P
T
I
M
A
Mathematical Programming Society Newsletter
DECEMBER2007
Rapid Development of an Opensource
Minlp Solver with COINOR
Pierre Bonami, John J. Forrest, Jon Lee and Andreas Wachter
Abstract
We describe the rapid development of
Bonmin (Basic Opensource Nonlinear
Mixed INteger programming) using the
community and components of COIN
OR (COmputational INfrastructure
for Operations Research).
1. Introduction
As famous examples like Linux have
proven, opensource communities
can produce rapidlydeveloping high
quality software, often competitive with
expensive commercial solutions. COIN
OR (COmputational INfrastructure
for Operations Research), originally
founded by IBM in 2000 and a nonprofit
foundation since 2004, provides such
a platform for us optimization folks.
Since its conception, a number of
optimizationalgorithm implementations
and related software projects have been
contributed to COINOR. From a web
server hosted by INFORMS, at the URL
www.COINOR.org, COINOR offers
online repositories for sharing source code,
web sites with information, documentation
and bug tracking systems, as well as mailing
lists for users and developers. Many of
the optimization codes have been used
extensively by academics and in commercial
products.
Currently, COINOR hosts many
optimizationrelated codes:
Clp, DyLP, and Vol are solvers
for Linear Programming
Bcp, Cbc, Cgl, Cops, and Symphony
are solvers, frameworks and tools
for (serial and parallel) Mixed
Integer Linear Programming,
75
Csdp, Dfo, Ipopt, LaGO, and Svm
QP are solvers for different types of
Nonlinear Optimization (semidefinite,
derivativefree, general largescale
interior point, global optimization, QP
solver for support vector machines)
CoinMP, FlopC++, GAMSlinks,
NLPAPI, Osi, OS and Smi are problem
modeling and interface tools.
Even more, COINOR has become the
platform that facilitates the collaboration
of optimization researchers to start entirely
new projects, exploring and developing new
algorithmic ideas, by providing both a sound
software basis that can be reused, as well as
the technical forum to coordinate the effort.
In order to clarify a common
misconception, we want to emphasize that
the term "open source" does not simply refer
to software that is made available to some
users in sourcecode format. According
to the definition put forth by the non
profit corporation Open Source Initiative
(OSI) [26], a number of conditions must
be met for a software license to be an
"opensource" license, including the right
for free distribution and usage (including
for commercial purposes), as well as the
right to modify the code. There are many
OSIcertified opensource licenses. The
one preferred on COINOR is the CPL
(Common Public License [12]), which was
chosen for its very liberal terms.
In this article, we tell the ongoing success
story of one such project. In 2004, IBM and
Carnegie Mellon University initiated a joint
study aimed at Mixed Integer Nonlinear
Programming (MINLP), with the goal of
releasing the resulting software under the
CPL on COINOR. In 2006, within a
new Open Collaborative Research (OCR) )
Georg Nenau' Cotiuin 5 MP Chi' Opim Coum 6 Prze and Cofrne
DECEMBER 2007
initiative at IBM, this relationship with
CMU was extended. This OCR program
builds on the Open Collaboration Principles
established in 2005 by IBM and eleven
other institutions to accelerate innovation
and foster opensource research.
Besides the joint activity of CMU and
IBM, the subject of MINLP has received a
lot of recent attention. For example,
The 2006 BealeOrchard Hays Prize
for Excellence in Computational
Mathematical Programming was
awarded by the Mathematical
Programming Society to Nick Sahinidis
and Mohit Tawarmalani for their
work on global optimization which
culminated in their software Baron [3,
27]; global optimization is one of the key
approaches to MINLP, and Baron can
be directly applied to MINLP problems.
The 2006 CORE Lecture Series
consisted of a series of lectures by
Robert Weismantel on MINLP.
Our goal was to develop algorithms
and software for this important class of
problems. We undertook to carry this out in
the context of COINOR. Our original goal
was to use existing software components of
COINOR to produce a simple NLPbased
branchandbound code aimed at MINLP
problems that have convex relaxations. As
we progressed, we expanded our goal and
developed instantiations of several MINLP
methods, aimed mostly at problems having
convex relaxations. We were able to carry
this out rather quickly, as we had at our
disposal the source code to manage a
branchandbound tree and to solve MILP
and NLP subproblems.
In Section 2, we describe the setting
of MINLP. In Section 3, we discuss the
COINOR NLP solver Ipopt and the
COINOR MILP solver Cbc, which were
the starting points of our implementation.
In Section 4, we describe the capabilities
of our MINLP solver Bonmin (Basic
Opensource Nonlinear Mixed INteger
programming) and how we constructed
it using other components of COINOR.
Finally, in Section 5, we discuss ongoing
and future plans for Bonmin.
2. MINLP
We consider MINLP problems of the form
min f(x)
g(x) < 0,
x I x ,
where: I ">R and g: I > I are
twice continuously differentiable.
MINLP is the mother of all deterministic
optimization problems. Many business
and engineering optimization problems
are naturally formulated as mathematical
programs involving both discrete and
continuous variables, as often decisions
have to be made that involve making
discrete decisions (e.g., yes/no) and setting
continuous levels (e.g., flowrates). Natural
phenomena often involve relationships
between variables which, while often
smooth, may be rather nonlinear. Accurate
models need to allow for all of these.
A common approach to such problems
is to reformulate them as Mixed Integer
Linear Programming (MILP) problems,
usually through the use ofpiecewiselinear
functions approximating the true functions.
In the presence of nonconvexity, this leads
to the introduction of additional binary
variables or nonstandard branching rules.
The inaccuracies that are introduced and
the added computational effort induced
by accommodating the piecewiselinear
functions may make the value of such a
solution approach rather limited.
Another approach is to relax the MINLP
to an NLP, apply an NLP solver, and then
use some heuristic method to reach a nearby
feasible point that respects the integrality
restrictions. Just as for MILP, this technique
is not very robust, often leading to very poor
solutions.
The alternative to methods that are based
strictly on either MILP and NLP is to
develop methods that are directly aimed at
MINLP formulations. In effect, one shifts
some of the burden from the modeler to
the solution software. Still, just as we know
from MILP and NLP, there are always
significant issues that cannot be ignored
by a modeler who seeks to solve difficult
instances.
It is useful to first consider methods
that are aimed at MINLP problems
having convex relaxations. In this sub
domain, there are natural methods that
generalize MILP methods (e.g., branch
andbound, using NLP subproblems) as
well as more sophisticated methods like
outer approximation (see [13]), generalized
Bender's decomposition (see [18]) and hybrid
methods (see [8], for example). Even in this
subdomain, problems can be much harder
than the MILP counterparts. A main reason
is that in the space of the natural variables,
the solution of a relaxation may be well
inside the interior of the feasible region,
and so the cutting planes that have been so
successful for MILP may have very limited
value (see [24], for example).
If the relaxation of the MINLP problem
is not convex, we are really in a different
world. There are several approaches. There
are ad hoc heuristics, which suffer from
being well suited to only very specific
models (see [23], for example). There is the
possibility of taking methods designed for
problems having convex relaxations and
heuristically adapting them to problems
that do not satisfy this restriction (e.g.,
the branchandbound option of Bonmin;
see [7]). There is the approach of global
optimization typified by the method of
spatial branchandbound wherein one
repeatedly subdivides the feasible region,
with the goal of producing better and
better convex underestimators (e.g., Baron;
see [28]). Finally, there is the approach of
systematically remodeling a broad subclass
ofMINLP problems as MILPs using
specialized techniques (e.g., see [16] for such
a method using linked ordered sets)
3. Building Blocks
As MINLP combines the worlds of
MILP and of NLP, a logical start for
the Bonmin project was to look at the
existing codes and libraries in COIN
OR related to MILP and NLP and to
see how they could be integrated.
Cbc (Coin Branch and Cut) is a powerful,
industrialstrength branchandcut code
for MILP problems. It is a component
of COINOR, and it uses several other
components of COINOR. It uses Clp (the
Coin Linear Programming solver) through
the Osi (Open Solver Interface) library. The
Osi is a layer used to insulate an application
that makes use of an LP solver from the
specific LP solver employed. Clp includes
primal and dual simplex solvers as well as
an interiorpoint method. As a fullfledged
branchandcut code, Cbc strengthens
relaxations through use of the COIN Cgl
(Cut Generation Library), accommodates
different nodeselection and branching rules
(e.g., via pseudocosts and strong branching),
and allows for SOS Type1 and Type2
modeling and branching. Cgl includes a
0SP T I MA 5
Date: November 2, 2007
PAGE 2
DECEMBER 2007
wide variety of cuttingplane generators
(including: knapsack cover, odd hole, flow
cover, lift and project, Gomory, reduce and
split, and MIR). Our interest in Cbc was
twofold. We sought to reuse infrastructure
that we would need (for example,
maintaining any needed branchandbound
enumeration trees, handling the branching
and fathoming decisions, etc.), but we also
simply wanted the Cbc functionality for
solving any MILP subproblems that we
might create.
To address the nonlinear aspect of
the problem, the Ipopt NLP solver from
COINOR was used. Ipopt implements
an algorithm for continuous nonlinear
optimization (see [29]). Its interior point
approach, together with a thirdparty sparse
linear algebra routine such as MA27 [20],
allows Ipopt to efficiently solve largescale
problems with up to millions of variables.
Global convergence (i.e.,convergence
to a stationary point from any starting
point) is achieved by a filter linesearch
procedure. Here, nonconvexities in the
problem formulation are tackled by possibly
regularizing the Hessian matrix of the
Lagrangian to ensure that the search
directions are descent directions in the
appropriate sense, so that convergence to
nonminimizers is less likely. Since Ipopt
is using second derivatives, fast local
convergence can be achieved. However, if
second derivatives are not available, they are
approximated by a quasiNewton approach.
The Ipopt code was originally written in
Fortran, but a complete rewrite in C++,
following an objectoriented approach, was
released on COINOR in 2004. The Ipopt
library includes a few utility functionalities,
such as output logging and algorithmic
options handling, which can be used by
other codes as well.
All of these basic components from
COIN are being actively developed and
have a wide user basis.
4. Bonmin
Using the building blocks described in
Section 3, we began to develop Bonmin
(Basic opensource nonlinear mixed
integer programming). In the first phase,
we sought to implement an MINLP
code targeted at problems with convex
relaxations. A straightforward approach
employing available components from
COINOR uses the Cbc framework to
120000
100000
80000
60000
40000 .
O'/2004
own t'l
01120061
20000 .
1?.;0i6
0
Bonmm
19,2006
11. 200
a ^^
Cbc Op Cgl CoinUtils Ipopt 0
Figure 1. Code reuse by Bonmin and growth
organize a branchandbound search,
where now all the nodes are NLPs. The
nonlinear relaxations are solved by Ipopt,
which has been connected to Cbc via an
Osi solver object. Attempts were also made
to handle nonconvex problem instances
by some heuristics, such as negative cut
off, multiple local optimization runs for
the node NLPs from randomly chosen
starting points, and continuing branching
even at (locally) infeasible nodes. The
integration of these features, which helped
to obtain better solutions, would not have
been possible if the source code of the
components had not have been available.
Another popular strategy for solving
convex MINLPs uses Outer Approximation
(OA). The basic idea is to approximate the
original nonlinear problem by a sequence
of MILPs, where, successively, linear cuts
are added based on linearizations of the
nonlinear functions. The OA algorithm
alternatingly solves the current MILP
approximation and an NLP (the MINLP
with all integer variables fixed). It can be
shown that this procedure with appropriate
linearization points converges to the optimal
solution of the MINLP in a finite number
of iterations (see [13]). Using Cbc with
Clp as the MILP solver and Ipopt as the
NLP solver, we could implement the OA
algorithm in Bonmin with existing COIN
OR components.
An intermediate hybrid approach between
a pure branchandbound and a pure OA
algorithm has also been developed and
implemented in Bonmin [8].
Trunk
Stable
1ii ,.J. I
iap Couenne
The success of the project is facilitated
by the fact that all essential components
were available as opensource code, so
that they could be changed (and changes
could be contributed back to the project
maintainers), and by the objectoriented
design of the codes. Figure 1 indicates how
Bonmin makes significant reuse of COIN
OR components. The ordinate indicates
the number of lines of code for the various
components. In the totals, we do not count
parts of some codes that we definitively
do not use. The bar for Bonmin refers to
the new code written specifically for the
MINLP project. Besides demonstrating
how Bonmin is reusing COINOR
components, Figure 1 also indicates how all
of these components (including Bonmin) are
growing. For each component, we indicate
the number of (relevant) lines of code as of
the mostrecent stable release on COIN
OR, as well as the current incremental size
(as of 21 May, 2007) in the \trunk" (i.e.,
development branch) of each project.
Bonmin can be accessed as a C++ library,
as a standalone code that can accept
an input file in .nl format [17] from the
mathematicalprogramming modeling
and scripting language Ampl [1], and via
the NEOS Server for Optimization [25].
Similar to Ampl, Gams is another widely
used scripting and modeling language for
manipulating mathematical programs.
The recently added COINOR project,
GAMSlinks, managed by Stefan Vigerske,
provides an interface to Bonmin (as well as
the COIN OR solvers Ipopt and Cbc).
10S T I MA 5
PAGE 3
DECEMBER 2007
5. Ongoing and future work
We are presently pursuing several
initiatives with the goal of extending
the applicability of our work.
One ongoing project involves using the
parallel MILP framework of the COIN
OR code Bcp (BranchCutand Price) to
develop a parallel version of Bonmin. Our
goal is to exploit parallelism on a small scale
as well as on massively parallel architectures
(like BlueGene; see [5]). Alternatively,
we might consider use of the relatively
new COINOR Cops (COINOR Open
Parallel Search) framework, to manage the
parallelization. The use of objectoriented
design for the basic Bonmin code facilitates
the easy adaption into different search tree
management frameworks (e.g., Cbc, Bcp,
Cops).
Another project jointly explored with
Carnegie Mellon University, involves
the use of globaloptimization ideas and
software. We are experimenting with the
usage of rigorous convex underestimators
and spatial decomposition. The resulting
software will be available on COINOR as
Couenne (Convex Over/Under ENvelopes
for Nonlinear Estimation). Similarly,
we are exploring the possibility to reuse
some globalization cuts from the recent
COIN addition LaGO (Lagrangian Global
Optimizer), authored by Ivo Nowak and
Stefan Vigerske. The LaGO code itself is a
MINLP solver that makes use of Clp, Cgl,
and Ipopt, and is based on some heuristics.
However, for many MINLP instances,
it is impractical to expect to be able to
reliably find the global optimum. So, an
extremely important area is the development
of generalpurpose heuristics for MINLP.
Already, in the context of the development
of the Bonmin project, some success in
this direction has been achieved (see [6]).
Still, there is much more work to do in this
direction.
We are also investigating the use of other
NLP solvers (e.g., activeset approaches like
that of FilterSQP [14]) as an alternative to or
in conjunction with Ipopt.
On the theoretical side, we are exploring
the use of curvature information from the
NLP in making branching decisions in the
branchandbound MINLP framework.
We are already finding this approach to be
useful for certain nonlinear facilitylocation
problems as well as certain portfolio
optimization problems.
We are also investigating the
strengthening of MINLP formulations,
following the success of such methods for
MILP. Already we have some success for
structured problems (see [19]) and we are
currently exploring more generic techniques.
Finally, there are other components of
COINOR that we may consider taking
advantage of. For example, Csdp (a
semidefiniteprogramming code authored by
Brian Borchers) and Dfo (a derivativefree
NLP code authored by Katya Scheinberg)
could be exploited in an attempt to solve
broader classes of MINLP problems than we
currently consider.
It is our fondest desire that this entire
manuscript be already obsolete by the
time that you are reading this. Indeed, for
our project to be succeeding, that should
be the case. The very latest information
concerning our efforts should be available
from the Bonmin web page [9] and
the IBM MINLP web page [21].
6. Acknowledgments
Thanks are due to the others who
have participated in the CMU/IBM
collaboration, namely Pietro Belotti, Larry
Biegler, Andrew Conn, G&rard Cornudjols,
Ignacio Grossman, Carl Laird, Andrea
Lodi, Franccois Margot and Nick Sawaya.
References
1. Ampl, www.ampl.com
2. Ampl Solver Library, www.
netlib.org/ampl/solvers
3. Baron, archimedes.scs.uiuc.edu/
baron.html
4. Bias, www.netlib.org/blas
5. BlueGene, IBM Journal of
Research and Development,
Volume 49, Number 2/3,
2005. www.research.ibm.
com/journal/rd4923.html
6. P. Bonami, G. Cornudjols,
A. Lodi and F. Margot. A
Feasibility Pump for Mixed
Integer Nonlinear Programs.
IBM Research Report
RC23862, 2006.
7. P. Bonami and J. Lee, BONMIN
Users' Manual. projects.
COINOR.org/Bonmin
8. P. Bonami, A. Wachter, L.T.
Biegler, A.R. Conn, G.
Cornudjols, I.E. Grossmann,
C.D. Laird, J. Lee, A. Lodi,
F. Margot and N. Sawaya,
An algorithmic framework
for convex mixed integer
nonlinear programs. IBM
Research Report RC23771,
October 2005. To appear in:
Discrete Optimization.
9. Bonmin, projects.COIN
OR.org/Bonmin
10. C. Bragalli, C. D'Ambrosio, J.
Lee, A. Lodi and P. Toth, An
MINLP model and solution
method for a water network
optimization problem. In
Algorithms ESA 2006, Y.
Azar and T. Erlebach, Eds.,
pages 696707. Springer,
2006.
11. COINOR, www.COIN
OR.org
12. Common Public License, www.
opensource.org/licenses/
cpll.0.php
13. M. Duran and I.E. Grossmann.
An outerapproximation
algorithm for a class of mixed
integer nonlinear programs.
Mathematical Programming,
36:307{339, 1986.
14. FilterSQP, wwwunix.mcs.anl.
gov/leyffer/solvers.html
15. C.A. Floudas, Deterministic
global optimization, Kluwer
Acad. Publ., Dordrecht, 2000.
16. J. Forrest and J. Lee (in
preparation).
17. D.M. Gay, Writing .nl files,
Technical Report 20057907P,
Sandia National Laboratories,
Albuquerque, NM, 2005.
www.cs.sandia.gov/dmgay/
nlwrite.pdf
18. A.M Geoffrion. Generalized
Benders decomposition.
Journal of Optimization
Theory and Applications,
10:237{260, 1972.
19. O. Gunliik, J. Lee and R.
Weismantel. MINLP
Strengthening for Separable
Convex Quadratic
TransportationCost UFL,
IBM Research Report
RC24213, March 2007.
20. Harwell Subroutine Library,
www.cse.clrc.ac.uk/nag/hsl
21. IBM MINLP, domino.research.
ibm.com/comm/research
projects.nsf/pages/minlp.
index.html
22. Lapack, www.netlib.org/lapack
23. J. Lee, In situ column
generation for a cutting
stock problem. Computers
& Operations Research,
34(8):23452358, 2007.
24. J. Lee, Mixed integer nonlinear
programming: Some modeling
and solution issues, to appear
in the IBM Journal of
Research and Development,
2007.
25. NEOS Server for Optimization,
wwwneos.mcs.anl.gov
26. Open Source Initiative, www.
opensource.org
27. N. Sahinidis and M.
Tawarmalani, A polyhedral
branchandcut approach
to global optimization,
Mathematical Programming,
103(2), pp. 225249, 2005.
28. M. Tawarmalani and N.V.
Sahinidis, Convexification )
0SP T I MA 5
PAGE4
DECEMBER 2007
Optimization Research Symposium Recognizing Professor George
Nemhauser's Contributions to the Field of Operations Research
July 27, 2007 marked the 70th birthday
of George Nemhauser. George has been
a longtime member of MPS, served as its
president from 19891992 and cochaired
the ISMP at Georgia Tech in 2000. To
mark the occasion of George's birthday, a
twoday symposium was held in his honor
at the Georgia Institute of Technology in
Atlanta, GA. The symposium included
a mix of serious scientific lectures and,
sometimes less serious presentations on
George's career and his contributions to
various fields and institutions. There
were 17 invited speakers and close to 100
attendees. The symposium was organized
by Mike Ball and Martin Savelsbergh.
The Thursday afternoon program
included one session on George's scientific
contributions to four disciplines: integer
programming, presented by Bob Bixby,
combinatorial optimization presented by
Gerard Cornudjols, airline optimization
presented by David Ryan and sports
scheduling presented by Mike Trick.
A second session covered George's
contributions to four institutions: Johns
Hopkins University presented by Mike
Thomas, Cornell University presented
by Dave Goldsman, CORE presented
by Lawrence Wolsey and Georgia Tech
presented by John Jarvis. There was a special
bonus presentation by Martin Gritschel and
the evening dinner included remarks by Bill
Pulleyblank, Don Ratliff and Mike Ball.
Several presentations included gifts such
as a selection of red and white New Zealand
wines by Ryan and Trick and the official
and global optimization in continuous and
mixedinteger nonlinear programming,
Kluwer Acad. Publ., Dordrecht, 2002.
29. A. Wachter and L. T. Biegler, On the
Implementation of an InteriorPoint Filter
LineSearch Algorithm for LargeScale
Nonlinear Programming, Mathematical
Programming 106(1), pp. 2557, 2006.
IBM T.J. Watson Research Center,
Yorktown Heights, New York. USA
email address: fpbonami,jjforre,jonlee,a
ndreasw@us.ibm.com
naming of a prime number in George's
honor by Groetschel. While Gerard
Cornuejols presented George with one
of his integer programming homework
assignments from Cornell, it did come
with the request that George regrade a
problem he had incorrectly graded 30 years
ago. Although Bob Bixby demonstrated
that certain mixed integer cuts developed
by Nemhauser and his coauthors were the
basis for a substantial speed improvement
in Cplex, no associated stock options
were provided. The institutional
presentations chronicled George's significant
contributions to four great institutions and
also demonstrated George's ability both
to solve and create problems. The dinner
presentations were lighter in tone and ended
with a presentation of a compilation of the
title pages and table of contents from the 50
PhD dissertations supervised by George.
The Friday program included five
scientific presentations: "Retail Assortment
Optimization" by Marshall Fisher, "Airline
Optimization: If we are so good at it, why
are things so bad?" by Cindy Barnhart,
"Corner Polyhedra and TwoEquation
Cutting Planes" by Ralph Gomory,
"Minimizing with Concave Costs?" by Tom
Magnanti, and "Integer Programming and
BranchWidth" by Bill Cunningham. Each
of the speakers has had close ties to George
over the years but more significantly the
body of results presented had numerous
connections to his work.
Overall the symposium proved to be
both intellectually stimulating and, at
times, highly entertaining. By examining
the contributions of a great individual in
this way, one can gain an appreciation for
the substantial positive impact one can
have through a career in mathematical
programming. The symposium was
certainly a fitting tribute to George
Nemhauser for providing such a positive role
model to our community.
Michael O. Ball
Orkand Corporation Professor
of Management Science
Robert H. Smith School of Business
and Institute for Systems Research
University of Maryland
College Park, MD 20742
1 PT I M A75
PAGE 5
DECEMBER 2007
MPS Chair's Optima Column 3rd issue of 2007 By Steve Wright
I am delighted to contribute the inaugural
MPS Chair's column for Optima. As
you can see, Optima is being published
on a regular schedule once again. Under
the leadership of editor Andrea Lodi,
the design of our Society's newsletter is
being reworked, and we look forward to
exciting innovations in the months ahead.
MPS has recently had a large increase
in membership due to the 2006 ISMP in
Rio de Janeiro and the 2007 ICCOPT in
Hamilton. Renewal notices for 2008 have
recently gone out. I hope that our new
members will stay involved in the Society
and help it play an ever more important role
in sustaining and building our international
community of optimizers and mathematical
programmers.
Preparations for the 2009 ISMP in
Chicago gathering momentum under
general chair John Birge. The symposium
will be held in the week of August 2328,
2009, with most sessions taking place at the
Gleacher Center (the downtown campus
of the University of Chicago Business
School) and the nearby Marriot Hotel.
Both locations are close to Millennium
Park, around which the recent revival of
downtown Chicago has been centered. The
meeting will include a commemoration of
the 60th anniversary of the "zeroth" ISMP,
which was held in Chicago in 1949. I hope
that all of you will be able to take advantage
of Chicago's pleasant summer weather and
its location as a global air hub to attend
a wonderful symposium and witness
the exciting recent changes to this great
American city.
We welcome new editors of our Society's
journal Mathematical Programming: Kurt
Anstreicher (Series A) and Danny Ralph
(Series B). The number of published issues
will increase from nine to ten in 2008 and
later years, a reflection of the robust health
of the journal. As always, subscription to the
print version of the journal is included in
MPS membership. The online version, along
with complete archives of the journal and
Mathematical Programming Studies, is also
free to members. We also welcome Philippe
Toint, the new editor of the MPSSIAM
book series. Please contact Philippe or one of
his editorial board members with any book
writing ideas you may have.
I take this opportunity to remind you of
the generous discounts available to MPS
members on all books published by Springer
and SIAM. A new SpringerToken will be
issued in 2008; look for it in your email
inbox.
Another snippet of news is that
MPS recently became a member of the
International Council of Industrial and
Applied Mathematics (ICIAM), a society of
societies whose main function is to organize
a congress every four years. I was privileged
to attend the most recent congress in Zurich
along with 3100 colleagues. The next
congress is planned for Vancouver in 2011.
MPS members have already played key roles
in the last two congresses, as organizers and
speakers, and our new status will formalize
and solidify our involvement.
I wish to thank the organizers of several
recent meetings sponsored by MPS in
2007. IPCO continued its tradition as a key
meeting of the integer programming and
combinatorial optimization community
with a highly successful meeting in June at
Cornell. Thanks to David Williamson and
Matteo Fischetti and all involved with the
organization and selection of papers for this
event. ICCOPT at McMaster University
in August was an outstanding success,
under general chair Tamas Terlaky and
his large and excellent cast of organizers.
The 420 attendees enjoyed good weather,
an excellent venue and program, and a
wonderful roster of social events.
Finally, I thank the officers and
councilors of MPS whose terms have ended
recently. Former Chair Rolf Mohring
continues as ViceChair, so we will
continue to have the benefit of his wisdom
and experience. Karen Aardal is stepping
down as Chair of the Executive Committee.
Karen has served MPS tirelessly for over
ten years as Councillor and Publications
Committee Chair, in addition to her most
recent role. Bill Cook did great work as
editor of Mathematical Programming Series
B, then Series A, maintaining the high
standards of our journal over a 7year span.
Special thanks go to these three colleagues!
Application for Membership
I wish to enroll as a member of the Society.
My subscription is for my personal use and not for the benefit of any library or institution.
O I will pay my membership dues on receipt of your invoice.
E I wish to pay by credit card (Master/Euro or Visa).
CREDIT CARD NO.
EXPIRATION DATE
FAMILY NAME
MAILING ADDRESS
Mail to:
Mathematical Programming Society
3600 University City Sciences Center
Philadelphia, PA 191042688 USA
Cheques or money orders should be
made payable to The Mathematical
Programming Society, Inc. Dues for 2007,
including subscription to the journal
Mathematical Programming, are US
80. Retired are 50. Student applications:
Dues are 25. Have a faculty member
verify your student status and send
application with dues to above address.
Faculty verifying status
Institution
0S TIM 75
TELEPHONE NO. TELEFAX NO.
EMAIL
SIGNATURE
PAGE
1 DEEMEI 200 PAGE
Prizes and Conferences by Katya Scheinberg
The Second Mathematical Programming
Society International Conference on
Continuous Optimization (ICCOPT II)
was held at McMaster University on August
1316, 2007. Even though only the second
of its kind this conference was a smashing
success. Instead of writing about it ourselves,
we let the professional speak for us.
Here is a excerpt from the
Hamilton Spectator:
"There is an endless number of optimization
applications", Terlaky says. Specialists
develop algorithms and use computer
networks to simulate, model and optimize
solutions to various problems, as well as
design new products and services. "Mostly,
you don't see it because it's behind the
scenes", Terlaky says. "But without it,
everything would be more expensive and
less efficient." It all has to do with a field of
research called continuous optimization, or
more formally, mathematical programming
 finding the most efficient and effective
means to an end. And big players like
NASA, Boeing, General Motors, which used
to have their own mathematics department,
Microsoft and Exxon Mobil rely on it to
improve their products or in designing new
ones. The companies have representatives
in Hamilton this week to attend a major
international conference on optimization
at McMaster. The conference has also
attracted NASA representatives, PhD
students, professors and researchers from
around the world, representing Ivy League
institutions such as Princeton, Oxford,
Stanford and MIT. About 450 specialists
are attending from 36 countries, including
New Zealand, Singapore, Brazil and Chile.
The Ivy Leagueness of Oxford, Stanford and
MIT is debatable, but everything else is true.
From ourselves we can add that with
22 streams on numerous subjects there
was never a dull moment at this well
attended conference. The quality of talks
and attendance was high and discussions
lively. One of the highlights of the
conference was the Young Researchers
Competition in continuous optimization.
The committee consisted of Kees Roos
(chair), Arnold Neumaier, Levent
Tuncel, Yinyu Ye and Akiko Yoshise.
Out of 23 submissions three
finalists were selected:
Alexandre Belloni (Duke University),
Mung Chiang (Princeton University)
Eissa Nematollahi (McMaster University)
who presented their work in a special
session during the conference. The prize
was awarded to Alexandre Belloni for
his paper "Norminduced Densities and
Testing the Boundedness of a Convex Set".
Here is how Kees Roos presented the
prize at the conference banquet:
Ludwig Wittgenstein, the most famous
logician and philosopher of language,
was awarded a PhD at the university of
Cambridge (UK), in 1929. As his PhD
thesis served the small booklet ''Tractatus
logicophilosophicus', a work that had been
in print already for seven years and that
was regarded at that time by many as a
philosophical classic. The examiners were
George E. Moore and Bertrand Russell.
The PhD defense was set for 18 June,
1929, and was conducted with an air of
farcical ritual. As Russell walked into the
examination room with Moore, he smiled
and said: 'I have never known anything
so absurd in my life'. The examination
began with a chat between old friends.
Then Russell, relishing the absurdity of the
situation, said to Moore: 'Go on, you've
got to ask him some questions you're the
professor'. Then followed a short discussion
in which Russell advanced his view that
Wittgenstein was inconstant in claiming to
have expressed unassailable truths by means
of meaningless propositions. He was, of
course, unable to convince Wittgenstein,
who brought the proceeding to an end
by clapping each of his examiners on the
shoulder and remarking consolingly: 'Don't
worry, I know you'll never understand it'.
In his examiner's report, Moore stated:
'It is my personal opinion that Mr.
Wittgenstein's thesis is a work of a genius,
but, be that as it may, it is certainly well
up to the standard for the Cambridge
degree of Doctor of Philosophy'. (See
Ray Monk, Ludwig Wittgenstein. The
duty of genius. Penguin Books, 1990).
The last sentence, slightly modified, well
expresses the committee's feelings on the
participants in this year's Young Researcher
Competition. Most of the 23 submissions
were of high quality. Also, the submissions
of the three finalists, Alex Belloni, Mung
Chiang and Eissa Nematollahi, who
presented their papers this afternoon,
where convincing enough for us to agree
that they are up to the standard in the
optimization community for being finalists
in the competition. But our committee
also agreed unanimously that the work of
Belloni certainly meets the standard for
being the winner of the competition.
Uneasy Relations by
Michael Bartholomew
Biggs: a Book Review
by Andrea Lodi
Uneasy Relations contains poems mainly
in haiku form which play with ideas
from computational mathematics. Non
specialists may choose to murmur them
as zenlike meditations. For readers of a
curious disposition, notes are provided
that may be no less informative than
those at the end of The Waste Land.
Our favorite?
AUXILIARIES
Lagrange Multipliers
These are our shadows,
dual personalities,
who know what we don't 
the consequences
of pushing our boundaries
or pulling them in.
Michael BartholomewBiggs has spent most
of his working life as a mathematician at
the University of Hertfordshire. As the
author himself says, his poetic ambitions
arrived as part of a midlife crisis and
this, his fourth poetry chapbook, is an
attempt to unite both halves of his brain.
Michael BartholomewBiggs, Uneasy
Relations, Hearing Eye London (UK),
2007. ISBN: 9781905082315.
DECEMBER 2007
PAGE 7
O P T I M A
MATHEMATICAL PROGRAMMING SOCIETY
UF UNIVERSITY of
UFFLORIDA
Center for Applied Optimization
401 Weil
PO Box 116595
Gainesville, FL 326116595 USA
FIRST CLASS MAIL
EDITOR:
Andrea Lodi
DEIS University of Bologna,
Viale Risorgimento 2,
I 40136 Bologna, Italy
email: andrea.lodi@unibo.it
COEDITORS:
Alberto Caprara
DEIS University of Bologna,
Viale Risorgimento 2,
I 40136 Bologna, Italy
email: acaprara@deis.unibo.it
Katya Scheinberg
IBM T.J. Watson Research Center
PO Box 218
Yorktown Heights, NY 10598, USA
katyas@us.ibm.com
FOUNDING EDITOR:
Donald W. Hearn
PUBLISHED BY THE
MATHEMATICAL PROGRAMMING SOCIETY &
University of Florida
Journal contents are subject to
change by the publisher.
