Most placement tools work in two ing in VLSI design.
University of Magdeburg/ FMA-IMO steps: After roughly spreading out the To solve the (integer) multicommod-
components over the chip ignoring dis- ity flow problem with congestion costs,
coauthor: Laszlo Szego jointness (global placement), the compo- we propose an efficient combinatorial
keywords: path-matching, nents are moved to their final position (le- fully polynomial approximation scheme
Gallai-Edmonds, structure theorem galization or detailed placement). for a fractional relaxation, followed by
As a generalization of matchings, Cun- In this talk, we describe how net- randomized rounding. The overall devia-
ningham and Geelen introduced the no- work flow algorithms can be applied both tion from the optimum can be bounded.
tion of path-matchings. We give a struc- to global and detailed placement. For This is the first approach to global
ture theorem for path-matchings which global placement, our approach leads to routing that takes coupling between adja-
generalizes the fundamental Gallai- minimum cost flow instances on bipar- cent wires, bounds on delays along criti-
Edmonds structure theorem for match- tite graphs with several millions of ver- cal paths, and overall capacitance (power
ings. Our proof is purely combinatorial. tices. However, we can give an algorithm consumption) into account.
that solves these instances efficiently in
near-linear time. For detailed placement,
MO3-302/44 COMBINATORIAL OPTIMIZATION

Capacity scaling algorithm for acomputation new reduction to a minimum cost flow
scalable M-convex submodular galizing a global is described that allows le- Combinatorial optimization II
placement with very chair : Pasquale Avella
flow problems small movement of circuits.
S ATOKO M ORIGUCHI On the facets and the diameter
Tokyo Institute of Technology Fast flow augmentation in bit- of the k-cycle polytope
coauthor: Kazuo Murota pattern-represented grid graphs
keywords: combinatorial optimization,
A NDREI H ORBACH
discrete convex function, submodular
D IRK M UELLER Christian-Albrechts-Universitaet zu Kiel
Research Institute for Discrete coauthors: Eberhard Girlich, Mikhail
flow, algorithms
Mathematics, Univ. of Bonn M. Kovalev
An M-convex function is a nonlinear dis- keywords: multicommodity flow, grid keywords: polyhedral combinatorics,
crete function defined on integer points, graphs, VLSI design, global routing k-cycle polytope, facets, diameter
46 Parallel Sessions: Monday 16:30 – 18:00 (MO3)

The k-cycle polytope represents the con- is effective to limit the growth of the mas- size of subproblems. For example, Tokyo
vex hull of all simple cycles of length ex- ter problem. consists of 56 cities and 25 districts.
actly k in a complete graph. This poly- Several families of cutting planes are The other is the difference in district-
tope generalizes the traveling salesman used to tighten the formulation. Particu- ing rule between Japan and well-studied
polytope (TSP). We present a class of larly we introduce a new family of Cycle countries such as the United States. In
facet inequalities for the k-cycle polytope inequalities, which turned out to be very US, compactness, geographical criteria,
derived from the path inequalities for the effective for several classes of instances. is emphasized, but it is of little signifi-
TSP, new classes of facet inequalities for cance in Japan. It is important not to di-
the 3-cycle polytope, and bounds on the MO3-302/45 COMBINATORIAL OPTIMIZATION vide city into smaller parts, and to find
diameter for the k-cycle polytope. the optimal solution to make an index of
Set partitioning applications gerrymander. Consequently, it is hard
chair : David M. Ryan to find the exact solution, not an ap-
Chromatic scheduling polytopes
coming from the bandwidth al- A column generation ap- proximate solution, with a conventional
location problem in point-to- proach for discrete lotsizing and method. However, we formulated sub-
multipoint radio access systems scheduling problem on identical problems as two combinatorial optimiza-
parallel machines tion problems, the set partition type and
JAVIER L EONARDO M ARENCO the graph partition type, solved them with
Universidad de Buenos Aires J UN I MAIZUMI some interesting idea, and then obtained
coauthor: Annegret Katrin Wagler Toyo University the optimal 300 districts for the first time.
keywords: bandwidth allocation, coauthors: Hiroaki Arai, Susumu
polyhedral combinatorics Morito
Improved solution techniques
The bandwidth allocation problem in keywords: parallel machines, lot sizing,
for multi-period area-based for-
point-to-multipoint radio access systems scheduling, column generation
est harvest scheduling problems
is NP-Hard in general and cannot be ap-
In this paper the Discrete Lotsizing and
DAVID M. RYAN
proximated in polynomial time with a University of Auckland
Scheduling Problem (DLSP) on identi-
guaranteed quality. One kind of algo- coauthors: Juan Pablo Vielma, Alan
cal parallel machines with setup times
rithms which turned out to be successful Murray, Andres Weintraub
is considered. The problem is a one
for many other combinatorial optimiza- keywords: forest harvesting, set packing
of determining the sequence and size of
tion problems uses cutting plane meth- and partitioning, integer programming
production batches for multiple items.
ods. For that, knowledge on the associ- Area-based forest harvest scheduling
The objective is to find a minimal cost
ated polytopes is required. The present models, where management decisions
production schedule such that determin-
work contributes to this issue, providing are made for relatively small forest units
istic, dynamic demand is fulfilled with
an initial study of chromatic scheduling subject to a maximum harvest area re-
backlogging. The problem is formulated
polytopes. We explore that these poly- striction, are known to be very diffi-
as an integer programming problem and
topes are symmetrical, which is a spe- cult to solve by exact techniques. Pre-
reformulated as a Generalized Set Par-
cial property having implications in the vious research has developed good ap-
titioning Problem (GSPP). We present
search for facets. We also present partial proaches for solving small and medium
a column generation heuristic to solve
results on the polytopes’ dimension, and sized forestry applications based on pro-
GSPP. GSPP is decomposed to subprob-
provide some classes of facet-inducing jecting the problem onto a cluster graph
lems and each subproblem is solvable
inequalities. for which cliques can be applied. The re-
with dynamic programming. The quality
of the solutions can be measured, since sulting set partitioning and packing mod-
Computational study of large- the heuristic generates lower and upper els can be solved relatively efficiently.
scale p-median problems bounds. Computational results on a per- However, as multiple time periods be-
sonal computer show that the heuristic is come of interest, current approaches en-
PASQUALE AVELLA counter difficulties preventing successful
Università del Sannio rather effective in terms of quality of the
solutions. identification of optimal solutions. Tim-
coauthors: Antonio Sassano, Igor ber demand constraints which are in-
Vasil’ev tended to smooth production over suc-
keywords: p-median, column-and-row Political redistricting: a case cessive time periods, create very difficult
generation, cutting planes study in japan integer programming problems. In this

Given a directed graph G V A  , the p- K EISUKE H OTTA paper we present an approach for elas-
Median problem consists of determining tizacing the timber demand constraints,
Bunkyo University
p nodes (median nodes) minimizing the which significantly improves the integer
total distance to the other nodes of the coauthor: Toshio Nemoto properties of the LP relaxation and lends
graph. The p-Median problem belongs keywords: political redistricting, itself to an efficient solution technique.
to the class of NP-hard problems. combinatorial optimization, exact It is also possible using this approach
We present a Branch-and-Cut-and- solution to examine trade-offs between objective
Price algorithm yielding provably good We consider the political redistricting value performance and maintaining de-
solutions for instances up to 3795 nodes problem for the 300 single-seat con- mand constraints.
(14 402 025 variables). In our algorithm stituencies of the House of Representa-
tives in Japan. This problem becomes MO3-302/49
INTEGER AND MIXED INTEGER

the LP relaxation is solved by a column- PROGRAMMING

and-row generation: for each variable, 47 redistricting subproblems because the Integer programming column
which is added to the current master 300 single-seats are apportioned to 47 generation II
problem, we add the related Variable Up- prefectures by law. There are two diffi- organizer : Marco Lübbecke
per Bound constraint too. This approach culties to solve them in Japan. One is the chair : Ralf Borndörfer
 47

Robust branch-and-cut-and- The German automobile associa- MO3-306/31 INTEGER AND MIXED INTEGER
PROGRAMMING
price for the capacitated min- tion ADAC (Allgemeiner Deutscher
imum spanning tree problem Automobil-Club), second in size only Integer and mixed integer
to the American Automobile Association programming I
R ICARDO F UKASAWA (AAA), maintains a heterogeneous fleet chair : Pedro Martins
GAPSO of over 1600 units (yellow angels) to help Reformulation techniques based
coauthors: Marcus Poggi de Aragão, people whose cars break down on their on group relaxations
Oscar Porto, Eduardo Uchoa way. As part of their daily work at one of
keywords: integer programming, the five ADAC help centers distributed ROBERT W EISMANTEL
network planning, column generation over Germany, human dispatchers con- University of Magdeburg/ FMA-IMO
stantly face large problem instances of coauthors: Matthias Koeppe, Quentin
We propose a formulation for the capac- the online VDP that they must sove in Louveaux
itated minimum spanning tree problem real time. We present preliminary results keywords: group relaxations
that leads to a robust branch-and-cut-and- on the development of an online algo-
price algorithm. By "robust", we mean rithm based on column generation for The paper addresses the question of how
that neither branching nor the addition automating the dispatching task. and why to use the non-decomposable
of cuts change the structure of the pric- solutions from knapsack group relax-
ing subproblem, a very desirable prop- ations for reformulating integer pro-
erty. The resulting pricing subproblem Combining column generation grams in a higher dimensional space.
is a minimum cost subtree problem with with memetic algorithms We present theoretical justifications and
a capacity constraint. Since such prob- some numerical evidence that this ap-
lem is strongly NP-hard, we generate proach is promising.
columns by solving a relaxation, find-
I VANA L JUBIC
Vienna University of Technology A polyhedral approach to hub
ing a minimum cost connected compo-
nent (not necessarily a tree) satisfying the coauthors: Petra Mutzel, Guenther location problems
capacity. This can be done by a pseudo- Raidl H IROO S AITO
polynomial dynamic programming pro- keywords: column generation, integer
The University of Tokyo
cedure. Our cut generation procedure programming, branch-and-cut-and-price,
uses generalized subtour elimination and applications coauthors: Tetsuya Fujie, Tomomi
root cutset cuts, basically the same ones Matsui
already employed in a branch-and-cut by We present new results on the com- keywords: hub location problem, mixed
L. Hall. Computation results on bench- bination of branch-and-cut-and-price integer programming, polytope, cutting
mark instances from the OR-Library sug- (B&C&P) algorithms with population planes
gest significant improvement over previ- based metaheuristics, in particular with Hubs are switching facilities for trans-
ously known algorithms. The combina- memetic or evolutionary algorithms portation in an airline network or a
tion of column and cut generation con- (MAs). We show how exact algorithms postal delivery system, etc. The prob-
sistently leads to tighter lower bounds, might benefit from metaheuristics, which lem that designs a hub-and-spokes net-
which allowed us to solve to optimality is, to our knowledge, the first survey on work is known as the Hub Location Prob-
several open instances in the set. this topic. lem (HLP). In this talk, we investigate
During the column generation phase, a polytope related to the feasible region
the way how the restricted subproblem of (HLP). We show some polyhedral as-
Online vehicle dispatching for is chosen, plays an important role in the pects of a mixed integer programming
service units design of a B&C&P. We consider the formulation. Furthermore, we propose
subset selection graph problems which a cutting plane method employing new
L UIS M IGUEL TORRES facets. We also show the strength of the
Konrad-Zuse Zentrum Berlin can be rephrased as 0/1 integer program-
ming problems, with one-to-one corre- cuts through computational experiments.
coauthors: Martin Groetschel, spondence between variables and edges
Benjamin Hiller, Sven O. Krumke, Jörg of the underlying graph. We usually
Rambau start from a set of edges/variables repre- Capacitated minimum spanning
keywords: vehicle routing, on-line senting a sparse graph. We maintain an tree problem: revisiting hop-
optimization, column generation additional set of promising edges (mak- indexed models
Given a set of guided service vehicles ing a so called reserve graph) first to
(units), a set of service requests arriving be checked with respect to their reduced
P EDRO M ARTINS
Coordinate Professor / ISCAC and CIO /
over time, and a set of service contrac- costs.
The standard method of B&C&P to Polytechnic Institute of Coimbra
tors (conts), the online vehicle dispatch-
ing problem VDP is the task of finding an create sparse and reserve graphs is to use coauthor: Luis Gouveia
assignment of requests to units and conts, k-nearest neighbor graphs. We suggest to keywords: capacitated tree problem,
as well as service tours for all units, such run a MA, and to take all edges present in subtour elimination constraint,
as to minimize a certain objective func- the last population into the sparse graph. root-cutset constraints, cutting plane
tion. The objective measures the ser- Additional edges which are present in algorithm
vice, overtime and driving costs for the earlier stages of the MA form the reserve We introduce hop-indexed generaliza-
units. Requests assigned to contractors graph. tions of the well known subtour elimina-
are charged with fixed costs. Moreover, Experimental comparisons show that tion constraints as well as the root-cutset
requests specify service deadlines, and the usage of MAs improves standard constraints and consider them in the con-
late costs are incurred whenever a dead- branch-and-cut and B&C&P approaches. text of the hop-indexed Model of Gou-
line is missed. veia and Martins. The inclusion of the
48 Parallel Sessions: Monday 16:30 – 18:00 (MO3)

two new classes of constraints has per- A catalytic surface reaction is considered in which stochastic quantities are treated
mitted us to solve instances not previ- whose chemical reactants are given in a as random variables and constraints are
ously solved in the literature. well mixed gas phase and can adsorb at a probabilistic. The design variables in
catalytic surface which subsequently en- RBDO problems are the means of the
MO3-306/32 NONLINEAR PROGRAMMING ables a reaction step due to lowering the random variables, while standard devia-
activation energy. Catalysts are usually tions are assumed to be known and con-
Selection and design very expensive so that it is of natural in- stant. The tools available for assessing
chair : Jens Starke terest to reduce the amount of catalyst, the reliability of satisfying the probabilis-
Local search methods for the i.e. to distribute a certain amount of a tic constraints, namely, quantifying the
subset selection problem with catalyst material on a cheaper carrier ma- probability of failure, require knowledge
minimum unit norm terial such that the output of the chem- of the statistical properties of the stochas-
ical reaction is maximized. For the ex- tic quantities that are inputs to the simu-
DAG H AUGLAND ample of the CO-oxidation on platinum lations. These are assumed to be known
Dept. of Informatics, University of surfaces reaction-diffusion equations are only at the bottom of the hierarchy. An it-
Bergen presented which well describe the exper- erative linearization algorithm is used to
coauthor: Sverre Storøy imentally observed rich dynamical prop- propagate them to higher levels in the hi-
keywords: subset selection, local erties including pattern formation. Bifur- erarchy. In this manner, we can quantify
search, pivoting cation analysis provides the local dynam- system output variations, which would
ical properties in dependence of the pa- not be identified accurately when consid-
Given a target vector of length m and a
rameters representing the respective ma- ering an “all-at-once” solution approach
matrix with m rows and n columns, we
terial of the surface compositon. Based in the presence of nonlinearities.
consider the problem of finding a subset
on these results concerning the dynamics
of no more than k columns such that a
of the chemical reaction, the output (CO2
closest possible approximation to the tar-
production) is maximized in dependence Stable noisy optimization by
get can be found in the subspace spanned
of the material compositon (catalyst or branch and fit
by the selected columns. The problem is
carrier) on the surface. The difficulty is
NP-hard when closeness is measured by
that the objective function depends on the WALTRAUD H UYER
any p-norm, and in the current work we Institut f. Mathematik, Univ. Wien
solution of a partial differential equation
suggest some local search techniques for coauthor: Arnold Neumaier
(reaction-diffusion equation) which itself
the unit norm case. The neighborhood keywords: derivative-free optimization,
depends on the high-dimensional vari-
structures on which the search is based, noisy function values
able representing the discretized material
are defined in terms of pivoting opera-
composition on the surface. Optimiza-
tions in a related linear program. Several We describe an algorithm for bound con-
tion results obtained by first a determin-
variants of the local search technique are strained optimization, where derivatives
istic and second a stochastic approach are
presented and compared through numer- are not available and the function values
presented.
ical experiments. are expensive and subject to noise. It pro-
ceeds by successive partition of the box
MO3-306/33 NONLINEAR PROGRAMMING
(branching) and building local quadratic
Computing the center of area of
a convex polygon Noisy problems models (fit) and combines global and lo-
chair : Antanas Zilinskas cal search. Trust regions are used for
L AURA N ICOLETA local optimization from the best point.
H EINRICH -L ITAN Analytical target cascading in In each step, a user-specified number of
Braunschweig University of Technology design optimization of hierar- suggested evaluation points is generated.
chical multilevel systems under Moreover, the algorithm allows for hid-
coauthors: Peter Brass, Pat Morin uncertainty
keywords: computational geometry den constraints and assigns to such points
The center of area of a planar convex set M ICHAEL KOKKOLARAS
a function value based on the function
University of Michigan values of nearby feasible points, and soft
C is the point p for which the minimum constraints are handled with the aid of
area cut off from C by any halfplane con- coauthor: Panos Y. Papalambros
keywords: design optimization, penalty and barrier functions.
taining p is maximized. Properties of this
point were studied already long ago in multilevel systems, target cascading,
classical geometry, but it is quite non- uncertainty propagation One dimensional global opti-
trivial to really determine this point for Analytical Target Cascading (ATC) is mization in the presence of noise
a given convex set. We describe a sim- a methodology for design optimization
ple randomized linear-time algorithm for of hierarchical multilevel systems. The A NTANAS Z ILINSKAS
computing the center of area of a convex original deterministic formulation of the Cardiff University UK, IMI Lithuania
n-gon. associated optimization problems does coauthor: James Calvin
not take into account uncertainties with keywords: global optimization,
Surface design for heteroge- respect to design variables and analy- statistical models, noisy observations,
neous catalytic reactions sis/simulation model parameters. There- testing
fore, the obtained optimal designs, which
J ENS S TARKE often lie on the boundaries of the ac- The problem of global optimization of
University of Heidelberg, Inst. of tive constraints, are questionable because multimodal functions in the presence of
Applied Math small variations may generate infeasibil- noise is one of the most difficult in op-
keywords: reaction diffusion equations, ity. In this work, we modify the ATC timization theory. An approach based on
differential equations, structural process to include reliability-based de- statistical models of an objective function
optimization, material optimization sign optimization (RBDO) formulations, is well-suited for incorporation of noisy
 49

observations. However, implementations keywords: protein folding, linear In this talk we show transition to su-
of algorithms based on statistical models programming, MAX FS, support vector perlinear local convergence for a modi-
are complicated. machines fied version of a trust-region filter-SQP
We consider a problem of one- We discuss the use of optimization tech- method for nonlinear programming intro-
dimensional global optimization in the niques such as linear programming and duced by R. Fletcher, S. Leyffer, and Ph.
presence of noise. A standard Wiener support vector machines for the design L. Toint.
process is accepted for a model of ob- of protein folding potentials. The key Remarkably the algorithm allows
jective functions. Noise is modeled by computational task lies in solving linear transition to superlinear local conver-
independent Gaussian random variables inequality systems with tens of millions gence for the original trust-region SQP-
with zero mean and known variance. To of constraints in a few hundred variables. steps without an additional second or-
construct a statistical model based al- The modeling approach and the resolu- der correction. The main modification
gorithm conditional mean and variance tion of computational challenges as well consists in using the Lagrangian function
should be calculated repeatedly. A spe- as results on "real" problems will be pre- value instead of the objective function
cial recurrent algorithm is developed to sented. value in the filter together with an appro-
calculate these characteristics of the sta- priate infeasibility measure. Moreover,
tistical model whose computational com- we show that the modified trust-region
plexity is of lower degree than that of Semidefinite programming for filter-SQP method has the same global
standard method. In the previous pub- electronic structure calculations convergence properties as the original al-
lications on similar subject the conver- gorithm of R. Fletcher, S. Leyffer, and
gence was considered assuming the fea-
BASTIAAN J. B RAAMS Ph. L. Toint.
Courant Institute, New York University
sible region finite, i.e. the minimization
interval has been discretized. Such an keywords: semidefinite programming,
quantum chemistry An interior-point filter line-
assumption greatly simplifies the analy- search method for large-scale
sis and implementation. However, the The core problem in quantum chemistry nonlinear programming
additional error implied by discretization is the computation of the ground state en-
may not always be acceptable. In the ergy of a system of electrons in an ex- A NDREAS WAECHTER
present paper we consider a version of a ternal potential. Reduced density matri- IBM T.J. Watson Research Center
P-algorithm searching over a continuous ces play an important role in the search coauthor: Lorenz Biegler
interval. Implementation issues are con- for methods of electronic structure cal- keywords: nonlinear optimization, filter
sidered and testing results are included. culation that offer a systematic route to methods, interior point method
better accuracy. In particular, the 2-
body reduced density matrix (2-RDM) Recent improvements of a barrier method
MO3-306/34 NONLINEAR PROGRAMMING

contains enough information to express for large-scale continuous nonlinear non-

Optimization applications in exactly (through a known linear func- convex optimization will be presented.
biology and chemistry tional) the electron Hamiltonian and all The algorithm computes search direc-
organizer/chair : Todd Munson other ground-state properties of interest. tions from a linearization of the primal-
However, the constructive use of the 2- dual equations, and global convergence
Computing transition states RDM is hindered by the problem of N- is guaranteed by a filter line-search pro-
representability that was first clearly de- cedure. Some details of the implemen-
TODD M UNSON tation, particularly of a new feasibility
Argonne National Laboratory scribed and studied by Coleman and by
Garrod and Percus. We describe re- restoration phase, will be discussed. Nu-
coauthor: Jorge More’ cent work that uses semidefinite pro- merical results on a large set of test prob-
keywords: transition states, saddle gramming to solve the variational prob- lems, as well as real-life applications
points, nonlinear optimization lem for the 2-RDM subject to a sub- (such as tuning of transistors in integrated
The calculation of transition states is a set of representability conditions, and we digital circuits) with problem sizes with
fundamental problem in computational discuss the quality of the resulting ap- up to several hundred thousand variables,
chemistry where the transition state cor- proximation. This is joint work with will be presented. The code is released
responds to a mountain pass between to Mituhiro Fukuda, Michael Overton, and as open source under the name IPOPT at
minimizers of a potential energy func- Zhengji Zhao. Supported by NSF/DMS- www.coin-or.org.
tion for a chemical reaction. An algo- 0113852.
rithm based on nonlinear optimization for On the convergence of a filter al-
finding mountain passes is presented with MO3-306/35 NONLINEAR PROGRAMMING
gorithm with independent feasi-
some convergence results. Numerical re- Recent developments in filter bility and optimality phases
sults for this algorithm on benchmark methods II
problems in computational chemistry are E LIZABETH K ARAS
organizers : Michael Ulbrich, Stefan Federal University of Parana
given.
Ulbrich and Luís N. Vicente
chair : Luís N. Vicente coauthors: Clovis C. Gonzaga, Marcia
Design of protein folding po- Vanti
tentials using mathematical pro- On the superlinear local conver-
gence of a filter-SQP method
keywords: filter methods, nonlinear
gramming programming, global convergence
S TEFAN U LBRICH In this talk we describe a filter algo-
M ICHAEL WAGNER Zentrum Mathematik, Technische rithm for nonlinear programming based
Cincinnati Children’s Hospital / Univ. of Universitaet Muenchen on independent restoration and tangen-
Cincinnati cial steps. Some simple conditions on
keywords: nonlinear programming,
coauthors: Jarek Meller, Ron Elber filter, SQP, superlinear convergence the efficiency of each of these steps lead
50 Parallel Sessions: Monday 16:30 – 18:00 (MO3)

to globally convergent algorithms. We S AMIR E LHEDHLI Y URIY Z INCHENKO

show how to ensure that these conditions University of Waterloo Cornell University
are satisfied by available codes for the keywords: Lagrangean relaxation, keywords: hyperbolic polynomials,
subproblems, discuss how to deal with ACCPM, branch-and-price elementary symmetric functions, dual to
Maratos effect and describe a practical the hyperbolicity cone
application to the Optimal Power Flow When Lagrangean relaxation is used,
obtaining the best Lagrangean bound Elementary symmetric functions (poly-
problem.
amounts to solving a non-differentiable nomials) are building blocks for hy-
optimization (NDO) problem. Within perbolic polynomials. While one can
MO3-306/36 NONSMOOTH OPTIMIZATION
easily construct a logarithmic self-
a Lagrangean-based branch-and-bound
Recent advances in framework (branch-and-price), the calcu- concordant barrier (SCB) functional
non-smooth optimization II lation of the Lagrangean bound is done for the hyperbolicity cone K p associ-
organizers : Jean-Philippe Vial and at every node of the branch-and-bound ated to an arbitrary hyperbolic polyno-
Jean-Louis Goffin tree. Thus, the success of such algo- mial p, little is known about the dual
chair : Jean-Philippe Vial rithms relies heavily on the fast solu- cone K p , even when p is the k-th
tion of these NDO problems. We show derivative (1  k  n 2) in the direc-
Using the central path for com- how ACCPM (an interior-point cutting tion e  1 1    1  of the polynomial
puting the analytic center of plane method) exploits past information x1


xn , the standard hyperbolic polyno-
apolytope and quickly solves the NDO problems mial for the nonnegative orthant.
C LOVIS C. G ONZAGA both after adding cuts and after branch- For the case k    1 – that is, for
Federal University of Santa Catarina ing, using a primal and a dual interior- the polynomial p x   Πn  1 x  
point method respectively. Promising nu- ∑1  i  n ∏ j   i x j – we give an algebraic
coauthor: Marli Cardia characterization of the dual cone and
keywords: cutting plane methods, merical results are provided.
show how one can easily construct an
analytic center, path following SCB functional for it. We also indicate
algorithms MO3-306/37 CONVEX PROGRAMMING
extensions of the result.
We describe an algorithm for comput- Semidefinite and conic
ing the analytic center of a polytope af- optimization Optimal mean-variance upper
ter the addition of a deep cut. When deep bounds on the expectation of
cuts are considered, the greatest difficulty chair : Karthik Natarajan
highest order statistics
lies in retrieving feasibility to initiate a Nested optimization for experi-
centering algorithm. Our strategy con- mental inference problems K ARTHIK NATARAJAN
sists in starting at a nearly centered point Singapore-MIT Alliance
for the old polytope, and then following L AURA W YNTER coauthors: Dimitris Bertsimas,
the primal-dual central path for an auxil- IBM Watson Research Center Chung-Piaw Teo
iary linear programming problem. When coauthors: Zhen Liu, Cathy H. Xia, Fan keywords: mean, variance
this path following algorithm stops, the Zhang In this research, we provide a new ap-
resulting point is a nearly central point keywords: estimation from proach to solving the classical problem in
for the new polytope. We show that experiments, inference, parameter statistics of computing the tightest pos-
this method performs well for shallow or estimation, semidefinite quadratic progr. sible upper bound on the expectation of
deep cuts, and prove that its application the maximum of n real random variables
to the convex feasibility problem matches We present a parameter estimation, or in- given mean and variance information on
the best existing complexity bounds. ference, problem, obtained from an ap- the individual random variables. This op-
plication in queueing networks. The de- timal upper bound is computable in poly-
Nonlinear multicommodity flow rived model is a semi-definite quadratic nomial time by solving a semidefinite
and traffic equilibrium program, whose solution gives the pa- program. By further exploiting the struc-
rameters of the underlying (queueing) ture, we reformulate the problem as an
F RÉDÉRIC BABONNEAU system, for any particular experiment. univariate convex minimization problem.
Université de Genève Each experiment provides the input data An efficient bisection search algorithm is
coauthor: Jean-Philippe Vial of each QP. proposed as a solution technique. Us-
keywords: Lagrangian relaxation Given a set of many such experi- ing the formulation, we propose two new
We formulate the lagrangian relaxation ments, we would like to best estimate the closed form bounds, that are important
of the nonlinear multicommodity flow parameters of the inference model. The due to simplicity and also compare well
problem (nlmcf) as a sum of a smooth non-strict convexity of the quadratic ob- with previously known bounds based on
convex function and a nondifferentiable jective poses the problem of nonunique- preliminary empirical testing. An appli-
one. We extend the proximal analytic ness of the solution for any single ex- cation of the work in an option pricing
center cutting plane method to handle the periment, and therefore complexifies the context will also be discussed.
smooth term directly. We apply method task of combining the obtained parame-
to large scale problem. Since Wardrop ters from all experiments. We propose MO3-306/38 PARALLEL COMPUTING

traffic equilibria can reduced to the solv- a series of nested optimization problems, Parallel methods for
ing of nlmcf, we specialise our nlmcf and their properties as well as two vari- combinatorial and global
method to this particular case. ants, for determining the best parameters
for the set of experiments. optimization
organizer/chair : Jonathan Eckstein
ACCPM-based branch-and-
bound methods for integer pro- The dual cone for the derivative Parallel constrained shortest
gramming of the non-negative orthant path
 51

D IEGO K LABJAN MO3-308/11 Crew scheduling problems at the plan-

AIRLINE OPTIMIZATION APPLICATIONS

University of Illinois at ning level are typically solved in two

Urbana-Champaign Airline scheduling II steps: first, creating working patterns,
keywords: shortest paths, parallel organizer/chair : Stefan Karisch and then assigning these to individual
computing crew. The first step is solved with a set
Large scale airline crew roster- covering model, and the second with a
Parallel algorithms for shortest path have ing problems in practice set partitioning model. At the operational
been extensively studied in the past. To
the contrary, we are not aware of paral- M AREK O ZANA level, the (re) planning period is consid-
lel algorithms to the constrained shortest Carmen Systems AB erably smaller than during the strategic
path problem. We propose several algo- coauthor: Sami Spjuth planning phase. This paper presents an
rithms and variants suitable for acyclic keywords: real world application, integrated model to solve time critical re-
graphs. We present computational exper- combinatorial optimization covey problems arising on the day of op-
iments showing additional issues that are erations. we describe how pairing con-
The Carmen Crew Rostering system struction and pairing assignment are done
not present in parallel shortest path im- could efficiently solve crew assignment is a single step. We describe solution
plementations. problems with up to 3000 crew members. techniques based on simple tree search
We present heuristic enhancements to the and more sophisticated column genera-
Computational grid imple- standard generate and optimize approach tion and shortest-path algorithms
mentations of branch-and- used in the production system. The im-
bound methods for nonconvex provements are based on a greedy ini-
quadratic programming tial algorithm and a stochastic trip lock-
ing mechanism. The current system is MO3-308/12
STOCHASTIC PROGRAMMING

J EFF L INDEROTH
Lehigh Univeristy capable of solving large scale rostering Real-life applications in
problems of up to 6500 crew to produc- stochastic programming
keywords: global optimization, tion quality in less than 14 hours.
computational grid, parallel computing, chair : Sin C. Ho
branch-and-bound
A stochastic programming
We describe branch-and-bound methods A proximal analytic center cut- model for strategic capacity
for solving nonconvex quadratic pro- ting plane method to solve the planning in the metal-working
grams. The methods make use of novel Lagrangian relaxation of the industry
relaxations of the nonconvexities and are aircraft rotation problem
implemented to run on a Computational O LIVIER DU M ERLE
Grid computing platform. Solution of Air France N INA L INN U LSTEIN
large-scale instances will be reported. Norwegian University of Science and
coauthors: Sébastien Lemaire,
Jean-Philippe Vial Technology, NTNU
A framework for scalable paral- keywords: proximal analytic center,
lel tree search coauthors: Marielle Christiansen, Bjørn
column generation, Lagrangian Nygreen, Kari M. Reikvam, Marianne
T ED R ALPHS relaxation, aircraft rotation problem Storhaug Jensen
Lehigh University Proximal ACCPM is a variant of the an- keywords: stochastic programming,
coauthors: Laszlo Ladanyi, Matthew alytic center cutting plane method, in facility location
Saltzman, Yan Xu which a proximal term is added to the
keywords: parallel algorithm, barrier function that defines the center. In Dramatic changes in the markets for sil-
branch-and-bound, branch-and-cut, this talk we will give a presentation of the icon and ferrosilicon metals in recent
software framework method and of its implementation. We years, has resulted in major restructuring
We describe our latest progress devel- will apply the method to solve the La- processes in the industry. We present a
oping the Abstract Library for Parallel grangian relaxation of a real life problem: stochastic programming model used by a
Search (ALPS), a software framework the aircraft rotation problem. This prob- large metals company with a number of
for implementing parallel tree search al- lem arises after we have considered the production facilities in Europe and North
gorithms. ALPS is designed specifically schedule design and the fleet assignment. America. The model support strategic ca-
to achieve scalability for methods such as The aim of this problem is to built suc- pacity planning in the company. It mod-
branch and cut in which large amounts of cessions of flights according to the num- els decisions for acquisitions and clo-
"knowledge," such as dynamically gen- ber of aircrafts and some operational con- sures of plants, investments in new equip-
erated valid inequalities, are generated straints. Numerical results concerning ments and production processes and in-
and must be shared as the algorithm pro- real problems are given comparing prox- vestments in and operation of furnaces.
gresses. Sharing this knowledge effec- imal ACCPM and column generation ap- The model is implemented as a large MIP
tively is the main challenge of achieving proaches. and solved by B&B. The robust solutions
parallel efficiency in such algorithms. In produced by the model are more at terms
addition to standard facilities for imple- Airline crew scheduling: from with the philosophy of the company than
menting search algorithms, ALPS pro- planning to operations deterministic solutions for individual sce-
vides a framework for defining new types narios.
of knowledge, along with methods for C LAUDE P H . M EDARD
storing and sharing this knowledge effi- Carmen Systems
ciently. Additional layers built on top of coauthor: Nidhi Sawhney
ALPS allow the user to implement spe- keywords: crew recovery scheduling, Optimization based derivative
cific classes of algorithms based on these pairing/rostering integration, leg base pricing in incomplete and im-
general principles. rostering perfect markets
52 Parallel Sessions: Monday 16:30 – 18:00 (MO3)

J ÖRGEN B LOMVALL the expected value function (EVF) of this specified threshold. Uncertainty arises
Linköpings universitet problem is non-convex in general. because the interdictor has incomplete
keywords: stochastic programming, We identify a class of probability knowledge of the topological or numer-
derivative pricing density functions (pdf) which yield con- ical data for the network, or because in-
vexity of the EVF. Moreover, by choos- terdiction attempts may fail. We de-
With an optimization based framework, ing a suitable pdf from this class to ap- vise and implement a row generation al-
derivatives can be priced in a general set- proximate any given pdf, we obtain con- gorithm for the basic formulation and
ting by using utility functions. Within vex approximations of the corresponding use it to solve several problem vari-
this framework pricing algorithms can be EVF. In doing so, we use the same ap- ants. We present computational results,
developed. We will present some prelim- proach that was successful for pure inte- and discuss extensions of the modeling
inary results from pricing in both incom- ger recourse models, also with more gen- paradigm to other interdiction problems.
plete and imperfect markets. eral recourse structures.
MO3-308/T1 FINANCE AND ECONOMICS

Designing routing zones for A stochastic programming ap- Finance and economics
VRP with stochastic demands proach to the generalized as-
signment problem chair : Alan King
S IN C. H O
Department of Informatics, University of J OYCE Y EN CAMBO: combinatorial auc-
Bergen University of Washington
tion using matrix bids with or-
der
coauthors: Dag Haugland, Gilbert coauthors: Berkin Toktas, Zelda B.
Laporte Zabinsky S. R AGHAVAN
keywords: vehicle routing, stochastic keywords: stochastic programming, University of Maryland
programming, districting, heuristics integer programming coauthor: Robert W. Day
We consider the problem of partitioning The Generalized Assignment Problem keywords: auctions
a set of customers into districts such that (GAP) addresses the assignment of a
all customers within a district are ser- Since combinational auctions allow bid-
given set of tasks to a set of agents, which
ders to bid on any combinations of goods,
viced by exactly one vehicle. The alloca- use a single resource that is individu-
tion of customers to districts is constant the bid data can be of exponential size.
ally capacitated for each agent. In this
over time, even if an occurring demand To combat this problem, restrictions have
study, we introduce the Collectively Ca-
been proposed limiting the number of
pattern indicates that a regrouping of the pacity Multi-Resource Generalized As-
customers is cost effective. When the dis- bundles a bidder is allowed to bid on (as
signment Problem (CCP) as an extension
tricts, or routing zones, are defined, the in FCC Auction #31). We describe a new
to the GAP, in which the objective is to
combinatorial auction format, CAMBO,
actual demand is observed, and a travel find a minimum-cost assignment of tasks
plan must be produced within each dis- for conducting combinatorial auctions
to agents that collectively use multiple
trict. that restricts bidders to submit a fixed
capacitated resources. We address the
number of "matrix bids with order". The
In this work, we propose a heuris- CCP with uncertain resource capacities,
advantage of this (restricted preference)
tic solution method for solving the out- where the resource capacities have an un-
approach is that it provides bidders a
lined problem. The procedure first finds known distribution that can be sampled.
mechanism to compactly express bids on
an initial feasible solution, and next it We propose four stochastic programming
every possible bundle. We describe many
improves the solution by a sequence formulations of this problem, which dif-
different types of preferences that can be
of moves to neighboring solutions of fer mainly in their recourse models. We
modeled using matrix bids and show that
smaller costs. We also give an analysis identify subgradient search techniques
matrix bids are quite flexible, support-
of experimental results on problems with that are based on the Lagrangian relax-
ing additive, subadditive, and superaddi-
100 customers. ation of the capacity constraints to solve
tive preferences simultaneously. Further,
each formulation, and provide numerical
auctions that restrict the number of bun-
MO3-308/13 STOCHASTIC PROGRAMMING results on a number of test cases.
dles can be considered a special case of
Stochastic integer programs CAMBO. We model the NP-hard winner
and network interdiction A stochastic network- determination problem for CAMBO as a
organizer : COSP (Maarten H. van der interdiction problem: maximiz- polynomially-sized integer program con-
Vlerk) ing the probability of sufficient sisting of assignment constraints and side
chair : Stein W. Wallace delay constraints that enforce ordering. We de-
scribe a computationally effective tech-
On six-pack recourse: convex K EVIN W OOD nique for solving this problem consist-
approximations for a simple Naval Postgraduate School ing of the addition of cover inequalities,
mixed-integer recourse model coauthors: Harald Held, David an objective perturbation heuristic, and a
Woodruff specialized branching scheme. Computa-
M AARTEN H. VAN DER V LERK keywords: stochastic programming, tional experiments will be discussed.
University of Groningen network interdiction, shortest paths
keywords: stochastic programming,
mixed integer recourse
We describe a new model formulation Parametric optimization meth-
for the following stochastic network- ods for visualization of frontier
We consider a mixed-integer recourse interdiction problem: Using limited re- in DEA and productivity analy-
model with simple recourse structure and sources, interdict network nodes or arcs sis of Russian banks
continuously distributed right-hand side in order to maximize the probability that
parameters. Due to the integer variables, the shortest-path length is greater than a V LADIMIR K RIVONOZHKO
 53

Institute for Systems Analysis of Russian u  U, f : E

R is a convex continuous, show that the refined solution always ex-
Academy of Sciences 
but not necessarily differentiable func- ists, contrary to many well known refine-
coauthors: Oleg B. Utkin, Andrei tion. In addition, we assume that h u
 ment concepts in the field.
Volodin is convex for each u  E. Such EPs have
keywords: DEA a great number of applications in Eco- MO3-308/T3 MODELING LANGUAGES AND SYSTEMS

nomics, Mathematical Physics, Trans- Modeling languages and

The proposed family of paramet- portation, Operations Research and other
ric optimization methods constructs systems II
fields. Moreover, they are closely re-
cross-sections of the frontier by the lated to other general problems of Non- organizer/chair : Robert Fourer
two-dimensions plane or by three- linear Analyses. In particular, EP(*) gen-
dimensional hyperplane. This approach
Large-scale optimization with
visualizes the frontier in multidimen-
eralizes fixed point, saddle point, com- TOMLAB / MATLAB
plementarity and variational inequality
sional space of inputs and outputs and problems. We propose to apply the D- K ENNETH H OLMSTRÖM
reduces the analysis of complex systems gap function approach, which was pro- Dept. of Mathematics and Physics /
to the investigation of the well-known posed for mixed variational inequalities Mälardalen University
functions in economics such as produc- in a finite-dimensional space, to the solu- coauthor: Anders Göran
tion function, isoquant, isocost and so tion of EP(*). We show that, under cer- keywords: optimization software,
on. Moreover, we can calculate various tain additional assumptions, EP(*) can be MATLAB, nonlinear programming,
economic factors in DEA applications, converted into a problem of finding a sta- integer programming
for example, marginal rate of substitu- tionary point of a differentiable function, The TOMLAB Optimization environ-
tion, marginal product, elasticity and so which is nothing but a nonlinear equa- ment in MATLAB (http://tomlab.biz),
on. We apply our approach to efficiency tion. Being based on this property, we first presented at ISMP97, has seen a
analysis of Russian banks. describe a descent type algorithm whose tremendous growth during the last years.
iteration sequence converges strongly to With the help of MEX-file interfaces
Calibration of options prices a solution to the initial problem. most state-of-the-art optimization soft-
to market data using stochastic ware has been hooked up to TOMLAB,
programming e.g. KNITRO, SNOPT and CONOPT for
Perfection and stability of sta- large-scale nonlinear programming, and
A LAN K ING tionary points with applications CPLEX and Xpress-MP for large-scale
IBM Research to noncooperative games mixed-integer programming. The unified
coauthors: Teemu Pennanen, Matti input-output format in TOMLAB makes
Koivu D OLF TALMAN it easy to use and compare the solvers.
keywords: options prices Dept. of Econometrics & OR, Tilburg This development also makes MATLAB
We develop a model for the calibration University a possible tool for large-scale optimiza-
of options prices to market data that re- coauthors: Gerard van der Laan, Zaifu tion and the talk will discuss some inter-
quires no assumptions about the underly- Yang esting TOMLAB user examples, experi-
ing stochastic price processes other than keywords: stationary point, perfectness, ence, solver comparisons and results.
that they be free of arbitrage. The im- stability
plemention of this model is a stochastic The open outer approximation
linear program that can be solved using Let be given an upper-semicontinuous MINLP solver in AIMMS
standard packages. mapping from a convex, compact set X in J OHANNES B ISSCHOP
Euclidean space to the collection of non- Paragon Decision Technology B.V.
empty, compact, convex sets in the same
MO3-308/T2 COMPLEMENTARITY AND INEQUALITIES
VARIATIONAL

space. In this paper we refine the solution keywords: modeling system, open
Complementarity and concept of stationary point by perturbing algorithm
variational inequalities simultaneously both the set X and the so- Instead of offering a black-box imple-
lution concept. In case a stationary point mentation of the outer approximation
chair : Dolf Talman
is the limit of a sequence of perturbed so- algorithm for Mixed-Integer NonLinear
Application of the d-gap func- lutions on a sequence of sets converging Programming (MINLP) models, AIMMS
tion approach to geheral equlib- continuously to X we say that the station- offers an open technology to allow for
rium problems in Banach spaces ary point is stable with respect to this se- customized and efficient implementa-
quence of sets and the mapping which de- tions of the algorithm. A novel feature is
O LGA P INYAGINA that each variant of the algorithm is writ-
Kazan State University fines the perturbed solution. It is shown
that stable stationary points exist for a ten in the AIMMS modeling language it-
coauthor: Igor Konnov self. The presentation will introduce, mo-
large class of perturbations. A specific
keywords: equilibrium problems, tivate and demonstrate this new technol-
refinement, called robustness, is obtained
nonsmooth functions, Banach spaces ogy.
if a stationary point is the limit of station-
We consider a general equilibrium prob- ary points on a sequence of sets converg-
lem (EP) of the form: Find an element ing to X. We discuss several applications Update on the AMPL/solver in-
u  U such that in noncooperative game theory. We show terface library
h u
 v  f v  f u

0  v U  that both perfect Nash equilibrium and DAVID G AY
proper Nash equilibrium are special cases AMPL Optimization LLC
where U is a nonempty closed and con- of our robustness concept. Applying our coauthor: Robert Fourer


vex subset of a real Banach space E, results to the field of evolutionary game
h : E  E R is an differentiable
keywords: nonlinear programming,
 equlib- theory yields a refinement of the station- linear programming, stochastic
rium bifunction, i.e. h u u   0 for every ary points of the replicator dynamics. We programming, complementarity
54 Parallel Sessions: Monday 16:30 – 18:00 (MO3)

Many solvers, particularly nonlinear A globally convergent (to KKT points), The facility location problem described
ones, have idiosyncratic requirements for locally quadratically convergent primal- in this paper comes from an industrial ap-
how problems are presented to them and dual interior-point method for nonlin- plication with economics of scale. The
how certain of their features are exer- ear programming is refined for the case objective function is first concave but
cised. The AMPL modeling language where the objective function is quadratic changes to convex when marginal costs
and environment seeks to provide a sim- and the constraints linear. Close to a start to increase.
ple view of solvers that hides the id- strong minimizer, the primal-dual New- We present an approach based on lin-
iosyncrasies. The AMPL/solver interface ton direction for the set of perturbed KKT earization of the facility costs and La-
library facilitates this by providing in- conditions is well defined and, for small grangean decomposition. We also de-
terface routines and data structures that enough values of the perturbation param- velop a greedy heuristic to find upper
simplify the task of providing a solver eter µ, its primal component is a direc- bounds. We use the method to solve
what it needs while letting AMPL users tion of descent for the objective func- a problem instance for the Norwegian
maintain a simplified view of the solver. tion. Away from such solutions, the same Meat Co-operative and compare our re-
This talk reviews some of these facili- holds provided the Hessian matrix is suit- sults to previous results achieved using
ties, focusing in particular on recent and ably modified by adding to it a small pos- standard branch-and-bound in commer-
forthcoming additions, such as suffix no- itive definite matrix. An efficient scheme cial software.
tation for auxiliary information and ex- for determining an appropriate such cor- We examine a location-allocation
tensions for dealing with constraint pro- rection is proposed and analyzed. Special problem focusing on the location of
gramming. Sometimes a solver inter- consideration is given to the “affine scal- slaughterhouses, their size and the allo-
face must do a custom "tree walk" of ex- ing” case (µ  0). cation of animals in the different farming
pression graphs. Examples include in- districts to these slaughterhouses. Inves-
terfaces to some global optimization and tigations show that the slaughterhouse in-
constraint programming solvers.
Primal-dual interior-point
methods based on a new class dustry experiences economics of scale in
of barrier functions the production facilities, due to high in-
MO3-341/21 INTERIOR POINT ALGORITHMS
vestment costs. The model is general and
Primal-dual algorithms M OHAMED E L GHAMI also have applications within other indus-
Tudelft tries that experience economics of scale.
chair : Mohamed El ghami
coauthors: J. B. M. Melissen, Kees
An extension of a predictor- Roos Terminal location integrated
corrector method variant from keywords: primal-dual interior point with ship routing
linear programming to semidef- alg, kernel function, proximity function,
inite programming complexity H ELENE G UNNARSSON
A NA PAULA A IRES B ORGES Department of Mathematics
In this paper we present a new class
T EIXEIRA of barrier functions for primal-dual in- coauthors: Mikael Rönnqvist, Dick
UTAD (Trás-os-Montes e Alto Douro terior point methods (IPMs). The pro- Carlsson
University)/Lisbon’s University posed functions have a finite value at the keywords: integer programming, ship
Operational Research Centre boundary of the feasible region. They routing, transportation, location problem
coauthor: Fernando Bastos are not exponentially convex and also not Södra Cell AB owns three pulp-mills.
keywords: semidefinite programming, strongly convex. In spite of this, the Ships, trains and trucks are used for sup-
interior point method, primal-dual iteration bound of large-update interior- plying customers. Coupled with each
algorithms point methods based on these functions
 11p n harbour there is a terminal. From each
We intend to present an extension of a is


shown to be O σn log ε 
 with p  terminal, the products are transported to
predictor-corrector algorithm from Lin- 0 1 and σ
1. For small-update inte- delivery points by truck, train or both
ear Programming to Semidefinite Pro-

rior point methods the iteration bound is combined. Recently, two new pulp-mills
gramming, where the predictor is cal- O n log nε  which is the currently best in Norway were acquired, which has led
culated as an iteration of a primal-dual known bound for primal-dual IPMs. In to a surplus of terminals. We have devel-
interior-point algorithm to Semidefinite this talk we restrict to linear optimiza- oped a mathematical model that selects
Programming and where the corrector tion, but extension of the methods to which terminals are to be used. A new
is calculated in the classical way. To other cone optimization problem can be feature combines a facility location as-
implement this predictor-corrector vari- derived in a natural way. pect with ship routing. The solution pro-
ant we modified the source code of the cedure includes route generation. Results
package CSDP, version 3.2, and we used based on evaluations and implementation
Lanczos’ method to find primal and dual MO3-341/22 LOGISTICS AND TRANSPORTATION

at Södra are presented.

steplengths. We will present some com- Facility location problems
putational experience. chair : Jakob Krarup
The p  q-active facility location
Primal-dual interior-point algo- Facility location under eco- problem
rithms for indefinite quadratic nomics of scale JAKOB K RARUP
programming A SGEIR TOMASGARD DIKU (Dept. of Computer Science,
A NDRE T ITS NTNU Univ. of Copenhagen)
University of Maryland coauthors: Leen Stougie, John van der coauthors: U. Leopold-Wildburger,
coauthor: Pierre-Antoine Absil Broek David Pisinger, J. Wanscher
keywords: quadratic programming, keywords: facility location, economics keywords: locational decisions, integer
nonlinear programming of scale, integer programming programming
 55

In almost all literature on locational de- relaxation for obtaining lower bounds. The first strategy allows resorting by
cisions where a prespecified number p of When branching, we implement con- color interchanges between identical let-
facilities are to be located, facilities are straint propagation techniques for fixing ters only. This results in a combinato-
either closed or open. Depending on the variables. We present preliminary results rial problem that is NP-complete even
cost structure, however, there may be op- for our exact algorithm. if the size of the underlying alphabet or
timal solutions where some of the open the number of colors is restricted to two.
facilities serve no users at all. To this Exact solution of the CVRP for It can be solved in polynomial time if
end a refinement of the notion “open” a given set of seed customers: a both parameters are bounded. The sec-
is called for. We say that an open fa- branch-and-bound approach ond strategy assumes the existence of a
cility is either active if at least one user set of parallel queues that can be used for
is allocated to it and passive otherwise. J ENS LYSGAARD resorting.
The p q-active facility location problem Aarhus School of Business
is the problem of locating p active facili- keywords: vehicle routing, Pursuit-evasion with imprecise
ties, each serving at least q users at a min- branch-and-bound, heuristics target location
imum total cost. For the resulting model The well-known Capacitated Vehicle G ÜNTER ROTE
we present a mixed 0-1 programming for- Routing Problem (CVRP) is the problem Freie Universität Berlin, Institut für
mulation, account for its solution prop- of routing a fleet of vehicles from a de- Informatik
erties and propose an exact branch-and- pot to a number of customers each requir-
bound algorithm with bounds generated keywords: search and surveillance,
ing delivery of a given quantity of a prod- on-line scheduling
via a dual ascent heuristic. uct. Some of the heuristic approaches to We consider a game between two persons
the CVRP are socalled two-phase heuris- where one person tries to chase the other,
MO3-341/23 LOGISTICS AND TRANSPORTATION
tics, among these are a number of cluster but the pursuer only knows an approxi-
Vehicle routing first – route second heuristics. The clus- mation of the true position of the flee-
chair : Jens Lysgaard tering phase in these heuristics are typ- ing person. The two players have iden-
ically based on an initial identification tical constraints on their speed. It turns
A two-commodity flow ap- of seed customers, such that the result- out that the fugitive can increase his dis-
proach to vehicle routing prob- ing routes each service exactly one seed tance from the pursuer beyond any limit.
lem customer. The identification of seed cus- However, when the speed constraints are
tomers as well as the subsequent cluster- given by a polyhedral metric, the pursuer
PAULA L ORENA Z ABALA ing and routing decisions are heuristic in can always remain within a constant dis-
Universidad de Buenos Aires nature in classical cluster first – route sec- tance of the other person.
coauthors: Abilio Lucena, Isabel ond heuristics. As such, there are poten- This can be applied to buffer mini-
Mendez-Diaz tially three sources of inoptimality of the mization in an on-line scheduling prob-
keywords: vehicle routing, resulting routes, namely i) the choice of lem with conflicts.
branch-and-cut, facets seed customers, ii) the clustering deci-
A two-commodity flow formulation for sions given the seed customers, and iii) A proximal trust-region algo-
the Vehicle Routing Problem (VRP) is the routing decisions given the clusters. rithm for stabilized column gen-
considered and a polyhedral study of the In this presentation we consider a solu- eration
associated polytope is carried out. Based tion approach to the symmetric CVRP JACQUES D ESROSIERS
on the results obteined from the polyhe- where only the choice of seed customers HEC Montreal and GERAD
dral study a branch and cut algorithm for is heuristic. In particular, we present a coauthor: Hatem ben Hamor
VRP is implemented and computationaly branch-and-bound procedure which, us- keywords: proximal methods, bundle
tested for instances from the literature. ing an additive bounding procedure, pro-
vides the exact solution to the symmetric methods, column generation, linear
CVRP for a given set of seed customers. programming
New valid inequalities and an This paper proposes a proximal point al-
exact algorithm for the ATSP- gorithm using both penalty and trust re-
TW MO3-321/053 PRODUCTION AND SCHEDULING

gion concepts. We establish finite con-

Sequencing and scheduling vergence while assuming the successive
V ICKY M AK chair : Jacques Desrosiers trust regions are of full dimension and
University of Melbourne never degenerate to a single point. The
coauthor: Andreas Ernst
Theoretical and practical as- approach is specialized to the column
keywords: cutting plane/facet, integer,
pects of color sequencing generation context. The resulting algo-
branch-and-bound, algorithms T HOMAS E PPING rithm converges to a pair of primal and
BTU Cottbus dual solutions. Computational experi-
In this talk, we present four new classes ments with large-scale multi-depot vehi-
of inequalities for the ATSP-TW. These coauthor: Winfried Hochstättler cle scheduling instances show substantial
new constraints are based on the Dk and keywords: flexible manufacturing, stabilizing and accelerating effects on the
Dk inequalities, and are strengthened due deterministic sequencing column generation approach.
to precedence restrictions. Driven by an application in the paint shop
We implement these inequalities on of an automobile plant, we consider the MO3-321/033 NETWORKS

a hybrid constraint programming and issue of resorting a sequence of colored Constrained paths
branch-and-bound algorithm for solving letters in order to minimize its number
the problems. At each node of the of color changes. We resume theoretical chair : Scott Provan
branch-and-bound tree, we implement an and practical results that arise from two On hop-constrained walk poly-
Assignment Problem-based Lagrangean different resorting strategies. topes
56 Parallel Sessions: Monday 16:30 – 18:00 (MO3)

N JÅL F OLDNES description of traffic flows. In this more fixed dimensions. Especially we discuss
University of Oslo general model, multiple Nash equilibria what kind of rooted circuits can be real-
coauthors: Geir Dahl, Luis Gouveia may exist and an arbitrary equilibrium ized in the affine space. We call a set X
does not need to be efficient. Neverthe- a stem if X r is a rooted circuit with a

keywords: hop-constrained walks,

complete linear description, projection less, our main result shows that the best root r, and the stems with a fixed root r
theorem equilibrium is as efficient as in the model a stem clutter. We start from a character-
without capacities. The improvements ization of stem clutters in 2-dimensional
We study hop-constrained walks from result from a simplified proof, which also case. We then show several generaliza-
a polyhedral point of view. A hop- allows us to work with broader classes of tions to higher dimensions. It includes a
constrained walk between two given latency functions. sufficient condition of the stem clutter to
nodes 0 and n is a 0-n-walk with at most be connected, a bound of the number of
H arcs. We briefly review the case H  3 connected components of the clutter, and
and consider here the case H  4. An A primal-dual algorithm for
finding maximum capacity the size of kernels needed in the realiza-
extended formulation for 4-walks is pre- tions.
sented. We then use the projection theo- paths with recourse
rem of Balas and Pulleyblank to derive a S COTT P ROVAN
complete linear description of the 4-walk The max-flow min-cut theorem
Univ. North Carolina for stems of rooted circuits of
polytope. The dominant of the 4-walk
polytope is also briefly discussed. Our coauthor: Rob Pratt affine convex geometries
main conclusions are: keywords: stochastic networks,
maximum capacity path, primal-dual M ASATAKA NAKAMURA
Describing the convex hull of hop- algorithms University of Tokyo
constrained paths, hop-constrained The Maximum Capacity Path with Re- coauthor: Masahiro Hachimori
walks and their dominant for H  4 course Problem involves finding a max- keywords: convex geometry
is much more complex than giving imum expected capacity path through a (antimatroid), closed-set system,
similar descriptions for H  3. stochastic network, in the case where an packing, max-flow min-cut property
arc capacity is known only when the tail A convex geometry is an abstraction of
As expected, describing the con-
node of that arc is reached. We study convex hulls in affine space. The comple-
vex hull of 4-walks is a lot easier
the problem under the assumption that ment of a convex geometry is known as
than describing the convex hull of
arc capacities reset randomly after each an antimatroid. We firstly introduce the
4-paths.
step. (All known no-reset versions of notion of closed-set systems, which leads
Quite surprisingly, for H = 4 it is this problem are NP-hard.) In this case to a common generalization of rooted cir-
easier to describe the convex hull a compactly-described strategy can be cuits both of matroids and convex geome-
of the 4-walks than their dominant. given for determining the optimal choice tries. In a matroid, (X,e) is a rooted cir-
of next path-arc at a node, based on cuit if X is a set not containing element
knowing the current arc values at that e and X e
is a circuit. We call X a

Selfish routing in capacitated node together with certain predetermined stem of e. A stem clutter is the collec-
networks values computed for the forward nodes of tion of stems of a fixed element. Seymour
these arcs. Finding these predetermined (1977) proved that a stem clutter of a bi-
N ICOLAS S TIER values involves solving an exponentially nary matroid has the max-flow min-cut
MIT large LP. We give a polynomial-time al- property if and only if it does not con-
coauthors: Jose Rafael Correa, Andreas gorithm to do this, based on a primal-dual tain a minor isomorphic to Q6 . We shall
S. Schulz solution technique, for implicitly solving present an analogue of this result in affine
keywords: price of anarchy, traffic this LP. convex geometries. Precisely, we shall
assignment, noncooperative game, show that a stem clutter of a convex ge-
multicommodity flow MO3-321/133 GRAPHS AND MATROIDS
ometry defined from 2-dimensional point
According to Wardrop’s first principle, Matroids and antimatroids configuration has the max-flow min-cut
agents in a congested network choose chair : Abdón Sanchez-Arroyo property if and only if the configuration
their routes selfishly, a behavior that is has no subset forming a ‘Pentagon’ with
captured by the Nash equilibrium of the Rooted circuits of convex ge- center e.
underlying noncooperative game. Be- ometries from affine point con-
cause Nash equilibria do not optimize figurations On independence systems and
any global criterion per se, there is no ap- M ASAHIRO H ACHIMORI Rado-Hall theorems
parent reason why they should be close Institute of Policy and Planning
to a solution of minimal total travel time, A BDÓN S ANCHEZ -A RROYO
Sciences, University of Tsukuba Banco de Mexico
i.e. the system optimum. In this talk,
we offer extensions of recent positive re- keywords: convex geometry, keywords: matroid, SDRs
sults on the efficiency of Nash equilib- antimatroid, rooted circuits
Rado-Hall Theorems concern systems of
ria. In contrast to previous work, we A convex geometry, whose complement
representatives of finite families of sub-
present results for networks with capac- is known as an antimatroid, is a set sys-
sets of a fixed finite set. Rado’s Theo-
ities and for latency functions that are tem arising from many important objects
rem states conditions under which it is
non-convex, non-differentiable and even such as posets graphs, point configura-
possible to find transversals of families
non-continuous. tions, oriented matroids, and so on. In
of independent subsets on a matroid. In
The inclusion of upper bounds on arc this work we focus convex geometries
this talk we present some partial results
flows has early been recognized as an im- arising from affine point configurations
for non-matroidal independence systems
portant means to provide a more accurate with additional fixed points (kernel) in
in which Rado’s Theorem holds.
 57

TU1-302/41 COMBINATORIAL OPTIMIZATION network flow problems. Here we pro- for improving the linear programming
pose an extension of most known results relaxations. Finally, we show the impact
Submodular functions and of this type to more general classes of of the presented methods on the solution
discrete convexity III linear and nonlinear problems. Further- of very large benchmark instances.
organizers : Satoru Fujishige and more, we discuss some economic appli-
Kazuo Murota cations of our results. A lower bound and test problem
chair : Satoru Fujishige generator for the group Steiner
A general two-sided matching Matroid representation of minimal tree problem
market with discrete concave clique complexes B OTING YANG
utility functions University of Regina
YOSHIO O KAMOTO
A KIHISA TAMURA Institute of Theoretical Computer coauthor: P. Gillard
RIMS, Kyoto University Science, ETH Zurich keywords: group Steiner minimal tree,
VLSI design, lower bound, test problem
coauthor: Satoru Fujishige coauthors: Kenji Kashiwabara, Takeaki generation
keywords: discrete convex analysis, Uno
The group Steiner minimal tree problem
stable marriage model, assignment keywords: independence systems,
is an extension of the standard Steiner
model matroid intersection
minimal tree problem in graphs, moti-
In the theory of two-sided matching In this paper, we approach the qual- vated by the problem of wire routing in
markets there are two standard models ity of a greedy algorithm for the maxi- the area of physical design of very large
called the stable marriage model, due to mum weighted clique problem from the scale integration (VLSI). This problem
Gale and Shapley, and the assignment viewpoint of matroid theory. More pre- is NP-hard, even if there are no Steiner
model, due to Shapley and Shubik. Re- cisely, we consider the clique complex nodes and the graph is a tree; moreover,
cently, Eriksson and Karlander have in- of a graph (the collection of all cliques there exists no polynomial time approxi-
troduced a hybrid model of these two of the graph) and investigate the mini- mation algorithm with a constant bound
and Sotomayor also considered the hy- mum number k such that the clique com- on the relative error under the hypothe-
brid model with full generality. In this plex of a given graph can be represented sis that P is not equal to NP. Hence, fast
talk, we propose a common generaliza- as the intersection of k matroids. This and good heuristic algorithms are needed
tion of these models by utilizing a frame- number k can be regarded as a measure in practice. In this talk, we describe an
work of discrete convex analysis intro- of “how complex a graph is with respect integer programming formulation of the
duced by Morota, and verify the exis- to the maximum weighted clique prob- problem. Using Lagrangean relaxation
tence of a stable solution in our general lem” since a greedy algorithm is a k- and subgradient optimization, we derive
model. approximation algorithm for this prob- a lower bound. In order to test the lower
lem. We characterize graphs whose bound, we present a procedure for gener-
clique complexes can be represented as ating test problems for the group Steiner
Substitutes and complements in the intersection of k matroids for any k minimal tree problem that have known
linear and nonlinear program- 0. Moreover, we determine the neces- optimal solutions. The computational ex-
ming sary and sufficient number of matroids periments for the test problems demon-
FABIO TARDELLA for the representation of all graphs with strate that the lower bound is very tight
Faculty of Economics - University of n vertices. This number turns out to be and differs from the optimal solutions by
Rome n 1. Other related investigations are only a few percent on average for sparse
also given. graphs.
keywords: complement, substitute,
submodular, parametric LP
TU1-302/42 COMBINATORIAL OPTIMIZATION
On the multi-weighted Steiner
Consider an optimization problem (P) tree problem: a reformulation
whose value depends on some parame- Steiner trees by intersection
ters in the objective function and /or in organizer/chair : Gunnar W. Klau
the constraints (e.g., the values of the J OÃO T ELHADA
right-hand sides in the constraints). Two
Practical algorithms for the CIO / DEIO - FCUL
parameters A and B are called comple-
Steiner problem coauthor: Luis Gouveia
ments if the sum of the improvements S IAVASH VAHDATI keywords: networks/graphs, Steiner
in the optimal value of (P) when each DANESHMAND trees, extended formulations
parameter is separately increased is not University of Mannheim Previously, Gouveia and Telhada (2003)
greater than the improvement obtained have presented an augmented arbores-
when they are jointly increased. Con- coauthor: Tobias Polzin cence formulation (AArb) for the multi-
versely, A and B are called substitutes keywords: Steiner problem, relaxation weighted Steiner tree problem. However,
if the reverse inequality holds. The State-of-the-art exact algorithms for this formulation is not symmetric, in the
concepts of complementarity and substi- N P -hard problems like the Steiner prob- sense that the linear programming (LP)
tutability can be equivalently expressed lem consist of many different compo- bound depends on the choice of the root
by saying that the optimal value func- nents, playing together as an "orchestra". node. In this talk we present a new for-
tion of (P) is supermodular or submod- We begin with giving an overview over mulation, which is compact although the
ular with respect to A and B. Starting our past and current work on such com- (AArb) model is not, with the property
with the work of Shapley, several authors ponents, namely relaxations and lower that its LP feasible set is obtained by
have established conditions guaranteeing bounds, reduction methods and upper intersecting the LP feasible sets of the
the complementarity or substitutability of bounds. Then we look more closely at (AArb) formulation for all possible root
some pairs of parameters, especially in one part of the work, namely methods choices. We prove that the LP bound of
58 Parallel Sessions: Tuesday 9:00 – 10:30 (TU1)

the new model is at least as good as the problem asks for a T -odd cut minimizing TU1-302/45 COMBINATORIAL OPTIMIZATION

LP bound obtained by the (AArb) formu- the weight c U : Ū  :  ∑e  U:Ū  ce . The
lation with the best root choice ( a few ex- Machine scheduling
problem occurs, e.g., in separation rou-
periments, indicate that in many cases, a tines for routing problems. chair : Jose Rafael Correa
strict improvement is obtained ). Compu- Two algorithms to compute a mini- A branch-and-cut approach for
tational results with instances involving mum T-odd cut are known (Padberg, Rao coupled task problems
up to 100 nodes demonstrate the compu- 1982; Rizzi 2003). Both require the solu-
tational advantage of the new modelling tion of up to  T  1 Max-Flow problems, M ARCUS O SWALD
Institute of Computer Science /
scheme.

leading to a  worst-case running time of
O  T  nm log n2 m   , if G has n nodes University of Heidelberg
TU1-302/44 COMBINATORIAL OPTIMIZATION
and m edges. We present an algorithm coauthors: Dino Ahr, Jozsef Bekesi
Combinatorial optimization based on the one by Rizzi and with the keywords: coupled task problem,
III same worst-case running time, but en- scheduling, branch-and-cut
hanced significantly. Following shrink- The coupled task problem (CTP) is to
chair : Klaus Michael Wenger
ing ideas known for the Min-Cut prob- schedule jobs on a single machine where
Fast algorithms for planar lem, we found ways to notably reduce each job consists of two subtasks and
graphs, Part I: cycles and cuts the average number of Max-Flow com- where the second subtask has to be
putations which are required. The con- started after a given time gap with re-
A DAM L ETCHFORD sequence is a clear decrease of running spect to the first subtask. The problem
Lancaster University times. We give computational results in- has several applications and is NP-hard.
coauthor: Nicholas Pearson dicating that our method is considerably Here we discuss an approach to solve the
keywords: analysis of algorithms, faster than both known algorithms. coupled task problem to optimality. We
combinatorial optimization, planar model the CTP as a linear 0 1-program
graphs, traveling salesman problem which is the basis for a branch-and-cut
Maximally violated mod-k cuts
We present improved algorithms for the and the capacitated vehicle algorithm. We describe the implemen-
following problems when the underlying routing problem tation of the algorithm and report about
graph is planar: computational results.
i) Minimum weight cut ii) Minimum K LAUS M ICHAEL W ENGER
weight circuit iii) minimum weight odd Institute of Computer Science, Mathematical models and meth-
cut iv) Minimum weight odd circuit. University of Heidelberg ods for balancing production
For the first and second problems lines with parallel blocks of op-
(which are equivalent under planar du- coauthor: Gerhard Reinelt erations at workstations
ality), the current best known algorithm keywords: CVRP, branch-and-cut,
is  due to Shih, Wu & Kuo and runs in cutting planes, mod-k G ENRIKH L EVIN
O n3 2 logn  time. We present a much In the CVRP the demands of a set of cus-
United Institute of Informatics problems
simpler algorithm with the same running of the National Academy of Sciences of
tomers have to be met at minimum total Belarus
time, and a more  complicated version travel cost using vehicles of identical ca-
which runs in O n3 2  time. pacity based at a single depot. We re- coauthors: Alexandre Dolgui, Nikolai
For the third and fourth problems port on the usefulness of maximally vi- Guschinsky
(which are again equivalent under planar olated mod-k cuts in branch-and-cut al- keywords: line balancing, machinery
duality), no specialised algorithm cur- gorithms for the CVRP. Let Ax  b be a We consider a balancing problem for
rently exists for the planar case. We system of valid inequalities with integer production lines when all operations of
present a simple algorithm which runs coefficients. Assume there is an integer the same workstation are partitioned into
in O n3 2 log n  time and a more compli- vector µ
0 such that the coefficients of several subsets (blocks), and all opera-

cated version which runs in O n3 2  time.  by the integer k
2 but
µT A are divisible tions of the same block are executed si-
These algorithms will be referred to µT b  k 1 mod  k  . Then the inequal- multaneously by one spindle head. The
in a companion talk by Nicholas Pear- ity 1k µT Ax  1k µT b  k 1   is valid. Its operation time of a block is usually
son, entitled "Fast algorithms for planar maximal violation is k 1  k which is more than maximal value among opera-
graphs, Part II: the traveling salesman achieved if Ai  x  bi for µi 0 where x tion times of its operations. All blocks
problem". is the fractional point to be separated. We of the same workstation are executed si-
take the binding constraints of an LP with multaneously as well. The problem is
solution x as input and follow the sep- to determine the quantity of the work-
Fast computation of a minimum stations and the blocks at every worksta-
t-cut in an undirected graph aration procedure suggested by Caprara,
Fischetti and Letchford (2000) with re- tion and to assign the operations to the
D IRK T HEIS striction to k prime. What can such blocks to minimize the total cost of all
University of Heidelberg general-purpose cuts of Chvátal-Gomory workstations and blocks. The following
type do for the CVRP? We answer this constraints have to be taken into account:
coauthor: Gerhard Reinelt (1) a partial order relation on the opera-
keywords: polynomial time algorithms, question by a computational study tak-
ing the problem-specific rounded capac- tion set; (2) the necessity or the impos-
graph algorithms sibility to assign some operations to the
ity, framed capacity, and multistar cuts as
Given an undirected graph G, positive foundation. We use an exact separation same block or workstation; (3) required
edge weights c and a set of “odd” nodes procedure for rounded capacity inequal- line cycle time; (4) limits on the quan-
T with  T  even, a T -odd cut is a cut ities. The conclusion is that maximally tity of the workstations and the blocks at
edge set U : Ū  such that  T
U  is an violated mod-k cuts are clearly too useful every workstation. Two approaches for
odd number. The minimum T-odd cut for the CVRP to be ignored. solving the problem are proposed. The
 59

approaches reduce the initial problem (1) We present mathematical models and Robust branch-and-cut-and-
to a mixed integer program problem and solution algorithms for staff scheduling price for the capacitated vehicle
(2) to a constrained shortest path prob- problems arising in real life applications. routing problem
lem. Both approaches are combined with In these problems, the daily assignments
heuristics techniques for solving large- to be performed are given and the dura- M ARCUS P OGGI DE A RAGAO
scale problems. Experimental results are tions (in days) of the working and rest pe-Pontificia Universidade Católica do Rio
discussed. riods for each employee in the planning de Janeiro
horizon are specified in advance, whereas coauthors: Eduardo Uchoa, Ricardo
the sequence in which these working Fukasawa, Marcelo Ladeira Reis
Single machine scheduling with and rest periods occur, as well as the keywords: column generation, routing,
precedence constraints daily assignment for each working pe- integer programming
J OSE R AFAEL C ORREA riod, have to be determined. The objec- We propose a branch-and-cut-and-price
MIT tive is to minimize the number of em- algorithm for the capacitated vehicle
ployees needed to perform all required routing problem. The column generation
coauthor: Andreas S. Schulz daily assignments.
keywords: scheduling theory, linear scheme follows the standard set partition-
We decompose the problem into two ing formulation(Balinsky and Quandt)
programming relaxations, approximation steps: the definition of the sequence of
algorithms where the node set is partitioned over the
working and rest periods (called pattern) routes (columns). We keep the edge set
We consider the classic problem of for each employee, and the definition of associated to each column to recover the
scheduling precedence-constrained jobs the daily assignment to be performed in original formulation where the variables
on a single machine to minimize the av- each working period by each employee. correspond to edges. Capacity cuts are
erage weighted completion time. This For step one, we present ILP models then separated on the original formula-
problem has attracted much attention by and exact enumerative algorithms; prac- tion, translated and added back to the col-
the mathematical programming commu- tical experience shows that the best ap- umn generation scheme. These cuts are
nity since Sidney’s pioneering work in proach is based on models involving a defined on the edge variables, the new
1975. large number of variables, thus requiring dual variables added can be distributed
In this talk, we discuss connections the adoption of column generation tech- over the edges in the cut. This implies
between various linear programming re- niques. Step two is stated as a feasibil- that no change is imposed to the column
laxations and some approximation algo- ity problem and is solved heuristically. generation subproblem. The same oc-
rithms. Among other things, we present We also propose an ILP formulation for curs when branching on edge variables.
a generalization of Sidney’s Decomposi- the entire problem, and present computa- This leads to a robust branch-and-cut-
tion Theorem and show that all known tional results on real life instances from and-price, where neither branching nor
2-approximation algorithms comply with an emergency call center. the addition of cuts change the struc-
this decomposition. Moreover, this the- ture of the pricing subproblem. Compu-
orem also yields a simple proof of the Solving duty scheduling prob- tational results on benchmark instances
polynomial-time solvability of the prob- lems in public transit from the literature are shown. Signifi-
lem when the precedence constraints are cant improvement was obtained on previ-
series-parallel. R ALF B ORNDÖRFER ously known lower bounds for instances
In addition, we establish a connec- Zuse Institute Berlin with a high number of vehicles. On
tion between the single-machine schedul- coauthor: Andreas Löbel all open instances tested, this algorithm
ing problem and the vertex cover prob- keywords: manpower planning, outperformed recently implemented pure
lem. Indeed, in the special case of series- combinatorics, column generation, branch-and-cut algorithms, even when
parallel precedence constraints, we prove public transportation these make use of several other families
that the sequencing problem can be seen of cuts besides the capacity cuts. We also
The duty scheduling problem in pub-
as a special case of vertex cover. More- closed a number of open instances.
lic transit is to string a set of duty el-
over, we argue that this result is true ements, i.e., units of driving work, into
for general precedence constraints if one a minimum cost set of shifts of work TU1-306/31
INTEGER AND MIXED INTEGER
PROGRAMMING
can prove that a certain integer program for bus, tram, or subway drivers. Such Integer and mixed integer
represents a valid formulation of the se- problems can be modelled as set parti- programming II
quencing problem. tioning problems and solved by column chair : Robert Nauss
generation techniques. To tackle large-
TU1-302/49 INTEGER AND MIXED INTEGER
PROGRAMMING
scale instances with several thousands of Simplifying branch-and-price
Integer programming column duty elements and complex labour reg- algorithms through lexico-
generation III ulations, special algorithmic techniques graphic ordering
are needed.
organizer : Marco Lübbecke The talk discusses LP solution, re-
E UGENE Z AK
chair : David M. Ryan TietoEnator Majiq
source constraint shortest path compu-
Models and algorithms for staff tation, and primal schedule construction keywords: cutting stock problem,
scheduling problems methods that aim at exploiting all degrees branch-and-price, column generation,
of freedom in the problem. The methods lexicographic ordering
M ICHELE M ONACI have been implemented in the optimiza- In the last decade, several exact algo-
DEIS, University of Bologna tion system DS-OPT. Computational re- rithms for solving the one-dimensional
sults for a number of real-world problems cutting stock problem have emerged.
coauthors: Alberto Caprara, Paolo Toth
will be reported. Most of these algorithms fall into the
keywords: combinatorial optimization,
column generation, branch-and-bound promising branch-and-price framework
60 Parallel Sessions: Tuesday 9:00 – 10:30 (TU1)

that uses Gilmore-Gomory column gen- solvers or even special purpose codes. rapid reliable optimization of
eration throughout a branch-and-bound For such problems excellent feasible so- systems described by paramet-
search tree. A new technique, proposed lutions with small relaxation gaps can ric partial differential equations
in this talk, has similarities with branch- generally be found, but the difficulty lies
and-price. The main difference, though, in proving optimality. An elastic ver- I VAN O LIVEIRA
comes from imposing a cutting pattern sion of the GAP where the resource con- Massachusetts Institute of Technology
order relation based on a lexicographic straints may be violated (at a cost) is ex- coauthor: A. T. Patera
comparison of vectors. This relation is amined. We explore this approach since keywords: reduced-basis, engineering
further drawn into a lexicographic or- a number of researchers have reported optimization, a posteriori error
dering over all feasible solutions of the on the success of solving elastic (versus estimation, partial differential equations
original problem. Due to this lexico- nonelastic) versions of various classes of
graphic ordering, branching rules, that integer programs. We perform compu- We present a technique for the rapid and
happen to be a tricky part in the ex- tational tests on corresponding cases of reliable optimization of systems charac-
isting branch-and-price algorithms, be- classical and elastic GAP problems to de- terized by linear-functional outputs of
come intuitive here. Evidently, the pric- termine if the elastic version is easier to partial differential equations with affine
ing problem turns into a knapsack prob- solve. (input) parameter dependence. The
lem with an additional constraint that lex- critical ingredients are: (Lagrangian)
icographically restricts the knapsack so- reduced-basis approximation to effect
TU1-306/32 NONLINEAR PROGRAMMING

significant reduction in state-space di-

lutions. A computer implementation of Optimization of engineering
the algorithm has been devised and the mensionality; a posteriori error bounds
systems governed by to provide rigorous error estimation and
computational results prove the competi- simulations/PDE, part 1
tiveness of the proposed technique. control; “offline/online” computational
organizer/chair : Natalia Alexandrov decompositions to permit rapid evalua-
tion of output bounds, output bound gra-
Eliminating redundant solu- Optimization with multiple dients, and output bound Hessians in
tions of combinatorial integer scales using wavelet based re- the limit of many queries; and refor-
programs duced order models mulation of the approximate optimiza-
PABLO A. R EY M ARTIN B ENDSOE tion statement to ensure (true) feasibil-
Universidad de Chile Dept. of Mathematics, Technical ity and control of suboptimality. To en-
University of Denmark sure reliable model behavior, we have
coauthor: Oscar Porto
developed a directionally-scaled trust re-
keywords: integer programming, coauthors: A. R. Díaz, S. Chellappa gion algorithm which we employ to solve
redundant solutions keywords: multiple scales, MPEC, the resulting mathematical program. To
The presence of equivalent solutions wavelets illustrate the method we consider the
in linear integer programming problems In this work we address the possibil- design of a three-dimensional thermal
make their solution difficult for enumer- ity of designing structures with multi- fin: Given volume and power objective-
ative algorithms. This work presents a ple scales using concepts from topology function weights, and root temperature
way to exploit known symmetries of lin- design coupled with multi-scale resolu- “not-to-exceed” limits, the optimal ge-
ear integer models to eliminate redundant tion methods that can account for the ometry and heat transfer coefficient can
equivalent solutions. This information is presence of finite scales. The approach be determined — with guaranteed feasi-
used to generate appropriate inequalities taken is to use a wavelet-Galerkin tech- bility — in only 2.3 seconds on a 500
which added to the original formulation nique and an associated multi-resolution MHz Pentium machine. Our method thus
enhance their solvability. In the exam- numerical homogenization meyhod to re- permits not only interactive optimal de-
ples considered, this approach leads to an place fine-scale operators with coarse- sign at conception and manufacturing,
impressive reduction of the search tree of scale averages, that is, reduced order but also real-time (reliable) adaptive op-
enumerative algorithms. models of the original system in a coarser timal design in operation.
scale. In general, these homogenized op-
A computational study of elastic erators are dense and cannot be used eas- Robust optimal design of exper-
and classical generalized assign- ily in optimization. They are used in- iments for DAE and PDE
ment problems stead in an inverse procedure that ob-
tains an equivalent coarse-scale elasticity
ROBERT NAUSS problem. Thus the complete process as-
E KATERINA KOSTINA
University of Missouri-St. Louis IWR, University of Heidelberg
sociates an equivalent coarse-scale mate-
keywords: optimization, integer rial layout with each element in a library keywords: design of experiments,
programming, mixed integer of fine-scale operators. The topology numerical methods
programming, assignment problems optimization problem uses this coarse Estimating model parameters from ex-
The classical generalized assignment material layout to parameterize the de- perimental data is crucial to reliably sim-
problem (GAP) is NP-complete. While sign space, and the optimization pro- ulate dynamic processes. In practical ap-
certain random problem generators yield cess from there on resembles a standard plications, however, it often appears that
GAP problems that are relatively easy to sizing problem – that is, mathematical the experiments performed to obtain nec-
solve, there are at least two generators programming problems with equilibrium essary measurements are expensive, but
that produce problems that are very dif- constraints (MPEGs). This is joint work nevertheless do not guarantee sufficient
ficult to solve to optimality. Indeed prob- with A. R. Díaz and S. Chellappa, Michi- identifiability. The optimization of one
lems on the order of only 1000 binary gan State University, USA. or more dynamic experiments in order
variables may not admit a provably op- to maximize the accuracy of the results
timal solution using general-purpose IP Reduced-basis techniques for of a parameter estimation subject to cost
 61

and other technical inequality constraints a sequence of primal quadratic subprob- TU1-306/34 NONLINEAR PROGRAMMING

leads to very complex non-standard opti-  bound constraints. These large

lems with
Advances in algorithms and
mal control problems. size 20000 unknowns) ill-conditioned
software for nonlinear
Newly developed successful meth- ( 107 ) subproblems are solved combin- programming
ods and software for design of optimal ing efficiently an active set method with
experiments for nonlinear processes are a gradient projection method and a pre- organizer/chair : Stephen Wright
based on the expansion of the problem at conditioned conjugate gradient method.
the nominal value of parameters which
LOQO: an interior-point
We use the standard multiplier approach method for nonconvex nonlin-
lie in a (possibly large) confidence re- to maximize the augmented Lagrangian
gion. Better experiments should be ob-
ear programming
dual function.
tained if we optimize the experiments The strategic choices of the algorithm H ANDE B ENSON
in minimax fashion (worst-case design) will be discussed, theoretically motivated United States Naval Academy,
over the whole range (confidence region) and illustrated by extensive numerical ex- Mathematics Dept.
of an uncertainty set. Such a mini- periments on problems arising in 3D re-
max problem, as a semi-infinite program- coauthors: David Shanno, Robert J.
flection tomography. Vanderbei, Igor Griva, Arun Sen
ming problem, has extremely high degree
of computational complexity. We have keywords: optimization, nonlinear
developed efficient and reliable meth- Finite time identification of ac- optimization, interior point method
ods for design of robust optimal experi- tive constraints with an aug- In this talk, we will present the lat-
ments that reduce the complexity of the mented Lagrangian algorithm est developments to LOQO, an interior-
semi-infinite programming problem. The J EAN C HARLES G ILBERT point method for nonconvex nonlinear
methods are successfully applied to real- INRIA Rocquencourt programming. Extensive numerical re-
life problems from chemistry and chem- sults will be provided.
ical engineering (joint work with Georg coauthor: Frederic Delbos
Bock, Stefan Koerkel and Johannes P. keywords: augmented Lagrangian,
Schloeder). active constraints, proximal algorithm, Improving to nonlinear interior
strongly convex optimization methods
TU1-306/33 NONLINEAR PROGRAMMING We consider an augmented Lagrangian J ORGE N OCEDAL
Augmented Lagrangian algorithm for minimizing a strongly con- Northwestern University
methods I vex function subject to linear inequality coauthors: Richard Waltz, Richard
constraints. It is shown that the active Byrd
chair : Otero Rolando Gárciga constraints at the solution are identified keywords: large scale optimization
Application of an SQP aug- after a finite number of iterations. This
result is a consequence of the proximal We begin by discussing strategies for
mented Lagrangian method to a adaptively choosing the barrier param-
large-scale problem in 3d reflec- interpretation of the method. Algorith-
mic implications are discussed, in par- eter in nonlinear interior methods. A
tion tomography framework for guaranteeing global con-
ticular when the augmented Lagrangian
F REDERIC D ELBOS method is used to solve the quadractic vergence is described, and numerical re-
Institut Français du petrole programs arising in the SQP approach. sults are presented. We also make some
remarks on how to handle cases when the
coauthors: Jean Charles Gilbert, central path does not exist in a neigh-
Delphine Sinoquet A strongly convergent aug- borhood of an iterate. We end the talk
keywords: SQP, Gauss-Newton mented Lagrangian method by evaluating the performance of a merit
algorithm, augmented Lagrangian, function approach, and in particular, by
seismic tomography OTERO ROLANDO G ÁRCIGA questioning whether filter mechanisms
Seismic reflection tomography allows UFRJ/IE are, in fact, more effective.
the determination of a subsurface model keywords: augmented Lagrangians,
from the traveltimes of seismic waves re- cone-constrained optimization, strong
flecting on geological interfaces. From convergence, inexact solutions An LCL algorithm for con-
an optimization viewpoint, the problem
strained optimization
We consider a general optimization prob-
consists in minimizing a nonlinear least- lem  of the form: min g(x) subject to M ICHAEL PAUL F RIEDLANDER
squares functional measuring the mis- G x   K, with g : B1

 , G:B
Mathematics and Computer Science Div,
1
match between observed traveltimes and B2 , where B1 and B2 are real reflex- Argonne Natl Labs
those calculated by raytracing in this ive Banach spaces and K is a nonempty
model. We introduce geological a priori coauthor: Michael Alan Saunders
closed convex cone in B2 . We present keywords: large scale optimization,
information thanks constraints in the op- an inexact hybrid augmented Lagrangian
timization problem. nonlinear optimization, augmented
method for solving this problem. The Lagrangian, elastic variables
We use a Gauss-Newton SQP method method generates a primal-dual sequence
globalized by line search to solve this which converges strongly to an optimal For optimization problems with non-
optimization problem. At each Gauss- pair. Moreover, we can establish the rela- linear constraints, linearly constrained
Newton step, a solution to a strictly tion between the strong limit and the ini- Lagrangian (LCL) methods sequentially
convex quadratic model of the objective tial iterate. Results are guaranteed under minimize an augmented Lagrangian sub-
function subject to linearized constraints G-differentiability and convexity of the ject to linearized constraints. Conver-
is computed using a dual approach. data functions (K-convexity for G) and gence is rapid near a solution (as proved
This one is based on an augmented La- existence of KKT-pairs provided B2 uni- by Robinson and often observed with MI-
grangian relaxation that requires solving formly convex and uniformly smooth. NOS).
62 Parallel Sessions: Tuesday 9:00 – 10:30 (TU1)

To induce global convergence algorithm is based on general New- the minimization of a separa-
and to unify the approaches used in ton method which uses inner iterative ble convex function on a con-
LANCELOT and MINOS, we work with method for calculating approximation of vex polyhedron generated by a
an elastic form of the linearized con- the inverse Jacobian matrix. Local q- parallel-consecutive structure
straints (by adding an L1 penalty term to superlinear and q-quadratic convergence
the augmented Lagrangian). Our stabi- is proved. Numerical experiments and Y URI L EVIN
lized LCL method possesses two impor- computational efficiency are also pre- Queen’s University School of Business
tant properties: the linearly constrained sented. coauthors: Adi Ben-Israel, Genrikh
subproblems are always feasible, and Levin, Boris Rozin
they may be solved inexactly. keywords: convex programming,
Object-oriented parallel inte- decomposition, large scale optimization
The current implementation is effi- rior point solver for structured
cient on large problems, using MINOS to nonlinear programs We consider a problem of minimization
solve the subproblems. Only first deriva- of a separable, strictly convex, monotone
tives are required. We present numerical A NDREAS G ROTHEY and differentiable function on a convex
results on the COPS and CUTE test prob- University of Edinburgh polyhedron generated by a system of m
lems. linear inequalities. The variables are di-
coauthor: Jacek Gondzio
keywords: parallel programming, vided into n disjoint subsets, and vari-
TU1-306/35 OTHER
interior point method, nonlinear ables from each inequality are elements
Software session, Dash programming, large scale of a direct product of these subsets.
An approximate solution of this prob-
chair : James Tebboth Very large optimization problems with lem can be obtained by solving at most
millions of constraints and decision vari- m subproblems; each subproblem has ex-
TU1-306/36 PARALLEL COMPUTING ables display usually some special struc- actly one constraint and at most n vari-
ture. The key to efficient solution of ables. The authors’ Projected Newton
Implementation such problems is the ability to exploit the Bracketing (PNB) method for linearly
chair : Andreas Grothey structure. We have developed a structure- constrained convex minimization is used
Parallel implementation of exploiting parallel primal-dual interior- to solve these subroblems. We review the
point solver for nonlinear programming PNB method during the talk.
metaheuristics for combinato- problems. The solver allows in particu-
rial optimization problems on lar For the exact solution of this problem
a nested embedding of structures and we propose the decomposition scheme
global computing systems by this means very complicated real-life based on the constraint relaxation idea.
K AZUHIRO KOBAYASHI optimization problems can be modeled. This decomposition scheme calls for so-
Department of Mathematical and Its design uses object-oriented pro- lution of a dynamically generated se-
Computing Sciences/Tokyo Institute of gramming techniques. Different matrix quence of subproblems with equality
Technology structures are implemented as subclasses constraints. A subroblem in the sequence
of an abstract matrix class. This abstract is obtained from the previous one by
coauthor: Katsuki Fujisawa matrix class contains a set of virtual func- adding and/or deleting one of the con-
keywords: parallel computing, tions (methods in the object-oriented ter- straints. A modification of the PNB
metaheuristics minology) that: (i) provide all the nec- method efficiently solves these subprob-
We describe parallel implementation of essary linear algebraic operations for an lems.
metaheuristics for combinatorial opti- IPM, and (ii) allow self-referencing. The The described problem arises in op-
mization problems. We use global com- program OOPS (Object-Oriented Parallel timization of parameters of mechani-
puting system called Ninf. Ninf system Solver: cal systems with a parallel-consecutive
enables us to utilize the computational re- structure subject to reliability of their
sources connected via local area network http://www.maths.ed.ac.uk/ components.
or wide area network. Ninf system pro- ˜lkgondzio/parallel/solver.html)
vides an easy-to-use programming inter-
face and enables us to build Grid-enabled can efficiently handle very large nonlin-
Solving system design using a
ear problems and achieves scalability on type of convex submodels: com-
applications. We present some computa-
a number of different computing plat- putational experience
tional experiments on PC Cluster.
forms.
The efficiency of the solver is il- Á NGEL M ARÍN
Newton-like method for special lustrated with problems known from Universidad Politécnica de Madrid
class of nonlinear systems the literature: applications arising from coauthor: Javier Salmerón
telecommunication and financial engi- keywords: system design, location,
Z ORANA L UZANIN neering. Numerical results are given for convex programming, decomposition
Institute of Mathematics and the solution of nonlinear financial plan- In this presentation we present an ex-
Informatics, University of Novi Sad, ning problems with sizes over 50 million plicit method for solving a convex model
Serbia decision variables. which involves a convex objective func-
coauthor: Natasa Krejic tion and cumulative bound constraints.
keywords: Newton-like methods, block TU1-306/37 CONVEX PROGRAMMING
These models may arise, for example,
diagonal structure, computational
efficiency, q-convergence
Decomposition in convex when Benders’ Decomposition or La-
programming grangean Relaxation-Decomposition is
We present an efficient algorithm for chair : Kaj Holmberg applied to solve large design problems
solving system of nonlinear equations such as the facility location and capac-
with a block diagonal structure. The A decomposition approach to ity expansion problems. To attain the
 63

optimal solution of the model we ana- that the performance of IHR is for many We address the challenging informat-
lyze its Karush-Kuhn-Tucker optimality problem classes comparable to or better ics problems of: (a) using group contri-
conditions and develop an explicit for- than that of existing algorithms, and that bution data to design chemicals with de-
mulation for the optimal primal and dual it is more universal in the sense that it sired properties, (b) using X-ray diffrac-
solutions. This approach yields higher can treat many types of objective func- tion data to infer three-dimensional struc-
performance when compared to other tions. We discuss benefits and drawbacks tures of crystals, and (c) using DNA se-
convex optimization techniques, MINOS of both deterministic and stochastic solu- quencing data to predict DNA-protein
and parametric optimization. tion approaches to fractional programs. binding sites. For each problem, we
present novel mathematical program-
Mean value cross decomposition Minimal ellipsoid circumscrib- ming models and algorithms for their so-
for nonlinear convex problems lution.
ing a polytope defined by a sys- The main conclusion of the paper
K AJ H OLMBERG tem of linear inequalities is that branch-and-reduce algorithms for
Linkoping University, Department of the global optimization of the underlying
Mathematics J UN - YA G OTOH nonlinear and mixed-integer nonlinear
coauthor: Krzysztof Kiwiel University of Tsukuba programming models have now reached
keywords: nonlinear programming, coauthor: Hiroshi Konno the level of maturity required to solve re-
decomposition methods keywords: minimal ellipsoid, alistic informatics problems.
Mean value cross decomposition is a computational geometry, global
symmetric primal-dual decomposition optimization
TU1-308/11 AIRLINE OPTIMIZATION APPLICATIONS

method, suitable for optimization prob- Calculation of a minimal sphere circum- Disruption management
lems with both primal and dual struc- scribing a polytope has been studied by
tures. It uses only easily solvable sub- organizer/chair : Niklas Kohl
many authors. In particular, if we know
problems and no difficult master prob- all extreme points of a polytope and if its Degradable airline schedules
lems. Originally developed for linear cardinality is not excessively large, then
problems, it is in this paper extended to the problem can be solved efficiently by a
nonlinear convex optimization problems.
J OHN -PAUL C LARKE
number of algorithms. On the other hand, MIT
Convergence is proved for a somewhat if a polytope is defined by a linear system
generalized version, allowing more gen- coauthor: Laura Kang
of inequalities, the problem is NP com-
eral weights. Computational results are keywords: robust scheduling, airline
plete. We are concerned with a problem
presented for a network routing problem scheduling, airline operations
for calculating a minimal ellipsoid cir-
with congestion costs, a large-scale non- cumscribing a polytope defined by a sys-
linear problem with structures that enable We present a methodology for deriving
tem of linear inequalities, which is a di- airline schedules that are robust to disrup-
decomposition both with respect to vari- rect extension of a minimal sphere prob-
ables and constraints. The main goals tions caused by bad weather, and that al-
lem. In this talk we present a global opti- low airlines to market itineraries and sell
of the tests are to illustrate the proce- mization algorithm for the minimal ellip-
dure and to indicate that this decomposi- tickets based on passenger preference for
soid problem where a polytope is of rel- reliability. In this methodology, an ex-
tion approach is more efficient than direct atively low dimension. We also present
solution with a well established general isting schedule is partitioned into inde-
some numerical results to show its per- pendent sub-schedules or layers – prior-
code. formance. itized on the basis of revenue – where the
highest priority or "protected" layers are
TU1-306/38 GLOBAL OPTIMIZATION

designed to both contain the highest rev-

Global optimization I Global optimization in infor- enue itineraries and to be operable in the
chair : Nikolaos Sahinidis
matics problems in biology, worse weather. The problem is modeled
chemistry, and physics as an integer program where the candi-
Solving fractional programs date aircraft routings are created using a
with stochastic algorithms N IKOLAOS S AHINIDIS greedy heuristic and a tabu search. Re-
University of Illinois sults indicate that no additional aircraft
M IRJAM D UER
Department of Mathematics, Darmstadt keywords: nonconvex optimization, are required to increase the number of
University of Technology mixed integer nonlinear prog., itineraries in the protected layers from
coauthor: Zelda B. Zabinsky bioinformatics, chemical informatics 891 to 938 and to increase the corre-
keywords: fractional programming, sponding revenue in the protected layers
With the recent accumulation of vast from 60% to 76.3%.
nonlinear programming amounts of chemical, biological, and
Maximization of a sum ratios is an opti- clinical data, many scientific fields are
mization problem which appears in many becoming increasingly data-driven as op- A branch-and-price approach
economic and engineering situations, posed to model-driven. This paradigm for operational aircraft mainte-
whenever efficiency of a system is to be shift has brought about many challenging nance routing
optimized. Traditional methods to solve computational problems. Even though
these problems were mainly Branch-and- these problems originate from very dis- R AJAN BATTA
Bound type algorithms. In this talk, a parate fields, they have very similar University at Buffalo (SUNY)
new approach is discussed: We report mathematical structures. In particular, coauthors: Abdulkadir Sarac,
on experience treating fractional prob- they involve the use of a merit function Christopher M. Rump
lems with stochastic algorithms like Im- to evaluate alternatives from very large, keywords: airline applications,
proving Hit-and-Run (IHR). It turns out often combinatorial, search spaces. maintenance routing
64 Parallel Sessions: Tuesday 9:00 – 10:30 (TU1)

We focus on the problem of aircraft rout- The unit commitment problem is an im- keywords: stochastic integer
ing, which involves generating and se- portant problem for electric power util- programming, integer duality, global
lecting a particular route for each aircraft ities. The problem is to determine the optimization
of a sub-fleet that is already assigned to schedule of power generating units and
a set of feasible sequences of flight legs. the generating level of each unit. The We consider a two-stage pure integer pro-
Similar studies typically focus on long- decisions are which units to commit at gram with stochastic right-hand sides.
term route planning. However, stochastic each time period and at what level to gen- We give a dual reformulation with which
events such as severe weather changes, erate power meeting the electricity de- the value functions of two stages can be
equipment failures, variable maintenance mand. The objective is to minimize the evaluated separately. Two cases are ad-
times, or even new regulations mandated operational cost which is given by the dressed with respect to the storage of
by the Federal Aviation Administration sum of the fuel cost and the start up cost. value functions. When the value func-
(FAA) play havoc on these long-term The electric power industry is undergoing tion can be explicitly stored, we use a
plans. As a result, these plans are of- restructuring and deregulation. In this pa- superadditive dual algorithm that is of-
ten ignored by personnel in airline op- per we develop a stochastic programming ten able to obtain the exact value func-
erations who are forced on a daily ba- model which incorporates power trading. tions with millions of right-hand sides by
sis to develop quick, ad hoc methods to It is assumed that demand and price un- only solving very few deterministic inte-
address these maintenance requirements certainty can be represented by a scenario ger programs. Once the value functions
and other irregular events. To address tree. We propose a stochastic integer pro- of both stages are found, the objective
this problem, we develop an operational gramming model in which on/off deci- function evaluation becomes straightfor-
aircraft maintenance routing problem for- sions of generators are made in the first ward. When the value function cannot
mulation that includes maintenance re- stage. The solution approach is based be explicitly stored, in addition to the
source availability constraints. We pro- on Lagrangian relaxation method and dy- superadditive dual approach, we present
pose a branch-and-price algorithm for namic programming. a global branch-and-bound framework to
solving this problem, which, due to the bound the objective functions on the sub-
resource constraints, entails a modifica- problems. We develop several techniques
tion of the branch-on, follow-on branch-
SPAR: stochastic programming to speed up the procedure of finding the
ing rule typically used for solving similar
with adversarial research value functions. A level set approach
problems. Through computational test- is also considered to reduce the number
A NDREW S CHAEFER of objective function evaluations needed.
ing, we explore the efficiency of this so- University of Pittsburgh
lution approach under a combination of We provide computational results on in-
heuristic choices for column (route) gen- coauthors: Matthew Bailey, Steven stances that are several orders of magni-
eration and selection. Shechter tude larger than problems found in the lit-
keywords: stochastic programming, erature.
dynamic programming, Markov decision
Fast aircraft schedule recovery processes, stochastic games
with crew and passenger con- TU1-308/13 STOCHASTIC PROGRAMMING

straints We introduce a general modeling tech- Sampling and control in

nique for problems involving the design stochastic programming
DARREN D EJUN H ANG of a system with the knowledge that an
Carmen Systems AB chair : Andreas Eichhorn
adversary may attempt to destroy or sig-
keywords: aircraft schedule recovery, nificantly decrease the performance of Integration quadratures in dis-
airline operations the system after installation. We consider cretization of stochastic pro-
The aircraft schedule recovery problem multi-stage problems in which an adver- grams
is solved daily by airlines. The objec- sary decides future stages. The decision
tive is to minimize the impact of disrup- maker must choose a system configura- M ATTI KOIVU
tions on the airline schedule taking into tion so as to minimize the long-run dam- Helsinki School of Economics
account essential aircraft, crew and pas- age inflicted by the adversary. The ad- coauthor: Teemu Pennanen
senger data and constraints. Our algo- versary then solves a multi-stage stochas- keywords: discretization, integration
rithm generates structurely different so- tic optimization problem to maximize the quadratures
lutions within seconds. Experimental re- expected damage to the decision maker.
sults on real life data from major airlines We formulate this problem as a two-stage Modern integration quadratures are de-
will be presented. stochastic program with Markov decision signed to produce finitely supported ap-
process recourse. This method gener- proximations of a given (probability)
TU1-308/12 STOCHASTIC PROGRAMMING alizes other approaches, including many measure. This makes them well suited
stochastic interdiction problems. We dis- for discretization of stochastic programs.
Computational stochastic cuss health care and military applica- We give conditions that guarantee the
integer programming tions, and preliminary computational re- epi-convergence of resulting objectives
organizer/chair : Andrew Schaefer sults. to the original one. Our epi-convergence
Lagrangian relaxation method result is closely related to existing ones
for price-based unit commit- Two-stage integer programming but it is easier to apply to discretizations.
ment problem As examples, we will verify the condi-
with stochastic integer right- tions in three different models of portfo-
TAKAYUKI S HIINA hand sides lio management and we study the behav-
Central Research Institute of Electric NAN KONG ior of various discretizations numerically.
Power Industry University of Pittsburgh In our tests, modern quadratures clearly
keywords: electric power, unit outperform crude Monte Carlo sampling
coauthor: Andrew Schaefer in discretization of stochastic programs.
commitment
 65

coauthor: Alexander Slastnikov spect to second degree stochastic domi-

Nondeterministic dynamic pro- keywords: investment under nance, while at the same time preserving
gramming uncertainty, optimal stopping problem, simplicity and LP computability of the
taxation original models. The models are tested
S EIICHI I WAMOTO We propose a model that can evaluate on real-life historical data.
Kyushu University the influence of tax system and stochas-
coauthor: Takayuki Ueno tic environment on investor bahavior in How serious are Wets’ et al “se-
keywords: non deterministic, stochastic, real sector of economics. As an object of rious zeros curves”?
deterministic investment it will be considered a project
We consider dynamic programming from of a creation of a new enterprise. Invest- K EN KORTANEK
a view point of transition law. A non- ments, necessary for such a project are University of Iowa
deterministic dynamic programming is considered to be instantaneous and irre-
keywords: discount function, forward
proposed in a finite (stage, state and deci- versible. At any moment the investor can
rate function, spot rate function
sion) framework which includes both de- either accept the project and start with the
terministic and stochastic dynamic pro- investment, or delay the decision before The recent manuscript by Wets et al, "
gramming. The non-deterministic dy- obtaining new information on the envi- Serious Zero Curves" [2] of "EpiSolu-
namic programming represents a large ronment. The purpose of the investor is tions Inc" offers an interesting contrast to
class of multi-stage decision problems to find a moment for investment, which the Kortanek/Medvedev approach devel-
where today’s state may go to more than depends on previous environment obser- oped in [1]. Both methods rely on contin-
two tomorrow’s states with one-step tran- vations, so that it maximizes expected net uous compounding relationships between
sition weights. While the stochastic present value. An investor problem is (a) the forward rate FR, (b) the discount
transition probability has both unit-sum reduced to an optimal stopping problem DF, and (c ) the spot rate SR functions:
property and non-negativity, the non- for two-dimensional geometric Brownian
motion and homogeneous terminal func-
deterministic one-step transition weight
tional. We propose the new (variational)  
t 
does not have unit-sum property and has DF t   exp FR s  ds 
non-negativity. An objective function in approach for solving this problem. It is 0
non-deterministic dynamic programming based on the connection between bound-
  
is a weighted value of additive reward ary problem for diffusion processes and SR t   ln DF t   t
function. It is a generalization of ex- Dirichlet problem for PDE of an elliptic
type, especially Feynman-Kac formula and
pected value in stochastic dynamic pro-

 
and variational inequalities.
gramming. Solving a recursive equation, FR t   DF t  DF t 
we can find an optimal policy in non-
deterministic dynamic programming. On extending the LP com-
putable risk measures to ac- Both methods incorporate standard
count downside risk nonarbitrage constraints and smoothness
Stochastic integer program- proxies.
ming: limit theorems and confi- W LODZIMIERZ O GRYCZAK The former approach is based upon
dence intervals Warsaw University of Technology, a finite Taylor series approximation to
Institute of Control & CE the DF function, leading to an LP model
A NDREAS E ICHHORN which seeks an optimal piecewise con-
Humboldt-University Berlin coauthor: Adam Krzemienowski
keywords: portfolio optimization, linear stant third derivative of DF. The latter
coauthor: Werner Roemisch programming, mean-risk, downside risk approach is based upon a dynamic model
keywords: stochastic programming, under uncertainty with perturbations for
mixed integer, empirical process, A mathematical model of portfolio opti- the FR function leading to a geometric
bootstrap mization is usually quantified with mean- programming model which seeks optimal
risk models offering a lucid form of two piecewise constant perturbations.
We consider empirical approximations of criteria with possible trade-off analysis.
two-stage stochastic mixed-integer pro- In the classical Markowitz model the risk After the models are compared,
grams and derive central limit theo- is measured by a variance, thus result- preliminary numerical comparisons are
rems for the objectives and optimal val- ing in a quadratic programming model. given using Wall Street Journal data for
ues. The limit theorems are based on There were introduced several alterna- about 133 Treasury Bills, Notes, and
empirical process theory and the func- tive risk measures which are computa- Bonds.
tional delta method. We also show how tionally attractive as (for discrete ran-
such limit theorems for the bootstrapping dom variables) they result in solving lin- [1] K. O. Kortanek and V. G.
method can be used to derive confidence ear programming (LP) problems. Typi- Medvedev. Building and Using
intervals for optimal values. Dynamic Interest Rate Models.
cal LP computable risk measures, like the
Wiley Finance. John Wiley &
mean absolute deviation (MAD) or the
Sons Ltd, Baffins Lane, Chich-
TU1-308/T1 FINANCE AND ECONOMICS Gini’s mean absolute difference (GMD)
ester, West Sussex PO19 1UD,
are symmetric with respect to the below-
Modeling under uncertainty mean and over-mean performances. The
England, 2001.
chair : Ken Kortanek paper shows how the measures can be [2] R. J. B. Wets, S. W. Bianchi,
Stochastic model of optimal in- further combined to extend their model- and L. Yang. Serious zero
vestor behavior in real sector ing capabilities with respect to enhance- curves. Technical report, EpiSo-
ment of the below-mean downside risk lutions, Inc, El Cerrito, California,
VADIM A RKIN aversion. The resulting mean-risk mod- 2002.
Russian Academy of Sciences els generate efficient solutions with re-
66 Parallel Sessions: Tuesday 9:00 – 10:30 (TU1)

TU1-308/T2 COMPLEMENTARITY AND VARIATIONAL

INEQUALITIES
enables practitioners and developers to
Interior point methods for vari- remotely access (expensive) software
Regularization methods tools such as solvers and modelling sys-
ational inequalities with weak
organizer/chair : Rainer Tichatschke regularisation tems as well as complete optimisation-
based vertical applications. In this paper
Convergence and stability of a T IM VOETMANN we tackle the issue of integrating these
regularization method for max- University of Trier remote tools into distributed applications.
imal monotone inclusions For this purpose we consider a web ser-
coauthor: Rainer Tichatschke
DAN B UTNARIU keywords: variational inequalities, vice model based on XML representation
University of Haifa, Israel interior point method, proximal point of the model instances and discuss the re-
methods, Bregman distances quirement for a common solver interface
coauthors: Yakov Alber, Gabor Kassay based on the SOAP protocol.
keywords: maximal monotone operator, The talk is devoted to the numeri-
regularization method, convex cal solution of variational inequalities
optimization problem, Mosco with set-valued monotone operators in a Analyzing submissions to the
convergence of sets Hilbert space. In contrast to the classi- NEOS server for the purpose of
We consider the inclusion f  Ax where
cal proximal point method that employs recommending solvers
the Hilbert space norm as a regularis-
A is a maximal monotone point-to-set op-
erator from the reflexive Banach space
ing functional we investigate the use of ROBERT F OURER
weaker norms. This leads to a signifi- Northwestern University
with the Kadec-Klee property X to its
cant acceleration of the convergence pro- coauthor: Dominique Orban
dual. We assume that the data A and f are
cess without losing the favorable stabil- keywords: internet, servers
given by approximations Ak and f k con- ity properties of the classical proximal
verging to the original data in the sense of method. The NEOS Server has made optimization
a Mosco type topology. We prove that the In addition we analyse criteria for the methods widely available over the Inter-
 1 k
sequence xk  
Ak  αk J µ 
f which inexact computation of the proximal it- net, but it continues to rely on individ-
results from a Tychonov type regulariza- erates. These allow us to incorporate ual users to select solvers appropriate to
tion process converges weakly and, under approximations, such as enlargements of their problems. We describe new utili-
some conditions, converges strongly to the monotone operators, and to compute ties that can assist in solver selection, by
the minimum norm solution of the orig- the iterates by a suitable discretisation. examining a submitted AMPL problem
inal inclusion. We show how this re- For solving the discretised problems we prior to optimization. Given a problem
sults can be applied in order to regular- make use of generalised distances, such instance produced by the AMPL trans-
ize convex optimization problems with as Bregman distances, forcing the iterates lator, an analyzer first checks for prop-
perturbed data, that is problems whose to stay in the interior of the "feasible" set. erties ranging from size and sparsity to
objective functions and/or the constraints Finally, we address implementation linearity and convexity. A second utility
are given by approximations. matters and present a scheme coupling then compares the results to a database of
a successive discretisation of the varia- solver characteristics, to produce a list of
A proximal point method for tional inequality with regularisation by recommended solvers. Experiments on
convex vector-valued optimiza- the generalised proximal point method. standard nonlinear problem libraries and
tion Conditions on the choice of the regularis- on the archive of previous NEOS submis-
ing distance functions and monitoring of sions are reported.
A LFREDO N OEL I USEM control parameters ensuring linear con-
Instituto de Matematica Pura e Aplicada vergence are discussed. The practical ef-
ficiency of the algorithm is demonstrated Interactive, synchronous nim-
coauthors: Benar Svaiter, Henri Bonnel bus method for multiobjective
keywords: multi objective optimization, by application to contact problems from
mathematical mechanics. optimization
vector valued optimization, proximal
point methods K AISA M IETTINEN
TU1-308/T3 MODELING LANGUAGES AND SYSTEMS

University of Jyvaskyla
We present a proximal point method for
finding weakly effficient points of a func- Modeling languages and coauthor: Marko M. Makela
tion defined on a Hiilbert space with val- systems III keywords: multiple criteria
ues in a Banach space, with an order in- organizer/chair : Robert Fourer programming, interactive methods,
duced by a closed convex cone. Our XML representation of mathe- nonlinear programming, software
method defines a sequence whose func- matical programming models
tional values decrease in this order, with- We describe an unconventional philoso-
out imposing a predetermined scalariza-
for distributed optimisation ap- phy in the methodology development for
tion. We analyze first an exact version,
plications multiobjective optimization. The presen-
demanding exact solution of the subprob- PATRICK VALENTE tation is given in the context of the inter-
lems, and then move over to an inex- CARISMA, Brunel University active NIMBUS method, where the so-
act one, wihich allows for constant rela- lution process is based on the classifi-
coauthor: Gautam Mitra cation of objective functions. The idea
tive errors in the solution of the subprob- keywords: modelling, internet, linear
lems, following the criterion introduced is to formulate several scalarizing func-
programming tions, all using the same preference infor-
by Solodov and Svaiter for scalar-valued
optimization. In both cases we estab- DRAFT subject to changes mation of the decision maker. Thus, op-
lish weak convergenge of the generated In recent years, there has been grow- posed to fixing one scalarizing function
sequence to a weakly efficient point, as- ing interest on the provision of optimisa- (as is done in most methods), we utilize
suming existence of such points. tion tools via the internet. This paradigm several functions in a synchronous way.
 67

This means that we as method develop- the average number of iterations in the the supply chain without assuming a spe-
ers do not make the choice but calculate first and the third phase by looking at cific distribution, while remaining highly
the results of different scalarizing func- random figures of the underlying poly- tractable and providing insight into the
tions and leave the final decision to the hedron. These bounds show, that IPMs corresponding optimal policy. It also al-
expert, the decision maker. Simultane- solve LPs in strongly polynomial time lows to adjust the level of robustness of
ously, (s)he obtains a better view of the in the average case, so only the dimen- the solution, in order to trade off perfor-
potential compromises corresponding to sion parameters and not the encoding mance and protection against uncertainty.
her/his preferences expressed once dur- length of the problem determine the av- An attractive feature of the proposed ap-
ing each iteration. erage behaviour of IPMs. So far, only proach is that the robust problem is of
Here we describe a synchronous vari- high probability results resp. conditional the same difficulty as the nominal prob-
ant of the classification-based NIMBUS expected values were known under the lem, that is, a linear programming prob-
method. In addition, we introduce its Rotation-Symmetry-Model. The new re- lem when there are no fixed costs and
implementation WWW-NIMBUS oper- sults on the average number of steps are a mixed integer programming problem
ating on the Internet. WWW-NIMBUS total expectations and are valid for all when fixed costs are present. Further-
is a software system capable of solving dimensions. more, we show that the structure of the
even computationally demanding nonlin- optimal robust policy is of the same base-
ear problems. Besides the synchronous Comparison of complexity of stock character as the optimal stochastic
philosophy, the new versions of NIM- ipms and the ellipsoid method policy for a wide range of inventory prob-
BUS and WWW-NIMBUS also include for conic feasibility and opti- lems in single installations, series sys-
other desirable features increasing user- mization tems and general supply chains.
friendliness such as a flexible solution
database. ROBERT F REUND
MIT New math programming models
TU1-341/21 INTERIOR POINT ALGORITHMS

keywords: convex optimization, interior

for revenue management
Complexity of interior point point method, computational complexity
methods K ALYAN TALLURI
We compare different methods for an- Universitat Pompeu Fabra
chair : Robert Freund alyzing the complexity of interior-point
Expected number of iterations methods for conic and non-conic opti- keywords: pricing, revenue
of interior point algorithms for mization and feasibility problems. Mov- management, applications of
linear programming ing beyond the bit-model and condition optimization
measures for conic problems, we propose
S IMING H UANG to evaluate the complexity of convex op- We present some new stochastic pro-
Institute of Policy and Management timization using geometric measures of gramming based models for O&D rev-
keywords: interior point algorithms, the feasible region or the level sets of the enue management. Comparison with per-
linear programming, complexity feasible region. Using such measures, we fect information bounds on a hotel data
show that the computational complexity set will be presented.
We show that the expected number of it-
erations of some interior point algorithms of interior-point methods and the ellip-
is bounded above by O n  . The prob-
1 5 soid method are both bounded by the
same geometric measures. A variational inequality model
abilistic LP model is Borgwordt’s model for dynamic pricing under com-
with stable distribution. petition
TU1-341/22 LOGISTICS AND TRANSPORTATION

Average complexity of interior Applications of optimization G EORGIA P ERAKIS

point methods: the expected in pricing MIT
number of steps organizer/chair : Georgia Perakis
coauthor: Anshul Sood
P ETRA H UHN A robust optimization approach keywords: pricing, optimization,
Augsburg University to supply chain management dynamic pricing
keywords: interior point method, linear
programming, average complexity D IMITRIS B ERTSIMAS In this talk we study a model of dynamic
MIT pricing for perishable products in a com-
We are interested in the average be-
coauthor: Aurelie Thiele petitive and dynamically changing mar-
haviour of Interior-Point-Methods
keywords: robust optimization, supply ket. We take a dynamic optimization ap-
(IPMs) for Linear Programming prob-
chain management proach to build a model that ties the com-
lems (LPs). We use the Rotation-
petitive with the dynamic nature of pric-
Symmetry-Model as probabilistic model We propose a general methodology based
ing using ideas from variational inequal-
for the average case analysis. This model on robust optimization to address the
ities. We discuss the model and estab-
has been used by Borgwardt in his aver- problem of optimally controlling a sup-
lish existence results on the equilibrium
age case analysis of the Simplex-Method. ply chain subject to stochastic demand
of the competitive market. We intro-
IPMs solve LPs in three phases. First, in discrete time. This problem has
duce an algorithm for determining opti-
one has to find an appropriate starting been studied in the past using dynamic
mal pricing policies and establish its con-
point, then a sequence of interior points programming, which suffers from di-
vergence. We show some computational
is generated, which converges to the op- mensionality problems and assumes full
results and establish insights. Our results
timal face. Finally, the optimum has knowledge of the demand distribution.
apply in a number of application areas in-
to be calculated, as it is not an inte- The proposed approach takes into ac-
cluding the airline, service and retail in-
rior point. We present upper bounds on count the uncertainty of the demand in
dustries.
68 Parallel Sessions: Tuesday 9:00 – 10:30 (TU1)

TU1-341/23 pickups are performed before the corre-

LOGISTICS AND TRANSPORTATION coauthors: Mikhail Y. Kovalyov, Erwin
sponding delivery. The routes should sat- Pesch
Vehicle routing with time isfy a number of additional constraints keywords: lot sizing, approximation
windows such as time window and capacity con- schemes
chair : Oli B.G. Madsen straints.
A generalization of a single-item eco-
Solving truck scheduling prob- The implemented model is quite rich,
nomic capacitated lot-sizing problem to
lems with branch-and-price this allows us to transform several well-
the case of a non-uniform resource used
known vehicle routing problems to the
M YRNA PALMGREN for production is studied. While the
model. For example, we can solve vehi-
Division of optimization, University of general problem is shown to be non-
cle routing problems with time windows,
Linköping approximable with any constant relative
site dependent vehicle routing problems
error in polynomial time, fully poly-
coauthors: Mikael Rönnqvist, David M. and vehicle routing problems with back- nomial time approximation schemes are
Ryan hauls and time windows using just one
presented for important NP-hard special
keywords: vehicle routing, scheduling, heuristic.
cases, in which no backlogging is al-
modeling Our heuristic uses large neighbor-
lowed or holding costs are all equal to
hood search embedded in a simulated an-
We consider the log truck scheduling zero.
nealing framework. A drawback of the
problem (LTSP) which is a variant of
the pick up and delivery problem with heuristic is that it is controlled by a rather
time windows. Our application arises
large number of parameters. This prob- A Lagrangean relaxation based
in forestry and it consists of transport-
lem is partially solved by automatically branch-and-bound heuristic for
ing logs of different assortments from tuning some of the parameters while the lot sizing with setup times
heuristic is running.
harvesting points to customers such as H ALDUN S URAL
The heuristic is quite successful as
pulp and sawmills. The set of vehicles Middle East Technical University
that perform this transportation contains it has found many new best solutions to
standard benchmark problem-instances. coauthor: Meltem Denizel
trucks of several types and capacities and
these are located at different home bases. keywords: inventory/production,
The LTSP consists in finding a set of Vehicle routing with time win- branch-and-bound, heuristics
minimal cost routes, one for each vehi- dows We consider a lot sizing problem with
cle in order to satisfy the demands of setup times where the objective is to min-
the customers. Each route should start O LI B.G. M ADSEN imize the total inventory carrying cost.
and end at the driver’s home base and CTT - Centre for Traffic and Transport The problem is difficult to solve; besides,
it should meet a number of constraints coauthor: Brian Kallehauge the feasibility problem is NP-complete.
such as the working hours and customers keywords: vehicle routing, time Using the lagrangean relaxation of the in-
opening hours. We propose a mathemati- windows ventory balance constraint, we develop
cal model where each column represents This paper reports some recent results lower bound and upper bound proce-
one feasible route. In order to obtain fea- of the application of a non-differentiable dures. These are combined in a branch
sible routes we apply a hybrid column optimisation method in connection with and bound heuristic where we seek im-
generation method that starts by build- the vehicle routing problem with time provement on the initial solution. Our
ing clusters of supply and demand points windows (VRPTW). The VRPTW is an enumerative procedure is compared with
and then enumerate a number of feasible extension of the vehicle routing prob- the Trigeiro-Thomas-McClain heuristic
routes within these clusters. All routes lem. In the VRPTW the service at each on test problems taken from literature.
generated are kept in a pool. We use customer must start within an associated Computational results are given and the
branch and price with constraint branch- time window. performance of heuristics is discussed.
ing and where the pricing is applied on The shortest path decomposition of
the pool of columns. Numerical results the VRPTW by Lagrangian relaxation
are presented. Big bucket lot sizing problems
requires determining the optimal La- with changeover times
grangian multipliers. This problem is a
A general heuristic for vehicle concave non-differentiable maximization H AMISH WATERER
routing problems problem. We propose a cutting-plane al- CORE/UCL
gorithm with a trust region stabilizing de-
S TEFAN RØPKE vice for finding the optimal multipliers. coauthors: Yves Pochet, Laurence A.
Department of Computer Science at the Computational experiences will be Wolsey
University of Copenhagen reported. We have succeeded in solving keywords: lot sizing, mixed integer
coauthor: David Pisinger some problems with 400 and 1000 cus- rounding, extended formulations
keywords: vehicle routing, tomers to optimality. The use of mixed-integer rounding in-
metaheuristics equalities and a multi-commodity flow
The pickup and delivery problem with TU1-321/053 PRODUCTION AND SCHEDULING extended formulation has been success-
time windows is the problem of serv- ful in solving practical instances of big
ing a number of requests using a lim-
Lot sizing bucket lot sizing problems with setup
ited amount of vehicles. Each request in- chair : Hamish Waterer times. We report on the effectiveness of
volves moving goods from a pickup site A FPTAS for a generalized this approach when solving instances of
to a delivery site. Our task is to con- single-item lot-sizing problem these problems with sequence indepen-
struct routes that visit all sites such that dent and dependent changeover times.
corresponding pickups and deliveries are S ERGEI C HUBANOV We observe that the cuts generated by a
placed on the same route and such that University Siegen commercial solver can give lower bounds
 69

comparable to those obtained using spe- in the design of telecommunications net- K ATHIE C AMERON
cial purpose cut generation routines. The works. The originality of the model lies Wilfrid Laurier University
key is to provide a formulation which al- in the combination of investment fixed coauthors: Elaine Eschen, Chinh T.
lows the solver’s cut generator to exploit costs with congestion costs. The combi- Hoang, R. Sritharan
the appropriate variable upper bounds. natorial nature of the problem, related to keywords: computational complexity,
arc expansion decisions, is embedded in polytime algorithm, combinatorics,
TU1-321/033 NETWORKS a continuous objective function that en- stable set
compasses congestion and investment arc
Network flow problems costs. The resulting objective function is
The k-partition problem is: Given a
chair : Dorit Hochbaum graph G and a positive integer k, parti-
nonconvex and nondifferentiable though
tion the vertices of G into at most k sets,
Flows over time with flow- separable with respect to the arcs of the A(1), A(2),...,A(k), where some of the
dependent transit times network. In this work, we develop local
sets may be required to be stable sets and
optimality conditions based on the flow
K ATHARINA L ANGKAU some may be required to induce complete
structure.In particular, we show that the
TU Berlin graphs, and some pairs A(i),A(j) may be
negative cycle optimality conditions are
required to be completely joined and oth-
coauthors: Ekkehard Köhler, Martin necessary and sufficient for the routing
ers may be required to not be joined at all,
Skutella, Alex Hall and capacity expansion problem. These
or decide that no such partition exists.
keywords: networks/graphs:flow conditions extend some previous results
The list k-partition problem general-
algorithm, transportation:vehicle valid in the convex and smooth case only,
izes the k-partition problem by specify-
routing, transportation:models:traffic, and they induce a new cycle-cancelling
ing for each vertex a list of sets in which
networks/graphs:multicommodity algorithm. That algorithm is then proved
it may be placed.
Motivated by applications in road traffic to be convergent with a linear rate. We Several well-known problems can
control, we consider network flows fea- compare the new algorithm with a clas- be formulated as list k-partition prob-
turing special characteristics. In contrast sical approach which alternates capacity lems: clique cutset, stable cutset and 3-
to classical static flow theory, time plays and flow assignments and present some colourability are 3-partition problems; 2-
a decisive role. Firstly, the flow value numerical experiments obtained on real clique cutset and skew partition are 4-
on an arc may change over time. In the life large size networks. The effective- partition problems.
context of road traffic, this property re- ness of the proposed algorithm is es- We classify the list 4-partition
flects that the traffic volume on a street tablished and some extensions are dis- problems as polytime-solvable or NP-
changes throughout the day. Secondly, cussed.. complete, with a single exception. In
there are transit times on the arcs which doing so, we give polytime algorithms
may vary with the current amount of flow for many problems whose polytime-
using this arc. The latter feature is crucial A faster algorithm for bipartite solvability was open, including list 2-
for various real-life applications of flows matching and for the maximum clique cutset. Previously, Feder, Hell,
over time; yet, it dramatically increases flow problem on closure graphs Klein and Motwani had classified the
the degree of difficulty of the resulting list 3-partition problems as polytime-
optimization problems. D ORIT H OCHBAUM solvable or NP-complete, and the list
Most problems dealing with flows UC berkeley 4-partition problems as "quasipolyno-
over time and constant transit times can keywords: maximum flow, parametric mial" or NP-complete. De Figueiredo,
be translated to static flow problems in analysis, scaling, pseudoflow Klein, Kohayakawa and Reed had given a
time-expanded networks. We develop an polytime algorithm for the skew partition
alternative time-expanded network which We show that the pseudoflow algorithm problem.
implicitely models flow-dependent tran- (Hochbaum 1997) runs  on simple bipar-
sit times. Again, the whole algorithmic tite networks in time O nn1 logn  for bi- Chromatic characterization of
toolbox developed for static flows can be partite networks on n nodes with n1 nodes biclique cover
applied. This approach does not entirely on the smaller side of the bipartition.
capture the behavior of flows over time This algorithm uses a word boolean op- J EAN F ONLUPT
with flow-dependent transit times. How- eration to identify a “merger”. Université Paris 6
ever, we present approximation results For the general maximum flow mini- coauthor: Denis Cornaz
for the quickest (multicommodity) flow mum cut problem on a closure graph with keywords: networks/graphs,
problem which affirm its surprising qual- m arcs and maximum source and sink computational complexity, linear
ity.  2 capacity U, our algorithm
adjacent arcs’ programming
 in O n logU  , and with additional
runs
Solving multicommodity flow O mn  time to identify mergers without
A biclique of a simple graph G is the
edge-set of a complete bipartite (not nec-
problems with separable piece- the use of boolean word operations. The essarily induced) subgraph of G. We will
wise convex costs algorithm makes use of a novel combi-
nation of the pseudoflow scaling algo- study the two following problems: given
P HILIPPE M AHEY rithm and parametric maximum flow pro- a graph G, find the maximum weighted
ISIMA - Universite Blaise Pascal cedure. biclique of G and find the minimum num-
ber of bicliques which cover the edge-set
coauthor: Mauricio C. de Souza of G. These two problems are in general
keywords: multicommodity flow, cycle TU1-321/133 GRAPHS AND MATROIDS
NP-Hard; we introduce linear program-
cancelling algorithm, network capacity Combinatorial algorithms ming relaxations of these two problems
expansion . The two linear programs may have an
organizer/chair : Kathie Cameron
We study here a continous model for the exponential number of constraints with
routing and capacity expansion problem Finding list partitions of graphs respect to the number of nodes of G;
70 Parallel Sessions: Tuesday 9:00 – 10:30 (TU1)

however we prove that these linear pro- model conclusions in relation to uncer- of positive semidefiniteness. This sub-
grams can be solved in polynomial time tain model parameters. This methodol- class games can be considered as ana-
by the use of the ellipsoid method. We ogy is applied to a mathematical model log of convex programming problems in
prove that the separation problem can be of world regions forming a coalition, for nonlinear programming. Of course, ev-
reduced to the following problem: find cooperatively reducing CO2 emissions. ery game from this subclass is a system
in a network the shortest weighted odd This model first determines the world re- of two convex programming problems.
path between a source an a sink; in our gions’ payoffs for each possible coali- To compute Nash equilibrium we scalar-
case this last problem may be solved in tion structure. Then the model verifies ize the game and present it as equilib-
polynomial time and as a consequence coalition stability by selecting the Nash- rium problem with functional constraints.
the separation problem for our linear pro- equilibrium coalition structures from the The Lagrange function of this equilib-
grams can be solved in polynomial time. set of all possible coalition structures. rium problem is the function of three
The applied methodology consid- variables and generate a double saddle
ers all uncertain model parameters as point. To solve the game it needs to com-
Revisiting oracle methods for stochastic variables. As a consequence, pute a double saddle point of Lagrange
well-described polytopes the stability of a specific coalition struc- function for equilibrium problem.
ture can be interpreted as a bernoulli vari- For solving any two-person nonzero-
JACK E DMONDS able. For estimating this stability likeli- sum game in above subclass we offer new
None hood, monte carlo simulation can be ap- extra-proximal method. The convergence
keywords: oracle methods, plied independent of the probability dis- of this method is proved. Derived re-
well-described polytopes, tribution type of the the stochastic model sults spread to a games with coupled con-
combinatorics, ellipsoid method parameters. An efficient algorithm is dis- straints, that is, if a functional constraints
cussed for the specific case when model depend on a parameters.
It is well known that Khachian’s ellipsoid parameters have a symmetric probabil-
method, and other methods, using a sep- ity distribution, such as the multivari-
aration oracle for a polytope P, will find ate normal- and multivariate t probability On properties of several refine-
either a point in P or else a polynomial distribution. ments of optimal solutions in
size set L of inequalities valid for P such linear programming
that the volume of the solution-set of L M. A RÁNZAZU
is zero. We describe a simple combinato- Extra-proximal method for
E STÉVEZ -F ERNÁNDEZ
rial algorithm which uses this to optimize computing Nash equilibria for
two-person nonzero-sum game University of Tilburg
over P or determine that P is empty.
coauthor: M.Gloria Fiestras-Janeiro
A NATOLY A NTIPIN keywords: linear programming, linear
TU1-305/205 GAME THEORY
Computing Center of Russian Academy complementary, perfect solution, proper
Game programming of Sciences solution
chair : M. Aránzazu Estévez-Fernández keywords: equilibrium programming, In this paper we investigate the proper-
game programming, Nash equilibrium,
Stability of coalitions in a Nash- computing equilibria ties of the optimal solutions we obtain
cartel game: a probability anal- when we translate the concepts of per-
ysis We present new results on convergence fect, proper and weakly proper solutions
the extra-proximal method to compute a from the context of Linear Complemen-
N IELS O LIEMAN Nash equilibrium point for two-person tary into the framework of Linear Pro-
Wageningen University nonzero-sum game:find x1 x2 such that gramming. Other refinements of optimal
coauthor: Eligius Maria Theodorus solutions can be formulated using the re-
x1 Argmin
f1 x1  x2  ϕ1 x1  g1 x1  0  x1 X1

lationship between Linear Programming

Hendrix
keywords: model robustness, stability



and Game Theory. In this paper is also
x2 Argmin
f2 x1  x2  ϕ2 x2  g2 x2  0  x2 X2

likelihood, efficient algorithm, Nash analysed the relationship between the so-
In this class of games it is very natu- lutions we obtain when we use the rela-
equilibrium ral to highlight smaller subclass of games tionship between Linear Complementary
In this paper, a methodology is pre- with so called convex structure. It can and Linear Programming or Game The-
sented for specifying the robustness of be done by means of special inequality ory and Linear Programming.
 71

TU2-302/41 present several approaches for coping

COMBINATORIAL OPTIMIZATION coauthors: Abilio Lucena, Nelson
with this difficulty. Maculan, Mauricio G.C Resende
Flows over time and fluid Firstly, we show that static, length- keywords: Lagrangean relaxation,
networks bounded flows lead to provably good cutting planes, prize collecting, network
organizer/chair : Martin Skutella multicommodity flows over time. Sec- design
ondly, we investigate ‘condensed’ time-
Multicommodity flows over expanded networks using a rougher dis- In this paper we address the problem
time: efficient algorithms and cretization of time. We prove that a of finding a minimal weight tree in an
complexity solution of arbitrary precision can be undirected graph with nonnegative edges
M ARTIN S KUTELLA computed in polynomial time through costs and nonnegative vertex penalties,
MPI Saarbruecken an appropriate discretization leading to named the Prize Collecting Steiner Prob-
a condensed time-expanded network of lem in Graphs (PCSPG). In this problem,
coauthors: Alex Hall, Steffen Hippler polynomial size. In particular, our ap- the weight of the tree is the sum of its
keywords: networks/graphs:flow proach yields fully polynomial approxi- edges plus the sum of penalties of those
algorithm, network/graphs: mation schemes for the NP-hard quickest vertices not spanned by the tree. Using an
multicommodity min-cost and multicommodity flow prob- approach proposed by Beasley, we for-
Flow variation over time is an important lems. For single commodity problems, mulate the PCSPG as a restricted mini-
feature in network flow problems arising we show that storage of flow at interme- mum forest problem, which is amenable
in various applications such as road or air diate nodes is unnecessary; and our algo- to be solved by a Lagrangean relaxation
traffic control, production systems, com- rithms do not use any. procedure. In our approach, some in-
munication networks (e. g., the Internet), equalities are dualized when they first
and financial flows. The common char- Approximately optimal control become violated and are dropped when
acteristic are ‘dynamic’ networks with of fluid networks their corresponding multipliers become
capacities and transit times on the arcs zero. As soon as new lower bounds
which specify the amount of time it takes JAY S ETHURAMAN are found, we properly temporarily mod-
for flow to travel through a particular arc. Columbia University ify the costs of the edges present in the
Moreover, in contrast to static flow prob- coauthor: Lisa Fleischer Lagrangean solution and call an upper
lems, flow values on arcs may change keywords: fluid models, dynamic flows, bounding procedure based on the Goe-
with time in these networks. time expanded networks, continuous mans and Williamson approximative al-
While the ‘maximum s-t-flow over linear programming gorithm. Computational results are pre-
time’ problem can be solved efficiently sented.
and ‘min-cost flows over time’ are known We consider a broad class of separated
to be NP-hard, the complexity of (frac- continuous linear programs (SCLPs) that
tional) ‘multicommodity flows over time’ arise as fluid relaxations of multiclass Primal-dual algorithms for
has been open for many years. We prove queueing networks; such relaxations are prize-collecting Steiner tree
that this problem is NP-hard, even for typically used to find approximate so-
series-parallel networks, and present new lutions to complex job shop scheduling C RISTINA F ERNANDES
and efficient algorithms under certain as- problems. The complexity of this prob- University of Sao Paulo
sumptions on the transit times or on the lem is not well understood, but there is
network topology. As a result, we can evidence to suggest that optimal solu- coauthors: Paulo Feofiloff, Carlos E.
draw a complete picture of the complex- tions may have exponential size. Ferreira
ity landscape for flow over time prob- For any given ε 0 and δ 0, we keywords: approximation algorithms,
lems.  finds a solu-
present an algorithm that primal-dual algorithms, prize-collecting
tion with value at most 1  ε  OPT  δ, Steiner trees
where OPT is the value of the optimal
Quickest flows over time solution. The complexity of our algo- The primal-dual scheme has been
rithm and the size of the solution we used to provide approximation algo-
L ISA F LEISCHER produce are polynomial in the size  of rithms for many problems.  Goemans
the input network, 1 ε, and log 1 δ  .
Carnegie Mellon University and IBM
n 1  -
1
and Williamson gave a 2
coauthor: Martin Skutella We introduce a natural discretization of approximation for the Prize-Collecting
keywords: multicommodity, flow polynomial size and prove that this dis- 
Steiner Tree Problem that runs in
algorithms, approximation schemes, cretization produces a solution with low O n3 logn  time—it applies the primal-
polynomial time cost. This is the first polynomial time dual scheme once for each of the n ver-
Flows over time (also called dynamic algorithm with a provable approximation tices of the graph. We present  a primal-
flows) generalize standard network flows guarantee for this class of problems. dual algorithm that runs in O n2 logn  ,
by introducing an element of time. They as it applies this scheme only once, and
naturally model problems where travel TU2-302/42 COMBINATORIAL OPTIMIZATION
achieves the slightly better ratio of 2 2n .
and transmission are not instantaneous. Prize-collecting Steiner trees We also present a correct ratio and analy-
Traditionally, flows over time are solved sis of an algorithm by Johnson, Minkoff
in time-expanded networks that contain organizer/chair : Gunnar W. Klau and Phillips that uses the primal-dual
one copy of the original network for each A relax and cut algorithm for scheme only once. Tight examples are
discrete timestep. While this method the prize collecting Steiner given for each of the algorithms.
makes available the whole algorithmic problem in graphs
toolbox developed for static flows, its
often fatal drawback is the enormous A LEXANDRE C UNHA The fractional prize-collecting
size of the time-expanded network. We Federal University of Rio de Janeiro Steiner tree problem on trees
72 Parallel Sessions: Tuesday 13:15 – 14:45 (TU2)

G UNNAR W. K LAU optimization problems over oriented ma- is proved. Computational results are re-
Konrad-Zuse-Zentrum fuer troids and more. ported and compared with those obtained
Informationstechnik with competing heuristics.
coauthors: Ivana Ljubic, Petra Mutzel, On the non-rank facets of the
Ulrich Pferschy, Rene Weiskircher stable set polytope of claw-free TU2-302/45 COMBINATORIAL OPTIMIZATION

keywords: networks: tree algorithms, and circulant graphs Traveling salesman I

analysis of algorithms, discrete location,
engineering G AUTIER S TAUFFER chair : Genevieve Benoit
ROSO-IMA-FSB-EPFL On the subtour-elimination
We consider the fractional prize-
collecting Steiner tree problem on trees. coauthors: Thomas M. Liebling, Bianca polytope
This problem asks for a subtree T con- Spille, Gianpaolo Oriolo
taining the root of a given tree G  V E  keywords: stable set polytope, claw-free K EVIN C HEUNG
graphs, circulant graphs, clique family University of Waterloo
maximizing the  ratio of the vertex prof-
its ∑v  V  T  p v  and the edge costs inequality coauthor: Bill Cunningham
∑e  E  T  c e  plus a fixed cost c0 and We deal with the non-rank facets of the keywords: polytope, traveling salesman
arises in energy supply management. We stable set polytope (SSP) of claw-free problem, sphere
experimentally compare three algorithms graphs. We first state some known results The subtour-elimination polytope (SEP)
based on parametric search: the binary and we try to capture the difficulty of is the feasible region of the linear pro-
search method, Newton’s method, and characterizing this SSP for the wide class gramming relaxation of an important
a new algorithm based on Megiddo’s of claw-free graphs. Then we present integer linear programming formulation
parametric search method. We show im- and discuss a conjecture of Oriolo about of the Symmetric Travelling Salesman
proved bounds on the running time for the SSP of quasi-line graphs. Finally we Problem. In this talk, we look at some
the latter two algorithms. The best theo- extend some results of Giles and Trotter properties of the SEP when the graph is

retical worst case running time, namely showing that for any non-negative integer not necessarily complete. In particular,
O  V  log  V   , is achieved by our new a, there exists a circulant graph (a sub- we consider the question of the existence
algorithm. A surprising result of our class of quasiline-graphs) whose  SSP has of “inner points”, which, by a characteri-
experiments is the fact that the simple facet inducing inequality with a a  1  - zation of Rivin, is related to the problem
Newton method is the clear winner of the valued coefficients, suggesting that this of inscribing a 3-dimensional polytope in
tested algorithms. class of graph is of interest in order to un- a sphere.
derstand the conjecture.
TU2-302/44 COMBINATORIAL OPTIMIZATION
Finding the exact integrality gap
Combinatorial optimization On the problem of finding mini- for small travelling salesman
IV mum fundamental cycle bases problems
chair : Francesco Maffioli F RANCESCO M AFFIOLI S YLVIA B OYD
Vertices, edge-directions and Politecnico di Milano SITE, University of Ottawa
combinatorial optimization coauthors: Edoardo Amaldi, Giulia coauthor: Genevieve Benoit
Galbiati, Leo Liberti, Nelson Maculan keywords: combinatorial optimization,
U RIEL G. ROTHBLUM keywords: biconnected graphs, integer programming, linear
Technion fundamental cycle basis, programming relaxations
coauthors: Shmuel Onn, Yoav Tangir approximability results, heuristics
keywords: combinatorial optimization, The Symmetric Travelling Salesman
We consider the problem of finding a cy- Problem (STSP) is to find a minimum
vertices, convexity, zonotopes cle basis of minimum total weight in the weight Hamiltonian cycle in a weighted
We consider problems where a function cycle space associated with an undirected complete graph on n nodes. One direc-
h is to be maximized over a (large) fi- biconnected graph G, where a nonnega- tion which seems promising for finding
nite set B of vectors. Even when h is re- tive weight is assigned to each edge of improved solutions for the STSP is the
stricted to be convex, this framework cap- G and the total weight of a basis is de- study of a linear relaxation of this prob-
tures difficult problems like the traveling fined as the sum of the weights of all lem called the Subtour Elimination Prob-
salesman problem. Still, we develop an the cycles in the basis. This problem is lem (SEP). A well known conjecture in
approach that facilitates the efficient so- polynomially solvable in the general case combinatorial optimization says that the
lution of subclass of these problems. Our [Horton 87] but NP-hard if the basis is integrality gap of the SEP is 4/3 in the
approach consists of extending the func- required to be fundamental, i.e., to con- metric case. Currently the best upper
tion h to the convex hull P of B , deter- sist of the fundamental cycles with re- bound known for this integrality gap is
mining conditions that suffice for h to at- spect to the chords of a spanning tree of G 3/2.
tain a maximum of h at a vertex (e.g., [Deo et al. 82]. We prove that this prob- Finding the exact value for the inte-
when the extension of h is (edge-quasi) lem is MAXSNP-hard (hence does not grality gap for the SEP is difficult even
convex on P ), and the development of admit a polynomial-time approximation for small values of n due to the exponen-
an efficient method for enumerating the scheme unless P=NP) and we derive the tial size of the data involved. We describe
vertices of P . Our vertex-enumeration first upper bounds on its approximability how we were able to overcome such dif-
method relies on a mapping of the nor- for several subclasses of graphs. We then ficulties and obtain the exact integrality
mal fan of the zonotope spanned by the propose a tabu search heuristic based on gap for all values of n up to 10. Our re-
edge-directions of P onto the normal fan particular edge swaps in the correspond- sults give a verification of the 4/3 conjec-
of P and the availability of an efficient ing spanning tree. The impact of these ture for small values of n and also give
solution of linear programs over P . Ap- edge swaps is investigated and a struc- rise to a new stronger form of the conjec-
plications include partitioning problems, tural property leading to major speed-up ture which is dependent on n.
 73

the Master problem, which is a set cov- CVRP for each day of the planning pe-
Finding violated cut constraints ering problem with additional constraint riod. The MDVRP is defined on a sin-
for the STSP using a decomposi- on the number of utilized snowploughs, gle day as the VRP but the vehicles oper-
tion approach generated routes are combined to cover ate from different depots and each route
the road segments. In order to obtain must start and end at the same depot. The
G ENEVIEVE B ENOIT an integer solution in the Master prob- MDVRP can be viewed as a special case
University of Ottawa lem we have applied two heuristic pro- of the PVRP. The exact method proposed
coauthor: Sylvia Boyd cedures, one using branch-and-bound on involves the computation of a valid lower
keywords: traveling salesman problem, a subset of the columns and the other one bound by means of an additive procedure
parallel cutting plane, decomposition a greedy procedure. The solution method which combines different relaxations to
The Symmetric Travelling Salesman is implemented on the operation district derive an effective feasible solution of the
Problem (STSP) is to find a minimum of Eskilstuna located in the middle of dual of the LP-relaxation of the integer
cost Hamiltonian cycle in the weighted Sweden. This talk is connected to Lind- program. The dual solution is used by an
complete graph on n nodes. A well bergs talk on Prize Collecting Connected iterative exact procedure. The computa-
known relaxation of this problem is Subgraph. tional results on both PVRP and MDVRP
the Subtour Elimination Problem (SEP) test problems show the effectiveness of
which provides very good lower bounds Graph-coloring and linear pro- the proposed method.
for the STSP. gramming
We present a new parallel cutting
plane approach for solving the SEP. In DAVID S CHINDL TU2-306/31 INTEGER AND MIXED INTEGER
PROGRAMMING

this approach, many violated cuts are Swiss Federal Institute of Technology Integer and mixed integer
found at each stage by decomposing the coauthors: Pierre Hansen, Martine programming III
current solution into a set of nicely struc- Labbé chair : Katalin Meszaros
tured points, and then searching for vi- keywords: graph-coloring, column
olations in these points in parallel by generation, facets Branch-and-cut for cyclic staff
using an algorithm which exploits this scheduling
nice structure. We report on our results The graph  coloring problem of a graph
and discuss how these ideas could be G  V E  can be formulated as a set
covering problem, where the ground set A NDREW M ASON
adapted to other separation problems for Dept of Engineering Science
the STSP as well as other combinatorial is V , the available sets are the maximal
optimization problems. (inclusionwise) stable sets of G. Alterna- coauthor: David Panton
tively, it can be expressed as a set pack- keywords: staff scheduling,
ing problem on the same ground set, and branch-and-bound, cyclic rostering,
TU2-302/49 INTEGER AND MIXED INTEGER
PROGRAMMING
where the sets are all stable sets of G. workforce planning
Integer programming column Since the number of variables grows ex-
generation IV ponentially with the size of G, both lin- This paper considers the problem of con-
organizer : Marco Lübbecke ear relaxations require column genera- structing cyclic staff schedules. In cyclic
chair : Jacques Desrosiers tion for their optimization. Neverthe- staff schedules, a single sequence of
less, they give a tight lower bound (called shifts and days off must be generated that
Routing of snowploughs- a col- the fractional chromatic number) on the
umn generation approach defines a schedule for n staff working
chromatic number of G, hence effective over n weeks. (This is in contrast to dy-
N IMA G OLBAHARAN branch and price algorithms can be ob- namic schedules where each staff mem-
Mathematics Institution, Devision of tained with it. Both formulations are ber works their own individual sequence
Optimization compared and polyhedral results as well of shifts and days off.) We present a new
as preprocessing procedures, expressed solution approach using integer program-
coauthor: Per Olov Lindberg in simple graphical terms, are presented.
keywords: column generation, ming with 3-way branching and cuts.
branch-and-bound, set covering, set This approach includes feature branch-
partitioning An exact algorithm for period ing in which aggregate features of the so-
In the present talk we study optimal
and multi-depot VRP lution are branched on to more rapidly
focus the search. Computational results
routing of snowploughs after and during A RISTIDE M INGOZZI are presented using a customised branch
snowfall. In the first case one has to de- University of Bologna, Italy and bound implementation with a stan-
sign a set of routes starting and ending dard LP solver.
at given depots, such that each road seg- keywords: periodic vehicle routing,
ment gets ploughed within a prescribed multi depot vehicle routing, Lagrangean
time window. In the latter case the time relaxation, column generation
windows of each road segment give a We present an exact method for solv- Computational experience with
frequence according to which the road ing two generalizations of the Vehicle a cutting plane algorithm for the
segment need to be ploughed. Our so- Routing problem (VRP): the Period VRP university timetabling problem
lution approach for solving both cases (PVRP) and the Multi-Depot VRP (MD-
is based on Dantzig-Wolf decomposition VRP). The PVRP consists of designing a I GOR VASIL’ EV
combined with column generation. The set of routes for each day of a given plan- Centro di Ricerca in Matematica Pura e
subproblem of the first case is a Re- ning period of p days. The PVRP con- Applicata, Universita di Salerno
source Constrained Shortest Path Prob- sists of simultaneously selecting a day-
coauthor: Pasquale Avella
lem and in the second case a Prize Col- combination for each customer and de-
keywords: branch-and-cut, university
lecting Connected Subgraph Problem. In signing the vehicle routes by solving a
timetabling
74 Parallel Sessions: Tuesday 13:15 – 14:45 (TU2)

School timetabling problems consist of TU2-306/32 NONLINEAR PROGRAMMING approach has been shown to be effec-
assigning a set of lectures to rooms and tive in solving optimization problems in-
time periods, satisfying some basic and Optimization of engineering volving both deterministic and nondeter-
specific constraints, depending on the or- systems governed by ministic merit functions. We will de-
ganization of the educational system. simulations/PDE, part 2 scribe the mathematical formulation of
In this paper we report on a suc- organizer/chair : Natalia Alexandrov the surrogate-based optimization under
cessful experience with a cutting plane uncertainty (SBOUU) algorithm, and we
Effective parallel optimization will present several examples that illus-
algorithm for a real-world university of expensive functions
timetabling problem arising at the Fa- trate the utility of applying the SBOUU
coltá di Ingegneria of the Universitá del J OERG G ABLONSKY algorithm to engineering design studies.
Sannio, a small university in the south The Boeing Company
of Italy. The formulation of the problem
coauthors: Paul Frank, Margaret Curtin Aerodynamic design optimiza-
is based on an extension of class-teacher
keywords: surrogate model, parallel tion using the Navier-Stokes
model, where the assignment of lectures
computing, engineering equations
to rooms is considered and where classes
can be joined together for some lectures. This talk presents an approach for opti- E RIC N IELSEN
Also we consider additional constraints, mizing expensive functions that are given NASA Langley Research Center
reflecting particular requirements of the by computer simulations. It introduces keywords: design, Navier-Stokes
university. Design Explorer, a Boeing suite of re- equations
We study the polyhedral structure of sponse modeling tools for performing de- Much effort has recently focused on de-
the problem and we introduce several sign studies and optimization on such veloping a gradient-based design opti-
families of cutting planes and the related problems. The use of a newly developed mization capability based on the Navier-
separation algorithms. The cutting planes parallel computing framework leads to a Stokes equations. The presentation will
turned out to be crucial to solve all the significant reduction in the time needed introduce the fundamental components
test instances to optimality. for optimization. We will conclude the necessary to achieve such a tool and the
talk with examples of this improvement, inherent difficulties that arise with each.
and of the use of Design Explorer within The unstructured-grid analysis code that
A fuzzy logic application as a the Boeing Company. forms the foundation of the current work
contribution to solving the bus is described, followed by a discussion on
crew scheduling problem a discrete adjoint-based approach to sen-
Surrogate-based optimization
under uncertainty: formula- sitivity analysis and grid adaptation. Sev-
K ATALIN M ESZAROS tions and applications eral mesh movement schemes will also
University of Novi Sad, Faculty of be presented.
Economics A NTHONY G IUNTA
keywords: public transportation, bus Sandia National Laboratories TU2-306/33 NONLINEAR PROGRAMMING

crew scheduling, heuristic algorithm, coauthor: Michael Eldred Augmented Lagrangian

theory of fuzzy sets keywords: optimization, surrogate methods II
The scheduling of public transport in- model, uncertainty, engineering chair : Giampaolo Liuzzi
volves difficulties. The problems of crew Noisy or nonsmooth merit functions are An algorithm for generalized
scheduling take a special place in com- often encountered in computational en- semi-infinite optimization prob-
binatorial optimization. Each of them is gineering design studies due to phys- lems based on augmented La-
a special type in a manner. It is impos- ical phenomena (e.g., shocks in high- grangians
sible to exactly define them by models. speed fluid flow) or artificial phenomena
In the past a heuristic method was devel- (e.g., adaptive grid refinement, incom- E LIJAH P OLAK
oped (by the author of this abstract) for plete convergence of an iterative solver). University of California, Berkeley
the solving of bus crew scheduling prob- Numerical noise can also arise in op- coauthor: Johannes Royset
lem, which is working using a special timization under uncertainty problems keywords: generalized min-max, semi
matrix. An improvement of this heuris- when the merit function is nondetermin- infinite optimization, augmented
tic method as a step before initialization istic and is computed using a small num- Lagrangian functions
of old heuristic method is given in this ber of random samples, i.e., the statis- We present an approach for the solution
paper. This step could be useful in other tical metrics of the merit function are of a class of generalized semi-infinite
heuristics also. Each element of this ma- under-resolved. The numerical noise cre- optimization problems. Our approach
trix is an index of preference. Their eval- ates discontinuities in the value and/or uses augmented Lagrangians to trans-
uation is made on the basis of fuzzy logic gradient of the merit function, and in form the generalized semi-infinite opti-
taking into consideration the quality of many cases produces multiple local op- mization problems into ordinary semi-
buses which are possible to assign to a tima. The usual approach to solving such infinite optimization problems, with the
trip , the significance of trip and the time an optimization problem is to apply a same set of local and global solutions as
difference between arrival of i-th trip and global optimization method. However, in well as the same stationary points. Once
departure of j-th trip. The quality of the many computational engineering studies the transformation is effected, the gen-
buses and the significance of trips are de- global optimization methods are unac- eralized semi-infinite optimization prob-
scribed linguistically. This is why they ceptably expensive. We have developed lems can be solved using any available
are defined by the theory of fuzzy set. an optimization algorithm that employs semi-infinite optimization algorithm. We
The linguistic information and the fuzzy surrogate models of the merit functions, illustrate our approach with a numerical
rules bases are given by the experts of and permits the use of efficient gradient- example taken structural design subject
traffic. based local optimization methods. This to reliability constraints.
 75

and that the unit stepsize is always ac- Specifying the results presented in B.
Entropy-like Lagrangian cepted thus retaining the superlinear con- Kummer’s talk, we characterize strong
method as a unifying approach vergence rate. Preliminary numerical re- Lipschitz stability of stationary solutions
for solving convex programs sults on the COPS collection of test ex- to nonlinear programs and discuss the
amples shows the viability of the method. relation to other Lipschitz stability con-
E VA KOMAROMI cepts. We do not suppose that the
Budapest University of Economics and TU2-306/34 NONLINEAR PROGRAMMING constraints satisfy the Linear Indepen-
Public Administration dence Constraints Qualification (LICQ).
Stability of stationary points Main tool is again the analysis of suit-
keywords: augmented Lagrangian
organizer/chair : Bernd Kummer able generalized derivatives of the sta-
method, penalty/barrier methods,
entropy function tionary solution map defined via the so-
Strong Lipschitz stability of sta- called Kojima function. The condi-
Certain penalty/barrier multiplier meth- tionary solutions in variational tions obtained in this way can be formu-
ods have been shown to belong to the analysis under MFCQ: Part I lated and interpreted in terms of the ini-
family of augmented Lagrangian meth- the general condition tial data of the given optimization prob-
ods in which the quadratic term is re- B ERND K UMMER lem. Our approach and results also ap-
placed by a distance-like entropy func- Humboldt-University Berlin ply to variational inequalities, nonlinear
tion. Convergence analysis and compu- complementarity problems and equilib-
tational experiments concerning the so- coauthor: Diethard Klatte
rium models. This continues the studies
lution of linear programs and probabilis- keywords: stationary point, strong
in the authors’ book "Nonsmooth Equa-
tic constrained linear programs are pre- Lipschitz bahavior, NLO, variational
tions in Optimization: Regularity, Calcu-
sented. inequalities
lus, Methods and Applications", Kluwer,
We completely characterize  strong Lips- May 2002.
chitz stability of solutions  x a b  to sys-
Use of a truncated Newton di- tems "a  h x   NC x b  " where
rection in an augmented La- the B-normal cone at x to C b    xN CXis:
grangian framework  Newton methods for optimiza-
gi x   bi i  1    m
dimX finite, h g tion problems without con-
G IAMPAOLO L IUZZI sufficiently smooth.  Supposing  MFCQ straint qualifications
University of Rome at a solution z  x 0 0  and g z   0 (not
essential), we show that x    is locally
coauthors: Gianni Di Pillo, Stefano single-valued and Lipschitz iff both M IKHAIL S OLODOV
Lucidi, Laura Palagi (i) solutions are persistent near z for small IMPA
keywords: nonlinear programming, parameters and
large scale problems, exact augmented coauthor: Alexey Izmailov
(ii) for every Lagrange multiplier y keywords: Newton method, constraints
Lagrangians, truncated Newton methods to z and nonzero u  X such that
   degeneracy, second order sufficiency,

We 
 consider the problem y i Dg i z  u  0   i  the n-vectors v x w 
P  min f x  g x   0
, where f : Rn

Dh z   yD2 g z   u  wDg x  cannot
MPEC


R and g : R n R are three times con- vanish (in the limit) if wi Dgi z  u
0  i 
m
We consider equality-constrained opti-
tinuously differentiable functions. More and x converge to z. mization problems, where a given solu-
in particular, we are interested in the tion may not satisfy any constraint quali-
“large-scale” case, that is, when n  m is Replacing here x by z and showing fication, but satisfies the standard second-
large. that (ii) implies (i) is possible in several order sufficient condition for optimal-
In order to solve problem (P) we (not all !) situations. Other discussions of ity. Based on local identification of the
adopt an efficient and locally convergent the conditions will be given in Diethard rank of the constraints degeneracy via the
algorithm and then globalize it by means Klatte’s talk. The present one gives an singular-value decomposition, we derive
of an appropriately chosen merit func- idea of a proof. It applies tools as stud- a modified primal-dual optimality system
tion. ied for deriving sharp stability conditions whose solution is locally unique, non-
As concerns the local algorithm, on in the authors’ book "Nonsmooth Equa- degenerate, and thus can be found by
the basis of an active set strategy, we tions in Optimization: ...", Kluwer, 2002. standard Newton-type techniques. We
solve a nonlinear system approximating However, for elaborating (ii), new inves- also discuss applications to mixed equal-
the KKT conditions of optimality. More tigations of generalized derivatives at dif- ity and inequality-constrained problems,
precisely, we use a truncated Newton’s ferent points are necessary. and to mathematical programs with com-
method, based on a projected conjugate plementarity constraints (MPCC). In par-
gradient scheme, which, under reason- Strong Lipschitz stability of sta- ticular, for MPCC we obtain a local al-
able assumptions, converges with a su- tionary solutions in variational gorithm with quadratic convergence un-
perlinear convergence rate, provided that analysis under MFCQ, Part II: der the second-order sufficient condition
the starting point is sufficiently close to application to nonlinear opti- only, without any constraint qualifica-
the solution point. mization tions, not even the special MPCC con-
As concerns the globalization tech- straint qualifications.
nique, we resort to a linesearch approach D IETHARD K LATTE
based on a continuously differentiable Institut Operations Research,
exact augmented Lagrangian function. Universität Zürich
We show that, under suitable assump- coauthor: Bernd Kummer
TU2-306/35 OTHER

tions, the directions produced by the lo- keywords: strong Lipschitz stability, Software session, ILOG
cal algorithm are eventually good de- perturbed nonlinear programs, organizer/chair : Irvin Lustig
scent directions for the merit function generalized derivatives, MFCQ
76 Parallel Sessions: Tuesday 13:15 – 14:45 (TU2)

TU2-306/36 NONSMOOTH OPTIMIZATION cutting planes that give high dimensional some blocks and equations — is in a cer-
tangent planes, and show how such cut- tain neighborhood of the original prob-
Recent advances in nonsmooth ting planes can be found efficiently. lem. Then starting from the old solu-
optimization III tion, the semismooth Newton’s iterates
organizers : John Mitchell, Jean-Louis converge quadratically to the solution for
Goffin and Jean-Philippe Vial Cut and column generation the perturbed problem. Numerical ex-
chair : John Mitchell based on analytic centers for amples are provided to support the the-
large scale integer program- oretical results. The algorithm is further
The face simplex method in the ming
cutting plane framework extended to SDP, which also enjoys the
above properties.
C ESAR B ELTRAN FATMA G ZARA
Logilab - HEC - University of Geneva McGill University, Faculty of
keywords: nonsmooth optimization,
Management Approximation algorithms for
Lagrangian relaxation, cutting plane coauthor: Jean-Louis Goffin conic program with extreme ray
methods, subgradient methods keywords: cut and column generation, constraints
ACCPM, minimum spanning trees
The maximization of piecewise affine PAUL T SENG
concave functions arise, for example, in We consider a general class of integer Department of Mathematics, University
the Lagrangian relaxation of linear in- programming with an exponential num- of Washington
teger programming problems. In order ber of constraints (that can be families of keywords: conic program,
to maximize a piecewise affine concave valid inequalities). When relaxing these approximation algorithms, SDP
function, one can basically use the sim- constraints, the resulting master problem relaxation, SOCP relaxation
plex method or an interior point method. has an exponential number of variables
As an alternative, we propose the face as well as an exponential number of con- Consider a problem of minimizing a lin-
simplex method. The vertex to vertex straints. We propose a cut and column ear function subject to linear constraints
scheme of the simplex method is replaced generation algorithm based on analytic and a constraint of belonging to an ex-
by the more general face to face scheme centers to solve it. We present numeri- treme ray of a closed convex cone K. We
in the face simplex method. To improve cal results for the capacitated minimum study approximation algorithms based on
the current iterate, in the face simplex spanning tree problem that outperform relaxing the extreme ray constraint. Ex-
method, one computes the steepest ascent the literature amples with K being the semidefinite
on the current face of the objective func- cone or the second-order cone will be dis-
tion graph and then an exact line search cussed. We also present improved SDP-
determines next iterate. This new pro- TU2-306/37 CONVEX PROGRAMMING

based approximation of quadratic mini-

cedure can be used in the cutting plane Second-order cone mization with ellipsoid constraints.
framework as a substitute of the sim- programming
plex method. As a preliminary numer- chair : Michael Todd
ical test, this new version of the cut- Distance weighted discrimina-
ting plane method is compared with three An algorithm for perturbed tion: applying SOCP to pattern
other methods: subgradient, Kelley cut- second-order cone programs recognition
ting plane and ACCPM. with application to the Steiner
minimal tree problem M ICHAEL TODD
Cornell University
Properties of a cutting plane Y U X IA
method for semidefinite pro- Rutgers University coauthor: J. S. Marron
gramming keywords: pattern recognition, second
coauthor: Farid Alizadeh order cone programming, data
J OHN M ITCHELL keywords: second order cone classification
Rensselaer (RPI) programming, semismooth,
coauthor: Kartik Krishnan complementarity, semidefinite We describe a variant of the well-known
keywords: semidefinite programming, programming support vector machine (SVM) method
nonsmooth optimization, column for classifying points into one of two
The problem of minimizing a convex classes based on a training set. Instead
generation, cutting plane methods function subject to second-order cone of maximizing the minimum (perturbed)
A semidefinite programming problem is and linear consraints is reformulated into residual, or equivalently minimizing the
a nonsmooth optimization problem, so it a system of nonlinear equations. For all maximum inverse (perturbed) residual,
can be solved using a cutting plane ap- the systems of equations, the iterates of we minimize the sum of the inverse (per-
proach. In this talk, we analyze prop- semismooth Newton’s method converge turbed) residuals. We call this distance
erties of such an algorithm. We discuss quadratically, and the solution is stable weighted discrimination: it amounts to
characteristics of good polyhedral repre- and accurate. The number of variables using an L1 as opposed to L-infinity cri-
sentations the semidefinite program. We and equations of this reformulation is terion. The resulting optimization prob-
show that the complexity of an interior about half of that used by other meth- lem can be formulated as a second-order
point cutting plane approach based on a ods. Furthermore, strict complementarity cone programming (SOCP) problem. We
semi-infinite formulation of the semidefi- is not required at optimum and the initial also describe and interpret the optimality
nite program has complexity comparable points needn’t to be in the cone. Hence conditions and the dual problem. Numer-
with that of a direct interior point solver. it is good for “warm start”. Assume the ical results demonstrate the usefulness
We show that cutting planes can always perturbation of the data — including per- of this approach in the high-dimension,
be found efficiently that support the fea- turbing cost and right hand side parame- low-sample size setting. The L2 version
sible region. Further, we characterize the ters, the matrix, and adding and deleting also leads to a SOCP problem.
 77

TU2-306/38 GLOBAL OPTIMIZATION it follows that pseudoconvexity is pre-

Smoothing local searches: a new served under T. This last result is used to
Global optimization problems characterize pseudoconvexity and pseu-
with special structure approach to global optimization
dolinearity of the sums of two linear frac-
organizer/chair : Marco Locatelli B ERNARDETTA A DDIS tional functions. This last result is used
Dip. Sistemi e Informatica to find necessary and suffficient condi-
Global optimization method for tions for the sums of two linear fractional
solving the minimum maximal coauthors: Marco Locatelli, Fabio
Schoen functions to be pseudoconvex or pseudo-
flow problem linear.
keywords: smoothing method, local
YOSHITSUGU YAMAMOTO searches, basin hopping
From Kato’s theorem to a new
Institute of Policy and Planning Most methods for global optimization class of generalized convex func-
Sciences, University of Tsukuba rely on local optimization: a good global tions?
coauthors: Jun-ya Gotoh, Nguyen van optimization method should be com-
posed of a local search phase (which is
I MMANUEL M. B OMZE
Thoai Telekom Austria AG, Vienna, Austria
keywords: minimum maximal flow, best accomplished via non linear opti-
optimization on efficient set mization) and of a global phase, which keywords: generalized convexity, game
usually consists of random sampling. Re- dynamics, stability, spectral radius
The problem of minimizing the flow cently some attempts of using smooth- Motivated by research on the balance be-
value attained by maximal flows plays an ing techniques have been proposed, in the tween mutation and frequency-dependent
important and interesting role to investi- hope that both sampling and local opti- selection in evolutionary game dynamics,
gate how inefficiently a network can be mization are easier on a smoothed func- we study the following generalization of
utilized. It is a typical multiextremal op- tion; however, because of the difficulty of Kato’s theorem: if a family of bounded
timization problem, which can have local tracking the optimum towards the orig- linear operators on a vector space de-
optima different from global optima. We inal one and of the high computational pends on the parameter in a log-convex
formulate this problem as a global opti- cost, these methods display only lim- way, then so does their spectral radius.
mization problem with a special structure ited computational success. In this pa- The proof and its generalization leads to
and propose a method to combine differ- per we will present an innovative strat- the natural (and seemingly open) ques-
ent techniques in local search and global egy based on the idea of smoothing the tion how to describe a convex cone of
optimization. Within the proposed algo- results of local searches. We apply a real-valued functions with certain prop-
rithm, the advantageous structure of net- smoothing method not to the function, erties. This cone contains all convex
work flow is fully exploited so that the but to the results of a local optimization functions and and is in turn contained
algorithm should be suitable for handling routine. Obviously an analytical smooth- in the (non-convex) cone of all quasi-
the problem of moderate sizes. ing is impossible in this context; thus we convex functions. Hence, this function
use sampling in the neighborhood of the class could turn out to be important for
optimization purposes, too.
Undominated D.C. decompo- current solution in order to build an ap-
sitions of quadratic functions proximate smoothing function which is Pairs of compact convex sets
and applications to branch-and- used to guide random explorations to-
bound approaches wards the global optimum. The details of D IEHARD E RNST PALLASCHKE
this method will be presented and com- University of Karlsruhe, Inst. f.
M ARCO L OCATELLI ments on numerical results will be pro- Statistics & Math. Econ
University of Torino vided which show an extremely high ef- coauthor: Ryszard Urbański
ficiency of the proposed method in high- keywords: generalized convexity,
coauthor: Bomze Immanuel dimensional test functions. quasidifferentiable functions
keywords: global optimization, DC Pairs of compact convex sets arise in
programming the quasidifferential calculus of V.F. De-
TU2-308/11 /
GENERALIZED
CONVEXITY MONOTONICITY

We analyze various difference-of-convex Generalized convexity I myanov and A.M. Rubinov as sub- and
(d.c.) decompositions for indefinite superdifferentials of quasidifferentiable
organizer/chair : Laura Martein functions and in the formulas for the
quadratic functions. Some decomposi-
tions are dominated, in the sense that Pseudomonotonicity and pseu- numerical evaluation of the Aumann-
other decompositions exist with a lower doconvexity under the Charnes- Integral.
curvature. Obviously, undominated de- Cooper transformation Of special importance are the in-
compositions are of particular interest. clusion minimal elements in a class of
We provide three different characteriza- L AURA M ARTEIN equivalent pairs of compact convex sets.
tions of such decompositions, and show Dept. of Statistics and Applied Different types of sufficient criteria for
that there is an infinity of undominated Mathematics, University of Pisa inclusion minimal representations will be
decompositions for indefinite quadratic presented.
coauthors: Alberto Cambini, Schaible
functions. Moreover, two different pro- In the plane equivalent inclusion min-
Siegfried
cedures will be suggested to find an un- imal pairs of compact convex sets are
keywords: pseudoconvexity
dominated decomposition starting from uniquely determined up to translations.
a generic one. Finally, we address It is shown that the pseudomonotonic- This is not true for higher dimensional
applications where undominated d.c.d.s ity of the gradient of a differentiable spaces.
may be helpful: in particular, we show function is preserved under a generalized Algebraic and geometric characteri-
how to improve bounds in branch-and- Charnes-Cooper transformation T. Since zation of minimality lead to the separa-
bound procedures for quadratic optimiza- a function is pseudoconvex if and only tion property of convex sets by sets and
tion problems. if its gradient map is pseodomonotone, to the separation law. The equivalence of
78 Parallel Sessions: Tuesday 13:15 – 14:45 (TU2)

the separation law with the order cancel- problem solving environment for model- we use a simpler model for simplicity.
lation law will be shown. ing and solving large stochastic programs We solve this simple problem using both
Moreover, inclusion minimality un- will be discussed. techniques and use this example to show
der constraints is considered. how policy optimization and stochastic
Furthermore invariants of a class of A stochastic programming programming may be used in tandem to
equivalent pairs of compact convex sets model for asset and liability provide us with the best of two worlds.
are considered and it is show that the management of a finnish pen-
affine dimension of the minimal repre- sion company TU2-308/13 OTHER

sentant of an equivalence class is invari- Tucker prize

ant and that each equivalence class has P ETRI H ILLI
invariant convexificators. chair : Rainer Burkard
Helsinki School of Economics
Finally the problem pairs of con-
coauthors: Matti Koivu, Teemu
vex sets is considered in the more gen- Pennanen, Antero Ranne TU2-308/T1 FINANCE AND ECONOMICS

eral frame of a commutative ordered keywords: asset liability management, Equilibrium models
semigroups with cancellation law, which
stochastic programming, scenario tree chair : Jorge Rivera
leads to a fractional arithmetic for convex generation, out-of-sample tests
sets. A sequential land use equilib-
This paper describes a stochastic pro- rium model with endogenous
TU2-308/12 STOCHASTIC PROGRAMMING gramming model that was developed for incomes
asset liability management of a Finnish
Asset-liability management pension company. In many respects the P EDRO DANIEL JARA M ORONI
chair : Fredrik Altenstedt model resembles those presented in the Universidad de Chile
Application of dependent dis- literature, but it has some unique features coauthors: Alejandro Jofré, Francisco
crete 2-coupling with reduction stemming from the statutory restrictions Martínez
to financial management for Finnish pension companies. Partic- keywords: Walras equilibrium, urban
ular attention is paid to modeling the economics, land use
RONALD H OCHREITER stochastic factors, implementation and to In this talk we introduce a new sequen-
Department of Statistics and Decision numerical testing. Out-of-sample tests tial land use equilibrium model which
Support, University of Vienna clearly favor the strategies suggested by is based on good and location exchange
our model over fixed-mix strategies. process including endogenous incomes.
coauthors: Georg Pflug, David Giczi
We define two possible market equilib-
keywords: stochastic programming,
A comparison between rium conditions: bid-rent theory and util-
scenario generation, probability
stochastic programming and ity maximization approach. We prove in
distributions, asset liability management
parametrized policies for asset the first part of this paper that both ap-
In this paper we present a comparison of proaches are equivalent under weak con-
different algorithms to solve the discrete
liability management
ditions. In the second part we establish an
dependent 2-coupling problem with re- F REDRIK A LTENSTEDT existence result for the utility maximiza-
duction, i.e. given two discrete marginal Department of Mathematics tion equilibrium model. In the third part
probability distributions and their cor- we discuss how a dynamic version of this
relation we calculate the joint distribu- keywords: stochastic programming,
equilibrium model could work under the
tion matrix subject to maximizing the parametrized policies, asset liability
rational expectation hypothesis.
amount of zeros in the joint distribution management
to achieve a reduction. Implementations Lately, stochastic programming have General equilibrium analysis
of various algorithms - non-linear opti- been used as an effective tool for as- in ordered topological vector
mization, branch and bound and stochas- set liability management. SP does how- spaces
tic approximation - will be presented and ever have a number of drawbacks, it is
compared. The coupling problem orig- computationally expensive, the solution M ONIQUE F LORENZANO
inates from the field of stochastic pro- is rather sensitive to how the random- CNRS
gramming when two or more univariate ness is expressed and the resulting solu- coauthors: C.d. Aliprantis, R. Tourky
scenario trees have to be coupled to a tion is not always easy to interpret. As keywords: equilibrium, Edgeworth
multivariate tree during the scenario gen- the main purpose of an SP based system equilibrium, properness,
eration phase. The reduction is necessary is to recommend an action based on a Riesz-Kantorovich formula
to ensure the computational tractability state of the world, a natural alternative The second welfare theorem and the
when solving the stochastic model. Two is a parametrized policy. Here an ex- core-equivalence theorem have been
examples from the field of financial engi- plicit policy function maps the state of proved to be fundamental tools for
neering, scenario generation for asset lia- the world to an action, and a set of pa- obtaining equilibrium existence theo-
bility management based on minimizing rameters of the policy are optimized to rems, especially in an infinite dimen-
of probabilty metrics between the origi- find a policy which performs well over sional setting. For well-behaved ex-
nal distribution and its approximation as a set of test-scenarios. Such an approach change economies that we call proper
well as pricing of multivariate options, suffer from other drawbacks than an SP economies, this paper gives (minimal)
e.g. basket options, substantiate the ap- model, parametrized policy problems are conditions for supporting with prices
plicability of discrete 2-coupling with re- non-convex and a user must still choose Pareto optimal allocations and decentral-
duction and conclude the paper. Addi- the general shape of the policy. Our in- izing Edgeworth equilibrium allocations
tionally implementational issues of these terest in this line of research is driven as non-trivial quasi-equilibria. As we as-
algorithms within the AURORA Finan- by our ALM-model of a swedish life in- sume neither transitivity nor monotonic-
cial Management System, a Grid-based surance company, but in this presentation ity on the preferences of consumers, most
 79

of the existing equilibrium existence re- PAULO J OSE DA S ILVA E S ILVA of Douglas-Rachford methods, but allow
sults are a consequence of our results. A University of Sao Paulo the proximal parameter to be changed an
natural application is in Finance, where coauthor: Jonathan Eckstein infinite number of times, and in one case
our conditions lead to new equilibrium keywords: coercive regularizations, to vary even between the operators A and
existence results, and also explain why monotone operators, multiplier methods B. Morover, some of the methods permit
some financial economies fail to have a relative error criterion for approximat-
For a given real interval, we consider
equilibria. ing the resolvents of A and B.
proximal kernel functions lying between
A possible additional topic is the use
a certain upper and lower envelope, each
of Solodov-Svaiter strong convergence
The second welfare theorem of which contains a quadratic portion
forcing as an anti-spiraling technique in
with public goods in general and a coercive portion. We show that
finite dimension.
economies all kernels in this range lead to conver-
gent algorithms when applied to mono-
J ORGE R IVERA tone complementarity problems that may TU2-308/T3 MODELING LANGUAGES AND SYSTEMS

DECON, Universidad de Chile not have optimization

structure. For Modeling languages and
coauthor: Alejandro Jofré the interval 0  ∞  , the lower bound- systems IV
keywords: non convex separation, ary of the envelope is the log-quadratic chair : Bob Daniel
Pareto optimum, equilibrium kernel already shown to have desirable
qualities by Auslender and Teboulle.
Testing global optimization soft-
In this paper we prove an extension of When applied to the dual of a comple-
ware
the Second Welfare Theorem for a fi-
nite non-convex, non-transitive economy,
mentarity problem, this kernel yields a A RNOLD N EUMAIER
multiplier method related to the Chen- Universitaet Wien
where there exist externalities and public Harker-Smale-Kanzow plus function. In
goods. For those purposes we will use coauthor: Oleg Shcherbina
the same context, another kernel lying keywords: testing software
the subdifferential to the distance func- strictly within the envelope gives rise to
tion to both production and preferences a a multiplier method related to the "neu- The available test problem collections for
sets as the main tool to define the tarifica- ral network” smooth plus function. Ex- constrained global optimization and con-
tion pricing rule. Computation for quasi- perimenting with complementarity prob- tinuous constraint satisfaction problems
equilibrium vector prices and an interpre- lems from the MCPLIB, we find that the are discussed. Test results obtained on
tation of the quasi-equilibrium allocation neural network plus method is somewhat these problems for several global opti-
as a solution of a perturbed optimization faster and more reliable in practice than mization codes available are presented.
problem for each agent is also given in the loq-quadratic method, although it is
this paper. still best to use the log-quadratic method Xpress-mosel: a modular envi-
to set the multipliers in the very first it- ronment for modelling and solv-
TU2-308/T2 COMPLEMENTARITY AND VARIATIONAL
INEQUALITIES
eration. These results resemble empirical ing optimization problems
Developments in proximal results obtained by Chen and Mangasar-
ian for smoothing methods not using La-
B OB DANIEL
algorithms Dash Optimization
grange multipliers.
organizer/chair : Jonathan Eckstein keywords: modeling languages
Interior proximal methods for Separation-based proximal Mosel is a new environment for mod-
non-paramonotone variational splitting methods elling and solving problems that is pro-
inequalities vided in the form of the Xpress-IVE de-
B ENAR S VAITER velopment environment (GUI), libraries
R AINER T ICHATSCHKE IMPA- Instituto de Matematica Pura e for embedding, or a standalone program.
Dept. of Mathematics, University of Aplicada Mosel includes a language that is
Trier coauthor: Jonathan Eckstein both a modelling and a programming
coauthor: A. Kaplan keywords: splitting methods, proximal language, combining the strengths of
keywords: regularization, Bregman point, projection, separator these two concepts. Unlike traditional
function, monotone operators Given two maximal monotone oper- modelling environments like AMPL, for
which the problem is described using a
ators A and B, a splitting  method modelling language and algorithmic op-
for the problem 0  A x   B x  may
For variational inequalities characteriz-
ing saddle points of Lagrangians associ-
evaluate only resolvents of A and B, erations are written with a scripting lan-
ated with convex programming problems
but not of A  B. Methods of this guage, Mosel has no separation between
in Hilbert spaces the convergence of an
type have previously fallen into three model definition statements (e.g., declar-
interior proximal method based on Breg-
classes: the little-used double-backward ing a decision variable or expressing a
man distance functionals is studied. The
class, the forward-backward class gen- constraint) and execution statements (e.g.
convergence results admit a successive
eralizing gradient projection, and the optimizing the problem).
approximation of the variational inequal- Thanks to this synergy, one can pro-
Douglas/Peaceman-Rachford class.
ity and an inexact treatment of the prox- gram a complex solution algorithm by
We propose a new class of splitting
imal iterations. An analogous analysis interlacing modelling and solving state-
methods based on separators for a certain
is performed for finite-dimensional com- ments.
higher-dimensional set whose projection
plementarity problems with multi-valued
is the solution set. The separators are
monotone operators.
computed using (approximate) resolvents TU2-341/21
INTERIOR POINT ALGORITHMS

of A and B. Second order cone

Double regularization proximal These methods retain the attrac- programming
methods for complementarity tive theoretical convergence properties chair : Farid Alizadeh
80 Parallel Sessions: Tuesday 13:15 – 14:45 (TU2)

The primal-dual second-order FARID A LIZADEH levels. Feasible solutions are obtained
cone approximations algorithm Rutgers University by a primal heuristic based on Simulated
coauthor: Yu Xia Annealing. Some preliminary computa-
C HEK B ENG C HUA keywords: semidefinite programming, tional results are presented.
University of Waterloo
interior point method, second order cone
keywords: second order cones, programming, symmetric cones
symmetric cone programming, interior
Design of virtual private net-
point method We present a new algorithm for solv- works under aggregate traffic
ing optimization problems over symmet- uncertainty
We explore the idea of second-order cone ric cones called the Q method. Sym-
approximations for convex conic pro- metric cones are the natural generaliza- M USTAFA C. P INAR
gramming. Given any open convex cone tion of both the cone of positive semidef- Bilkent University
K, a logarithmically homogeneous self- inite matrices and the second order cone. coauthors: Francesco Maffioli, Edoardo
concordant barrier for K and any pos- Our algorithm uses as its main tool box Amaldi, Pietro Belotti, Oya
itive real number r 1, we associate, properties of Euclidean Jordan Algebras. Ekin-Karasan, Francesco Maffioli
with each direction x  K, a second-order The algorithm is a direct generalization keywords: virtual private network
cone K̂r x  containing K. We show that the Q method for semidefinite program- design, aggregate traffic uncertainty,

K is the interior of the intersection of ming. The main idea behind it is us- multicommodity flow, mixed integer
the second-order cones K̂r x  , as x ranges ing the spectral decomposition of the de- programming
over all directions in K. cision variables and search in the space
Using these second-order cones as of eigenvalues and Jordan frames asso- We consider a setting in which a group of
approximations to cones of symmet- ciated with the variables, rather than di- nodes, situated in a large, widely accessi-
ric positive definite matrices, we de- rectly working with the variables them- ble network, wishes to reserve bandwidth
velop a new polynomial-time primal-dual selves. We compare this method with on this network for secure private com-
interior-point algorithm for semi-definite the special case of semidefinite program- munication. A virtual private network
programming. The algorithm is extended ming and second order cone program- (VPN) is a service that supports such
to symmetric cone programming via the ming, review its advantages and discuss requirements by building a virtual sub-
relation between symmetric cones and its local and global convergence. network where bandwidth is reserved for
Euclidean Jordan algebras. this particular group of nodes. In a re-
cent paper, Gupta et al. (2000) addressed
TU2-341/22 LOGISTICS AND TRANSPORTATION
this problem when the traffic generated
On treating second order cone Design in communication and among the nodes are not known with cer-
problem as a special case of allocation tainty prior to designing the VPN. They
semidefinite problem defined different versions of the prob-
chair : Chung-Piaw Teo
G ONGYUN Z HAO lem according to desired VPN topology,
National University of Singapore Internet protocol network de- under the assumption that an aggregate
sign and routing traffic matrix is given, and investigated
coauthor: Chee-Khian Sim the complexity and approximability sta-
keywords: second order cone problem, P ETER B ROSTROM tus of the problems. In the present paper,
semidefinite problem Dept. of Mathematics we propose an algebraic representation of
It is well known that a vector is in coauthor: Kaj Holmberg the VPN design problem under aggregate
a second order cone if and only if its keywords: internet protocol, network traffic demand uncertainty as a mixed-
"arrow" matrix is positive semidefinite. design, heuristics, routing integer programming problem with mul-
But much less well-known is about the ticommodity flow structure. We investi-
In the Internet Protocol Network De-
relation between a second order cone gate properties of the formulation as well
sign and Routing problem, we design
problem (SOCP) and its corresponding as practical solvability issues and exten-
networks for data communication at the
semidefinite problem (SDP). The corre- sions.
same time as deciding a suitable metric
spondence between the dual problem of for data distribution. Network demands
SOCP and SDP is quite direct and the
correspondence between the primal prob-
are routed according to the OSPF pro- Berth allocation planning opt-
tocol, i.e. shortest paths with respect to mization in container terminal
lems is much more complicated. Given the metric has to be used for each ori-
a SDP primal optimal solution which is
not necessarily "arrow-shaped", we can
gin/destination pair. Each link is associ- C HUNG -P IAW T EO
ated with a fixed charge and a stepwise National University of Singapore
construct a SOCP primal optimal solu- capacity extension cost. The total design
tion. The mapping from the primal op- coauthors: Jim Dai, Wugin Lin,
cost and the level of reliability compare Rajeeva Moorthy
timal solution of SDP to the primal opti- solutions. A design is considered as reli-
mal solution of SOCP can be shown to be keywords: berth allocation
able when a large amount of the demands
unique. Conversely, given a SOCP pri- can be redistributed even if a single arc is In this talk, we present a study on the
mal optimal solution, we can construct removed, i.e. when a link failure occurs. problem of allocating berth space for
a SDP primal optimal solution which is Demands are rerouted with the same met- in-coming vessels in a container termi-
not an "arrow" matrix. Indeed, in general ric during failures as when the network is nal, which is referred to as the berth
no primal optimal solutions of the SOCP- in normal state. We present a mathemati- allocation planning problem. Combin-
related SDP can be an "arrow" matrix. cal formulation for the design and routing ing techniques from sequence-pair con-
problem and we suggest a Lagrangean re- cept in rectangle packing, and stochas-
The q method for optimization laxation that provides a lower bound of tic network stability analysis, we design
over symmetric cones the total design cost for different security an effective berth allocation system for
 81

this problem. We address the trade- George Mason University TU2-321/053 PARALLEL COMPUTING

offs between throughput, berth utiliza-

tion, choice of planning time window,
coauthor: Bradford Wood Parallel computation in
berth-on-arrival etc. In a moderate load
keywords: treatment, PDE-constrained optimization: semidefinite
setting, extensive simulation results show
optimization programming and related
that the proposed berthing system is able issues
to allocate space to most of the call- Radiofrequency ablation (RFA) is a min- organizers : Mituhiro Fukuda and
ing vessels upon arrival, with majority of imally invasive technique for killing tu- Masakazu Kojima
them allocated to the preferred berthing mors. A needle diode is placed at the chair : Mituhiro Fukuda
location. In a heavy load setting, we tumor site, and alternating current in the
range of radiofrequency is applied. This High performance grid comput-
show that surprisingly, sometime it is ing for mathematical program-
beneficial to deliberately delay vessels in causes ionic agitation, which in turn cre-
ates friction heat. Temperatures in excess ming
order to achieve higher throughput in the
berthing system. of 50 Celsius kill tissue. RFA has re- K ATSUKI F UJISAWA
cently emerged as a leading method for National Institute of Advanced Industrial
treatment of hepatic tumors, since most Science and Technology, Grid
TU2-341/23 BIOINFORMATICS AND OPTIMIZATION
liver cancer patients are not candidates Technology Research Center
Optimization in medicine I for surgical resection. The ablation treat-
ment plan is to determine the number of coauthors: Masakazu Kojima, Satoshi
organizer/chair : Panos M. Pardalos Matsuoka
needles and their positions, to guarantee
Survival-time classification of that the entire tumor is killed while dam- keywords: grid computing, cluster
breast cancer patients and age to vital healthy tissue (such as vital computing, optimization problem
chemotherapy blood vessels or the colon) is minimized. Grid computing has recently received
Since the spread of heat within the organ much attention as a powerful and inex-
O LVI M ANGASARIAN is governed by the bio-heat equation, this pensive methodology for solving large
University of Wisconsin is a PDE-constrained optimization prob- optimization problems that an existing
coauthors: Y.- J. Lee, W. H. Wolberg lem. We discuss the problem and present single CPU cannot process. Ninf is a
keywords: breast cancer, chemotherapy, initial solution approaches. grid computing infrastructure which en-
classification, support vector machines ables us to easily access computational
The identification of breast cancer pa- resources including hardware and soft-
tients for whom chemotherapy could pro- Beam geometry and intensity ware library distributed across a wide
long survival time is treated here as a map optimization in IMRT via area network. The Ninf system employs
data mining problem. This identification mixed integer programming a client-server model, where the server
is achieved by clustering 253 breast can- and client machines are connected via a
cer patients into three prognostic groups: local area network or the Internet. We
E VA L EE have been applying the Ninf system to
Good, Poor and Intermediate. Each of School of Industrial and Systems
the three groups has a significantly dis- optimization problems and polynomial
Engineering, Georgia Institute of systems of equations. Among others, we
tinct Kaplan-Meier survival curve. Of Technology
particular significance is the Intermediate implemented highly parallel algorithms
group, because patients with chemother- coauthors: Tim Fox, Ian Crocker which utilize several PC clusters con-
apy in this group do better than those keywords: mixed integer programming, nected via the high speed local area net-
without chemotherapy in the same group. cancer treatment, radiation therapy work and/or the Internet. In this talk,
This is the reverse case to that of the we discuss grid computing for mathemat-
overall population of 253 patients for ical programming showing some numer-
In Intensity-modulated radiation therapy
which patients undergoing chemotherapy ical results.
(IMRT) not only is the shape of the beam
have worse survival than those who do controlled, but combinations of open and
not. We also prescribe a procedure that closed multileaf collimators modulate the Parallel computation for
utilizes three nonlinear smooth support intensity as well. In this talk, we offer semidefinite programming
vector machines (SSVMs) for classifying a mixed integer programming approach M AKOTO YAMASHITA
breast cancer patients into the three above which allows optimization over beam-
prognostic groups. These results sug- Tokyo Institute of Technology
let fluence weights as well as beam and
gest that the patients in the Good group couch angles. Computational strategies, coauthors: Katsuki Fujisawa, Masakazu
should not receive chemotherapy while including a constraint and column gen- Kojima
those in the Intermediate group should erator, a specialized set-based branching keywords: nonlinear programming,
receive chemotherapy based on our sur- scheme, a geometric heuristic procedure, semidefinite programming, parallel
vival curve analysis. To our knowledge and the use of disjunctive cuts, are de- computation, interior point method
this is the first instance of a classifiable scribed. Our algorithmic design thus far One approach to solve large-scale SDPs
group of breast cancer patients for which has been motivated by clinical cases. Nu- at high speed is to combine primal-dual
chemotherapy can possibly enhance sur- merical tests on real patient cases reveal interior point methods and parallel com-
vival. that good treatment plans are returned putation on PC cluster. We developed
within 30 minutes. The MIP plans con- a parallel version of the SDPA, the SD-
Optimal treatment planning for sistently provide superior tumor cover- PARA (SemiDefinite Programming Al-
radiofrequency ablation of liver age and conformity, as well as dose ho- gorithm PARAllel version) on PC cluster.
tumors mogeneity within the tumor region while The main feature of the SDPARA is to
maintaining a low irradiation to impor- process the linear system with the Schur
A RIELA S OFER tant critical and normal tissues. complement coefficient matrix which is
82 Parallel Sessions: Tuesday 13:15 – 14:45 (TU2)

solved at each iteration to generate a I RINA D UMITRESCU

search direction. This part is known to be TU Darmstadt Increasing distances in graphs
the most time consuming when solving coauthor: Natashia Boland by deleting minimum edge sets
many large scale SDPs by primal-dual in- keywords: networks, Lagrangean
terior point methods. The SDPARA com- relaxation, constrained shortest path
S TEFAN K RAUSE
putes the elements of the Schur comple- TU Braunschweig
ment matrix in parallel and applies a par- Given a directed graph that has a cost and keywords: min-cut, shortest paths, set
allel Cholesky factorization to the linear an amount of consumed resource asso- cover
system on distributed memory. Through ciated with each arc, the Resource Con- Given a simple and undirected graph G
numerical results on 64 PCs, we show straint Shortest Path Problem (RCSPP) the distance increasing problem (DIP) is
that the SDPARA can solve large-scale consists of finding the least cost path be- to find a minimum edge set such that
SDPs at high speed that we could not tween two specified nodes such that the deleting these edges from G increases the
achieve before. The SDPARA also at- total amount of resource consumed is less distance of given pairs of vertices by at
tains a high scalability on parallel com- than a specified value. The RCSPP is least a given amount, where vertices of
putation. NP-hard. Preprocessing can be very im- different components have distance ∞. In
portant when a RCSPP is solved. We this talk some special cases of DIP are
will briefly show the effect of preprocess- discussed. We give polynomial time al-
Parallel semidefinite program- ing and propose a new Lagrangean relax-
ming algorithm using matrix gorithms and proofs of NP-hardness.
ation approach to solving the RCSPP. We
completion will present an improved label setting al-
gorithm that uses preprocessing informa- TU2-321/133 GRAPHS AND MATROIDS

K AZUHIDE NAKATA tion as well as all Lagrangean multiplier Graphs and discovery
Tokyo Institute of Technology information collected in a Lagrangean re- organizer/chair : Pierre Hansen
coauthors: Katsuki Fujisawa, Masakazu laxation step. Numerical results obtained
Kojima, Makoto Yamashita for randomly generated and real life test Computer-assisted research in
keywords: semidefinite programming, problems will be provided. graph theory
sparsity, matrix completion, parallel D RAGAN S TEVANOVIC
computing
The minimum congestion short- Faculty of Science and Mathematics
When we solve large scale SDPs by est path routing problem keywords: graph theory, computers in
primal-dual interior-point methods, the research, graph editing, conjecture
positive semidefinite matrix variable of- A NDREAS B LEY testing
ten becomes large and fully dense. To Konrad-Zuse-Zentrum Berlin One of the first computer programs to as-
overcome this difficulty, we proposed to keywords: network design, shortest path sist graph theorists in research, the sys-
use a positive semidefinite matrix com- routing, mixed integer programming tem GRAPH was developed at the Uni-
pletion technique a few years ago. This versity of Belgrade, Yugoslavia, by Dra-
technique worked very effectively for Given a capacitated network and some gos Cvetkovic and his collaborators dur-
SDPs with large scale sparse data matri- traffic demands between its nodes, the ing 1981–1984. We present the sys-
ces satisfying special structures. In gen- minimum congestion shortest path rout- tem GRAPH and, based on experience
eral sparse cases, however, it requires ing problem is to find non-negative with it, study two important aspects of
more cpu time to compute search di- lengths for the network’s edges such that such programs: graph editing and con-
rections at each iteration than the stan- for the resulting shortest path routing the jecture testing from the viewpoint of
dard primal-dual interior-point method. maximum flow/capacity ratio of the links user-friendliness and user-time manage-
An important feature of this technique is is minimized. This problem is of great ment. Implementing these observations,
considerably less memory spent. To take practical interest, since the most com- we also present a first working pro-
full advantage of this feature and to re- monly used routing protocol in the Inter- totype of GRAPH 2, which is to be
duce much cpu time to compute search net, OSPF, is based on shortest path rout- linked to the world-class isomorphism-
directions, we incorporate a parallel ma- ing. We consider only unsplittable short- tester nauty, various graph generators and
trix completion technique into the SD- est path routing, i.e., for each demand Mathematica package Combinatorica.
PARA, a parallel version of the SDPA there must be a unique shortest path be-
(Semidefinite Programming Algorithm), tween its terminals.
which employs a parallel Cholesky fac- We discuss the relation between path What forms do interesting con-
torization of the Schur complement ma- systems that arise from shortest path jectures have in graph theory?
trix. Our numerical results show that routings and stable sets in some asso- G ILLES C APOROSSI
the parallel matrix completion technique ciated hypergraphs, formulate the min- HEC Montreal
works very efficiently for some large imum congestion shortest path routing coauthors: Mustapha Aouchiche, Pierre
scale and sparse SDPs. problem as a mixed-integer linear pro- Hansen, Dragan Stevanovic
gram, and present a branch-and-cut-and- keywords: graphs, conjecture
price algorithm for its solution. Com-
TU2-321/033 NETWORKS

putational results for various real-world Conjectures in graph theory have multi-
Shortest path problems problem instances are reported. ple forms and involve graph invariants,
chair : Stefan Krause We also show that it is NP-hard graph classes, subgraphs, minors and
to approximate the minimum congestion other concepts in premises and/or conclu-
A Lagrangean relaxation ap- problem within a sions. Various abstract criteria have been
proach to solving the resource 
routing
shortest path
  
constrained shortest path prob- factor of O I  1 2 ε  , ε 0, where I  proposed in order to find interesting ones
with computer-aided or automated sys-
is the binary encoding size of the prob-
lem lem.
tems for conjecture-making. Beginning
 83

with the observation that famous theo- Universidad de Sevilla (2002) construct other sets of marginal
rems (and others) have first been con- coauthors: Eligius Maria Theodorus vectors such that the requirement that
jectures, if only in the minds of those Hendrix, Niels Olieman, Michael Finus these marginal vectors are core elements
who obtained them, we review forms that keywords: cooperative, multiple is a sufficient condition for convexity
they take. We also give examples of con- coalitions, Nash equilibrium of a game. This construction is based
jectures of such forms obtained with the on a neighbour argument, i.e. it is
help of, or by, computers when it is the Our study is based on work by several shown that if two specific neighbours of
case. It appears that many forms are un- researchers on stability of coalition for- a marginal vector are core elements, then
explored and so computer-assisted and mation in the Kyoto protocol (see papers this marginal vector is a core element as
automated conjecture-making in graph from Carraro, Finus, Eyckmans, Olie- well. In this way they characterize con-
theory, despite many successes, is pretty man). We consider multiple coalitions vexity using a fraction of the total num-
much at its beginning. which groups of countries can join. A ber of marginal vectors. In this paper
new notation is presented for specifying we use combinatorial arguments to ob-
the stability and facilitating the imple- tain sets of marginal vectors that charac-
How far should, is and could mentation into computer coding. The re- terize convexity. We characterize these
conjecture-making be auto- sults are shown of an application to study sets of marginal vectors and present the
mated in graph theory? stability conditions to coalitions struc- formula for the minimum cardinality of
P IERRE H ANSEN tures in an Open Membership Game and sets of marginal vectors that characterize
GERAD and HEC Montreal Exclusive Membership Game with Mul- convexity.
tiple Coalitions.
keywords: graphs, conjecture, assisted,
automated Extremal convolution and
Computer-assisted and automated Characterizing convexity of games
conjecture-making in graph theory is games using marginal vectors
reviewed, focusing on the three opera-
H ERBERT H AMERS S JUR D IDRIK F LÅM
tional systems GRAPH,Graffiti and Au- Economics Dep. Bergen University
Tilburg University
tographix (AGX). A series of possible keywords: extremal convolution,
enhancements,mostly through hybridis- coauthors: Bas van Velzen, Henk Norde
production games, core solutions,
ation of these systems, are proposed as keywords: convex cooperative games
duality gap
well as several research paths for devel- This paper studies the relation between
opment of the area. Main objects here are so-called market
convexity of TU games and marginal
or production games. These are coopera-
vectors. Shapley (1971) and Ichiishi
tive, transferable-utility games for which
TU2-305/205 GAME THEORY
(1981) showed that a game is convex if
core solutions emerge, in concave in-
Cooperative games and only if all marginal vectors are core
stances, via shadow prices on endow-
elements. Rafels,Ybern (1995) show that
chair : Sjur Didrik Flåm ments. Such games can model mutual in-
if all even marginal vectors are core ele-
surance, financial equilibrium and coor-
Determination of stable struc- ments, then all odd marginal vectors are
dinated production. They can also figure
tures in a multiple coalition core elements as well, and vice versa.
as subgames of noncooperative, strategic
game Hence, if all even or all odd marginal
interaction. We explore all this and elab-
vectors are core elements, then the game
orate on what happens when payoffs are
E LENA S ÁIZ P ÉREZ is convex. Van Velzen, Hamers, Norde
not concave.
84 Parallel Sessions: Wednesday 9:00 – 10:30 (WE1)

WE1-302/41 COMBINATORIAL OPTIMIZATION from a finite range. We apply this theo- paths from s to t. A q-path flow from s
rem to TSP and the minimum bisection to t is a non-negative linear combination
Domination analysis of problem. The practical message of this of elementary q-path flows from s to t,
combinatorial optimization work is that the greedy algorithm should adhering to edge-capacities. We consider
algorithms be used with great care, since for many the following network synthesis problem:
organizer/chair : Gregory Gutin optimization problems its usage seems We are given an nxn, non-negative, sym-
impractical even for generating a starting metric integer-valued matrix R, where
Introduction to domination solution (that will be improved by a local each non-diagonal element, R(ij), repre-
analysis search or another heuristic). sents the required value of q-path flow
G REGORY G UTIN between nodes i and j in an undirected
Dept of Computer Science, Royal Approximate local search network on the node set N=1,2,...,n. We
Holloway, University of London want to construct a network G =[N, E]
A NDREAS S. S CHULZ with integer-valued edge capacities u(e)
keywords: heuristics, approximation :e in E so that each of these requirements
algorithms, domination analysis, Massachusetts Institute of Technology
can be realized (one at a time) and the
combinatorial optimization coauthors: James B. Orlin, Abraham P.
sum of all edge-capacities is minimum.
The study of Domination Analysis (DA) Punnen
We present a strongly polynomial algo-
keywords: theory of local search,
was initiated by F. Glover and A.P. Pun- rithm for the problem and show that, ex-
analysis of algorithms, 0/1-integer
nen in 1997. DA provides an alterna- cept for a very special case, the problem
tive and a complement to Aproximation programming, approximation algorithms has integer rounding property.
Analysis. In DA, we are normally inter- Local search algorithms for combinato-
ested in the minimum possible number rial optimization problems are in general Minimal multicut and maximal
(proportion, respectively) of feasible so- of pseudopolynomial running time and integer multiflow in rings
lutions that are worse or equal in qual- polynomial-time algorithms are often not
ity to the heuristic one. This is called known for finding locally optimal solu- L UCAS L ÉTOCART
the domination number (domination ra- tions for NP-hard optimization problems. CEDRIC CNAM
tio, respectively) of the heuristic solution. We introduce the concept of ε-local op- coauthors: Marie-Christine Costa,
In many cases, DA is very useful. For timality and show that an ε-local opti- Frédéric Roupin
example, it is proved in 2002 by A. Yeo mum can be identified in time polyno- keywords: multicommodity flow,
and myself, that the greedy algorithm has mial in the problem size and 1 ε when- multicut, rings
domination number 1 for many combi- ever the corresponding neighborhood can This presentation deals with the mini-
natorial optimization problems. In other be searched fast. mization of multicuts and with the max-
words, the greedy algorithm, in the worst We also discuss various extensions of imization of integral multiflows. A ring
case, produces the unique worst possible our main result, we relate it to the com- is a connected graph where all vertices
solution (for any size of instances). This plexity class PLS introduced by Johnson, have degree 2. Several simplifications
is reflected in one of the latest computa- Papadimitriou and Yannakakis, and we can be made before solving the mini-
tional experiments with the greedy algo- point out that it is best possible. In partic- mal multicut and the maximal integer
rithm, see Chapter 10 (by D.S. Johnson ular, it takes in general exponential time multiflow problems in rings. The main
et al.) in the 2002 TSP book, where it to compute a local optimum for a prob- one is that a path without terminals, ex-
was concluded that the greedy algorithm lem in PLS if it is just described by a cept for its endpoints, may be reduced
’might be said to be self-destruct’. neighborhood-search oracle. to a single edge, which is the lowest
The aim of the talk is to give a short weighted edge of the path. Moreover
introduction to DA. We’ll discuss some problems in bidirectional rings can be
results obtained recently by several re-
WE1-302/42 COMBINATORIAL OPTIMIZATION

transformed in equivalent problems in di-

searchers including N. Alon, J. Bang- Combinatorial optimization V rected rings by doubling the number of
Jensen, S. Kabadi, M. Krivelevich, F. chair : Marek Libura commodities. In fact, without loss of
Margot, A.P. Punnen, A. Vainshtein, A. A strongly polynomial algo- generality, one can assume that there is a
Yeo and the speaker. Open problems and rithm for integer version of the source and/or a sink located at each ver-
applications will also be discussed. multipath flow network synthe-  propose
tex. We  a polynomial algorithm
in O min Kn2 n3   to solve the minimal
sis problem multicut problem in ring networks. The
When the greedy algorithm fails algorithm is based on the enumeration of
S ANTOSH K ABADI
J OERGEN BANG -J ENSEN University of New Brunswick several minimum cuts associated with an
University of Southern Denmark arbitrary path, each one containing one
coauthors: R. Chandrasekaran, K. P. K. different edge of the path; these cuts are
coauthors: Gregory Gutin, Anders Yeo Nair, Y. P. Aneja obtained by using the algorithm given for
keywords: greedy algorithms, keywords: network synthesis, graphs, rooted trees. We propose also a polyno-
independence systems, combinatorial strongly polynomial algorithm, integer mial algorithm in O(n) to solve the max-
optimization, traveling salesman rounding property imal integer multiflow problem in rings
problem Given an undirected network G = [N, E], with uniform capacities. This algorithm
We provide a characterization of the a source-sink pair (s,t) of nodes in N, a is based on the continuous solution of the
cases when the greedy algorithm may non-negative edge capacity u(i,j) for each problem.
produce the unique worst possible solu- edge (i,j) in E, and a positive integer q,
tion for the problem of finding a mini- an elementary q-path flow from s to t is Adjustment problem for bi-
mum weight base in a uniform indepen- a flow of q units, with one unit of flow nary constrained linear pro-
dence system when the weights are taken along each path in a set of q edge-disjoint gramming problems
 85

M AREK L IBURA with the presentation of a new exact algo- M ICHAL S TERN
Systems Research Institute, Polish rithm which is capable of solving large- Ben Gurion University
Academy of Sciences scale QKP problems with up to 1000 coauthor: Ephraim Korach
keywords: inverse optimization, linear variables. keywords: clustering spanning trees,
programming, combinatorial clustering TSP path, stars, polynomial
optimization A survey of stochastic aspects of graph algorithms
We consider a pair of optimization prob- knapsack problems
lems with the same linear objective func- We consider the following problem: Let
tion: an initial problem and its restric- (V,S) be a given hypergraph where V is a
U LRICH P FERSCHY ground set of elements and S is a simple
tion. For the initial problem we propose University of Graz
a generalization of standard inverse prob- collection of subsets of V. Let G=(V,E)
keywords: knapsack problem, be a complete graph with a cost func-
lem and call it the adjustment problem. stochastic analysis, on-line problem
The adjustment problem consists in find- tion. Find in G a minimum cost span-
ing such minimum norm perturbations of Although the knapsack problem is NP- ning tree T such that each subset of S
the objective vector, which guarantee that hard, highly successful optimal algo- induces a subtree in T. In this talk we
an optimal solution of the perturbed ini- rithms are available for almost all types consider two special cases of this prob-
tial problem is also an optimal solution of instances. Hence, it is of particular in- lem: (i) The clustering-TSP-path prob-
of the given restriction. We propose a terest to study stochastic properties of the lem in which the spanning tree is re-
method of solving the adjustment prob- knapsack problem to gain more insight stricted to be a Hamiltonian path, (ii) The
lem when the initial optimization prob- into its structure. optimal-stars-clustering-tree problem in
lem is a linear programming problem and In this survey we will review the usu- which each subset induces a star. For the
all variables in the restriction are binary ally applied probabilistic models and the first problem finding a feasible solution is
variables. We illustrate the approach with main structural results describing e.g. the equivalent to the Consecutive Ones Prop-
two examples: For a given graph we cal- value of the optimal solution, the num- erty solved by Booth and Lueker. We
culate such minimum norm perturbations ber of items where integer and continu- present a polynomial algorithm for find-
of arc lengths which guarantee that the ous optimal solutions differ, and the re- ing an optimal solution for a restricted
shortest path connecting two given ver- sulting integrality gap. Furthermore, a case. For the latter problem we present
tices belongs to some specified subset of number of algorithmic results will be in- a structure theorem and a polynomial al-
paths. Similarly, we calculate such min- cluded, concerning both the expected so- gorithm. A motivation is to construct a
imum perturbations of objective coeffi- lution quality of simple algorithms, in minimum cost communication network
cients in the continuous knapsack prob- particular greedy-type methods, and the with the above restrictions for a collec-
lem, which guarantee that an optimal so- expected running time of optimal algo- tion of non-disjoint groups of customers
lution of the perturbed problem is inte- rithms, which can be bounded by a poly- such that each group induces a subtree so
gral. nomial. that the network will provide "group fault
Related problems such as the sub- tolerance" and "group privacy". For the
clustering-TSP-path problem there is an-
WE1-302/44 COMBINATORIAL OPTIMIZATION
set sum problem and the d-dimensional
other motivation from robotics.
Knapsack problems knapsack problem will also be consid-
ered. Special emphasis will be put on the
organizer/chair : David Pisinger
on-line knapsack problem, where an un- Solving the TSP with
A survey of upper bounds derlying time-line defines the availability decomposition-based pricing
and exact algorithms for the of items.
quadratic knapsack problem J OHN R AFFENSPERGER
University of Canterbury
DAVID P ISINGER A survey of approximation algo-
DIKU, University of Copenhagen rithms for knapsack-type prob- keywords: programming, integer,
lems algorithms, relaxation/subgradient
coauthors: Anders Bo Rasmussen,
Rune Sandvik In this paper, we propose to solve the
keywords: quadratic knapsack problem, H ANS K ELLERER Traveling Salesman Problem using two
integer programming, upper bounds, University of Graz methods. First, we generate columns us-
knapsack problem keywords: knapsack, approximation ing the Held & Karp 1-tree subgradi-
The quadratic knapsack problem (QKP) algorithms, dynamic programming ent optimisation procedure, to be used in
maximizes a quadratic objective function the standard Dantzig-Fulkerson-Johnson
We give a survey on approximation al-
subject to a linear capacity constraint. (DFJ) formuation for the TSP. We restrict
gorithms (heuristics, PTAS, FPTAS) for
Due to its simple structure and challeng- the DFJ master to arcs found in the sub-
knapsack-type problems. Especially, we
ing difficulty it has been studied inten- gradient optimisation procedure. Second,
treat the classical knapsack problem,
sively during the last two decades. We we use the same procedure for generating
the subset sum problem, the multiple
will give a survey of upper bounds pre- columns, but we use a master formulation
knaspack problem, the multidimensional
sented in the literature, show the relative for the TSP based on Martin?s LP for-
knapsack problem and other variants.
tightness of several of the bounds, and mulation for the minimum spanning tree.
experimentally compare their strength In both cases, we get good bounds on
and computational effort. Techniques WE1-302/45 COMBINATORIAL OPTIMIZATION
the optimal tour. Use of the DFJ formu-
for deriving the bounds include relax- Traveling salesman II lation requires standard subtour break-
ation from upper planes, linearisation, re- ing constraints. The second method does
chair : Nicholas Pearson
formulation, Lagrangian relaxation, La- not use the standard subtour breaking
grangian decomposition, and semidefi- On the optimal clustering TSP constraints, but rather employs auxiliary
nite programming. The talk is concluded path and stars problems variables to restrict the assignment to
86 Parallel Sessions: Wednesday 9:00 – 10:30 (WE1)

a 1-tree. The two algorithms resem- early 1990s and has its roots in disjunc- WE1-306/31 INTEGER AND MIXED INTEGER
PROGRAMMING
ble Mamer & McBride?s decomposition- tive programming from the 1970s. Sev-
based pricing algorithm for linear pro- eral papers have been published with Integer and mixed integer
gramming. promising results, but these implementa- programming IV
tions have all relied on an external linear chair : Agostinho Agra
or mixed integer solver. Exact solution to separable inte-
Fast algorithms for planar We present here the first commercial
graphs, Part II: the traveling ger programming
implementation of lift-and-project cuts
salesman problem where the cut generation is integrated D UAN L I
N ICHOLAS P EARSON with the solver. It relies on perform- Chinese University of Hong Kong
Lancaster University ing regular simplex iterations of the lin- coauthors: Jun Wang, Xiaoling Sun
ear programming relaxation instead of keywords: integer programming,
coauthor: Adam Letchford the classical approach of solving a sepa-
keywords: traveling salesman problem, Lagrangian duality, objective level cut,
rate linear program. We compare the per- dynamic programming
combinatorial optimization, planar formance on various classes of publicly
graphs, branch-and-cut available problem instances. A convergent Lagrangian and objective
level cut method is proposed for solv-
It is well-known that the TSP remains ing separable integer programming prob-
NP-hard even when the underlying graph A branch-and-cut algorithm for lems. The method exposes an optimal
is planar. However, a polynomial-time nonconvex quadratic program- solution to the convex hull of a revised
approximation scheme is known for the ming perturbation function by successively re-
planar TSP, whereas the general TSP is shaping the perturbation function. The
APX-hard. G EORGE N EMHAUSER objective level cut is used to eliminate the
We present evidence to suggest that ISYE - Georgia Tech duality gap and thus to guarantee the con-
the planar TSP may also be easier to coauthor: Dieter Vandenbussche vergence of the Lagrangian method on a
solve to optimality than in the general keywords: branch-and-cut, nonconvex, revised domain. Computational results
case. In particular, we show that the quadratic programming, polyhedron are reported for a variety of separable in-
separation problem for the subtour elim- teger programming problems with up to
By reformulating quadratic programs us-

ination and 2-matching constraints can 5000 integer variables.
be solved in O n3 2  time in the planar ing necessary optimality conditions, we
case, whereas  for general sparse graphs present a branch-and-cut approach to
Discrete time LQR optimal con-
the bound is O n2 logn  for
 subtour elim-
solve nonconvex instances. For the
trol problem with logical con-
ination constraints and O n3 logn  for 2- bound constrained case, we study a relax-
straints
ation based on a subset of the optimality

matching constraints. Moreover, for pla-
nar graphs there is an O n3  separation conditions. By characterizing its convex D ORIN P REDA
algorithm for the domino-parity inequal- hull, we obtain a large class of valid in- ENSEEIHT-INP
ities, which include the comb inequali- equalities. These inequalities are tested
within a branch-and-cut scheme and the coauthor: Joseph Noailles
ties as a special case, whereas in the gen- keywords: mixed integer programming,
eral case no polynomial-time algorithm is computational results demonstrate the ef-
ficiency of the algorithm. discrete time LQR, logical constraints,
known. GOA GBD
We also present a small number of We are interested in a discrete time
graphs for which the subtour elimination On the use of Gomory’s cyclic LQR optimal control problem with log-
and domino-parity inequalities do not group polyhedra in cutting- ical constraints. These constraints re-
suffice to describe the (graphical) TSP plane generation strain the total time of control on each
polyhedron, and conjecture that these are fixed length time interval. Therefore the
the excluded minors for this property. S ANJEEB DASH problem naturally belongs to the Mixed
This talk will draw on results given IBM T. J. Watson Research Center
Integer Programming class; each con-
in a companion talk by Adam Letch- coauthor: Oktay Gunluk trol variable has its correspondent binary
ford, entitled "Fast algorithms for planar keywords: integer programming, variable expressing his presence or ab-
graphs, Part I: cycles and cuts". cutting planes, cyclic group sence in the LQR dynamics.
Gomory introduced certain cyclic group We first explore the integer feasible
WE1-302/49 INTEGER AND MIXED INTEGER
PROGRAMMING polyhedra in 1969 as relaxations of sin- space and present mathematical consid-
Integer programming gle constraints in integer programs. Go- erations allowing us to reduce the size
mory showed that facets of these poly- of this space and to give useful proper-
organizer/chair : Daniel Bienstock ties of an admissible integer solution. We
hedra yield cutting planes for the associ-
A practical implementation of ated integer programs; additional proper- then present two different formulations
lift-and-project cuts ties, and cutting plane ideas, can be found of our problem. As usual mixed integer
in some recent papers. We continue the approach involves the generation, at each
M ICHAEL P ERREGAARD study of these polyhedra. In particular, iteration, of a lower bound and an up-
Dash Optimization we study a class of facets described in per bound of the optimal solution. Our
Araoz, Evans, Gomory, and Johnson (the first model, formulated as a Mixed Inte-
keywords: cutting planes,
two-slope facets), and present a simple ger Quadratic Problem, can naturally be
lift-and-project, mixed integer
principle - analogous to the basic MIR solved using classical algorithms: Gener-
programming
principle but with three variables and one alized Benders Decomposition and Gen-
Lift-and-project cuts for mixed integer constraint - and derive the validity of eralized Outer Approximation. We show
programs were first introduced in the these facets from this principle. why this is not the best approach and
 87

propose, for our second model (a Mixed complex design problems. Shape op- Complex engineering design optimiza-
Integer Nonlinear Problem), an effective timization of existing surface ships is tion problems governed by computa-
way of estimating the successive lower performed through the solution of CFD tional simulations rarely lend themselves
bounds wich ensures a rapid convergence solvers, included a high-fidelity finite- to straightforward formulation and so-
of the algorithm. volume RANSE code for the prediction lution by conventional methods of non-
For both formulations we present nu- of the free surface flow past the ship free linear programming. Two of the most
merical results that seem to be very to dynamically sink and trim. In the difficult aspects are the multidisciplinary
promising for the approach that we pro- adopted formulation, the accuracy of the nature of the problems and the expense
pose. metamodels evolve during the optimiza- of repeated function evaluation via high-
tion cycles. Results show up to a factor fidelity simulations, such as the Navier-
MIP cuts based on knapsacks of 5 in terms of CPU time. Stokes equations of computational fluid
with 2 integer variables dynamics. Such problems are not only
Boundary shape optimization prohibitively expensive, but they can also
AGOSTINHO AGRA for fluids and waves thwart optimization because function and
University of Aveiro constraint evaluation is not robust. We
M ARTIN B ERGGREN address the expense of simulations via a
coauthor: Miguel Constantino Division of Scientific Computing, rigorous use of models simplified both in
keywords: knapsack sets, valid Department of Information Technology, physics and in mesh resolution to obtain
inequalities Uppsala University solutions to models of higher physical fi-
We consider MIP relaxations leading to keywords: shape optimization, adjoint delity with significant savings in terms of
integer knapsacks with 2 integer vari- equations, Helmholtz equations, Euler high-fidelity simulations. We address the
ables and a continuous variable. The con- equations multidisciplinary nature of the problem
vex hull of solutions for these models can We discuss similarities and differences via a modular approach to problem for-
be obtained in polynomial time. In order between two shape optimization prob- mulation, in which basic computational
to extend these inequalities to the orig- lems: drag reduction of airfoils or air- components can be combined into differ-
inal model, we consider sequence inde- craft wings, and transmission optimiza- ent optimization formulations and solu-
pendent lifting functions. Computational tion of acoustic or microwave transduc- tion algorithms with relative ease.
experience is reported. ers. A common problem is how to trans-
late the intension of changing the geome- WE1-306/33 NONLINEAR PROGRAMMING

WE1-306/32 NONLINEAR PROGRAMMING

try into a mathematical formulation that Least square problems
Optimization of engineering make sense in a nonlinear optimization
context. We choose to consider normal chair : Jerry Eriksson
systems governed by
simulations/PDE, part 3 deflections from a given, fixed reference On global minimization of
configuration. The normal deflection, in procrustes-penrose regression
organizer/chair : Natalia Alexandrov turn, is the solution of a second-order problems
High-fidelity solvers in the de- elliptic boundary-value problem defined
sign optimization of ships on the reference configuration. This for- T HOMAS V IKLANDS
mulation has the advantage of being (i) Department of computing science
E MILIO F. C AMPANA general and parameter free, (ii) produc- coauthor: Per Åke Wedin
INSEAN - The Italian Ship Model Basin ing smooth design updates with speci- keywords: weighted orthogonal, least
coauthor: D. Peri fied behavior at edges of the boundary, squares, several minima, iterative
keywords: approximation management, (iii) being able to handle geometric con- algorithms
nonlinear programming, ship design straints such as convexity, volume, and
Consider the set of matrices Q with or-
local displacement in a uniform way. An
The use of high fidelity, CPU time expen- thonormal
 columns defined by Vm n 
issue that differs between the fluid and
Q  Rm n : QT Q  In 1  n  m
.

sive solvers is now quite common in the wave application is how easy it is to
trial-and-error design process. Their use accurate compute gradients by the dis- The set Vm n is called the Stiefel mani-
in a simulation-based design process is crete adjoint-equation approach. Fully fold. We want to minimize a real val-
however still reduced by the large amount discrete expressions for Jacobians and ued quadratic or linear function of Q
of computational effort needed in a de- gradients are complicated for the fluid over the Stiefel manifold. One problem
sign problem. To reduce the computa- equations (compressible Euler). A hy- of that kind is the Procrustes problem
tional cost of running complex engineer- brid approach combining hand-derived min   AQ B   2F s.t. Q  Vn n . This prob-
ing simulations, different techniques may expressions with Automatic Differentia- lem has an unique minimum value. The
be adopted to obtain global approxima- tion seems to be promising: almost as ef- solution can be computed by using the
tion of the high-fidelity solvers, as the ficient as fully hand-coded Jacobians, but SVD of AT B. Most quadratic problems
widely used metamodelling techniques. much less error-prone to implement. defined on the Stiefel manifold have sev-
An alternative approach is the Variable eral minima. We have made a detailed
Fidelity Modeling (VFM), which com- study of algorithms for one problem of
bine less accurate but also incomparably
Problem formulation and mod- some practical interest, the Penrose re-
less expensive solvers with high-fidelity
eling in simulation-based design gression problem min   AQX B   2F s.t.
models. The use of trust region strate- NATALIA A LEXANDROV Q  Vm n . For this problem we have con-
gies and of consistency conditions finally NASA Langley Research Center structed an algorithm that computes all
ensure the accuracy of the optimal solu- keywords: simulation based the minima with a method of a new kind.
tion. Applications of metamodels and of optimization, PDE-constrained Having computed one minimum we use
VFM with trust region strategies are pre- optimization, variable fidelity modeling, the "normal plane" to the surface defined
sented for both single and multi-objective multidisciplinary optimization by AQX, Q  Vm n for the minimizer Q
88 Parallel Sessions: Wednesday 9:00 – 10:30 (WE1)

to compute a set of orthonormal matrices. methods such as a Gauss-Newton method limits on the size of problems that ap-
Remarkably each of these matrices lies in performs well. proach can efficiently handle. For that
the vicinity of other minima due to the The theory presented also reveals in- reason we have developed SLIQUE, an
special geometry of the surface. formation of the curvature properties in active set algorithm that uses solution of a
both the parameter space and the function linear programming subproblem to deter-
The least square values and the space. This information is potentially mine a working set, on which a quadratic
Shapley value very useful already in an early phase approximation of the Lagrangian is mini-
when modelling an inverse problem. mized. In this talk we study the behavior
I RINEL D RAGAN We will give practical algorithms of this algorithm, considering its global
University of Texas, Mathematics and their performance based on Gauss- convergence and the asymptotic choice
keywords: Shapley value, least square Newton, Quasi-Newton, and conjugated of the working set. We also analyze ex-
values gradient methods. We also present nu- perimental results on the stability of the
merical results when optimizing feed- active set, and on the cost of linear pro-
The Least Square Values of cooperative
forward neural networks, which includes gram solution.
TU games (briefly LS-values), have been
convergence rates but also the problems
introduced as optimal solutions depend-
curvature properties.
ing on parameters of a quadratic opti-
mization problem by M. Keane (1969), An f nonmonotonic filter algo-

and axiomatized by L. Ruiz, F. Valen- WE1-306/34

NONLINEAR PROGRAMMING
rithm for NLP
ciano and J. Zarzuelo (1998). The Shap- Large scale nonlinear
ley value is a member of the family. programming I ROGER F LETCHER
In an earlier paper of the author (Dra- organizers : Sven Leyffer and Richard University of Dundee
gan,1992), a so-called average per capita Waltz
formula for the Shapley value was proved chair : Richard Waltz coauthor: Sven Leyffer
and a computational algorithm has been keywords: filter algorithm,
derived. In the present paper, we prove An interior-point L1-penalty nonmonotonicity, SQP, trust region
an Average per capita formula for the LS- method for nonlinear optimiza-
values. The main result is that any LS- tion This paper extends the global conver-
value is the Shapley value of a game eas- D OMINIQUE O RBAN gence proof for an SQP trust region fil-
ily obtained from the given game. This Northwestern Univeristy ter algorithm given by Fletcher, Leyf-
fact led to the so-called poten- tial ba- coauthors: Nick Gould, Philippe L. fer and Toint, SIAM J. Optim., 13,2002,
sis of the space of games relative to LS- Toint 44-59. This algorithm exhibits a cer-
values and to the solution of the inverse keywords: primal-dual interior method, tain degree of nonmontonicity, but reverts
problem: for a given n-vector, find out L1 penalty, mixed interior/exterior to a monotonic algorithm when applied
the set of games for which the LS-value We discuss the merits of a mixed to an unconstrained problem. For such
is the given vector. The inverse prob- interior/exterior-point method for non- problems, nonmonotonic line searches,
lem for Shapley value was earlier solved linear programming in which all non- such as that proposed by Grippo, Lam-
(Dragan,1991). pariello and Lucidi, have proved very
linear constraints are treated by an 1 effective. In this talk it is first shown
penalty function. Inspired by a pro- how the GLL modification can be applied
The superiority of using regu- posal by Mayne and Polak, a suitable de- in an unconstrained trust region context.
larization over trust-region for composition of the constraints allows us A simple proof is sketched, which il-
ill-conditioned nonlinear least to derive an exact differentiable penalty lustrates a feature of the GLL algorithm
squares function involving only inequality con- that may not be well understood. It is
straints, which may then be treated using shown how these ideas can be introduced
J ERRY E RIKSSON a logarithmic barrier. Exactness of the
Umeå University into the FLT filter algorithm, in such a
exterior penalty function eliminates the way that nonmonotonicity in the uncon-
keywords: regularization, inverse need to drive the corresponding penalty
problems, Gauss-Newton, nonlinear strained case is allowed. Some familiar-
parameter to infinity. Global and fast lo- ity with the FLT proof will be assumed.
least squares cal convergence of the proposed scheme
Trust-region methods are very efficient are exposed.
and a natural first choice when solving
nonlinear least squares. Their main ad- Convergence and constraint ac- WE1-306/35 OTHER

vantages include strong global conver- tivity in a successive LP algo-

gence properties, simplicity and relia- rithm
Software session, COIN-OR
bility. However, for problems that are organizer/chair : Ted Ralphs
ill-conditioned in the neighbourhood of R ICHARD B YRD
a solution point (local minimizer) these University of Colorado
methods suffer from weak theoretical and coauthors: Nick Gould, Richard Waltz WE1-306/36 NONSMOOTH OPTIMIZATION

practical convergencee rate. As alter- Jorge Nocedal Cutting plane methods for
nativ, for such ill-conditioned problems, keywords: nonlinear programming, conic optimization problems
we propose methods based on regulariza- active set, trust region organizers : John Mitchell, Jean-Louis
tion. We use two well-known techniques; Although it is sometimes desirable to Goffin and Jean-Philippe Vial
Tikhonov (lagrangian) regularization and solve large-scale constrained nonlin- chair : John Mitchell
truncation to augment (or modify) the ear optimization problems by an ac-
objective function. This leads to well- tive set method, techniques of active- Cutting plane methods for
defined problems on which first-order set quadratic programming appear to put semidefinite programming
 89

K ARTIK K RISHNAN coauthor: Jean-Louis Goffin problem. These bounds also depend on
Rice University keywords: second order cone, other natural problem parameters such as
coauthor: John Mitchell semidefinite cut, analytic center, cutting the size of the data A b c  and the com-
keywords: semidefinite programming, planes plexity parameter for the self-concordant
nondifferentiable optimization, cutting Recent applications of nonsmooth opti- barrier of the convex set P.
plane methods mization deal with nonpolyhedral cutting
Interior point methods for semidefinite planes such as semidefinite cuts (max-
programming (SDP) are fairly limited in cut problem). An efficient cutting plane Convex separable optimization
the size of problems they can handle. technique in nonsmooth optimization is problems with bounded vari-
Cutting plane methods provide a means the analytic center cutting plane method ables
to solve large scale SDP’s cheaply and (ACCPM). This method uses subgradient
quickly. We give a survey of various cut- information to generate cutting planes.
Incorporating semidefinite cuts into the S TEFAN M S TEFANOV
ting plane approaches for SDP in this pa- Neofit Rilski South-West University
per. These cutting plane approaches arise ACCPM results in a polynomial approx-
from two perspectives: the first is based imation but yields to computational dif- keywords: convex programming,
on the polynomial separation oracle for ficulties in practice. Replacing semidef- separable programming, algorithms,
the SDP that is utili zed by polynomial inite cuts by second-order cone cuts is computational complexity
interior point cutting plane methods; the considered. Theoretical issues as well
second rewrites an SDP with a bounded as pros and cons of this replacement are Consider minimization problems with a
feasible set as an eigenvalue optimiza- discussed. We show that an approxi- convex separable objective function sub-
tion problem, which in turn is solved us- mate analytic center can be recovered af- ject to a separable convex inequality con-
ing bundle methods for nondifferentiable ter adding a second-order cone cut, in one straint of the form "less than or equal
opti mization. Newton step and that the analytic center to" / linear equality constraint / linear in-
We present an accessible and unified cutting plane method with second order equality constraint of the form "greater
introduction to various cutting plane ap- cone cuts is a fully polynomial approxi- than or equal to", and bounds on the vari-
proaches that have appeared in the litera- mation scheme. ables. The three problems are denoted by
ture; in particular we show how each ap- (C), C   and C  , respectively. Such
proach arises as a natural enhancement WE1-306/37 CONVEX PROGRAMMING
problems arise, e.g., in scheduling theory,
of a primordial LP cutting plane scheme Complexity in convex in allocation of resources, in inventory
based on a semi-infinite formulation of programming control, in facility location problems, in
the SDP. chair : Stefan M Stefanov the theory of search, in subgradient op-
timization, in implementation of projec-
Condition number complexity tion methods when the feasible region is
Analytic center cutting plane for non-conic convex optimiza- of the same form as the feasible sets un-
method in conic programming tion
 etc. For each of prob-
der consideration,
VASILE BASESCU lems (C) and C   , a necessary and suf-
F ERNANDO O RDONEZ ficient condition is proved for a feasible
Rensselaer Polytechnic Institute ISE- University of Southern California
coauthor: John Mitchell solution to be an optimal solution to the
coauthor: Robert Freund respective problem, and a sufficient con-
keywords: conic programming, analytic keywords: condition number, convex
center 
dition is proved for a feasible solution
optimization, complexity to problem C  to be an optimal solu-
We analyze the problem of finding a The condition number of optimization tion. Algorithms of polynomial complex-
point strictly interior to a bounded, fully- problems in conic form, as introduced by ity for solving these problems are sug-
dimensional set from a finite dimensional Renegar, has been shown to bound the gested and convergence of these algo-
Hilbert space. We try to generalize the re- sizes of feasible and optimal solutions, rithms is proved. Some important convex
sults obtained for the LP, SDP and SOCP the changes to solutions due to changes functions for the three problems as well
cases. The cuts added by our algorithm in the data, and the complexity of algo- as computational results are presented.
are central and linear. In our analysis, we rithms to solve the problem. Although
will find an upper bound for the number any convex problem can be transformed
of Newton steps required to compute an to conic form, such transformations are WE1-306/38 GLOBAL OPTIMIZATION

’approximate’ analytic center. Also, we neither unique nor natural given the natu- Bilevel programming
will provide an upper bound for the total ral description of many problems. There-
number of cuts added to solve the prob- fore the relevance of condition number chair : Anton Evgrafov
lem. This bound depends on the quality bounds of equivalent conic formulations
of cuts, the dimensionality of the prob- for non-conic problems is questionable. A bilevel model and solution al-
lem and the ’thickness’ of the set we are Recent research has extended the def- gorithm for the stochastic taxa-
considering. inition of the condition number to non- tion problem
conic convex  optimization
 problems of
Second-order cone cuts vs. the form min c x  Ax b  CY x  P
. M ARCIA FAMPA
semidefinite cuts in the analytic We show that the condition number for Universidade Federal do Rio de Janeiro
center cutting plane method non-conic convex optimization problems
bounds the sizes of feasible and optimal coauthors: Nelson Maculan Filho,
M OHAMMAD R. O SKOOROUCHI solutions, the change in solution due to Luidi Simonetti
Assistant Professor, Cal State University, changes in the data, and the complexity keywords: global optimization, bilevel
San Marcos of interior point algorithms to solve the programming, taxation problem,
stochastic model
90 Parallel Sessions: Wednesday 9:00 – 10:30 (WE1)

Bilevel programming problems are opti- them. The resulting problem belongs to
mization problems where a subset of de- the class of bilevel programming prob- Mixed type duality for multi-
cision variables is not controlled by the lems, or mathematical programs with objective optimization problems
principal optimizer or by the leader, but equilibrium constraints (MPEC). Prob- with set constraints
by a second agent, the follower, who opti- lems of this type violate the standard non-
mizes his own objective function with re- linear programming constraint qualifica- R ICCARDO C AMBINI
spect to this subset of variables. We con- tions, are known to be non-smooth and Dept. Statistics and Applied
sider the class of bilinear bilevel prob- non-convex, and, as a result, are hard to Mathematics
lems obtained when the leader imposes solve. coauthor: Laura Carosi
taxes on a specified set of services, while We propose a smooth approximation keywords: vector optimization,
the follower makes his decision, taking of the problem, which allows us to solve generalized convexity, duality, set
into account the taxation scheme devised a sequence of smooth problems instead. constraints
by the leader and also by his competitors. We study the sequences of stationary (re-
An important obstacle observed in spectively, globally optimal) solutions to We propose a pair of vector dual pro-
practical implementation of those taxa- the approximating problems, and show grams where the primal has a feasible re-
tion problems is the uncertainty of the that their limit points are stationary (re- gion defined by a set constraint, equal-
taxes imposed by the leader’s competi- spectively, globally optimal) for the orig- ity and inequality constraints. The sug-
tors. We deal with this obstacle, consid- inal MPEC. gested dual problem can be classified as a
ering a stochastic taxation model and we “mixed type” one. The duality results are
The approach is illustrated with nu-
present an exact algorithm for this prob- proved under suitable generalized con-
merical examples.
lem. Finally, we show how the stochastic cavity properties.
taxation program can be used to model As parameters in the “mixed type”
the strategic bidding problem in whole- WE1-308/11 /
GENERALIZED
CONVEXITY MONOTONICITY
dual program take different values, dif-
sale energy markets and present some nu- Generalized convexity II ferent dual problems can be obtained and
merical results for this problem. different generalized concavity proper-
organizer : Laura Martein ties can be used in order to get the duality
chair : Riccardo Cambini results.
Heuristics for toll setting prob- Discontinuity properties of con- Furthermore, we deep on the role of
lem vex polynomial mappings vector ρ-quasiconcavity and, in the case
S OPHIE D EWEZ the feasible region has no set constraints,
Service d’optimisation / Université Libre E VGENY B ELOUSOV 
we state duality results assuming the vec-
de Bruxelles Moscow State University, Faculty of tor F ρ  -quasiconcavity.
Economics
coauthors: Martine Labbé, Patrice
Marcotte, Gilles Savard keywords: convex mappings, convex WE1-308/12 STOCHASTIC PROGRAMMING

keywords: bilevel programming, polynomial mappings, Hausdorff Modeling and computation

transportation, networks, pricing (semi)continuity issues in stochastic
We consider the problem of determining We consider a convex polynomial map- programming
a set of optimal prices on a subset of all pings (CP-mappings) given with a finite organizer/chair : Gus Gassmann
arcs of a highway network in order to system of convex polynomial inequalities
maximize the revenu of the government SMPS and SMPSReader: an in-
or private socities. A λ



x En  fi x  λi  i

1      m
 1

put format and user routines for
We can formulate this problem as stochastic programs
bilinear bilevel problem. We devel- The main goal of this work is to de-
scribe the possible structure of set Bω G US G ASSMANN
opp heuristics to solve the problem and Dalhousie University
present computational results. of partial limits for
 lower
 excess func-
tion dω λ   ρ A ω  A λ   and set Bω keywords: stochastic programming,
Smoothing by relaxing the equi- of  limits for
partial  upper
 excess
func- input format, computation
tion d ω λ   ρ A λ  A ω   as λ ω  It The SMPS format is a modification of
librium conditions in topology
optimization problems for Truss turns out that for CP-mappings inclusion
0  Bω
Bω always holds.It was shown
the MPS format for stochastic linear
structures in contact programs. It was recently extended
that for mapping of form (1) the set Bω to include chance-constrained programs,
A NTON E VGRAFOV has fairly simple structure: it can be ei- stochastic problem dimensions, quadratic
Department of Mathematics/Chalmers ther zero or a closed interval of the form objectives and free-format input.
University of Technology 0  b  b0 or ray b
0  For the set This talk gives an overview of the for-
coauthor: Michael Patriksson Bω we still do not know the full descrip- mat and describes a collection of user
keywords: structural optimization, tion of its possible structure, but we con- routines that convert the format to inter-
MPEC, bilevel programming, smoothing structed the examples of CP-mappings nal data structures, explicitly or implic-
with the following structure of set Bω :1) itly building the deterministic equivalent
We consider the problem of maximizing Bω is the union of zero and arbitrary fi-
the mechanical performance of a truss in nite set of positive numbers, 2) Bω is the linear program. These routines are avail-
unilateral, frictionless contact with rigid union of zero and either a closed interval, able without restriction to algorithm de-
obstacles, by the means of redistributing or a closed ray, 3) Bω consists of two el- velopers and other researchers.
the available material among the struc- ements - finite (zero) and infinite   ∞  In
tural members. Some parts of the struc- this talk we suggest the conjecture about Scalability and implementation
ture can be completely removed by al- possible structure of the set Bω for CP- issues in stochastic program-
locating the material of volume zero to mappings. ming algorithms
 91

C HANDRA P OOJARI keywords: natural gas, decision support more than 2.000 km of pipelines and 10
Centre for the analysis of risk and systems, dynamic markets, value chain compressor stations.
optimisation modelling applications, optimization
Brunel university This presentation gives an overview of
coauthors: Frank Ellison, Gautam a 4-year ongoing project SINTEF and WE1-308/T1 FINANCE AND ECONOMICS

Mitra, Suvrajeet Sen NTNU are doing for Statoil. Portfolio planning
keywords: decomposition, multi-stage, The aim is to develop methodology chair : Gautam Mitra
warm start and decision support models that incor-
We present an overview of the current porate an overall view of production, A review of risk measures with
progress in solving Stochastic program- transportation and market behavior for application in financial portfolio
ming problems and discuss the outstand- the supply of natural gas from the Nor- analysis
ing issues the needs to be addressed for wegian continental shelf. A major chal-
the further growth of the field. We em- lenge is to establish decision-making pro- D IANA ROMAN
phasis the need for decomposition based cesses that facilitate a dynamic coordina- Brunel University, Department of
structure exploiting algorithms to pro- tion of the activities aimed at the market Mathematical Sciences
cess practical instances of two stage and place. This is especially important con-
coauthors: Ken Darby-Dowman,
multi-stage SP problems. Using indus- sidering the gas directive from the Euro-
Gautam Mitra
trial models and models from the SP pean commission. It is expected that the
keywords: risk measure, expected
test library, we will compare the perfor- directive and the initiatives to liberalize
utility, stochastic dominance, coherence
mances of two such approaches- Ben- the European energy markets will lead to
ders’ decomposition and Stochastic de- a new situation for the gas producers with
We consider the general problem of com-
composition. Our industrial models are a particular focus on increased short-term
paring random variables assuming an
from the domains of finance, supply variation in supply and demand.
outcome as large as possible is preferred.
chain and telecom. We will discuss Statoil is Norway’s largest gas pro-
This problem has widespread application
the techniques such as parallelisation and ducer, producing about 30% of the nat-
in decision making in general and is of
warm starts through which such SP algo- ural gas on the Norwegian shelf and
particular relevance in the area of port-
rithms can be scaled up to process large selling about 70% of the Norwegian
folio selection. Mean-risk models and
instances of practical problems. gas. Norwegian gas sale is approximately
expected utility/ stochastic dominance
10% of the west European consumption.
models are important methods for mod-
Subtree decomposition for mul- We give an overview of tactical and
elling choice under uncertainty. With
tistage stochastic programs operational decision support models for
mean-risk models, the choice of the risk
coordination of production, transporta-
S HANE DYE measure to be used plays an important
tion and sales of natural gas in Sta-
University of Canterbury role in capturing a decision maker’s pref-
toil. The models include physical prop-
erences. Commonly used risk measures:
keywords: stochastic programming, erties of the transportation network, con-
variance, mean absolute deviation, lower
decomposition tracts and forward markets. The model-
partial moments, value-at-risk, condi-
An algorithm for solving multistage ing framework is stochastic mixed inte-
tional value-at-risk are reviewed, proper-
stochastic recourse problems is de- ger programming.
ties are stated and a short historical per-
scribed. The scenario tree is decom- spective is given. The compatibility of
posed using a cover of subtrees. The A two-stage stochastic MILP for mean-risk models with expected utility
progressive hedging algorithm is used to natural gas transmission opti- maximization and stochastic dominance
ensure implementability across the en- mization with transient flow is explored. Lastly, the coherence of risk
tire tree. The approach leads to a class measures is examined. Risk measures
of methods based on the subtree cover M ARKUS W ESTPHALEN (which evaluate the overall seriousness of
chosen (the class includes the original University Duisburg-Essen
possible losses) and deviation measures
implementation of the progressive hedg- coauthors: Andreas Maerkert, Ruediger (which measure the uncertainty in terms
ing algorithm). Computational testing Schultz of variability around the mean) are often
indicates that the method can provide keywords: stochastic programming, considered to be equivalent terms. How-
improved performance over the original integer programming, natural gas ever, when addressing the issue of co-
progressive hedging algorithm. We model natural gas transmission and herence, this distinction may be impor-
distribution in a pipeline system by lin- tant. Deviation measures fail to satisfy
WE1-308/13 STOCHASTIC PROGRAMMING
ear mixed-integer optimization. Physical the risk-free condition. The derivation
Stochastic programming laws and technological layout determine of risk measures (some of which are co-
applications in the power and the transient gas flow in pipelines and the herent) from deviation measures is exam-
gas industries features of other network elements, e. g. ined.
compressor stations. In presence of un-
organizer : COSP (Markus Westphalen) certain consumer demand we develop a
chair : Markus Westphalen stochastic integer program for minimiz- Fin4cast: from data mining to
Value chain management in the ing transportation costs. Our algorithm optimal portfolio selection
liberalized natural gas market employs a decomposition method based
on Lagrangian relaxation of nonantici- E RANDA C ELA
F RODE RØMO pativity and branch-and-bound to solve Siemens Austria AG
SINTEF Industrial Management the stochastic program. We report com-
coauthors: Asgeir Tomasgard, Matthias putational experiences with a transmis- keywords: data mining, input selection,
Peter Nowak, Marte Fodstad sion network in Germany, which includes forecast models, optimal portfolio
selection
92 Parallel Sessions: Wednesday 9:00 – 10:30 (WE1)

fin4cast is a complex system which can 19016 rows ,23963 variables and 129099
forecast financial instruments and use the nonzeroes. The linear complementarity
produced forecasts to select an optimal problem with interval data
portfolio. An optimal portfolio would
maximize the difference between protfo-
WE1-308/T2 COMPLEMENTARITY AND VARIATIONAL
INEQUALITIES
U WE S CHAEFER
lio return and portfolio risk, where the Complementarity principle in Universitaet Karlsruhe
portfolio risk is measured in a classical games keywords: linear complementarity
manner. fin4cast is used to (a) do the organizer/chair : Samir Kumar Neogy problem, enclosure method,
data mining and the input selection for applications, free boundary problem
some prespecified target(s) to be forecast Complementarity principle
and some structured stochastic We consider an LCP where the entries of
(b) build the forecast model(s) (c) ana- the matrix M and the entries of the vector
lyze the performance of the later and, (d) games
q under consideration are not explicitly
select the optimal portfolio over a set of known but can be enclosed in intervals;
prespecified financial instruments. The S AMIR K UMAR N EOGY
Indian Statistical Institute i.e., given an interval matrix [M] and an
selection of inputs is based on statistical interval vector [q] we are interested in the
methods, sensitivity analysis and search keywords: structured stochastic game, set S([M],[q]) where S([M],[q]) is the set
algorithms. The backbone of this pro- linear complementarity problem, of all vectors z such that z is a solution of
cess is the fin4cast own database which switching control game, ARAT game the LCP defined by M and q where M be-
practically covers the whole universe of longs to [M] and q belongs to [q]. Since it
Solving a general zero-sum stochastic
financial data. The forecast models range is difficult to describe S([M],[q]) we are
game problem is a very complex prob-
from classical statistical and economet- satisfied to calculate an interval vector [z]
lem. The game may not possess the or-
rical models to neural network models. which includes S([M],[q]) for the case
dered field property. For undiscounted
The portfolio selection is based on clas- that S([M],[q]) is bounded. As an appli-
games, the problem is more complex
sical quadratic programming algorithms, cation we present an enclosure method
since in general optimal (or equilibrium)
as far as quadratic optimization prob- for an ordinary free boundary problem
stationary strategies do not exist. For
lems are concerned. More complex op- which is based on an LCP with interval
the zero-sum stochastic games with some
timization problems including combina- data by discretizing the differential equa-
structural assumptions have led to algo-
torial and ranking constraints are solved tion without neglecting the discretization
rithmic results. Examples of the struc-
by heuristic algorithms developed with error. We present some numerical ex-
tured zero sum stochastic games are
fin4cast. Other important ingredients of amples which illustrate the theoretical re-
ARAT(additive reward and additive tran-
fin4cast are the tool for the selection of sults.
sition) and Switching control Stochastic
the best forecast model and the tool for
game. For structured zero sum stochas-
the optimal tuning of the portfolio selec- ON - LINE OPTIMIZATION
tic games, the scope of applicability WE1-308/T3
tion strategy.
of Lemke’s and other pivot algorithms Deterministic on-line
which uses complementarity principle optimization
Algorithms for the solution of are discussed. We show that Lemke’s chair : Joan Boyar
large-scale quadratic program- algorithm solves discounted switching
ming (QP) and quadratic mixed control games and Cottle-Dantzig’s algo- On-line matching on a line
integer programming (qmip) rithm(a generalization of Lemke’s algo- W INFRIED H OCHSTÄTTLER
models rithm) solves undiscounted ARAT games BTU Cottbus
under some mild assumptions.
G AUTAM M ITRA coauthors: Bernhard Fuchs, Walter
CARISMA, Department of Mathematical Kern
Sciences Some classes of matrices in lin- keywords: on-line algorithms,
coauthors: Frank Ellison, Marion ear complementarity problem matchings, competitive analysis
Guertler and matrix games Given a set S  R of points on the line, we
keywords: portfolio planning, quadratic 
consider the task of matching a sequence
programming, discrete optimization, A RUP K UMAR DAS r1 r2     of requests in R to points in
large scale optimization Indian Statistical Institute S. It has been conjectured that there ex-
keywords: matrix game, value of the ists a 9-competitive on-line algorithm for
The increasing acceptance of M-V model f this problem, similar to "hide and seek"
and its extensions in the finance indus- game, almost N̄ matrix, E0 matrix on a line. We disprove this conjecture
try has rekindled interest in the solution Von Neumann’s Minimax Theorem and and show that no on-line algorithm can
of large-scale convex QP models. There Kaplansky’s result on completely mixed achieve a competitive ratio strictly less
is also growing interest in the process- game were used to derive certain results than 9.001.
ing of QMIP models and their use as in Linear Complementarity. In this pa-
a practical portfolio planning tool. In per some more results are proved which
this talk we consider the algorithmic is- further establishes the connections be-
On-line bleaching control at
sues of the sparse simplex method and tween LCP and matrix game. We show
pulp mills
interior point method for the solution that an almost matrix with value greater PATRIK F LISBERG
of large QPs. We also consider how than zero is a Q matrix. Certain new Division of Optimization, Linkoping
the dual SSX is adopted to solve QMIP sub-classes of (introduced by Stone) are University
models. We provide computational ex- also studied using the game theoretic ap- coauthor: Mikael Rönnqvist
perience of processing index tracking proach. The classes of matrices studied keywords: nonlinear optimization,
models, with 22 factors and a universe here are also important from an algorith- process control, industrial application,
of 4752 stocks leading to models with mic point of view. on-line optimization
 93

We present an on-line control system for Preprocessing sparse semidefi- the global convergence of the proximal
the bleaching process of pulp mills. It nite programs by the conversion method for complete Riemannian mani-
supplies target values for the chemical method folds, with non positive curvature. An
charges in order to minimize the cost important example of this kind of space
of used bleaching chemicals. Approxi- M ITUHIRO F UKUDA is the positive definite matrix set. We ap-
mate functions to describe each stage of New York University, Courant Institute ply that method to create a hybrid bundle
the process are established dynamically of Mathematical Sciences algorithm in semidefinite programming.
by solving constrained least square prob- coauthors: Katsuki Fujisawa, Kazuhide Some constructions are made, in order to
lems. Then an overall nonlinear opti- Nakata get a computable algorithm.
mization problem is solved to generate keywords: semidefinite programming,
the target values. This problem makes chordal graph, preprocessing, matrix
use of the approximate functions together completion Robust search directions for
with linking constraints between several We present a method to convert certain large sparse semidefinite pro-
process stages. The system is imple- sparse semidefinite programs (SDPs) into gramming (SDP)
mented at a Swedish pulp mill and shows equivalent SDPs with small block ma-
a 10 percent decrease of the mill’s cost of trices which can be solved more effi-
bleaching chemicals. ciently by any primal-dual interior-point H ENRY W OLKOWICZ
method code. The conversion method University of Waterloo
New paging results using the rel- can be viewed as a preprocessing of the keywords: large sparse systems,
ative worst order ratio SDP data. The method is based on the optimization, algorithms
matrix completion theory, and it directly
J OAN B OYAR manipulates the sparsity structure repre-
University of Southern Denmark Current paradigms for search directions
sented by a graph of a given SDP. The
for primal-dual interior-point methods
coauthors: Lene Favrholdt, Kim S. conversion of the problem is performed
for SDP use: (i) symmetrizethe lin-
Larsen on this graph exploiting basic properties
earization of the optimality conditions
keywords: on-line algorithms, relative of chordal graphs and clique trees. A
at the current estimate; (ii) form and
worst order ratio, paging, look-ahead heuristic procedure to estimate the com-
solve the Schur complement equation
putational cost of the converted problem
We consider a measure for the quality of for the dual variable dy; (iii) back
is employed in this method.
on-line algorithms, the relative worst or- solve to complete the search direction.
der ratio, which was recently proposed by These steps result in loss of sparsity and
Boyar and Favrholdt and shown to pro- New scaling algorithms and a ill-conditioning/instability, in particular
vide new and easier separations for on- hybrid bundle method for SDP when one takes long steps and gets close
line algorithms for two variants of bin programming to the boundary of the positive semidef-
packing. This new ratio is used to com- PAULO ROBERTO O LIVEIRA inite cone. This has resulted in the ex-
pare on-line algorithms directly by taking COPPE/Federal University of Rio de clusive use of direct, rather than iterative
the ratio of their performances on their Janeiro methods, for the linear system.
respective worst orderings of a worst- We look at alternative paradigms
case sequence. Here we consider the coauthor: Ronaldo Gregorio based on least squares, an inexact Gauss-
paging problem, where a processor has keywords: nonlinear programming, Newton approach, and a matrix-free pre-
a cache of size k, pages requested must semidefinite programming, proximal conditioned conjugate gradient method.
be in the cache to be processed, and methods, non smooth optimization This avoids the ill-conditioning in the
there is unit cost for bringing a page This talk has two parts, both motivated nondegenerate case. We emphasize ex-
into the cache. The goal is to minimize by some applications of Riemannian ge- ploiting structure in large sparse prob-
the total cost of a sequence of page re- ometry tools. In the first, we present lems. In particular, we look at LP
quests. Using standard competitive anal- a new class of scaling metrics, defined and SDP relaxations of the: Max-Cut;
ysis, Least-Recently-Used (LRU) is an in the Rn hypercube. For that class, Quadratic Assignment; Theta function;
optimal deterministic algorithm for the we show that the corresponding Riemann and Nearest Correlation Matrix prob-
problem, obtaining exactly the same ra- manifold is complete, and has null sec- lems.
tio, k, as many other algorithms. We tional curvature. Those metrics, when
propose a new algorithm which is better applied to continuous optimization, in
than LRU according to the relative worst proper settlings, lead to projective (gradi- WE1-341/22 LOGISTICS AND TRANSPORTATION

order ratio. In addition, using standard ent) or sub-gradient geodesic algorithms, Practical routing problems
competitive analysis, look-ahead, allow- in a metric dependent scheme. As a con-
ing the algorithm to see the next s page sequence, we can apply some previous chair : K. Ganesh
requests before deciding which page to results, where we ensured global con- Decision support for real-time
evict, cannot produce better algorithms. vergence in the convex case, for com- ambulance planning and con-
In contrast, we show that with the rela- plete Riemannian manifolds, with non- trol
tive worst order ratio, look-ahead helps negative curvature. Besides, in the pro-
significantly. jective case, we have also the linear rate
of convergence for strongly convex func- TOBIAS A NDERSSON
tions. As an outcome, we assure the Dep of Science and Technology /
WE1-341/21 INTERIOR POINT ALGORITHMS

linear rate of convergence for the usual University of Linkoping

Methods in semidefinite gradient method, with line search. The
programming second part deals with the application coauthor: Peter Värbrand
chair : Henry Wolkowicz of a previous result, where we proved keywords: ambulance service,
transportation, decision support systems
94 Parallel Sessions: Wednesday 9:00 – 10:30 (WE1)

SOS Alarm AB is the company control- keywords: vehicle routing, evolutionary time periods. These results allow us to
ling all ambulance movements in Swe- algorithms, contraceptive logistics, generate practical rules of thumb that can
den. In case of an emergency, an ambu- supply chain be immediately implemented in current
lance controller working in an SOS cen- In today’s era of globalization and cut- planning technologies.
tral quickly has to decide which unit (or throat competition, customer satisfaction
units) to dispatch. This allocation of an has become an issue of prime concern Optimization techniques in
ambulance to the emergency may how- and that is where the role of supply chain, seizure prediction
ever affect the ability to handle future which link together the suppliers, manu-
emergencies, since some area might be facturers, distribution channels and cus- WANPRACHA
left uncovered as a result. This makes tomers, as one single organization of C HAOVALITWONGSE
it necessary to relocate other available pooled resources and skills. Logistics Department of Industrial and Systems
units for better coverage of the consid- costs constitute a major portion of the to- Engineering
ered area. Ambulances are not only used tal costs in a supply chain. Transporta- coauthors: Panos M. Pardalos, J. Chris
for emergencies, but also for planned, not tion plays an important role in logistics Sackellares, Leonidas D. Iasemidis, D.-
urgent, patient transports. In these cases and another element of logistics system S. Shiau, Paul R. Carney
the objective is to execute the assignment is the allocation and routing of vehicles keywords: quadratic programming,
as close to the planned time as possible, for the purpose of collection and delivery EEG time series, seizure prediction,
and not like in the urgent cases where it is of goods and services on regular basis. chaos theory
crucial to get an ambulance to the patient The system may involve a single depot There is growing evidence that temporal
as quickly as possible. The situation may or multiple depots; the objectives may be lobe seizures are preceded by a preictal
be further complicated by restrictions on aimed at cost minimization; time mini- transition, characterized by a gradual dy-
what kind of assignments an ambulance mization or distance minimization. This namical change in electroencephalogram
unit can serve, the working schedule of problem involves the design of several (EEG) time series from asymptomatic in-
the ambulance crew and so on. Here, a vehicle tours to meet a given set of re- terictal state to seizure. We present a
mathematical model and a solution algo- quirements for customers with known lo- new methodology of studying the spa-
rithm providing decision support for the cations, subjected to several constraints. tiotemporal patterns in EEG by employ-
ambulance controllers are presented. The main focus of this paper is to opti- ing chaos theory and optimization tech-
mize the vehicle routing for contracep- niques, specifically multi-quadratic inte-
Simultaneous optimization of tive logistics (Family Planning Logistics) ger programming. The application of this
school starting times and public and hence solving the problem using evo- methodology is to select brain electrode
bus services lutionary algorithms like Genetic Algo- sites that exhibit the dynamical conver-
rithms (GA) and Simulated Annealing gence in the measure of chaos, which can
A RMIN F UEGENSCHUH (SA) and meet the objective of vehicle be identified by optimization techniques,
Darmstadt University of Technology routing problem. for the prediction of epileptic seizures.
coauthor: Alexander Martin The algorithm was tested in continuous,
keywords: transportation, vehicle WE1-341/23 BIOINFORMATICS AND OPTIMIZATION
long-term EEG recordings obtained from
scheduling, time windows Optimization in medicine II 5 patients with temporal lobe epilepsy.
The results suggest that it may be pos-
In many rural areas, the public bus ser- organizer/chair : Panos M. Pardalos sible to develop automated seizure warn-
vice is demand-oriented: By far the ing devices for diagnostic and therapeutic
biggest group of customers are pupils
Fractionation in radiation treat-
ment planning purposes.
who are transported to their schools
within certain strict time limits. Usually, M ICHAEL F ERRIS
all schools start around the same time,
A column generation approach
University of Wisconsin to aperture modulation in radia-
which causes a morning peak in the num-
ber of deployed buses. However, schools
coauthor: Meta M. Voelker tion therapy treatment planning
keywords: radiation therapy, stochastic
are allowed to change their starting times programming, treatment planning E DWIN ROMEIJN
within some intervall. The question is, University of Florida, Department of
how to simultanenously rectify the start- In many cases a radiotherapy treatment Industrial and Systems Engineering
ing times for all schools and bus trips in is delivered as a series of small dosages
over a period of time. New generations coauthors: Ravindra K. Ahuja, James F.
a certain county so that the number of Dempsey, Arvind Kumar, Jonathan G. Li
scheduled buses is minimal. We present of machines generate accurate informa-
tion of the actual dose delivered, allowing keywords: health care - treatment,
a mixed-integer programming formula- radiation therapy, column generation
tion for this optimization problem and a planner to compensate for errors in de-
address its solution for some real-world livery in an adaptive fashion. We formu- We consider the problem of radiation
instances. late a model of the day-to-day planning therapy treatment planning for cancer pa-
problem as a stochastic linear program tients. During radiation therapy, beams
and exhibit the gains that can be achieved of radiation pass through a patient, killing
Optimization of vehicle routing by incorporating uncertainty about errors both cancerous and normal cells. Thus,
using evolutionary algorithms during treatment into the planning pro- the radiation treatment must be care-
for contraceptive logistics in a cess. Due to size and time restrictions, fully planned so that a clinically pre-
supply chain the model becomes intractable for real- scribed dose is delivered to targets con-
istic instances. We show how neuro- taining cancerous cells, while nearby or-
K. G ANESH dynamic programming can be used to ap- gans are spared. Using a modern tech-
Indian Institute of Technology Madras proximate the stochastic solution, and de- nique called intensity modulated radia-
coauthor: T. T. Narendran rive results from our models for realistic tion therapy (IMRT), each of the beams
 95

is decomposed into many small beamlets, YOSHIKO I KEBE definition, description of the solution ap-
the intensities of which can be controlled Dept. of Management Science, Tokyo proach, and some simulation and compu-
individually. When all beamlet intensi- Univ. of Science tational results.
ties (together called the fluence map) are keywords: sports timetableing, round
given, this fluence map is, for delivery robin tournament, graphs, list coloring WE1-321/033 MULTICRITERIA OPTIMIZATION

purposes, decomposed into (and approx- Efficiency analysis in

imated by) a limited number of apertures In sports timetabling, a frequently used
approach for constructing schedules for multicriteria optimization III
for each beam, each with its associated
intensity. We propose a new formulation round robin tournaments is to (i) fix chair : Henri Bonnel
of the fluence map optimization problem travel patterns (Home-Away patterns) for Inferring efficient weights from
that integrates the two phases. To over- all teams, then (ii) enumerate schedules pairwise comparison matrices
come the problem that the number of al- which conform to the fixed pattern set.
lowed apertures is enourmous, we pro- Thus, the problem of deciding whether E MILIO C ARRIZOSA
pose a column generation approach. We a given pattern set is feasible or not university of seville
study different pricing problems, each is significant, and there have been nu- coauthors: Rafael Blanquero, Eduardo
corresponding to a particular choice for merous studies devoted to this problem. Conde
the set of feasible apertures, and pro- Of special interset have been the pattern keywords: AHP, vector optimization,
vide polynomial time algorithms. Exper- sets which have a minimum number of eigenvector method
iments show that this approach can re- breaks. In this talk we discuss this prob-
lem, and show some results for the case Several Multi-Criteria-Decision-Making
duce the number of apertures required for methodologies assume the existence of
treatment without a clinically significant when the number of teams is a multiple
of four. We also explore relations to the weights associated with the different cri-
reduction in treatment plan quality. teria, reflecting their relative importance.
list coloring problem for graphs.
One of the most popular ways to in-
WE1-321/053 PRODUCTION AND SCHEDULING
fer such weights is the Analytic Hierar-
Real-life scheduling Airline schedule recovery - mod- chy Process, which constructs first a ma-
chair : Matthew Berge
els and algorithms for air traffic trix of pairwise comparisons, from which
management concept analysis weights are derived following one out of
Characterizing feasible pattern many existing procedures, such as the
sets with a minimum number of M ATTHEW B ERGE eigenvector method or the least (loga-
breaks Boeing, Phantom Works rithmic) squares. Since different proce-
keywords: airline scheduling, air traffic dures yield different results (weights) we
RYUHEI M IYASHIRO management, Lagrangian relaxation, pose the problem of describing the set of
University of Tokyo dynamic programming weights obtained by "sensible" methods:
coauthor: Tomomi Matsui those which are efficient for the (vector-
keywords: sports scheduling, integer This paper presents a modeling method- ) optimization problem of simultaneous
programming ology to assess a range of operational minimization of discrepancies.
concepts for collaborative air traffic flow A characterization of the set of effi-
Creating a round-robin schedule with management. The focus is on the prob-
home/away assignment is a significant cient solutions is given, which enables
lem of airline schedule recovery in con- us to assert that the least-logarithmic-
problem of sports scheduling. When ditions where both airports and airspace
tournament organizers construct such squares solution is always efficient,
sectors are capacity limited due to con- whereas the (widely used) eigenvector
schedules, they often create home/away ditions such as weather events or system
assignment first, i.e. fix a pattern set, solution is not, in some cases, efficient,
outages. This model is embedded in a thus its use in practice may be question-
then assign an opponent to the pattern dynamic simulation environment repre-
set. However, not all pattern sets can able.
senting the US National Airspace Sys-
be completed into a schedule; valid pat- tem. The airline schedule recovery model
tern sets called feasible pattern sets. Al- is based on an optimization formulation Efficiency and generalized con-
though finding feasible pattern sets is that allows a representation of adaptive cavity in stochastic multiobjec-
at the heart of many scheduling algo- airline behavior in current and future op- tive programming
rithms, good characterization of feasible erations. The recovery options consid-
pattern sets is not known yet. We con- S TEFAN T IGAN
ered include ground delay, re-routing and University of Medicine and Pharmacy,
sider the feasibility of pattern sets and flight cancellation. Constraints include
propose a new necessary condition for Cluj-Napoca
airport arrival and departure rates, sec-
feasible pattern sets. For a special class tor occupancy limits, and aircraft flow. coauthor: I.m. Stancu-Minasian
of pattern sets, called pattern sets with The problem is formulated as a large- keywords: stochastic programming,
a minimum number of breaks, we de- scale integer linear program and solved multiobjective programming,
veloped a polynomial-time algorithm to using an efficient Lagrangian Relaxation minimum-risk problem,
check whether a given pattern set satisfies based approximation method. This relax- pseudomonotonic programming
the necessary condition. Computational ation method enables the dual problem The purpose of this paper is to derive
experiment shows that, when the number to be decomposed into airplane itinerary relations among several efficiency con-
of teams is up to 26, the proposed con- subproblems which can be solved by dy- cepts (local and global weakly efficient,
dition characterizes feasible pattern sets namic programming. Efficiency is par- efficient and properly efficient solutions)
with a minimum number of breaks. ticularly important because this recovery to some multiobjective minimum-risk
problem must be solved many times in problems with given levels of minimum
Feasibility problems in sports the context of a single simulation exper- satisfaction and stochastic programming
scheduling iment. The paper presents a problem problems of Kataoka type with given
96 Parallel Sessions: Wednesday 9:00 – 10:30 (WE1)

probabilities of satisfaction. For the Leontief matrix structure. We prove that, of crossover events introduced by Vava-
multiobjective linear-fractional stochas- when the discount factor θ is strictly sis and Ye. For a standard form lin-
tic programming problem with given less than 1, the  problem
 can be solved ear program min cT x : Ax  b x
0

probabilities of satisfaction, in the case of in at most O n1 5 log 1  1 θ  log n  clas-

with decision variable x  ℜn , we show
normal distribution of the random vectors that the MTY P-C algorithm started
 interior-point
sical  method iterations and
in the objectives, we obtain a local-global O n4 log 1  1 θ  logn   arithmetic oper-from a well-centered interior-feasible so-
property. We also derive a result pro- ations. Our method is a combinato- lution with duality gap nµ0 finds an
viding sufficient conditions on the distri- rial interior-point method related to the  2 with
interior-feasible solution  duality
 gap
bution functions of the objectives in or- work of Ye and Vavasis and Ye. To 
less than nη in O n log log µ 0 η   
der to be equal the sets of the properly my knowledge, this is the first strongly n3 5 log chistarA  n   iterations, where
efficient solutions for the multiobjective polynomial-time algorithm for solving chistarA is a scaling invariant condition
minimum-risk problem with given lev- the MDP when the discount factor is a number associated with the matrix A.
els of minimum satisfaction and multi- constant less than 1. More specifically, chistarA is the infi-
objective stochastic programming prob- mum of all the conditions numbers ¯χAD ,
lem with given probabilities of satisfac- where D varies over the set of positive
tion. Finaly, via multi-criteria pseudo- Polynomial-time computation of diagonal matrices. Under the setting of
monotonic programming, we give suffi- optimal policies in Markov deci- the Turing machine model, our analysis

cient conditions of generalized concavity sion chains yields an O n3 5 LA  n2 logL   iteration-
type in order to the minimum-risk linear complexity bound for the MTY P-C al-
fractional problem has the property that M ICHAEL O’S ULLIVAN gorithm to find a primal-dual optimal so-
any its efficient solution is a properly ef- University of Auckland lution, where LA and L are the input sizes
ficient solution too. coauthor: Arthur F. Veinott Jr. of the matrix A and the data A b c  , re-
keywords: dynamic programming, spectively. This contrasts well with the
Semivectorial bilevel optimiza- stochastic, infinite horizon classical iteration-complexity bound for
tion problem This paper shows that the problem of the MTY P-C algorithm which depends
finding an n-optimal (resp., n-present- linearly on L instead of logL.
H ENRI B ONNEL value optimal, n-Cesàro-overtaking-
University of New Caledonia
optimal) policy for a finite-state-and- WE1-305/205
GAME THEORY

coauthor: Jacqueline Morgan action infinite-horizon stationary Markov Optimization in electricity

keywords: vector optimization, bilevel decision chain is solvable in polynomial markets
optimization, multiobjective time. (Where n is the number of states, organizer/chair : Andy Philpott
optimization this is equivalent to finding a policy that
We consider a bilevel optimization prob- has maximum present value for all suf- Estimation of market distribu-
lem, where the lower level is given by a ficiently small positive interest rates.) tion functions in electricity pool
vector (in particular : multicriteria) opti- This is done by decomposing the prob- markets
mization problem. The upper level is a lem into a sequence of at most 3n + G OLBON Z AKERI
real valued function. The bilevel prob- 5 subproblems. Each subproblem is University of Auckland
lem contains as a particular case the well either a maximum-transient-value or a
known problem of optimizing a scalar maximum-reward-rate problem, and can coauthors: Geoffrey Pritchard, Andy
function over an efficient (Pareto) set. Of be solved by linear programming or pol- Philpott
course, our problem generalizes also the icy improvement. Moreover, for the case keywords: electricity market, market
usual (salar) bilevel optimization prob- of unique transition systems, i.e., each distribution function, non parametric
lem. We present some optimality condi- action in a state sends the system to at estimation
tions and a penalty approach. most one state, the problem is shown to The market distribution function is a
be solvable in strongly polynomial time. probabilistic device that can be used to
The last case includes standard determin- model the randomness in dispatch and
WE1-321/133 LINEAR PROGRAMMING

istic dynamic programs. clearing price that generators in electric-

Linear programming and ity pool markets must take account of
Markov decision when submitting offers. We discuss tech-
A new iteration-complexity niques for estimating the market distri-
organizer/chair : Yinyu Ye bound for the MTY predictor- bution function, and ways of measuring
A new complexity result on solv- corrector algorithm the quality of these estimators, using both
ing the Markov decision prob- classical statistical approaches and in the
lem TAKASHI T SUCHIYA context of optimization.
The Institute of Statistical Mathematics
Y INYU Y E coauthor: Renato Monteiro Optimization models for hydro-
Stanford University keywords: interior point algorithms, reservoir operations in electric-
keywords: linear programming, Markov primal-dual algorithms, condition ity pools
decision problem, computational number, crossover events
complexity
A NDY P HILPOTT
In this talk, we present a new iteration- University of Auckland
We present a new complexity result on complexity bound for the Mizuno- coauthors: Geoffrey Pritchard, Philip
solving the Markov decision problem Todd-Ye predictor-corrector (MTY P- Neame
(MDP) with n states and a number of C) primal-dual interior-point algorithm keywords: electricity market,
actions for each state, a special class for linear programming. Our analy- hydroelectric reservoirs, dynamic
of real-number linear programs with the sis is based on the important notion programming, equilibria
 97

In electricity pool markets, generators of- Learning collusive strategies in optimal for a one stage game. In this pa-
fer supply functions in each trading pe- electricity markets per we use a coevolutionary genetic al-
riod to an independent system operator gorithm framework to show how agents
who dispatches these to minimize the E DWARD A NDERSON with limited memory can learn to play
cost of meeting an observed load. The Autralian Graduate School of collusive strategies. The collusive strate-
optimal supply function for each hydro Management gies which emerge can have a rich struc-
generator must take account of the oppor- coauthor: Thai Doan Hoang Cau ture (for example they are more complex
tunity cost of using water for generation keywords: collusion, electricity market, than the forgiving trigger strategies that
now as compared with releasing it later. genetic algorithm, co-evolution occur in the literature). Besides explor-
To compute this opportunity cost we con- ing the learning behaviour we also use
sider a dynamic programming model in Wholesale electricity markets operate as this framework to investigate the mar-
which generators optimize with stochas- repeated games between relatively small ket characteristics which make collusion
tic electricity prices, and compare this numbers of participants in which we can more likely to occur. For example we ask
with models in which generators offer expect to see implicit collusion emerg- whether collusion is more likely when
supply functions as participants in a dy- ing. This may in practice amount to no market participants are similar or when
namic game. more than market participants being less they are different.
aggressive in their bidding than would be
98 Parallel Sessions: Wednesday 13:15 – 14:45 (WE2)

WE2-302/41 COMBINATORIAL OPTIMIZATION coauthors: Anuj Mehrotra, Hakan problem may be defined in general tables
Yildiz in which cells can take any real value and
Computational methods for keywords: graph-coloring, graph in nonnegative tables where all cell val-
graph-coloring I theory, heuristics ues are nonnegative. We present com-
organizer/chair : Michael Trick pact formulations for the exact disclosure
Column generation and branch-and-price
is an effective approach for exact graph problem in general and nonnegative ta-
A systematic approach to gener- bles and use a lagrangean approach to ob-
ate very hard 3COL instances coloring. We extend this work to color-
ing generalizations such as multicoloring tain near optimal solutions and an upper
K AZUNORI M IZUNO and bandwidth problems. bound on their deviations from the opti-
Institute of Information Sciences and mum. Computational results will be pre-
Electronics, University of Tsukuba sented to show the effectiveness of this
WE2-302/42 COMBINATORIAL OPTIMIZATION
new approach.
coauthor: Seiichi Nishihara Combinatorial optimization
keywords: computational complexity, VI
graph-coloring, phase transition, Perfect sampling algorithm for
NP-complete chair : Tomomi Matsyu two-rowed contingency tables
Graph colorability, COL, and proposi- The distribution of values in the TOMOMI M ATSYU
tional satisfiability, SAT, are typical con- quadratic assignment problem University of Tokyo
straint satisfaction problems to which
phase transition phenomena, PTs, are im- TAMON S TEPHEN coauthor: Shuji Kijima
portant in the computational complexity Institute for Mathematics and its keywords: sampling, Markov chain,
of combinatorial search algorithms. Sec- Applications contingency table, counting problem
ondary PT, rather than primary PT, is sig- coauthor: Alexander Barvinok This paper proposes a polynomial time
nificant and subtle because it claims that, keywords: combinatorial optimization, perfect (exact) sampling algorithm for
in the secondary PT region, extraordi- quadratic assignment problem 2  n contingency tables. Our algorithm
narily hard instances are found that may We survey some results on the distri- is a Las Vegas type randomized algo-
require exponential-order computational bution of objective values in the objec- 
rithm and the expected running time is
time to solve. To clarify the PT mecha- tive function of the Quadratic Assign- bounded by O n3 ln N  where n is the
nism, many studies have been undertaken ment Problem (QAP). We are interested number of columns and N is the total
to produce very hard instances, many of in how many permutations σ produce a sum of whole entries in a table. The al-
which were based on generate-and-test value f σ  within
 a given range of the
gorithm is based on monotone coupling
approaches. We propose a rather sys- optimal value f τ  and how far such per- from the past (monotone CFTP) algo-
tematic or constructive algorithm that re- mutations σ are from τ in the Hamming rithm and new Markov chain for sam-
peats the embedding of 4-critical graphs distance on the symmetric group. pling two-rowed contingency tables uni-
to arbitrarily generate large extraordinar- formly. We employed the path cou-
ily hard 3-colorability (3COL) instances.
Exact disclosure prevention in
pling method and showed that the mix- 
We demonstrate experimentally that the ing  rate of  our chain is bounded by n n
computational cost of our 3COL in- general and nonnegative statis- 1  2 1  ln nN   . Our result shows that
stances is of an exponential order by us- tical tables uniform generation of two-rowed con-
ing a few actual coloring algorithms. F ILIPA D UARTE DE C ARVALHO tingency tables is easier than the cor-
Instituto Superior de Economia e responding counting problem, since the
Graph-coloring in the estima- Gestão/UTL counting problem is known to be #P–
tion of mathematical derivatives coauthor: M. T. Almeida complete.
keywords: combinatorics, integer
S HAHADAT H OSSAIN programming, relaxation/subgradient, WE2-302/44 COMBINATORIAL OPTIMIZATION

University of Lethbridge
applications Networks and communication
coauthor: Trond Steihaug
keywords: sparse jacobians, patterns Statistical offices release information in chair : Fulvio Piccinonno
two-dimensional tables. To prevent dis-
We describe graph coloring instances closure of confidential data, these offices Nonlinear optimization in gas
arising in numerical estimation of math- often use a technique called cell suppres- network operation
ematical derivatives. The coloring in- sion. As the name suggests, this tech- M ARC C. S TEINBACH
stances are obtained from a matrix par- nique consists of suppressing from the ta- Zuse Institute Berlin
titioning problem. The size of the gener- bles the confidential values, before the ta-
ated graphs is dependent on the number bles are released. Suppressing only con- coauthor: Klaus Ehrhardt
of segmented columns which can be var- fidential data cells from the tables may keywords: nonlinear programming,
ied between the number of columns and not prevent their values from being dis- network operation, natural gas
the number of nonzeros in the associated closed, due to the linear relationships that A routine task in providing natural gas
matrix. can be deduced from table row and col- is the operative planning of the load dis-
umn subtotals. Some non-confidential tribution in the pipeline network to sat-
Column generation approaches data cells must then be suppressed as isfy the current demand. Variable oper-
for graph-coloring generaliza- well. The selection of a set of non- ating costs are dominated by the energy
tions confidential data cells to be suppressed, for the gas transport; thus minimizing the
in order to prevent exact disclosure of the cost requires sophisticated control strate-
M ICHAEL T RICK confidential data at minimum cost, is a gies for the technical equipment. Mathe-
Carnegie Mellon NP-hard problem. The exact disclosure matically this leads to highly dimensional
 99

mixed-integer optimization problems in- codes and turbo-codes, obtained usu- consists of finding k 1 tours (starting
volving various physical, technical, and ally with the concatenation of two or and ending at the depot node) such that
contractual restrictions. more convolutional codes, are the most each edge is traversed by at least one
Due to the enormous complexity of used classes of codes when good perfor- tour and the length of the longest tour is
the overall problem we pursue a hier- mances at low signal-to-noise ratios are minimized.
archical solution approach. Binary de- needed. The turbo-codes decoding proc-
We will present a Branch-And-Cut
cisions are determined by mixed inte- cess is based on soft-decoding compo-
approach which is based on a sparse IP-
ger linear programming using approxi- nent convolutional codes and needs both
formulation of the MM k-CPP. Besides
mations of the nonlinearities; continu- a lot of computing time and memory. We
the usual parity and connectivity con-
ous control variables are then optimized will propose a new family of algorithms,
straints we use a new class of constraints
using a nonlinear control problem with derived form MAP algorithm, that with
which is similar to the capacity con-
fixed combinatorial decisions. The lec- a computational complexity of the same
straints used in the context of the Capaci-
ture concentrates on the modeling and order and much reduced memory require-
tated Arc Routing Problem (CARP) and
numerical solution of the latter problem, ments will obtain identical results.
the Capacitated Vehicle Routing Prob-
addressing in particular the hyperbolic
lem (CVRP), respectively. We will report
gas flow equations with underlying net- WE2-302/45 COMBINATORIAL OPTIMIZATION

computational results on an extensive set

work topology and the nonlinear behav- Postman optimization of instances from literature as well as self
ior of compressors. Computational opti- problems generated instances.
mization results will finally be presented.
chair : Julián Arturo Aráoz
New integer linear program- Restricted postman problems on
ming formulations for the SRAP mixed graphs
Privatized rural postman prob-
N ELSON M ACULAN F RANCISCO JAVIER Z ARAGOZA lems
Politecnico di Milano M ARTINEZ
coauthors: Elder Macambira, Cid C. de Department of Combinatorics and
Souza Optimization, University of Waterloo
keywords: integer formulation, J ULIÁN A RTURO A RÁOZ
coauthor: Bill Cunningham Universidad Politecnica de Cataluña
telecommunication, node partitioning keywords: postman problems, mixed
In this work, we consider the problem of graphs, combinatorial optimization, coauthors: Elena Fernández, Cristina
assigning customer sites with given de- approximation algorithms Zoltan
mands to SONET rings of equal capac- Given a strongly connected mixed graph keywords: postman problems, graphs,
ity. The so-called SONET Ring Assign- M  V E A  , with nonnegative costs c combinatorial optimization, integer
ment Problem (SRAP) is known to be on its edges E and arcs A, the mixed programming
NP-hard and can be described formally postman problem is to find a tour T of
as a node-partitioning problem. Each site M, traversing all its vertices, edges and
is assigned to exactly one ring and a spe- arcs at least once, and with minimum cost
In this work we analyze the Privatized
cial ring, called federal ring, intercon- c T  . There exists
 a 3
2 -approximation
Rural Postman Problem which is the
nects the others rings together. The ob- algorithm for c T  due to Raghavachari
edge version of the Traveling Salesman
jective is to minimize the number of rings and Veerasamy [IPCO 1998]. We are
Problems With Profits. Given an undi-
in the network.

interested  in approximating
 the measure
rected graph G(V,E) with a distinguished
We propose two new integer linear r T  c T  c E A  , that is, the cost
vertex d, called the Depot, and two non

programming formulations for the SRAP. of the repeated edges and arcs. In this
negative real functions from the edge set
Computational results for both formula- talk, we introduce a generalization of the
E, the profit function b and the cost func-
tions are discussed. Comparisons with mixed postman problem where we re-
tion c, such that every time an edge e
other formulations presented earlier in quire that the tours traverse a given subset
is traversed a cost c(e) is incurred, but
the literature are made. Empirical evi- R  E A exactly once. In particular, we
the profit b(e) is collected only the first

dence indicates that our formulations per- discuss the special cases of this restricted
time the edge is traversed. The Privatized
form much better than the existing ones. mixed postman problem when R  E and
Rural Postman Problem is to find a cy-
While previous models cannot be solved R  A, and relate  our results to the ap-
cle C, passing through d and not neces-
to optimality in a reasonable amount of proximation of r T  . sarily simple, which maximizes the value
time, our models are computed much
of the sum of the profit of the traversed
faster and optimal solutions are obtained
A branch-and-cut approach for edges (counting only one time the profit)
in minutes in a typical desktop machine.
the min-max k-Chinese postman minus the sum of the cost of the tra-
versed edges, but this time counting ev-
A new map-like algorithm with problem
 the edge
ery time  is  traversed.
 That is,
low memory requirements for D INO A HR maxC ∑e  C b e  t e  c e  
where t(e)
decoding error correcting codes Institute of Computer Science, is the number of times that edge e is tra-
F ULVIO P ICCINONNO University of Heidelberg verse in C. We consider the properties of
Università di Venezia coauthor: Gerhard Reinelt the problem, its relation with known and
keywords: computational complexity, keywords: Chinese postman problem, new problems, special cases with good
coding, turbo codes branch-and-cut algorithms, the polyhedral structure and
the possibility of the use of Branch and
In this work we propose a new MAP- Given an undirected edge-weighted
Cut methods and some Heuristics for the
like algorithm for decoding error cor- graph and a depot node, the Min-Max k-
problem. We finish with remarks about
recting codes. Actually convolutional Chinese Postman Problem (MM k-CPP)
future research.
100 Parallel Sessions: Wednesday 13:15 – 14:45 (WE2)

WE2-302/49 INTEGER AND MIXED INTEGER

PROGRAMMING
a hybrid Relax and Cut - Linear Pro- The semimetric polytope associated to a
gramming Relaxation exact solution al- complete undirected graph G is an impor-
Decomposition algorithms and gorithm is proposed. Such an algorithm tant structure in combinatorial optimiza-
dynamic cut generation has proved capable of solving NCRPP in- tion, and it is related to several relevant
organizer/chair : Ted Ralphs stances more than twice as large as those optimization problems such as the max-
found in the literature. cut problem. A family of facet-defining
Decomposition and dynamic cut inequalities for the semimetric polytope,
generation in integer program- the so-called triangle inequalities, of-
ming Solving lexicographic multiob-
jective MIPs with branch-cut- ten provide a quite tight approximation
M ATTHEW G ALATI price of the polytope, which results in strong
Lehigh University bounds on the related optimization prob-
L ASZLO L ADANYI lems. Unfortunately, the corresponding
coauthor: Ted Ralphs IBM Research linear program, which has a number of
keywords: integer programming, coauthors: Marta Eso, David L. Jensen variables and constraints that are respec-
Dantzig-Wolfe, cutting planes, keywords: mixed integer programming, tively quadratic and cubic on the num-
relaxation branch-and-cut-and-price, ber of nodes of G, is usually very de-
Decomposition algorithms such as La- Dantzig-Wolfe decomposition, generate and almost impossible to solve
grangian relaxation and Dantzig-Wolfe lexicographic multiobjective efficiently, for large graphs, even when
decomposition are well-known methods In this talk we will first describe an ap- using state-of-the-art linear programming
that can be used to develop bounds for in- plication, the FCC Auction 31, and its software and constraint-generation tech-
teger programming problems. We draw two formulation: one with a compact rep- niques. We show how to interpret the
connections between these classical ap- resentation solvable via branch-and-cut, relaxation corresponding to the polytope
proaches and techniques based on gener- and one with column generation using defined by all the triangle inequalities as-
ating strong valid inequalities. We also branch-and-price. We will show how the sociated to one fixed node as a classi-
discuss several methods for integrating second formulation can be interpreted as cal transshipment problem, and how to
dynamic cut generation with decomposi- a Dantzig-Wolfe decomposition and that efficiently solve this relaxation, called
tion and present a decomposition-based this interpretation leads to a branch-cut- rooted semimetric polytope. Building on
separation algorithm called decompose price algorithm. Then we will demon- this, we construct Lagrangean-based ap-
and cut. The algorithm takes advantage strate how to apply the same principle proaches for maximizing a linear func-
of the fact that separation of an integer to a general Mixed Integer Programming tion over the semimetric polytope, i.e.,
solution is often much easier than sepa- problem. Finally, we extend the applica- by considering all the triagle inequali-
ration of an arbitrary fractional solution. bility of linear complementarity to inte- ties. We report extensive computational
ger programming to solve problems with results on several classes of graphs show-
Optimal rectangular partitions multiple (lexicographic) objective. ing that, in general, the proposed ap-
proaches clearly outperform ordinary lin-
A BILIO L UCENA WE2-306/31 INTEGER AND MIXED INTEGER
PROGRAMMING
ear programming techniques.
Universidade Federal do Rio de Janeiro Computational methods for
coauthors: Felipe Calheiros, Cid C. de solving IPs I Mixing two sets of mixed integer
Souza organizer/chair : Eva Lee inequalities
keywords: rectangular partition, relax
and cut, Lagrangian relaxation, cutting Generating cutting planes using M ATHIEU VAN V YVE
planes master cyclic group polyehdra CORE - Universite Catholique de
Louvain
Assume that a rectangle R is given on the L ISA E VANS coauthors: Andrew Miller, Miguel
Euclidean plane together with a finite set University of Minnesota Constantino
P of points that are interior to R. A rect- coauthor: Ellis Johnson keywords: mixed integer programming,
angular partition of R is a partition of the keywords: integer programming, mixed integer rounding, convex hull
surface of R into smaller rectangles. The cutting planes
We will review the master cyclic group following mixed-integer set X    s z  
length of such a partition equals the sum We study the polyhedral structure of the
of the lengths for the line segments that polyhedron, and discuss how its facets I1  I2  s  z
b i  I s  Cz

define it. The partition is said to be feasi- can be used to generate cutting p lanes R   Z i i 1 i
ble if no point of P is interior to a partition for general integer programming prob- b i i  I2
, where C is a positive integer.
rectangle. The Rectangular Partitioning lems. Some practical implementation is- The set X naturally appears as a sub-
Problem (RPP) seeks a feasible rectan- sues will be discussed, and co mputa- model of production planning problems
gular partition of R with the least length. tional results will be given. involving variable lower bounds on pro-
Computational evidence from the litera- duction (i.e. the production must be ei-
ture indicates that RPPs with noncorecti- ther zero or higher than some value L) or
linear points in P, denoted NCRPPs, are Lagrangean approaches for op- start-ups.
timizing over the semimetric
the hardest to solve to proven optimality.
polytope We characterize the extreme  points
In this paper, some structural properties and extreme rays of conv X  . We pro-
of optimal feasible NCRPP partitions are A NDREA L ODI vide a description of conv X  by linear
presented. These properties allow sub- DEIS, University of Bologna inequalities  and show  how to separate
stantial reductions in problem input size coauthors: Fred Glover, Antonio over conv X  in O n log n  time. We also
to be carried out. Additionally, a stronger Frangioni, Giovanni Rinaldi provide  an integral extended formulation
formulation of the problem is also made keywords: linear programming, of conv X  of polynomial size. Finally,
possible. Based on these ingredients, Lagrangean techniques, cutting planes we give some conditions under which
 101

 
conv X 
Nz  d
is an integral poly- rather than vectors. Under a Slater- absolute values of the function deriva-
hedron. These results generalize earlier type constraint qualification a semidefi- tives is to be added to each function.
results of Gunluk and Pochet for the case nite program is equivalent to the associ- Suppose, however, that only values
I2  0./ ated (primal-dual) optimality conditions. but not derivatives of the object and
In this talk we consider a semismooth constraint functions are available. In
reformulation of these optimality condi- this case the calculation of finite dif-
WE2-306/32 NONLINEAR PROGRAMMING

 as a nonsmooth system of equations

tions ference approximations of the deriva-
Algorithms for degenerate Θ W   0. This reformulation is done tives of “modified” functions may be
nonlinear optimization using the matrix valued minimum func- computationally expensive since the sec-
organizer/chair : Olga Brezhneva tion. In order to prove local fast conver- ond derivatives of the “original” function
gence of a nonsmooth Newton method must be estimated. In this paper, an ef-
Behavior of NLP codes on a de- applied to this reformulated system, we ficient scheme for such approximation is
generate optimal control prob- first investigate the B-subdifferential of Θ proposed. It normally requires n  k  3
lem at a solution W . calculations of the “original” function,
The nonsmooth Newton-type method and never needs more than 2n  k  3 cal-
C ATARINA AVELINO presented here achieves a local quadratic culations, where n is the number of opti-
University of Coimbra rate of convergence under a linear inde- mization variables and k is the number of
coauthor: Luís N. Vicente pendency assumption and a (modified) parameters. Thus the proposed scheme
keywords: nonlinear optimization, nondegeneracy condition. We try to allows robust zero-order optimization at
degeneracy, optimal control avoid strict complementarity. almost the same expense as the conven-
Our test problem is a discretized optimal tional non-robust approach, while using
control problem governed by a semilin- Methods for nonlinear pro- the same optimization packages.
ear elliptic equation. When there are only gramming without strict com- The method has successfully been
bounds on the control variables this dis- plementarity assumption applied to the design of high-frequency
cretized problem can be shown to be al- electrical filters.
ways nondegenerate (in the sense that the
O LGA B REZHNEVA
IMA, University of Minnesota
gradients of the functions defining the ac- Analysis of response surface
tive constraints are linearly independent). keywords: nonlinear programming,
However, the problem becomes degener- superlinear convergence, degeneracy, problems using quadratic pro-
strict complementarity gramming
ate by including an appropriate number
of bounds on the state variables (active at For nonlinear constrained optimization
the solution). problems, we consider the case when
R ENTSEN E NKHBAT
National University of Mongolia
We ran the following NLP codes the strict complementarity condition does
on our problem instances: IPOPT (line not hold. In this situation, only linear coauthor: T. Ibaraki
search, filter, barrier interior point), KNI- rate of convergence can be guaranteed for keywords: quadratic programming,
TRO (trust region, SQP, barrier interior most classical algorithms. In this talk, we optimization method, response surface,
point), and LOQO (line search, primal- present new local methods that are slight design of experiments
dual interior point). We tried a prelim- modifications of the well-known algo-
inary implementation of an augmented rithms. Proposed methods attain super- The mathematical theory of experimental
Lagrangian multipliers method using the linear convergence even when the strict design is divided into two parts: design of
optimal control structure (AUGCON). complementarity condition does not hold extremal experiments and response sur-
In the nondegenerate case we also ran and subsume the case when this condi- face problems. The main principle of ex-
TRICE (trust-region, SQP, affine-scaling tion holds. tremal experiment is to obtain the maxi-
interior point, optimal control structure). mum information about investigated pro-
In this talk we will report our numer- cess for a given number of experiments
WE2-306/33 NONLINEAR PROGRAMMING

and reduce the number of experiments for

ical results for different settings of the Approximation of gradients
problem (including different mesh sizes) a given precision for the model expressed
and provide explanations for some of nu- chair : Benoit Hamelin by nonlinear regression functions. Mean-
while, the response surface deals with op-
merical behavior that has been observed. Computationally efficient ap- timization problems defined over a given
proximation of derivatives for criteria of experiment and experimental
robust optimization region.
Quadratic convergence of a non- S ERGE S HISHKIN In this paper we consider the re-
smooth Newton-type method for Sawtek, Inc. sponse surface problems which are for-
semidefinite programs keywords: robustness, zero-order mulated as the general quadratic pro-
optimization, nonlinear programming gramming. The general quadratic pro-
C HRISTIAN NAGEL gramming is split into: convex quadratic
University of Wuerzburg theory, computational complexity
maximization, convex quadratic mini-
coauthor: Christian Kanzow Consider the optimization problem in mization and indefinite quadratic pro-
keywords: semidefinite program, which the variables and parameters may gramming. Based on optimality condi-
quadratic convergence, nondegeneracy, fluctuate randomly within given limits, tions, we propose finite algorithms for
strict complementarity but the optimization is to be performed solving those problems. As application,
by a “conventional” optimization pack- some real practical problems arisen in the
Semidefinite programs may be viewed as age. This can be achieved by replacing response surface, one of the main part of
generalizations of linear programs with all functions by their “robust” modifica- design of experiment, have been solved
the variables being symmetric matrices tions: some differentiable functional of numerically by the algorithms.
102 Parallel Sessions: Wednesday 13:15 – 14:45 (WE2)

decomposed SQP step. It is shown that show how the primal-dual Newton equa-
Automatic finite differences, an if ρ is bounded for each barrier param- tions can be solved efficently, by exploit-
alternative to automatic deriva- eter µ, then every limiting point of the ing the structure and solving three sys-
tives sequence generated by the algorithm is tems, two of which being totally seper-
a Karush-Kuhn-Tucker point; otherwise able with respect to the scenarios and the
B ENOIT H AMELIN the sequence has a limiting point which third one enjoying a very sparse struc-
Université de Sherbrooke is either a feasible point with gradients ture, again exploitable. We then give
coauthor: Jean-Pierre Dussault of the active constraints being linearly our general algorithm, whose inner itera-
keywords: unconstrained optimization, dependent or an infeasible point which tion solves a trust-region barrier subprob-
automatic differentiation, robust is also a stationary point of minimizing lem by a dogleg method that combines
implementation the 2 -norm of the infeasibility. The al- a quasi-Newton step (computed via the
gorithm produces the correct results for above-mentioned decomposition) with a
We previously proposed to adapt the for-
some hard problems, including the ex- projected Cauchy point. We draw the
ward automatic differentiation technique
amples provided by Wächter and Biegler main theoretical properties of the pro-
to compute finite differences instead of
and Byrd, Marazzi and Nocedal. posed class of algorithms and present
derivatives. We then used the finite dif-
some preliminary numerical experimen-
ferences to reduce the cancellation er-
tations.
ror when computing the actual descent On the solution of linear equa-
within a differentiable optimization algo- tions arising in interior methods
rithm, whether of line search or trust re- for nonconvex optimization WE2-306/35 NONLINEAR PROGRAMMING

gion type. In this presentation, we gen- A NDERS F ORSGREN NLP software - state of the art
eralize the use of automatic finite differ- Royal Institute of Technology (KTH) I
encing to compute accurate finite differ-
ence estimates of directional derivatives. keywords: nonlinear programming, organizer/chair : Arne Stolbjerg Drud
We thus propose robust and efficient trun- interior method Detecting unboundedness in
cated Newton and quasi-Newton algo- When solving smooth optimization prob- practical nonlinear models
rithms enhanced by the clever use of au- lems by interior methods, an important
tomatic finite differences. We analyze step at each iteration is the solution of A RNE S TOLBJERG D RUD
the respective benefits of the automatic fi- a system of linear equations. The ma- ARKI Consulting & Development A/S
nite difference and those of the automatic trix involved is typically symmetric and
derivative approaches. Numerical com- indefinite, often referred to as the bar- keywords: nonlinear programming,
parisons on a set of test functions are re- rier KKT matrix. For large-scale prob- unboundedness
ported for many algorithmic variants. lems, the barrier KKT matrix is often
Nonlinear models can be unbounded
sparse. An important feature for noncon-
without having unbounded rays. To
WE2-306/34 NONLINEAR PROGRAMMING vex problems is that it is of interest to
prove unboundedness it is therefore of-
know the inertia of the barrier KKT ma-
Large scale nonlinear trix, in order to determine if the problem
ten necessary to actually move towards
programming II may be considered locally convex around
infinity. This can cause numerical prob-
organizers : Sven Leyffer and Richard lems, e.g. with large derivatives, scaling,
the current iterate. We discuss methods
Waltz etc. and the chance of mislabeling un-
for solving the linear equations involving
chair : Richard Waltz bounded models is large, in practice often
the barrier KKT matrix which reveal the
over 50%.
A robust primal-dual interior inertia. We compare a set of NLP solvers on
point algorithm for nonlinear several simple unbounded models, clas-
programs On interior-point methods us- sify types of unboundedness, and suggest
ing a suitable decomposition for ways to improve the classification of the
J IE S UN multistage nonlinear stochastic
National University of Singapore models.
programming
coauthor: Xinwei Liu
keywords: nonlinear optimization, A NNICK S ARTENAER Mathematical programs with
interior point method, global FUNDP equilibrium constraints: auto-
convergence coauthors: Jie Sun, Fabian Bastin matic reformulation and solu-
We present a primal-dual interior point keywords: stochastic programming, tion via constrained optimiza-
algorithm for solving optimization prob- nonlinear programming, interior point tion
lems with nonlinear inequality con- method
straints. The algorithm has some of We consider a multistage stochastic opti- S TEVEN D IRKSE
the theoretical robustness of a trust re- mization problem, with nonconvex non- GAMS Development Corp.
gion method, but works entirely by line linear objective function, linear equality coauthors: Michael Ferris, Alex
search. Global convergence properties constraints and nonnegativity constraints. Meeraus
are derived without assuming regularity In order to take the uncertainty into ac- keywords: MPEC, NLP, reformulation,
conditions. The penalty parameter ρ in count, we make use of a scenario formu- complementarity
the merit function is updated adaptively lation that transforms the original prob-
and plays two roles in the algorithm. lem into a larger but highly structured Constrained optimization has long been
First, it guarantees that the search direc- linearly constrained problem. We then used to solve large scale (non)linear pro-
tions are descent directions for the up- present a feasible primal-dual barrier ap- gramming problems arising in a variety
dated merit function. Second, it helps to proach globalized by a trust-region tech- of disciplines. A separate set of models
determine a suitable search direction in a nique to solve this problem. We first have been generated more recently, using
 103

complementarity to model various phe- The prox-regularity condition, developed Some applications to solving prob-
nomenon, particularly in general equilib- by Poliquin and Rockafellar in 1996, is lems of classification and cluster analy-
ria. The unifying framework of mathe- a useful generalization of convexity. In sis are presented. Numerical results are
matical programs with equilibrium con- particular, in 2000, Poliquin, Rockafellar, demonstrated.
straints (MPEC) has been postulated for and Thibault showed that a prox-regular
problems that combine facets of both set has a locally single-valued projection A robust gradient sampling al-
optimization and complementarity. In mapping. gorithm for nonsmooth, non-
this talk we briefly review some methods In 2003, Lewis introduced the class convex optimization
available to solve these problems and de- of partly smooth functions. These func-
scribe a new suite of tools for working tions, though nonsmooth, have a smooth M ICHAEL OVERTON
with MPEC models. These tools auto- substructure. By adding a prox-regularity New York University
matically reformulate the MPEC model condition to his definition, we can exploit coauthors: James Burke, Adrian S.
as an NLP, solve the NLP, and recover this substructure to identify active con- Lewis
the MPEC solution. We present compu- straints, analyse critical points, and en- keywords: nonconvex, nonsmooth, non
tational results from a large set of test sure uniqueness of active manifolds. Lipschitz
problems demonstrating the potential of Let f : Rn

R be nonconvex, continu-
the different nonlinear programming re-
formulations of MPEC problems. In ad- Nonlocal methods for solving ous, and continuously differentiable on
dition, the advantages of using a domain- some classes of nonsmooth and an open dense subset of Rn where its gra-
specific function library to implement nonconvex optimization prob- dient is easily computed. We present
lems a practical, robust algorithm to locally
these reformulations are discussed. minimize such functions based on gradi-
A LEXEI D EMYANOV ent sampling.
Global optimization with Saint-Petersburg State University If, in addition, f is locally Lipschitz
GAMS - applications and per- keywords: sum-min function, max-min and coercive and the sampling diameter ε
formance functions, nonsmooth optimization, is fixed, then, with probability one, the al-
exchange algorithms gorithm generates a sequence with a clus-
M ICHAEL B USSIECK ter point that is Clarke ε-stationary. Fur-
GAMS Development Corp. Let X be a metric space, Ω X. Func- thermore, if f has a unique Clarke sta-
coauthors: Leon Lasdon, Janos Pinter, tions ϕik : X

R and index sets I  tionary point x̄, then the set of all cluster
 
Nikolaos Sahinidis 1    N
Ji  1    mi
i  I are points generated by the algorithm con-
keywords: global optimization, given. Define the functions verges to x̄ as ε is reduced to zero.
modeling system   We present numerical results that
F1 x   ∑ min ϕik x  demonstrate the robustness and utility of
i  I k  Ji
Mixed integer non-linear optimization
problems can be formulated and solved the algorithm in a variety of contexts,
with GAMS for more than a decade.   including cases where f is not Lips-
F2 x   max min ϕik x  chitz at minimizers. Approximate local
Users of non-linear models had to cope i  I k  Ji minimizers are reported for functions in
with the limits of available local solvers.
Lots of effort can go into finding a The problem is to minimize the func- the applications literature that have not,
“good” starting point. Furthermore, the tions F1 and F2 on the set Ω. In fact we to our knowledge, been obtained pre-
local solver stops at the first local opti- aim at being able to minimize not only viously. When the termination criteria
mum. Recent advances in global opti- these functions, but also more general of the algorithm are satisfied, a precise
mization (GO) made the introduction of functionals of the so-called "sum-min" statement about nearness to Clarke ε-
three solid GO solvers into the GAMS and "max-min" types. stationarity is obtained. A MATLAB im-
system possible: BARON, LGO, and plementation is freely available.
The problem of minimizing functions
OQNLP. These solvers help to overcome of sum-min, max-min or similar types
represent a very important, though ex- WE2-306/37
CONVEX PROGRAMMING
some of the limits of local optimization.
In this talk we will discuss different mod- tremely complicated class of optimiza- Duality and optimality
eling requirements for local and global tion problems. When dealing with such chair : Michael L. Flegel
codes. We will focus on differences be- functions one has to face the following
tween the three solvers, present favored difficulties: their nonsmoothness, non- Enhanced optimality conditions
application, and compare performance. convexity, and a large number of local and informative multipliers for
minimizers they possess. convex programming
WE2-306/36 NONSMOOTH OPTIMIZATION The following approach is suggested:
it is shown that the problem is equiva-
A SUMAN O ZDAGLAR
Non-convex problems lent to another problem, and this equiv-
MIT
chair : Michael Overton alence has remarkable consequences al- coauthors: Dimitri Bertsekas, Paul
Partial smoothness and prox- lowing to state conditions for a mini- Tseng
regularity in optimization mum of nonlocal type. The definition keywords: optimality condition,
of a stationary point is introduced, that Lagrange multiplier, sensitivity
WARREN H ARE enables one to decrease the number of In previous work, we presented a new
Simon Fraser University invloved local minimizers. Using these development of Lagrange multiplier the-
coauthor: Adrian S. Lewis conditions some methods for finding and ory for smooth problems, which is mo-
keywords: nonsmooth optimization, “improving” local minimizers are con- tivated by an enhanced set of optimality
prox-regularity, partly smooth, structed (Exchange and ε Exchange al- conditions. These optimality conditions
variational analysis gorithms). led to the introduction of new conditions,
104 Parallel Sessions: Wednesday 13:15 – 14:45 (WE2)


which generalize, unify, and streamline xs λs
to the primal-dual solution in Deterministic global optimiza-
the theory of constraint qualifications. value, i.e tion: a new class of improved
In this talk, we will focus on con- f x lim f xs lim d λs d λ
convex underestimators for con-
vex constrained optimization problems

s  ∞

s  ∞

strained NLPs
and present enhanced Fritz John type op- The max constraints violation   and the C HRISTODOULOS F LOUDAS
timality conditions for these problems duality gap are bounded by o µs   0 5  Princeton University
without assuming the existence of a pri-
mal optimal solution. In particular, un- Theorem 3: If the complementarity coauthors: Ioannis G. Akrotirianakis,
der general conditions, we show the ex-  in the strong form:
condition  satisfies Christodoulos A. Floudas
istence of informative multipliers that max1  i  m λi ci x 
0 then keywords: convex underestimators,
carry sensitivity information by indicat- d λ d λs o µs  1 global optimization, constrained NLPs
ing the constraints to relax to affect a

A novel class of convex underestima-

cost reduction. We also present enhanced tors for twice continuously differentiable
optimality conditions for the dual opti- Consider an LP:
nonconvex functions is introduced. The
mization problem and investigate exis- a x min
a  x Ax b


5 convex underestimating functions are

tence of informative dual optimal solu- based on a relaxation component that in-
tions, which provide sensitivity informa- Theorem 4: If the dual LP to (5) has a volves a sum of univariate parametric ex-
tion even when there is a duality gap, and unique solution and the kernel ϕ is well ponential terms. An efficient procedure
hence there exist no multipliers. defined(see[1]) then there exists c 0 based on interval arithmetic determines
This work is part of the new book independent on µ 0 that the parameter values and the same time
"Convex Analysis and Optimization" by b λ
b  λs 1  cµ  1 b  λ

b  λs 2 it verifies convexity. Computational stud-

D. Bertsekas, with A. Nedic, and A. ies and comparisons with the aBB global
Ozdaglar, published in April 2003.
Reference 1.Polyak R. Nonlinear Rescal- optimization approach illustrate the im-
ing vs. Smoothing technique in convex proved convex underestimators.
Lagrangean transformation in optimization. Math Programming Ser
convex optimization A92:197-235(2002) Global optimization in the co-
ROMAN P OLYAK conut project
George Mason University On the Guignard constraint H ERMANN S CHICHL
keywords: Lagrangian transformation, qualification for mathematical Universität Wien
prox method, duality programs with equilibrium con-
straints keywords: global optimization, verified
We assume that the optimal set of the computing, directed acyclic graphs
convex programming problem M ICHAEL L. F LEGEL In this talk, a solver platform for global
X arg max
f x ci x
0  i

1      m 1

University of Würzburg optimization will be presented, as it is de-
coauthor: Christian Kanzow veloped in the COCONUT project. There
is bounded and Slater conditions holds.
Let Ψ be a class of concave func- keywords: MPECs, constraint will be a short description of the basic
tions ψ  C2 with particular proper- qualifications, Guignard CQ, first order algorithmic concept and of all relevant

ties(see 
[1]).Along 
with Lagrangean conditions components, the strategy engine, infere-
L x λ  f x ∑ λi ci x  we consider Mathematical programs with equilibrium nence engines, and the remaining mod-
m

the Lagrangean Transformation(LT) constraints (MPECs) are nonlinear con- ules. A compact description of the search
Ł : R n  Rm  R  R given by for- strained optimization problems which, graph and its nodes and of the inter-
mula at no feasible point, satisfy any of the nal model representation using directed
Ł x λ k
f x

∑ ki 1 ψ ki λi ci x


standard constraint qualifications known acyclic graphs (DAGs) and their advan-
from nonlinear programming. One ex- tages completes the presentation.
Let       and µ 0 The
λo Rm ko Rm ception to this is the Guignard con-
LT method generates the primal,dual and straint qualification. Recent research in Investigating population algo-
the scaling vector sequences by formulas. the area of constraint qualifications for rithms; a spatial GA
xs  1
:
MPECs has uncovered a string of mod-
∇x Ł x s  1
λ k
s s

∇x L x s  1
λ
s 1 
0 2
ifications to standard constraint qualifi- E LIGIUS M ARIA T HEODORUS
 cations which lead to suitable first or- H ENDRIX
λis 1
ψ µ λsi  1 ci xs

1


λsi 3

der optimality conditions, the strongest Wageningen University

of which is a KKT-point of the MPEC, keywords: evolutionary algorithms,
kis  µ λsi  
1 1 2
4
  known as strong stationarity in the MPEC spatial optimization

Let d λ    infx L x λ  be the dual community. We show where the Guig-
function ψ s   inft st ψ t 
be the nard constraint qualification stands in re- Population (evolutionary) algorthims are
Fenchel conjugate for the transformation lation to the constraint qualifications in- applied nowadays for solving for in-
ψ  Ψ and ϕ  ψ be the kernel of the troduced in the context of MPECs. This stance optimal control, parameter estima-
second order Entropy-like  distance func- yields an elementary approach to strong tion and spatial optimization problems.
tion D u v   ∑ni 1 v2i ϕ ui  vi 1  stationarity. As more tradional approaches from non-
  linear and global optimization could also
Theorem 1:The LT method 2  4 be applied, it is useful to follow this de-
is well defined and equivalent to the In- WE2-306/38 GLOBAL OPTIMIZATION

velopment in a critical way: what could

terior Prox method with Entropy-like Global optimization II be advantages and drawbacks of such ap-
distance function D u v  (see [1]). proaches. We were involved in several
chair : Eligius Maria Theodorus
Theorem 2: The primal-dual sequence Hendrix projects where analysis and experiments
 105

of population algorithms play a role. A sets related to an allure. Some relations M ARKKU K ALLIO
recent application has been in spatial op- between them are also presented. Helsinki School of Economics
timization, where the evolutionary algo- keywords: real options, stochastic
rithms are designed especifically for de- On the connectedness of mix programming, asset pricing, arbitrage
cision variables that have a spatial struc- semi-efficient frontier theory
ture. We will report about the expe-
rience with applying and experimenting A NNA M ARCHI
A general approach for (real) option val-
with such algorithms. University of Pisa uation is presented in a discrete time, dis-
keywords: connectedness, efficiency crete probability space framework. An
WE2-308/11 GENERALIZED
CONVEXITY / MONOTONICITY
In this paper the concepts of mix semi- option is defined by a compact choice
Generalized convexity III efficint point and quasi-concave set are set of stochastic cash flow streams result-
organizer : Laura Martein introduced. By means of these concepts ing from feasible choices within an op-
chair : Rita Pini we will investigate the connectedness of tion. Based on the buying price princi-
ple, option value is determined as a com-
Coercivity conditions and equi- the mix semi-efficient frontier for vector petitive price; i.e. the maximum price
librium problems maximization problems, having an ob-
jective function whose components are an agent is willing to pay, given com-
R ITA P INI quasi-concave, strictly and strong quasi- peting investment opportunities, such as
Università di Milano-Bicocca concave functions. We will prove that publicly traded financial instruments. We
coauthor: Monica Bianchi the mix semi-efficient frontier of a mix show consistency of such valuation with
keywords: equilibrium problems, quasi concave problem is a closed and arbitrage pricing theory. For evaluations,
connected set. an agent’s multi-period utility is deter-
coercivity conditions, generalized mined by a yield stream, such as profit,
monotonicity consumption or wealth, and a multi-stage
We consider an equilibrium problem in a WE2-308/12
STOCHASTIC PROGRAMMING

portfolio model is employed to find op-

reflexive Banach space in the framework Option prices and timal levels of yield and investment in
of the formulation proposed by Blum and optimization competing assets. Additive exponential
Oettli, which includes, for instance, vari- chair : Teemu Pennanen utility functions allow major savings in
ational inequalities and vector optimiza- computational burden of valuation.
tion problems as particular cases. The Recovering risk-neutral prob-
study of the existence of solutions of such abilities from option prices-a
problems on unbounded domains usually convex quadratic programming Arbitrage pricing of American
involves the same sufficient assumptions approach contingent claims in incomplete
as for bounded domains together with a A NA M ARGARIDA M ONTEIRO markets
coercivity condition. We focus on two University of Coimbra
different conditions. The first one is ob-
tained by assuming the existence of a coauthors: Reha Tutuncu, Luís N.
T EEMU P ENNANEN
Helsinki School of Economics,
bounded set such that no elements out- Vicente
keywords: risk neutral pdf, estimation, Department of Management Science
side is a candidate for a solution; the
other one allows the solution set to be option prices, convex quadratic coauthor: Alan King
unbounded. Our results exploit the gen- programming keywords: stochastic programming,
eralized monotonicity properties of the We present a new approach to compute convex duality, American contingent
function f defining the equilibrium prob- the risk-neutral probability density func- claims, arbitrage
lem. It turns out that, both in the pseu- tion (pdf) of an underlying asset from
domonotone and in the quasimonotone option prices written on the asset. The Convex optimization provides a natural
setting, an equivalence can be stated be- estimation is carried out in the space of framework for studying the pricing of fi-
tween the nonemptyness and the bound- cubic splines functions, yielding appro- nancial derivatives in incomplete market
edness of the solution set, and these co- priate smoothness. The resulting opti- models. Besides its simplicity, one of
ercivity conditions. In the pseudomono- mization problem, used to invert the data the main advantages of the optimization
tone case, we compare various coerciv- and determine the corresponding density approach is that it is computational. In-
ity conditions appeared in literature with function, is a convex quadratic program- deed, many algorithms are available for
ours. ming problem, which can be efficiently pricing problems as soon as they are set
solved by nowadays numerical optimiza- up as optimization problems. Moreover,
Generalized convexity on allures tion software. The nonegativeness of the convex optimization duality theory has
in the sense of elena popoviciu risk-neutral pdf is ensured by imposing been shown to yield well-known expres-
linear inequality constraints at the nodes. sions for the prices of European contin-
DANIELA M ARIAN We tested our approach using data gen- gent claims in terms of martingale mea-
Grup Scolar de Transporturi Cai Ferate erated from Black-Scholes option prices sures. This paper extends the analysis
Cluj-Napoca and from the SPX 500 Index. The numer- to the case of American style contin-
keywords: convex sets ical results presented show the effective- gent claims. The pricing problems of the
seller and the buyer of an American con-
We use the notion of allure, in the ness of this methodology for estimating
the risk-neutral probability density func- tingent claim are shown to be expressible
sense of Elena Popoviciu, in definition
tion. as convex optimization problems. Af-
of many roughly convex sets. We de-
ter this, martingale expressions for the
fine and study ϕ-convex respectively δ-
buyer’s and seller’s prices are obtained
convex, midpoint δ-convex, lightly δ- Real option valuation via by inspecting the dual optimization prob-
convex, midpoint γ-convex, r-convexlike stochastic optimization
lems.
106 Parallel Sessions: Wednesday 13:15 – 14:45 (WE2)

WE2-308/13 STOCHASTIC PROGRAMMING metrics and their consequences on stabil- undersupplied" set of bidders. We con-
ity of the underlying stochastic programs. sider this concept to be the natural gener-
Stochastic programming with In particular, we present stability results alization of an “overdemanded" set of ob-
risk measures for models containing coherent risk mea- jects, introduced by Demange, Gale and
organizer/chair : Andrzej Ruszczynski sures. Furthermore, we discuss proper- Sotomayor for the one-to-one assignment
ties of risk functionals that are favourable problem.
Risk aversion via stochastic for maintaining decomposition structures Under a substitutes condition, the
dominance constraints (Part I) of (multistage) stochastic programs. The auction implements the Vickrey-Clarke-
A NDRZEJ RUSZCZYNSKI results are motivated by applications to Groves outcome. Furthermore, we show
Rutgers University stochastic power management models. that this substitutes condition is a neces-
sary condition to do so.
coauthor: Darinka Dentcheva
keywords: stochastic programming, WE2-308/T1 FINANCE AND ECONOMICS

WE2-308/T2 COMPLEMENTARITY AND VARIATIONAL

INEQUALITIES
risk, stochastic dominance, convex Combinatorial auctions I
analysis Complementarity methods for
organizer/chair : Susan Powell simulating nonsmooth
We introduce a new stochastic optimiza- mechanical systems
tion model involving stochastic domi- Pause: a general combinatorial
auction procedure organizer/chair : Mihai Anitescu
nance constraints and we discuss its rela-
tions to existing models. It is a convex in- Optimization problems with
finite dimensional optimization problem R ICHARD S TEINBERG complementarity constraints in
with partial order constraints in the space University of Cambridge mechanics
of integrable random variables. We de- coauthors: Ailsa Land, H.Paul Williams
keywords: bidding/auctions
M IHAI A NITESCU
velop necessary and sufficient conditions Argonne National Laboratory
of optimality for this problem. We show Pause is an auction procedure proposed
that the Lagrange multipliers correspond- keywords: nonsmooth mechanics,
by Kelly and Steinberg that permits all complementarity constraints
ing to dominance constraints are concave combinatorial bids while being computa- We discuss recent algorithmic results for
nondecreasing utility functions. tionally tractable for the auctioneer and complementarity-based simulation for
transparent to the bidders. We examinine time-dependent nonsmooth mechanics.
Risk aversion via stochastic in detail bidder behaviour and illustrate We present applications of our results to
dominance constraints (Part II) this with numerical examples. robotics and granular mechanics.
DARINKA D ENTCHEVA
Stevens Institute of Technology Combinatorial auctions in prac- Differential
ities
variational inequal-
tice
coauthor: Andrzej Ruszczynski J ONG -S HI PANG
keywords: stochastic programming, A RNE A NDERSSON The Johns Hopkins University
convex analysis, stochastic dominance, Uppsala University, Dept of Information coauthor: David E. Stewart
risk Technology keywords: variational inequalities,
We analyze a new class of stochastic op- keywords: auctions, combinatorial complementarity problem, differential
timization problems involving stochastic auctions, business applications systems
dominance constraints. The problem is This paper introduces the class of dif-
formulated as an optimization problem The presentation will cover some expe- ferential variational inequalities (DVIs)
with decisions in a separable locally con- riences from real-world use of combina- in a finite-dimensional Euclidean space.
vex Hausdorff vector space and with con- torial auctions, inculding some scientific The DVI provides a powerful modeling
straints in the form of a stochastic order challenges. paradigm for many applied problems in
between random outcomes. We develop which dynamics, inequalities, and dis-
duality theory for these models, involv- On ascending vickrey auctions continuities are present; examples of
ing utility functions as dual variables as- for heterogeneous objects such problems include constrained time-
sociated with dominance constraints. We dependent physical systems with unilat-
also construct specialized decomposition R AKESH VOHRA eral constraints, differential Nash games,
methods for the case of discrete distribu- Northwestern University and hybrid engineering systems with
tions. variable structures. The DVI unifies sev-
coauthors: Sven de Vries, James
eral mathematical problem classes that
Schummer
include ordinary differential equations
Risk functionals in stochastic keywords: auctions, duality
(ODEs) with smooth and discontinuous
programming: stability and al- We consider a setting in which bidders right-hand sides, differential algebraic
gorithmic issues have nonadditive valuations for sets of equations (DAEs), and dynamic comple-
W ERNER ROEMISCH heterogeneous, indivisible objects. We mentarity systems. We present condi-
Humboldt-University Berlin, Institute of construct an ascending auction by using tions under which the DVI can be con-
Mathematics a primal-dual algorithm associated with verted, either locally or globally, to an
a linear programming formulation of the equivalent ODE with a Lipschitz con-
coauthor: Andreas Eichhorn efficient-allocation problem for this set- tinuous right-hand function. For DVIs
keywords: stochastic programming, ting. The auction assigns nonanonymous that cannot be so converted, we con-
risk, stability, decomposition prices to sets, and asks bidders which sets sider their numerical resolution via an
We study the continuity of risk function- they demand in each round. Prices are Euler time-stepping procedure, which in-
als with respect to suitable probability increased by determining a “minimally volves the solution of a sequence of
 107

finite-dimensional variational inequali- given capacity C must transport objects Spielman and Teng introduced the con-
ties. Borrowing results from differential between points of a metric space. Each cept of smoothed analysis to explain the
inclusions (DIs) with upper semicontin- transportation request is defined by its re- success of algorithms that are known to
uous, closed and convex valued multi- lease time, its source, and its destination. work well in practice while presenting
functions, we establish the convergence An online algorithm learns about the ex- poor worst case performance. Spiel-
of such a procedure. Lastly, we present a istence of a request at its release time, man and Teng [STOC 01] proved that
class of DVIs for which the theory of DIs and it must serve each request by trans- the simplex algorithm runs in expected
is not directly applicable, and yet a simi- porting an object from the given source to polynomial time if the input instance is
lar convergence result can be established. the destination, never carrying more than smoothened with a normal distribution.
C objects at once. The goal is to mini- We extend this notion to analyze on-
mize the maximum flow time, where the line algorithms. In particular we intro-
A robust sequential quadratic flow time of a request is the difference duce smoothed competitive analysis to
programming method for math- between the time when the object reaches study the Multi-Level Feedback (MLF)
ematical programs with linear its destination and its release time. algorithm, at the basis of the scheduling
complementarity constraints It is easy to see that competitive anal- policies of Unix and Windows NT, when
ysis fails to distinguish online algorithms the processing times of jobs released over
X INWEI L IU for this problem unless the adversary is time are only known at time of comple-
Singapore-MIT Alliance, National restricted. We extend the fair adversary tion.
University of Singapore defined for the real line by Blom et al. to We show that, if the k least signifi-
coauthor: Jie Sun the uniform metric space and show that cant of K bits describing the processing
keywords: complementarity constraints, a simple first-come-first-serve strategy is time of a job are randomly changed, MLF
inequality relaxation, global 2-competitive and best possible, if source achieves a tight smoothed competitive ra-
convergence, regularity and and destination of each request coincide. tio of O 2K  k  to minimize the average
nondegeneracy For the case that they differ, we provide flow time. A direct consequence is a first
a worst-case construction that shows that constant approximation for this problem
The relationship between the mathemat-
no deterministic online algorithm can be in the average case under a quite general
ical program with linear complementar-
competitive for the OLDARP, not even class of probability distributions.
ity constraints (MPCC) and its inequal-
against the fair adversary. For various other smoothening mod-
ity relaxation is studied. A new se-
quential quadratic programming method els, including the one used by Spiel-
is presented for solving the MPCC based The generalized 2-server prob- man and Teng,  we present a higher lower
on this relationship. A function with lem bound of Ω 2K  .
a penalty term measuring the violation
of the complementarity constraint relax- R ENE S ITTERS WE2-341/21 INTERIOR POINT ALGORITHMS

ation is used to decide the stepsize in Technische Universiteit Eindhoven

each iteration. Global convergence re- coauthor: Leen Stougie Interior point methods in
sults are derived without assuming the keywords: on-line algorithms linear programming
linear independence constraint qualifica- We consider the generalized 2-server chair : Miguel Argaez
tion for MPCC and nondegeneracy of problem in which at each step both
the complementarity constraints. Prelim- servers receive a request, which is a point An iterative method for solving
inary numerical results show the valid- in a metric space. One of the two servers the search direction of linear
ity of the algorithm for the test problems has to move to its request. If the sequence programming
generated by QPECgen and the test prob- is known in advance, then the problem
lems with non-quadratic objective func- can be solved efficiently by dynamic pro-
H UA W EI
University of Waterloo
tions. gramming. It has been a well-known
coauthors: Maria Gonzalez-Lima,
open question in online optimization if a
WE2-308/T3 ON - LINE OPTIMIZATION Henry Wolkowicz
constant factor approximation algorithm
keywords: linear programming, large
On-line optimization exists for the generalized 2-server prob-
sparse problems, linear equations,
algorithms lem. We answer this question in the af-
interior point method
chair : Tjark Vredeveld firmative sense by providing a constant
competitive algorithm based on the so We give a new method to solve the per-
Minimizing the maximum flow called work function. turbed optimality condition to get the
time in the online dial-a-ride search direction in interior-point method
problem Smoothening helps: a proba- for linear programming. A typical
D IANA P OENSGEN bilistic analysis of the multi- method to solve the perturbed optimality
Konrad-Zuse-Zentrum fuer level feedback algorithm condition is to form the normal equation
Informationstechnik (ZIB) and then use Choleski factorization to di-
T JARK V REDEVELD rectly solve it. We show that the normal
coauthors: Sven O. Krumke, Alberto Konrad-Zuse-Zentrum equation approach may not be stable near
Marchetti-Spaccamela Luigi Laura, the boundary of the feasible region due
Maarten Lipmann, Willem E. de Paepe, coauthors: Luca Becchetti, Stefano
Leonardi, Alberto to the back solve. We then show that by
Leen Stougie using a different block elimination, we
keywords: on-line dial-a-ride problem, Marchetti-Spaccamela, Guido Schaefer
keywords: smoothed competitive can get another linear system that does
maximum flow time, fair adversary not become ill-conditioned under nonde-
analysis, on-line algorithms, multi-level
In the Online Dial-a-Ride Problem feedback algorithm, probabilistic generacy assumptions. This system is
(OLDARP), a server of unit speed and analysis a sparse non-symmetric system. It is
108 Parallel Sessions: Wednesday 13:15 – 14:45 (WE2)

equivalent to the so-called Quasi- definite An inner approximation algo- analysis of the fluid model in these two
system after a scaling. We then apply di- rithm for the dynamic user equi- application areas and establish existence
rect method as well as iterative method to librium problem of solution.
solve the system to get the search direc-
tion. The iterative method uses the fact T HANASIS Z ILIASKOPOULOS
that the sparse pattern of the system does Northwestern University On second-best toll pricing
not change. We construct an approximate keywords: flow algorithms, equilibrium problems
inverse of the system at each iteration as models
a preconditioner. The preconditioner in Traffic network equilibrium models,
S IRIPHONG L AWPHONGPANICH
next iteration can use the sparse pattern University of Florida
commonly used by planning agencies,
information of the preconditioner in the assume link travel time functions mono- coauthor: Don Hearn
current iteration. Thus we can construct tonically increasing with flow; this makes keywords: congestion pricing, traffic
a preconditioner very efficiently after the these models unsuitable for congested equilibrium, MPEC
first iteration. Numerical result is pre- networks for which such a monotonic
sented. relationship does not hold. This pa- In this talk, we address the second-
per introduces a Variational Inequality best toll pricing problems with fixed and
On the convergence of interior (VI) formulation for computing equilib- elastic travel demand, and discuss how
point methods for linear pro- rium flows that circumvents this draw- they can be applied to the current prac-
gramming back by relying on traffic flow theoretical tices such as cordon toll pricing, High-
models and non-steady state demand in- Occupancy Toll (HOT), and Fast and In-
C ORALIA C ARTIS tertwine Regular (FAIR) lanes. We pro-
University of Cambridge flow. An Inner Approximation (IA) algo-
rithm is proposed that efficiently solves vide three equivalent MPEC formula-
keywords: linear programming, interior tions, establish existence conditions for
point method the VI formulation; the formulation and
the algorithm can solve large networks optimal toll vectors, and relate them to
We consider a long-step primal-dual in- marginal social cost pricing tolls. In ad-
terior point method and a second-order for steady state or time varying origin-
destination demand. The IA equilib- dition, we develop a cutting constraint
variant of it, applied to linear program- algorithm to solve the problems and
ming problems. The iterates are feasible rium algorithm relies on a traffic sim-
ulator to evaluate the link travel times; present a framework for determining a
with respect to the primal and the dual second-best toll vector that both induces
equality constraints. The behaviour of we demonstrate that under some mild
assumptions, the algorithm converges to a user equilibrium flow and achieves a
the two methods is analyzed on an illus- desired (secondary) goal, such as mini-
trative problem. We address the problem a user equilibrium solution. Compu-
tational experiments on large networks, mizing the maximum toll, the number of
of the convergence of the sequences of it- toll facilities, and, for the fixed demand
erates generated by the two algorithms to such as the Chicago’s six-county network
, indicate reasonable convergence in ac- case, total toll revenue.
the analytic centre of the optimal primal-
dual set, both when the centring parame- ceptable CPU times.
ter is bounded away from zero and when WE2-341/23 APPROXIMATION ALGORITHMS

it converges to zero. A fluid dynamics model in dis- Approximation algorithms for

tribution systems and trans- metric facility location and
The role of the notion of quasi- portation spanning tree
central path for linear program- organizer/chair : Yinyu Ye
ming S OULAYMANE K ACHANI
Columbia University On universal facility location
M IGUEL A RGAEZ coauthor: Georgia Perakis
University of Texas at El Paso problem
keywords: fluid models, distribution
coauthors: Leticia Velazquez, Osvaldo systems, transportation, supply chain
Mendez
B O C HEN
In this talk, we present a fluid dynam- University of Warwick
keywords: linear programming, interior
point method ics model as it applies in anticipatory coauthors: Yinyu Ye, Jiawei Zhang
The notion of central path plays an im- route guidance as well as in supply chain keywords: facility location,
portant role in the development of most management. This fluid model is mo- approximation algorithms
primal-dual interior-point algorithms. In tivated from the dynamic network load-
this work we prove that a related notion ing model which is used for modeling We address the Universal Facility Loca-
called quasicentral path, introduced by the dynamic traffic assignment problem. tion (UniFL) problem, in which one is
Argaez and Tapia in nonlinear program- We consider the model’s application in a given a set of facility locations and a set
ming, is an enough region to guide the it- particular type of supply chain, that is, a of demands with locations. The objective
erates towards a solution of the problem. distribution system with several whole- is to assign the demands to facilities in
We use a new merit function for advanc- salers, intermediate distributors, and re- such a way that the sum of service and fa-
ing to the quasicentral path, and weighted tailers who are subsidiaries of the same cility costs is minimized. While the ser-
neighborhoods as proximity measures of company producing and selling multi- vice cost is proportional to the distance
this central region. ple products. Furthermore, we consider between the facility and the demand it
the model’s application in the anticipa- services, the facility cost of each facil-
tory route guidance problem which aims ity is a non-decreasing function of the
WE2-341/22 LOGISTICS AND TRANSPORTATION

to provide consistent messages to drivers amount of demand it services. A most

Applications of optimization based on forecasts of traffic conditions in difficult special case of UniFL problem
in transportation order to assist the drivers in their path se- is when the facilities have arbitrary hard
organizer/chair : Georgia Perakis lections. We design a framework for the capacities.
 109


Following the recent first constant 2 in O n log n  time, where n denotes the each job is associated with a due date
(8.53 + eps)-approximation algorithm for size of the associated Hanan grid. and a weight, whcih will be used to pe-
the problem with arbitrary hard capaci- Moreover, we prove that a certain nalize a job when it is not completed by
ties and then the first constant (7.88 + variant of the Hanan grid always contains the due date. This problem is already
eps)- approximation algorithm for UniFL an optimal solution. Based on this struc- known to be strongly NP-hard. To design
problem, we present improved approxi- tural result, we can improve the perfor- branch-and-bound algorithms for deriv-
mation results on both upper and lower mance guarantee of approximation algo- ing exact solutions, we develop several
bounds. rithms for the special case that all obsta- properties to curtail unnecessary explo-
cles are of rectangular shape or of con- ration. Our properties include two dom-
An efficient polynomial time  
stant complexity. For such a scenario, we inance rules and a lower bound. Be-
approximation scheme for the give a 2k  2k 1  α-approximation for sides, we propose a branching scheme
constrained minimum spanning any integral k
4, where α denotes the that not only provides a new tree struc-
tree using matroid intersection performance guarantee for the ordinary ture but also facilitate the enhancement
Steiner tree problem in graphs. of our lower bounds. We performed a set
A SAF L EVIN of computational experiments to study
Tel-Aviv university the performances of the proposed prop-
WE2-321/053 PRODUCTION AND SCHEDULING

erties. Computational experiments show

coauthor: Refael Hassin Flow shop scheduling
keywords: approximation algorithms, that our proposed properties can reduce
matroid intersection, Lagrangean chair : Débora Pretti Ronconi computational efforts required by tree ex-
relaxation Scheduling with simultaneous ploration. Statistics also suggest compli-
task processing for two-machine mentary results between the conventional
Given an undirected graph G=(V,E) with depth-first-search and our new scheme.
|V|=n and |E|=m, non-negative integers ce no-wait flowshop problem
This result can suggest a line of thinking
and de for each edge e  E, and a bound O ULAMARA A MMAR in constructing search tree as well as im-
D, the constrained minimum spanning LORIA-INRIA proving lower bounds for some optimiza-
 (CST) is to find a spanning
tree problem tion problems.
tree T  V ET  such that ∑e  ET de  D keywords: no-wait flowshop,
and ∑e  ET ce is minimized. We present complexity, batch processing machines,
A branch-and-bound algorithm
an efficient polynomial time approxima- Lagrangian relaxation
to minimize the makespan in a
tion scheme (EPTAS) for this problem. In this paper, we consider a no-wait flowshop with blocking

Specifically, for every ε 0 we present flowshop problem on two sum-batch ma-
an 1  ε  -approximation algorithm with D ÉBORA P RETTI RONCONI
 1 
chines. A machine is called a sum-batch
time complexity O 1ε  O  ε  n4 log2 n   . processing machine, if it can process sev- UNIVERSIDADE DE SÃO PAULO
Our method is based on Lagrangean re- eral tasks sequenctially. The set of tasks keywords: multiple machines,
laxation and matroid intersection. processed together is called a batch. The scheduling, branch-and-bound,
processing time of a batch is equal to the makespan
sum of the processing times of tasks in This paper analyzes the minimization of
Approximation of rectilinear batch. A batch can only begin to pro-
Steiner trees with length restric- the makespan criterion for the flowshop
cessed when all tasks of this batch are problem with blocking. In this environ-
tions on obstacles available on the corresponding machine ment there are no buffers between suc-
M ATTHIAS and all tasks assigned to the same batch cessive machines, and therefore interme-
M UELLER -H ANNEMANN start and completed at same time. A setup diate queues of jobs waiting in the sys-
Research Institute for Discrete time s (depending only on machine) is tem for their next operations are not al-
Mathematics, Bonn required on such a machine before each lowed. While there is a considerable
execution of a batch. It correspondins to amount of research that deals with flow-
coauthor: Sven Peyer tool changes or to machine cleaning and
keywords: rectilinear Steiner trees, shop with no storage constraints, few
adjustment. On the second machine, we works have addressed the flowshop with
obstacles, VLSI design, approximation study the case in which the setup time is
algorithms blocking in process. The particular case
processed with anticipation (as) and the of three machines, which is NP-hard in
We consider the problem of finding a case in which the setup is processed with- the strong sense for such criterion, can
shortest rectilinear Steiner tree for a out anticipation(ns). indicate the difficulty level of this prob-
given set of points in the plane in the lem. This paper proposes a lower bound
presence of rectilinear obstacles. The which exploits the occurrence of block-
Steiner tree is allowed to run over obsta-
A branch-and-bound algo-
rithm for the minimization of ing. A branch-and-bound algorithm that
cles; however, if we intersect the Steiner uses this lower bound is described and its
tree with some obstacle, then no con-
weighted tardiness in a two-
machine flowshop efficiency is evaluated on several prob-
nected component of the induced subtree lems. Numerical experiments indicated
must be longer than a given fixed length. B ERTRAND L IN that the bounding scheme works better
This kind of length restriction is moti- National Chi Nan University than the lower bound proposed by Ron-
vated by its application in VLSI design coni and Armentano (2001).
where a large Steiner tree requires the coauthor: Kenny K.Y. Chen
insertion of buffers (or inverters) which keywords: scheduling, flowshop,
must not be placed on top of obstacles. weighted tardiness, branch-and-bound WE2-321/033 MULTICRITERIA OPTIMIZATION

We show that the length-restricted There is a set of jobs to be processed in Efficiency analysis in
Steiner tree problem can be approxi- a two-machine flowshop. In addition to multicriteria optimization I
mated with a performance guarantee of the processing times on two machines, chair : Fernanda Maria Pereira Raupp
110 Parallel Sessions: Wednesday 13:15 – 14:45 (WE2)

A biobjective method for sample solutions for a linear multiobjective prob- R EHA T UTUNCU
allocation in stratified sampling lem, which has many practical applica- Carnegie Mellon University, Dept of
tions in Economy, specially in Industry . Math. Sciences
D OLORES ROMERO M ORALES We build a cone such that every belong- keywords: semidefinite programming
Universiteit Maastricht ing vector gives at least one weakly ef-
coauthor: Emilio Carrizosa ficient solution. We also show that each We discuss our recent work related to
keywords: stratified random sampling, weakly efficient solution corresponds to SDPT3, a MATLAB based software
sample allocation, biobjective integer a vector in such cone. We claim that in for semidefinite-quadratic-linear opti-
program the case that the original problem has a mization. This work includes new rou-
solution, this solution is obtained solving tines for special handling of fixed or free
The two main and contradicting criteria variables, an interface for control prob-
guiding sampling design are accuracy of a linear programming problem, where the
objective function is composed by a vec- lems, and specialized routines for opti-
estimators and sampling costs. In strat- mizing linear functions over an intersec-
ified random sampling, the sample size tor in such cone.
tion, rather than product, of two or more
must be allocated to strata in order to op- cones.
timize both objectives. WE2-321/133 LINEAR PROGRAMMING

In this presentation we address, fol- Computational and WE2-305/205 GAME THEORY

lowing a biobjective methodology, this algorithmic progress in cone

allocation problem. A two-phase method programming Game theory applications I
is proposed to describe the set of Pareto- organizer/chair : Jos F. Sturm organizer/chair : Daniel Granot
optimal solutions of this nonlinear in-
teger biobjective problem. In the first A new variant of the interior On the complexity of the kernel
phase, all supported Pareto-optimal so- point method with rank-1 up- WALTER K ERN
lutions are described via a closed for- dates University of Twente
mula, which enables quick computation.
J OS F. S TURM coauthors: U. Faigle, J. Kuipers
Moreover, for the common case in which
Department of Econometrics, Tilburg keywords: cooperative games, kernel,
sampling costs are independent of the
University complexity
strata, all Pareto-optimal solutions are
shown to be supported. For more general keywords: interior point method We study classes of cooperative games
cost structures, the non-supported Pareto- with the property that the surplusses si j
TBA
optimal solutions are found by solving a (relative to a given game in the class and
parametric knapsack problem. a given allocation x) are efficiently com-
The methods and results obtained are Semidefinite programming in putable. We show that for such classes of
also extended to more general settings, the space of partial positive games also the problem of computing an
such as estimation problems with miss- semidefinite matrices element in the kernel is easy.
ing values and various types of multi-
response sampling.
S AM B URER
University of Iowa Buying and selling strategies in
keywords: semidefinite programming,
the assignment game
Duality in risk minimization interior point algorithms, sparsity M ARILDA S OTOMAYOR
and vector optimization University of São Paulo - Brazil
We build upon the work of Fukuda et
F RANK H EYDE al. and Nakata et al., in which the the- keywords: optimal matching, core
Martin Luther University ory of partial positive semidefinite ma- allocation, competitive price, Nash
Halle-Wittenberg trices has been applied to the semidef- equilibrium
keywords: coherent risk measure, inite programming (SDP) problem as a For a version of the "Assignment mar-
duality, vector optimization technique for exploiting sparsity in the ket" of Shapley and Shubik, the strate-
data. In contrast to their work, which im- gic games induced by a class of market
We investigate duality in minimization of proves an existing algorithm that is based
coherent risk measures as introduced by on the HRVW/KSH/M search direction, clearing price mechanisms are analyzed.
Artzner/Delbaen/Eber/Heath and point we present a primal-dual path-following In these procedures, buyers and sellers in
out connections with proper efficient algorithm that is based on a new search different stages reveal their demand and
points of appropriate vector optimization direction, which, roughly speaking, is supply functions, and a competitive equi-
problems. librium rule is used. We obtain a neces-
defined completely within the space of sary and sufficient condition for the exis-
partial symmetric matrices. We show tence of subgame perfect equilibria. Pre-
On minimization over weakly that the proposed algorithm computes a cise answers can be given to the strategic
efficient sets primal-dual solution to the SDP prob- questions raised.
lem having duality gap less than a frac-
F ERNANDA M ARIA P EREIRA
 ε  0 of the initial duality gap in
tion
R AUPP O n log ε  1   iterations, where n is the Chinese postman games on a
LNCC size of the matrices involved. More- class of Eulerian graphs
coauthor: Wilfredo Sosa over, we present computational results DANIEL G RANOT
keywords: weakly efficient points, showing that the algorithm possesses sev- University of British columbia
multicriteria optimization, global eral advantages over other existing im-
optimization plementations. coauthors: Herbert Hamers, J. Kuipers,
M. Maschler
We are interested in minimizing a linear keywords: Eulerian graphs, cooperative
function over the set of weakly efficient New developments in SDPT3 games, core, nucleolus
 111

The Extended Chinese Postman (ECP) players reside. The issue is to allocate linear time for an ECP enterprize induced
enterprize is induced by a connected and the net cost of travelling among all play- by a simple cycle or a chain, and we
undirected graph with a special vertex, ers. We show that for ECP enterprizes develop a quadratic algorithm to calcu-
called the post office, wherein a server induced by Eulerian graphs having the 4- late the nucleolus of such an enterprize.
is located. Each edge has a prize, which cut property, the core as well as the nucle- These results imply that we can check
can be collected only once, and a cost at- olus are the Cartesian products of cores core membership in linear time and we
tached to it, and each player resides in an and nucleoli of ECP enterprizes associ- can calculate the nucleolus in quadratic
edge. The server has to travel along the ated with simple cycles derived from the time for ECP enterprizes induced by Eu-
edges, starting and returning to the post original graph. We further demonstrate lerian graphs having the 4-cut property.
office, and visiting all edges in which that one can check core membership in
112 Parallel Sessions: Thursday 9:00 – 10:30 (TH1)

TH1-302/41 COMBINATORIAL OPTIMIZATION I SABEL M ENDEZ -D IAZ

Universidad de Buenos Aires On semidefinite relaxations for
Computational methods for
graph-coloring II coauthor: Paula Lorena Zabala the linear ordering problem
keywords: graph-coloring,
organizer/chair : Michael Trick branch-and-cut, facets A LANTHA N EWMAN
Graph-coloring heuristics based We present a Branch-and-Cut algorithm MIT
on lovasz theta number for the graph coloring problem. In a pre- keywords: linear ordering problem,
vious work, we studied the facet structure maximum acyclic subgraph, semidefinite
I GOR D JUKANOVIC of the 0/1-polytope associated with an programming, approximation algorithms
University of Maribor / School for integer programming formulation of the
Economics and Business graph coloring problem. Based on these The linear ordering problem is easy to
coauthor: Franz Rendl theoretical results a Branch-and-Cut al- state: Given a complete weighted di-
keywords: graphs, heuristics, vertex gorithm is developed. Our computational rected graph, find an ordering of the ver-
coloring, SDP experiences compare favorably with the tices that maximizes the weight of the
well-known exact graph coloring algo- forward edges. Although the problem is
SDP formulation of Lovasz theta num- NP-hard, it is easy to estimate the op-
rithm DSATUR.
ber does not provide us only with a lower timum to within a factor of 1/2. (Take
bound on the chromatic number of a any ordering of the vertices; either the set
graph, but also with a matrix which ap- TH1-302/42 COMBINATORIAL OPTIMIZATION

of forward edges or the set of backward

proximately describes the optimal color- Combinatorial optimization edges accounts for at least half of the to-
ing. This matrix is used to define sev- VII tal edge weight.) It is not known whether
eral recursive heuristics. Tighter SDP re- chair : Marco Lübbecke the maximum can be estimated to a bet-
laxations for the chromatic number then ter factor using a polynomial-time algo-
theta number may also be employed. A fast parametric assignment
algorithm with applications in rithm.
max-algebra Recently it was shown that widely-
Cliques, holes and lower bounds studied polyhedral relaxations for this
for the vertex coloring problem B ETTINA K LINZ problem cannot be used to approximate
TU Graz the problem to within a factor better than
R ICARDO C ORRÊA coauthor: Elisabeth Gassner 1/2. This was shown by demonstrating
Federal University of Ceará keywords: networks/graphs, matchings, that the integrality gap for these relax-
coauthors: Manoel Campêlo, Yuri Frota parametric programming, max-algebra ations is 2 on certain classes of random
keywords: graph-coloring, integer graphs. We present a new semidefinite
This talk deals with parametric assign- relaxation for the linear ordering prob-
programming, facets of polyhedra ment problems and their application to
 lem and show that the gap is strictly less
Given a graph G  V E  , where V is its max-algebra where “max´´ plays the role than 2 on the classes of random graphs
set of vertices and E, its set of edges, and of “ 
´´ in linear algebra and “  ´´ the role referred to above.
a positive integer k, a k-coloring of “ ´´.  The max-characteristic polyno-
 of G is
an assignment of colors from 1


k
mial ξA λ  of an n  n matrix A serves
to the vertices of G so that each vertex re- a similar role in max-algebra than the On the stabbing number of
ceives at least one color and the endpoints characteristic  polynomial in linear alge- matchings, trees, and triangula-
of any edge are assigned different colors. bra. ξA λ  can be computed by com- tions
The vertex coloring problem is defined puting the optimal value function of a

as the problem of finding the minimum nonlinear parametric assignment   prob- M ARCO L ÜBBECKE
number  of colors χ G  such that G ad- lem with cost  matrix B λ   bi j  where Braunschweig University of Technology
mits a χ G  -coloring. Observe that there bii  max aii λ
and bi j  ai j for i  j. coauthors: Sandor Fekete, Henk Meijer
exists a coloring of G with χ G  colors We start with investigating the fol-
keywords: perfect matching, line
in which each vertex is assigned exactly lowing linear parametric assignment stabbing number, NP-hardness, integer
one color. problem (PAP):  We are given a bipartite
graph G  V1 V2 E  , a subset P  E programming

In this paper we propose a new 0-1

formulation of the vertex coloring prob- and edge weights ci j for i j   E.  The Given a planar point set we investi-
lem. We show that certain particular sub- parametric weight di j λ   of edge i j   gate perfect matchings, spanning trees,
structures of G induce classes of facet E is given by ci j  λ if i j   P and by and triangulations of minimum stabbing
defining inequalities of the associated ci j otherwise. The PAP is to find a perfect number. The stabbing number of a
polytope, particularly maximal cliques matching with maximum weight with  re- matching (tree, triangulation) is the max-
and odd hole and anti-holes. Some com- spect to the parametric weights d i j λ  for imal number of intersections of any line
putational experiments were carried out all values of λ. We present  an algorithm with matching edges (tree edges, trian-
in order to evaluate the lower bound pro- which solves the PAP in O mn  n logn 
2
gles). A small stabbing number is a de-
vided by this formulation, and their re- time for a graph with 2n vertices and m sirable property of data structures for lo-
sults are presented. A comparison of the edges. cation queries in the widest sense. We
results obtained are made with those of We then show how this algorithm can prove that minimizing the stabbing num-
the column generation approach from the be extended to solve the nonlinear para- ber is NP hard for all three problems.
literature. metric assignment problem  which arises We present integer programming formu-
in the computation  of ξA λ  . In this man- lations and evaluate their performance
ner we obtain an O n3  algorithm for this computationally. We outline how to ob-
Application of mathematical task which improves upon the previously tain approximation results from the LP
programming to graph-coloring best-known algorithm by a factor of n. relaxation.
 113

TH1-302/44 COMBINATORIAL OPTIMIZATION TAKEAKI U NO discriminant rule generated. Computa-

National Institute of Informatics tional tests performed on several well-
Counting problems known benchmark datasets indicate that
keywords: maximal clique,
chair : Takeaki Uno enumeration, algorithms, bipartite our algorithm consistently outperforms
An algorithm for enumerating other classification approaches in terms
A clique of a graph is a vertex subset such of accuracy, and is therefore capable
all substitutions a rooted or- that any its two vertices are connected by
dered tree into a term tree of good generalization on future un-
an edge. A bipartite clique of a graph is seen data. Further testing on marketing
YASUKO M ATSUI a pair of vertex subsets such that any ver- datasets of realistic size have proven the
Dept . of Mathematical Sciences, Tokai tex of a vertex set of them is connected to applicability of our algorithm to massive
University any vertex of the other. In this talk, we classification tasks.
coauthors: Satoshi Matsumoto, propose new algorithms for enumerating
Tetsuhiro Miyahara, Takayoshi Shoudai cliques and bipartite cliques of a given
keywords: graphs, enumeration, graph G  V E  These problems can be So it’s facet-defining... big, fat,
dynamic programming      however
solved by an existing algorithm, hairy deal!
We consider an ordered tree with internal the algorithm takes O V E  time per
structured variables, called a term tree. output clique. The time  4 complexities of DAVID WARME
A term tree is a rooted tree that consists our algorithms are O ∆  per clique and
L-3 Communications Analytics Corp.
of tree structures with ordered children O ∆3  per bipartite clique, respectively,
and internal variable, and it is suited for where ∆ is the maximum degree of G  keywords: facets
representing tree-structured data such as We show some result of computational
HTML files. In learning theory, there are experiments for both random instances Decades of practical experience with
many problems on term trees. In this talk, and real world instances, and show that cutting-plane algorithms has taught us
we deal with a problem on a term tree. the practical computation time of our al- that “all inequalities are not created
Let a rooted ordered tree T be a term tree gorithm is proportional to the square of equal.” Certain inequalities slash away
such that T does not including internal the average degree, which is quite less major portions of the LP/IP gap, while
variable. We define a replacing operation than ∆4  By using this algorithm, several thousands of others collectively produce
is as follows: (1)A variable of a term tree huge enumeration problems taken from very little. (Quite fortunate then that
is replaced by an edge. (2)A variable of application areas can be solved in a suffi- we call them “inequalities”!) It is
a term tree is replaced by a subgraph of ciently short time. well known that facet-defining inequali-
a rooted ordered tree. Then, a term tree ties generally outperform those that are
T1 and a rooted ordered tree T2 are given, TH1-302/45 INTEGER AND MIXED INTEGER
merely valid, but it is less well appre-
we consider a problem for enumerating
PROGRAMMING ciated how wide the variance in gap-
all substitutions T2 into T1 where the re- Integer and mixed integer closing performance can be even within
placing operation is permitted and an or- programming V a single family of facet-defining inequal-
der relation of all leaves on T1 satisfy an chair : Atsushi Kawamoto ities. How do we design separation pro-
order relation of them on T2 We propose On classes of mixed integer pro- cedures that are effective — i.e., that
an algorithm with polynomial delay for gramming problems arising in produce not just violations, but very
the above problem using a dynamic pro- discriminant analysis strong violated inequalities that converge
gramming technique. rapidly?
C ARLO V ERCELLIS We address this question by present-
Counting the vertices of certain Politecnico di Milano ing several metrics based on intuitive
classes of polyhedra properties that “strong” inequalities are
coauthor: Carlotta Orsenigo likely to possess. We apply two of these
S AMMANI DANWAWU keywords: classification, segmentation, metrics analytically to several families of
A BDULLAHI support vector machines, heuristics inequalities, including subtour inequal-
UAE University We consider the new concept of dis- ities for both the TSP and the span-
coauthors: Martin Dyer, Les Proll crete support vector machines for solv- ning tree in hypergraph polytopes, and
keywords: counting, approximation, ing classification problems in discrimi- the 3-toothed comb inequalities for the
randomization nant analysis. It is a discrete version of TSP. We present computational evidence
The problem of counting the vertices of support vector machine (SVM) in which that corroborates the predictive power of
arbitrary polyhedra has various compu- the error is properly expressed as the these metrics, and we show how such
tational applications. The problem is count of misclassified instances, in place metrics can be used both to design more
known to be #P-complete. We discuss of the misclassification distance consid- effective separation procedures, and to
different fully polynomial randomized ered by traditional SVMs. This results theoretically justify existing “constraint
approximation schemes (fpras) for count- in a mixed integer programming prob- strengthening” methods.
ing the vertices of polyhedra associated lem whose properties are investigated,
with;0-1 Permanent, 0-1 Knapsack, 2 x and whose approximate solution is effi-
n transportation problems, matroids and ciently obtained by means of a sequential Articulated mechanism designs:
matchings in non-bipartite graphs. We LP-based heuristic. Alternative approxi- topology and geometry varia-
show that the vertices of a polyhedron as- mation algorithms based upon truncated tions
sociated with regular matroids can be ex- branch and bound and tabu search meth-
actly counted. ods are also discussed and compared. ATSUSHI K AWAMOTO
Further, an extension of the model is ob- MAT/DTU
Fast algorithms for enumerat- tained by introducing an additional term keywords: mechanism design, topology
ing maximal cliques in order to reduce the complexity of the
optimization, geometry optimization
114 Parallel Sessions: Thursday 9:00 – 10:30 (TH1)

This paper deals with the design of ar- of the simplex or proves no such point ex-
ticulated mechanisms using a truss-based The integral basis method and ists. Efficiency of the approach is demon-
ground-structure representation with ge- extensions strated through numerical results.
ometric non-linearity taken into consid-
eration. It is essential to consider both M ATTHIAS KOEPPE
topology and geometry variations for ob- University of Magdeburg/ FMA-IMO TH1-306/31 INTEGER AND MIXED INTEGER
PROGRAMMING

taining an effective mechanism, and this coauthors: Robert Weismantel, Quentin Computational methods for
research is an initial attempt at develop- Louveaux solving IPs II
ing a systematic method for optimizing keywords: primal methods, integral organizers : Eva Lee and Andrew
both topology and geometry of mecha- bases Miller
nisms. This can be performed sequen- The Integral Basis Method by Haus, chair : Eva Lee
tially or simultaneously and takes into ac- Köppe, Weismantel (2001) is an exact
count design requirements such as max- Strong formulations and sepa-
primal-type algorithm for solving inte- ration for multi-level lot-sizing
imization of the output displacement or ger linear programs. It is based on the
generation of a specified path. Re- problems
computation of integral generating sets of
search on design of compliant mecha- cones and polyhedra. Using these gen-
nisms using topology optimization tech- K EREM A KARTUNALI
erating sets, the integer program to be University of Wisconsin-Madison
niques in continuum structures suggests solved is reformulated in an iterative way,
that it should also be possible to obtain until augmentation vectors are found or coauthor: Andrew Miller
results for articulated mechanisms by ap- the optimality of the current solution is keywords: lot sizing, combinatorial
plying such techniques. On the other proved. optimization, integer programming,
hand, dimensional synthesis techniques In this talk we present extensions of production planning
based on optimization methods can pro- the algorithm to mixed integer programs.
vide a broad range of tools especially Much of the difficulty in solving practical
We introduce a variant of the method that lot-sizing problems arises because strong
for mechanism path generation. How- uses a different kind of discrete relax-
ever, this method often encounters non- formulations for the underlying multi-
ation of integer programs. Moreover, we level problems are usually not used. Such
assembly problems during the optimiza- present new computational results for the
tion process because the method employs problems have been studied by Afentakis
augmentation phase of the optimization and Gavish, Tempelmeier and Derstroff,
rigid bars and their use as design vari- process of hard integer programs from
ables involves the process of assembling and Stadtler, among others. We discuss
the MIPLIB. computational results obtained by us-
of the mechanism in all possible configu-
rations. The proposed method with truss ing strong reformulations, including us-
representation uses cross-sectional areas A simplicial approach to integer ing efficient new methods to separate for
and nodal positions as design variables, programming strong valid inequalities.
and thus assembly is unnecessary during C HUANGYIN DANG
the optimization process. City University of Hong Kong Lifting, superadditivity and sin-
coauthor: Hans van Maaren gle node flow sets revisited
TH1-302/49 INTEGER AND MIXED INTEGER
PROGRAMMING keywords: integer programming,
Alternative methods simplicial approach, integer point, Q UENTIN L OUVEAUX
polytope CORE, Universite catholique de Louvain
chair : Chuangyin Dang
Integer programming is equivalent to de- coauthor: Laurence A. Wolsey
An augmentation framework termining whether there exists an inte- keywords: mixed integer programming,
for integer programming ger point in a polytope. By increas- lifting, single node flow sets
ing dimension, we show that determin- This talk focuses on the lifting of valid
U TZ -U WE H AUS ing whether there exists an integer point
University of Magdeburg/ FMA-IMO inequalities for mixed integer programs.
in a polytope is equivalent to determin- We review the work of Gu, Nemhauser
coauthor: Robert Weismantel ing whether there exists an integer point and Savelsbergh on the lifting of valid in-
keywords: lattice reduction, in a simplex. Based on an integer la- equalities for single node flow sets and
augmentation algorithm beling rule and the D1-triangulation of try to give a general presentation of it. In
A number of practical algorithmic build- the space, we develop a simplicial ap- particular, we show that the lifting can al-
ing blocks that combine into an augmen- proach to determining whether there ex- ways be carried out by fixing a set of vari-
tation framework for solving general in- ists an integer point in a simplex in the ables to 0, after an appropriate transfor-
teger programs are presented. The main canonical form, which is obtained from mation of the variables. We show how to
ingredients are the generation of augmen- the application of a unimodular trans- apply this theory to both particular cases
tation vectors using Lovász’s lattice re- formation. The approach consists of and generalizations of single node flow
duction and the Integral Basis Method, two phases. The first phase forms a sets. Special empahasis is placed on su-
which are used to find an initial feasi- homotopy-like simplicial method, which peradditive lifting that allows one to si-
ble solution, augment to an optimal point, generates (n+1)-dimensional simplices, multaneously lift different sets of vari-
and prove optimality. Other techniques and the second phase forms a piv- ables.
can be integrated seamlessly at any point, oting procedure, which generates n-
if additional information (e.g. combina- dimensional simplices. Starting at an ar-
torial structure) for the problem is avail- bitrary integer point on the artificial level, Problems of optimization: an
able. Computational experience on prob- the approach interchanges between those exact algorithm for finding a
lems from the MIPLIB library is reported two phases and follows a finite simplicial maximum clique optimized for
upon. path that either leads to an integer point the dense graphs
 115

D ENISS K UMLANDER (MMA) which show that a substantial re- smooth transition of optical power across
Tallinn Technical University duction in total mass of the rotor can be the PAL. Unfortunately, some optical er-
keywords: maximum clique, achieved. ror, know as astigmatism , is unavoid-
independent sets, dense graph able for this kind of lenses and compli-

In this paper we present an algorithm

A food problem for a predator- cates the design of the PAL. Usually, one
prey interaction model tries to get minimal astigmatism at the far
that solves one of the problems arising and near vision area and on a central zone
from the optimization task, namely find- O LEG M AKARENKOV of progression which connects both ar-
ing maximum clique from an arbitrary Department of Mathematics, Voronezh eas. The non-avoidable astigmatism has
graph. The optimization problem is a State University to be reduced as far as possible and to
common problem in many areas includ-
keywords: nonlinear programming, cost be placed near the boundary of the lens.
ing theory of programming. Our algo-
analysis, predator-prey interaction mode, This talk will first consider the PAL de-
rithm is an optimization for dense graphs sign when spherical coordinates are used
periodic solutions
of one of the quickest algorithms for in the modelling of the PAL. We will
a maximum clique finding - Carraghan Consider the following predator-prey in- compare the optimization problem and
and Pardalos algorithm. An optimiza- teraction model its solution in spherical coordinates with
tion effect starts to appear from graphs

k
Ṅ1 k1 N1 k  2N N1 N2 k3 N12 the classical frame using cartesian coor-
with average density and reach maximum 
0 1
εa1 F1 t  N1  N2 N1    εan Fn t  N1  N2 N1 


dinates. A new approach for solving the

on graphs with high density. On dense

k
k5 N2  k  4N N1 N2 PAL problem which combines nonlinear

Ṅ2



0 1
graphs maximum clique is found up to εb1 F1 t  N1  N2 N2    εbn Fn t  N1  N2 N2 

optimization and perturbation techniques

1000 times faster than for initial algo- will then be proposed. We conclude by
rithm. where N1 and N2 are the prey and preda- testing the computational effectiveness of
tor species concentrations, Fi is the quan- the new suggested strategies.
tity of the food number i the species
TH1-306/32 NONLINEAR PROGRAMMING
eat during the unit time and the coef-
Applications ficients k0    k4 εa1    εan εb1    εbn TH1-306/33 NONLINEAR PROGRAMMING

are positive. Assume that the parame-

ters k0    k4 satisfy the condition under PDE-constrained optimization
chair : Katia Demaseure
Shape optimization in rotor- which the system () with ε  0 has a limit (1)
dynamics with eigenvalue and cicle Ñ and denote by T it’s period. Then, organizers : Michael Ulbrich and
eigenvector constraints we assume that the functions F1    Fn are Stefan Ulbrich
T -periodic with respect to the first vari- chair : Stefan Ulbrich
F RANK S TRAUSS able. Finally, we consider a1  0 and the
University of Heidelberg parameter ε 0 small. Introduce the cost Optimization of time-dependent
function partial differential equations
coauthors: Jens Starke, Mizuho Inagaki
keywords: finite element model, f a 1      a n  b1      b n a 1  b1 c 1      a n  bn c n 
M ATTHIAS H EINKENSCHLOSS

sensitivity analysis, shape optimization,

where ci 0 express the price of the Rice University
eigenvector constraints
food number i. In the present report we keywords: dynamic programming,
We study design optimization problems solve the following problem: To mini- optimal control, nonlinear programming
for general rotating bodies including mize the function f on the set of such
gyroscopic effects and consider a tur- a1    an and b1    bn that the system () The size and structure of optimization
bocharger as example. The target of our has an assimptotically stable T -periodic problems governed by time-dependent
research is to reduce rotor weight and ro- solution Ñε and Ñε
Ñ geometrically as partial differential equations (PDEs)
tor vibrations leading to higher efficiency ε
0  The work supported by RFBR, makes the direct application of existing
and reliability and less noise. This can be projects 02-01-00189, 02-01-00307 and optimization algorithms to the fully dis-
achieved by minimizing the total mass of 03-01-06308. cretized PDE constrained problem or the
the rotor while reducing the vibration re- adapt ion of optimization methods for
sponse of the first bending mode in the ODE constrained problems impractical
resonance case below a certain thresh- The progressive addition lens for realistic problems.
old and at the same time increasing the problem : an hybrid approach In this talk we discuss some of the is-
critical speed on a given level. In our combining optimization and sues arising in the optimization of time-
model the rotor is described by beam el- perturbation techniques dependent PDE constrained optimiza-
ements for the shaft and two rigid disks K ATIA D EMASEURE tion problems. We present our solu-
for impeller and turbine. Design vari- FUNDP tion approach, which is based on an in-
ables are the diameters of the beam el- between formulation of the infinite di-
ements as well as mass and moments coauthor: Annick Sartenaer mensional problem, solved using sequen-
of inertia of the rigid disks. This leads keywords: progressive addition lenses, tial quadratic programing methods.
to an optimization problem with eigen- nonlinear optimization, perturbation
value and eigenvector constraints. Spe- tecnhiques
cial attention is paid to differentiability Progressive Addition Lenses (PAL for Optimal design of groundwater
aspects and a fast and precise calcula- short) represent the best solution for pres- remediation systems
tion of the eigenvector derivatives. We byopia. A PAL has a far vision area with
will present numerical results obtained low refractive power in the upper part, C ARL K ELLEY
by sequential linear programming (SLP), and a near vision area with higher optical North Carolina State University
sequential quadratic programming (SQP) power in the lower part. It is important keywords: optimal design, hydrology,
and the method of moving asymptotes for the visual quality of the lens to have implicit filtering
116 Parallel Sessions: Thursday 9:00 – 10:30 (TH1)

In this talk we discuss several problems keywords: interior point method, trust H ANS D. M ITTELMANN
in optimal design of subsurface reme- region method, Newton method, Arizona State University
diation and flow control systems. The primal-dual method keywords: optimization software, conic
objective functions and constraints are programming
We present a new methodology for solv-
constructed from PDE solvers for the ing general nonlinear programs. We pro- For the Seventh DIMACS Implementa-
flow and species transport equations. We pose the use of interior-point methodol- tion Challenge in SDP&SOCP we had
show how to implement problems us-
ogy, trust-region globalization strategies evaluated all ten submitted codes. The
ing standard production codes, how these and the conjugate gradient method to find results appeared in early 2003 in Math-
codes and the underlying physics present a solution to large-scale problems. As a ematical Programming B. Several of the
obstacles to the optimization algorithms,
central region to guide the iterates to an codes have not been updated since. The
and how sampling methods, in particu- optimal solution of the problem, we use others, however, are under development.
lar implicit filtering, can overcome these the notion of the quasi-central path intro- As part of our ongoing benchmarking ef-
problems.
duced by Argaez and Tapia. fort we are evaluating those, especially
The objective functions and con-
on large and/or sparse problems. Several
straints in these problems are typically
non-smooth in all the design parameters An inexact Newton trust region authors have been using our benchmark
and discontinuous in some of them. We interior-point algorithm for problems to improve their codes. We will
will discuss the physical reasons for these nonlinear programming prob- report on the current state.
properties and how optimization meth- lems
ods can deal with them and compare the Functional versus conic opti-
results from several optimization algo- JAIME , J R . H ERNANDEZ mization: what is the better
rithms. University of Texas at El Paso
E RLING A NDERSEN
coauthors: Miguel Argaez, Leticia MOSEK ApS
Velazquez
Multilevel and preconditioning keywords: interior point, conjugate keywords: quadratic cone, conic
approaches for inequality con- gradient, trust region optimization, convex, convex quadratic
strained optimization problems constraints
governed by PDEs We present an algorithm for solving large Convex optimization problems can be
scale nonlinear programming problems. formulated and solved on either func-
M ICHAEL U LBRICH We use interior-point methodology, a tional or conic form. In this talk we will
Fachbereich Mathematik, Universitaet trust region globalization strategy, and discuss the advantages and disadvantages
Hamburg the conjugate gradient method with Stei- of the two approaches.
keywords: optimal control, nonlinear haug’s ending conditions. Some prelimi- Finally using the software pack-
programming, multilevel methods, nary numerical results are presented. age NOSEK we will present computa-
preconditioning tional results comparing whether con-
The numerical solution of PDE- When sparsity counts: optimal vex quadratically constrained optimiza-
constrained optimization problems re- direct Jacobian computation tion problems should be solved in in their
quires efficient solvers for linear or non- traditional functional form or on conic
linear equations of huge size. These sys- T ROND S TEIHAUG form.
tems are usually structured and sparse, University of Bergen
but due to their size they cannot be coauthor: Shahadat Hossain Automated performance analy-
solved by direct methods. We discuss keywords: nonlinear programming, sis in the evaluation of nonlinear
reformulation approaches and interior- automatic differentiation programming solvers
point techniques that are well suited for When Coleman and More in 1982 ana-
PDE-constrained optimization problems lyzed the Curtis-Powell-Read technique A RMIN P RUESSNER
with inequality constraints. It is shown to estimate an matrix they showed that GAMS Development Corp.
how these methods can be combined this is equivalent to a graph coloring coauthor: Hans D. Mittelmann
with inexact iterative solvers that are problem. In the same paper they show keywords: performance analysis,
preconditioned by multilevel and other that the CPR technique is a direct deter- automation server, visualization,
techniques. The proposed methods are mination method, but they also give an optimization software
applied to the solution of optimization example showing that the CPR technique As part of Performance World, we de-
problems in engineering and finance. is not equivant to direct determination. scribe an automation server PAVER:
In this paper we characterize optimal di- http://www.gamsworld.org/performance/paver
TH1-306/34 NONLINEAR PROGRAMMING
rect determination and give an algorithm
Large scale problems I for computing the minimum number of to help facilitate reproducible perfor-
function evalutions to compute a Jaco- mance testing and benchmarking of op-
chair : Trond Steihaug bian matrix with direct determination. timization software. Our online server
automates the task of performance analy-
On the development of a trust- sis and visualization, taking into account
region interior-point method for TH1-306/35 NONLINEAR PROGRAMMING

various performance metrics, including

large-scale nonlinear programs NLP software - the state of the not only robustness and efficiency but
art II also quality of solution. We discuss the
C RISTINA V ILLALOBOS
The University of Texas-Pan American organizer/chair : Hans D. Mittelmann tools and the performance metrics used
and give benchmark examples using an
coauthors: Miguel Argaez, Leticia The state of the art in software instance of the COPS test case for non-
Velazquez for SDP&SOCP problems linear programming.
 117

TH1-306/36 NONSMOOTH OPTIMIZATION Furthermore, we use only few vectors to programs introduced by Ben-Tal and Ne-
represent the variable metric approxima- mirovski. In contrast to this previous
Bundle-based methods tion of the Hessian and, thus, we avoid work, where it is shown that such prob-
chair : Marko M. Makela storing and manipulating large matrices lems can be formulated as semidefinite
An infeasible bundle method for as is the case in variable metric bundle programs, our focus is to identify un-
nonsmooth convex constrained methods. These improvements make the certainty sets that allow this class of
optimization without a penalty limited memory bundle method suitable problems to be formulated as second-
function or a filter for large-scale optimization: The num- order cone programs. We propose three
ber of operations needed for the calcu- classes of uncertainty sets that satisfy this
C LAUDIA A LEJANDRA lation of the search direction is only lin- criterion and discuss examples from fi-
S AGASTIZÁBAL early dependent on the number of vari- nance, classification, and signal process-
IMPA ables while, for example, with variable ing where these classes of uncertainty
coauthors: Pablo A. Rey, Mikhail metric bundle methods this dependence sets are natural.
Solodov is quadratic. The efficiency of the new
keywords: bundle methods method is shown by giving some results
Robust optimization approach
from the numerical experiments. to multi-stage uncertain deci-
Global convergence in constrained op-
timization algorithms has traditionally sion problems
been enforced by the use of penalty A hybrid of simulated annealing A HARON B EN -TAL
functions. Recently, the filter strategy and proximal bundle method for Technion-Israel Institute of technology
has been introduced as an alternative, continuous global optimization
with the motivation to alleviate the need keywords: robust optimization,
for estimating a suitable penalty param- M ARKO M. M AKELA multi-stage decision problems
eter (which is often problematic). In University of Jyvaskyla, Department of We use the Robust Optimization(RO)
this paper, we demonstrate that the use Mathematical Information Technology approach to solve multi-stage decision
of a parametrized penalty function in coauthors: Kaisa Miettinen, Heikki problems with uncertain data. New fea-
nonsmooth convex optimization can be Maaranen tures of the RO approach are the use
avoided without using the relatively com- keywords: global optimization, hybrid of adjustable variables (recourse actions)
plex filter methods. In particular, we pro- methods, metaheuristics and the employment of linear decision
pose an approach which appears to be We introduce several hybrid method for rules,to obtain comutationally tractable
more direct and easier to implement, in global continuous optimization. They convex programs. The methodology
the sense that it is closer to the well- combine simulated annealing and a local is illustrated by studying a problem in
developed unconstrained bundle meth- proximal bundle method. Traditionally, supply-chain management.
ods. Preliminary computational results the simplest hybrid of a global and a lo-
are also reported. cal solver is to call the local solver after Robust discrete optimization
the global one. This does not necessar- and network flows
Large-scale nonsmooth opti- ily produce good results. Besides, using
mization: new limited memory a gradient-based solver implies that the
M ELVYN S IM
bundle method MIT & NUS Business School
hybrid can only be applied to differen-
tiable problems. We show several ways coauthor: Dimitris Bertsimas
M ARJO H AARALA how to integrate the local solver as a gen- keywords: robust optimization, discrete
University of Jyvaskyla
uine part of simulated annealing to en- optimization, stochastic optimization
keywords: nondifferentiable able both efficient and reliable solution We propose an approach to address data
programming, large scale optimization, processes. Using the proximal bundle uncertainty for discrete optimization and
bundle methods, limited memory method as a local solver makes it possible network flow problems that allows con-
methods to solve even nondifferentiable problems. trolling the degree of conservatism of the
Many practical optimization problems We present results of numerical experi- solution, and is computationally tractable
involve nonsmooth (that is, not necessar- ments giving evidence of the applicabil- both practically and theoretically. When
ily differentiable) functions of hundreds ity of the hybrids. both the cost coefficients and the data
or thousands of variables. In such prob- in the constraints of an integer program-
lems, the direct application of smooth TH1-306/37 CONVEX PROGRAMMING ming problem are subject to uncertainty,
gradient-based methods may lead to a we propose a robust integer program-
failure due to the nonsmooth nature of the Robust optimization ming problem of moderately larger size
problem. On the other hand, none of the organizer/chair : Garud Lyengar that allows controlling the degree of con-
current nonsmooth optimization methods servatism of the solution in terms of
is efficient in large-scale settings. In this
Robust quadratically con-
presentation we introduce a new limited
strained quadratic program- probabilistic bounds on constraint vio-
lation. When only the cost coefficients
memory variable metric -based bundle
ming are subject to uncertainty and the prob-
method for nonsmooth large-scale opti- G ARUD LYENGAR lem is a 0-1 discrete optimization prob-
mization. Columbia University lem on n variables, then we solve the ro-
The new limited memory bundle bust counterpart by solving at most n  1
method uses limited memory approach coauthor: Don Goldfarb instances of the original problem. Thus,
to calculate the search direction. Thus, keywords: robust optimization, QCQP the robust counterpart of a polynomially
the time-consuming quadratic direction In this talk we will dicsuss robust con- solvable 0-1 discrete optimization prob-
finding problems appearing in standard vex quadratically constrained programs, lem remains polynomially solvable. We
bundle methods need not to be solved. a subset of the class of robust convex also show that the robust counterpart of
118 Parallel Sessions: Thursday 9:00 – 10:30 (TH1)

an NP-hard α-approximable 0 1 dis- programming techniques, a study of con-

crete optimization problem, remains α- vex extensions enables us to construc- Generalized vector variational
approximable. Finally, we propose an tively determine the convex and concave and quasi-variational inequali-
algorithm for robust network flows that envelopes of many classes of nonlinear ties with operator solutions
solves the robust counterpart by solving a functions. We define a polyhedral basis
polynomial number of nominal problems as a set of convex extensions defined over S ANGHO K UM
on a modified network. a disjunctive set. We study conditions un- Department of Mathematics Education/
der which a polyhedral basis is composed Chungbuk National University
TH1-306/38 GLOBAL OPTIMIZATION
of polynomial functions. Then, we de-
velop polyhedral basis for certain classes coauthor: Won Kyu Kim
Global optimization III keywords: variational inequality,
of polytopes. Finally, we investigate ap-
chair : Ivo Nowak plications of polyhedral basis in devel- monotonicity, fixed point, equilibrium
Relaxation of the semicontinuity oping lower bounds for many classes of
for marginal functions global optimization problems. In a recent paper, Domokos and Kolum-
ban gave an interesting interpretation of
V INCENZO S CALZO Rounding and partitioning variational inequalities (VI) and vector
Dipartimento di Matematica e Statistica, heuristics for solving nonconvex variational inequalities (VVI) in Banach
Università di Napoli space settings in terms of variational
minlps inequalities with operator solutions (in
coauthor: Jacqueline Morgan
keywords: value functions, marginal I VO N OWAK short, OVI). They presented a general
functions, maxinf problems, upper Humboldt University Berlin version of Yu and Yao in a Banach space
quasicontinuity coauthors: Hernan Alperin, Stefan as a main application and gave some
Vigerske other applications such as the solvabil-
Let f be an extented real valued func- ity of variational inequalities defined on
tion defined on X  Y , K be a set- keywords: nonconvex programming,
convex relaxation, rounding heuristic, Hausdorff topological vector space,  and

valued function from  X to Y and  let that of variational inequality on L∞ Ω  .
vS x   supy  K  x  f x y  and vI x   mixed integer programming
 In this paper, in the spirit of Domokos
infy  K  x  f x y  . In light of the Berge In this talk, we present solution ap-
and Kolumban, we consider a multi-
Maximum Theorem, when X and Y are proaches to nonconvex mixed integer
nonlinear programs (MINLPs). First we valued version of (OVI), i. e., the case of
topological spaces, the upper semicon- T being multi-valued, which is called the
tinuity of the function f plays a cen- describe decomposition methods for gen-
erating nonlinear convex relaxations and generalized variational inequalities with
tral role for the upper semicontinuity of operator solutions (simply, GOVI). In ad-
the functions vS and vI . More recently polyhedral relaxations. Both relaxations
are constructed via a block-separable re- dition, we introduce the quasi-version of
(see Lignola and Morgan in Optimiza- (OVI) in multi-valued settings, i.e., the
tion (1992) for sequential spaces and in formulation of a given MINLP. Then we
present rounding and partitioning heuris- case of T and A being multi-valued called
J. Math. Anal. Appl. (1990) for topo- the generalized quasi-variational inequal-
logical spaces), weaker sufficient condi- tics for retrieving solution candidates
from a convex relaxation. We report ities with operator solutions (simply, GO-
tions on the function f have been pre- QVI). Some applications are also pro-
sented which also guarantee existence re- numerical results on instances from the
MINLPLib and on an engineering de- vided.
sults for the MaxSup problem (maximize
the function vS ) and the MaxInf problem sign problem using the recently devel-
(maximize the function vI ). Here, with oped software package LaGO. Well-posed vector variational
reference to sequential spaces, we con- inequalities
sider an other property called sequential TH1-308/11 /
GENERALIZED
CONVEXITY MONOTONICITY

upper quasicontinuity and we obtain new Generalized convexity IV A NGELO G UERRAGGIO

existence results for the MaxSup and the organizer : Laura Martein Universita’ Bocconi
MaxInf problems. Among the others, it chair : Nicolas Hadjisavvas
is shown that sequential upper quasicon- keywords: well-posedness, vector
tinuity cannot be later weaken using the Quasimonotone variational in- optimization, vector variational
minimal condition for existence of maxi- equalities inequality, convexity
mum points and that the functions vS and N ICOLAS H ADJISAVVAS The notion of well posedness of a scalar
vI are sequentially upper quasicontinuous University of the Aegean minimization problem dates back to J.
if f is sequentially upper quasicontinu-
coauthor: Didier Aussel Hadamard at the beginning of the past
ous.
keywords: variational inequality, century. It has been revisited by A.
quasimonotone operator, generalized Tykhonov in the early sixties. Both no-
Convex extensions and polyhe- monotonicity tions are now fully accepted in the litera-
dral basis An existence theorem for the Stampac- ture and their relevance (as their motiva-
M OHIT TAWARMALANI chia variational inequality for quasi- tions) is quite evident also because of the
Purdue University monotone multivalued operators is implications on the stability of the prob-
shown, under very weak conditions. No lem. More recently, by means of Eke-
keywords: convexification, disjunctive land variational principle, the same no-
programming, multilinear envelopes, assumption on the existence of inner
points is made. Also, the assumptions tion was linked to variational inequali-
nonlinear programming ties. As far as we know, the vector ex-
of hemicontinuity and compactness of
We recently developed a theory of con- values of the operator are relaxed, even tension is less developed. However some
vex extensions. In this talk, we demon- with respect to known theorems on pseu- definitions have been proposed for a vec-
strate that along with facial disjunctive domonotone operators. tor minimization problem (a nice survey
 119

is due to J. Loridan) and few compar- Value-at-Risk and Conditional Value-at- 5. Extensions to the Algebraic mod-
isons have been made between the defi- Risk constraints. The model is sup- elling languages, such as AMPL,
nitions itselves and their scalar counter- plemented with control limits, which GAMS, MPL, for Stochastic pro-
part. The aim of this talk is to present a are used to signal misdirected parame- gramming.
generalisation of the scalar results to the ter estimates,and correspondingly deter-
vector case. We present two suitable no- mine times for portfolio rebalancing. Op- 6. Process for collecting the models
tions of well-posed vector variational in- timal strategies and control limits are de- from diverse industrial domain.
equalities. Moreover, by means of some termined for a geometric Brownian mo- Structure of session:
vector Ekeland variational principle, we tion asset pricing model with random pa-
prove the relations with well-posedness rameters. The approaches to risk control Short presentation by three to four
of the primitive vector optimization prob- are applied to the fundamental problem plane members (30-45 minutes).
lem. Some examples are provided to il- of investment in stocks, bonds, and cash This will possibly cover the items
lustrate the concepts. over time. 1...6 above.
Open forum and discussions (30-
TH1-308/12 STOCHASTIC PROGRAMMING

Planning of supply chains un- 45 minutes).

Dynamic stochastic der uncertainty: interworking
programming: theory and of stochastic programming and Follow up action plan.
applications simulation Any other business.
organizer/chair : Lisa Korf
A LEXEI G AIVORONSKI
Dynamic stochastic program- Norwegian University of Science and TH1-308/T1 FINANCE AND ECONOMICS

ming: models, information Technology Combinatorial auctions II

structures, and duality keywords: supply chain management, organizer/chair : Susan Powell
L ISA KORF uncertainty, stochastic programming,
Department of Mathematics, University production planning Secure multi-agent dynamic
of Washington programming and its applica-
We consider the problem of supply chain tion to combinatorial auction
keywords: stochastic optimal control, design and related production decision
dynamics, scenario tree, structure of for an enterprise which is engaged in M AKOTO YOKOO
information production of components for a complex NTT
Dynamic stochastic programming mod- product. We are specifically interested in keywords: dynamic programming,
els encompass discrete time stochastic the case when demand for the product is information security, combinatorial
optimal control problems. To be compre- uncertain. Stochastic programming ap- auction, multi-agent systems
hensive, these models ought to include proach based on conbination of simula- Combinatorial auctions have recently at-
formulations in which the controller has tion and optimization is considered and tracted the interests of many researchers
only partial state information, or may re- numerical experiments are reported with due to their promising applications such
spond to information exogenous to the the use of Stochastic Quasigradient meth- as the spectrum auctions recently held
system state. In the stochastic program- ods by the FCC. In a combinatorial auction,
ming setting, this would require the con- multiple items with interdependent val-
sideration of scenario trees in which each ues are sold simultaneously and bidders
time stage contains not only nodes, but TH1-308/13 STOCHASTIC PROGRAMMING

Stochastic programming: a are allowed to bid on any combination of

also the description of how information items.
evolves withi a time stage. A gen- panel discussion
This paper presents a method for im-
eral stochastic programming formulation organizer/chair : Gautam Mitra plementing several secure combinatorial
of stochastic optimal control problems auction protocols based on our newly
is presented that encompasses this en- panel members: Alan King, Julia developed secure dynamic programming
hanced information structure, and a gen- Higle, Robert Fourer, Maarten van der protocol. Dynamic programming is a
eral duality theory is developed for these Vlerk, H. Gassmann, A. Felt, S. Ahmed very effective, widely used technique for
problems. and Gautam Mitra. tackling various combinatorial optimiza-
The purpose of the panel will be to: tion problems, including several types of
Risk control of dynamic invest- combinatorial auctions. We developed
ment models 1. Identify the need for benchmarks two alternative methods for implement-
L EONARD M AC L EAN models for Stochastic program- ing a secure dynamic programming pro-
Dalhousie University ming and Stochastic integer pro- tocol. One method utilizes homomor-
gramming. phic encryption and the other method uti-
coauthors: Yonggan Zhao, William
Ziemba lizes secret sharing techniques. Our se-
2. Discuss processes for collecting cure dynamic programming protocol can
keywords: investment, risk, stochastic, and adding to the bench mark li-
control obtain the optimal solution of a combina-
brary and maintaining a repository. torial optimization problem, i.e., result of
The risk inherent in the accumulation of
a combinatorial auction, without reveal-
investment capital in a dynamic, stochas- 3. Revisit the SMPS representation ing the inputs of the problem, i.e., bid-
tic capital market depends on the true re- standard for Stochastic program- ding prices. We discuss the application of
turns distribution of risky assets, the ac- ming models. the method to several combinatorial auc-
curacy of estimated returns, and the in-
tions, i.e., multiple-unit single-item auc-
vestment strategy. This paper consid- 4. Discuss the role of XML and data
tions, linear-goods auctions, and general
ers risk control in a decision model with exchange across the net.
combinatorial auctions.
120 Parallel Sessions: Thursday 9:00 – 10:30 (TH1)

TH1-308/T2 COMPLEMENTARITY AND VARIATIONAL

INEQUALITIES
of the MPCC by solving these regular-
Winner determination in combi- ized nonlinear programs for a decreasing
Topics in mathematical sequence of t values. In this talk we ex-
natorial auctions: Lagrangean programs with equilibrium
heuristic for a generalization of amine several issues related to the reg-
constraints ularization approach, including the ex-
the WJISP
organizers : Angel-Victor DeMiguel istence of minimizers to the regularized
T HOMAS E LENDNER and Michael Paul Friedlander problem near a B-stationary point of the
Inst. for Business Administration, chair : Michael Paul Friedlander MPCC, variants on the definition of the
University of Kiel A superlinearly convergent regularized nonlinear program, and iden-
keywords: winner determination, interior-point method for tification of the zero components of the
combinatorial auctions, scheduling MPECs two complementary vectors by inspec-
We present an allocation problem an tion of the solution of a regularized non-
point out that combinatorial auctions are A NGEL -V ICTOR D E M IGUEL linear program for small nonzero values
well–suited for solving it. We show that London Business School of t ("finite identification").
the winner determination problem can coauthors: Michael Paul Friedlander,
be modelled as a generalization of the Francisco J. Nogales, Stefan Scholtes
Weighted Job Interval Scheduling Prob- keywords: MPEC, interior point TH1-308/T3 TELECOMMUNICATION NETWORK
DESIGN

lem. Here, positive–weighted jobs with method, convergence

We propose a new interior-point method
Integer programming for the
release and due dates compete for time– design of ring networks
slots. The generalization is based on for mathematical programs with equilib-
budgets which the bidders can addition- rium constraints (MPECs). At each iter- chair : Rosemary T. Berger
ally submit. Since the WJISP is proven ation, a single log-barrier Newton step is
to be np–hard we suggest a Lagrangean derived from a nondegenerate relaxation A column generation approach
Heuristic for obtaining upper and lower of the MPEC. The barrier and relaxation for a network design problem
bounds and give empirical results. parameters are updated simultaneously to
maintain a nonempty strict interior of the I RENE L OISEAU
Branch-and-price and new test- relaxed MPEC at each iteration and in the Depto. de Computación, Facultad de
problems for spectrum auctions limit. We discuss the convergence prop- Ciencias Exactas y Naturales,
erties of the proposed algorithm. Universidad de Buenos Aires
S VEN DE V RIES
Zentrum Mathematik, TU Muenchen Algorithms for solving EPECs coauthors: Nelson Maculan, Gerard
coauthors: Oktay Gunluk, Laszlo Plateau
Ladanyi C HE -L IN S U keywords: column generation, integer
keywords: spectrum auctions, Stanford University programming, network design
combinatorial auctions, keywords: NCP, MPEC, EPEC
branch-and-price An equilibrium problem with equilib- We present a column generation method
When combinatorial bidding is permit- rium constraints (EPEC) is a mathe- for a problem arising in ring architecture
ted in Spectrum Auctions, such as the matical program to find an equilibrium networks, based in a mathematical model
upcoming FCC auction #31, the result- point which simultaneously solves sev- presented in previous work by Lisser,
ing winner-determination problem can be eral MPECs, each of which is param- Maculan, Passini and Brito. Studies
computationally challenging. We present eterized by decision variables of other about network dimensioning and surviv-
a branch-and-price algorithm based on MPECs. EPECs often arise from non- ability focused mainly on ring based and
a set-packing formulation originally pro- cooperative multi-leader-follower games mesh based architectures. Most works
posed by Dietrich and Forrest (2002). where each leader is solving a Stackel- treated both models separately. But sev-
This formulation has a variable for ev- berg game formulated as an MPEC. In eral companies deploy networks based on
ery possible combination of winning bids this talk, we propose a new algorithm rings, and at the same time they have
for each bidder. Our algorithm ex- for solving EPECs by solving a sequence meshed networks in operation or being
ploits the structure of the XOR-of-OR- of nonlinear complementarity problems. built. In large metropolitan centers it is
bidding-language used by the FCC. We Numerical results on randomly generated easy to identify clusters of nodes gener-
also present a new methodology to pro- EPECs will be presented. ating big data traffic among themselves
duce realistic test problems based on If the area under study in already served
the round-by-round-results of FCC auc- Local properties of MPCC by a meshed network it is possible to
tion #4. We generate 2639 test prob- methods deal with the growth of traffic demands
lems which involve 99 items and are sub- DANIEL R ALPH by superimposing self-healing rings on
stantially larger than most of the previ- Cambridge University the existing mesh. The mesh part of the
ously used benchmark problems. Since problem, can be modeled as a capacitated
there are no real-life test problems for coauthor: Stephen Wright multicommodity flow problem with ex-
combinatorial spectrum auctions with the keywords: complementarity constraints, pandable arc capacities. The other way
XOR-of-OR language, we used these test nonlinear programming, constraint the flow can circulate on the net is on an
problems to observe the computational identification additional net of rings where only flow
behavior of our algorithm. Our algo- A regularisation scheme for MPCC, re- between nodes on each ring is allowed.
rithm can solve all but one test problems cently analysed by Scholtes, approxi- The formulation of the problem as integer
within ten minutes, appears to be very ro- mates the original problem by a nonlin- linear programming problem includes a
bust and compares favorably to the natu- ear program parametrized by a nonneg- column associated with each ring. A pro-
ral formulation solved using a commer- ative scalar t. The intent is to find B- cedure for generating these columns is
cial optimization package. stationary points that are local solutions proposed.
 121

the formulations are described, and com- conjugate gradient method with a termi-
An exact algorithm for optimal putational results comparing the two for- nation test proposed by Steihaug. Some
routing and wavelength assign- mulations are reported. Future research preliminary numerical results on large
ment in ring networks directions are discussed. scale problems are reported.

H ERNAN A BELEDO TH1-341/21 INTERIOR POINT ALGORITHMS

Interior-point solution of large-

George Washington University Large scale problems scale entropy maximization
coauthor: Rudy Kusnadi Sultana chair : Michael Alan Saunders problems
keywords: integer programming
Smoothed analysis of condition M ICHAEL A LAN S AUNDERS
In this routing and wavelength assign- numbers of linear programs Stanford University
ment problem, traffic demands on a
ring network must be routed on colored
S HANGHUA T ENG coauthor: John Tomlin
Boston University keywords: convex optimization, interior
paths, where paths that share an edge
must be colored differently and each col- coauthors: John Dunagan, Daniel point method, entropy
ored path can carry one unit of traf- Spielman
We describe the application of a Primal-
fic. Colors here represent wavelengths keywords: smoothed analysis, linear
Dual Separable Convex Objective
in an all-optical communication network programming, interior point algorithms
(PDSCO) method to a class of extremely
based on wavelength division multiplex- For any linear program, we show that large entropy maximization problems.
ing (WDM). The optimization problem the slight random perturbation of that lin- Separability allows the Newton direc-
seeks to minimize the number of required ear program has small condition num- tions to be computed from (even larger)
wavelengths and is known to be NP- ber with high probability. In particu- least-squares problems by the iterative
Hard. Our approach has two phases. The lar, we prove that the expectation of the solver LSQR. With the entropy objective,
first one solves a standard network de- log of the condition number of any ap- LSQR and PDSCO converge remarkably
sign problem that seeks a set of routes propriately scaled linear program subject rapidly.
that minimizes the necessary capacity of to a Gaussian perturbation of variance
the ring edges. The second phase deter- sigma squared is at most O(log nd/sigma)
mines an optimal coloring for the paths with high probability. By combining this TH1-341/22 LOGISTICS AND TRANSPORTATION

selected in the first stage. Computational result with a recent result of Spielman Railways / public
results show that this approach yields op- and Teng on the smoothed complexity of transportation
timal solutions in almost all cases. In the the termination phase of an interior-point organizer/chair : Leo Kroon
cases where a gap persists we use con- method, we show that the smoothed com-
straint aggregation to solve the problem plexity of an interior-point algorithm
 for Rolling stock circulations for
to proven optimality. linear programming is O m3 log m σ   . passenger trains
In contrast, the best known bound on the
worst-case complexity of linear program- L EO K ROON
Designing stacks of intercon- ming is O m3 L  , where L could be as Netherlands Railways
nected bidirectional sonet rings large as m. coauthor: Lex Schrijver
The smoothed complexity of an algo- keywords: railway systems, rolling
ROSEMARY T. B ERGER rithm is defined to be the maximum over
Industrial and Systems Engineering, stock circulation, public transportation,
its inputs of the expected running time of optimization
Lehigh University the algorithm under slight perturbations
coauthor: Karen A. Kelly of that input. It has been proposed as In this presentation, we describe opti-
keywords: networks/graphs an alternative to worst-case and average- mization models that can be used to find
applications, integer programming, case analyses, in an attempt to explaining high quality rolling stock circulations for
communications the good practical performance many al- passenger trains that may consist of train
gorithms. The result of this talk may help units of different types. Performance in-
The problem of designing stacks of in- explain the observed fast convergence of dicators for evaluating these rolling stock
terconnected bidirectional SONET rings the interior-point algorithms. circulations are: (i) the quality of the
arises in the capacity planning of matching of demand and supply for ca-
telecommunications networks. Given a pacity per trip, (ii) the total number of
topology of interconnected rings and a A trust region inexact Newton
interior-point method for solv- carriage kilometers, and (iii) the num-
set of demands between pairs of nodes, ber of shunting movements. Since the
the objective of the problem is to de- ing large scale unconstrained
minimization problems trains may consist of train units of dif-
termine how to route each demand and ferent types, the positions of the train
where to place add/drop multiplexers L ETICIA V ELAZQUEZ units in the trains need to be modeled
(ADMs) on each stack of each ring in University of Texas at El Paso explicitly. In this way, infeasible shunt-
such a way as to minimize the total num- ing movements can be excluded a pri-
ber of ADMs required. Two linear in- coauthors: Miguel Argaez, Roberto
Saenz ori, but this also leads to huge optimiza-
teger programming formulations are pre- tion models. Computational results are
sented. The formulations are extensions keywords: unconstrained minimization,
interior point method presented, based on several real-life case
of path-based and edge-based formula- studies from Netherlands Railways.
tions for the multi-commodity flow prob- We present a trust region inexact Newton
lem. A set of constraints that strengthens interior-point algorithm for solving un-
the linear programming relaxation bound constrained minimization problems. We Multiple-depot integrated vehi-
is identified. Solution algorithms to solve propose to solve the linear system using cle and crew scheduling
122 Parallel Sessions: Thursday 9:00 – 10:30 (TH1)

D ENNIS H UISMAN YOSHIKO WAKABAYASHI describe some lifting procedures and dis-
Econometric Institute, Erasmus University of São Paulo cuss separation heuristics. We finally de-
University Rotterdam coauthor: Liliane R. Benning Salgado scribe a Branch-and-Cut algorithm based
coauthors: Richard Freling, Albert keywords: approximation algorithms, on these results and present some compu-
Wagelmans graph partitioning, connectivity tational results for both the bipartite in-

keywords: vehicle scheduling, crew
scheduling, column generation, a weight function w : V


Let G  V E  be a connected graph with


 , and let
duced subgraph and the constrained via
minimization problem.
Lagrangian relaxation p, q  be positive integers. For X  V ,
This paper presents two different mod- let w X  denote the sum of the weights A genetic algorithm for the vehi-
els and algorithms for integrated vehi- of the vertices in X. We consider the cle routing problem
cle and crew scheduling in the multiple- following two NP-hard problems on G:
depot case. The algorithms are both the max balanced connected q-partition  K IMMO TAPIO N IEMINEN
based on a combination of column gener- problem and the max connected p q  - Helsinki University of Technology
ation and Lagrangian relaxation. Further- partition problem. In the first, defined
more, we compare those integrated ap- for q
2, we are  interested in finding a
coauthor: Istvan Maros
proaches with each other and with the tra- q-partition

P  V1 V2    Vq  of V such keywords: genetic algorithm, vehicle
  (1  i  q) and
ditional sequential one on random gener- that G Vi  is connected routing
ated as well as real-world data instances P maximizes min w Vi  : 1  i  q
. In In this study we present a new genetic al-
for a suburban/extra-urban mass transit the second, defined for 1  p q, we gorithm for the capacitated vehicle rout-
system. To simulate such a transit sys- want to
find a partition

V1 V2  of V such ing problem (CVRP) in which a fixed
tem, we propose a new way of generating  G V1  and G V2  are connected
that  and fleet of delivery vehicles of uniform ca-
randomly data instances such that their  V1  is maximum subject to w V1  
w pacity, a common depot and several geo-
properties are the same as for our real- p q  w V  . These two problems are graphically scattered customer demands
world instances. equivalent when q  2; for this case ap- are given and the problem is to find the
proximation algorithms have appeared in set of routes with overall minimum route
Shunting passenger train units the literature. We consider this more gen- cost which service all the demands. In a
eral setting and present approximation al- practical application we noticed that em-
R AMON L ENTINK gorithms with performance ratio at most
Erasmus University Rotterdam bedded travelling salesperson problems
2. We show that the ratio decreases as the (TSP) problems involves less than 11 lo-
coauthor: Leo Kroon connectivity of G increases. cations. It justifies the use of an exact
keywords: railway transportation TSP algorithm. In this way we can sub-
In this talk, we introduce Train Unit The multi-layer constrained via stantially reduced the search space of the
Shunting Problem (TUSP). This prob- minimization problem: formu- CVRP problem. The solution space is
lem occurs when train units are tem- lation andpolyhedral issues reduced significantly because during the
porarily not necessary to operate a given evolutionary process we need to focus on
timetable. The problem is to park the P IERRE F OUILHOUX finding good subsets instead of ordered
superfluous train units at a shuntyard as Université Blaise Pascal -
subsets of the set of all customers. Ex-
efficiently as possible. One element of Clermont-Ferrand II
perimental results of this method are en-
this problem is the matching of arriving coauthor: Ali Ridha Mahjoub couraging.
train units to departing train units. Fur- keywords: VLSI design, bipartite
thermore, for each train unit, we need to induced subgraph, polytope,
find routes from it’s arrival platform to branch-and-cut TH1-321/053 PRODUCTION AND SCHEDULING

it’s shunt track, and from this shunt track The multi-layer constrained via mini- Scheduling with complex
to it’s departure platform. Moreover, de- mization problem can be seen as the last constraints
ciding at which shunt track a unit will be step in the design of integrated circuits chair : Chris Potts
parked is a major part of the problem. Fi- and printed circuit boards. It consists in
nally, there are several processes that take positionning wires on a card so that no A time staged linear program-
place during the night, such as cleaning two wire nets cross. For this, we should ming model for multi-site aggre-
the train units. These processes might use the different card layers which can gate production planning prob-
require moving train units to dedicated be drilled for connecting wires that are lems
tracks and therefore they can complicate on different layers. We call these holes
the problem considerably. In this talk, we vias, and, for fiability reasons, we want Y UE W U
will discuss the elements of the problem to place a minimum number of vias. We London School of Economics
and discuss an application of our solution have shown that the constrained via min- coauthor: H.Paul Williams
approach to real life data of a station in imization on k layers can be reduced to keywords: aggregate production
the Netherlands. the k-partite induced subgraph problem planning, time stage linear
in an appropriate graph G. For the partic- programming, supply chain management
TH1-341/23 APPROXIMATION ALGORITHMS
ular case when k equals 2, we study the
Algorithms for routing and facial structure of the bipartite (2-partite) This paper is concerned with the problem
partitioning induced subgraph polytope P(G) associ- of aggregate production planning (APP)
chair : Kimmo Tapio Nieminen ated with the solutions of that problem. for a multinational fashion company in
We describe several classes of valid in- China. The multi-site production plan-
Approximation algorithms for equalities for this polytope and give nec- ning problem considers the production
partitioning a graph into con- essary and sufficient conditions for these loading plans among manufacturing fac-
nected subgraphs inequalities to be facet defining. We also tories subject to certain restrictions, such
 123

as production import/export quotas im- cost objective, when a new set of jobs ar- Multicriteria model of interchange debts
posed by regulatory requirements of dif- rives and creates a disruption. The deci- repayment is considered. The mathemat-
ferent nations, the use of manufactur- sion maker needs to insert the new jobs ical formulation of the problem is caused
ing factories/locations with regard to cus- into the existing schedule without exces- by existing of crises of interchange non-
tomers’ preferences, as well as produc- sively disrupting it. Two classes of mod- payment in Ukraine.
tion capacity, workforce level, storage els are considered. First, we minimize Let’s consider
 weighed oriented
space and resource conditions of the fac- the scheduling cost of all the jobs, subject graph G  S;U
. S  S1    Sn
is
tories. A time-stage linear programming to a limit on the disruption caused to the set of graph nodes – enterprises. The
model is developed to minimize the total original schedule, where this disruption weight vector (debts values) is given on
cost, while satisfying the different pro- is measured in various ways. In the sec- set U directed lines of orgraph. It is
duction constraints. To enhance the prac- ond class a total cost objective, which in- necessary to distribute debts among the
tical implications of the proposed model, cludes both the original cost measure and given objects to discharge maximally
different managerial production loading the cost of disruption, is minimized. For the debts by transferring of some debts
plans are evaluated according to changes both classes and various costs based on units through m nodes on a contour.
in future production policy and situation. classical scheduling objectives, and for The vector  criterion  function is given:
Numerical results demonstrate the effec- almost all problems, we provide either an f x  f1 x     f5 x   . Five criteria
tiveness of the developed model. efficient algorithm or a proof that such an are proposed with its economic interpre-
algorithm is unlikely to exist. We also tation.
show how to extend both classes of mod- The optimization algorithms are con-
Scheduling to simultaneously els to deal with multiple disruptions in structed for each criterion and for two-
optimize two objectives the form of repeated arrivals of new jobs. criteria problem. The lexicographic
Our work refocuses the extensive litera- method of multicriteria choice is applied
C LIFFORD S TEIN ture on scheduling models towards issues for the problem decision in uncertainty.
Estimations of the alternatives j  1 J
Columbia University
of rescheduling which are important be-
coauthors: April Rasala, Eric Torng, cause of the frequency with which dis- by criteria fi is function of uncertainty
Pratchawat Uthaisombut ruptions occur in manufacturing practice. preference relation, when the criterion 
 to a maximum: µ fi j x 
keywords: multiple criteria, approaches
approximation algorithms TH1-321/033 MULTICRITERIA OPTIMIZATION
fi j sup j fi j  . Uncertainty information
Scheduling algorithms are designed to Planning by multicriteria is accounted by consideration of nature
optimize many different optimality cri- optimization conditions – bankruptcy or sudden in-
teria in a wide variety of scheduling chair : Alexander Scherrer crease of depts. The solution can also be
models. Unfortunately, an algorithm considered with use of separating func-
that works well for one scheduling prob- An MCDM approach to portfo- tion containing potentials with weak sin-
lem/objective may perform poorly on a lio optimization gularity. Calculation method of the po-
different scheduling problem/objective. tential is also developed. Computer pro-
K ATHRIN K LAMROTH grams on the algorithms are made in VB.
We give very general results about the Institute of Applied Mathematics,
existence of schedules which simultane- University of Erlangen-Nuremberg
ously minimize two criteria. We will fo- Intensity modulated radiation
coauthors: Matthias Ehrgott, Christian
cus mainly on results that apply to al- therapy - a multicriteria opti-
Schwehm
most any scheduling environment, and mization problem
keywords: portfolio, multiple criteria,
apply to all pairs of metrics in which
utility function, metaheuristics A LEXANDER S CHERRER
the first metric is one of maximum flow
time, makespan, or maximum lateness We propose a model for portfolio opti- Fraunhofer Institute for Industrial
and the second metric is one of average mization extending the Markowitz mean- Mathematics
flow time, average completion time, aver- variance model. Based on a case study coauthors: Karl-Heinz Kuefer, Michael
age lateness, number of on-time jobs. We in cooperation with Standard & Poor’s Monz, Fernando Alonso, Thomas
will show that for almost all such pairs we use five specific objective functions Bortfeld, Christian Thieke
of metrics there exist schedules which are related to risk and return and allow the keywords: Pareto domain
simultaneously close to optimal for both consideration of individual preferences representation, hierarchical clustering,
metrics. through the construction of decision- adaptive problem approximation,
We will also discuss algorithms for maker specific utility functions. Numeri- multicriteria decision support
simultaneously optimizing two metrics. cal results using customized local search,
simulated annealing, tabu search and ge- In oncological intensity modulated radi-
netic algorithm heuristics show that good ation therapy, inverse treatment planning
Rescheduling for new orders solutions can be obtained quickly for strives for ideal treatment plans with suf-
problems of practically relevant size. ficiently high doses in the tumour and
C HRIS P OTTS ideally low doses in the organs at risk. In
Faculty of Mathematical Studies, general, there are no intensity profiles to
Optimization of debt manage- realize such dose distributions, hence ac-
University of Southampton ment in uncertainty by algo- ceptable compromises have to be found.
coauthor: Nicholas G. Hall rithms on graph We model this as a convex multicriteria
keywords: scheduling, single machine,
disruption G ANNA S HYSHKANOVA optimization problem, where the criteria
Zaporizhzhya National Technical are defined as relative deviations from the
This paper considers scheduling prob- University ideal doses. A representative set of sev-
lems where a set of original jobs has al- keywords: decision analysis, multiple eral hundred clinically meaningful Pareto
ready been scheduled to minimize some criteria, uncertainty, graphs optimal solutions is then computed and
124 Parallel Sessions: Thursday 9:00 – 10:30 (TH1)

stored in a database, in which the physi- for the aimed distance are provided, and blocked Cholesky factorization of matri-
cian, supported by a navigation tool, can different examples show the difficulties ces arising in time phased or dynamic
search for a suitable treatment plan. The underlying this analysis. The formula LP models solved by interior point algo-
discrete problem formulations are large for the distance to insolvability consists rithms.
scale, but the use of a combined hier- of the minimum of two ingredients; the
archical clustering and local refinement first one is the distance to infeasibility,
strategy yields adaptive problem approx- whereas the other ingredient turns out to
imations of decisively reduced size and be the ’distance to ill-posedness’ of the Application of robust counter-
thus acceptably small computation times. dual problem. This second ingredient can part technique to production
Numerical experiments on real clinical be also used for finding balls, centered on management
data showed, that the problems can then the origin, containing the optimal set or
be exactly solved on the basis of adapted certain lower level sets, which constitute
approximations with 5-10% of the origi- key ingredients in certain Lipschitz con- V INCENT G UIGUES
nal size and the computation of the whole stant for the optimal value function ob- IMAG-LMC
database requires approximately 2 hours. tained in the paper.
The talk provides a generalized mathe- keywords: uncertain linear programs,
matical embedding of the problem, gives Gaussian elimination as a com- robust optimization, stochastic
detailed information about computational putational paradigm programming, AARC
complexity and presents numerical ex- E TIENNE L OUTE
amples. CORE/UCL The problem of production management
keywords: parallel programming, large can be often cast in the form of linear
TH1-321/133 LINEAR PROGRAMMING

scale linear programming,

Aspects of linear program with uncertain parameters. Typ-
decomposition algorithms, Cholesky ically, such problems are treated in the
programming factorization framework of multi-step stochastic pro-
chair : Vincent Guigues An abstract view of the symmetric gaus- gramming. Recently, a new robust coun-
sian elimination is presented. Problems terpart (RC) approach has been proposed
Distance to ill-posedness for lin- are viewed as an assembly of compu-
ear optimization problems: dis- in the literature, in which the decisions
tational entities whose interdependence are optimized for the worst realizations of
tance to insolvability is modeled by a graph. An algorithmic problem parameters. However, an appli-
F.JAVIER TOLEDO -M ELERO transformation on the entities which can cation of the RC technique often results
Operations Research Center/Miguel be associated with vertex removal, is as- in very conservative solutions. To tackle
Hernández University sumed to exist. The elimination tree of this drawback, an adjustable robust coun-
coauthors: M. J. Cánovas, M. A. López, the symmetric gaussian elimination fig- terpart (ARC) approach has been pro-
J. Parra ures the order in which these transforma- posed in [1]. In ARC, some decision
keywords: distance to insolvability tions are applied and captures any poten- variables are allowed to depend on past
This talk is concerned with the param- tial parallelism. The inherently sequen- values of uncertain parameters. A re-
eter space of all the linear optimiza- tial part of the computational effort de- stricted version of ARC, introduced in
tion problems, in the n-dimensional Eu- pends on the height of the tree. The [1], is the affinely adjustable robust coun-
clidean space, whose constraint systems paradigm is illustrated by block struc- terpart (AARC). The implementation of
are indexed with a fixed and arbitrary in- tured LP problems with nested decompo- robust programs relies on the description
dex set. In this context, the boundary of sition and basis factorization approaches, of parameter uncertainty sets, which can
the set of solvable (resp. bounded) prob- problems of blocked symmetric and un- be defined using the a priori information
lems may be considered as the set of ill- symmetric systems of linear equations, or available past observations of problem
posed problems. In a first stage, the pa- with respectively blocked Cholesky fac- parameters. In this paper, we consider an
per shows that, both, the boundary of the torization and blocked gaussian elimina- application of RC and AARC methodol-
set of solvable problems and the one of tion. Contributions are: demonstration of ogy to the problem of annual electricity
bounded problems coincide. The second the paradigm expressive power through production management and gas distribu-
stage consists of characterizing the prob- graph concepts (eliminations sets, elimi- tion management in France. We compare
lems in this boundary. The paper also nation chains, etc.); emphasis on patterns the quality of robust solutions obtained
provides a formula in terms of certain of similarity in the use of the graph con- by polyhedral or ellipsoidal uncertainty
subsets of R and Rn 1 for the distance cepts (with connections between nested sets which are estimated using historical
to this boundary, provided that the nom- decomposition and blocked basic factor- data. [1] Ben-Tal, Goryashko, Guslitzer,
inal problem is solvable. The remain- ization for the revised simplex) and fi- Nemirovski. Adjustable robust solutions
ing case is also considered; lower bounds nally an effective parallelization tool for of uncertain linear programs.
 125

TH2-302/41 COMBINATORIAL OPTIMIZATION Reconstructing (0,1)-matrices Shellability is one of basic properties

using Lagrangean decomposi- in the field of topological and geomet-
Computational methods for tion ric combinatorics, and it has been stud-
graph-coloring III ied extensively by many researchers un-
organizer/chair : Michael Trick T RULS F LATBERG til today. However, it is hard to de-
Department of Informatics, University of
cide whether a given simplicial complex
Lightpath coloring in all-optical Oslo
is shellable or not. Deciding shellabil-
networks coauthor: Geir Dahl ity of a simplicial complex is one of
A RIE M.C.A. KOSTER keywords: integer programming, difficult problems which are unknown
discrete tomography to be solved in polynomial time or not.
Zuse Institute Berlin (ZIB)
We consider a variant of the NP-hard Shellability is a concept asking whether
coauthor: Adrian Zymolka problem of reconstructing hv-convex a simplicial complex has a shelling or
keywords: communications, (0,1)-matrices from known row and col- not, where a shelling is a total ordering
networks/graphs applications, integer umn sums. Instead of requiring the ones F1 F2



to occur consecutively in each row and ing that  F1


Fk  1 
Fk is a pure
programming Ft of the facets satisfy-
column, we maximize the number of  dimFk 1  -dimensional simplicial com-

In optical telecommunication net-

works, wavelength division multiplexing neighboring ones. This is reformulated plex for every 2  k  t. Instead of using
(WDM) allows for the efficient usage of as an integer programming problem. A such kind of total ordering, we charac-
fibers by transmitting optical signals in solution method based on variable split- terize the shellability of simplicial com-
parallel on different wavelengths. Opti- ting is proposed and tested with good re- plexes in terms of acyclic orientations of
cal Cross Connects provide full optical sults on moderately sized test problems. the Hasse diagram of their face posets.
switching in the nodes. This enables We show that shellability is described
ongoing optical channels across several by a problem of minimization of an ob-
links, so-called lightpaths. A lightpath A new class of matrices for half- jective function among acyclic orienta-
is a path together with a wavelength integrality tions satisfying some conditions. We also
associated to each of its links. At a show that this characterization gives a
node the wavelength of operation can be BALAZS KOTNYEK
INRIA simple integer programming type formu-
exchanged at the cost of a wavelength lation of shellability.
converter. The wavelength assignment coauthor: Gautam Appa
problem consists in the conflict-free allo- keywords: networks/graphs, matrices,
cation of the available wavelengths to a generalized networks TH2-302/44 COMBINATORIAL OPTIMIZATION

set of lightpaths with a minimum number We present a class of matrices, called bi-
net matrices, that ensure half-integral op-
Paths in graphs
of converters.
In this talk, we discuss this gener- timal solutions. They furnish a direct chair : L. Sunil Chandran
alized coloring problem and the rela- generalisation of totally unimodular net- NP-hardness of some maximum
tion with both vertex and edge coloring. work matrices and arise from the node- edge-disjoint paths problem
We present integer programming formu- edge incidence matrices of bidirected
lations for the minimum converter wave- graphs in the same way as the network E UGENE BARSKY
length assignment problem and discuss matrices do from directed graphs. Net- NACE(Negev Academic College of
their properties. For the most promising work matrices are shown to be a special Engineering)
formulation, we present a column gen- case of binet matrices, as are the two im-
eration algorithm to obtain good assign- portant non-network but totally unimod- coauthor: Ilani Hagai
ments. ular matrices. We discuss ways in which keywords: edge-disjoint paths
binet matrices preserve the advantages 
of network matrices, and provide theo- Given a graph G  V E  , a set of ter-
Hybrid local search on two mul- 
rems revealing the structure of these ma- minals T  V and a demand graph H 
ticolouring models trices. Connections are made to signed T F  , we call a path in G an H-path if
graphs, monotone inequalities and sub- its ends are adjacent terminals in H. We
S TEVEN P RESTWICH prove that if H is fixed and has 3 inter-
Cork Constraint Computation Centre spaces with well-scaled frames respec-
tively. Finally, for the LP or IP problem secting maximal stable sets , A , B , C ,
keywords: heuristics, stochastic
defined over a binet constraint matrix, such that A
B  A
C then, the problem
This paper describes a hybrid local we provide adaptations of algorithms for of finding maximum number of edge-
search approach to pure and bandwidth generalised networks and matching as disjoint H-paths in an Eulerian graph G
multicolouring problems. It compares well as a proof that the strong Chvatal is NP-Hard.
two models of multicolouring: a 0/1 rank of the relevant polytope is one.
integer linear program, and a binary
constraint satisfaction model derived by Two edge-disjoint hop-
Shellability of simplicial com- constrained paths and polyhe-
transforming multicolouring problems to plexes and optimization in dra
colouring problems. The latter approach acyclic orientations
yields smaller models and competitive
results. S ONOKO M ORIYAMA DAVID H UYGENS
Dept. of Computer Science, Faculty of Université Libre de Bruxelles
TH2-302/42 COMBINATORIAL OPTIMIZATION Information Science and Technology, coauthors: Ali Ridha Mahjoub, Pierre
The University of Tokyo Pesneau
Combinatorial optimization
VIII coauthor: Masahiro Hachimori keywords: survivable network,
keywords: simplicial complex, edge-disjoint paths, hop-constraints,
chair : Sonoko Moriyama shellability, acyclic orientation polyhedron
126 Parallel Sessions: Thursday 13:15 – 14:45 (TH2)

Given a graph G with distinguished problem of allocating a set of jobs to a the generalized assignment problem, in
nodes s and t, a cost on each edge of G, set of machines, where each job must which agents can perform tasks at differ-
and a fixed integer L
2, the Two edge- be assigned to a machine, or shared be- ent efficiency levels. MGAP arises, for
disjoint Hop-constrained Paths Problem tween two adjacent machines, and there example, in the context of manufactur-
(THPP for short) consists in finding a are constraints on the availability of re- ing systems. As reported in the litera-
minimum cost subgraph such that, be- sources for job allocation. For reason- ture, due to the large number of binary
tween s and t, there exist at least two ably large values of m and n the NP- variables involved, it is impractical to use
edge-disjoint paths of length at most L. In hard combinatorial problem GAPS2 be- standard ILP software to achieve opti-
this talk, we consider the problem from comes intractable for standard mathemat- mal solutions to these problems. A tabu
a polyhedral point of view. We give an ical programming software, hence there search algorithm (Laguna et al., 1995),
integer programming formulation for the is a need for heuristic algorithms to solve obtained good results for instances with
problem and discuss the associated poly- such problems. It will be shown how an up to 40 tasks, 4 agents and 4 effi-
tope, P G L  , when L  2 3. In partic- LP-based heuristic developed previously ciency levels. Two recent heuristic al-
ular, we show that in this case the linear for the well-established generalized as- gorithms (French and Wilson, 2002) ob-
relaxation of P G L  , Q G L  , given by signment problem can be modified and tained good approximations on problems
the trivial, the st-cut, and the so-called extended to solve GAPS2. Encouraging with up to 200 tasks, 30 agents and 5
L-path-cut inequalities, is integral. As results, in terms of speed and accuracy, efficiency levels. We present a branch-
a consequence, we obtain a polynomial in particular when compared to an exist- and-price algorithm that, at the best of
time cutting plane algorithm for the prob- ing heuristic for GAPS2, are described. our knowledge, is the first exact algo-
lem when L  2 3. We also give nec- rithm for MGAP. We report computa-
essary and sufficient conditions  for these tional experience on many sets of ran-
inequalities to define facets of P G L  for Cut-and-branch for the general- domly generated instances, with different
L
2. We  finally investigate the domi- ized assignment problem kind of correlations between cost and re-
nant of P G L  and give a complete de- source consumption for each agent-task-
scription of this polyhedron
 for L
2, A NDREAS D REXL level assignment, and with different ra-
when P G L   Q G L  . University of Kiel tio between the number of agents, tasks
coauthor: Andreas Klose and levels. Experimental results show
Isoperimetric inequalities and keywords: generalized assignment that our method is very effective, and it
the width parameters of graphs problem, extended cover inequalities, solves to proven optimality instances of
cut and branch the maximum size proposed in the litera-
L. S UNIL C HANDRAN The generalized assignment problem
ture.
Max Planck Institute for Informatik (GAP) examines the maximum profit as-
keywords: graph theory, treewidth, path signment of jobs to agents such that each TH2-302/49 INTEGER AND MIXED INTEGER
PROGRAMMING
width, hypercube job is assigned to precisely one agent Routing in networks
We relate the isoperimetric inequalities subject to capacity restrictions on the
with many width parameters of graphs: agents. The GAP has many applica- chair : Bruce Shepherd
treewidth, pathwidth and the carving tions and several algorithms have been
proposed for solving this NP-hard opti-
A prize collecting connected
width. Using these relations, we deduce subgraph problem
1. A lower bound for the treewidth in mization problem. Recently, Gabrel and
terms of girth and average degree 2. The Minoux (2002) have developed an al-
exact values of the pathwidth and carving gorithm for the exact separation of ex- P ER O LOV L INDBERG
tended cover inequalities (ECI). Further- Department of Mathematics, Linköping
width of the d–dimensional hypercube,
 more, they have shown that ECI speed-up University
Hd 3. That treewidth Hd   Θ 2 d .
d

the convergence of Cplex when solving coauthor: Nima Golbaharan

Moreover we study these parameters the multidimensional knapsack problem. keywords: prize collecting, subgraph,
in the case of a generalization of hyper- In this research we adapt the generation routing
cube, namely the Hamming graphs. of ECI to the GAP. Additionally, ECI are
embedded within a cut-and-branch algo- We consider a new (?) NP-hard optimiza-
TH2-302/45 INTEGER AND MIXED INTEGER
PROGRAMMING
rithm for solving the GAP. We show by tion problem, the Prize Collecting Con-
Generalized assignment means of a computational study that ECI nected Subgraph Problem, which appears
improve convergence of Cplex substan- as a subproblem in routing of snow plows
chair : Alberto Ceselli tially. during snow fall. In this problem we have
An LP-based heuristic proce- a set of arcs a  A in a network with
dure for the generalized assign- A branch-and-price algorithm arc “costs” ca , and arc times ta . More-
ment problem with special or- for the multilevel generalized as- over there is a time budget b. The prob-
dered sets signment problem lem is to find a connected subset Ā 
A (corresponding to a snow plow tour)
J OHN W ILSON A LBERTO C ESELLI such that ∑ ca is minimized subject to the
Loughborough University University of Milano, Department of budget constraint ∑ ta  b. This prob-
coauthor: Alan French Information Technologies lem can be modeled by introducing mul-
keywords: assignment, generalized coauthor: Giovanni Righini ticommodity constraints or adding valid
assignment, heuristics, special ordered keywords: generalized assignment, inequalities. We have mainly used the
sets column generation, branch-and-bound second approach. We will present results
for the classical Sioux Falls network, and
The generalized assignment problem The multilevel generalized assignment
for a real life case of the region of Eskil-
with special ordered sets (GAPS2), is the problem (MGAP) is a variation of
stuna in Sweden.
 127

TH2-306/31 INTEGER AND MIXED INTEGER

PROGRAMMING
talk, we report on the progress of COIN-
The quadratic selective travel- OR and its future direction. We will
Open-source software for announce the winners of the COIN-OR
ling salesman problem mathematical programming Open Source Coding Contest. The con-
organizer/chair : Robin Lougee-Heimer test was a joint effort of INFORMS,
TOMMY T HOMADSEN
Technical University of Denmark The COIN-OR open solver in- MPS, and COIN-OR with prizes donated
by IBM.
coauthor: Thomas Stidsen terface: a progress report
keywords: traveling salesman, M ATTHEW S ALTZMAN
branch-and-cut, ring networks, Prox-Accpm: a cutting plane
Clemson University solver for convex optimization
multicommodity
keywords: software, integer
A well-known extension of the Travel- programming, linear programming, open C LAUDE M ARTIN TADONKI
ling Salesman Problem (TSP) is the Se- source University of Geneva
lective TSP: Each node has an associated When the COIN-OR project debuted at coauthors: Jean-Philippe Vial, Frédéric
profit and instead of visiting all nodes, ISMP 2000, a key component of the Babonneau, Cesar Beltran, Olivier du
only profitable nodes are visited. The project was the Open Solver Interface Merle
Quadratic Selective TSP (QSTSP) adds (OSI). The OSI is intended to be a com- keywords: convex, cutting planes, non
the additional complication that each pair mon application program interface (API) smooth, analytic center
of nodes have an associated profit which for calling any of a variety of embed-
can be gained only if both nodes are vis- ded solvers in an algorithm. The cur- Prox-Accpm is an extension of the ana-
ited. The QSTSP problem is motivated rent version includes support for CPLEX, lytic center cutting plane method to solve
by being a subproblem when construct- XPRESS-MP, Soplex, the GNU LP Kit, general (nondifferentiable) convex opti-
ing hierarchical ring networks. Hafer’s DyLP, and COIN-OR native mization problems, whose components
We describe a mixed integer linear solvers CLP (the COIN-OR LP solver), (objective and constraints) are described
program model for the QSTSP. The prob- SBB (Simple Branch and Bound), and by first order oracles. The addition of
lem is solved by Branch and Cut using Vol (an implementation of Barahona and a proximal term allows a better con-
the COIN/BCP framework. Computa- Anbil’s volume algorithm). A user can trol of the algorithm behavior. Prox-
tional results are presented. write a single implementation of an algo- Accpm applies to problems arising from
rithm calling any of these solvers through decomposition approaches (Benders, La-
the same interface. grangian, Column generation) or to equi-
Maximum multicommodity de- This presentation describes features librium problems, or to cutting plane ap-
mand flows of the OSI and an outline of the design proaches to difficult mathematical pro-
of a new version of the OSI (under devel- gramming problems (e.g., some SDP’s).
B RUCE S HEPHERD opment). The new design will improve In view of the wide range of applications,
Bell Laboratories flexibility and efficiency and simplify the Prox-Accpm is conceived as a parame-
coauthors: C. Chekuri, M. Mydlarz process of embedding new solvers. It will tizable solver. It is written in MATLAB,
keywords: integrality gap, knapsack, offer a consistent (solver-independent) but some routines are written in C. All
problem representation and access to a developments are tested against families
integer multicommodity flow,
b-matching broad set of solver capabilities through an of benchmark problems: Column gen-
efficient, open, standard, portable API. eration (cutting stock, linear separation
We consider requests  for capacity in a in data mining), Lagrangian relaxations
tree network T  V E  where each edge (p-median, unit commitment, nonlinear
of the tree has a bounded capacity ue .
The COIN-OR initative: open- multi-commodity flow), equilibria (two-
Each request consists of an integer de-
source tools for mathematical person games, traffic equilibria), non-
mand d f and a profit w f which is ob-
programmers linear convex problems (optimization on
tained if the request is satisfied. The ob- ROBIN L OUGEE -H EIMER positive polynomials, quadratically con-
jective is to find a set of demands that can IBM Research strained problems) etc. Prox-Accpm can
be feasibly routed in the tree and which be used as a standalone binary version to
keywords: software, optimization, be included in any standard program for
provide a maximum profit. This general- mixed integer programming, open
izes well-known problems including the user-specific purposes.
source
knapsack and b-matching problems.
When all demands are 1, we have Much of the mathematical programming
research and application relies on soft- TH2-306/32
NONLINEAR PROGRAMMING

the integer multicommodity flow prob-

lem. In this case, Garg, Vazirani, and ware. And yet, there is relative low num- Surrogate modelling for
Yannakakis had shown an integrality gap ber of reference implementations, com- engineering optimization I
of 2 for the cardinality version of the mon interfaces, re-usable frameworks, organizer/chair : Kaj Madsen
problem. We establish factor 4 gap for and open standards for the specific soft-
the case of arbitrary profits. ware needs of the mathematical program- On multivariate polynomial
ming community. Open source is an al- interpolation and its use in
For arbitrary demands (with maxi-
ternative style of software development derivative-free optimization
mum demand at most the minimum ca-
with some attractive benefits. Three (parts I and II)
pacity) we show that the integrality gap
years ago at ISMP 2000, the COIN-
of the LP is at most 48. This is obtained
OR project was launched as an exper- A NDREW C ONN
by showing that the integrality gap for de-
iment with the mission of promoting IBM T.J. Watson Research Center
mand version of such problems is at most
open-source software for the mathemat- coauthors: Katya Scheinberg, Luís N.
12 times that for their unit demand ver-
ical programming community. In this Vicente
sion.
128 Parallel Sessions: Thursday 13:15 – 14:45 (TH2)

keywords: multivariate interpolation, to the key ingredients for attaining global multipliers in L2 -spaces is proven under
derivative-free optimization, trust region convergence. A new approach to solve natural assumptions. This question is an-
method, bilevel optimization derivative-free bilevel optimization prob- swered by investigating a Lagrange dual
We will introduce a new, comprehensive lems based on interpolation-based mod- quadratic problem. The theory is applied
approach for multivariate polynomial in- els for both upper and lower level ob- to different optimal control problems for
terpolation under which it is possible to jective functions is also presented. The elliptic and parabolic partial differential
derive appropriate bounds for the error underlying algorithms are of trust-region equations with mixed pointwise control-
between the function being interpolated type and the setting is such that both state constraints.
and its interpolating polynomial. This function and gradient values can be inex-
derivation is based on a new concept of act.
well-poisedness for the interpolation set, Space mapping for optimal con-
directly connecting the accuracy of the Constrained global optimization trol of partial differential equa-
error bounds with the geometry of the with radial basis functions tions
points in the set. Our approach includes
error bounds for function values as well JAN -E RIK K ÄCK
as for derivatives. We show how to de- M ICHAEL H INTERMÜLLER
Department of Mathematics and Department of Mathematics / University
sign algorithms to build sets of well- Physics, Mälardalen University
poised interpolation points or to modify of Graz
coauthor: Kenneth Holmström
existing interpolation sets to ensure well- coauthor: Luís N. Vicente
keywords: global optimization,
poisedness. keywords: space mapping, optimal
response surface, costly constraints
We will also talk about derivative- control of PDEs, trust regions
free optimization techniques, relating ge- Response surface methods show promis-
ometrical concepts like well-poisedness ing results for global optimization of Numerically solving optimal control
to the key ingredients for attaining global costly nonconvex objective functions, i.e. problems for nonlinear partial differential
convergence. A new approach to solve the problem of finding the global mini- equations represents a significant chal-
derivative-free bilevel optimization prob- mum when there are several local min- lenge due to the tremendous size and pos-
lems based on interpolation-based mod- ima and each function value takes con- sible model difficulties (e.g., nonlineari-
els for both upper and lower level ob- siderable CPU time to compute. Such ties) of the discretized problems. In this
jective functions is also presented. The problems often arise in industrial and talk, a novel space-mapping technique
underlying algorithms are of trust-region financial applications, where a func- for solving the aforementioned problem
type and the setting is such that both tion value could be a result of a time- class is introduced, analyzed, and tested.
function and gradient values can be inex- consuming computer simulation or opti- The advantage of the space-mapping ap-
act. mization. Derivatives are most often hard proach compared to classical multigrid
to obtain. The problem is here extended techniques lies in the flexibility of not
On multivariate polynomial with linear and nonlinear constraints, and only using grid coarsening as a model
interpolation and its use in the nonlinear constraints can be costly or reduction but also employing (perhaps
derivative-free optimization not. A new algorithm that handles the less nonlinear) surrogates. The space
(parts I and II) constraints, based on radial basis func- mapping is based on a regularization ap-
tions (RBF), and that preserves the con- proach which, in contrast to other space-
L UÍS N. V ICENTE vergence proof of the original RBF algo-
University of Coimbra mapping techniques, results in a smooth
rithm is presented. The algorithm takes mapping and, thus, avoids certain irreg-
coauthors: Andrew Conn, Katya advantage of the optimization algorithms ular situations at kinks. A new Broy-
Scheinberg in the Tomlab optimization environment den update formula for the sensitivities
keywords: multivariate interpolation, (http://tomlab.biz). Numerical results are of the space map is also introduced. This
derivative-free optimization, trust region presented for standard test problems. quasi-Newton update is motivated by the
method, bilevel optimization usual secant condition combined with a
We will introduce a new, comprehensive secant condition resulting from differen-
approach for multivariate polynomial in- TH2-306/33 NONLINEAR PROGRAMMING

PDE-constrained optimization tiating the space-mapping surrogate. The

terpolation under which it is possible to overall algorithm employs a trust-region
derive appropriate bounds for the error (2)
framework for global convergence. We
between the function being interpolated organizers : Michael Ulbrich and highlight some of the issues involved in
and its interpolating polynomial. This Stefan Ulbrich the computations and we report a few il-
derivation is based on a new concept of chair : Michael Ulbrich lustrative numerical tests.
well-poisedness for the interpolation set,
directly connecting the accuracy of the
Optimal control of PDEs with
error bounds with the geometry of the
mixed control-state constraints
points in the set. Our approach includes Software for rapid source loca-
F REDI T ROELTZSCH tion in chemical/biological de-
error bounds for function values as well TU Berlin, Institut fuer Mathematik
as for derivatives. We show how to de- fense
sign algorithms to build sets of well- keywords: optimal control, Lagrange
multiplier, regularity
poised interpolation points or to modify PAUL T. B OGGS
existing interpolation sets to ensure well- A class of quadratic optimization prob- Sandia National Laboratories
poisedness. lems with pointwise box constraints and
coauthors: Kevin R. Long, Stephen
We will also talk about derivative- constraints ofbottleneck type is consid-
Margolis
free optimization techniques, relating ge- ered in Hilbert spaces. The main sub-
keywords: PDE-constraints, nonlinear
ometrical concepts like well-poisedness ject is the existence of regular Lagrange
optimization, software
 129

The problem to be discussed is to find the many master problems and subproblems. therefore lead to very large-scale opti-
source of a toxic release in a large build- Furthermore, computation of Lagrange mization problems. Their solution re-
ing, e.g., an airport. In such cases, time multipliers for the quadratic program- quires carefully designed optimization
is of the essence and is clearly more im- ming problems is performed based on techniques and the use of high perfor-
portant than accuracy, since even a rela- any available non degenerate extreme mance computing. These optimization
tively crude approximation of the source point of the optimum set. Numerical re- problems display exploitable structure:
will allow timely response by emergency sults for several nonlinear programming the blocks corresponding to scenarios are
managers to contain the release and to example problems demonstrate compu- linked through common variables corre-
take appropriate crowd control actions. tational saving up to 60% due to the sponding to earlier stage investment de-
At the beginning, we are unsure of many proposed procedure. cisions.
of the aspects of the problem, e.g., how Financial planning problems are of-
many sensors are needed, how much ac- Dynamic scaling based precon- ten solved by decomposition meth-
curacy is needed in the prediced loca- ditioning for truncated Newton ods. However, decomposition ap-
tions, how much accuracy is needed in methods in large scale uncon- proaches tend to be slower than any di-
the flow models, and what is the sensi- strained optization rect approach (if the latter can handle the
tivity of the predictions with respect to large problem).
these questions. Our approach has been M ASSIMO ROMA An alternative to a development of a
to start with something relatively simple Universita’ di Roma "La Sapienza" parallel implementation of a decompo-
and move to more complex models. To keywords: truncated Newton methods, sition method is to exploit structure di-
do this, we used several software tools conjugate gradient method, rectly in the linear algebra operations.
that made changing the models and do- preconditioning, row-column scaling We explore such alternative in this paper.
ing the analyses much easier. We re- We have developed a structure-exploiting
This work deals with the precondition-
port specifically on Sundance, a tool that parallel interior-point solver for nonlin-
ing of truncated Newton methods for the
allows symbolic description of the PDE ear programming problems. Its design
solution of large scale nonlinear uncon-
and related constraints as well as easy uses object-oriented programming tech-
strained optimization problems. We fo-
connection to the optimization software. niques. The program OOPS (Object-
cus on preconditioners which can be nat-
Oriented Parallel Solver:
urally embedded in the framework of
TH2-306/34 NONLINEAR PROGRAMMING
truncated Newton methods, i.e. which http://www.maths.ed.ac.uk/
Large scale problems II can be built without storing the Hes- ˜lkgondzio/parallel/solver.html)
sian matrix of the function to be min-
chair : Jacek Gondzio
imized, but only based upon informa- can efficiently handle very large nonlin-
Approximation of optimum sets tion on the Hessian obtained by the prod- ear problems and achieves scalability on
of extreme points in a SQP inte- uct of the Hessian matrix times a vec- a number of different computing plat-
grated simple decomposition tor. In particular we propose a diag- forms.
onal preconditioning which enjoys this We illustrate its performance on
M EHDI L ACHIHEB feature and which enables us to exam- a collection of quadratic programming
Département de Génie Civil, ENIT ine the effect of diagonal scaling on trun- problems with sizes over 50 million deci-
coauthor: Hichem Smaoui cated Newton methods. In fact, this new sion variables arising from asset liability
keywords: simple decomposition, large preconditioner carries out a scaling strat- management and portfolio optimization.
scale, SQP, extreme point egy and it is based on the concept of
equilibration of the data in linear sys- TH2-306/35 NONLINEAR PROGRAMMING

Sacher’s simple decomposition, orig-

tems of equations. An extensive numeri- Newton-like algorithms I
inally developped for large scale
cal testing has been performed showing
quadratic programming problems, has
that the diagonal preconditioning strat- chair : Oleg Burdakov
recently been integrated into a sequential
egy proposed is very effective. In fact, on Newton-like search directions
quadratic programming (SQP) algorithm
most problems considered, the resulting
in order to handle large nonlinear pro-
diagonal preconditioned truncated New- J ELENA N EDIC
gramming problems. The resulting al-
ton method performs better than both the Oxford University
gorithm is particularly efficient when the
unpreconditioned method and the one us- coauthor: Raphael Hauser
number of passive constraints is a small
ing an automatic preconditioner based on keywords: Newton method, q-quadratic
fraction of the number of variables and
limited memory quasi–Newton updating convergence, flux lines, stiffness of
the total number of constraints. However,
(PREQN) recently proposed. vector field
when the standard simple decomposition
is applied, the number of extreme points We consider problems of nonlinear con-
that are generated throughout the SQP it- Parallel solution techniques in tinuous unconstrained optimization. New
erations is usually very large, with many financial planning problems approximations of the local minimiser
points leaving and entering the set. In or- JACEK G ONDZIO 
are obtained via the updating rule xk  1 
der to reduce the computational effort de- School of Mathematics, University of x k  v x k 
 k  0 1 2    where v xk  is a
voted to the generation of extreme points, given vector field. It is a well known fact
Edinburgh
a procedure is developped for initiating that the sequence of iterates obtained by
the simple decomposition with a whole coauthor: Andreas Grothey using the Newton method converges to
set of extreme points that approximates keywords: financial engineering, an optimal solution q-quadratically if the
the optimum set. This set is determined parallel computing, interior point initial approximation is in some neigh-
at the start of each new iteration of the method, large scale optimization bourhood of a local minimiser x . But
SQP sequence, based on the results of the Financial planning problems by their are there other conceptually different di-
preceding one, bypassing the solution of very nature involve uncertainty and rection vector fields that will also lead to
130 Parallel Sessions: Thursday 13:15 – 14:45 (TH2)

q-quadratic convergence? To answer this TH2-306/36 NONSMOOTH OPTIMIZATION q circles. The mathematical modeling
question, necessary and sufficient condi- of that problem leads to a min max
tions for q-quadratic convergence of such Covering problems min p problem, where min p denotes the
methods are derived. They are given chair : Frits Spieksma pth least value. That problem, in ad-
in terms of the Jacobian of the direc-
tion field and are very simple to check.
Global optimality conditions for dition to its intrinsic multi-level nature,
Moreover, we analyse the convergence
discrete and nonconvex optimi- has the significant characteristic of be-
rates of the direction from which xk ap-
ation, with applications to La- ing strongly non-differentiable. To ad-
proaches x to give deeper insight into
grangian heuristics and column dress that difficulty we smooth the min
generation max min p problem by using a special
the problem. In particular, we are in- family of smoothing function of class
terested in providing conditions that will M ICHAEL PATRIKSSON C∞ . The use of this technique, called
completely determine whether the given Department of mathematics Hyperbolic Smoothing, permits to over-
direction field is actually the Newton one come the main difficulties presented by
for some function or not. coauthor: Torbjörn Larsson
keywords: integer programming, the original problem. The final solution is
Lagrangian relaxation, Lagrangian obtained by solving a sequence of differ-
Value function and error heuristics, set covering problems entiable sub-problems which gradually
bounds of a basic trust region approaches the original problem. Prelim-
model The well-known and established global inary computational results obtained for
optimality conditions based on the La- real test problems are also presented to
J IAN -Z HONG Z HANG grangian formulation of an optimization illustrate the potentiality of the method.
City University of Hong Kong problem are consistent if and only if the
coauthor: Chang-Yu Wang duality gap is zero. We develop a set
keywords: trust region method, error of global optimality conditions which are
An approximation algorithm for
bounds, constrained optimization structurally similar but which are con-
an interval stabbing problem
We study some properties of a trust re- sistent for any size of the duality gap. F RITS S PIEKSMA
gion model for solving constrained min- This system characterizes a primal-dual Katholieke Universiteit Leuven
imization problems. We first consider optimal solution by means of primal and coauthor: Sofia Kovaleva
the optimal value function defined by the dual feasibility, primal Lagrangian ε- keywords: combinatorial optimization
predicted reduction of the trust region optimality, and, in the presence of in-
subproblem and the ratio between the equality constraints, δ-complementarity, We consider the following problem.
value function and the trust region radius. that is, a relaxed complementarity condi- Given is a grid consisting of columns and
We expose the close relationship between tion. At an optimal solution the value of rows. The grid contains intervals; each
this ratio and the projected gradient of the ε  δ equals the size of the duality gap. interval intersects a set of consecutive
objective function. We then show under The characterization is further equivalent columns and lies on a single row. We say
different conditions that this ratio and the to a near-saddle point condition which that, if interval i intersects column j, col-
value function provide global or local er- generalizes the classic saddle point char- umn j stabs interval i. We also say that,
ror bounds for the distances from the pro- acterization of a primal-dual optimal so- if interval i lies on row j, row j stabs in-
jected gradient to the origin, and from the lution in convex programming. The sys- terval i. The aim is to select as few rows
iterative point to the solution set of the tem developed can be used to explain, to and columns as possible such that each
minimization problem. Using these error a large degree, when and why Lagrangian interval is stabbed at least once. This
bounds, we give a condition for conver- heuristics for discrete optimization are problem is known to be APX-hard. We
gence of the iterative sequence generated successful in reaching near-optimal so- describe a polynomial-time algorithm,
from the trust region model. Especially, lutions. Further, experiments on a set called STAB, that always outputs a so-
we give a new equivalent condition for fi- covering problem illustrate how the new lution  with a value that is bounded by
nite convergence. optimality conditions can be utilized as e e 1 
1  582 times the optimal value
a foundation for the construction of La- of the problem. This algorithm is based
On limited memory methods grangian heuristics. Finally, we outline on solving the LP-relaxation of an integer
programming formulation of the problem
with shape changing trust re- possible uses of the optimality conditions
in column generation algorithms. described above. We also give extensions
gion to weighted variants of the problem.
O LEG B URDAKOV
Linkoping University Optimum order p covering of
plane domains by circles via hy- TH2-306/37
CONVEX PROGRAMMING

coauthor: Ya-xiang Yuan perbolic smoothing method Conic programming and

keywords: large scale optimization, positive polynomials
limited memory, quasi-Newton methods, A DILSON E LIAS X AVIER
trust region organizer/chair : Javier Pena
COPPE/Federal University of Rio de
We consider a trust-region (TR) ap- Janeiro A conic programming approach
proach. In this approach, the TR- coauthor: Antonio Alberto Fernandes to generalized Tchebycheff in-
subproblem is defined by a limited mem- de Oliveira equalities
ory model function and a trust region, keywords: non differentiable
which shape depends on the model func- programming, semi infinite L UIS Z ULUAGA
tion. This special shape allows to solve programming, location, smoothing Carnegie Mellon University
efficiently the TR-subproblem. We prove coauthor: Javier Pena
global convergence for our limited mem- In this work, we will consider the par- keywords: Tchebycheff bounds, cones
ory methods with shape changing trust ticular problem of obtaining an optimum of moments, positive polynomials,
region. order p coverage of a plane region by semidefinite programming
 131

Consider the problem of finding optimal hierarchies for mixed and pure 0-1 pro- nonconvex quadratic programming, a
bounds on the expected value of piece- gramming from an algebraic geometry box region is easy to get in which this
wise polynomials over all measures with point of view. As a result, we obtain local minimizer is global. Research on
a given set of moments. This is a spe- a general approximation scheme from such method using for general global
cial class of Generalized Tchebycheff In- which these relaxation hierarchies arise nonlinear programming is presented too.
equalities in probability theory. We study as special cases. For specific mixed and
this problem within the framework of pure 0-1 programs, this generalization
conic programming. Relying on a gen- can be applied to strengthen the sequence A new interval selection tech-
eral approximation scheme for conic pro- of relaxations obtained by any of the nique for global optimization
gramming, we show that these bounds known relaxation hierarchies. Moreover,
can be numerically computed or approx- from this general setting, both new relax- T IBOR C SENDES
imated via semidefinite programming. ation hierarchies and natural generaliza- University of Szeged
We also describe some applications of tions of the existing ones are obtained. keywords: global optimization, interval
this class of Generalized Tchebycheff In- selection
equalities in probability, finance, and in- The convergence properties of interval
ventory theory.
TH2-306/38 GLOBAL OPTIMIZATION

global optimization algorithms are stud-

Global optimization IV ied that select the next subinterval to be
Global optimization of rational chair : Tibor Csendes subdivided with the largest value of the
functions: a semidefinite pro- Global optimization methods indicator p f fk X  . This time the more
gramming approach for approximation by consistent general case is investigated, when the
matrices global minimum value is unknown, and
E TIENNE DE K LERK thus its estimation f k in the iteration k has
Delft University of Technology JÁNOS F ÜLÖP an important role. We present a new tech-
Computer and Automation Institute, nique to ensure convergence to global
coauthor: Dorina Jibetean
Hungarian Academy of Sciences minimizer points even for this case.
keywords: semidefinite programming,
Extensive numerical tests will be
global optimization, rational functions keywords: global optimization,
completed on 40 problems to clear the
We consider the problem of global min- branch-and-bound, decision theory,
positive efficiency impact of this new
imization of rational functions on Rn pairwise comparison
technique.
(unconstrained case), or on a connected The lecture considers a classic problem Csendes, T.: Convergence proper-
semi-algebraic subset of Rn . We show of decision theory, when the priority ties of interval global optimization algo-
that in the univariate case (n  1), these weights of alternatives are determined by rithms with a new class of interval se-
problems have exact reformulations as using pairwise comparisons. We review lection criteria, J. Global Optimization
a semidefinite programming problems the methods proposed for the approxima- 19(2001) 307-327.
(SDP), by using reformulations intro- tion of a pairwise comparison matrix by
duced in the PhD thesis of Jibetean. This consistent matrices. We consider then a TH2-308/11 GENERALIZED
CONVEXITY / MONOTONICITY
extends the analogous results by Nes- special problem, when a consistent ma- Generalized convexity V
terov for global minimization of univari- trix with the least sum of squares of dif-
ate polynomials.  ferences is to be determined to an in- organizer : Laura Martein
For the bivariate case n  2  , we consistent pairwise comparison matrix. chair : Piera Mazzoleni
obtain a fully polynomial time approxi- This is a nonconvex optimization prob- Generalized convexity and de-
mation scheme (FPTAS) for the uncon- lem, and multiple global optimal solu- pendence structures in equilib-
strained problem, if an a priori lower tions are possible. We propose two ap- rium problems
bound on the infimum is known, by us- proaches to solve the problem. First, we
ing results by De Klerk and Pasechnik. formulate the problem as a generalized P IERA M AZZOLENI
For the NP-hard multivariate case, we convex multiplicative problem solvable Catholic University, Milan
discuss semidefinite programming-based as a concave minimization subject to con- keywords: generalized convexity,
heuristics for obtaining lower bounds vex constraints. The classic branch-and- interval orders, equilibrium problems
on the infimum, by using results by bound approach is also applied, where Ambiguous and imprecise knowledge of-
Lasserre. the partition elements are rectangles, and ten affect economic problems and they
to compute bounds, a convex underesti- can be formalized by using fuzzy num-
mation and the concave envelope of the bers. In this approach there are diffi-
Using algebraic geometry in 0-1 ratio function are also used. Computa- culties to estimate the membership func-
programming tional experience is presented. tions. Therefore the theory of interval
J UAN V ERA orders is applied to give a more flexi-
Carnegie Mellon University Box region in which the local ble tool in order to represent a whole
range of values and to compare them.
coauthors: Javier Pena, Luis Zuluaga minimizer is global The role of order relations is fundamen-
keywords: integer programming, tal to develop monotonicity and concav-
relaxation hierarchies, algebraic YOUKANG FANG
University of Botswana & Shenyang ity properties and several applications
geometry can be given to state a qualitative analy-
Institute of Technology,China
Recently, Algebraic Geometry has been sis for parametric optimization problems.
successfully applied to develop new so- keywords: global minimizer, Studies in this area are now in progress
lution schemes for a large class of op- quadratically constrained QP, box region and new results are suggested by several
timization problems. Motivated by this, Using information from computing a lo- fields. It is well known the generaliza-
we study several well-known relaxation cal minimizer of quadratially constrained tion from complementarity problems to
132 Parallel Sessions: Thursday 13:15 – 14:45 (TH2)

variational inequalities up to equilibrium This talk addresses various risk modeling problems of random variables. We re-
problems: indeed, solvability conditions paradigms for two-stage stochastic pro- view the concept of stochastic domi-
are weakened and the range of applica- gramming models. We show that tradi- nance and several related mean-risk mod-
tions is growing. Interval orders offer a tional mean-variance approach may lead els. When turning to specific random
fruitful tool to make comparisons more to non-convex optimization models and variables induced by two-stage random
flexible. New monotonicity and concav- severe computational difficulties. On the mixed-integer programs, besides consis-
ity properies can be stated and an applica- other hand, many risk paradigms that tency with stochastic dominance, struc-
tion is given to a vector equilibrium prob- are consistent with second order stochas- tural properties of risk measures become
lem. tic dominance lead to tractable convex crucial. We investigate two properties
optimization problems. Furthermore, which yield, when satisfied, well-defined
On generalized equilibrium the original decomposability property of and algorithmically amenable stochastic
points the stochastic program is preserved. A programs.
computational comparison of several risk The material is presented in two con-
G ABOR K ASSAY modeling paradigms will be presented. secutive talks.
Eastern Mediterranean University In this talk we investigate mean-
coauthor: Peter Kas Risk aversion in stochastic pro- risk models, which are consistent with
keywords: approximate equilibrium grams with mixed-integer re- stochastic dominance, for the specific
points, approximate saddlepoints, course, Part I random variables induced by two-stage
perturbed convex functions random mixed-integer programs. We
A NDREAS M AERKERT present structural results, an algorithm,
Generalizations of the usual definitions University Duisburg-Essen and computational experience with aca-
of saddlepoint and equilibrium point are demic and real-world problems.
introduced. The existence of these points coauthors: Ruediger Schultz, Stephan
is proved using the Ky Fan’s minimax in- Tiedemann
equality theorem and shown to be related keywords: stochastic programming,
mean-risk models, mixed integer
TH2-308/T1 DYNAMIC PROGRAMMING AND
OPTIMAL CONTROL

to a class of functions that we call per- Dynamic programming and

turbed convex functions. First and sec- programming
optimal control
ond order conditions regarding the exis- Many decision problems involving ran-
tence of these points are also proved. chair : Julien Laurent-Varin
dom data can be regarded as ordering
problems of random variables. We re- Robust dynamic programming
On generalized convexity for set view the concept of stochastic domi- for polytopic systems with piece-
valued maps nance and several related mean-risk mod- wise linear cost
els. When turning to specific random
DAVIDE L A TORRE variables induced by two-stage random M ORITZ D IEHL
Department of Economics mixed-integer programs, besides consis- University of Heidelberg
keywords: generalized convexity, set tency with stochastic dominance, struc-
keywords: dynamic programming,
valued analysis tural properties of risk measures become
robust optimization, polytopic systems,
crucial. We investigate two properties
It is well known that convexity and gen- which yield, when satisfied, well-defined convex programming
eralized convexity play a crucial role in and algorithmically amenable stochastic
We present a new method for robust dy-
many fields. Recently, some papers deal programs.
namic programming that is suitable for
with some extensions of these notions The material is presented in two con- polytopic linear systems and piecewise
to set valued maps. In this talk some secutive talks.
linear convex cost functions. The method
definitions of convexity and generalized In this talk we associate two-stage performs exact dynamic programming
convexity are considered and compared. stochastic mixed-integer programs with
without any discretization of the state
Some characterizations are given by Dini stochastic ordering principles, in partic-
space by a using piecewise linear repre-
and Dini-Hadamard first order general- ular, with stochastic dominance. We use
sentation of the cost-to-go function. It is
ized derivatives. Applications of these the mean-variance model to illustrate that
shown that the robust dynamic program-
concepts in optimization problems and the choice of the risk measure requires
ming recursion preserves this piecewise
variational inequalities are shown. special care when integer variables enter linear representation. Inclusion of lin-
stochastic programs. ear inequality constraints on controls and
TH2-308/12 STOCHASTIC PROGRAMMING
states leads to a simultaneous recursion
Risk issues in stochastic Risk aversion in stochastic pro- of the cost-to-go and of feasible sets in
programming grams with mixed-integer re- state space. It is shown that a polyhedral
organizer : COSP (Stephan Tiedemann) course, Part II representation of the feasible sets is pre-
chair : Stephan Tiedemann served as well under the robust dynamic
S TEPHAN T IEDEMANN programming recursion.
Risk modeling in stochastic pro- University Duisburg-Essen Finally, we show how the method
gramming coauthors: Andreas Maerkert, Ruediger can be applied to an example problem,
Schultz namely a parking car with uncertain mass
S HABBIR A HMED keywords: stochastic programming, that shall quickly reach the end of a
Georgia Institute of Technology garage and stop there without hitting the
mean-risk models, mixed integer
coauthor: Andrzej Ruszczynski programming wall. We also discuss the involved com-
keywords: stochastic programming, putational costs and possible dynamic
risk, stochastic dominance, Many decision problems involving ran- programming approximations based on
decomposition dom data can be regarded as ordering this approach.
 133

rithm, the process being monitored by a problem characteristics that cause diffi-
A relaxation theorem for nonlin- local error estimate. culties in practice.
ear control systems with sliding We will present numerical results
modes showing how behaves the automatic re- Solving complementarity con-
finement procedure for some flight me- straints via nonlinear optimiza-
A NDREI D MITRUK chanics problems such as the atmo- tion
Russian Academy of Sciences, CEMI spheric reentry. The extension of these
keywords: nonlinear operator, metric techniques for future space launcher tra- S VEN L EYFFER
regularity, control system, sliding modes jectory optimization, involving multi-arc Argonne National Laboratory
On a fixed time interval we consider a problems, will be discussed.
keywords: complementarity constraints,
nonlinear control system with additional nonlinear programming
equality constraints on the endpoints of TH2-308/T2 COMPLEMENTARITY AND VARIATIONAL
INEQUALITIES
An exciting new application of nonlinear
trajectory, and also consider a relaxed Mathematical programs with programming techniques is mathemati-
(convexified) system with so-called slid- equilibrium constraints cal programs with complementarity con-
ing modes. The question is, when the organizers : Sven Leyffer and Richard straints (MPCC). Problems of this type
passage to the relaxed system is valid, Waltz arise in many engineering and economic
i.e., when a given trajectory of relaxed chair : Sven Leyffer applications. Recently, it has been shown
system can be approximated by trajecto-
ries of initial system? We prove that if the Solution and application of that MPCCs can be solved efficiently and
operator of relaxed system is nondegen- MPECs in process engineering reliably as nonlinear programs.
We examine several different nonlin-
erate, i.e., its derivative at a given relaxed
trajectory is surjective (the Lyusternik L ORENZ B IEGLER ear formulations of the complementarity
Carnegie Mellon University constraints. Unlike traditional smoothing
condition), and the weight coefficients
coauthor: Arvind Raghunathan techniques, our reformulations do not re-
of sliding modes are essentially positive,
keywords: process engineering, MPEC, quire the control of a smoothing param-
then the relaxed trajectory can be approx-
barrier methods, phase equilibrium eter. Thus they have the advantage, that
imated (uniformly in state and control
the smoothing is exact in the sense, that
variables, and weakly-* in weight coef- The talk focuses on the development of Karush-Kuhn-Tucker points of the refor-
ficients) by trajectories whose weight co- nonlinear programming methods for the mulation correspond to strongly station-
efficients take only two values: 0 or 1. solution of MPECs in process engineer- ary points of the MPCC.
These new trajectories actually present ing. Complementarity constraints for
trajectories of the initial system. This We show how the formulations can be
these problems arise from nested opti- integrated into an SQP method and report
theorem strengthens the known results of mization problems due to phase equilib- on numerical experience on a large range
Bogolyubov, Warga, Gamkrelidze et al, rium, feedback control or phenomeno-
related to systems without additional ter- of test problems.
logical models. The addition of these fea-
minal equality constraints. tures leads to richer optimization models
that cover a wider range of process be- TH2-308/T3
TELECOMMUNICATION NETWORK
The proof is based on a nonlocal esti- DESIGN

mate to the level set of a nonlinear oper- havior. Scheduling and routing in
ator, which in turn is based on a general- We apply an approach that extends networks
ization of the known Lyusternik theorem directly from the IPOPT barrier method chair : Hans-Florian Geerdes
on the tangent subspace. with a filter line search algorithm. The
extended algorithm, called IPOPT-C, has Design of survivable IP-over-
An interior points algorithm for desirable local and global convergence optical networks
optimal control problems properties and performs very well on nu-
merical comparisons. Several examples S YLVIE B ORNE
J ULIEN L AURENT-VARIN drawn from metabolic flux analysis, dis- Université Blaise Pascal -
CNES/ONERA/INRIA tillation optimization and two phase flow Clermont-Ferrand II
coauthors: Frederic Bonnans, Nicolas behavior will be presented. coauthors: Eric Gourdin, Bernard Liau,
Berend, Mounir Haddou, Christophe Ali Ridha Mahjoub
Talbot Interior point methods for keywords: IP-over-optical network,
keywords: interior point, optimal MPECs survivability, integer programming
control, discretization In the past years, telecommunication net-
In this talk we will analyze interior-
G ABRIEL L OPEZ -C ALVA works have seen an important evolution
Northwestern University
point algorithms for solving optimal con- wih the advances in optical technologies
trol problems, and their application to coauthors: Sven Leyffer, Jorge Nocedal and the explosive growth of the Inter-
aerospace problems. The algorithm is keywords: MPEC, interior point net. Several optical systems allow a very
based on a Runge Kutta discretization of method, L1 penalty, nonlinear large transport capacity, and data traffic
the problem and its approximation using programming has dramatically increased. Telecommu-
logarithmic penalties. We discuss different interior-point ap- nication networks are now moving to-
An important feature is that all linear proaches for solving MPECs, paying par- ward a model of high-speed routers in-
systems to be solved involve band matri- ticular attention to an L1-penalty formu- terconnected by intelligent optical core
ces, for which efficient solvers are avail- lation. We present global and local con- networks. Moreover, there is a general
able. Equally important is the quality of vergence results and compare them with consensus that the control plan of the
the discretization. The interior-point al- classical results for nonlinear program- optical networks should utilize IP-based
gorithm allows easily to introduce addi- ming. Numerical experiments on large- protocols for dynamic provisioning and
tional points at any iteration of the algo- scale MPECs are presented. We identify restoration of lightpaths. The interaction
134 Parallel Sessions: Thursday 13:15 – 14:45 (TH2)

of the IP routers with the optical core TH2-341/21 in contrast to primal-dual interior point
INTERIOR POINT ALGORITHMS

networks permits to achieve end-to-end methods for CPs, the analysis deeply de-
connections. And the lightpaths of the Recent development in pends on the monotonicity and the exten-
optical networks provide the topology of interior-point methods sion to more general classes of CPs, e.g.,
the virtual network interconnecting the organizer/chair : Jiming Peng P CPs or P0 CPs, is difficult as it is. In
IP routers. This new infrastructure has
to be sufficiently survivable, so that net-
What is special with the loga- this paper, we propose a new homoge-
neous model which can be applied to P0
work services can be restored in the event
rithmic barrier function in opti- CPs, and show that (i) there exists a tra-
of a catastrophic failure. This paper ad-
mization? jectory which leads to a complementar-
dresses survivability issues that may be K EES ROOS ity solution of the homogeneous model,
of practical interest for IP-over-optical TU Delft (ii) from the solution, we can obtain a
neworks. We consider some optimiza- complementarity solution of the original
tion problems concerning the topology coauthors: Y. Bai, M. Elghami problem, under the assumption that the
of these networks. We give integer pro- keywords: primal-dual method, interior original problem is a strictly feasible P
gramming formulation of the problem. point method, barrier function CP and (iii) any positive point, feasible or
We also present some computational re- After its introduction by Frisch in 1955, infeasible, can be set as a starting point of
sults and discuss some extensions. the logarithmic barrier function (LBF) the trajectory and it does not need to use
has played a major role in the field any big-M parameter.
optimization. The revolutionary devel-
The potential of relaying cellu- opments of the past two decades in Self-regular proximity based
lar wireless computer networks this field, which gave rise to the sub- dynamic large-update primal-
field of interior-points (IP) methods, dual ipms for linear optimiza-
H ANS -F LORIAN G EERDES has re-emphasized its importance. The tion
Konrad-Zuse-Zentrum für search directions in all state-of-the-art IP-
Informationstechnik Berlin solvers for linear, and also for second- TAMAS T ERLAKY
order cone and semidefinite optimiza- Dept. of Computing and Software,
coauthor: Holger Karl tion problems are explicitly or implic- McMaster University
keywords: telecommunication, integer itly based on an LBF, and the analysis coauthor: Maziar Salahi
programming, scheduling of these methods strongly depends on keywords: linear optimization, large
properties of such functions. Other bar- update, interior point method
Bandwidth is a scarce resource in wire- rier functions have been proposed, but Primal-Dual Interior-Point Methods
less computer networks. Relaying is a both from a theoretical and computa- (IPMs) are the method of choice when
protocol extension designed for utilizing tional viewpoint LBF’s always were win- one solves large scale optimization prob-
radio resources more efficiently in a cel- ning, at least surviving. It has often lems. The gap between large-update
lular setting: several hops are allowed been asked what makes LBF so special. and small-update IPMs was narrowed re-
within one cell. In traditional networks In this talk we deal with this question. cently by Peng, Roos and Terlaky by in-
only direct transmission between base We focus on primal-dual methods for lin- troducing and analyzing a novel class of
station and mobile terminals is feasible. ear optimization. It is probably for the so-called Self-Regular IPMs (SR-IPMs).
It has been shown that relaying can in- first time that alternative barrier functions In this paper a family of dynamic,
crease data throughput significantly for have been found that in some cases pro- large-update primal-dual IPM is pro-
special cases, but the average potential is vide better theoretical complexity results posed where the proximity functions and
an open question. than the LBF. The results can be extended the search directions are defined by one
In our work we discuss the gen- to other conic optimization problems; it of the specific families of SR func-
eral potential of relaying. We ap- is an open question if the new barrier tions. The results are generalizations
proach this question by defining the op- functions can be adapted to primal meth- of the dynamic SR-IPM algorithm pro-
timization problem of finding a feasi- ods and dual methods, respectively. posed recently by Peng and Terlaky.
ble transmission schedule that maximizes  q 1

the data throughput under fairness con- An O qn 2q log nε  worst case iteration
ditions. A special mathematical model A homogeneous model for p0 bound is established, where q is the bar-
and a column-generation based algorithm and p  nonlinear complemen- 
rier degree of the SR proximity. For
is presented for this problem. This en- tarity problems the special case q  log n  the
 so far
ables us to quantitatively compare the best known complexity O n log n  log nε 
A KIKO YOSHISE of large-update IPMs is established.
data transmission capacity with and with- Institute of Policy and Planning
out relaying of concrete networks con- Sciences, University of Tsukuba
sisting of multiple cells. TH2-341/22 LOGISTICS AND TRANSPORTATION

We apply our methodology to ex- keywords: homogeneous model Routing by mixed integer
ample networks and assess the average The homogeneous model for linear pro- programming
potential of relaying under the assump- grams provides a most simple and firm chair : Jean-Francois Cordeau
tion of uniform node placement. This theory in interior point algorithms, and
yields statistical evidence that relaying several computational experiences report Modelling of road investments
increases the average data throughput its practical efficiency. In 1999, An- J ENNY K ARLSSON
in an example network class by 30%. dersen and Ye generalized this model Linkøping University
In addition, the method presented here to monotone complementarity problems coauthors: Mikael Rönnqvist, Mathias
can serve as an engineering benchmark (CPs) and showed that the model inher- Henningsson
for any concrete scheduling/routing algo- its most of desirable properties under the keywords: forestry, road investments,
rithm for relaying. monotonicity of the problems. However, modeling, mixed integer programming
 135

 
Public and private roads, which have an solution approach based on column gen- O 2d n2  and space O 22d n  . We also
insufficient bearing capacity or blocking eration and some computational results give a partial answer to the second ques-
due to thawing or heavy rains, annu- are presented. tion, by presenting a lower bound of
ally contribute to a considerable loss in 1.666 on the approximation ratio of the
Swedish forestry sector. The forestry is optimal double-tree solution. Compu-
forced to build up large stocks of raw ma-
A branch-and-cut approach for tational experiments show that the op-
terial in order to secure a continuous sup-
the dial-a-ride problem timal double-tree algorithm is practical,
ply during periods of uncertain accessi- J EAN -F RANCOIS C ORDEAU providing one of the best known tour-
bility to parts of the road network. One HEC Montreal constructing heuristics.
method to reduce costs related to road
keywords: dial-a-ride, branch-and-cut,
blocks, is to made investments, i.e. up-
vehicle routing, time windows Covering minimum spanning
grade the road network, to a standard trees of induced subgraphs
that will guarantee accessibility. We con- In the dial-a-ride problem, users formu-
sider the road investment problem from late transportation requests between an M ICHEL G OEMANS
the perspective of Swedish forest compa- origin and a destination. Transportation M.I.T.
nies. The objective is to minimize the in- is carried out by a fleet of vehicles that coauthor: Jan Vondrak
vestment and overall transportation costs, provide a shared service. The problem is keywords: combinatorial optimization,
given that demanded volumes of raw ma- to design a set of minimum cost vehicle minimum spanning trees, probabilistic
terial must be accessible along roads of routes satisfying capacity, duration, time analysis
a certain bearing class. This problem is window, pairing, precedence and ride In a variety of optimization settings, one
strategic. The planning period is typi- time constraints. We propose a mixed- has to repeatedly solve instances of the
cally one decade. We present a math- integer programming formulation of the same problem in which only part of the
ematical model for this problem. The problem. We then describe a branch-and- input is changing. It is important in
problem becomes a mixed integer prob- cut approach that uses new valid inequal- such cases to perform a precomputation
lem or a network design problem. Pre- ities for the problem as well as known involving only the static part and as-
liminary numerical results will be pre- valid inequalities for the TSP with pre- sumptions on the dynamic part, allowing
sented. cendence constraints and the VRP with to speed-up the repeated solution of in-
time windows. stances.
We consider here the minimum span-
A ship routing problem in the ning tree problem, when instances are re-
pulp industry TH2-341/23 APPROXIMATION ALGORITHMS

peatedly drawn either randomly or deter-

Approximation and trees ministically from a given graph. The goal
H ENRIK A NDERSSON chair : Abdel Lisser is to prove the existence and exhibit a
Department of Science and Technology / The double-tree approximation sparse subgraph of edges guaranteed to
Linköping University for metric TSP: is the best good contain the MST of all (or almost all)
coauthor: Peter Värbrand enough? subgraphs generated. The two random
keywords: ship routing, column models that we consider are subgraphs
generation V LADIMIR D EINEKO induced by a random subset of vertices
The University of Warwick or edges, each present with probability
In this presentation, a ship routing prob- coauthor: Alexander Tiskin p. In both random cases, we prove that
lem in the pulp industry is introduced. keywords: traveling salesman problem,  is a covering set Q of cardinality
there
This year, the Swedish forest company double-tree algorithm O n logb n  where b  1 1 p  and this
Södra will produce around two million is asymptotically optimal. We also de-
tons of pulp in over 20 different quali- The Metric Traveling Salesman Problem rive a randomized algorithm which calls
ties. The pulp is produced in five dif- (TSP) is a classical NP-hard optimiza- an MST subroutine only a polylogarith-
ferent mills, three in Sweden and two tion problem. The double-tree heuristic mic number of times, and finds such a
in Norway. The companies buying the allows one to approximate the solution sparse covering set with high probability.
pulp are located all over the world, but of the Metric TSP within a factor of 2, The random results also imply results in
most of them are located in Europe. This by picking an arbitrary solution from a the deterministic setting.
study concerns the distribution of pulp restricted space of solutions to the orig-
to European customers. The distribution inal problem. Such an approach raises Robust minimum spanning tree
of pulp from the mills is done either by two natural questions: can we find effi-
trains or trucks to domestic customers, ciently a solution that is optimal in the re- A BDEL L ISSER
or by ships to harbor terminals in Eu- stricted space? will this optimal solution LRI, Universite de Paris IX
rope. At the harbors, the pulp is either provide a better approximation ratio than coauthors: Lisser Abdel, Kenyon Claire
picked up by the customers or delivered an arbitrary solution? Paper R.E Burkard keywords: robust optimization,
to the customers by Södra. The distri- et al. (1999), "The traveling salesman approximation algorithms, minimum
bution planning problem is to combine problem and the PQ-tree", Mathemat- spanning trees
demands from different customers into ics of Operations Research, 24(1), an- We Propose a new randomized algorithm
shiploads and to decide where to load and swers the first question in the affirma- for solving the robust spanning tree prob-
land the ships and how to coordinate the tive, by presenting an algorithm that finds lem, hereafter called RST P. The robust-
land and sea transports. Another question 
the optimal double-tree solution to Met-  ness in our problem is expressed by the
is when and where to use the own fleet of ric TSP in time O 2d n3  and space O n2  , means of a set of weights for each edge of
ships and when to buy transports on the where d is the maximal vertex degree the underlayer graph. We model RST P as
spot market. The problem is modeled as a in the corresponding minimal spanning a 0 1
linear programming problem.
column based mixed integer program. A tree. We improve these bounds to time As the size of the LP is very large, we use
136 Parallel Sessions: Thursday 13:15 – 14:45 (TH2)

Benders Decomposition to solve its re- the concepts are equivalent but all four of We are concerned with multiobjective
laxation in order to get the lower bounds. them are not equivalent in general. It is fractional variational programming prob-
We then apply our randomized algorithm NP-complete to decide whether a single lems. A class of generalized convexity
to compute the upper bounds. Complex- job is critical, whereas it is easy to find a functions are introduced. Based upon
ity as well as Numerical results are given critical set. this concept of generalized convex, opti-
for instances with up to 100 nodes fully We also focus on scheduling mality conditions and duality results are
meshed graphs. AND/OR-networks on identical paral- established.
lel machines. For the objective of mini-
TH2-321/053 PRODUCTION AND SCHEDULING mizing the makespan, the problem with A projected gradient method for
AND/OR precedence constraints can be vector optimization
Scheduling AND/OR reduced to the problem with standard
constraints precedence constraints. In contrast to
L UIS M AURICIO G RAÑA
chair : Valery Gordon this, a reduction from a special disassem-
D RUMMOND
Rio de Janeiro Federal University
bly problem provides us with a strong
Scheduling with or-precedence inapproximability result for the total coauthor: Alfredo Noel Iusem
constraints weighted completion time objective. keywords: vector optimization, weak
efficiency, projected gradient method,
B ERIT J OHANNES k-convexity
Technische Universität Berlin Assignment of due dates in We propose a method for solving con-
keywords: scheduling, OR networks, scheduling partially ordered strained vector optimization problems in
computational complexity jobs a finite dimensional linear space, with
In various scheduling contexts, VALERY G ORDON preference order induced by a closed
AND/OR-networks are a relevant gen- United Institute of Informatics Problems, convex cone. When the objective is
eralization of ordinary precedence con- National Academy of Sciences of scalar-valued, the procedure coincides
straints (AND constraints). Accordingly, Belarus with the classical projected gradient
scheduling problems with AND/OR- method. For convex objectives (respect
coauthors: Jean-Marie Proth, Vitaly to the cone), we show that the algorithm
constraints inherit the computational Strusevich
complexity of their classic ancestors generates sequences which converge to
keywords: scheduling, due date weak efficient points, no matter how poor
with AND-constraints. On the other assignment, precedence constraints,
hand, the complexity status of var- is the initial guess.
single machine
ious scheduling problems with OR-
constraints has remained open. In this A single machine due date assignment On convex functions with values
talk, we present a couple of complex- and scheduling problem is considered in semi-linear spaces
ity results for scheduling unit-time jobs where due dates are determined by in-
creasing the processing times of jobs by
A NDREAS L OEHNE
subject to OR-precedence constraints. Martin-Luther-Universitaet
In particular, we give a polynomial-time a common positive slack and jobs are
Halle-Wittenberg
algorithm for minimizing the makespan partially ordered according to the given
precedence constraints. The objective is keywords: set valued optimization, semi
and the total completion time on identical
to explore the trade-off between the size linear spaces, convex programming
parallel machines. However, minimizing
the total weighted completion time is of the slack and the arising holding costs It holds the following well–known re-
NP-hard, even on a single machine. for the early jobs, and to find an optimal

sult of convex
analysis: If the func-
schedule that minimizes holding costs tion f : X  ∞  ∞ is convex
 and
provided that there are no tardy jobs and some x0  core dom f  satisfies f x0 
Scheduling with AND/OR- precedence constraints are respected. A ∞, then f never takes the value ∞.
networks polynomial-time solution is proposed for From a corresponding theorem for con-
the series-parallel and somewhat wider vex functions with values in semi–linear
VANESSA K AEAEB spaces a variety of results are deduced,
Technische Universitaet Berlin class of precedence constraints.
among them the mentioned theorem, a
keywords: networks, sensitivity theorem of Deutsch and Singer on the
analysis, scheduling, approximation TH2-321/033 MULTICRITERIA OPTIMIZATION

single–valuedness of convex set–valued

In contrast to standard digraphs, an Convexity and generalized maps as well as a result on the compact–
AND/OR-network has two different sorts convexity in multicriteria valuedness of convex set–valued maps.
of nodes, AND- and OR-nodes, which optimization
represent the following precedence con- chair : Andreas Loehne TH2-321/133 LINEAR PROGRAMMING

straints. An AND-node is just a normal Efficiency conditions and dual- Numerical approaches I
node with the meaning that the job it rep- ity for multiobjective fractional chair : Neelam Gupta
resents can be processed as soon as ev- variational programming prob- Solving DEA models by the
ery job preceding it in the graph has been lems
completed. An OR-node on the other Fourier Motzkin elimination
hand can be scheduled as soon as at least K HADIJA K HAZAFI method
one of its predecessors has been finished. Université Hassan II, faculté des G AUTAM A PPA
First, we define four versions of crit- sciences Ain Chock, Casablanca, London School of Economics
ical sets in AND/OR-networks. The sys- Morocco
coauthor: H.Paul Williams
tems of critical sets of an AND/OR- keywords: multiobjective fractional, keywords: data envelopment analysis,
network form two pairs of blocking clut- variational programming, duality, linear programming, Fourier Motzkin
ters. It follows that in each case two of optimality method
 137

When solving Data Envelopment Anal- G IDEON W EISS In this paper, we present a strikingly sim-
ysis (DEA) models it is usual to solve University of Haifa ple numerical algorithm called UNA for
a Linear Programme (LP) many times, keywords: simplex, continuous linear computing a feasible solution or detect-
with different right-hand-side (RHS) vec- programming ing infeasibility of linear constraints in
tors: once for each Decision Making Unit We consider the separated continuous real variables. We show that UNA con-
(DMU) being evaluated. This iterative linear programming problem (SCLP) verges for solving linear constraints in
approach gives little insight into the over- with linear data as formulated originally real variables. We present experimen-
all structure of the model for the organi- by Bellman 1953 and by Anderson 1979 tal comparison of the average time per-
sation. Instead, by projecting out all the and investigated by Pullan 1993. We formance of UNA, with two commercial
variables of the LP which are common characterize the form of its optimal so- linear programming packages XA and
to all LP runs, we obtain a formula into lution, and present an algorithm which CPLEX, for solving randomly generated
which we can substitute the inputs and solves it in a finite number of steps, using constraint sets.
outputs of each DMU in turn to obtain simplex pivot operations. To do so we:
its efficiency number and efficient com- In our experiments, CPLEX per-
(a) Formulate a symmetric dual problem. formed better than UNA, and UNA per-
parators. In addition we also obtain the (b) Find that a finite sequence of LP bases
best weightings which the DMU would formed better than XA in solving larger
plays the role of a basis for SCLP. (c) De- feasible constraint sets. As the size of
choose to put on its inputs and outputs. fine neighboring base-sequences through
The method of projection, which we use, constraint sets increases, the average time
their validity regions. (d) Construct an taken by UNA is only about 2 to 3
is Fourier-Motzkin Elimination. This algorithm to pivot between neighboring
provides us with a finite number of ex- times the average time taken by CPLEX,
base-sequences. (e) Find an algorithm whereas the time performance of XA
treme rays of the elimination cone. These to solve SCLP by a sequence of pivots
rays give the dual multipliers which can becomes much worse than both UNA
analogous to the parametric simplex al- and CPLEX. In experiments with detect-
be interpreted as weights which will ap- gorithm of LP.
ply to the inputs and outputs for partic- ing infeasibility of large constraint sets,
CPLEX performed far better than UNA
ular DMUs. As the approach provides Una: a simple numerical algo- and XA.
all the extreme rays of the cone, multi- rithm to solve linear constraints
ple sets of weights, when they exist, are in real variables Our experiments indicate that UNA is
explicitly provided. In addition it is pos- N EELAM G UPTA a simple but practical algorithm to solve
sible to construct the skeleton of the effi- The University of Arizona linear constraints. Most of its computa-
cient frontier of efficient DMUs. tion time is spent when it is already very
coauthors: YongJun Cho, Mohammad
close to a feasible solution. It would be
Z. Hossain
worthwhile to explore variants of UNA
A simplex algorithm for sep- keywords: solving linear constraints,
that cut down the computation time after
arated continuous linear pro- numerical algorithms, linear
it is in close proximity to a feasible solu-
grams programming
tion.
138 Parallel Sessions: Thursday 16:30 – 18:00 (TH3)

TH3-302/41 COMBINATORIAL OPTIMIZATION We present preliminary computa- presents a mathematical model for the
tional results where we combine inte- whole problem of finding a protected
Semidefinite liftings in rior point methods with the bundle con- pattern with minimum loss of informa-
combinatorial optimization cept. These results indicate that this tion, and proposes a branch-and-cut al-
organizer/chair : Miguel Anjos model is manageable for sparse graphs gorithm to solve it. It also describes
on a few hundred nodes and yields very a new methodology closely related with
Solving the satisfiability prob- tight bounds. the classical controlled rounding meth-
lem using semidefinite program- ods but with several advantages. The
ming new methodology leads to a different op-
Semidefinite representations for
M IGUEL A NJOS finite varieties timization problem which is simpler to
Faculty of Mathematical Studies solve than the previous problem. We
M ONIQUE L AURENT present a cutting-plane algorithm for
keywords: semidefinite programming CWI finding an exact solution of the new prob-
The satisfiability (SAT) problem is a keywords: semidefinite relaxation, sum lem, which is a pattern guaranteeing the
central problem in mathematical logic, of squares of polynomials same protection level requirements but
computing theory, and artificial intelli- We propose a semidefinite formulation, with smaller loss of information when
gence. An instance of SAT is speci- involving combinatorial moment matri- compared with the classical controlled
fied by a set of boolean variables and ces, for the problem of optimizing a poly- rounding optimal patterns. Considering
a propositional formula in conjunctive nomial function over a semi-algebraic the solutions of the here-introduced two
normal form. Given such an instance, set  defined by polynomial equalities problems, the auditing phase is unneces-
the SAT problem asks whether there is (h j x   0, j  1   m) and inequali- sary. We conclude with computational
a truth assignment to the variables such ties, in the case when the polynomials results on randomly generated instances
that the formula is satisfied. The gen- h1   hm have a finite number of common proving the good performance of the pro-
eral SAT problem is well known to be posal.

zeroes with single multiplicities. In the
NP-complete, although several important 0/1 case (h j x   x2j x j ) and 1 case
special cases can be solved in polynomial 
time. (h j x   x2j 1), we find the semidef-
inite representation underlying lift-and- The graph disconnectivity prob-
We propose semidefinite program- lem: an improved algorithm
ming relaxations for the satisfiability project methods (Sherali-Adams, Anjos-
Wolkowicz,..) and algebraic methods via and new models
problem within a paradigm of “higher
liftings” for combinatorial optimization sums of squares of polynomials (Nes-
problems. We first discuss the ability terov, Lasserre, Parrilo,..). Semidefi- YOUNG -S OO M YUNG
of these relaxations to prove whether or nite approximations are obtained by con- Dankook University
not a given instance of SAT is satisfi- sidering truncated moment matrices; we
able. Then, from a more practical point present rank conditions ensuring that the keywords: graph disconnectivity
of view, we explore the amenability of approximation solves the original prob-
these relaxations to practical computa- lem at optimality, with applications to The graph disconnectivity problem is to
tion. max-cut and the maximum stable set find a set of edges such that the total cost
problem. of removing the edges is no more than a
given budget and the weight of nodes dis-
Semidefinite relaxations for
sparse max-cut problems TH3-302/42 COMBINATORIAL OPTIMIZATION
connected from a designated source by
Combinatorial optimization removing edges is maximized. Martel et
A NGELIKA W IEGELE IX al. have shown that the problem with unit
University of Klagenfurt capacity and unit demand is NP-hard and
chair : Safia Kedad-Sidhoum
Myung and Kim present an integer pro-
coauthor: Franz Rendl Controlled rounding and other gramming formulation and develop an
keywords: semidefinite programming new statistical disclosure limita- algorithm that includes a preprocessing
Let be given an edge-weighted undi- tion methods for tabular data procedure and lower and upper bound-
rected graph. The Max-Cut Problem con- ing strategies. In this talk, we present
sists in finding a partition S and V § of the
J UAN J OSÉ S ALAZAR new findings on the properties of the op-
vertex set V in such a way that the to- G ONZÁLEZ timal solution, based on which a new al-
University of La Laguna
tal weight of the edges joining S and V § gorithm is developed. We also introduce
is maximized. The Max-Cut Problem is keywords: statistical disclosure control new models that have different objectives
known to be NP-complete. Rounding methods are common tech- and required conditions.
We investigate semidefinite relax- niques in many statistical offices to pro-
ations of the Max-Cut problem, which tect sensitive information when publish-
are formulated in terms of the edges of ing data in tabular form. These methods A new lower bound for the one
the graph, thereby exploiting the (poten- do not consider protection levels while machine problem with earliness
tial) sparsity of the problem. We show searching patterns with minimum infor- andtardiness penalties
how this is related to higher order liftings mation loss, and therefore typically the
of Anjos and Wolkowicz and Lasserre. so-called auditing phase is required to
Contrary to the basic semidefinite relax- check the protection of the proposed pat- S AFIA K EDAD -S IDHOUM
ation, which is based on the nodes of the terns. This work addresses new ver- LIP6 - Pierre Marie Curie University
graph, the present formulation leads to a sions of the classical Controlled Round-
model which is significantly more diffi- ing Methodology which guarantee pro- keywords: one machine scheduling,
cult to solve. tection levels implicetely. We will also earliness/tardiness, weighted stable set,
interval graphs
 139

We address the one-machine problem in to develop a model with a stronger linear problem of finding an edge set of min-
which the jobs have distinct due dates, programming relaxation. The linear pro- imum weight that partitions into k dis-
earliness costs and tardiness costs. More gramming gaps for the tightened model joint spanning trees. We give an algo-
precisely, a set of jobs has to be sched- are very small, typically less than 0.5%, rithm whose complexity is dominated by
uled on a single machine. Each job has and are usually one third to one tenth of V  applications of the minimum cut algo-
a processing time. A schedule is a se- the gaps of the best model previous de- rithm of Goldberg and Tarjan.
quence of the starting times of all the jobs scribed in Gouveia and Magnanti (2003).
subject to disjunctive constraints, that is Moreover, using the new model, we are
two jobs cannot be processed at the same able to solve large Euclidean problem in- TH3-302/45 INTEGER AND MIXED INTEGER
PROGRAMMING

time by the machine. A job is then sched- stances that are not solvable by the previ- Industrial applications
uled, without preemption, from its start- ous approaches. chair : Ken McKinnon
ing time to its completion time. A cost
function is attached to each job includ- An integer programming solu-
Flexibility of Steiner minimum tion to optimally locating ra-
ing the earliness and tardiness penalties. trees in uniform orientation dioactive waste facilities
The total cost of the schedule is the sum metrics
of the costs of all the jobs. The prob-
lem is to find a schedule with a mini- J OHN K ARLOF
M ARCUS B RAZIL University of North Carolina at
mum cost. This problem arises in just- The University of Melbourne
in-time production environment. A time- Wilmington
coauthors: Pawel Winter, Martin
indexed formulation is considered and a keywords: integer programming,
Zachariasen
new lower bound based on lagrangian re- applications
keywords: geometric network design,
laxation is proposed. This bound is based
networks, combinatorial optimization, The American 1980 Low-level Radioac-
on the observation that the lagrangian re-
VLSI design tive Waste (LLRW) Policy Act suggests
laxation of a set of job constraints leads
to a minimum weighted stable set prob- We present some fundamental flexibil- several constraints for choosing LLRW
lem on an interval graph. This problem is ity properties for minimum length net- dump sites, such as three states con-
efficiently solvable by dynamic program- works (known as Steiner minimum trees) taining dump sites, single sourcing, and
ming. Computational results will be re- interconnecting a given set of points in equally distributing the waste between
ported. an environment in which edge segments the three sites. The optimization criteria
are restricted to λ uniformly oriented di- is to minimize the number of ton-miles
rections. We refer to these networks of LLRW transported. There is an ex-
TH3-302/44 COMBINATORIAL OPTIMIZATION

as λ-SMTs. They will play an increas- tremely large number of ways to choose
Minimum trees ingly important role in the future of op- the three states for sites and the site each
chair : Francisco Barahona timal wire routing in VLSI physical de- state will send its waste to. We model
A 2-paths approach for odd- sign, particularly for the next generation this problem as an integer programming
diameter-constrained minimum of VLSI circuits, which will make perva- problem and present a branch and bound
sive use of "diagonal" interconnects. algorithm using penalty fathoming for its
spanning and Steiner trees
In this paper we develop the con- solution.
C RISTINA R EQUEJO cept of an extremal polygon for a λ-SMT,
University of Aveiro which is a region representing the union Optimal routing of oil and gas
of all λ-SMTs with the same topology on from wells to separators
coauthors: Luis Gouveia, Thomas
a given set of points. We show that this
Magnanti
polygon can be constructed, for a given B JØRN N YGREEN
keywords: diameter-constrained trees,
point set and given topology, in linear NTNU
network flows, hop-indexed models
time. We the discuss the applications of
Using underlying graph theoretical prop- this polygon, which can be thought of as coauthor: Ken McKinnon
erties, Gouveia and Magnanti (2003) de- a geometric representation of the amount keywords: mixed integer programming,
scribed several network flow-based for- of flexibility inherent in a given λ-SMT. petroleum/natural gas
mulations for diameter-constrained tree In particular, we show that it can be used Networks of pipelines are used to connect
problems. Their computational results to develop more powerful pruning tech- oil and gas wells to central production
showed that, even with several enhance- niques for exactly constructing a λ-SMT plants where the oil, gas and water are
ments, models for situations when the for a given set of points. We also dis- separated. We consider networks where
tree diameter D is odd proved to be more cuss the ways in which it can be used to there are alternative routings between the
difficult to solve than those when D is solve multi-objective optimisation prob- wells and the separation plants. The out-
even. We provide an alternate model- lems in VLSI physical design, involving puts of wells depend nonlinearly on the
ing approach for situations when D is minimising congestion or signal delay as pressure in the gathering pipelines.
odd. The approach is based on a new a secondary objective. We discuss the problem finding the
(although easily proved) graphical prop- valve setting which maximize the output
erty and views the diameter-constrained
minimum spanning tree as being com-
Network reinforcement from the network and show how to model
this as a MILP. We describe how to ob-
posed of a variant of a directed span- F RANCISCO BARAHONA tain sharper formulations of parts of the
ning tree (from an artificial root node) IBM Research problem.
together with two constrained paths, a
keywords: spanning trees
shortest and a longest path, from the root 
node to any node in the tree. We show Given a graph G  V E  with edge- Branch-and-price to solve a
how to tighten the resulting formulation weights, and a number k, we study the MILP feed mill problem
140 Parallel Sessions: Thursday 16:30 – 18:00 (TH3)

K EN M C K INNON with graph-density equal to one. In this it NP-hard. We compare several res-
Edinburgh University talk, we demonstrate that high density is olution techniques such as linear pro-
coauthor: Andreas Grothey not a feature that is special to ’combina- gramming (standard linear relaxation, la-
keywords: branch-and-price, MILP, torial’ 0/1-polytopes. In fact, the graph grangian relaxation) and constraint pro-
decomposition, feed mill 
density of a d-dimensional random 0/1- gramming schemes (standard constraints,
polytope with n d  vertices tends to one
This talk will describe a problem which   global constraint). Finally we present
arises in feed mills of producing mini- (with d going to infinity) if n d   2 an innovative way of hybridizing global
ε  d holds for some ε 0,  while it tends constraints and local search making use
mum cost diets when there are some dis-  
to zero in case of n d 
2  ε  d . The
of incremental flow algorithm.
crete constraints on the use of raw mate-

rials. cut-polytopes have n d   const d ver-
A Branch and Price method for the tices. A combined mixed integer pro-
problem in which the price step is solved gramming and constraint pro-
as a MILP will be described. This allows gramming approach to planning
the problem to be decompose so that each Point containment in the integer and scheduling in a paint com-
diet can be treated separately. A compar- hull of a polygon pany
ison will be given of this with a standard
MILP approach to the non-decomposed E RNST A LTHAUS M IGUEL C ONSTANTINO
problem. MILP Max-Planck Institute University of Lisbon
coauthors: Friedrich Eisenbrand, Stefan keywords: production planning,
TH3-302/49 INTEGER AND MIXED INTEGER
PROGRAMMING
Funke, Kurt Mehlhorn scheduling, mixed integer programming,
keywords: integer programming, constraint programming
Theory of 0/1 and integer complexity, fixed dimension, separation
programming   We consider the integrated planning and
organizer/chair : Friedrich Eisenbrand We provide a O m  log m  s  time algo- scheduling of production lots in a paint
rithm to decide whether a given rational company. While the production plan es-
Fast integer programming in point is contained in the integer hull of tablishes the production lots for a hori-
fixed dimension a two-dimensional convex polygon, de- zon of 1 to 2 weeks, the scheduling of the
fined by m constraints, each involving  co- lots (tasks) on the machines is done daily.
F RIEDRICH E ISENBRAND efficients of binary encoding length O s  .
MPI-Informatik The approach considered uses a Mixed
For comparison, the equivalence of sepa- Integer Programming model to obtain so-
keywords: integer programming, lattice ration and optimization yields a running
  lutions for the planning problem, and
basis reduction time of O m s  logm  s2  for this task. uses Constraint Programming to check
It is shown that the optimum of an in- We certify containment with a fea- for the existence of daily feasible sched-
teger program in fixed dimension with a sible unimodular triangle. This can ules. In case such a schedule does not
fixed number of constraints can be com- be applied to obtain a faster algorithm exist, this information is incorporated in
puted with O s  basic arithmetic oper- for multiprocessor scheduling with two the MIP model, and the procedure is it-
ations, where s is the binary encoding job lengths, previously studied by Mc- erated. We present results with real data
length of the input. This improves on the Cormick, Smallwood and Spieksma. from the paint company.
quadratic running time of previous algo-
rithms, which are based on Lenstra’s al-
gorithm and binary search. TH3-306/31 CONSTRAINT LOGIC PROGRAMMING
Scheduling the national football
It follows that an integer program Constraint programming league with constraint program-
in fixed dimension, which is defined by chair : Irvin Lustig ming
m constraints, each of binary encoding
length at most s, can be solved with Optimizing TV-break packages I RVIN L USTIG
an expected number of O m  log m  s  offered by satellite channels ILOG Direct
arithmetic operations using Clarkson’s keywords: constraint programming,
random sampling algorithm. E RIC B OURREAU sports scheduling
Bouygues e-lab
The National Football League (NFL)
The graph-density of random coauthors: Bourreau Eric, Rottembourg consists of 32 teams, with each team
0/1-polytopes Benoit, Benoist Thierry
keywords: constraint programming, playing a predetermined set of 16 games
VOLKER K AIBEL linear relaxation, combinatorial and one bye over 17 weeks. The NFL
TU Berlin optimization has to schedule these games to meet the
demands of the teams as well as the tele-
coauthor: Anja Remshagen The TV breaks packing problem con- vision networks. We describe how con-
keywords: combinatorics, polytope, sists, for a TV channel, in building straint programming has been success-
random spot packages satisfying marketing con- fully applied to solve this problem.
One of the discoveries that were es- straints whilst maximizing the total sell-
pecially exciting for the early polyhe- ing price. A package is a fixed amount
dral combinatorialists was the observa- of messages satisfying both campaign TH3-306/32 NONLINEAR PROGRAMMING

tion that many 0/1-polytopes associated shapes constraints and audience require- Surrogate modelling for
with combinatorial optimization prob- ment (minimum number of forecasted engineering optimization II
lems have graphs with very small di- spectators). This problem is modelled organizer/chair : Kaj Madsen
ameters. Particularly striking examples with a single flow together with ad-
are the cut-polytopes of complete graphs ditional covering constraints that make Surrogate modelling by Kriging
 141

H ANS B RUUN N IELSEN In this paper we use Lagrange interpola- utilized methods (discretization of state
IMM tion polynomials to obtain good gradient AND of admissible controls) -Approach
coauthors: Søren N. Lophaven, Jacob estimations when the underlying func- applicable in 1,2 and 3 spatial dimen-
Søndergaard tion is noisy. Good gradient estimation sions, and also for Galerkin type dis-
keywords: surrogate modeling, is e.g. important for nonlinear program- cretization schemes of time dependent
response surface, Kriging ming solvers. As an error criterion we state equations -Numerical analysis se-
take the mean squared error. This error riously simplifies compared to the com-
The Kriging method has wide applica- can be split up into a deterministic and a mon approach
tions in statistics, and was introduced as stochastic error. We analyze these errors As numerical solution algorithms
a tool for modelling response surfaces using (N times replicated) Lagrange in- primal-dual active set strategies and
about 15 years ago. terpolation polynomials. We show that semi-smooth Newton methods are dis-
the MSE is of order N  1  2d if we use
1
In the talk we present the Kriging cussed. Several numerical examples will
method in a common framework with ra- 2d evaluations in each replicate. The be presented that confirm the theoretical
dial basis functions and multivariate in- value d  1 corresponds with finite for- investigations.
terpolating polynomials. Further, we give ward differencing, and the corresponding References: M. Hinze: A generalized
a numerical analyst’s view of some of the discretization concept for optimal con-
MSE order N  2 is well-known in the lit-
1
problems that we had to cope with dur- trol problems with control constraints,
ing the development of DACE: a Matlab erature. Moreover, note that as a result Preprint Math–NM–02-2003, Institut für
toolbox for Design and Analysis of Com- the order of the MSE converges to N  1 if Numerische Mathematik, TU Dresden,
puter Experiments. By proper choice of the number of evaluation points increases Germany.
a correlation model it is possible get very to infinity. Moreover, we show that our
good approximations, but you have to approach is also useful for deterministic
functions in which numerical errors are Adaptive finite elements for
work with a very ill conditioned matrix. control of convection-diffusion
Finally, we give some preliminary re- involved. Finally, we consider the case
of a fixed budget of evaluations. For this equation
sults from the use of the Kriging model as
surrogate model in connection with opti- situation we provide an optimal division ROLAND B ECKER
mization of "expensive functions". between the number of replicates and the Universitaet Heidelberg
number of evaluations per replicate.
keywords: finite elements, a posteriori
A derivative-free algorithm TH3-306/33 NONLINEAR PROGRAMMING error estimates, optimal control
based on radial basis functions PDE-constrained optimization We consider selve-adaptive construction
(RBF) (3) of appropriate meshes for the optimal
control of scalar elliptic equations. A
organizers : Michael Ulbrich and
RODRIGUE O EUVRAY Stefan Ulbrich
general framework for deriving a poste-
ROSO-IMA-SB-EPFL riori error estimates is presented. With
chair : Stefan Ulbrich the help of such an estimator we con-
coauthor: Michel Bierlaire
keywords: derivative-free algorithm, A generalized discretization struct an adaptive algorithm which com-
trust region method, radial basis concept for control constrained bines optimization and local mesh refine-
functions, DFO
optimal control problems and ment allowing for quantitative error con-
its numerical realization trol and efficiency. We further focus on
Derivative-free optimization involves all M ICHAEL H INZE the additional difficulties introduced by
the methods used to minimize an objec- TU Dresden stabilization of the Galerkin finite ele-
tive function when its derivatives are not ment scheme.
keywords: discretization concept,
available and when the function is expen-
control of PDEs, primal-dual active set
sive. We present here a trust region algo- Very large scale nonlinear pro-
method, control constraints
rithm based on radial basis functions in- gramming by scp methods
In my talk I will discuss a new discretiza-
stead of the classical second order poly-
tion concept for abstract control prob-
nomials. Actually, our surrogate of the
lems with control constraints which ex-
C HRISTIAN Z ILLOBER
objective function is a mixed radial and University of Bayreuth
tends the common discrete approaches.
polynomial model. We present encourag- keywords: nonlinear programming,
Discretization only is applied to the state
ing preliminary results. The overall effi- applications, very large scale problems,
variables, which in turn implicitly yields
ciency of the algorithm is comparable to convex approximations
a discretization of the control variables
DFO. On some instances, the new algo-
by means of the first order optimality We consider the general nonlinear pro-
rithm performs even better.
condition. For discrete controls obtained gramming problem with smooth objec-
in this way an optimal error estimate in tive function and constraints. Sequen-
Gradient estimation using La- terms of the state-discretization parame- tial convex programming (SCP) methods
grange interpolation polynomi- ter is presented. have been developed in the context of
als Applied to control of partial differen- structural optimization problems where
tial equations combined with finite ele- the methods are known to work well
D ICK DEN H ERTOG ment discretization of the state the key since some kind of frequently used con-
Tilburg University features of the new control concept in- straints are approximated very well. The
coauthors: Ruud Brekelmans, Lonneke clude most important properties of these first
Driessen, Herbert Hamers -decoupling of finite element grid order approximations which are applied
keywords: gradient estimation, and discrete active set -numerical im- for the objective function and inequal-
Lagrange interpolation, nonlinear plementation requires only small addi- ity constraints are convexity and sepa-
optimization, noisy function tional overhead compared to commonly rability. The convex subproblems that
142 Parallel Sessions: Thursday 16:30 – 18:00 (TH3)

arise in this framework are solved by a M ARIA H ELENA C AUTIERO of nonmonotone line searches. A spe-
primal-dual interior point method. The H ORTA JARDIM cific algorithm, the Inexact Spectral Pro-
main advantage is that sparsity of the Ja- Universidade Federal do Rio de Janeiro jected Gradient method (ISPG) uses in-
cobian can be exploited very efficiently. coauthors: Nelson Maculan, Claudio exact projections computed by Dykstra’s
The method had has been applied to very Prata Santiago alternating projection method and gen-
large scale nonlinear programming prob- keywords: orthogonal projection, erates interior iterates. This algorithm
lems with up to one million variables computational linear algebra, is implemented and tested and numeri-
arising in topology optimization. For computational complexity cal results are shown. The ISPG method
these applications one could expect good is a generalization of the Spectral Pro-
performance of SCP methods. But most The orthogonal projection of a vector on jected Gradient method (SPG), but can
recently the methods have been applied the intersection of two hyperplanes and a be used when projections are difficult to
to very large scale nonlinear program- box arises as a subproblem of some op- compute. Convergence properties and
ming problems from optimal control for timization methods, for example, the so- numerical results are presented.
which the motivation for the special ap- lution of the dual Lagrangian problem in
proximation does not hold. Nevertheless, integer programming using e-subgradient
SCP methods perform superior to other techniques. In this work, for this pro- TH3-306/35 NONLINEAR PROGRAMMING

methods that have been tested in a bench- jection we present an O n2 logn  time Newton-like algorithms II
mark study by Maurer and Mittelmann. algorithm . We extend a previous re- chair : Mehiddin Al-Baali
The largest problem that has been solved sult of Maculan et al. [1], with simi-
lar approach relating to the complexity A nonmonotone line search
in this context has about 720000 vari- technique and its application
ables and 360000 constraints. of the algorithm . With reference to pro-
jection, the solution is based in finding to unconstrained optimization
Lagrange multipliers associated with the
TH3-306/34 NONLINEAR PROGRAMMING

hyperplanes. Details of the complexity

H ONGCHAO Z HANG
Large scale problems III University of Florida
of the algorithm for this quadratic pro-
chair : Ernesto G. Birgin gramming problem are presented. Com- coauthor: William Hager
putational results are also presented in or- keywords: nonmonotone line search,
Issues on conjugate gradient- der to evaluate the algorithm and its time r-linear convergence, unconstrained
type algorithms within trun- complexity. optimization, L-BFGS method
cated Newton methods, in large
scale unconstrained optimiza- A new nonmonotone line search algo-
[1] MACULAN N.,SANTIAGO
rithm is proposed and analyzed. In
tion C.P., MACAMBIRA,E.M.and our scheme, we require that an aver-
G IOVANNI FASANO JARDIM,M.H., An O(n) algorithm
age of the successive function values de-
University of Rome for projecting a vector on the in-
creases, while the traditional nonmono-
tersection of a hyperplane and a
tone approach of Grippo, Lampariello,
coauthor: Massimo Roma box in Rn- Journal of Optimiza-
keywords: large scale optimization, and Lucidi [A nonmonotone line search
tion and Applications , Vol117, N3
technique for Newton’s method, SIAM
conjugate gradient method, truncated pp553-574 (2003)
Newton methods J. Numer. Anal., 23 (1986), pp. 707–
[2] MINOUX,M., Mathematical Pro- 716] requires that a maximum of re-
In this work we aim at studying some cent function values decreases. We
new features of conjugate gradient type gramming, John Wiley(1986)
prove global convergence for noncon-
methods based on planar schemes. The
latter algorithms are used for solving [3] SANTIAGO,C.P. Projeção de um vex, smooth functions, and R-linear con-
the Newton’s equation within truncated ponto sobre a interseção de um vergence for strongly convex functions.
Newton methods. The interest on pla- hiperplano e o Rn+; complexidade For the L-BFGS method and the un-
nar conjugate gradient algorithms has numérica e implementação. Mas- constrained optimization problems in the
been recently rekindled, since they have ter’s thesis, COPPE-UFRJ (2000) CUTE library, the new nonmonotone line
search algorithm used fewer function and
a twofold advantage. On one hand they
gradient evaluations, on average, than ei-
avoid the well known shortcomings of ther the monotone or the traditional non-
the standard conjugate gradient methods, Inexact spectral projected gra-
monotone scheme.
in dealing with indefinite problems. On dient methods on convex sets
this purpose we investigate the possibility
to overcome these drawbacks and to ex- E RNESTO G. B IRGIN High order Newton-penalty al-
tend the relationship between the Lanc- Computer Science Dept. - IME - gorithms
zos and the planar conjugate gradient University of Sao Paulo
type methods, in the indefinite case. On coauthors: J. M. Martinez, M. Raydan J EAN -P IERRE D USSAULT
the other hand, planar conjugate gradient keywords: projected gradient, convex Université de Sherbrooke
methods provide a set of conjugate di- constrained, nonmonotone line search, keywords: quadratic penalty function,
rections which seem to be promising, in spectral gradient Newton method, predictor-corrector
defining effective preconditioning strate- A new method is introduced for large methods
gies. scale convex constrained optimization. Recent efforts in differentiable non-linear
The general model algorithm involves, programming have been focused on in-
An o n2 log n algorithm for pro-

at each iteration, the approximate min- terior point methods, akin to penalty
jecting a vector on the intersec- imization of a convex quadratic on the and barrier algorithms. In this pa-
tion of two hyperplanes and a feasible set of the original problem and per, we address the classical equal-
box in rn global convergence is obtained by means ity constrained program solved using
 143

the simple quadratic loss penalty func- number of hyperplanes generating piece- minimizer nor a maximizer. For a con-
tion/algorithm. The suggestion to use wise linear function. If the intersection vex function of the above type there ex-
extrapolations to track the differentiable of these sets is empty then they can be ists a "fast track" to a minimizer which
trajectory associated with penalized sub- strictly separated by a function repre- an algorithm can follow by alternating
problems goes back to the classic mono- sented as a max-min of linear functions. "V-steps" (each produced by a bundle
graph of Fiacco & McCormick. This idea The problem of max-min separability is algorithm subprocess) with orthogonal
was further developed by Gould who ob- reduced to a certain nonsmooth optimiza- "U-quasi-Newton-steps" (each based on
tained a two-steps quadratically conver- tion problem. We propose an algorithm a VU-decomposition and a U-gradient
gent algorithm using prediction steps and for solving max-min separability prob- that depends on cutting-plane subgradi-
Newton correction. Dussault interpreted lem and apply it to solve supervised data ents generated by the bundle subroutine).
the prediction step as a combined extrap- classification problems. Results of nu-
olation with respect to the penalty param- merical experiments are presented. TH3-306/37 CONVEX PROGRAMMING

eter and the residual of the first order op- Geometry and duality in
timality conditions. Extrapolation with Large-scale Lagrangian decom- convex optimization
respect to the residual coincides with a position problems solved by cut-
Newton step. organizer/chair : Gabor Pataki
ting plane methods
We explore here higher order ex- The D-induced duality and its
trapolations, thus higher order Newton- M ATTHIAS K NOBLOCH applications
like methods. We first consider high Chemnitz University of Technology
order variants of the Newton-Raphson S HUZHONG Z HANG
keywords: nonsmooth programming, The Chinese University of Hong Kong
method applied to non-linear systems of
convex programming, cutting plane coauthor: Jan Brinkhuis
equations. Next, we obtain improved
methods keywords: dual cone, robust
asymptotic convergence results for the
quadratic loss penalty algorithm by using We consider general convex programs optimization
high order extrapolation steps. with binding constraints suitable for La- In this talk we introduce the notion of D-
grangian decomposition. The proposed induced duality for a convex cone, using
oracle-type method is capable to man- a pre-described conic ordering induced
Quasi-Wolfe conditions for age infinite function values of the dual
Newton-like methods by the convex cone D and a fixed bilin-
function. For linear and convex quadratic ear mapping. This is an extension of the
M EHIDDIN A L -BAALI programming problems a complete real- standard definition of dual cones, in the
Sultan Qaboos University ization of the oracle will be described. sense that the nonnegativity of the inner-
Several generalizations for generic con- product is replaced by the conic ordering,
keywords: nonlinear optimization, vex programs will be given. and the inner-product itself is replaced
Newton-like methods, inexact line
by a general multi-dimensional bilinear
searches
Properties of functions with mapping. The resulting cone is called the
Quasi-Wolfe conditions for practical primal-dual gradient structure D-induced dual cone. Furthermore, this
Newton-like methods of nonlinear opti- notion is extended to the D-induced po-
mization will be introduced. These con- ROBERT M IFFLIN lar set. Basic properties of the D-induced
ditions suggest a modification technique Washington State University duality, including the extended bi-polar
to these methods to rectify certain dif- theorem, are proven. We shall further
coauthor: Claudia Alejandra
ficulties which usually arise when the show that this notion of duality is an es-
Sagastizábal
Wolfe line-search conditions are either sential ingredient in optimization. As ex-
keywords: second order derivatives,
weak or not satisfied at some iterations. amples, we mention its applications in ro-
VU-decomposition, nonsmooth analysis,
It will be shown, in particular, that the bust optimization and multiple-objective
convex minimizaton
modification technique speeds up several programming.
quasi-Newton algorithms and maintains This talk concerns further developments
their useful properties. for nonsmooth functions with primal- On the block-structured dis-
dual gradient (pdg) structure[SIOPT tance to infeasibility
13(2003) 1174-1194]. Such func-
TH3-306/36 NONSMOOTH OPTIMIZATION

tions have an underlying C 2 substruc- JAVIER P ENA

Nonsmooth optimization ture which differs from that of fully Carnegie Mellon University
chair : Robert Mifflin amenable functions, because the class keywords: distance to infeasibility,
Eckart and Young identity, sparse
Max-min separability and its contains maximum eigenvalue functions
and other "infinitely-defined" functions. matrices, conic systems
applications Under certain conditions, including "pri- We show that the block-structured dis-
A DIL BAGIROV mal feasibility", such a function has "U- tance to infeasibility of a conic system
School of Information Technology and subspace" Hessians that give second- of constraints equals the reciprocal of a
Mathematical Sciences, University of order epi-derivatives in U-space direc- suitable block-structured norm of a cer-
Ballarat tions. With the addition of a "dual feasi- tain inverse operator. This gives a natu-
blity" condition at a point the function is ral generalization of the classical Eckart-
keywords: nondifferentiable partly smooth relative to a certain man- Young identity, which states that the (un-
optimization, data analysis, pattern ifold. The pdg structure at a point of a structured) distance to singularity of a
analysis nonconvex partly smooth example func- square matrix equals the reciprocal of the
We consider the problem of discriminat- tion due to Lewis is interesting because norm of its inverse. We also discuss our
ing between two or more finite point sets it has two U-Hessians whose eigenvalues results in the equivalent context of set-
in the n-dimensional space by the finite indicate that the point is locally neither a valued sublinear mappings, where a dual
144 Parallel Sessions: Thursday 16:30 – 18:00 (TH3)

counterpart of the generalized Eckart- restoration, we apply DC programming coauthor: Le Dung Muu
Young identity is revealed. and DC algorithms (DCA), introduced keywords: maximal flow, D.C.
Our results unify and extend the by T. PHAM DINH in 1986 and exten- optimization, global optimization
Eckart and Young identity, Renegar’s sively developed by H. A. LE THI and In this talk, we consider the minimum
characterization of the distance to infea- T. PHAM DINH since 1992, to globally maximal flow problem, i.e., minimiz-
sibility, Cheung and Cucker’s characteri- minimize the Gibbs energy function in a ing the flow value among maximal flow,
zation of the normalized distance to ill- Markov random field model. Numerical which is an N P -hard problem. After for-
posedness, and Rohn’s characterization simulations show the efficiency, reliabil- mulating the problem, we introduce a du-
of the componentwise distance to singu- ity and robustness of DCA with respect alformulation and cast the problem into a
larity. We also present some connections to the standard GNC method. minimization of a concave functionover a
with the mu-number in robust control, convex set. We propose a D.C. optimiza-
and the sign-real spectral radius.
DC programming and DCA ap- tion algorithm for solving the problem
proaches to variational inequal-
Bad semidefinite systems: they ity problems TH3-308/11 /
GENERALIZED
CONVEXITY MONOTONICITY

look all the same Generalized convexity VI

TAO P HAM D INH
G ABOR PATAKI National Institute for Applied organizer : Laura Martein
Operations Research, University of Sciences-Rouen chair : Matteo Rocca
North Carolina at Chapel Hill coauthor: Hoai An Le Thi On a dual method for quasi-
keywords: semidefinite programming, keywords: DC programming/DCA, variational inequalities
second order cone programming, conic variational inequality problem,
duality, exact characterization nonconvex programming, global M ARIELLA ROMANIELLO
optimization Università della Calabria, facoltà di
We give a surprisingly simple, exact, "ex- Economia, Dip. di Organizzazione
cluded minor" type characterization of DC programming and DCA have been Aziendale e Amministrazione Pubblica
semidefinite systems that have a badly introduced by T. Pham Dinh and exten-
coauthor: Jacqueline Morgan
behaved dual for some objective func- sively developed by the authors to solve
keywords: quasi-variational
tion. We also prove that a semidefinite dc programs. DCA are based on lo-
inequalities, dual method
system has a well behaved dual for all ob- cal optimality conditions and dc duality
jectives, iff the constraint matrices can be and their constructions depend on dc de- In “Generalized Quasi-Variational In-
made block-diagonal; the block structure compositions. Our approaches have been equalities and Duality” (J. Morgan -
is given by the maximum rank semidefi- successfully applied to solve real-world M. Romaniello, to appear on JIPAM) a
nite slack. dc programs for which DCA converge duality scheme and Generalized Kuhn-
The characterization is based on new quite often to global solutions. By us- Tucker conditions have been presented.
results in convex analysis about the ing gap functions, variational inequality This scheme allows to solve the primal
closedness of the linear image of a closed problems can be reformulated either as and the dual problems in the spirit of
convex cone. These results lead to exact problems of finding critical points of dc the classical Lagrangian duality for con-
characterizations for other badly behaved functions or as those of globally solv- strained optimization problems and ex-
convex inequality systems as well, most ing dc programs with known optimal val- tends, in non necessarily finite dimen-
notably second order conic systems. ues. In this talk, we will present dc tional spaces, the duality approach ob-
reformulations of variational inequality tained by A. Auslender for generalized
problems and our DC programming and variational inequalities. Stability of the
TH3-306/38 GLOBAL OPTIMIZATION

DCA approaches to solving the result- dual problem with respect to perturba-
Global optimization ing dc programs. Numerical simulations tions on the data will be investigated.
techniques via reformulations, are reported. References Pham Dinh Tao,
smoothing approaches and DC Le Thi Hoai An : Convex analysis ap- Minty variational inequalities
programming proach to dc programming: Theory, Al- and vector optimization
organizer/chair : Tao Pham Dinh gorithms and Applications. Acta Math-
ematica Vietnamica, Vol. 22, Number M ATTEO ROCCA
A difference of convex func- 1 (1997), pp. 289-355 Pham Dinh Tao, University of Insubria, Department of
tions programming approach Le Thi Hoai An: DC Optimization Al- Economics
to image restoration via Markov gorithm for Solving The Trust Region coauthors: Giovanni Paolo Crespi, Ivan
model Problem. SIAM Journal on Optimiza- Ginchev
tion, Vol. 8, Number 2, (1998), pp. keywords: vector variational inequality,
H OAI A N L E T HI vector optimization, pointwise
National Insitute for Applied 476-505 Pham Dinh Tao, Le Thi Hoai
An: Large Scale Molecular Optimization well-posedness
Sciences-Rouen
From Distance Matrices by a DC Opti- Vector extensions of Stampacchia and
coauthors: Pham Dinh Tao, Pham Tien mization Approach. (40 pages), To ap- Minty variational inequalities have been
Son pear in SIAM Journal on Optimization. introduced by F. Giannessi and have re-
keywords: Gibbs energy, Markov vealed useful in vector optimization. In
model, DC programming, DCA/GNC this talk we focus on Minty variational
D.C. and smooth optimization
Image restoration problems can be solved methods for solving minimum inequalities.
using optimization techniques. They lead maximal network flow problem Starting from classical results for scalar
often to the solution of a nonconvex variational inequalities, stating that
and nondifferentiable minimization prob- J IANMING S HI Minty variational inequality is a suffi-
lem. In this paper, to carry out image Muroran Institute of Technology cient optimality condition for a primitive
 145

minimization problem (that is the prob- Monte Carlo scenario tree gen- coauthors: Frode Rømo, Asgeir
lem
 of minimizing a function f such that eration for stochastic program- Tomasgard, Lars Harald Vik
f  F, where F is the function involved ming keywords: stochastic programming,
in the inequality), we remark that for scenario generation, principal
vector optimization this result holds only DAVID M ORTON component analysis
under C-convexity assumption. The University of Texas at Austin
This paper presents an approach of gener-
Therefore, since solutions of a Minty coauthor: Anukal Chiralaksanakul ating scenariotrees for multidimensional
vector variational indeed are solution of keywords: stochastic programming, stochastic processes. Multidimensional
a linearly scalarized problem (under con- Monte Carlo simulation, scenario tree processes appear in real world applica-
vexity assumption), we use the oriented generation tions for example as inflow in water mag-
distance function, introduced, to obtain
a scalar Minty variational inequality re- Solving a multi-stage stochastic program azines, electricity and gas prices, and gas
lated to the vector optimization problem. with a large number of scenarios and a demands. Typically many of the stochas-
We prove that the solutions of the for- moderate-to-large number of stages can tic variables are highly correlated.
be computationally challenging. We de- We present an approach for generat-
mer variational inequality characterize ing multidimensional scenrio trees based
weakly efficient and efficient solution velop Monte Carlo-based methods that
(without need for convexity) and, under exploit special structures to generate fea- on the following following steps:
convexity assumption, are equivalent to sible policies. To establish the quality
of a policy, we use a Monte Carlo-based Forcasting the uncertainty for
the solution of the classical Minty vector these multidimensional variables
variational inequality. Furthermore, we lower bound (for minimization problems)
deduce also pointwise well-posedness in constructing a confidence interval on Principal component analysis and
of vector optimization problem which the policy’s optimality gap. We allocate dimension reduction to identify
admits solutions to this inequality. sample sizes in generating nonuniform uncertain factors
scenario trees in order tighten the lower
bound and hence the resulting confidence Generating samples for uncertainty
Minty variational inequalities interval. Computational results are pre- factors
and well-posedness in scalar op- sented for a stochastic lot-sizing problem
timization and for pricing an American-style option. Combining the forecasts and the
samples in order to build a scenario
G IOVANNI PAOLO C RESPI tree.
Université de la Vallée d’Aoste - Faculty Scenario generation in stochas-
of Economics tic programming - what can we First numerical results will be presented.
coauthors: Matteo Rocca, Ivan Ginchev reasonably ask for?
keywords: Minty variational inequality, TH3-308/13 - SEMI INFINITE AND INFINITE

optimization, generalized convexity, S TEIN W. WALLACE DIMENSIONAL PROGRAMMING

well-posedness Molde University College Complexity in semi-infinite

coauthor: Michal Kaut
programming
Minty Variational Inequalities are widely
studied in relation to minimization prob- keywords: stochastic programming, chair : Klaus Meer
lems. When the function involved in the scenarios Distance to ill-posedness for
inequality has a primitive f this is natu- As stochastic programming matures, the linear semi-infinite inequality
rally the objective function of the prim- focus is shifting in the direction of appli- systems: applications to lips-
itive minimization problem over the fea- cations. It then becomes clear that op- chitzian analysis
sible region K and Minty inequality is a timal solutions can be very sensitive to
sufficient optimality condition. We deal how randomness is described, in partic- M ARCO A. L ÓPEZ -C ERDÁ
with a more general case, namely when ular, how the random variables are dis- Aicante University
f admits lower Dini derivative. In this cretized in the form of scenario trees. On coauthors: M. J. Cánovas, J. Parra,
setting we define a Generalized Minty one hand, there is a need to have a tree F.Javier Toledo-Melero
Variational Inequality as refered to the that properly describes the important as- keywords: semi infinite programming,
minimization of (non differentiable) f . pects of the underlying random variables. stability, metric regularity, distance to
The main result of the talk is that, if the On the other hand, we must realize that ill-posedness
set K is star-shaped at x and f is ra- the actual distributions are rarely known
dially lower semicontinuous in K along We consider the parameter space of all
with any large degree of accuracy. This the linear inequality systems, in the
rays starting at x , then x is a solu- talk will discuss, from a very practical
tion of the genralized GMV I f K  if perspective, how we can determine if we
n-dimensional Euclidean space, with a
and only if f is increasing along rays fixed and arbitrary (possibly infinite) in-
have a good scenario generation proce- dex set, and endowed with the uniform
(starting at x ). The relevance of In- dure at the same time as we try to take
creasing Along Rays function (which can convergence topology. Since the parame-
into account that there is a limit to how ter space is not a normed space, a special
be viewed as a generalization of convex much we can know about future stochas-
functions) is stressed porving some prop- subset arises, formed by those systems
tic events. Examples will be given. whose distance to ill-posedness is infi-
erties they have with regards to optimiza-
tion and well-posedness. nite. This fact gives rise to a generalized
Generating scenario trees for concept of ill-posedness, which has been
multidimensional stochastic characterized in terms of the so-called as-
TH3-308/12 STOCHASTIC PROGRAMMING

processes sociated hypographical set, allowing for

Scenario tree generation an explicit formula of the ’distance to
organizer : COSP (David Morton) M ATTHIAS P ETER N OWAK generalized ill-posedness’. On the other
chair : David Morton SINTEF hand, the consistency value of a system,
146 Parallel Sessions: Thursday 16:30 – 18:00 (TH3)

also introduced in the talk, provides an K LAUS M EER algorithms with complexity of order n3
alternative way to determine its distance University of Southern Denmark can be derived, where n is the number
to ill-posedness (in the original sense). keywords: computational complexity, of states. In the second part we con-
This talk is also devoted to some applica- interval arithmetic, semi infinite sider other problems, like mean-variance
tions of the ’distance to ill-posedness’, in optimization trade-offs and multi-armed bandit prob-
relation to some Lipschitzian properties lems. The mean-variance problem can be
(pseudolipschitz, calmness) of the feasi- We study some problems in interval formulated as a nonlinear programming
ble set mapping, the corresponding regu- arithmetic treated by Kreinovich et al. problem. Fortunately, using the structure
larity concepts of the inverse multifunc- First, we consider the best linear approx- of the problem, this nonlinear program
tion, and some associated Hoffman type imation of a quadratic interval function can efficiently be solved by parametric
inequalities. From an algorithmic point (a semi-infinite optimization problem). linear programming. The multi-armed
of view, the distance to ill-posedness al- Whereas this problem (as decision prob- bandit problem is a high-dimensional
low us to analyze, in our semi-infinite lem) is known to be NP-hard in the Tur- problem, say of dimension m with in
context, the complexity of the ellipsoid ing model, we analyze its complexity in each dimension n states, which generate
algorithm. the real number model and the analogu- an LP-formulation with mn variables and
ous class NP  Our results substantiate constraints. It is known that the multi-
that most likely it does not any longer
An outer approximation method capture the difficulty of NP in such a
armed bandit problem can be solved as
for general semi-infinite pro- real number setting. More precisely, we
a sequence of m one-dimensional prob-
gramming without discretiza- give upper complexity bounds for the ap-
lems. We show that each of these prob-
tion proximation problem for interval func-
lems can be solved with complexity of
order n3 by parametric linear program-
O LIVER S TEIN tions by locating it in Σ2 (a real analogue ming.
Aachen University of Σ2 
keywords: interior point method, This result allows several conclu-
Stackelberg game, optimality condition, sions: Scheduling jobs with uncertain
convexity processing times on parallel ma-
Today applications of semi-infinite pro- - the problem is not (any more) chines
gramming range from approximation, ro- NP -hard under so called weak
bust optimization and minimax problems polynomial time reductions and
to design centering, disjunctive optimiza- likely not to be NP -hard under M ARC U ETZ
tion, data envelopment analysis and de- (full) polynomial time reductions; Uni Maastricht / FdEWB/ Quantitative
Economics
fect minimization for operator equations. - for fixed dimension the problem is coauthors: Rolf H. Möhring, Andreas
In many of these examples the in- polynomial time solvable; this ex- S. Schulz, Martin Skutella
dex set of the inequality constraints in- tends results by Koshelev et al. and keywords: stochastic scheduling,
evitably depends on the decision vari- answers a question left open by dynamic scheduling policy, linear
able, yielding a general semi-infinite Kreinovich et al.
optimization problem, as opposed to programming relaxations, approximation
standard semi-infinite problems, where We also study several versions of in-
this index set is assumed to be fixed. We consider scheduling problems where
terval linear systems and show similar the processing times of jobs are uncer-
For problems of the latter type opti- results as for the approximation prob-
mality conditions and solution methods tain, and scheduling decisions have to
lem. Our methods combine structural be made dynamically over time. In
are well-known, including discretization complexity theory with issues from semi-
methods for the index set which naturally contrast to the pessimistic assumptions
infinite optimization theory. in on-line optimization, we assume that
lead to outer approximations of the feasi-
ble set. On the other hand, general semi- job characteristics, particularly their ex-
infinite programs have turned out to be TH3-308/T1 DYNAMIC PROGRAMMING AND
OPTIMAL CONTROL pected processing times as well as an up-
surprisingly difficult to treat theoretically Linear programming per bound on their variability are known
as well as computationally. approaches to dynamic and to the scheduler. In this scenario, we talk
Recent results admit to solve a large stochastic optimization of a non-anticipatory scheduling policy
subclass of these problems by a numer- if, at any point in time, only the avail-
organizer/chair : Jose Nino-Mora able information about the past is an-
ical method which bases on a reformu-
lation of the semi-infinite problem as a (Parametric) linear program- ticipated. The goal is to find a non-
Stackelberg game. Although then an in- ming for Markovian control anticipatory scheduling policy that min-
terior point approach is used in the lower problems imizes expected performance. We de-
level problem, it turns out that one ob- rive the first approximation results for
tains an outer approximation of the feasi-
L ODEWIJK K ALLENBERG this setting. Our results are based on two
Mathematical Institute, University of main ingredients: On the one hand, this
ble set for the semi-infinite problem.
Leiden is a linear programming relaxation that is
We give convergence properties for
the method and illustrate it with a number keywords: Markov decision processes, formulated in terms of the expected pro-
of numerical examples. The presented al- parametric linear programming cessing times of jobs. On the other hand,
gorithm is easy to implement, and our ex- we use certain list scheduling algorithms
It is well known that discrete time, fi-
amples indicate that it works well even that have been used before also in deter-
nite Markov decision problems with dis-
for high dimensional index sets. ministic settings. Our results lead to con-
counted or average rewards can be solved
stant performance guarantees whenever
by linear programming. This talk con-
the coefficients of variation of the jobs’
On the complexity of some prob- sists of two parts. In the first part we
processing times are bounded by a con-
lems in interval arithmetic consider some special models for which
stant.
 147

We also discuss extensions to Optimization of pilot power in

Dynamic allocation indices for semidefiniteness constraints. 3G mobile networks
restless projects and queueing
admission control: a polyhedral Nonsmooth matrix valued func- P ETER VÄRBRAND
approach tions defined by singular values Linköping University
D EFENG S UN coauthor: Di Yuan
J OSE N INO -M ORA keywords: mobile communications,
Depto. de Estadistica, Universidad National University of Singapore
resource management, relaxation,
Carlos III de Madrid coauthor: Jie Sun
heuristics
keywords: stochastic scheduling, keywords: semismoothness, FB
Markov decision processes, index rules, function, SVD, complementarity In 3G mobile networks, a cell uses a pi-
polyhedral methods A class of matrix valued functions de- lot signal to provide channel estimation
This work develops a polyhedral ap- fined by singular values of nonsymmetric to the mobile terminals. The problem of
proach to the design, analysis and matrices are shown to have many proper- choosing the power levels of the pilot sig-
computation of dynamic allocation in- ties analogous to matrix valued functions nals involves the trade-off between ser-
dices for scheduling binary-action (en- defined by eigenvalues of symmetric ma- vice coverage and power consumption.
gage/rest) Markovian stochastic projects trices. In particular, the (smoothed) In this presentation, we consider the op-
that can change state when rested (rest- matrix valued Fischer-Burmeister func- timization problem of minimizing the to-
less projects/bandits), based on the tion is proved to be strongly semismooth tal pilot power subject to coverage con-
framework of partial conservation laws everywhere. This result is also used straints. We present two linear integer
(PCLs). Satisfaction of PCLs by per- to show the strong semismoothness of formulations for the problem, and de-
formance measures implies that the the (smoothed) vector valued Fischer- scribe a Lagrangean-based solution ap-
achievable performance region admits a Burmeister function associated with the proach. We present computational results
tight polyhedral relaxation having strong second order cone. The strong semis- using real-life data. The results show
structural and algorithmic properties (F- moothness of singular values of a non- that the proposed method finds optimal or
extended polymatroid). This allows symmetric matrix is discussed and used near-optimal solutions.
to identify a class of projects (PCL- to analyze the quadratic convergence of
indexable) that are optimally controlled Newton’s method for solving the inverse
by an index policy having a postulated singular value problem. Scheduling of spatial time di-
structure, where the indices are effi-
vision multiple access in multi-
ciently computed by an adaptive-greedy Finite Newton’s methods for LP hop radio networks
algorithm. PCL-indexability is shown to and LCP
represent a form of the classic economic D I Y UAN
V ITALIJ Z HADAN Linköping University
law of diminishing marginal returns, with Dorodnicyn Computing Centre of
the indices representing optimal marginal Russian Academy of Sciences coauthor: Peter Värbrand
cost/reward rates. The approach is de- keywords: wireless communications,
ployed to give a novel analysis of a ver- keywords: linear programming scheduling, column generation,
satile queueing admission control model. problem, linear complementarity heuristics
problem, Newton method, finite
convergence We consider resource optimization in
We consider the primal-dual technique multi-hop radio networks that use Spa-
TH3-308/T2 COMPLEMENTARITY AND VARIATIONAL
INEQUALITIES
tial Time Division Multiple Access
for linear programming problem in which
Linear complementarity the non-perturbed system of optimal- (STDMA). In STDMA, transmission
problems ity conditions is solved by Newton’s must be scheduled to avoid access colli-
chair : Vitalij Zhadan method. We suppose that starting points sions and interference. We consider two
and all cosequent points are feasible. The scheduling problems, where the objec-
On polyhedral games, linear step length is chosen from the steepest tives are minimizing the schedule length
complementarity problems, and descent approach basing on minimization and maximizing the network through-
interior point methods of the dual gap. We have not introduced put, respectively. We provide linear in-
S IMON S CHURR the safity factor and allow trajectories in teger formulations for the two problems.
University of Maryland at College Park primal and dual spaces to move along the We then present a column generation
method for solving the LP-relaxations,
coauthors: Dianne O’Leary, Andre Tits boundaries of feasible sets. In the case
where the Newton system is underdeter- and heuristic algorithms for generating
keywords: game theory, linear
mined the auxiliary linear complementar- feasible schedules. Computational re-
programming, linear complementarity
ity problem (LCP) is solved for obtain- sults for real instances are presented.
problem, interior point method
ing unique search directions. Conditions
It is known that a two-person zero-sum for local and non-local finite convergence
game with linear constraints, can be for- of the method are presented. The gener- Planning of a wireless LAN
mulated as a linear programming prob- alization of the method for solving LCP
lem (LP), and therefore a linear com- with a positive definite matrix is given. F EDERICO M ALUCELLI
plementarity problem (LCP). We use the DEI - Politecnico di Milano
equivalence between different forms of
LCP, to induce special structure in the
TH3-308/T3 TELECOMMUNICATION NETWORK
DESIGN coauthors: Antonio Capone, Matteo
Optimization problems in Cesana
aforementioned LP. Such structure al- keywords: set covering, quadratic
lows for a more efficient solution via in- wireless telecommunications programming, combinatorial
terior point methods. organizer/chair : Federico Malucelli optimization, telecommunication
148 Parallel Sessions: Thursday 16:30 – 18:00 (TH3)

Wireless LANs are assuming a growing Lartge-update interior-point methods present solution methods based on op-
importance either in firm organization, or (IPMs) working in a large neighborhood timization. The company has five pulp
in public areas like airports and railway of the central path are among the most mills in Scandinavia. On each pulp mill
stations. Their planning (indoor or out- efficient IPMs for linear optimization they produce several different pulp types
door) is usually made by means of em- (LO) problems. The best known worst- that they deliver to customers mostly lo-
pirical methods and experience. How- case complexity of this type of methods cated in Europe. The transport is done
ever the diffusion of this kind of net-

based on the standard Newton search di- mainly by cargo vessels that they hire on
rection is O n log  x ε s  . Recently, the
0 T 0
works foreseen in the next future and long term contracts, but also by railway,
the need of efficiency both in terms of author, with his co-coauthors, proposed trucks and spot vessels. It is a ship rout-
cost and of service quality suggests the another kind of large-update IPMs on the ing and scheduling problem where the
use of quantitative methods. Consider- notion of self-regular functions with an constraints are to satisfy demand given
nlog n log  x ε s
ing the rather simple transmission proto-  0 T 0 production, storage and transport capac-
col, the maximization of the network ca- improved O

itera- ities, and the objective to minimize the
pacity gives rise to two contrasting ob- tion bound. In this paper, we suggest a overall cost. We present optimization
jectives: on one hand we aim at cover- hybrid method, which employs the stan- models and compare the results from dif-
ing the service area with as many access dard Newton search direction or the self- ferent heuristic solution approaches.
points (antennas) as possible to increase regularity induced search direction alter-
the parallelism in transmissions, on the natively, depending on the position of the
current iterate. We show that the hybrid
Inventory placement problem
other hand we would like to minimize for a serial supply chain with the
n logn log  x ε s
 0 T 0
the overlapping between adjacent anten- IPMs enjoy an O satisficing objective
nas to minimize the interference. We pro-

pose a mathematical model based on a iteration bound, matched the best know
complexity result for IPMs. C HIA -S HIN C HUNG
particular set covering formulation with Cleveland State University
a quadratic objective function to account coauthors: James Flynn, Piotr Stalinkki
for the interference. We will propose ad Introducing interior-point keywords: inventory, supply chain,
hoc heuristics and bounds. Some prelim- methods into the first operations single period, budget
inary computational results will be also research course
presented on a set of realistic data. Along a long supply chain, it is certainly
G ORAN L ESAJA unnecessary and wasteful to store inven-
Georgia Southern University tory in every stage of the chain. In stead,
TH3-341/21 INTERIOR POINT ALGORITHMS

keywords: interior point method, a more efficient policy is to place in-

Advances in interior point operations research ventory in only a few strategic locations.
methods This paper addresses this problem in a se-
organizer/chair : Tamas Terlaky In recent years the introduction and de-
velopment of Interior-Point Methods has rial supply chain facing a stochastic de-
Exploiting sparsity in sum of had a profound impact on optimization mand for a single planning period. All
squares of polynomials theory and practice, and therefore, has customer demand is served from stage
impacted the field of Operations Re- 1, which contains the product in its fi-
M ASAKAZU KOJIMA search as a whole. Thus, there has nal form. If the demand exceeds the sup-
Tokyo Institute of Technology been an increasing need to introduce ply at stage 1, then stage 1 is resupplied
coauthors: Sunyoung Kim, Hayato theory and methods of this new area from stocks held at the upstream stages
Waki in optimization into introductory Oper- 2 through N, where the product may be
keywords: sum of squares, polynomial ations Research and/or Linear Program- stored in finished form or as raw materi-
optimization, sparsity, semidefinite ming courses. The objective of this pa- als or sub-assemblies. All stocking deci-
program per is to discuss the ways of simplifying sions are made before the demand occurs.
In recent developments of sum of squares the introduction of Interior-Point Meth- All unsatisfied demand is lost. The objec-
optimization and SDP (semidefinite pro- ods for students who have various back- tive is to select the stock quantities that
gramming) relaxation of polynomial op- grounds or who are not necessarily math should be placed at different stages so as
timization problems, representing a given majors. to maximize the probability of achieving
nonnegative multivariate polynomial in a budgeted profit level B. For the case of
terms of sum of squares of polynomial TH3-341/22 LOGISTICS AND TRANSPORTATION
positive lost sales costs, we prove that it
has become an important subject. We is optimal to consider at most three stock-
discuss effective methods to get a simpler Supply chain optimization ing locations. We also characterize its
representation of a “sparse” polynomial chair : Huifu Xu properties, and provide an algorithm for
in sum of squares of sparse polynomials Ship routing and scheduling in its computation.
by eliminating redundancy. the pulp mill industry
Optimal supply functions in
Hybrid primal-dual interior- DAVID B REDSTRÖM electricity markets with piece-
point methods based on self- Division of Optimization / Department wise continuously differentiable
regular functions of Mathematics / Linköping University profit functions
J IMING P ENG coauthor: Mikael Rönnqvist H UIFU X U
keywords: ship routing, ship
Department of Computing and Software, University of Southampton
McMaster University scheduling, real world application,
supply chain optimization coauthor: Edward J. Anderson
keywords: interior point method, large keywords: optimal supply function,
update methods, large neighborhood, We consider the delivery problem for electricity market, one way contracts,
polynomial complexity a large pulp producer in Sweden and piecewise marginal costs
 149

Optimal supply functions for generators keywords: scheduling problems, known results. We give both LP round-
who sell electricity into wholesale spot computational experience, ing and primal-dual algorithms that build
markets have been investigated over the approximation algorithms upon techniques recently applied to fa-
past few years. In previous work a gener- In this paper we consider a computa- cility location problems. These methods
ator’s profit function has been assumed to tional study of approximation algorithms have natural extensions to more general
be continuously differentiable. However for scheduling problems in unrelated ma- variants of the JRP and other inventory
in some interesting instances, a genera- chines minimizing the average weighted models.
tor’s profit function may not be differen- completion time. We implemented two
tiable. In this paper we consider the opti- known approximation algorithms for the TH3-321/053 PRODUCTION AND SCHEDULING

mal supply function when the generator’s problems R   ∑ w j C j and R  r j  ∑ w jC j . Industrial scheduling
profit function is merely piecewise con- The first is a 2-approximation algorithm problems
tinuously differentiable. This allows us
to deal with one-way hedging contracts
presented by Skutella using a semidef-  chair : Johannes Hatzl
inite program and the second is a 2 
that a generator may sign before bidding ε  -approximation algorithm presented by Optimal machine scheduling in
into the spot market, as well as a situation Schulz and Skutella using linear pro- a shipyard
in which a generator owns several gener- gramming. We also present a modi-
ation units with different marginal costs. J ENS B UCHHOLZ
fied algorithm based on the algorithm of Universität Duisburg-Essen
We derive globally optimal supply func- Schulz and Skutella.
tions under these circumastances. We generated computational in- coauthor: Ruediger Schultz
stances using 2, 5, 7 and 10 machines. keywords: machine scheduling,
TH3-341/23 APPROXIMATION ALGORITHMS
For all instances, the practical perfor- machine groups, shift schedules
Approximation algorithms for mance of the algorithms was very sat- The talk deals with real-life machine
scheduling problems isfactory. When there are no presence scheduling in the manufacturing unit of
chair : David Shmoys of release dates, the algorithms perform a major german shipyard. Standard prob-
very well. In many cases, the solutions lems are often strongly influenced by ad-
Approximation schemes for the have value very close to the optimum. ditional restrictions which arise from the
open shop scheduling prob- For problems with release dates, solu- application. In this case the standard job
lem with non-availability con- tions with value under 10 percent from shop scheduling problem is modified due
straints the optimum was generated. The mod- to the fact that we have a machinery con-
ified algorithm generates solutions with sisting of several machine groups with
V ITALY S TRUSEVICH better quality in almost all instances con- identical machines. Furthermore the so-
University of Greenwich sidered. lution has to fulfil a given shift schedule.
coauthors: Mikhail Kubzin, Joachim The shift schedule can differ from ma-
Briet, Guenter Schmidt Approximation algorithms for chine group to machine group. The op-
keywords: scheduling, open shop, the joint replenishment problem timization goal is non-standard, too. We
approximation are heading for a schedule including as
We consider the two-machine open shop DAVID S HMOYS many jobs completed in due time as pos-
scheduling problem to minimize the Cornell University sible. We present a linear mixed integer
makespan in which the machines are not coauthors: Retsef Levi, Robin Roundy model for the described problem. More-
continuously available during the plan- keywords: inventory/production, over we report first computational expe-
ning period. The location of each non- analysis of algorithms, combinatorial rience with real data and compare our re-
availability interval, also called a hole, optimization sults with schedules obtained by an in-
is known in advance. We study the re- house heuristic of the shipyard.
We consider the joint replenishment
sumable scenario which allows the op- problem (JRP) in discrete time, with
eration affected by a hole be resumed a finite horizon and deterministic non- Industrialising the application
from the point of interruption as soon stationary demand. For each of N items of mathematical models and
as the machine becomes available again. and T time periods, we have a given de- methods
The problem is NP-hard even for a single mand that must be satisfied on time. In
hole, and is not approximable within a J ESPER H ANSEN
each time period we can order any sub- Carmen Consulting AB
constant ratio in polynomial time if there set of the items, paying a joint fixed
is at least one hole on one machine and cost in addition to an individual fixed keywords: software development,
at least two holes on the other machine. cost for each item ordered. We can also methodology, packing, crane scheduling
In this talk we present polynomial-time hold inventory while incurring an item- The purpose of the CIAMM project has
approximation schemes for the problem dependent linear holding cost. The goal been to industrialise the application of
with one hole on each machine and with is to find a policy that satisfies all de- mathematical models and methods for
any fixed number of holes on one of the mands on time and minimizes the overall solving these models within the manufac-
machines. fixed and holding cost. turing industry. Two problem cases has
This classical deterministic inventory been considered in the project posed by
Computational experience on model is known to be NP-hard; many the industrial partners and a methodology
approximation algorithms for heuristics for finding optimal and near- for developing combinatorial optimiza-
scheduling unrelated machines optimal solutions have been developed. tion software has been proposed based on
We will show how LP-based techniques the experiences gained from the cases.
F LAVIO K EIDI M IYAZAWA yield efficient approximation algorithms The first case was an investigation of
State University of Campinas with constant performance guarantees possible improvements for handling the
coauthor: Eduardo C. Xavier significantly improving on previously storage of steel plates at Odense Steel
150 Parallel Sessions: Thursday 16:30 – 18:00 (TH3)

Shipyard. The second case was posed M ATTHIAS E HRGOTT recently developed algorithms for updat-
by Bang & Olufsen concerning packing The University of Auckland ing a sparse Cholesky factorization after
of 3-dimensional packets in boxes and keywords: multiple criteria, a small rank change. We compare so-
boxes on pallets. The talk will introduce combinatorial optimization, scalarization lution times for our implementation to
the problem cases and discuss the experi- those of simplex and barrier algorithms,
ences gained in the project leading to the Multiobjective programming problems as implemented in CPLEX, using the
proposed methodology. are generally solved using a scalarization. Netlib LP Test Problems.
Most scalarization methods use an ag-
gregation of the objectives or/and trans-
Makespan minimization for form some of the objectives into con- An enhanced piecewise linear
chemical batch processes straints. Both techniques have drawbacks dual phase-1 algorithm for the
when applied to combinatorial problems: simplex method
J OHANNES H ATZL not all Pareto optimal solutions can be
TU Graz Institut für Mathematik B I STVAN M AROS
found or the scalarized problems are Imperial College
coauthor: Rainer Burkard much harder to solve than the original
keywords: heuristics, applications (single objective) problems. We discuss keywords: linear programming
algorithms, dual simplex method,
In this talk we discuss different opti- these issues and point out how the diffi- phase-1
mization methods for batch processing culties can be overcome at least to some
problems occurring in the chemical in- extent. A dual phase-1 algorithm for the sim-
dustry, where the objective is to mini- plex method that handles all types of vari-
mize the make-span. Up to now, most A new density result in multicri- ables is presented. In each iteration it
papers formulate this kind of problem teria optimization maximizes a piecewise linear function of
as a MILP model based on a discrete dual infeasibilities in order to make the
time representation. However, due to the C HRISTIANE TAMMER largest possible step towards dual feasi-
large number of binary variables various University Halle-Wittenberg, Dept. bility with a selected outgoing variable.
LP-based heuristics do not obtain good Mathematics and Computer Science The algorithm can be viewed as a gen-
results within reasonable computational keywords: multiple criteria eralization of traditional phase-1 proce-
time. We propose a new greedy type al- programming, density theorem, dures. It is based on the multiple use of
gorithm which produces solutions of a variational principle, properly efficient the expensively computed pivot row. By
good quality within very short computa- point a relatively small amount of extra work
tion times. Finally, we suggest ideas how per iteration, the progress it can make is
Arrow, Barankin and Blackwell (1953) equivalent to many iterations of the tra-
the fastness of the greedy approach can
presented a deep density assertion in ditional method. While this is its most
be exploited to devise iterated greedy al-
multicriteria optimization concerning the important feature, it possesses some ad-
gorithms and improved heuristics.
approximation of the efficient points of a ditional favorable properties, namely, it
compact convex subset of a finite dimen- can be efficient in coping with degener-
TH3-321/033 MULTICRITERIA OPTIMIZATION
sional space by points that are maximiz- acy and numerical difficulties. Both the-
Scalarized functions and ers of some strictly positive functional oretical and computational issues are ad-
functionals in multicriteria on this set. We give a new version of dressed. Some computational experience
optimization the Arrow-Barankin-Blackwell theorem is also reported which shows that the po-
chair : Christiane Tammer in normed vector spaces. The novelty of tentials of the method can materialize on
our result is represented by the fact that real world problems.
A method for calculating trade- we do not assume compactness of the set;
offs in multicriteria optimiza- in fact it can be an unbounded asymptot-
tion ically compact set. Our result subsumes
Average-case-behaviour of three
convex-hull-algorithms under
several generalizations of this important rotation-symmetry
P ETRA W EIDNER theorem.
FH Hildesheim/Holzminden/Göttingen
K ARL H EINZ B ORGWARDT
keywords: multiple criteria TH3-321/133 LINEAR PROGRAMMING
University of Augsburg
In the paper, we present a scalarization Numerical approaches II keywords: convex hull, average case
for multicriteria optimization problems chair : Karl Heinz Borgwardt complexity, linear programming,
with the following property: The parame- polyhedra
ters of the scalarization allow one to cal- A sparse implementation of the We analyze three different algorithms
culate the tradeoffs in the optimal solu- LP dual active set algorithm (Beneath-Beyond, Gift-Wrapping and
tion of the scalar problem under rather Fukuda-Avis), which can all be applied
mild assumptions. Here, tradeoffs are un- W ILLIAM H AGER to answer the following question: "Given
derstood to be the efficient points of the University of Florida m points a1    am  Rn , which are
contingent cone. coauthor: Timothy A. Davis the vertices of the polyhedron X : 

For this scalarization, we give a ge- keywords: dual active set algorithm, x  Rn  a1 T x  1    am T x  1
? We
ometrical interpretation and discuss the linear programming, numerical compare their expected number of cal-
control of parameters taking into consid- experimentation culations under the Rotation-Symmetry-
eration the reachability of each efficient We present a factorization-based imple- Model: a1    am are distibuted indepen-
point and numerical difficulties. mentation of the LP Dual Active Set Al- dently, identically and symmetrically un-
gorithm. This implementation exploits a der rotations. The incremental Beneath-
Scalarizations in multiobjective proximal approximation to the dual func- Beyond-Algorithm determines sequen-
combinatorial optimization tion, a multilevel solution strategy, and tially m-n+1 auxiliary polyhedra Xk 
 151


x  Rn  a1 T x  1    a k T x  1
 k  cent vertex until every vertex  is discov-

lection of Simplex-Paths starting from
n     m,
which
 causes an expected effort ered.  Its expected effort is C n  m2 n  any vertex and running to the optimal
of C n  m2 n2 n  ln m    . The sequen- mn2 ln m   . The Fukuda-Avis-Algorithm vertex forms a sspanning tree). This
tial Gift-Wrapping-Algorithm moves on avoids storing the data of the traversed 

algorithm causes an expected effort of
the surface ov X from vertex to adja- path (by exploiting the fact that the col- C n  m2 n  .
152 Parallel Sessions: Friday 9:00 – 10:30 (FR1)

FR1-302/41 COMBINATORIAL OPTIMIZATION M ICHAL P ENN L EONIDAS P ITSOULIS

Technion London School of Economics
Packing, partitioning and
trees coauthor: Stas Rozenfeld keywords: multidimensional
keywords: Steiner, group, assignment, matroid intersection, greedy
chair : Michal Penn approximation algorithms The k-dimensional assignment problem
Phi-functions for complex 2d- In this talk we study the group Steiner is a natural extension of the well
objects network problem, which is the following. known 2-dimensional assignment prob-
G UNTRAM S CHEITHAUER Given a graph G=(V,E), a partition of its lem, where a matching of optimum
Dresden University of Technology vertices into K groups and connectivity weight is sought in a k-partite hyper-
requirements between the groups the aim graph. In this talk the problem is formu-
keywords: cutting and packing is to find simultaneously a set of repre- lated as an independence system arising
Within two-dimensional cutting and sentatives, one for each group, and a min- from the intersection of k partition ma-
packing problems with irregular shaped imum cost connected subgraph that sat- troids. Worst case bounds of greedy type
objects, the concept of Phi-functions isfies the connectivity requirements be- algorithms for finding a minimum weight
has been proven to be very helpful for tween the groups (representatives). This basis and various properties of the set of
several solution approaches. Based on problem is a generalization of the Steiner bases of the resulting independence sys-
Phi-functions for pairs of so-called pri- network problem (termed also the sur- tem are investigated.
mary objects, now Phi-functions are con- vivable network design problem) and the
structed for pairs of objects which can group Steiner tree problem (called also Scheduling without waiting pe-
be represented as a finite combination the generalized minimum spanning tree riods
(union, intersection, complement) of pri- problem), two known NP-complete prob-
mary objects which allows the handling lems. B JARNE TOFT
of many arbitrary shaped objects by ap- We present an approximation algo- University of Southern Denmark
propriate approximations of sufficient rithm for the group Steiner network keywords: graph-coloring, interval
accuracy. problem with an approximation ratio of coloring
min2(1+2x),2I where I is the cardinality The problem of scheduling parent con-
of the largest group and x is a parameter sultations at a school without waiting pe-
The densest packings of 28, 29 that depends on the cost function.
and 30 congruent circles in the riods was first presented to me by Jesper
unit square - a reliable optimal- Bang-Jensen around 1990. It had been
ity proof FR1-302/42 COMBINATORIAL OPTIMIZATION studied a couple of years earlier by As-
ratyan, Karmalyan and Sevastyanov in
Combinatorial optimization X the Sovjet Union. We stated the problem
M IHALY C SABA M ARKOT chair : Bjarne Toft
University of Szeged as 12.23 in the book [Jensen and Toft,
coauthor: Tibor Csendes K-edge-connected graphs and Graph Coloring Problems, Wiley 1995].
positive semidefiniteness The purpose of the present talk is to give
keywords: interval analysis, an overview and update of the problem.
branch-and-bound, circle packing,
computer-assisted proofs A LEXANDRE L AUGIER
France Telecom Recherche & FR1-302/44 COMBINATORIAL OPTIMIZATION

We present a new verified optimiza- Developpement Maximum feasible subsystem

tion method for the problem of finding keywords: edge-connectivity, positive problem
the densest packing of non-overlapping semi-definite matrix, cut metrics, cones
equal circles within a square. In order organizer/chair : Edoardo Amaldi
of semi-metrics
to provide reliable numerical results, the The maximum feasible subsys-
developed algorithm is based on interval It is well known that the problem of find- tem problem and related poly-
analysis. As a demonstration of the capa- ing a minimum weight k-edge-connected hedra
bilities of the new algorithm the problems subgraph of a given graph G is NP-hard.
of packing 28, 29, and 30 circles were In this work we modelise the problem of M ARC P FETSCH
solved within very tight tolerance values. finding a k-edge-connected subgraph as a Konrad-Zuse-Zentrum
For all the above problem instances we problem over the cone of the cut metrics, coauthors: Edoardo Amaldi, Jr. Leslie
stated that - apart from symmetric con- then we embed this formulation into the E. Trotter
figurations - all the global optimizers are cone of negative type metrics. This em- keywords: infeasible linear systems,
located within a very small box. More- bedding defines a positive semi-definite feasible subsystem polytope, alternative
over, we have proved that the exact global relaxation of the original problem, then polyhedron
optimum differs from the currently best we exploit this relaxation in order to de-
known packing value by at most 7e-15, sign a heuristic for the minimum weight Given an infeasible linear inequality sys-
2e-14, and 3e-15, respectively. Our veri- k-edge-connected subgraph problem. At tem, the maximum feasible subsystem
fied procedure decreased the uncertainty last we extend the formulation to the mul- problem (MaxFS) is to find a feasi-
in the location of the optimal packings by ticommodity flow problem in the case ble subsystem of maximum cardinality.
more than 700 orders of magnitude in all where the demand graph satisfies some MaxFS is NP-hard, has many interesting
cases. given properties. applications, and generalizes well-known
combinatorial optimization problems, as
we will briefly present.
An approximation algorithm for Bounds for greedy type algo- For the investigation of the 0-1 poly-
the group Steiner network prob- rithms for the k-dimensional as- tope corresponding to MaxFS, i.e., the
lem signment problem convex hull of the incidence vectors of
 153

feasible subsystems, results about inde- a factor 2 but does not admit a polyno- E MILIE DANNA
pendence system polyhedra can be spe- mial time approximation scheme unless ILOG
cialized. For example, facets arising P=NP, has interesting applications in a coauthors: Edward Rothberg, Claude
from generalized antiwebs, which gener- variety of fields such as computational Le Pape
alize cliques, odd holes, and antiholes in biology, linear programming, machine keywords: mixed integer programming,
graphs, only occur under very special cir- learning and telecommunications. To heuristics
cumstances. The tools used for such re- tackle medium-to-large scale instances
sults are based on a (combinatorial) con- of MaxFS, we consider randomized and Given a feasible solution to a Mixed Inte-
nection between the minimal infeasible thermal variants of the classical relax- ger Programming (MIP) model, a natural
subsystems and the support of vertices of ation method for solving systems of lin- question is whether that solution can be
an alternative polyhedron. ear inequalities. Based on the general- improved using local search techniques.
Finally, we will also report on com- ized boundedness theorem and the con- Local search has been applied very suc-
putational results to solve MaxFS by dition number presented in the previous cessfully in a variety of other combina-
branch-and-cut. talk, we establish an upper bound on the torial optimization domains. Unfortu-
expected number of iterations needed by nately, local search relies extensively on
these randomized relaxation methods to the notion of a solution neighborhood,
Generalized boundedness of the find an optimal solution of MaxFS. This and this neighborhood is almost always
relaxation method with an ex- implies, in particular, that they are guar- tailored to the structure of the particular
plicit bound in terms of a con- anteed to determine within a finite num- problem being solved. A MIP model typ-
dition number ber of iterations a solution which is op- ically conveys little information about the
R APHAEL H AUSER timal with probability one. Some com- underlying problem structure. We con-
Oxford University Computing putational results are also reported for in- sider a new approach, which we call Re-
Laboratory stances arising in planning digital video laxation Induced Neighborhood Search
broadcasts and in modeling the energy (RINS), to exploring interesting, domain-
coauthor: Edoardo Amaldi independent neighborhoods in MIP. It
function underlying the folding of pro-
keywords: infeasible linear systems, constructs a promising neighborhood us-
teins.
relaxation methods, condition number, ing information contained in the continu-
MAX FS FR1-302/45 INTEGER AND MIXED INTEGER
ous relaxation of the MIP model. Neigh-
PROGRAMMING

We study the behavior of relaxation Improved algorithms for borhood exploration is then formulated
methods when applied to infeasible sys- mixed-integer programming as a MIP model itself and solved recur-
tems of linear inequalities. We show that sively. Extensive computational experi-
organizer/chair : John W. Chinneck ments on very difficult MIP models show
the sequence of iterates generated by the
algorithm remains confined within a ball Pivot, cut, and dive: a class of that RINS outperforms default CPLEX
whose radius is a function of a condi- heuristics for general MIP MIP and two other recent strategies (lo-
tion number of the input data. This ra- cal branching and guided dives) with re-
dius plays a key role as a complexity J ONATHAN E CKSTEIN spect to several different metrics: qual-
measure that determines the convergence Rutgers University ity of the best integer solution produced
rates of an interesting class of algorithms coauthor: Mikhail Nediak within a time limit, ability to improve a
for solving the MaxFS problem. Our keywords: heuristics, mixed integer given integer solution (of both good and
result does not depend on the order in programming, intersection cuts, poor quality), and time required to diver-
which inequalities are considered and re- branch-and-bound sify the search in order to find a new so-
quires only very weak assumptions on the The pivot, cut, and dive heuristic frame- lution.
step lengths. This generalizes a classical work for general mixed integer program-
theorem by Block and Levin (1970) and ming was originally intended for use
answers their long-standing open ques- within exact branching methods, but may Faster MIP solutions through
tion to give an explicit bound as a func- also “stand alone”. The heart of the better variable ordering
tion of the input data. method is a simple integrality merit func-
tion which is concave when all integer
J OHN W. C HINNECK
variables are binary. We consider vari- Carleton University
Randomized relaxation meth- ous multi-objective pivoting and Frank- coauthor: Jagat Patel
ods for the maximum feasible Wolfe schemes to locate local minima of keywords: mixed integer programming,
subsystem problem this merit function without excessively branch-and-bound, branching variable
E DOARDO A MALDI degrading the original objective. To es- selection
DEI, Politecnico di Milano cape local minima of the merit function The selection of the branching variable
that are not integer feasible, several ad- can greatly affect MIP solution speed.
coauthors: Raphael Hauser, Pietro ditional levels of the heuristic can come We present new algorithms for choos-
Belotti into play: probing of adjacent vertices, ing the branching variable based on the
keywords: infeasible linear systems, intersection cuts which may be a natural impact of each variable on the active
randomized relaxation methods, consequence of probing, and a method of constraints in the current LP relaxation.
convergence analysis diving simultaneously on multiple vari- This constraints-oriented approach is a
In the MaxFS problem, given an infea- ables and introducing vertex/facial cuts. significant departure from the traditional
sible system of linear inequalities, one objective-oriented approach. We fo-
wishes to find a Feasible Subsystem con- cus on improving the speed to the first
taining as many inequalities as possi- Exploring relaxation induced integer-feasible solution. Empirical com-
ble. This combinatorial optimization neighborhoods to improve MIP parisions to solvers such as Cplex are
problem, which is approximable within solutions presented.
154 Parallel Sessions: Friday 9:00 – 10:30 (FR1)

FR1-302/49 INTEGER AND MIXED INTEGER

PROGRAMMING
improve the performance of mixed inte- We provide a continuous relaxation for
ger Gomory cuts. Mixed integer Gomory the cumulative constraint, one of the
Integer programming cuts are an integral part of state-of-the- most important global constraints used
organizer/chair : Ismael de Farias art solvers for MILP problems. Any im- in constraint programming (CP) sys-
provement in the performance of these tems. The constraint is used to for-
Semi-continuous cuts for mixed- cuts is therefore of great practical value. mulate resource-constrained scheduling
integer programming problems. We apply the relaxtion to solve
I SMAEL DE FARIAS FR1-306/31 CONSTRAINT LOGIC PROGRAMMING a multiple-machine scheduling problem
GSIA, Carnegie Mellon University with a hybrid CP/MILP method, based
Hybrid methods on a generalized Benders decomposi-
keywords: mixed integer programming, organizer/chair : John Hooker
branch-and-cut, polyhedral tion in which the subproblem separates
into cumulative constraints. Previous
combinatorics, semi continuous Constraint and integer pro- work suggests that a suitable relaxation
constraints gramming for the orthogonal of the subproblem could result in dra-
We study the convex hull of the fea- latin squares problem matic speedups relative to CP and MILP
sible set of the semi-continuous knap- acting alone. Since the cumulative con-
sack problem, in which the variables be-
D IMITRIS M AGOS
Technological Educational Institute of straint is difficult to formulate with inte-
long to the union of two intervals. Be- ger variables, we provide valid inequali-
Athens
sides being important in its own right, ties solely in terms of the original contin-
the semi-continuous knapsack problem is coauthors: Gautam Appa, Ioannis
uous variables.
a relaxation of general mixewd-integer Mourtos
programming, problems with partial in- keywords: orthogonal latin squares,
teger variables, and discrete optimiza- constraint programming, integer FR1-306/32 NONLINEAR PROGRAMMING

tion. We show how strong inequalities programming Surrogate modelling for

can be generated and used in a branch- Constraint programming (CP) and In-
engineering optimization III
and-cut scheme to solve these problems teger programming (IP) constitute two organizer/chair : Kaj Madsen
efficiently. generic frameworks dealing with hard Motivations and applications of
problems of combinatorial nature. IP has the space mapping technique
A polyhedral approach to ro- been developed within the Operational
bust integer programming Research discipline, whereas CP is con- K AJ M ADSEN
sidered to fall into the field of Artificial Technical University of Denmark
A LPER ATAMTURK Intelligence. In many cases, similar prob- keywords: non linear programming,
University of California at Berkeley lems have been addressed by CP and IP, surrogate modeling, space mapping
keywords: cutting plane/facet independently. Consequently, the inte-
The space mapping technique is a use-
The min-max robust optimization mod- gration of CP and IP methods deserves ful tool for pre-processing an optimiza-
els aim to find solutions minimizing the further investigation. Along this line of tion of engineering models which in-
maximum damage by the randomness of research, we present a hybrid algorithm, volve very cpu-intensive function evalua-
data. In this talk we present a polyhedral incorporating constraint and integer pro- tions. We use two different models of the
approach to robust integer programming. gramming techniques, for the Orthogo- same physical system: Besides the cpu-
nal Latin Squares (OLS) problem. We
briefly introduce some definitions. A intensive model of primary interest (de-
Improving the performance of Latin square is a square matrix of order n noted the fine model), access to a less ex-
mixed integer Gomory cuts where each value 1    n appears exactly pensive (coarse) model is assumed which
once in each row and column. Consider may be less accurate. We use the coarse
K ENT H ØJ A NDERSEN two Latin squares and the set of all the model to gain information about the fine
GSIA, Carnegie Mellon University model - thus quickly finding a rough es-
pairs each of which consists of the ele-
coauthors: Gerard Cornuejols, Yanjun ments of the two squares in the same row timate of the solution sought. For en-
Li and column. The two Latin Squares are gineering purposes the rough estimate is
keywords: mixed integer Gomory cuts, often sufficient - otherwise the method is

called 
Orthogonal if each of the n2 pairs
split cuts, intersection cuts, distance cut 1 1     n n  appears exactly once in considered a pre-prossing technique for
off . Finding such a structure constitutes one of the traditional optimization meth-
In the seventies, Balas introduced inter- the OLS problem. As a general result we ods. The Space Mapping principle has
section cuts for a Mixed Integer Linear report that the hybrid method performs been used by engineers for many years,
Program (MILP), and showed that these substantially better than "pure" CP and but the automatic way to do it - utilizing
cuts can be obtained by a closed form IP schemes, especially, as the size of the Broyden updates for estimating a connec-
formula from a basis of the standard lin- problem grows. tion between the two models - has only
ear programming relaxation. In the same been known since the first paper by Ban-
paper, Balas demonstrated that the well- A relaxation of the cumula- dler et al. in 1994. Since then a lot of ap-
known mixed integer Gomory cuts can be tive constraint for resource- plications have been published within the
viewed as intersection cuts. In the early constrained scheduling engineering society, and recently mathe-
nineties, Cook, Kannan and Schrijver in- maticians have shown interest in the the-
troducet the split closure of an MILP, and J OHN H OOKER oretical aspects of the method.
showed that the split closure is a poly- Carnegie Mellon University
hedron. It was shown recently that the coauthor: Hong Yan Space mapping: engineering
split closure can be obtained using only keywords: constraint programming, modeling and optimization ex-
intersection cuts. We use these facts to machine scheduling, hybrid methods ploiting surrogates
 155

J OHN BANDLER parameters. The method forms a pa- from a first estimate. We propose a La-
McMaster University rameter mapping that links a low fidelity grangian dualization which yields a con-
keywords: engineering, space mapping, (coarse) model to the fine model. Using vex differentiable dual problem. Numer-
surrogate modeling, trust regions the parameter mapping together with the ical results assess the method, even for
coarse model when performing the opti- very large matrices.
The space mapping approach to engi- mization saves fine model evaluations.
neering model enhancement and design SMO usually converges to a parame-
optimization intelligently links compan- ter set close to the fine model optimizer.
Gabor frames with optimal time
ion "coarse" (ideal or low-fidelity) and Convergence can be ensured by either
and frequency localization
"fine" (practical or high-fidelity) mod- performing a switch to a direct optimiza- J OACHIM DAHL
els of different complexities. Exam- tion method, or if the surrogates gener- Aalborg University, Dept. of Comm.
ples include full-wave electromagnetic ated by the method interpolate with the Tech.
(fine) simulations with empirical circuit- fine model function value and first deriva-
coauthors: Søren V. Andersen, Søren.
theory based (coarse or surrogate) sim- tives.
H. Jensen
ulations, or an engineering device un- This talk introduces work aimed at keywords: Gabor frames, localization,
der test coupled with a suitable simu- generating interpolating surrogates us-
global optimization
lation surrogate. Our methodology has ing a Space Mapping approach. Re-
been adopted for diverse design applica- sults showing significant increase in the We consider the design of Gabor analysis
tions: electronic components, magnetic performance of the method will be pre- and synthesis filterbanks with oversam-
systems, civil and mechanical engineer- sented. More practical ways of generat- pling (Gabor frames). We design Gabor
ing structures. Space mapping facilitates ing interpolating surrogates will be dis- frames with optimal trade-off between
efficient optimization while avoiding di- cussed. time- and frequency localization. This is
rect optimization of the fine model. It is a a difficult non-convex optimization prob-
simple CAD methodology, which closely FR1-306/33 NONLINEAR PROGRAMMING lem, but suprisingly it can be solved glob-
follows the traditional experience and in- ally and efficiently via a dual problem us-
tuition of engineers, yet is amenable to Semidefinite problems ing semidefinite programming. As an in-
rigorous mathematical treatment. Fol- chair : Joachim Dahl teresting consequence of the problem for-
lowing the original concept of Bandler in A global algorithm for nonlin- mulation, the optimal analysis and syn-
1993, algorithms have utilized Broyden ear semidefinite programming thesis frames become identical which is
updates, trust regions, and artificial neu- appealing in many engineering applica-
ral networks. New developments include H ECTOR R AMIREZ C ABRERA tions.
(1) implicit space mapping, in which we INRIA-Rocquencourt
allow preassigned parameters not used coauthor: Rafael Correa FR1-306/34 NONLINEAR PROGRAMMING

in optimization to change in the coarse keywords: nonlinear semidefinite prog., Large scale problems IV
model, and (2) output space mapping, sequentially programming, global chair : Gerardo Toraldo
where a transformation is applied to the convergence
response of the model. We present il- PENNON - a code for large-
lustrative examples such as the cheese- In this paper we propose a global algo- scale nonconvex npl and SDP:
cutting problem, and new results applica- rithm for solving nonlinear semidefinite algorithm, theory, numerial re-
ble to RF, wireless and microwave circuit programming problems. This algorithm, sults
design, integrating electromagnetic sim- inspired in the classic SQP (Sequentially
ulations. Quadratic Programming) method, mod- M ICHAEL S TINGL
ifies the S-SDP (Sequentially Semidefi- Institute of Applied Mathematics II,
nite Programming) local method by using University of Erlangen-Nuremberg
Space mapping optimization us- a nondifferentiable merit function com- coauthor: Michal Kocvara
ing interpolating surrogates bined with a line search strategy. keywords: nonlinear programming,
semidefinite programming, augmented
F RANK P EDERSEN Covariance matrices calibration Lagrangian method
Civil Engineering Department
via semidefinite least-squares A class of iterative methods for convex
keywords: non linear programming,
M ALICK J EROME nonlinear programming problems was in-
trust region method, space mapping
INRIA troduced by Ben-Tal and Zibulevsky and
optimization, interpolating surrogates
named PBM. The framework of the al-
Designing a system in an optimal way keywords: matrix calibration, gorithm is given by the augmented La-
can be done by applying numerical op- Lagrangian duality, semidefinite grangian method; the difference to the
timization methods to a mathematical optimization classic algorithm is in the definition of
model of the system. Using a time con- We present a method to solve the follow- a special penalty/barrier function satisfy-
suming, high fidelity (fine) model of the ing problem, called semidefinite least- ing certain properties. A generalization
system when optimizing it, is problem- squares: to project a matrix onto the of the PBM method for convex semidefi-
atic since solving an optimization prob- intersection of the cone of positive nite programming problems was recently
lem requires a potentially large num- semidefinite matrices and an affine sub- proposed by the authors. The algorithm
ber of fine model evaluations to be per- space. An important application of was implemented in the code PENNON,
formed. semidefinite least-squares is the so-called that proved to be very efficient for convex
The Space Mapping Optimization calibration of covariance matrices, which NLP and linear SDP problems.
(SMO) method specifically aims at bring- arises for example in portfolio risk anal- In this talk, we present a general-
ing the number of fine model evaluations ysis: to find a "good” approximation ization of the algorithm for large-scale
down, required to find the optimal model covariance matrix between some assets, nonconvex optimization problems with
156 Parallel Sessions: Friday 9:00 – 10:30 (FR1)

NLP and(generally nonlinear) SDP con- an interior-point method and a gradient- purpose the tools of quantifier elimina-
straints. We propose two strategies for projection strategy. The basic idea is tion over real closed fields with infinites-
treatment of nonconvexity. In the first to perform interior-point iterations un- imals are involved. The latter are used to
one, we replace the Newton method, used til a point sufficiently close to the so- bring the problem into general position,
for approximate minimization of the aug- lution is obtained. Starting from this in certain well-defined sense.
mented Lagrangian, by the Levenberg- point, gradient-projection steps are per- In contrast, the best previously
Marquardt algorithm. In the second one, formed in order to refine the final com- known procedure only works in  the
we propose to use for this purpose the full puted solution. We focus on the large homogeneous case, and needs nO k2 
trust region method. For the latter ap- and sparse case, and on the use of iter- operations just for deciding feasibility,
proach we will give a convergence the- ative techniques to solve the linear sys- while no polynomial-time procedure pre-
ory. tem arising at each iteration of the algo- viously known is able to find a feasible
Both approaches were implemented rithm. In order to deal with the increasing point with prescribed precision.
in a code PENNON. We will compare ill-conditioning of such systems when Our result implies a complexity
PENNON to state-of-the-art codes on a the iterates get close to the boundary of bound for semidefinite programming an-
large set of standard NLP problems. Re- the feasible region, we use precondition- nounced in 1997 by Khachiyan and
sults for nonconvex SDP problems will ing techniques based on incomplete fac- Porkolab without a proof.
be presented in a separate talk. torizations with predictable memory re-
quirements. Moreover, we propose suit- On the rank-reducibility of bar-
A parallel NPL code for optimal able termination rules allowing to adapt rier functions
control problems with bounded the accuracy requirement in solving the
linear systems on the quality of the cur- Y VAN H ACHEZ
states Université catholique de Louvain
rent iterate. We show some results of
J OAO -L AURO FACO ´ computational experiments carried out an coauthor: Yurii Nesterov
Universidade Federal do Rio de Janeiro a IBM SP system and on a cluster of keywords: semidefinite programming,
keywords: dynamic systems, nonlinear PC’s, variyng the accuracy requirements nonnegative polynomials, spectral
optimization, parallel computing, for both methods. We also compare factorization
reduced gradient the proposed algorithm, with the package In this talk, we present a new class of
MOSEK. convex (dual) barrier functions, which
Nonlinear dynamic systems are consid-
permit low-rank representations in the
ered as discrete-time optimal control
(primal space. These barrier functions
problems with bounded state variables. FR1-306/35 NONLINEAR PROGRAMMING

are obtained from “rank-reduced” self-

A specialized Generalized Reduced Gra- Semidefinite programming concordant barrier functions, but they
dient method that exploits the staircase and nonnegative polynomials do not to enjoy this important property
structure of the jacobian matrix of the
organizer/chair : Etienne de Klerk in general. Although our results are
dynamic equations by using some prior-
strongly related to spectral factorization
ity principles in the choice of the basic Optimization subject to a fixed of nonnegative matrix polynomials, they
variables generating factorized represen- number of quadratic con- can be applied to other convex cones
tations of the GRG basic matix. Linear straints in polytime as well. Because these barrier func-
systems of equations are solved by par-
tions are obtained from specific linear
allel computing techniques for the LU- D MITRII V. PASECHNIK matrix inequalities, low-rank representa-
decompositions of the diagonal blocks in Theoretische Informatik, FB 15, tions can be efficiently computed using
the basic matrix factorization improving Universitaet Frankfurt any semidefinite programming solver.
efficiency and numerical stability. In the
reduced dimension space of the indepen- coauthors: Etienne de Klerk, Dima
dent variables we have a differentiable
Grigoriev Sum of squares decompositions
nonlinear objective-function. The mini-
keywords: quadratic constraints, for structured polynomials
mization search directions can be com- polynomial time, symbolic computation,
real algebraic geometry PABLO PARRILO
puted by an adapted parallel conjugate- ETH Zürich
gradient method or a limited-memory It is shown that quadratic programming keywords: sum of squares, algebraic
quasi-Newton method. A computer code

in n variables subject to k quadratic con- optimization, semidefinite programming,
is presented, and numerical experiments straints has time complexity nO k  . More convex optimization
with practical optimal control problems generally, given a quadratic map Q :
are discussed. Rn

Rk , optimization over a semialge- The decomposition of a multivariate
polynomial as a sum of squares is a ba-
 
braic set specified by s polynomial in- sic question in applied mathematics. It
An hybrid algorithm for large- equalities fi Q X  
0, and any  number is also one with important consequences,
scale quadratic programs 
of polynomial equations f i  X    0,
Q as such decompositions can be used via
has time complexity snd  O k  , where d the Positivstellensatz as easily verifiable
G ERARDO TORALDO is the maximal degree of the f i ’s. suboptimality certificates in optimiza-
University of Naples
The described procedure reports the tion. Particularly exciting is the recent
coauthor: Marco D’Apuzzo range of the objective function on the fea- availability of efficient techniques, based
keywords: hybrid algorithms, interior sible region. For any value of the objec- on semidefinite programming, for its ef-
point method, gradient-projection tive function in its range the algorithm fective computation. In this talk, we
method, adaptive termination rule can find a feasible point with this pre- present an overview of the available al-
We present an hybrid algorithm for solv- scribed value. Optimal points and values gorithms, emphasizing our recent devel-
ing convex quadratic programming prob- are algebraic numbers represented by the opments oriented towards the exploita-
lems. The algorithm suitably combines algorithm in a symbolic way. For this tion of additional algebraic properties,
 157

such as symmetries and ideal structure. We present preconditioned Krylov sub-

The ideas and algorithms will be illus- A new feasible directions algo- space methods for the augmented and
trated with examples from a broad range rithm for nonsmooth convex op- Schur complement equations arising
of domains, and the use of the SOS- timization from interior-point methods for solv-
TOOLS software (developed in collabo- ing large scale semidefinite program-
ration with Stephen Prajna and Antonis J OSÉ H ERSKOVITS ming. When the SDP is primal-dual
Papachristodoulou). COPPE/Federal University of Rio de non-degenerate and strict complementar-
Janeiro ity condition holds, the preconditioned
FR1-306/36 NONSMOOTH OPTIMIZATION coauthors: Wilhelm Passarella, Susana matrices are shown to have bounded con-
Scheimberg, Regina Burachik dition numbers even when the barrier pa-
Nonsmooth optimization II keywords: nonsmooth optimization, rameter decreases to 0. Numerical exper-
chair : José Herskovits nonlinear programming, nonlinear iments on some large scale SDPs arising
Twice differentiable spectral optimization from maximum clique problems show
functions We consider the Problem that the preconditioners are effective. In
 particular, we are able to solve such an
H RISTO S ENDOV min F x  , x   n SDP with more than 100,000 constraints
University of Guelph where F is convex and not necessar- in under 10 hours to a relative accuracy
of less than 10  6.
coauthor: Adrian S. Lewis ily differentiable. Let be the Equivalent
keywords: spectral functions, Problem EP:
eigenvalue optimization, symmetric  A conic bundle approach to lin-
functions, Hessian min z   , such that F x   z
ear programming over symmet-
A function, F, on the space of n  n real   an algorithm that builds a se- ric cones
We present
symmetric matrices is called spectral if quence xk zk 
at the interior of the
C HRISTOPH H ELMBERG
it depends only on the eigenvalues
 of its epigraph of F, such that zk  1 zk .
argument, that is F A   F UAU T  for Our method employs the Feasible Di- Chemnitz University of Technology
every orthogonal U and symmetric A in rections Interior Point Algorithm, FD- keywords: Lagrangean relaxation,
its domain. Spectral functions are in one- IPA, for constrained smooth optimiza- primal aggregation, cutting models,
to-one correspondence with the symmet- tion, [1]. At each iteration, by solving software
ric functions, f , on  n , those invariant two linear systems, FD-IPA computes a In combinatorial optimization and
under permutations of their arguments. Feasible Descent Direction. Making a stochastic programming Lagrangean re-
We show that a spectral function is twice line search, a new interior point with a laxation is frequently employed to ex-
(continuously) differentiable at a matrix lower objective is obtained. ploit structural properties of the under-
if and only if the corresponding symmet- The present algorithm defines a se- lying problem. Bundle methods allow
ric function is twice (continuously) dif- quence of Auxiliary Problems AP, where to optimize over the multipliers while
ferentiable at the vector of eigenvalues. the constraints of EP are approximated generating, at the same time, approxi-
We give a concise and usable formula for by cutting planes. mate primal solutions via primal aggre-
the Hessian. In the case of convex func- At each iteration a Search Direction gation. Convergence properties of the
tions we also show that a positive definite for EP is obtained by computing with bundle method and the quality of this
Hessian of f implies positive definiteness FD-IPA a Feasible Descent Direction of primal approximation largely depend on
of the Hessian of F. AP. If the step length is "short", AP is up- the quality of the cutting model. For
dated and a new search direction is com- linear programming and general convex
Characterization of error puted. This procedure is repeated until a optimization via first order oracles the
bounds for lower semicontin- "good" step is obtained. When this hap- linear cutting plane model is state of
uous functions pens, the search direction is a Feasible the art. For semidefinite programming
Descent Direction of EP. We prove global a specialized model was introduced in
D OMINIQUE A ZÉ convergence and solve several test prob- the spectral bundle method. Recently,
University Paul Sabatier, Toulouse, lems very efficiently. we proposed a similar model for sec-
Laboratoire MIP [1] Herskovits J. "A Feasible Direc- ond order cone programming. We report
keywords: error bounds, Hoffman type tions Interior Point Technique For Non- on our efforts to combine these features
estimates linear Optimization", JOTA, v99-1, 1998. (primal aggregation, linear, semidefinite,
We give a characterization of the exis- and second order models) in a "Conic
tence of an error bound for lower semi- FR1-306/37 CONVEX PROGRAMMING Bundle Package" for linear programming
continuous functions. More precisely, Large-scale semidefinite over symmetric cones and present some
given a l.s.c. function f : X

 preliminary numerical results.


programming
 ∞
defined on a complete metric
space X and given α   , we charac- organizer/chair : Madhu Solving large-scale SDP’s in
  of τ 0 such that
Nayakkankuppam

the existence
terize parallel
τd x f  γ  f  x 
γ   for all x  X Solving large scale semidefinite
and γ
α where d x f  γ  denotes the programming via Krylov sub- M ADHU NAYAKKANKUPPAM
distance function to the upper-level set of space methods Department of Mathematics & Statistics,
f at height γ. Several applications are UMBC
given to metric regularity of multifunc- K IM -C HUAN TOH coauthor: Yevgen Tymofyeyev
tions and to the existence of a Hoffman National University of Singapore keywords: semidefinite programming,
type estimate in semidefinite program- keywords: large scale SDP, subgradient bundle methods, Lanczos
ming. preconditioners method, parallel computing
158 Parallel Sessions: Friday 9:00 – 10:30 (FR1)

We describe a portable, parallel and Convex- and monotone-

distributed implementation of the spec- A solution algorithm for a class transformable mathematical
tral bundle method for semidefinite pro- of box constrained quadratic programming problems and a
grams (SDP) with the usual block diag- programming problem proximal-like point method
onal structure. Subgradients are com-
puted via a Lanczos method with im- C LAUDIO S ODINI O RIZON F ERREIRA
plicit restarts, but specially tailored to University of Pisa Federal University of Goias
handle block structure. Based on the coauthor: Riccardo Cambini
Message Passing Interface, this code ex- keywords: quadratic programming, coauthors: Joao Xavier da Cruz Neto,
ploits parallelism to solve (to low ac- optimal level solutions, D.C. Luis Roman Lucambio Perez, Sandor Z.
curacy) problems larger than previously optimization Nemeth
possible. (For instance, a Lovasz theta keywords: monotone vector fields,
The aim of this paper is to suggest an
function SDP on a graph with 2000 nodes Hadamard manifolds, non convex
algorithm to solve a box constrained
and 20,000 edges is solved to about 3 dig- constrained problem, proximal point
quadratic problem where the objective
its of accuracy in under 12 minutes on 16 algorithm
function is given by the sum of a
processors). Our current efforts are di- quadratic strictly convex separable func-
rected towards improvements necessary The problem of finding singularities of
tion and the square of an affine function monotone vectors fields on Hadamard
to achieve scalability and superior paral- multiplied by a real parameter.
lel performance on very large-scale prob- manifolds will be considered and solved
Depending on the value of the real by extending the well-known proximal
lems. parameter, the problem might be con- point algorithm. For monotone vec-
vex or not. The particular structure of tor fields the algorithm will generate a
FR1-306/38 GLOBAL OPTIMIZATION the problem allows to suggest an efficient well defined sequence, and for mono-
Quadratic optimization parametric algorithm solving the prob- tone vector fields with singularities it will
lem in a finite number of steps. converge to a singularity. It will be
chair : Andreas Brieden
also shown how tools of convex analy-
Solving nonconvex quadratic Clustering in agriculture by sis on Riemannian manifolds can solve
programming problems with means of quadratic optimiza- non-convex constrained problems in Eu-
simple bound constraints tion clidean spaces. To illustrate this remark-
A NDREAS B RIEDEN able fact examples will be given.
J OHANNES J. DE N IJS TU München
Old Dominion University
keywords: quadratic optimization, A projection-type method with
keywords: quadratic programming, randomized approximation, convex
unconstrained optimization, Newton variational metric for solving
maximization a general variational inequality
method, quadratic spline
In the process of passing farmland to problem
A nonconvex quadratic programming heirs a formerly large piece of land has
problem with simple bound constraints under the heritage laws of certain coun-
can be reformulated as an unconstrained tries been split over the centuries into
S USANA S CHEIMBERG
Universidade Federal do Rio de
minimization problem with a (onconvex) small lots. In fact, for this and other rea-
Janeiro,UFRJ/COPPE-PESC
quadratic spline as the objective func- sons today’s farmers may own a num-
tion. Newton-type methods can be used ber of small sized lots that are distributed coauthor: Paulo Sergio Santos
to define descent directions for finding over a wide range leading to high agricul- keywords: general variational
minimizers of unconstrained optimiza- tural production costs. inequality, projection method,
tion problems, but this usually requires a For various reasons classical forms of variational metric, cocoercivity
positive definite Hessian, as is an O n3  land consolidation become less and less
algorithm for finding a solution through feasible. It has hence been suggested In this work we consider the Gen-
a Cholesky factorization. In this pa- to support the voluntary lend-lease based eral Variational Inequality problem in a
per we propose an algorithm that uses exchange of agricultural acreage. Due Hilbert space, (GVI), introduced by M.
a global line search method to find a to combinatorial explosion the underly- A. Noor in 1988. We present an iterative
stationary point satisfying the secon or- ing combinatorial optimization problem scheme based in a projection on an inte-
der necessary optimality condition, us- becomes hard to solve for almost all sam- rior approximation of the constraint set,
ing modifications to the Cholesky fac- ple regions. Hence, in order to fully ex- that can be interesting from a practical
torization. The method can be consid- ploit the potential of the method mathe- point of view. The convergence analy-
ered as an implicit active set method with matical optimization is required. sis requieres the existence of a solution
the novel feature of a mixed primal-dual In this talk a quadratic optimization of (GVI) and a usual cocoercivity condi-
approach for identifying active indices, model is presented that also intractable in tion of the underlying operator. We also
and a global exact line search strategy theory, is successfully solved by means consider a variational metric introduced
for global solutions. The first feature of a randomized approximation algo- by H. Attouch and R. J. B. Wets in 1986.
provides a dynamic balance between the rithm for Euclidean norm-maximization Preliminary computational experience is
need of minimizing the original function in practice. reported. We compare our method with
and the need of forcing the iterates to stay the algorithms given by M. A. Noor et al.
in the feasible region. Directions of neg- FR1-308/11 GENERALIZED
CONVEXITY / MONOTONICITY
and by B. He using test problems.
ative curvature at stationary points and Generalized
global search help to find a global solu- convexity/monotonicity
tion of the nonconvex quadratic program- A method for solving monotone
ming problem. chair : Evgeny G. Gol’shtein variational inequalities
 159

E VGENY G. G OL’ SHTEIN We consider the equation whith un- as a membrane. In addition, there are
Central Economics @ Mathematics known variables varying in a given con- rigid obstacles that unilaterally limit the
Institute of Russian Academy of Sciences vex closed cone. This cone defines the deformation. This gives rise to a con-
keywords: monotone mapping, oracle, enequality type constraints. The corre- strained, non convex variational prob-
measure of proximity sponding nonlinear mapping is assumed lem. We derive the optimality system
We consider a class of variational in- to be sufficiently smooth. The inverse and its interpretation in terms of equilib-
equalities each of which is defined by function and implicit function theorems rium of forces. A second order sensitiv-
a monotone mapping with convex val- are proved. The main distinction of pre- ity analysis, allowing to compute deriva-
ues and by a convex polyhedron compris- sented results from wellknown is that the tives of solutions and a second order
ing the domain of the mapping. In the point under consideration may be abnor- Taylor expansion of the cost function,
paper, we describe an oracle-type algo- mal. In other words we don’t a’priori as- is performed, in spite of the fact that
rithm solving the variational inequalities sume that Robinson’s constrained quali- the cost function is not twice differen-
of the above-mentioned class. Under the fication are satisfyed. tiable. We also study the finite elements
assumption of the mapping being upper discretization, introduce a decomposition
semi-continuous, the algorithm’s conver- Sensitivity analysis for abnor- algorithm for the numerical computation
gence is established. We also present a mal optimization problems of the solution, and display numerical re-
way to calculate the measure of proxim- sults.
ity of the current iteration point to the A LEXEY I ZMAILOV
set of solutions of the variational inequal- Moscow State University
Solution of the optimal set par-
ity. This way is an extension of the func- coauthor: Aram Arutyunov titioning problem as problem of
tional proximity measure used in the op- keywords: sensitivity analysis, minimization the sets function
timization problems. In terms of the pro- parametric optimization,
posed proximity measure, we obtain the cone-constrained problem, constraint TATYANA S TEPANCHUK
algorithm’s convergence rate that proves qualifications Phd. student /Department of applied
its optimality on the considered class of
For the parametric cone-constrained op- mathematics and mathematical
variational inequalities.
timization problem, we consider the case cybernetics/Dnepropetrovsk national
when the customary (Robinson’s, direc- university
FR1-308/12 STOCHASTIC PROGRAMMING

tional) regularity conditions can be vio- coauthor: Elena Kiseleva

Sensitivity and stability lated at solution of the unperturbed prob- keywords: minimization the sets
chair : Alexey Izmailov lem. Under the assumptions substantially function, optimal set partitioning
weaker than those previously used in the
Nonparametric kernel estima- literature in this context, we develop a The problem of optimal set partitioning
tion via nonstationary stochastic reasonably complete local sensitivity the- of the set Ω from an n -dimensional Eu-
optimization ory for this class of problems, including clidean space En into its disjoint sub-
V LADIMIR N ORKIN upper and lower bounds for the rate of sets with unknown in advance subset
Institute of Cybernetics of the Ukrainian change of the optimal value function sub- center coordinates with constraints of
Academy of Sciences ject to parametric perturbations, as well the form of equalities and/or inequali-
as the estimates and the description of ties is considered. This problem is the
coauthors: Y. M. Ermoliev, M. A.
asymptotic behavior of solutions of the novel problem of the theory of opti-
Keyzer
perturbed problems. This development mization the sets function. The neces-
keywords: kernel estimation, stochastic
relies on certain genaralization of Robin- sary conditions for the extremum that
optimization, density/regression
son’s stability theorem, extending the lat- are generalized well-known fundamen-
estimation, tracking
ter to the nonregular case. tal Neyman-Pearson lemma are obtained
The report considers nonparametric ker- based on the mathematical tools of the
nel density/regression estimation from SEMI - INFINITE AND INFINITE theory of optimization the sets function.
stochastic optimization point of view. FR1-308/13 DIMENSIONAL PROGRAMMING
The algorithms for numerical solution of
The estimation problem is represented Semi-infinite and infinite the above-mentioned problem are con-
through a family of stochastic optimiza- dimensional programming structed, programmed and evaluated on a
tion problems. Recursive constrained chair : Tatyana Stepanchuk set of test problems.
estimators are obtained by application
of stochastic (quasi)gradient methods to The obstacle problem for water
these problems, classical kernel estimates tanks FR1-308/T1 DYNAMIC PROGRAMMING AND
OPTIMAL CONTROL

are derived as particular cases. Accuracy Large-scale dynamic

and rate of convergence of the obtained R ADHIA B ESSI programming - algorithms
estimates are established, asymptotically ENIT and approximation
optimal estimation procedure parameters coauthors: Frederic Bonnans, R. Bessi
are found. The case of moving den- Fourati, Hichem Smaoui organizer/chair : Marina Epelman
sity/regression is particularly studied. keywords: obstacle problem, nonconvex Efficient average-cost optima
variational problems, sensitivity for infinite horizon optimization
Stability theory for systems of analysis, finite elements
ineaqualities at abnormal points Abstract In this work we discuss the I RWIN S CHOCHETMAN
A RAM A RUTYUNOV problem of computing and analyzing the Oakland University
People’s Friendship University of Russia static equilibrium of a non rigid water coauthor: Robert L. Smith
tank. Specifically, we fix the amount keywords: average-cost, efficient, state
keywords: abnormal point, cone,
of water contained in the tank, modelled reachability, existence
inverse function
160 Parallel Sessions: Friday 9:00 – 10:30 (FR1)

Those infinite horizon optimization prob- provides a method to incorporate action In this work we show that the mixed
lems which do not admit a minimum- space restrictions which have dependen- nonlinear complementarity problem may
cost (strong) optimum, may also not ad- cies across states, a modification which be formulated as an equivalent nonlinear
mit a (minimum-) average-cost optimum. generally results in the loss of the Marko- bound-constrained optimization problem
Furthermore, if an average-cost optimum vian property. We demonstrate the aggre- that preserves the smoothness of the orig-
does exist, there will exist many such in gation method on a radiation treatment inal data. One may thus take advantage
general because of instability over finite planning problem. of existing codes for bound-constrained
horizons. Thus, it is desirable to seek optimization. This approach is imple-
average-cost optima which are also effi- mented and tested by means of an exten-
cient (finite optimal) over finite horizons,
FR1-308/T2 COMPLEMENTARITY AND VARIATIONAL
INEQUALITIES
sive set of numerical experiments, show-
i.e., attain each of their states optimally. Nonlinear complementarity ing promising results. The mixed nonlin-
We show that efficient optima exist in our problems ear complementarity problem considered
broad context and, under a general state- chair : Natasa Krejic in the tests arises from the discretization
reachability condition, conclude that ef- of a motion planning problem concern-
ficient optima are also average-cost opti- An investigation of a nonlin- ing a set of rigid 3D bodies in contact in
mal. ear obstacle plate optimization the presence of friction. For each time
problem frame one needs to solve a complemen-
Solution acceleration through J OAO PATRICIO tarity problem in order to calculate the
nested aggregation Escola Superior de Tecnologia de Tomar accelerations of the bodies involved.
coauthors: Joaquim Judice, Isabel
M ATTHEW BAILEY Figueiredo, Luis Merca Fernandes On a quasi-Newton method for
University of Pittsburgh keywords: variational inequalities, complementarity problems
keywords: dynamic programming, nonlinear programming, NATASA K REJIC
aggregation, deterministic complementarity problem, obstacle Department of Mathematics and
Dynamic Programming is a powerful, ro- problem Informatics, University of Novi Sad
bust modeling mechanism for sequential We discuss the solution of a nonlinear coauthor: Zorana Luzanin
decision processes. A significant deter- obstacle plate problem which describes keywords: complemetarity problems,
rent to its use is the computation time the equilibrium of a thin elastic clamped quasi-Newton methods
and computer storage requirements re- plate, that may come into contact with Nonlinear complementarity problems ap-
sulting from the exponential growth in a rigid obstacle, by the action of exter- pear in many engineering and economic
the state space of most models. One of nal forces. The discrete formulation of applications. Such problems are of-
the methods for mitigating this problem the problem is based on the finite el- ten solved by iterative methods based
is aggregation. The creation of aggregate ement method, using the Bogner-Fox- on generalization of classical Newton
states results in the loss of information Schmit and the Melosh rectangle, with 16 method and its modifications for smoorh
and hence an error in the resulting solu- and 8 degrees of freedom, respectively, equations. We consider the Modification
tion. This results in an approximate so- to approximate the vertical and horizon- of the Right-Hand Side Vector (MRV)
lution technique. Typically aggregation tal displacements of the middle plane of quase-Newton method for a semismooth
techniques have three steps: aggregation, the plate. Due to the symmetry of the reformulation of NCP. The local conver-
solution, and disaggregation to find the associated function of the variational in- gence is proved and some numerical re-
associated feasible solution. We discuss equality, it is possible to solve the cor- sults are presented.
an extension of sequential aggregation responding discrete model as a nonlinear
disaggregation to included a nested hier- program with simple bounds on the vari-
archy of aggregated macronodes in a pos- ables. A number of algorithms exploiting FR1-308/T3 TELECOMMUNICATION NETWORK
DESIGN

sibly cyclic network. the structure of the optimization prob- Models and simulations for
lem is investigated. Numerical results planning and control of UMTS
Aggregation in stochastic dy- of the application of these algorithms are (momentum)
namic programming included to highlight their efficiency in organizer/chair : Arie M.C.A. Koster
practice.
M ARINA E PELMAN Challenges in UMTS radio net-
University of Michigan work planning
Mixed nonlinear complemen-
coauthors: Theodore J. Lambert III, tarity problems via nonlinear A NDREAS E ISENBLÄTTER
Robert L. Smith optimization: numerical re- ZIB
keywords: dynamic programming, sults on multi-rigid-body con- coauthors: Alexander Martin, Thorsten
Markov: finite state, aggregation tact problems with Coulomb Koch, Roland Wessäly
We present a general aggregation method
friction keywords: radio network planning,
for Markov decision problems under S ANDRA AUGUSTA S ANTOS UMTS, momentum
which the aggregate solution is imme- State University of Campinas - The Universal Mobile Telecommunica-
diately applicable to the original prob- UNICAMP tions System (UMTS) is a 3rd genera-
lem (i.e., no disaggregation step is re- tion cellular system for mobile telecom-
coauthors: Roberto Andreani, Ana
quired). The method is based on select- munications. In UMTS cell coverage
Friedlander, Margarida P. Mello
ing a collection of “distinguished” states. and cell capacity are inversely coupled in
keywords: complementarity problem,
We provide error bounds for general time UMTS. The radio transmissions for dif-
bound-constrained minimization,
dependent “distinguished” state selection ferent mobiles are not separated in time
multi-rigid-body contact prob., Coulomb
method. The aggregation scheme also or frequency (unlike in GSM), but they
friction
 161

are separated by means of coding. For a objective is to minimize the total cost of F RANÇOIS G LINEUR
proper decoding, a minimum ratio, called the network. CORE / UCL
CIR, between the carrier’s strength at keywords: second order cone
the receiver and the interference has to optimization, conic optimization, convex
be achieved. The network’s capability Optimisation methods for
UMTS radio network planning optimization
to serve users’ demand depends on the
transmission powers required to satisfy We consider the standard conic opti-
all CIR requirements. A LEXANDER M ARTIN mization problem involving only a sin-
The typical network planning pro- TU Darmstadt gle second-order cone and do not assume
cess cycles in a time consuming trial- coauthors: Andreas Eisenblätter, any type of constraint qualification (e.g.
and-error fashion through tentative net- Roland Wessäly, Thorsten Koch Slater) condition.
work designs and simulation-based net- keywords: UMTS, radio network We first tackle the associated feasi-
work performance analysis. An auto- planning, momentum, mixed integer bility problem, which amounts to decid-
matic search for a low-cost network de- programming ing whether the problem is strictly fea-
sign, achieving the required quality stan- sible, weakly feasible, weakly infeasible
dard is therefore in high demand. The Universal Mobile Telecommunica- or strongly infeasible. We show that this
This talk introduces the context of tions System (UMTS) is a 3rd generation can be determined without any iterative
cellular system for mobile telecommu- process by solving a few linear equality
network planning. Service and source
nications. The planning of UMTS net- systems.
models, spatial load distributions, and
hardware characteristics are explained. works is a challenging problem. One We then outline how this procedure
key issue is to select base station loca- can be generalized to solve the original
The mathematical modelling and solving
tions and to configure base stations in- second-order cone optimization problem.
of the radio network planning challenge
are addressed in the two companion talks cluding antenna types, heights, azimuths, In particular, we obtain a proof of the
and tilts such that the required services fact that such second-order cone prob-
by Roland Wessäly and Alexander Mar-
are met. The problem leads to a huge lems can never exhibit a duality gap.
tin.
mixed integer program the modelling of We also discuss the algorithmic com-
which has been discussed in the preced- plexity viewpoint and possible general-
Models for UMTS radio net- ing presentations by Roland Wessäly and izations of this procedure.
work planning Andreas Eisenblätter. In this talk we deal
with solution methods of this model. We
ROLAND W ESSÄLY demonstrate success and limits of general Uniform boundedness of a pre-
Konrad-Zuse-Zentrum MIP solvers, we present model reformu- conditioned normal matrix used
coauthors: Andreas Eisenblätter, lations including valid and heuristic in- in interior point methods
Alexander Martin, Thorsten Koch equalities, and we discuss primal meta- J EROME O’N EAL
keywords: radio network planning, heuristics. Computational results on re- Georgia Institute of Technology
UMTS, momentum, mixed integer alistic data conclude the presentation.
programming coauthors: Renato Monteiro, Takashi
Tsuchiya
The Universal Mobile Telecommunica- FR1-341/21 INTERIOR POINT ALGORITHMS
keywords: linear programming, interior
tions System (UMTS) is a 3rd generation point method, network flow problems,
cellular system for mobile telecommuni-
Second-order cone
cations. Even though UMTS networks
optimization and extensions condition number
are currently being deployed in many organizer/chair : Erling Andersen Solving systems of linear equations with
European countries, the development of The “normal” matrices of the form AD2 AT
simplex
planning and optimization methodolo- semidefinite and second-order
method for is a key ingredient in the computation
gies which cover the many particulari- cone programming of search directions for interior-point al-
ties of these networks is still in its begin- gorithms. In this article, we establish
nings. The project MOMENTUM, which D ON G OLDFARB that a well-known basis preconditioner
is supported by the European Union un- Columbia University for such systems of linear equations pro-
der the Information Society Technologies duces scaled matrices with uniformly
(IST) Programme is one of the driving keywords: conic programming, simplex bounded condition numbers as D varies
forces to push these developments ahead. method, semidefinite programming, over the set of all positive diagonal matri-
This presentation focuses on model- second order cone programming ces. In particular, we show that when A
ing aspects related to planning and opti- A vast array of convex optimization prob- is the node-arc incidence matrix of a con-
mizing the radio interface of UMTS net- lems can be formulated as semidefinite nected directed graph with one of its rows
works. It is accompanied by the preced- and second-order cone programs. We deleted, then the condition number of
ing presentation of Andreas Eisenblätter present a natural and straightforward ex- the corresponding preconditioned  normal
introducing principles of the UMTS radio tension of the simplex method to these matrix is bounded above by m n m  1  ,
interface and the succeeding of Alexan- classes of problems and discuss how where m and n are the number of nodes
der Martin one explaining optimization to efficiently implement various "pivot and arcs of the network.
methods. The presented mixed-integer rules". We also present numerical results
linear programming model has been de- to illustrate the performance of these FR1-341/22 LOGISTICS AND TRANSPORTATION

veloped within MOMENTUM and cov- variants of our simplex method. Logistics and transportation
ers a broad range of planning issues: site
chair : Maria Daneva
selection for the UMTS base stations, cell
sectorization, antenna selection and con- Second-order cone optimization A robust optimization model for
figuration (height, tilt, azimuths). The with a single second-order cone empty container allocations
162 Parallel Sessions: Friday 9:00 – 10:30 (FR1)

K IN K EUNG L AI FR1-341/23 which capture the intrinsic power and

APPROXIMATION ALGORITHMS

City University of Hong Kong limitations of greedy algorithms. The

Priority algorithms: models original paper was limited to scheduling
coauthors: Y. Wu, Stephen C. H. Leung and approximation bounds for
keywords: empty container, robust problems; but subsequent workextended
greedy-like algorithms the model to other domains.
optimization, distribution planning
organizer/chair : Joan Boyar We generalize their notion to general
Shipping companies pay substantial op- optimization problems, and apply it, in
erational expenses to maintain its con- What is a greedy (approxima-
tion) algrorithm particular, to graph problems. We charac-
tainer fleet. It is important that shippers terize the best performance of algorithms
plan and schedule efficiently the move- in each model in terms of a combinatorial
ment and inventory of containers. Ow- A LLAN B ORODIN
University of Toronto game between a Solver and an Adversary.
ing to a trade imbalance, shipping lines
keywords: greedy algorithms, We prove bounds on the approxima-
accumulate a large number of unneces-
scheduling, priority algorithm, facility tion ratio achievable by such algorithms
sary empty containers in the Middle East,
location for basic graph problems such as short-
whilst some export ports such as Hong
est path, metric Steiner trees, indepen-
Kong and Japan often face a shortage of Many undergraduate algorithms courses dent set and vertex cover. For exam-
empty containers. In this paper, a well- and texts often organize the material in ple, we show no fixed priority algorithm
known container shipping line that pro- terms of “algorithmic paradigms”, such can achieve any approximation ratio for
vides scheduled sea transportation ser- as greedy algorithms, dynamic program- shortest paths (even a function of the
vices covering the Far East, the Mid- ming, local search, primal-dual, etc. We graph size). In contrast, the well-known
dle East and Europe is studied. We de- seem to be able to intuitively describe Dijkstra’s algorithm shows that an adap-
velop an optimisation model to dispatch what we have in mind when we discuss tive priority algorithm can be optimal for
empty containers from the Middle East these classes of algorithms but rarely (if this problem. We show that the known
to various export ports in the Far East re- ever) do we (or any of the standard texts) upper bound of 2 for the approximation
gion and reposition surplus empty con- attempt to define precisely what we mean ratio of an adaptive priority algorithm for
tainers from any port to shortage ports by greedy, dynmaic programming, etc. vertex cover is tight. We also give an ex-
in anticipation of future demand in sub- In a paper co-authored with Nielsen ample of a problem where memoryless
sequent periods. A set of real data from and Rackoff, we proposed a defini- algorithms are less powerful than general
one of the largest shipping companies in tion for greedy-like approximation algo- adaptive priority algorithms.
Hong Kong is used to test the efficacy rithms, called priority algorithms. We did
and robustness of the model. To enhance so in the context of classical scheduling
the practical implications of the proposed problems. Angelopoulos and Borodin Restricted models for priority
model, different logistics plans are eval- then applied the priority algorithm frame- algorithms on graphs
uated according to changes of future pol- work to the facility location and set cover
icy and situation, which are demonstrated problems. Impagliazzo and Davis have K IM S. L ARSEN
by numerical results. applied the framework to a number of University of Southern Denmark
traditional graph theoretic optimization coauthors: Allan Borodin, Joan Boyar
Conjugate direction Frank- problems. keywords: approximation algorithms,
Wolfe methods with applica- Similar to online algorithms, we want greedy algorithms, priority algorithm,
tions to the traffic assignment to derive limitations on the (approxima- graph problems
problem tion) power of priority algorithms based
only on the structure of the algorithm In the topic of priority algorithms,
M ARIA DANEVA (allowing unbounded time complexity). one explores the limits of greedy-like
MAI/Linköping University Hopefully such a study will also lead to techniques for solving approximation
coauthor: Per Olov Lindberg new algorithms. We motivate the pri- problems by establising general lower
keywords: traffic assignment problem, ority algorithm framework by discussing bounds. To capture the notion of greedy-
Frank-Wolfe method, conjugate some well known greedy algorithms and like, where irrevocable decisions must be
directions the corresponding lower bounds provable made for each input item without knowl-
We present versions of the classical within this framework. We also discuss edge of unprocessed items, very precise
Frank-Wolfe method for linearly con- some extensions of the model such as the specifications of the input format must
strained convex programs, in which con- “one-pass” framework of Erlebach and be given. We focus on graph algorithms
secutive search direction are made con- Spieksma. where this modelling aspect is particu-
jugate to each other. We apply the larly interesting. Among other things,
new algorithms to the Traffic Assignment the models must clarify if the identity
Models of greedy algorithms for of the other endpoint is revealed when
Problem and present preliminary com- graph problems
putational studies in a Matlab environ- presenting the edges of a vertex (vertex-
ment. In a limited set of computational RUSSELL I MPAGLIAZZO adjacency) or if the edges are given by
tests Conjugate direction Frank-Wolfe UCSD some edge identifier (edge-adjacency).
and Bi-conjugate Frank-Wolfe turn out In addition, we distinguish between fixed
coauthor: Sashka Davis priorities and adaptive priorities, where
to be quite efficient, outperforming pure keywords: approximation algorithms,
and “PARTANized” Frank-Wolfe meth- the algorithms may reconsider their pri-
greedy algorithm, priority algorithm, orities before each new item is chosen.
ods. Attempts to compare the new meth- graph problems
ods with the recent origin-based algo- We study unweighted vertex cover,
rithms, indicate that they are competitive Borodin, Nielsen, and Rackoff intro- independent set, and vertex coloring. For
for accuracy requirements suggested by duced the notion of priority algorithms. independent set, we obtain non-constant
Boyce et al. Priority algorithms are an abstract model results: No fixed priority algorithm can
 163


obtain an approximation ratio better than decisions. It is well known that tool C ÉSAR G UTIÉRREZ
order of 1 n  , where n is the number life is probabilistic which leads to proba- Universidad de Valladolid
of vertices. The same bound holds for bilistic processing time. Schedules made coauthors: Bienvenido Jiménez,
adaptive algorithms which accept all ver- without considering the probabilistic na- Vicente Novo
tices that they are going to accept before ture of the tool life may lead to exces- keywords: approximate solutions,
rejecting any. sive in-process tool failures, prolonged ε-Pareto solutions, proper ε-solutions,
waiting time, and delayed product deliv- ε-subdifferential
ery. For this reason, we propose an in-
FR1-321/053 PRODUCTION AND SCHEDULING

tegrated approach to solving the tool re- Kuhn-Tucker type rules are basic in op-
Scheduling with variable placement and JIT scheduling problems. timization because they describe condi-
parameters The integrated problem is formulated as tions for solutions in mathematical pro-
chair : Mohamed Haouari a probabilistic non-linear integer model. grams. Different authors have extended
Scheduling problems with con- The objective is to minimize the total these rules to approximate solutions in
trollable processing times and expected production costs. The model optimization problems.
preemption aimed to provide an optimal job sequence In convex scalar optimization, it
and the associated tool replacement inter- is possible to obtain multipliers rules
NATALIA S HAKHLEVICH vals. Due to the combinatory and proba- for approximate solutions using the ε-
University of Leeds bilistic nature of the model, optimal so- subdifferential. In nonconvex scalar op-
coauthor: Vitaly Strusevich lution cannot be obtained in polynomial timization, the general method to obtain
keywords: scheduling, controllable time. We therefore propose an efficient multipliers rules for approximate solu-
processing times, preemption tabu search approach to provide near op- tions is based on variational principles.
timal solutions. A problem with 30 op- Multiobjective optimization prob-
We consider the problem of optimal
erations and 20 tools has been solved the lems add an additional detail since in
scheduling a set of jobs on a single ma-
performance of the proposed tabu search this programs the notion of approximate
chine or on a set of parallel machines.
algorithm. efficiency is not unique. The concept
The processing time for each job may
vary within certain limits and it depends more used in the bibliography is due to
on the amount of a resource consumed. Optimal parallel machines Kutateladze (1979) and Loridan (1984).
Compressing processing times decreases scheduling with availability con- We analyze some connections be-
the completion times of the jobs but in- straints tween the notions of approximate ef-
curs additional costs. It is assumed that ficiency introduced by Kutateladze-
M OHAMED H AOUARI Loridan and Helbig (1992). As a conse-
preemption is allowed, i.e., the process- Ecole Polytechnique de Tunisie
ing of any job can be interrupted and re- quence of that we introduce a new notion
sumed later. The objective is to minimize coauthor: Anis Gharbi of proper approximate Pareto efficiency.
a function depending on job completion keywords: parallel machines We establish a characterization for proper
times (such as the makespan or max- scheduling, machine availability, release approximate Pareto solutions when the
imum lateness) together with the total date multiobjective problem is convex. We
compression cost. We describe a unified We address a generalization of the clas- provide a general method to transform a
approach that is based on reducing the sical multiprocessor scheduling problem multiobjective optimization problem into
scheduling problem to minimization of a with non-simultaneous machine avail- a scalar optimization problem in such a
linear function over a polymatroid. The ability times, release dates, and deliv- way that approximate Pareto solutions
approach is applicable to solve a range ery times. We develop new lower and for the first problem are approximate so-
of scheduling problems with controllable upper bounds as well as a branching lutions for the second problem. Finally,
processing times: with a single machine strategy which is based on an original in convex multiobjective optimization we
or several parallel machines, with equal representation of a schedule as a per- deduce multipliers rules for approximate
or different release dates, with or with- mutation of jobs. We show that em- Pareto solutions in the sense of Helbig
out deadlines, etc. We conclude with a bedding a semi-preemptive lower bound using the ε-subdifferential.
comprehensive overview of the results on based on max-flow computations in a
optimal processing jobs with controllable branch and bound algorithm yields very Second order optimality criteria
processing times when preepmtion is al- promising performance. Computational in nonsmooth vector optimiza-
lowed. experiments demonstrate that randomly tion
instances with up to 700 jobs and 20 ma-
Optimizing job sequening and chines are solved within moderate CPU
time. Moreover, the versatility of the
G IANCARLO B IGI
tool replacement decisions Dipartimento di informatica /
proposed approach is assessed through
Universita’ di Pisa
M ING L IANG its ability to solve large instances of two
University of Ottawa important particular cases P NCinc   Cmax coauthor: Davide La Torre
keywords: scheduling, modeling, tabu and P  r j q j  Cmax keywords: multicriteria optimization,
search nonsmooth analysis, second order
incremental ratio
Much research related to just-in-time FR1-321/033 MULTICRITERIA OPTIMIZATION

(JIT) production has been carried out. Efficiency analysis in Second order optimality conditions are
However, in most of the previous stud- multicriteria optimization II studied for a vector optimization problem
ies, the processing times and tool lives chair : Irina Pospelova without supposing twice differentiability.
are treated as constant and determinis- The standard approaches are component-
tic. The decisions are often made with- On a new concept of proper ap- wise since they try to employ nonsmooth
out considering detailed tool replacement proximate Pareto efficiency analysis for real-valued functions also in
164 Parallel Sessions: Friday 9:00 – 10:30 (FR1)

this framework, relying directly on com- of the normal form games and properties of observed input-output data, ii) an ex-
ponents or on suitable weighted sums. of the games with fixed order of moves. pansion of this convex hull justified by
Recently, a global approach based on the a maintained hypothesis of (strong) dis-
well-known concept of generalized Hes- FR1-321/133 LINEAR PROGRAMMING
posability of inputs and outputs, even-
sian has been investigated too. In this pa- Mathematical programming tually followed by iii) a radial scaling
per we consider a different kind of global techniques in data of input-output configurations in the ex-
approach: set-valued directional deriva- envelopment analysis panded convex hull. The faces and facets
tives of vector-valued functions are intro- of the resulting polyhedral production
duced relying on Kuratowski type limit organizer/chair : Niels Christian possibility set define the efficient produc-
of vector second order parabolic ratios. Petersen tion frontier. The paper demonstrates the
We develop second order necessary and A new algorithm for DEA use of some of the available (exponen-
sufficient optimality conditions for vector tial) convex hull algorithmn for an ex-
problems subject to either abstract con- N IELS C HRISTIAN P ETERSEN plicit identification of the facial structure
straints or inequality and equality con- University of Southern Denmark of the efficient frontier in fairly large em-
straints and we show that this approach coauthor: Jose Dula pitical DEA data sets, An outline of the
gives sharper results than the previous keywords: DEA, computational use of this type of information in a pro-
ones. geometry, polyhedral sets ductivity analysis is given.
Data Envelopment Analysis (DEA) is es-
Existence conditions of game so- sentially a clustering tool for data sets
lution in multicriteria case that identifies geometric outliers. Many How to bootstrap DEA effi-
of the insights about the geometry of ciency scores by using the slice
I RINA P OSPELOVA polyhedral sets from convex analysis and model approach
Moscow State University computational geometry are available for
coauthors: Natalia N. Novikova, DEA. These insights and other computa-
Elizabeth M. Kreines tional contributions have been used to de- TOR B ELTOV
keywords: game theory, multicriteria sign a new algorithm for DEA. This al- Department of Organization and
game, vector payoff function gorithm is faster than the traditional ap- Management, SDU
We investigate games with opposite pur- proach. In addition, its results provide keywords: DEA, slice models,
poses and vector payoff function. The additional useful information that can be bootstrap
presence of vector criteria generates am- used to accelerate complex analyses as
biguity in interpretation of opposite pur- well as eschew some of the complications
poses. In the case of all partial criteria of inherent to some of the more standard ap- Calculating Data Envelopment Analysis
one player are opposite to all partial cri- proaches. We present the new algorithm (DEA) Efficiency Scores can be a very
and report on computational results. time consuming task. The time involved
teria of the other we come to the game is especially high when standard mod-
that is conventionally called vector zero- elling systems like GAMS are used. One
sum game. Here one player maximizes Identification and use of effi- approach to deal with this issue is to use
the payoff function and the other mini- cient faces and facets in DEA a callable library like the one supplied by
mizes it. CPLEX. Another approach is to observe
In the case of opposite goals of the O LE B ENT O LESEN
Department of Organization and that a complete DEA model in fact con-
players, we come to the game conven- sists of multiple linear programs, each of
tionally called antagonistic (following D. Management, University of Southern
Denmark which has the same structure but differ-
Blackwell, G. Jentzsch and R.J. Au- ent data. It is known that all DEA mod-
mann). Here if the goal of one player is to coauthor: Niels Christian Petersen
els fits the definition of the socalled slice
maximize the vector payoff function then keywords: data envelopment analysis,
model approach. Thinking DEA mod-
the goal of the other is to minimize an efficiency analysis, polyhedral convex
els as slice models reduces the time in-
arbitrary partial criterion, which is non- analysis
volved for solving the problems signifi-
fixed before. Data Envelopment Analysis (DEA) can cantly. This issue is prominent when we
We analyze the problem of determi- be seen as a mathematical programming turn to an estimation of confidence inter-
nation of the described games’ values. technique designed for an identification vals for DEA Efficiency Scores by boot-
The possibility of generalization of the of the efficient production frontier in an strapping, since this procedure involves
concepts known for scalar games to vec- empirical production data set. The em- the solution of an LP each DMU more
tor games is investigated. In particu- pirical production possibility set is con- than 500 times. The paper considers how
lar, we consider saddle points and mixed structed in a three step procedure involv- to use the slice model approach for boot-
strategies; solution existence conditions ing i) the identification of the convex hull strapping in DEA.
 165

Index of Keywords

0 1  -matrices, 44 augmentation algorithm, 114
f augmented Lagrangian, 61
E0 matrix, 92
ε-Pareto solutions, 163 augmented Lagrangian functions, 74
ε-subdifferential, 163 augmented Lagrangian method, 75, 155
0/1-integer programming, 84 augmented Lagrangians, 61
1-median, 42 automated, 83
a posteriori error estimates, 141 automatic differentiation, 102, 116
a posteriori error estimation, 60 automation server, 116
AARC, 124 average case complexity, 150
abnormal point, 159 average complexity, 67
ACCPM, 50, 76 average-cost, 159
active constraints, 61 b-matching, 127
active set, 35, 88 Banach spaces, 53
active-set approach, 41 bandwidth allocation, 46
acyclic orientation, 125 barrier function, 134
adaptive problem approximation, 123 barrier methods, 133
adaptive termination rule, 156 barrier-penalty algorithm, 43
adjoint equations, 87 basin hopping, 77
aggregate production planning, 122 basis reduction, 34
aggregate traffic uncertainty, 80 batch processing machines, 109
aggregated interindustry model, 40 beam orientation, 43
aggregation, 160 behavioural econometric models, 39
AHP, 95 berth allocation, 80
air traffic management, 95 biconnected graphs, 72
aircraft rotation problem, 51 bidding strategies, 39
aircraft schedule recovery, 64 bidding/auctions, 106
airline, 38 bilevel optimization, 96, 128
airline applications, 63 bilevel programming, 89, 90
airline operations, 63, 64 biobjective integer program, 110
airline operations research, 38 bioinformatics, 63
airline scheduling, 63, 95 bipartite, 113
algebraic geometry, 131 bipartite induced subgraph, 122
algebraic optimization, 156 block diagonal structure, 62
algorithms, 37, 45, 55, 85, 89, 93, 113 bootstrap, 65, 164
almost N̄ matrix, 92 bound-constrained minimization, 160
alternative polyhedron, 152 box region, 131
ambulance service, 93 boxicity of a graph, 32
American contingent claims, 105 branch-and-bound, 34, 37, 51, 55, 59, 68, 73, 109,
analysis of algorithms, 58, 72, 84, 149 126, 131, 152, 153
analytic center, 36, 50, 89, 127 branch-and-cut, 33, 43, 51, 55, 58, 73, 86, 99, 112,
antimatroid, 56 122, 127, 135, 154
applications, 47, 92, 98, 139, 141, 150 branch-and-cut-and-price, 47, 100
applications of optimization, 67 branch-and-price, 33, 50, 59, 120, 140
approximability results, 72 branching variable selection, 153
approximate equilibrium points, 132 breast cancer, 81
approximate saddlepoints, 132 Bregman distances, 66
approximate solutions, 163 Bregman function, 79
approximation, 113, 136, 146, 149 budget, 148
approximation algorithms, 59, 71, 76, 84, 85, 99, 108, budget modeling, 39
109, 112, 122, 123, 135, 149, 152, 162 bundle methods, 36, 55, 117
approximation management, 87 bus crew scheduling, 74
approximation schemes, 68, 71 business applications, 106
ARAT game, 92 cancer treatment, 81
arbitrage, 105 capacitated tree problem, 47
arbitrage theory, 105 categorical variables, 34
asset liability management, 78 chaos theory, 94
asset pricing, 105 chemical informatics, 63
assignment, 126 chemotherapy, 81
assignment model, 57 Chinese postman problem, 99
assignment problems, 60 Cholesky factorization, 124
assisted, 83 chordal graph, 93
auctions, 52, 106 Chvatal-Gomory cuts, 34
166 Index of Keywords

circle packing, 152 conjecture, 82, 83

circulant graphs, 72 conjecture testing, 82
classification, 81, 113 conjugate directions, 162
claw-free graphs, 72 conjugate gradient, 116
clique family inequality, 72 conjugate gradient method, 129, 142
closed-set system, 56 conjugate sets, 35
cluster computing, 81 connectedness, 105
clustering spanning trees, 85 connectivity, 122
clustering TSP path, 85 constrained NLPs, 104
co-evolution, 97 constrained optimization, 130
cocoercivity, 158 constrained programming, 41
coding, 99 constrained shortest path, 82
coercive regularizations, 79 constraint identification, 120
coercivity conditions, 105 constraint programming, 140, 154
coherence, 91 constraint qualifications, 104, 159
coherent risk measure, 110 constraints degeneracy, 75
collusion, 97 contingency table, 98
column generation, 33, 46, 47, 51, 55, 59, 73, 76, 94, continuous linear programming, 71, 137
120, 122, 126, 135, 147 contraceptive logistics, 94
column-and-row generation, 46 control, 119
combinatorial algorithm, 45 control constraints, 141
combinatorial auction, 119 control of PDEs, 141
combinatorial auctions, 106, 120 control system, 133
combinatorial optimization, 31, 40, 42, 45, 46, 51, 58, controllable processing times, 163
59, 72, 84–86, 98, 99, 114, 130, 135, 139, 140, 147, convergence, 120
149, 150 convergence analysis, 153
combinatorial problems, 36 convex, 116, 127
combinatorics, 32, 59, 69, 70, 98, 140 convex analysis, 106
communications, 121, 125 convex approximations, 141
competitive analysis, 92 convex constrained, 142
competitive electricity market, 39 convex cooperative games, 83
competitive price, 110 convex duality, 105
complement, 57 convex function, 31
complementarity, 53, 76, 102, 147 convex geometry, 56
complementarity constraints, 106, 107, 120, 133 convex geometry (antimatroid), 56
complementarity problem, 106, 160 convex hull, 100, 150
complementary cone, 40 convex mappings, 90
complemetarity problems, 160 convex maximization, 158
complete linear description, 56 convex minimizaton, 143
complexity, 36, 54, 67, 89, 109, 110, 140 convex optimization, 67, 89, 121, 156, 161
computation, 90 convex optimization problem, 66
computational complexity, 67, 69, 89, 96, 98, 99, 101, convex polynomial mappings, 90
136, 142, 146 convex programming, 62, 89, 132, 136, 143
computational efficiency, 62 convex quadratic constraints, 116
computational experience, 149 convex quadratic programming, 105
computational geometry, 48, 63, 164 convex relaxation, 118
computational grid, 51 convex sets, 105
computational linear algebra, 142 convex underestimators, 104
computer aided design, 32 convexification, 118
computer-assisted proofs, 152 convexity, 31, 32, 37, 72, 118, 146
computers in research, 82 cooperative, 83
computing equilibria, 70 cooperative games, 110
condition number, 89, 96, 153, 161 core, 110
cone, 159 core allocation, 110
cone preserving transformation, 40 core solutions, 83
cone-constrained optimization, 61 correlated knapsack problems, 34
cone-constrained problem, 159 cost analysis, 115
cones of moments, 130 costly constraints, 128
cones of semi-metrics, 152 Coulomb friction, 160
congestion pricing, 108 counting, 113
conic duality, 144 counting problem, 98
conic optimization, 116, 161 coupled task problem, 58
conic program, 76 crane scheduling, 149
conic programming, 89, 116, 161 crew pairing, 38
conic systems, 143 crew recovery scheduling, 51
 167

crew scheduling, 38, 122 distance to ill-posedness, 145

crossover events, 96 distance to infeasibility, 143
cut and column generation, 76 distance to insolvability, 124
cut and branch, 126 distributed algorithms, 37
cut metrics, 152 distributed computing, 37
cutting and packing, 152 distribution planning, 162
cutting models, 157 distribution systems, 108
cutting plane algorithm, 47 districting, 52
cutting plane methods, 36, 50, 76, 89, 143 domination analysis, 84
cutting plane/facet, 32, 55, 154 double-tree algorithm, 135
cutting planes, 32–34, 46, 47, 58, 71, 86, 89, 100, 127 downside risk, 65
cutting stock problem, 59 dual active set algorithm, 150
CVRP, 58 dual cone, 143
cycle cancelling algorithm, 69 dual method, 144
cyclic group, 86 dual simplex method, 150
cyclic rostering, 73 dual to the hyperbolicity cone, 50
D.C. optimization, 144, 158 duality, 43, 90, 104, 106, 110, 136
Dantzig-Wolfe, 100 duality gap, 83
Dantzig-Wolfe decomposition, 33, 100 due date assignment, 136
data analysis, 143 dynamic flows, 71
data classification, 76 dynamic markets, 91
data envelopment analysis, 136, 164 dynamic pricing, 67
data mining, 91 dynamic programming, 41, 64, 85, 86, 95, 96, 113,
DC programming, 77, 144 115, 119, 132, 160
DC programming/DCA, 144 dynamic scheduling policy, 146
DCA/GNC, 144 dynamic systems, 156
DEA, 53, 164 dynamics, 119
decision analysis, 123 earliness/tardiness, 138
decision support systems, 91, 93 Eckart and Young identity, 143
decision theory, 131 economics of scale, 54
decomposition, 42, 62, 73, 91, 106, 132, 140 edge-connectivity, 152
decomposition algorithms, 124 edge-disjoint paths, 125
decomposition methods, 63 Edgeworth equilibrium, 78
degeneracy, 101 EEG time series, 94
dense graph, 115 efficiency, 105
density theorem, 150 efficiency analysis, 164
density/regression estimation, 159 efficient, 159
derivative free optimization, 35 efficient algorithm, 70
derivative pricing, 52 eigenvalue optimization, 157
derivative-free algorithm, 141 eigenvector constraints, 115
derivative-free optimization, 48, 128 eigenvector method, 95
design, 74 elastic variables, 61
design of experiments, 60, 101 electric power, 64
design optimization, 48 electricity market, 96, 97, 148
deterministic, 65, 160 elementary symmetric functions, 50
deterministic sequencing, 55 ellipsoid method, 70
DFO, 141 elusiveness, 44
dial-a-ride, 135 empirical process, 65
diameter, 45 empty container, 162
diameter-constrained trees, 139 empty distribution, 42
differential equations, 48 enclosure method, 92
differential systems, 106 engineering, 72, 74, 155
directed acyclic graphs, 104 engineering optimization, 60
discount function, 65 entropy, 121
discrete convex analysis, 57 entropy function, 75
discrete convex function, 45 enumeration, 113
discrete location, 42, 72 EPEC, 120
discrete optimization, 92, 117 equilibria, 96
discrete time LQR, 86 equilibrium, 78, 79, 118
discrete tomography, 125 equilibrium models, 108
discretization, 64, 133 equilibrium problems, 53, 105, 131
discretization concept, 141 equilibrium programming, 70
disjunctive programming, 118 error bounds, 130, 157
disruption, 123 estimation, 105
distance cut off, 154 estimation from experiments, 50
168 Index of Keywords

Euclidean Jordan algebra, 40 generalized convexity, 77, 90, 131, 145

Euler equations, 87 generalized derivatives, 75
Eulerian graphs, 110 generalized min-max, 74
evolutionary algorithms, 94, 104 generalized monotonicity, 105, 118
exact augmented Lagrangians, 75 generalized networks, 125
exact characterization, 144 genetic algorithm, 97, 122
exact solution, 46 geometric network design, 139
exchange algorithms, 103 geometry optimization, 113
existence, 159 Gibbs energy, 144
expected utility, 91 global convergence, 36, 49, 102, 107, 155
expensive simulation, 39 global minimizer, 131
extended cover inequalities, 126 global optimization, 48, 51, 63, 64, 77, 89, 103, 104,
extended formulations, 42, 57, 68 110, 117, 128, 131, 144, 155
extremal convolution, 83 global routing, 45
extreme point, 129 GOA GBD, 86
face, 40 gradient estimation, 141
facets, 32, 34, 45, 55, 73, 112, 113 gradient methods, 36
facets of polyhedra, 112 gradient-projection method, 156
facility location, 42, 51, 54, 108, 162 graph algorithms, 58
fair adversary, 107 graph disconnectivity, 138
fairness, 40 graph editing, 82
fault tolerance, 37 graph partitioning, 122
FB function, 147 graph problems, 162
feasible subsystem polytope, 152 graph property testing, 44
feed mill, 140 graph theory, 44, 82, 98, 126
filter, 36, 49 graph-coloring, 73, 98, 112, 152
filter algorithm, 36, 88 graphs, 32, 82–84, 95, 99, 112, 113, 123
filter methods, 36, 49 greedy, 152
filter ods, 34 greedy algorithm, 162
financial engineering, 129 greedy algorithms, 84, 162
finite convergence, 147 grid computing, 81
finite element model, 115 grid graphs, 45
finite elements, 141, 159 group, 152
first order conditions, 104 group relaxations, 47
fix and relax coordination, 38 group Steiner minimal tree, 57
fixed cost allocation, 40 Guignard CQ, 104
fixed dimension, 140 Hadamard manifolds, 158
fixed point, 34, 118 Hausdorff (semi)continuity, 90
flexible manufacturing, 55 health care - treatment, 94
flow algorithms, 45, 71, 108 Helmholtz equations, 87
flowshop, 109 Hessian, 157
fluid models, 71, 108 heuristic algorithm, 74
flux lines, 129 heuristics, 52, 55, 68, 72, 80, 84, 98, 112, 113, 125,
forecast models, 91 126, 147, 150, 153
forest harvesting, 46 hierarchical clustering, 123
forestry, 134 Hoffman type estimates, 157
forward rate function, 65 homogeneous model, 134
Fourier Motzkin method, 136 hop-constrained walks, 56
fractional programming, 63 hop-constraints, 125
Frank-Wolfe method, 162 hop-indexed models, 139
free boundary problem, 92 hub location problem, 47
frequency assignment, 32 hybrid algorithms, 156
fundamental cycle basis, 72 hybrid methods, 117, 154
Gabor frames, 155 hydroelectric reservoirs, 96
Gale-Ryser algorithm, 44 hydrology, 115
Gallai-Edmonds, 45 hyperbolic polynomials, 50
game dynamics, 77 hypercube, 126
game programming, 70 ill-conditioning, 35
game theory, 40, 147, 164 implicit filtering, 115
Gauss-Newton, 88 IMRT, 43
Gauss-Newton algorithm, 61 independence systems, 57, 84
general variational inequality, 158 independent sets, 32, 115
generalized convexity, 132 index rules, 147
generalized assignment, 126 industrial application, 92
generalized assignment problem, 126 inequality relaxation, 107
 169

inexact line searches, 143 Lagrangian relaxation, 33, 36, 42, 50, 51, 76, 95, 100,
inexact solutions, 61 109, 122, 130
infeasible linear systems, 152, 153 Lagrangian transformation, 104
inference, 50 Lanczos method, 157
infinite horizon, 96 land use, 78
information security, 119 large neighborhood, 148
input format, 90 large scale, 62, 129
input selection, 91 large scale linear programming, 124
integer, 55, 85 large scale optimization, 43, 61, 62, 92, 117, 129, 130,
integer duality, 64 142
integer formulation, 99 large scale problems, 42, 75
integer multicommodity flow, 127 large scale SDP, 157
integer point, 114 large sparse problems, 107
integer programming, 32–34, 41, 46, 47, 52–54, 59, large sparse systems, 93
60, 72, 85, 86, 91, 95, 98–100, 112, 114, 120, 121, large update, 134
125, 127, 130, 131, 133, 134, 139, 140, 154 large update methods, 148
integer programming duality, 40 lattice basis reduction, 140
integer rounding property, 84 lattice reduction, 114
integral bases, 114 least square values, 88
integrality gap, 127 least squares, 87
integration quadratures, 64 leg base rostering, 51
interactive methods, 66 lexicographic multiobjective, 100
interior method, 102 lexicographic ordering, 59
interior point, 116, 133 lift-and-project, 34, 86
interior point algorithms, 67, 96, 110, 121 lifting, 114
interior point method, 36, 49, 54, 61, 62, 66, 67, 80, limited memory, 130
81, 102, 107, 108, 110, 116, 120, 121, 129, 133, 134, limited memory methods, 117
146–148, 156, 161 line balancing, 58
internet, 66 line stabbing number, 112
internet protocol, 80 linear complementarity problem, 92, 147
interpolating surrogates, 155 linear complementary, 70
intersection cuts, 153, 154 linear equations, 107
interval analysis, 152 linear optimization, 134
interval arithmetic, 146 linear ordering problem, 112
interval coloring, 152 linear programming, 41, 49, 53, 55, 65–67, 69, 70, 85,
interval graphs, 32, 138 96, 100, 107, 108, 121, 127, 136, 137, 147, 150, 161
interval orders, 131 linear programming algorithms, 150
interval selection, 131 linear programming problem, 147
inventory, 148 linear programming relaxations, 59, 72, 146
inventory/production, 68, 149 linear relaxation, 140
inverse function, 159 linearization, 33
inverse optimization, 42, 85 Lipschitz stability, 39
inverse problems, 88 list coloring, 95
investment, 119 load balance, 37
investment under uncertainty, 65 local search, 48
IP-over-optical network, 133 local searches, 77
iterative algorithms, 87 localization, 155
iterative methods, 32 location, 62, 130
k-convexity, 136 location problem, 42, 54
k-cycle polytope, 45 locational decisions, 54
kernel, 110 logical constraints, 86
kernel estimation, 159 logistics and transportation, 42
kernel function, 54 look-ahead, 93
knapsack, 85, 127 lot sizing, 46, 68, 114
knapsack problem, 85 lower bound, 57
knapsack sets, 87 machine availability, 163
Kriging, 141 machine groups, 149
L-BFGS method, 142 machine scheduling, 149, 154
L1 penalty, 88, 133 machinery, 58
Lagrange interpolation, 141 maintenance routing, 63
Lagrange multiplier, 103, 128 majorization, 44
Lagrangean relaxation, 50, 71, 73, 82, 109, 157 makespan, 109
Lagrangean techniques, 100 manpower planning, 59
Lagrangian duality, 86, 155 marginal functions, 118
Lagrangian heuristics, 130 market distribution function, 96
170 Index of Keywords

Markov chain, 98 monotone mapping, 159

Markov decision problem, 96 monotone operators, 79
Markov decision processes, 64, 146, 147 monotone vector fields, 158
Markov model, 144 monotonicity, 118
Markov: finite state, 160 Monte Carlo simulation, 145
matchings, 92, 112 Mosco convergence of sets, 66
material optimization, 48 MPEC, 60, 75, 90, 102, 108, 120, 133
mathematics, 37 MPECs, 104
MATLAB, 53 multi depot vehicle routing, 73
matrices, 37, 125 multi objective optimization, 66
matrix calibration, 155 multi-agent systems, 119
matrix completion, 82, 93 multi-level feedback algorithm, 107
matrix game, 92 multi-rigid-body contact prob., 160
matroid, 31, 56 multi-stage, 91
matroid intersection, 57, 109, 152 multi-stage decision problems, 117
MAX FS, 49, 153 multicommodity, 71, 127
max-algebra, 112 multicommodity flow, 45, 56, 69, 80, 84
max-flow min-cut property, 56 multicommodity network flow, 42
max-min functions, 103 multicriteria decision support, 123
maximal clique, 113 multicriteria game, 164
maximal flow, 144 multicriteria optimization, 110, 163
maximal monotone operator, 66 multicut, 84
maximum acyclic subgraph, 112 multidimensional assignment, 152
maximum capacity path, 56 multidisciplinary optimization, 87
maximum clique, 115 multilevel methods, 116
maximum flow, 69 multilevel systems, 48
maximum flow time, 107 multilinear envelopes, 118
maxinf problems, 118 multiobjective fractional, 136
mean, 50 multiobjective optimization, 96
mean-risk, 65 multiobjective programming, 95
mean-risk models, 132 multiple coalitions, 83
measure of proximity, 159 multiple criteria, 123, 150
mechanism design, 113 multiple criteria programming, 66, 150
metaheuristics, 39, 62, 68, 117, 123 multiple machines, 109
methodology, 149 multiple maxima, 34
metric regularity, 133, 145 multiple scales, 60
MFCQ, 75 multiplier methods, 79
MILP, 140 multivariate interpolation, 128
min-cut, 82 Nash equilibrium, 70, 83, 110
minimal ellipsoid, 63 natural gas, 91, 98
minimax problems, 34 Navier-Stokes equations, 74
minimization, 31 NCP, 120
minimization the sets function, 159 network application, 42
minimum maximal flow, 77 network capacity expansion, 69
minimum spanning trees, 76, 135 network design, 33, 42–44, 71, 80, 82, 120
minimum-risk problem, 95 network flow problems, 161
Minty variational inequality, 145 network flows, 139
mixed graphs, 99 network interdiction, 52
mixed integer, 65 network operation, 98
mixed integer Gomory cuts, 154 network planning, 47
mixed integer nonlinear prog., 63 network synthesis, 84
mixed integer programming, 34, 37, 42, 47, 60, 80– network/graphs: multicommodity, 71
82, 86, 100, 114, 118, 127, 132, 134, 139, 140, 153, networks, 82, 90, 136, 139
154, 161 networks/graphs, 32, 57, 69, 112, 125
mixed integer recourse, 52 networks/graphs applications, 121, 125
mixed integer rounding, 68, 100 networks/graphs:flow algorithm, 69, 71
mixed interior/exterior, 88 networks/graphs:multicommodity, 69
mobile communications, 147 networks: tree algorithms, 72
mod-k, 58 Newton method, 75, 116, 129, 142, 147, 158
model robustness, 70 Newton-like methods, 62, 143
modeling, 68, 134, 163 night duty, 43
modeling languages, 41, 79 NLO, 75
modeling system, 41, 53, 103 NLP, 102
modelling, 66 no-wait flowshop, 109
momentum, 160, 161 node partitioning, 99
 171

noisy function, 141 optimal level solutions, 158

noisy function values, 48 optimal matching, 110
noisy observations, 48 optimal portfolio selection, 91
noisy problems, 34 optimal set partitioning, 159
non convex constrained problem, 158 optimal stopping problem, 65
non convex separation, 79 optimal supply function, 148
non deterministic, 65 optimality, 136
non differentiable programming, 130 optimality condition, 103, 146
non linear problems, 40 optimization, 38, 60, 61, 67, 74, 93, 121, 127, 145
non linear programming, 154, 155 optimization method, 101
non Lipschitz, 103 optimization on efficient set, 77
non parametric estimation, 96 optimization problem, 81
non smooth, 127 optimization software, 53, 116
non smooth optimization, 36, 93 option prices, 105
nonconvex, 86, 103 options prices, 53
nonconvex optimization, 63 OR networks, 136
nonconvex programming, 118, 144 oracle, 159
nonconvex variational problems, 159 oracle methods, 70
noncooperative game, 56 orthogonal latin squares, 154
nondegeneracy, 101 orthogonal projection, 142
nondifferentiable optimization, 39, 89, 143 out-of-sample tests, 78
nondifferentiable programming, 117 p-median, 46
nonlinear equations, 34, 42 P-property, 40
nonlinear least squares, 88 packing, 56, 149
nonlinear operator, 133 paging, 93
nonlinear optimization, 35, 43, 49, 61, 92, 101, 102, pairing/rostering integration, 51
115, 128, 141, 143, 156, 157 pairwise comparison, 131
nonlinear programming, 34–36, 41, 49, 53, 54, 62, 63, parallel algorithm, 37, 51
66, 75, 81, 87, 88, 93, 98, 101, 102, 115, 116, 118, parallel computation, 81
120, 133, 141, 155, 157, 160 parallel computing, 51, 62, 74, 82, 129, 156, 157
nonlinear programming theory, 101 parallel cutting plane, 73
nonlinear semidefinite prog., 155 parallel machines, 46
nonmonotone line search, 142 parallel machines scheduling, 163
nonmonotonicity, 88 parallel programming, 62, 124
nonnegative polynomials, 156 parameter estimation, 50
nonsmooth, 103 parametric analysis, 69
nonsmooth analysis, 143, 163 parametric linear programming, 146
nonsmooth functions, 53 parametric LP, 57
nonsmooth mechanics, 106 parametric optimization, 159
nonsmooth optimization, 36, 76, 103, 157 parametric programming, 112
nonsmooth programming, 143 parametrized policies, 78
NP-complete, 98 Pareto domain representation, 123
NP-hardness, 112 Pareto optimum, 79
nucleolus, 110 partial differential equations, 60
numerical algorithms, 40, 137 partly smooth, 103
numerical experimentation, 150 path following algorithms, 50
numerical methods, 60 path tracing, 34
nurse scheduling, 43 path width, 126
objective level cut, 86 path-matching, 45
obstacle problem, 159, 160 pattern analysis, 143
obstacles, 109 pattern recognition, 76
on-line algorithms, 92, 93, 107 pattern search, 34, 36
on-line dial-a-ride problem, 107 patterns, 98
on-line optimization, 47, 92 PDE-constrained optimization, 81, 87
on-line problem, 85 PDE-constraints, 128
on-line scheduling, 55 penalty methods, 36
one machine scheduling, 138 penalty/barrier methods, 75
one way contracts, 148 perfect graphs, 44
open algorithm, 53 perfect matching, 112
open shop, 149 perfect solution, 70
open source, 127 perfectness, 53
operations research, 148 performance analysis, 116
optimal control, 101, 115, 116, 128, 133, 141 periodic solutions, 115
optimal control of PDEs, 128 periodic vehicle routing, 73
optimal design, 115 permutations, 32
172 Index of Keywords

personnel scheduling, 38 production games, 83

perturbation tecnhiques, 115 production planning, 114, 119, 140
perturbed convex functions, 132 programming, 85
perturbed nonlinear programs, 75 programming involving graphs, 44
petroleum/natural gas, 139 progressive addition lenses, 115
phase equilibrium, 133 project scheduling, 32
phase transition, 98 projected gradient, 142
phase-1, 150 projected gradient method, 136
piecewise marginal costs, 148 projection, 79
pivoting, 48 projection method, 158
pivoting algorithm, 41 projection theorem, 56
planar graphs, 58, 86 proper ε-solutions, 163
pointwise well-posedness, 144 proper solution, 70
political redistricting, 46 properly efficient point, 150
polyhedra, 32, 150 properness, 78
polyhedral combinatorics, 45, 46, 154 protein folding, 49
polyhedral convex analysis, 164 prox method, 104
polyhedral homotopy cont., 34 prox-regularity, 103
polyhedral methods, 147 proximal algorithm, 61
polyhedral sets, 164 proximal analytic center, 51
polyhedral techniques, 33 proximal methods, 36, 55, 93
polyhedrally tight functions, 31 proximal point, 79
polyhedron, 86, 125 proximal point algorithm, 158
polynomial complexity, 148 proximal point methods, 66
polynomial graph algorithms, 85 proximity function, 54
polynomial optimization, 148 pseudoconvexity, 77
polynomial system, 34 pseudoflow, 69
polynomial time, 71, 156 pseudomonotonic programming, 95
polynomial time algorithms, 58 public transportation, 59, 74, 121
polytime algorithm, 69 q-convergence, 62
polytope, 47, 72, 114, 122, 140 q-quadratic convergence, 129
polytopic systems, 132 QCQP, 117
portfolio, 123 quadratic assignment problem, 98
portfolio optimization, 65 quadratic cone, 116
portfolio planning, 92 quadratic constraints, 156
positive polynomials, 130 quadratic convergence, 101
positive semi-definite matrix, 152 quadratic knapsack problem, 85
postman problems, 99 quadratic optimization, 158
power generation, 34 quadratic penalty function, 142
precedence constraints, 136 quadratic programming, 32, 33, 54, 86, 92, 94, 101,
preconditioners, 157 147, 158
preconditioning, 116, 129 quadratic spline, 158
predator-prey interaction mode, 115 quadratically constrained QP, 131
predictor-corrector methods, 142 quantum chemistry, 49
preemption, 163 quasi-Newton methods, 130, 160
preprocessing, 93 quasi-variational inequalities, 144
price of anarchy, 56 quasidifferentiable functions, 77
pricing, 67, 90 quasimonotone operator, 118
primal aggregation, 157 r-linear convergence, 142
primal methods, 114 radial basis functions, 141
primal-dual, 35, 36 radiation therapy, 81, 94
primal-dual active set method, 141 radiation therapy optimization, 43
primal-dual algorithms, 54, 56, 71, 96 radiation therapy planning, 43
primal-dual interior method, 88 radio network planning, 160, 161
primal-dual interior point alg, 54 railway, 38
primal-dual method, 116, 134 railway systems, 121
principal component analysis, 145 railway transportation, 42, 122
priority algorithm, 162 random, 140
prize collecting, 71, 126 randomization, 113
prize-collecting Steiner trees, 71 randomized approximation, 158
probabilistic analysis, 107, 135 randomized relaxation methods, 153
probabilistic constraints, 39 rational functions, 131
probability distributions, 78 reaction diffusion equations, 48
process control, 92 real algebraic geometry, 156
process engineering, 133 real options, 105
 173

real world application, 51, 148 second order derivatives, 143

rectangular partition, 100 second order incremental ratio, 163
rectilinear Steiner trees, 109 second order sufficiency, 75
reduced gradient, 156 segmentation, 113
reduced-basis, 60 seismic tomography, 61
redundant solutions, 60 seizure prediction, 94
reformulation, 102 semi continuous constraints, 154
regression function, 39 semi infinite optimization, 74, 146
regularity, 128 semi infinite programming, 130, 145
regularity and nondegeneracy, 107 semi linear spaces, 136
regularization, 35, 79, 88 semidefinite complementarity, 40
regularization method, 66 semidefinite cone, 40
relative worst order ratio, 93 semidefinite cut, 89
relax and cut, 100 semidefinite LCP, 40
relaxation, 32, 57, 100, 147 semidefinite optimization, 155
relaxation hierarchies, 131 semidefinite problem, 80
relaxation methods, 153 semidefinite program, 101, 148
relaxation/subgradient, 85, 98 semidefinite programming, 41, 43, 49, 54, 76, 80–82,
release date, 163 89, 93, 110, 112, 130, 131, 138, 144, 155–157, 161
resource constraints, 32 semidefinite quadratic progr., 50
resource management, 147 semidefinite relaxation, 138
response surface, 101, 128, 141 semismooth, 76
revenue management, 67 semismoothness, 147
Riesz-Kantorovich formula, 78 sensitivity, 103
ring networks, 127 sensitivity analysis, 115, 136, 159
rings, 84 separable programming, 89
risk, 106, 119, 132 separation, 140
risk management, 41 separation theorem, 31
risk measure, 91 separator, 79
risk neutral pdf, 105 sequencing and scheduling, 38
road investments, 134 sequential quadratic progr., 35
robust implementation, 102 sequentially programming, 155
robust optimization, 39, 67, 117, 124, 132, 135, 143, servers, 66
162 set constraints, 90
robust scheduling, 63 set cover, 82
robustness, 101 set covering, 73, 147
rolling stock circulation, 121 set covering problems, 130
root-cutset constraints, 47 set packing and partitioning, 46
rooted circuits, 56 set partitioning, 33, 73
round robin tournament, 95 set valued analysis, 132
rounding heuristic, 118 set valued optimization, 136
routing, 59, 80, 126 several minima, 87
row-column scaling, 129 shape optimization, 87, 115
saddle points, 49 Shapley value, 88
sample allocation, 110 shellability, 125
sampling, 98 shift schedules, 149
scalarization, 150 ship design, 87
scaling, 69 ship routing, 54, 135, 148
scenario generation, 78, 145 ship scheduling, 148
scenario tree, 119 shortest path routing, 82
scenario tree generation, 78, 145 shortest paths, 51, 52, 82
scenarios, 145 simple decomposition, 129
scheduling, 46, 58, 68, 109, 120, 123, 134, 136, 140, simplex, 137
147, 149, 162, 163 simplex method, 37, 161
scheduling problems, 149 simplicial approach, 114
scheduling theory, 59 simplicial complex, 125
SDP, 112 simulation based optimization, 39, 87
SDP relaxation, 76 single machine, 123, 136
SDRs, 56 single node flow sets, 114
search and surveillance, 55 single period, 148
second order cone, 89 slack balancing, 31
second order cone optimization, 161 slice models, 164
second order cone problem, 80 sliding modes, 133
second order cone programming, 41, 76, 80, 144, 161 SLP, 35
second order cones, 80 smoothed analysis, 121
174 Index of Keywords

smoothed competitive analysis, 107 structural optimization, 36, 41, 48, 90

smoothing, 90, 130 structure matrix, 44
smoothing method, 77 structure of information, 119
SOCP relaxation, 76 structure theorem, 45
software, 66, 127, 128, 157 structured stochastic game, 92
software development, 149 subgradient bundle methods, 157
software framework, 37, 51 subgradient methods, 76
solving linear constraints, 137 subgraph, 126
space mapping, 128, 154, 155 submodular, 57
space mapping optimization, 155 submodular flow, 45
spanning trees, 139 submodular function, 31, 45
sparse jacobians, 98 subproblem centric approach, 43
sparse matrices, 143 subset selection, 48
sparsity, 82, 110, 148 substitute, 57
spatial optimization, 104 subtour elimination constraint, 47
special ordered sets, 126 subtransformations, 40
spectral factorization, 156 sum of squares, 148, 156
spectral functions, 157 sum of squares of polynomials, 138
spectral gradient, 142 sum-min function, 103
spectral radius, 77 superlinear convergence, 49, 101
spectrum auctions, 120 supply chain, 94, 108, 148
sphere, 72 supply chain planning, 41
split cuts, 154 supply chain management, 67, 119, 122
splitting methods, 79 supply chain optimization, 38, 148
sports scheduling, 95, 140 support vector machines, 49, 81, 113
sports timetableing, 95 surrogate model, 74
spot rate function, 65 surrogate modeling, 141, 154, 155
SQP, 49, 61, 88, 129 survivability, 133
stability, 53, 77, 106, 145 survivable network, 125
stability likelihood, 70 SVD, 147
stability number, 44 switching control game, 92
stable marriage model, 57 symbolic computation, 156
stable set, 69 symmetric cone programming, 80
stable set polytope, 72 symmetric cones, 80
Stackelberg game, 146 symmetric functions, 157
staff scheduling, 43, 73 system design, 62
stars, 85 tabu search, 163
state reachability, 159 target cascading, 48
stationary point, 53, 75 taxation, 65
statistical disclosure control, 138 taxation problem, 89
statistical models, 48 Tchebycheff bounds, 130
Steiner, 152 telecommunication, 99, 134, 147
Steiner problem, 57 test problem generation, 57
Steiner trees, 57 testing, 48
stiffness of vector field, 129 testing software, 79
stochastic, 65, 96, 119, 125 theory of fuzzy sets, 74
stochastic analysis, 85 theory of local search, 84
stochastic constraints, 39 time expanded networks, 71
stochastic dominance, 91, 106, 132 time stage linear programming, 122
stochastic games, 64 time windows, 68, 94, 135
stochastic integer programming, 38, 64 topology optimization, 113
stochastic model, 39, 89 tracking, 159
stochastic networks, 56 traffic assignment, 56
stochastic optimal control, 119 traffic assignment problem, 162
stochastic optimization, 41, 117, 159 traffic equilibrium, 108
stochastic programming, 39, 41, 51–53, 64, 65, 78, 90, transition states, 49
91, 94, 95, 102, 105, 106, 119, 124, 132, 145 transportation, 33, 42, 54, 90, 93, 94, 108
stochastic scheduling, 146, 147 transportation:models:traffic, 69
stratified random sampling, 110 transportation:vehicle routing, 69
strict complementarity, 101 traveling salesman, 127
strong convergence, 61 traveling salesman problem, 58, 72, 73, 84, 86, 135
strong Lipschitz bahavior, 75 treatment, 81
strong Lipschitz stability, 75 treatment planning, 43, 94
strongly convex optimization, 61 treewidth, 126
strongly polynomial algorithm, 45, 84 truncated Newton methods, 75, 129, 142
 175

Truss topology design, 43

trust region, 35, 88, 116, 130
trust region method, 35, 116, 128, 130, 141, 155
trust regions, 128, 155
turbo codes, 99
UMTS, 160, 161
unbounded multipliers, 36
unboundedness, 102
uncertain linear programs, 124
uncertainty, 74, 119, 123
uncertainty propagation, 48
unconstrained minimization, 121
unconstrained optimization, 102, 142, 158
unit commitment, 64
university timetabling, 73
upper bounds, 85
upper quasicontinuity, 118
urban economics, 78
utility function, 123
valid inequalities, 87
value chain optimization, 91
value functions, 118
value of the game, 92
variable fidelity modeling, 87
variable fixing, 33
variance, 50
variational analysis, 103
variational inequalities, 66, 75, 106, 160
variational inequality, 118
variational inequality problem, 144
variational metric, 158
variational principle, 150
variational programming, 136
vector optimization, 90, 95, 96, 110, 118, 136, 144
vector payoff function, 164
vector valued optimization, 66
vector variational inequality, 118, 144
vehicle routing, 47, 52, 55, 68, 94, 122, 135
vehicle scheduling, 94, 122
verified computing, 104
vertex coloring, 112
vertices, 72
very large scale problems, 141
virtual private network design, 80
visualization, 116
VLSI design, 31, 32, 45, 57, 109, 122, 139
VU-decomposition, 143
Walras equilibrium, 78
warm start, 91
wavelets, 60
weak efficiency, 136
weakly efficient points, 110
weighted orthogonal, 87
weighted stable set, 138
weighted tardiness, 109
well-described polytopes, 70
well-posedness, 118, 145
winner determination, 120
wireless communications, 147
workforce planning, 73
zero-order optimization, 101
zonotopes, 72
176 Speakers, Coauthors and Chairpersons

Speakers, Coauthors and Chairpersons

Page numbers refer to speakers, coauthors and chairpersons.

Abdel, Lisser, 135 Barsky, Eugene, 125

Abdullahi, Sammani Danwawu, 113 Barvinok, Alexander, 98
Abeledo, Hernan, 121 Basescu, Vasile, 89
Abramson, Mark, 34 Bastin, Fabian, 102
Absil, Pierre-Antoine, 54 Bastos, Fernando, 43, 54
Addis, Bernardetta, 77 Batta, Rajan, 63
Agra, Agostinho, 86, 87 Becchetti, Luca, 107
Ahmed, Shabbir, 132 Becker, Roland, 141
Ahr, Dino, 58, 99 Bekesi, Jozsef, 58
Ahuja, Ravindra K., 94 Belotti, Pietro, 80, 153
Akartunali, Kerem, 114 Belousov, Evgeny, 90
Akrotirianakis, Ioannis G., 104 Beltov, Tor, 164
Al-Baali, Mehiddin, 142, 143 Beltran, Cesar, 76, 127
Alber, M., 43 ben Hamor, Hatem, 55
Alber, Yakov, 66 Ben-Israel, Adi, 62
Alexandrov, Natalia, 60, 74, 87, 87 Ben-Tal, Aharon, 117
Aliprantis, C.d., 78 Bendsoe, Martin, 60
Alizadeh, Farid, 76, 80, 80 Bengtsson, Lennart, 38
Almeida, M. T., 98 Benoit, Genevieve, 72, 72, 73
Alonso, Antonio, 38 Benoit, Rottembourg, 140
Alonso, Fernando, 123 Benson, Hande, 36, 61
Alperin, Hernan, 118 Berend, Nicolas, 133
Altenstedt, Fredrik, 78, 78 Berge, Matthew, 95, 95
Althaus, Ernst, 140 Berger, Rosemary T., 120, 121
Amaldi, Edoardo, 72, 80, 152, 152, 153, 153 Berggren, Martin, 87
Ammar, Oulamara, 109 Bertsekas, Dimitri, 103
Andersen, Erling, 116, 161 Bertsimas, Dimitris, 50, 67, 117
Andersen, Kent Høj, 154 Bessi, Radhia, 159
Andersen, Søren V., 155 Bianchi, Monica, 105
Anderson, Edward, 97 Biegler, Lorenz, 49, 133
Anderson, Edward J., 148 Bienstock, Daniel, 34, 86
Andersson, Arne, 106 Bierlaire, Michel, 34, 42, 141
Andersson, Henrik, 135 Bigi, Giancarlo, 163
Andersson, Tobias, 93 Birgin, Ernesto G., 142, 142
Andreani, Roberto, 160 Bisschop, Johannes, 53
Aneja, Y. P., 84 Blanquero, Rafael, 95
Angun, M. Ebru, 39 Bley, Andreas, 82
Anitescu, Mihai, 106, 106 Blomvall, Jörgen, 51
Anjos, Miguel, 138, 138 Boggs, Paul T., 128
Antipin, Anatoly, 70 Boland, Natashia, 82
Aouchiche, Mustapha, 82 Bomze, Immanuel M., 77
Appa, Gautam, 125, 136, 154 Bonnans, Frederic, 133, 159
Arai, Hiroaki, 46 Bonnard, Nicolas, 41
Aráoz, Julián Arturo, 99, 99 Bonnel, Henri, 66, 95, 96
Argaez, Miguel, 107, 108, 116, 121 Borgwardt, Karl Heinz, 150, 150
Arkin, Vadim, 65 Borndörfer, Ralf, 47, 59
Arutyunov, Aram, 159, 159 Borne, Sylvie, 133
Atamturk, Alper, 154 Borodin, Allan, 162, 162
Audet, Charles, 34, 36 Bortfeld, Thomas, 123
Aussel, Didier, 118 Bourreau, Eric, 140
Avelino, Catarina, 101 Boyar, Joan, 92, 93, 162, 162
Avella, Pasquale, 45, 46, 73 Boyd, Sylvia, 72, 73
Azé, Dominique, 157 Braams, Bastiaan J., 49
Babonneau, Frédéric, 50, 127 Brass, Peter, 48
Bagirov, Adil, 143 Brazil, Marcus, 139
Bai, Y., 134 Bredström, David, 148
Bailey, Matthew, 64, 160 Brekelmans, Ruud, 141
Bandler, John, 154 Brenner, Ulrich, 45
Bang-Jensen, Joergen, 84 Brezhneva, Olga, 101, 101
Barahona, Francisco, 139, 139 Brieden, Andreas, 158, 158
 177

Briet, Joachim, 149 Correa, Rafael, 155

Brinkhuis, Jan, 143 Corrêa, Ricardo, 112
Brostrom, Peter, 80 Costa, Marie-Christine, 84
Brualdi, Richard A., 44 Courdurier, Matias, 36
Buchholz, Jens, 149 Crespi, Giovanni Paolo, 144, 145
Burachik, Regina, 157 Crittin, Frank, 34, 42
Burdakov, Oleg, 129, 130 Crocker, Ian, 81
Burer, Sam, 110 Csendes, Tibor, 131, 131, 152
Burkard, Rainer, 42, 78, 150 Cunha, Alexandre, 71
Burke, James, 103 Cunningham, Bill, 32, 32, 72, 99
Bussieck, Michael, 103 Curtin, Margaret, 74
Butnariu, Dan, 66 D’Apuzzo, Marco, 156
Byrd, Richard, 35, 61, 88 da Cruz Neto, Joao Xavier, 158
Calheiros, Felipe, 100 da Silva e Silva, Paulo Jose, 79
Calvin, James, 48 Dahl, Geir, 44, 44, 55, 125
Cambini, Alberto, 77 Dahl, Joachim, 155, 155
Cambini, Riccardo, 90, 90, 158 Dai, Jim, 80
Cameron, Kathie, 69, 69 Daneva, Maria, 161, 162
Campana, Emilio F., 87 Dang, Chuangyin, 114, 114
Campêlo, Manoel, 112 Daniel, Bob, 79, 79
Cánovas, M. J., 124, 145 Danna, Emilie, 153
Capone, Antonio, 147 Darby-Dowman, Ken, 91
Caporossi, Gilles, 82 Das, Arup Kumar, 92
Caprara, Alberto, 59 Dash, Sanjeeb, 86
Cardia, Marli, 50 Davis, Sashka, 162
Cardoso, Domingos Moreira, 44 Davis, Timothy A., 150
Carlsson, Dick, 54 Day, Robert W., 52
Carney, Paul R., 94 de Aragão, Marcus Poggi, 33, 47
Carosi, Laura, 90 de Farias, Ismael, 154, 154
Carrizosa, Emilio, 95, 110 de Klerk, Etienne, 131, 156, 156
Cartis, Coralia, 108 de Nijs, Johannes J., 158
Castro, Clicia Stelling, 37 de Oliveira, Antonio Alberto Fernandes, 130
Cau, Thai Doan Hoang, 97 de Paepe, Willem E., 107
Cela, Eranda, 91 de Souza, Cid C., 99, 100
Cerveira, Adelaide, 43 de Souza, Mauricio C., 69
Cesana, Matteo, 147 de Vries, Sven, 106, 120
Ceselli, Alberto, 126, 126 Deineko, Vladimir, 135
Chandran, L. Sunil, 125, 126 Delbos, Frederic, 61, 61
Chandrasekaran, R., 84 Demaseure, Katia, 115, 115
Chaovalitwongse, Wanpracha, 94 DeMiguel, Angel-Victor, 120
Chebalov, Sergei, 38 Dempsey, James F., 94
Chekuri, C., 127 Demyanov, Alexei, 103
Chellappa, S., 60 den Hertog, Dick, 39, 141
Chen, Bo, 108 Denizel, Meltem, 68
Chen, Kenny K.Y., 109 Dennis, John, 34, 36
Chen, Ming Zhe, 42 Dentcheva, Darinka, 106, 106
Cheung, Kevin, 72 Desrosiers, Jacques, 55, 55, 73
Chinneck, John W., 153, 153 Dewez, Sophie, 90
Chiralaksanakul, Anukal, 145 Díaz, A. R., 60
Cho, YongJun, 137 Diehl, Moritz, 132
Christiansen, Marielle, 51 Dinh, Tao Pham, 144
Chua, Chek Beng, 80 Dirkse, Steven, 102
Chubanov, Sergei, 68 Djukanovic, Igor, 112
Chung, Chia-Shin, 148 Dmitruk, Andrei, 133
Claire, Kenyon, 135 Dolgui, Alexandre, 58
Clarke, John-Paul, 63 Dormer, Alan, 40
Clement, M. F., 38 Dragan, Irinel, 88
Cominetti, R., 36 Drexl, Andreas, 126
Conde, Eduardo, 95 Driessen, Lonneke, 141
Conn, Andrew, 127, 128 Drud, Arne Stolbjerg, 102, 102
Constantino, Miguel, 87, 100, 140 Drummond, Lucia, 37
Cordeau, Jean-Francois, 134, 135 du Merle, Olivier, 36, 51, 127
Cornaz, Denis, 69 Duarte de Carvalho, Filipa, 98
Cornuejols, Gerard, 154 Duer, Mirjam, 63
Correa, Jose Rafael, 56, 58, 59 Dula, Jose, 164
178 Speakers, Coauthors and Chairpersons

Dumitrescu, Irina, 82 Fonlupt, Jean, 69

Dunagan, John, 121 Forsgren, Anders, 102
Dussault, Jean-Pierre, 102, 142 Fortz, Bernard, 43
Dye, Shane, 91 Fouilhoux, Pierre, 122
Dyer, Martin, 113 Fourati, R. Bessi, 159
Eckstein, Jonathan, 50, 79, 79, 153 Fourer, Robert, 40, 41, 53, 53, 66, 66
Edmonds, Jack, 70 Fox, Tim, 81
Ehrgott, Matthias, 123, 150 Frangioni, Antonio, 100
Ehrhardt, Klaus, 98 Frank, Paul, 74
Eichhorn, Andreas, 64, 65, 106 Freling, Richard, 121
Eisenblätter, Andreas, 160, 161 French, Alan, 126
Eisenbrand, Friedrich, 140, 140, 140 Freund, Robert, 67, 67, 89
Ekin-Karasan, Oya, 80 Friedlander, Ana, 160
El ghami, Mohamed, 54, 54 Friedlander, Michael Paul, 61, 120, 120
Elber, Ron, 49 Frota, Yuri, 112
Eldred, Michael, 74 Fuchs, Bernhard, 92
Elendner, Thomas, 120 Fuegenschuh, Armin, 94
Elghami, M., 134 Fujie, Tetsuya, 37, 47
Elhedhli, Samir, 50 Fujisawa, Katsuki, 62, 81, 81, 82, 93
Ellison, Frank, 90, 92 Fujishige, Satoru, 31, 57, 57
Engel, Konrad, 43 Fukasawa, Ricardo, 47, 59
Enkhbat, Rentsen, 101 Fukuda, Mituhiro, 41, 81, 93
Epelman, Marina, 159, 160 Fülöp, János, 131
Epping, Thomas, 55 Funke, Stefan, 140
Eric, Bourreau, 140 Gablonsky, Joerg, 74
Eriksson, Jerry, 35, 87, 88 Gaivoronski, Alexei, 119
Ermoliev, Y. M., 159 Galati, Matthew, 100
Ernst, Andreas, 55 Galbiati, Giulia, 72
Eschen, Elaine, 69 Ganesh, K., 93, 94
Escudero, Laureano F., 38 Gárciga, Otero Rolando, 61, 61
Eso, Marta, 100 Gassmann, Gus, 90, 90
Estévez-Fernández, M. Aránzazu, 70, 70 Gassner, Elisabeth, 112
Evans, Lisa, 100 Gay, David, 53
Evgrafov, Anton, 89, 90 Geerdes, Hans-Florian, 133, 134
Faco´, Joao-Lauro, 156 Gharbi, Anis, 163
Faigle, U., 110 Giczi, David, 78
Fampa, Marcia, 89 Gil, M. L., 38
Fang, Youkang, 131 Gilbert, Jean Charles, 61, 61
Fasano, Giovanni, 142 Gillard, P., 57
Favrholdt, Lene, 93 Ginchev, Ivan, 144, 145
Fekete, Sandor, 112 Girlich, Eberhard, 45
Feofiloff, Paulo, 71 Giunta, Anthony, 74
Feremans, Corinne, 43, 43 Glineur, François, 161
Fernandes, Cristina, 71 Glover, Fred, 100
Fernandes, Luis Merca, 160 Goemans, Michel, 135
Fernández, Elena, 99 Goffin, Jean-Louis, 76, 89
Ferreira, Carlos E., 71 Gol’shtein, Evgeny G., 158, 158
Ferreira, Orizon, 158 Golbaharan, Nima, 73, 126
Ferris, Michael, 94, 102 Goldfarb, Don, 117, 161
Fiestras-Janeiro, M.Gloria, 70 Gondzio, Jacek, 62, 129, 129
Figueiredo, Isabel, 160 Gonzaga, Clovis C., 49, 50
Filho, Nelson Maculan, 89 González, Juan José Salazar, 43
Finus, Michael, 83 Gonzalez-Lima, Maria, 107
Flåm, Sjur Didrik, 83, 83 Göran, Anders, 53
Flatberg, Truls, 125 Gordon, Valery, 136, 136
Flegel, Michael L., 103, 104 Gotoh, Jun-ya, 63, 77
Fleischer, Lisa, 71, 71 Gould, Nick, 35, 88
Fletcher, Roger, 88 Gourdin, Eric, 133
Flisberg, Patrik, 92 Gouveia, Luis, 42, 47, 55, 57, 139
Florenzano, Monique, 78 Goux, Jean-Pierre, 41
Floudas, Christodoulos, 104 Gowda, Muddappa, 40
Floudas, Christodoulos A., 104 Graña Drummond, Luis Mauricio, 136
Flynn, James, 148 Granot, Daniel, 110, 110
Fodstad, Marte, 38, 91 Gregorio, Ronaldo, 93
Foldnes, Njål, 55 Grigoriev, Dima, 156
 179

Griva, Igor, 36, 61 Hossain, Mohammad Z., 137

Groetschel, Martin, 47 Hossain, Shahadat, 98, 116
Grothey, Andreas, 62, 62, 129, 139 Hotta, Keisuke, 46
Guerraggio, Angelo, 118 Hougardy, Stefan, 44
Guertler, Marion, 92 Huang, Siming, 67
Gueye, Serigne, 32, 33 Huhn, Petra, 67
Guigues, Vincent, 124, 124 Huisman, Dennis, 121
Gunluk, Oktay, 86, 120 Huyer, Waltraud, 48
Gunnarsson, Helene, 54 Huygens, David, 125
Gupta, Neelam, 136, 137 Iasemidis, Leonidas D., 94
Gurkan, Gul, 39 Ibaraki, T., 101
Guschinsky, Nikolai, 58 Ikebe, Yoshiko, 95
Gutiérrez, César, 163 Ikegami, Atsuko, 43
Gutin, Gregory, 84, 84, 84 Imaizumi, Jun, 46
Gzara, Fatma, 76 Immanuel, Bomze, 77
Haarala, Marjo, 117 Impagliazzo, Russell, 162
Hachez, Yvan, 156 Inagaki, Mizuho, 115
Hachimori, Masahiro, 56, 56, 125 Iusem, Alfredo Noel, 66, 136
Haddou, Mounir, 133 Iwamoto, Seiichi, 65
Hadjisavvas, Nicolas, 118, 118 Iwata, Satoru, 45
Hagai, Ilani, 125 Izmailov, Alexey, 75, 159, 159
Hager, William, 142, 150 Jara Moroni, Pedro Daniel, 78
Hall, Alex, 69, 71 Jardim, Maria Helena Cautiero Horta, 142
Hall, Nicholas G., 123 Jensen, David L., 100
Hamelin, Benoit, 101, 102 Jensen, Marianne Storhaug, 51
Hamers, Herbert, 83, 110, 141 Jensen, Søren. H., 155
Hang, Darren Dejun, 64 Jerome, Malick, 155
Hansen, Jesper, 149 Jibetean, Dorina, 131
Hansen, Pierre, 73, 82, 82, 83 Jiménez, Bienvenido, 163
Haouari, Mohamed, 163, 163 Joborn, Martin, 42
Hare, Warren, 103 Jofré, Alejandro, 78, 79
Hassin, Refael, 109 Johannes, Berit, 136
Hatzl, Johannes, 149, 150 Johnson, Ellis, 100
Haugland, Dag, 48, 52 Jorge Nocedal, Richard Waltz, 88
Haus, Utz-Uwe, 114 Judice, Joaquim, 160
Hauser, Raphael, 129, 153, 153 Kabadi, Santosh, 84
Hearn, Don, 108 Kachani, Soulaymane, 108
Heinkenschloss, Matthias, 115 Käck, Jan-Erik, 128
Heinrich-Litan, Laura Nicoleta, 48 Kaeaeb, Vanessa, 136
Held, Harald, 52 Kaibel, Volker, 140
Held, Stephan, 31 Kallehauge, Brian, 68
Helmberg, Christoph, 157 Kallenberg, Lodewijk, 146
Hendrix, Eligius Maria Theodorus, 70, 83, 104, 104 Kallio, Markku, 105
Henningsson, Mathias, 134 Kang, Laura, 63
Henrion, Rene, 39, 39 Kanzow, Christian, 101, 104
Hernandez, Jaime, Jr., 116 Kaplan, A., 79
Herskovits, José, 157, 157 Karas, Elizabeth, 49
Heyde, Frank, 110 Karisch, Stefan, 38, 51
Hiller, Benjamin, 47 Karl, Holger, 134
Hilli, Petri, 78 Karlof, John, 139
Hintermüller, Michael, 128 Karlsson, Jenny, 134
Hinze, Michael, 141 Kas, Peter, 132
Hippler, Steffen, 71 Kashiwabara, Kenji, 57
Hiriart-Urruty, Jean-Baptiste, 36, 37 Kassay, Gabor, 66, 132
Hjorring, Curt, 38 Katayama, Naoto, 42
Ho, Sin C., 51, 52 Kaut, Michal, 145
Hoang, Chinh T., 69 Kawamoto, Atsushi, 113, 113
Hochbaum, Dorit, 69, 69 Kawasaki, Hidefumi, 35
Hochreiter, Ronald, 78 Kedad-Sidhoum, Safia, 138, 138
Hochstättler, Winfried, 55, 92 Kellerer, Hans, 85
Høeg, Erik, 38 Kelley, Carl, 115
Holmberg, Kaj, 62, 63, 80 Kelly, Karen A., 121
Holmström, Kenneth, 53, 128 Kern, Walter, 92, 110
Hooker, John, 154, 154 Keyzer, M. A., 159
Horbach, Andrei, 45 Khazafi, Khadija, 136
180 Speakers, Coauthors and Chairpersons

Kijima, Shuji, 98 Langkau, Katharina, 69

Kim, Sunyoung, 34, 148 Laporte, Gilbert, 52
Kim, Won Kyu, 118 Larsen, Kim S., 93, 162
King, Alan, 52, 53, 105 Larsson, Torbjörn, 130
Kis, Tamas, 32 Lasdon, Leon, 103
Kiseleva, Elena, 159 Laugier, Alexandre, 152
Kiwiel, Krzysztof, 63 Laurent, Monique, 138
Klabjan, Diego, 38, 50 Laurent-Varin, Julien, 132, 133
Klamroth, Kathrin, 123 Lawphongpanich, Siriphong, 108
Klatte, Diethard, 75, 75 Le Thi, Hoai An, 144
Klau, Gunnar W., 57, 71, 71 Lee, Eva, 81, 100, 114
Kleijnen, Jack P. C., 39 Lee, Jon, 34
Kliewer, Georg, 33 Lee, Y.- J., 81
Klinz, Bettina, 112 Lemaire, Sébastien, 51
Klose, Andreas, 126 Lemaréchal, Claude, 36
Knobloch, Matthias, 143 Lentink, Ramon, 122
Kobayashi, Kazuhiro, 62 Leonardi, Stefano, 107
Koch, Thorsten, 160, 161 Leopold-Wildburger, U., 54
Kocvara, Michal, 41, 155 Lesaja, Goran, 148
Koeppe, Matthias, 47, 114 Leslie E. Trotter, Jr., 152
Kohl, Niklas, 63 Letchford, Adam, 43, 58, 86
Köhler, Ekkehard, 69 Létocart, Lucas, 84
Koivu, Matti, 53, 64, 78 Leung, Janny, 34
Kojima, Masakazu, 34, 81, 82, 148 Leung, Stephen C. H., 161
Kokkolaras, Michael, 48 Levi, Retsef, 149
Komaromi, Eva, 75 Levin, Asaf, 109
Kong, Nan, 64 Levin, Genrikh, 58, 62
Konno, Hiroshi, 63 Levin, Yuri, 62
Konnov, Igor, 53 Lewis, Adrian S., 103, 157
Korach, Ephraim, 85 Leyffer, Sven, 88, 133, 133, 133
Korf, Lisa, 119, 119 Li, Duan, 86
Kortanek, Ken, 65, 65 Li, Jonathan G., 94
Korte, Bernhard, 45 Li, Yanjun, 154
Koshlai, Ludmila, 39, 40 Liang, Ming, 163
Koster, Arie M.C.A., 32, 125, 160 Liau, Bernard, 133
Kostina, Ekaterina, 60 Liberti, Leo, 72
Kotnyek, Balazs, 125 Libura, Marek, 84, 84
Kounoike, Yuusuke, 37 Liebling, Thomas M., 72
Kovalev, Mikhail M., 45 Lin, Bertrand, 109
Kovaleva, Sofia, 130 Lin, Wugin, 80
Kovalyov, Mikhail Y., 68 Lindberg, Per Olov, 73, 126, 162
Krarup, Jakob, 54, 54 Linderoth, Jeff, 51
Krause, Stefan, 82, 82 Lipmann, Maarten, 107
Kreines, Elizabeth M., 164 Lisser, Abdel, 135, 135
Krejic, Natasa, 62, 160, 160 Liu, Xinwei, 102, 107
Krishnamoorthy, Bala, 34 Liu, Zhen, 50
Krishnan, Kartik, 41, 76, 88 Liuzzi, Giampaolo, 74, 75
Krivonozhko, Vladimir, 52 Ljubic, Ivana, 47, 71
Kroon, Leo, 121, 121, 122 Löbel, Andreas, 59
Krumke, Sven O., 47, 107 Locatelli, Marco, 77, 77, 77
Krzemienowski, Adam, 65 Lodi, Andrea, 100
Kubzin, Mikhail, 149 Loehne, Andreas, 136, 136
Kuefer, Karl-Heinz, 123 Loiseau, Irene, 120
Kuipers, J., 110 Long, Kevin R., 128
Kum, Sangho, 118 Lopes, Leonardo, 41
Kumar, Arvind, 94 López, M. A., 124
Kumlander, Deniss, 114 Lopez-Calva, Gabriel, 133
Kummer, Bernd, 75, 75, 75 López-Cerdá, Marco A., 145
La Torre, Davide, 132 Lophaven, Søren N., 140
Labbé, Martine, 43, 73, 90 Lougee-Heimer, Robin, 127, 127
Lachiheb, Mehdi, 129 Loute, Etienne, 124
Ladanyi, Laszlo, 51, 100, 120 Louveaux, Quentin, 47, 114, 114
Lai, Kin Keung, 161 Lübbecke, Marco, 33, 112, 112
Lambert III, Theodore J., 160 Lucambio Perez, Luis Roman, 158
Land, Ailsa, 106 Lucena, Abilio, 55, 71, 100
 181

Lucidi, Stefano, 75 Miettinen, Kaisa, 66, 117

Luigi Laura, Alberto Marchetti-Spaccamela, 107 Mifflin, Robert, 143, 143
Lustig, Irvin, 75, 140, 140 Mikhail, Mikhalevich, 40
Luzanin, Zorana, 62, 160 Mikhalevich, M., 39
Lyengar, Garud, 117, 117 Miller, Andrew, 100, 114
Lysgaard, Jens, 55, 55 Mingozzi, Aristide, 73
Maaranen, Heikki, 117 Mitchell, John, 76, 76, 88, 88, 89
Macambira, Elder, 99 Mitra, Gautam, 66, 90, 91, 91, 92, 119
MacLean, Leonard, 119 Mittelmann, Hans D., 116, 116, 116
Maculan, Nelson, 71, 72, 99, 120, 142 Miyahara, Tetsuhiro, 113
Madsen, Kaj, 127, 140, 154, 154 Miyashiro, Ryuhei, 95
Madsen, Oli B.G., 68, 68 Miyazawa, Flavio Keidi, 149
Maeght, Jean, 36 Mizuno, Kazunori, 98
Maerkert, Andreas, 91, 132, 132 Mohan, Srinivasa R., 40, 40
Maffioli, Francesco, 72, 72, 80 Mohri, Hiroaki, 42
Magnanti, Thomas, 139 Möhring, Rolf H., 146
Magos, Dimitris, 154 Monaci, Michele, 59
Mahey, Philippe, 69 Mongeau, Marcel, 37
Mahjoub, Ali Ridha, 43, 122, 125, 133 Monteiro, Ana Margarida, 105
Mak, Vicky, 55 Monteiro, Renato, 96, 161
Makarenkov, Oleg, 115 Monz, Michael, 123
Makela, Marko M., 66, 117, 117 Moorthy, Rajeeva, 80
Malik, Madhur, 40 More’, Jorge, 49
Malucelli, Federico, 147, 147 Morgan, Jacqueline, 96, 118, 144
Mangasarian, Olvi, 81 Moriguchi, Satoko, 45
Mannino, Carlo, 32 Morin, Pat, 48
Marchetti-Spaccamela, Alberto, 107 Morito, Susumu, 46
Marchi, Anna, 105 Moriyama, Sonoko, 125, 125
Marcotte, Patrice, 90 Morton, David, 145, 145
Marenco, Javier Leonardo, 46 Mourtos, Ioannis, 154
Margolis, Stephen, 128 Mueller, Dirk, 45
Margot, Francois, 34 Mueller-Hannemann, Matthias, 109
Marian, Daniela, 105 Munson, Todd, 49, 49
Marín, Ángel, 62 Muramatsu, Masakazu, 41
Markot, Mihaly Csaba, 152 Murota, Kazuo, 31, 31, 45, 45
Maros, Istvan, 122, 150 Murray, Alan, 46
Marron, J. S., 76 Mutzel, Petra, 47, 71
Martein, Laura, 77, 77 Muu, Le Dung, 144
Martin, Alexander, 94, 160, 161, 161 Mydlarz, M., 127
Martínez, Francisco, 78 Myung, Young-Soo, 138
Martinez, J. M., 142 Nagel, Christian, 101
Martins, Pedro, 47, 47 Nair, K. P. K., 84
Maschler, M., 110 Nakamura, Masataka, 56
Mason, Andrew, 73 Nakata, Kazuhide, 82, 93
Matsui, Tomomi, 47, 95 Narayanan, Harihar, 31
Matsui, Yasuko, 113 Narendran, T. T., 94
Matsumoto, Satoshi, 113 Natarajan, Karthik, 50, 50
Matsuoka, Satoshi, 81 Nauss, Robert, 59, 60
Matsyu, Tomomi, 98, 98 Nayakkankuppam, Madhu, 157, 157
Mazzoleni, Piera, 131, 131 Neame, Philip, 96
McCormick, S. Thomas, 43 Nediak, Mikhail, 153
McKinnon, Ken, 139, 139, 139 Nedic, Jelena, 129
Medard, Claude Ph., 51 Nemeth, Sandor Z., 158
Meer, Klaus, 145, 146 Nemhauser, George, 86
Meeraus, Alex, 102 Nemoto, Toshio, 46
Mehlhorn, Kurt, 140 Neogy, Samir Kumar, 92, 92
Mehrotra, Anuj, 98 Nesterov, Yurii, 36, 156
Meijer, Henk, 112 Neumaier, Arnold, 48, 79
Melissen, J. B. M., 54 Newman, Alantha, 112
Meller, Jarek, 49 Nielsen, Eric, 74
Mello, Margarida P., 160 Nielsen, Hans Bruun, 140
Mendez, Osvaldo, 108 Nieminen, Kimmo Tapio, 122, 122
Mendez-Diaz, Isabel, 55, 112 Nino-Mora, Jose, 146, 147
Meszaros, Katalin, 73, 74 Nishihara, Seiichi, 98
Michelon, Philippe, 33 Noailles, Joseph, 86
182 Speakers, Coauthors and Chairpersons

Nocedal, Jorge, 35, 35, 61, 133 Pham Dinh, Tao, 144
Nogales, Francisco J., 120 Philpott, Andy, 96, 96, 96
Norde, Henk, 83 Piccinonno, Fulvio, 98, 99
Norkin, Vladimir, 159 Pillo, Gianni Di, 75
Novikova, Natalia N., 164 Pinar, Mustafa C., 80
Novo, Vicente, 163 Pini, Rita, 105, 105
Nowak, Ivo, 118, 118 Pinter, Janos, 103
Nowak, Matthias Peter, 91, 145 Pinyagina, Olga, 53
Nygreen, Bjørn, 51, 139 Pisinger, David, 54, 68, 85, 85
O’Leary, Dianne, 147 Pitsoulis, Leonidas, 152
O’Neal, Jerome, 161 Plateau, Gerard, 120
O’Sullivan, Michael, 96 Pleschiutschnig, Carmen, 42
Obuchowska, Wieslawa, 37 Pochet, Yves, 68
Oeuvray, Rodrigue, 141 Poensgen, Diana, 107
Ogryczak, Wlodzimierz, 65 Poggi de Aragao, Marcus, 59
Okamoto, Yoshio, 57 Polak, Elijah, 74
Olesen, Ole Bent, 164 Polyak, Roman, 104
Olieman, Niels, 70, 83 Polzin, Tobias, 57
Oliveira, Ivan, 60 Poojari, Chandra, 90
Oliveira, Paulo Roberto, 93 Popescu, Ioana, 39
Onn, Shmuel, 72 Porto, Oscar, 47, 60
Orban, Dominique, 66, 88 Pospelova, Irina, 163, 164
Ordonez, Fernando, 89 Potts, Chris, 122, 123
Oriolo, Gianpaolo, 32, 72 Powell, Michael, 35
Orlin, James B., 84 Powell, Susan, 106, 119
Orsenigo, Carlotta, 113 Pratt, Rob, 56
Ortuño, M. T., 38 Preda, Dorin, 86
Oskoorouchi, Mohammad R., 89 Prestwich, Steven, 125
Oswald, Marcus, 58 Pritchard, Geoffrey, 96
Overton, Michael, 103, 103 Proll, Les, 113
Ozana, Marek, 51 Proth, Jean-Marie, 136
Ozdaglar, Asuman, 103 Provan, Scott, 55, 56
Palagi, Laura, 75 Pruessner, Armin, 116
Pallaschke, Diehard Ernst, 77 Punnen, Abraham P., 84
Palmgren, Myrna, 68 Raffensperger, John, 85
Pang, Jong-Shi, 106 Raghavan, S., 52
Panton, David, 73 Raghunathan, Arvind, 133
Papalambros, Panos Y., 48 Raidl, Guenther, 47
Pape, Claude Le, 153 Ralph, Daniel, 120
Pardalos, Panos M., 81, 94, 94 Ralphs, Ted, 51, 88, 100, 100
Parra, J., 124, 145 Rambau, Jörg, 47
Parrilo, Pablo, 156 Ramirez Cabrera, Hector, 155
Pasechnik, Dmitrii V., 156 Ranne, Antero, 78
Passarella, Wilhelm, 157 Rasala, April, 123
Pataki, Gabor, 34, 34, 41, 143, 144 Rasmussen, Anders Bo, 85
Patel, Jagat, 153 Raupp, Fernanda Maria Pereira, 110, 110
Patera, A. T., 60 Raydan, M., 142
Patricio, Joao, 160 Reemtsen, Rembert, 43, 43
Patriksson, Michael, 90, 130 Reikvam, Kari M., 51
Pearson, Nicholas, 58, 85, 86 Reinelt, Gerhard, 58, 99
Pedersen, Frank, 155 Reis, Marcelo Ladeira, 59
Pena, Javier, 130, 130, 131, 143 Remshagen, Anja, 140
Peng, Jiming, 134, 148 Renaud, Arnaud, 36, 41
Penn, Michal, 152, 152 Rendl, Franz, 112, 138
Pennanen, Teemu, 53, 64, 78, 105, 105 Requejo, Cristina, 139
Perakis, Georgia, 67, 67, 108, 108 Resende, Mauricio G.C, 71
Peri, D., 87 Rey, Pablo A., 60, 117
Perregaard, Michael, 86 Ricci, Federico, 32
Pesch, Erwin, 68 Righini, Giovanni, 126
Pesneau, Pierre, 43, 125 Rinaldi, Giovanni, 100
Petersen, Niels Christian, 164, 164, 164 Rivera, Jorge, 78, 79
Peyer, Sven, 109 Rocca, Matteo, 144, 144, 145
Pferschy, Ulrich, 71, 85 Roemisch, Werner, 39, 65, 106
Pfetsch, Marc, 152 Roma, Massimo, 129, 142
Pflug, Georg, 78 Roman, Diana, 91
 183

Romaniello, Mariella, 144 Sen, Suvrajeet, 90

Romeijn, Edwin, 94 Sendov, Hristo, 157
Romero Morales, Dolores, 110 Sethuraman, Jay, 71
Rømo, Frode, 91, 145 Shakhlevich, Natalia, 163
Ronconi, Débora Pretti, 109, 109 Shanno, David, 36, 61
Rönnqvist, Mikael, 54, 68, 92, 134, 148 Shcherbina, Oleg, 79
Roos, Kees, 54, 134 Shechter, Steven, 64
Røpke, Stefan, 68 Shepherd, Bruce, 126, 127
Rote, Günter, 55 Shi, Jianming, 144
Rothberg, Edward, 153 Shiau, D.- S., 94
Rothblum, Uriel G., 72 Shiina, Takayuki, 64
Roundy, Robin, 149 Shinano, Yuji, 37
Roupin, Frédéric, 84 Shioura, Akiyoshi, 31, 31
Royset, Johannes, 74 Shishkin, Serge, 101
Rozenfeld, Stas, 152 Shmoys, David, 149, 149
Rozin, Boris, 62 Shoudai, Takayoshi, 113
Rump, Christopher M., 63 Shyshkanova, Ganna, 123
Rustem, Berc, 34 Siegfried, Schaible, 77
Ruszczynski, Andrzej, 106, 106, 106, 132 Sigurd, Mikkel Mühldorff, 33
Ryan, David M., 33, 46, 46, 59, 68 Silva, Renata, 36
Sackellares, J. Chris, 94 Sim, Chee-Khian, 80
Saenz, Roberto, 121 Sim, Melvyn, 117
Sagastizábal, Claudia Alejandra, 117, 143 Simonetti, Luidi, 89
Sahinidis, Nikolaos, 63, 63, 103 Sinoquet, Delphine, 61
Saito, Hiroo, 47 Sitters, Rene, 107
Sáiz Pérez, Elena, 83 Skutella, Martin, 69, 71, 71, 71, 146
Salahi, Maziar, 134 Slastnikov, Alexander, 65
Salazar González, Juan José, 138 Smaoui, Hichem, 129, 159
Saldanha-da-Gama, Francisco, 42, 42 Smith, Robert L., 159, 160
Salgado, Liliane R. Benning, 122 Sodini, Claudio, 158
Salmerón, Javier, 62 Sofer, Ariela, 81
Saltzman, Matthew, 51, 127 Solodov, Mikhail, 75, 117
Sanchez-Arroyo, Abdón, 56, 56 Son, Pham Tien, 144
Sandvik, Rune, 85 Søndergaard, Jacob, 140
Santiago, Claudio Prata, 142 Sood, Anshul, 67
Santos, Paulo Sergio, 158 Sosa, Wilfredo, 110
Santos, Sandra Augusta, 160 Sotomayor, Marilda, 110
Sarac, Abdulkadir, 63 Spieksma, Frits, 130, 130
Sartenaer, Annick, 102, 115 Spielman, Daniel, 121
Sassano, Antonio, 46 Spille, Bianca, 45, 72
Saunders, Michael Alan, 61, 121, 121 Spjuth, Sami, 51
Savard, Gilles, 90 Sritharan, R., 69
Sawhney, Nidhi, 51 Stalinkki, Piotr, 148
Scalzo, Vincenzo, 118 Stancu-Minasian, I.m., 95
Schaefer, Andrew, 64, 64, 64 Starke, Jens, 48, 48, 115
Schaefer, Guido, 107 Stauffer, Gautier, 72
Schaefer, Uwe, 92 Stefanov, Stefan M, 89, 89
Scheimberg, Susana, 157, 158 Steihaug, Trond, 98, 116, 116
Scheinberg, Katya, 127, 128 Stein, Clifford, 123
Scheithauer, Guntram, 152 Stein, Oliver, 146
Scherrer, Alexander, 123, 123 Steinbach, Marc C., 98
Schichl, Hermann, 104 Steinberg, Richard, 106
Schindl, David, 73 Stepanchuk, Tatyana, 159, 159
Schmidt, Guenter, 149 Stephen, Tamon, 98
Schochetman, Irwin, 159 Stern, Michal, 85
Schoen, Fabio, 77 Stevanovic, Dragan, 82, 82
Scholtes, Stefan, 120 Stewart, David E., 106
Schrijver, Lex, 121 Stidsen, Thomas, 127
Schultz, Ruediger, 91, 132, 149 Stier, Nicolas, 56
Schulz, Andreas S., 56, 59, 84, 146 Stingl, Michael, 41, 155
Schummer, James, 106 Storøy, Sverre, 48
Schurr, Simon, 147 Stougie, Leen, 54, 107
Schütz, Peter, 38 Strauss, Frank, 115
Schwehm, Christian, 123 Strusevich, Vitaly, 136, 149, 163
Sen, Arun, 36, 61 Struzyna, Markus, 31
184 Speakers, Coauthors and Chairpersons

Sturm, Jos F., 110, 110 Uetz, Marc, 146

Su, Che-Lin, 120 Ulbrich, Michael, 36, 36, 116, 128
Sultana, Rudy Kusnadi, 121 Ulbrich, Stefan, 36, 49, 115, 141
Sun, Defeng, 147 Ulstein, Nina Linn, 51
Sun, Jie, 102, 102, 107, 147 Uno, Takeaki, 57, 113, 113
Sun, Xiaoling, 86 Urbański, Ryszard, 77
Sural, Haldun, 68 Uthaisombut, Pratchawat, 123
Svaiter, Benar, 66, 79 Utkin, Oleg B., 52
Szegedy, Christian, 32 Vahdati Daneshmand, Siavash, 57
Szego, Laszlo, 45 Valente, Patrick, 66
Tabbert, Eckhard, 43 van der Broek, John, 54
Tadonki, Claude Martin, 127 van der Laan, Gerard, 53
Takouda, Pawoumodom L., 37 van der Vlerk, Maarten H., 52
Talbot, Christophe, 133 van Maaren, Hans, 114
Talluri, Kalyan, 67 Van Slyke, Richard, 37, 37
Talman, Dolf, 53, 53 van Thoai, Nguyen, 77
Tammer, Christiane, 150, 150 van Velzen, Bas, 83
Tamura, Akihisa, 57 Van Vyve, Mathieu, 100
Tangir, Yoav, 72 Vandenbussche, Dieter, 86
Tao, Pham Dinh, 144 Vanderbeck, Francois, 33
Tardella, Fabio, 57 Vanderbei, Robert J., 36, 61
Tawarmalani, Mohit, 118 Vanti, Marcia, 49
Tebboth, James, 62 Värbrand, Peter, 93, 135, 147, 147
Teixeira, Ana Paula Aires Borges, 54 Vasil’ev, Igor, 46, 73
Telhada, João, 57 Vazacopoulos, Alkis, 40
Teng, Shanghua, 121 Veinott Jr., Arthur F., 96
Teo, Chung-Piaw, 50, 80, 80 Velazquez, Leticia, 108, 116, 121
Terlaky, Tamas, 134, 148 Vera, Juan, 131
Teytaud, Olivier, 41 Vercellis, Carlo, 113
Theis, Dirk, 58 Vial, Jean-Philippe, 36, 36, 50, 50, 51, 127
Thi, Hoai An Le, 144 Vicente, Luís N., 36, 49, 101, 105, 127, 128, 128
Thieke, Christian, 123 Vielma, Juan Pablo, 46
Thiele, Aurelie, 67 Vigerske, Stefan, 118
Thierry, Benoist, 140 Vik, Lars Harald, 145
Thomadsen, Tommy, 127 Viklands, Thomas, 35, 87
Tichatschke, Rainer, 66, 66, 79 Villalobos, Cristina, 116
Tiedemann, Stephan, 132, 132, 132 Viterbo, Jose, 37
Tigan, Stefan, 95 Voelker, Meta M., 94
Timajev, Larissa, 33 Voetmann, Tim, 66
Tiskin, Alexander, 135 Vohra, Rakesh, 106
Tits, Andre, 54, 147 Volodin, Andrei, 52
Todd, Michael, 76, 76 Vondrak, Jan, 135
Toft, Bjarne, 152, 152 Vredeveld, Tjark, 107, 107
Toh, Kim-Chuan, 157 Vygen, Jens, 31, 45
Toint, Philippe L., 88 Waechter, Andreas, 49
Toktas, Berkin, 52 Wagelmans, Albert, 121
Toledo-Melero, F.Javier, 124, 145 Wagler, Annegret Katrin, 44, 46
Tomasgard, Asgeir, 38, 54, 91, 145 Wagner, Michael, 49
Tomlin, John, 121 Wakabayashi, Yoshiko, 122
Toraldo, Gerardo, 155, 156 Waki, Hayato, 148
Torng, Eric, 123 Wallace, Stein W., 52, 145
Torre, Davide La, 163 Waltz, Richard, 35, 61, 88, 102
Torres, Luis Miguel, 47 Wang, Chang-Yu, 130
Toth, Paolo, 59 Wang, Jun, 86
Tourky, R., 78 Wang, Yaoguang, 32
Triadou, Coralie, 36 Wanscher, J., 54
Trick, Michael, 98, 98, 112, 125 Warme, David, 113
Triki, Chefi, 38, 39 Waterer, Hamish, 68, 68
Troeltzsch, Fredi, 128 Wedin, Per Åke, 34, 35, 87
Tseng, Paul, 76, 103 Wei, Hua, 107
Tsuchiya, Takashi, 96, 161 Weidner, Petra, 150
Tutuncu, Reha, 105, 110 Weintraub, Andres, 46
Tymofyeyev, Yevgen, 157 Weiskircher, Rene, 71
Uchoa, Eduardo, 33, 37, 47, 59 Weismantel, Robert, 47, 114
Ueno, Takayuki, 65 Weiss, Gideon, 137
 185

Wenger, Klaus Michael, 58, 58

Wessäly, Roland, 160, 161, 161
Westphalen, Markus, 91, 91
Wiegele, Angelika, 138
Williams, H.Paul, 39, 40, 106, 122, 136
Wilson, John, 126
Winter, Pawel, 139
Wolberg, W. H., 81
Wolkowicz, Henry, 93, 93, 107
Wolsey, Laurence A., 68, 114
Wood, Bradford, 81
Wood, Kevin, 52
Woodruff, David, 52
Wright, Stephen, 35, 61, 120
Wu, Y., 161
Wu, Yue, 122
Wynter, Laura, 50
Xavier, Adilson Elias, 130
Xavier, Eduardo C., 149
Xia, Cathy H., 50
Xia, Yu, 76, 80
Xu, Huifu, 148, 148
Xu, Yan, 51
Yamamoto, Yoshitsugu, 77
Yamashita, Makoto, 81, 82
Yan, Hong, 154
Yang, Boting, 57
Yang, Zaifu, 53
Yarmish, Gavriel, 37
Ye, Yinyu, 96, 96, 108, 108
Yen, Joyce, 52
Yeo, Anders, 84
Yildiz, Hakan, 98
Yokoo, Makoto, 119
Yoshise, Akiko, 134
Yuan, Di, 147, 147
Yuan, Ya-xiang, 130
Zabala, Paula Lorena, 55, 112
Zabinsky, Zelda B., 52, 63
Zachariasen, Martin, 139
Zak, Eugene, 59
Zakeri, Golbon, 96
Zakovic, Stanislav, 34, 34
Zaragoza Martinez, Francisco Javier, 99
Zhadan, Vitalij, 147, 147
Zhang, Fan, 50
Zhang, Hongchao, 142
Zhang, Jian-Zhong, 42, 130
Zhang, Jiawei, 108
Zhang, Shuzhong, 143
Zhang, Yin, 41
Zhao, Gongyun, 80
Zhao, Yonggan, 119
Ziemba, William, 119
Ziliaskopoulos, Thanasis, 108
Zilinskas, Antanas, 48, 48
Zillober, Christian, 141
Zimmermann, Uwe, 42, 42
Zinchenko, Yuriy, 50
Zoltan, Cristina, 99
Zuckerberg, Mark, 34
Zuluaga, Luis, 130, 131
Zymolka, Adrian, 32, 125
186 List of Exhibitors

List of Exhibitors
The following companies and organizations have exhibition stands at the conference.

Dash Optimization
DORS - Dansk Selskab for Operationsanalyse
GAMS Development Corp.
ILOG
Kluwer Academic Publishers
Princeton University Press
SIAM
Springer Verlag
Taylor and Francis