GdR blog

RSS feed

Latest news

MASCOT-NUM 2018 : demandez le programme !

Published on: Wed, 14 Mar 2018 09:10:00 +0100 by Julien Bect

Le programme complet des journées du GdR est consultable sur le site :

Rendz-vous la semaine prochaine à Nantes !

8th GECCO workshop on Black-Box Optimization Benchmarking (BBOB 2018)

Published on: Mon, 12 Mar 2018 21:58:00 +0100 by Julien Bect
Dear colleagues,

Please find below the announcement for the next GECCO workshop on Real-Parameter Black-Box Optimization Benchmarking (BBOB 2018).

Black-box optimization problems occur in many application areas and several types of optimization algorithms have been proposed for this class of problems. One of the main questions when having to solve a black-box problem in practice is to decide about which algorithm (type) to use.

Most of the time, we fall back on numerical benchmarking of such algorithms to understand and recommend algorithms but this is a non-trivial and tedious task. In recent years, the Comparing Continuous Optimizers platform (COCO, has been developed to free algorithm designers and users of optimization software alike from the tedious tasks of setting up experiments and analyzing the performance data of algorithms by automatizing the benchmarking process. Three black-box optimization benchmarking (BBOB) suites have been made available so far with (jointly) more than 200 algorithm data sets available to the optimization community.

Though the basis of the BBOB workhop series is the COCO platform, we are looking forward to any submission related to black-box optimization benchmarking of continuous optimizers in the widest sense, for example papers that:
  • discuss, compare, and improve upon any benchmarking methodology for continuous optimizers such as design of experiments, performance measures, presentation methods, benchmarking frameworks, test functions...
  • describe and benchmark new or not-so-new algorithms on one of the following testbeds,
    • bbob testbed with 24 noiseless single-objective functions
    • bbob-noisy with 30 noisy single-objective functions
    • bbob-biobj with 55 noiseless bi-objective functions
  • compare new or existing algorithms from the COCO/BBOB database,
  • analyze the data obtained in previous editions of BBOB.

In the latest COCO release (version 2.2) we offer quite some new features that will improve your efficiency when using COCO and that will allow in particular to focus on using and investigating the already existing data sets:
  • automated download of algorithm data sets via cocopp.data_archive,
  • updated convergence plots,
  • LaTeX tables are reorganized to have better control within LaTeX,
  • noisy LaTeX template moved from the svn repository to github and was updated,
  • postprocessing cocopp module now python 3 compatible,
  • all figures (and their inclusion to the LaTeX templates) have been updated  and beautified in order to comply with the newest matplotlib version,
  • zip files are supported as input files.

To be notified about further releases of the COCO code and information related to the workshop, please register at

Don't hesitate to forward this message to possibly interested people
and all our apologies already if you received this file multiple

The BBOBies

with a continued focus on multi-objective problems

to be held as part of the

2018 Genetic and Evolutionary Computation Conference (GECCO-2018)
July 15-19, Kyoto, Japan

Organized by ACM SIGEVO

8th Real-Parameter Black-Box Optimization Benchmarking (BBOB-2018)
Submission Deadline: Tuesday, March 27, 2017
Register for news via
Source code: (scroll down for a quick start)

The Black-Box-Optimization Benchmarking (BBOB) methodology associated to the BBOB GECCO workshops has become a well-established standard for  benchmarking stochastic and deterministic continuous optimization  algorithms in recent years ( So far,  the BBOB GECCO workshops have covered benchmarking of blackbox optimization algorithms for single- and bi-objective, unconstrained  problems in exact and noisy, as well as expensive and non-expensive  scenarios. A substantial portion of the success can be attributed to the Comparing Continuous Optimization benchmarking platform (COCO) that builds the basis for all BBOB GECCO workshops and that  automatically allows algorithms to be benchmarked and performance data to be visualized effortlessly.

Like for the previous editions of the workshop, we will provide source code in various languages (C/C++, Matlab/Octave, Java, and Python) to benchmark algorithms on three different test suites (single-objective with and without noise a well as a noiseless bi-objective suite). Postprocessing data and comparing algorithm performance will be equally automatized with COCO (up to already prepared LaTeX templates for writing papers). As a new feature for the 2018 edition, we provide significantly easier access to the already benchmarked data sets such that the analysis of already available COCO data becomes simple(r).

Analyzing the vast amount of available benchmarking data (from 200+ experiments collected throughout the years) will be therefore a special focus of BBOB-2018. Given that the field of (multiobjective) Bayesian optimization received renewed interest in the recent past, we would also like to re-focus our efforts towards benchmarking algorithms for expensive problems (aka surrogate-assisted algorithms developed for limited budgets). Moreover, several classical multiobjective optimization algorithms have not yet been benchmarked on the bbob-biobj test suite, provided since 2016, such that we encourage contributions on these three following topics in particular:
  • expensive/Bayesian/surrogate-assisted optimization,
  • multiobjective optimization,
  • analysis of existing benchmarking data.

Interested participants of the workshop are invited to submit a paper (not limited to the above topics) which might or might not use the provided LaTeX templates to visualize the performance of unconstrained single- or multiobjective black-box optimization algorithms of their choice on any of the provided testbeds. We encourage particularly submissions about algorithms from outside the evolutionary computation community as well as any papers related to topics around optimization algorithm benchmarking.

For details, please see the separate BBOB-2018 web page at:

We encourage any submission that is concerned with black-box optimization benchmarking of continuous optimizers, for example papers that:
  •  describe and benchmark new or not-so-new algorithms on one of the above testbeds,
  • compare new or existing algorithms from the COCO/BBOB database,
  • analyze the data obtained in previous editions of BBOB, or
  • discuss, compare, and improve upon any benchmarking methodology for continuous optimizers such as design of experiments, performance measures, presentation methods, benchmarking frameworks, test functions, ...

Paper submissions are expected to be done through the official GECCO  submission system at the (hard) deadline on March 27, 2018.

In order to finalize your submission, we kindly ask you to fill in addition the form at where you are  supposed to provide a link to your data as well if this applies. To upload your data to the web, you might want to use offers uploads of data sets up to 50GB in size or any other  provider of online data storage. Please let us know briefly in the mandatory Data field, why you do not provide any data for example in case you submit a paper unrelated to the above BBOB test suites.

Source Code:
The basis of the workshop is the Comparing Continuous Optimizer platform (COCO,, now rewritten fully in ANSI C with other languages calling the C code. Languages currently available are C, Java, MATLAB/Octave, and Python. Please use the latest version of the code, but at least COCO version 2.2.

Please note that the `bbob-noisy` test suite is not yet supported by the new code. In case you want to submit a paper related to this test suite, please use the previous code base at

Important Dates:
  • 2018-02-27 paper submission system opened
  • 2018-03-01 release 2.2 of the COCO platform: 
  • 2018-03-27 paper and data submission deadline (not extendable!)
  • 2018-04-10 decision notification
  • 2018-04-24 deadline camera-ready papers
  • 2018-07-15 or 2018-07-16 workshop

  • Anne Auger, Inria Saclay - Ile-de-France, France
  • Julien Bect, CentraleSupélec, France
  • Dimo Brockhoff, Inria Saclay - Ile-de-France, France
  • Nikolaus Hansen, Inria Saclay - Ile-de-France, France
  • Rodolphe Le Riche, Ecole Nationale Supérieure des Mines de  Saint–Etienne, France
  • Victor Picheny, INRA Occitanie-Toulouse, France
  • Tea Tušar, Jožef Stefan Institute, Ljubljana, Slovenia

GECCO is sponsored by the Association for Computing Machinery Special Interest Group on Genetic and Evolutionary Computation (SIGEVO). SIG Services: 2 Penn Plaza, Suite 701, New York, NY, 10121, USA, 1-800-342-6626 (USA and Canada) or +212-626-0500 (Global).
Uncertainty management for computational materials science (23-25 mai)

Published on: Mon, 12 Mar 2018 21:36:00 +0100 by Julien Bect
Une formation sur les incertitudes en mécanique des matériaux est organisée par EDF Lab les Renardières du 23 au 25 mai.

Le flyer et le formulaire d'inscription sont disponibles à cette adresse :

PGMO lectures on Bandit convex optimisation (13-14 mars)

Published on: Tue, 06 Mar 2018 13:04:00 +0100 by Julien Bect

Sébastien Bubeck donnera les 13 et 14 mars à l'Ecole Polytechnique, une série de cours intitulée "Bandit convex optimization", dont voici l'annonce :
Dear Colleagues,

The Gaspard Monge Program for Optimisation and Operations Research and their interactions with data sciences (PGMO) is pleased to advertise the exceptional lecture series of

Sebastien BUBECK (Microsoft)



at Ecole polytechnique, Palaiseau, March 13--14, 2018

This is jointly organized by Ecole polytechnique and the master
program ``Optimization'' of Université-Paris Saclay, in the framework
of the PGMO program of Fondation Mathematique Jacques Hadamard,
with the support of EDF, Orange and Thales.

This lecture is intended to graduate students and researchers
(from academics or industry).

The schedule is the following
Lecture 1 - Tuesday, March 13, 10h00-12h30 Amphi Becquerel
Lecture 2 - Tuesday, March 13, 14h00-16h30, Amphi Becquerel
Lecture 3 - Wednesday, March 14, 10h00-12h30, Amphi Becquerel
Lecture 4 - Wednesday, March 14, 14h00-16h30, Amphi Becquerel

Registration (free of charge) is requested; register by filling
the form:

Owing to ``vigipirate'' security measures, attendees may have to
show an identification document.

To access to school:

For local information, contact: (PGMO, FMJH) (DEPMAP, Ecole polytechnique).

The scientific organizers
P. Carpentier (ENSTA)
S. Charousset (EDF)
S. Gaubert (INRIA and Ecole polytechnique)
J.-C. Pesquet (CentraleSupélec)
F. Santambrogio (Université Paris Sud)
G. Stoltz (Université Paris Sud, CNRS)
T. Tomala (HEC)


Summary of the lecture

The multi-armed bandit and its variants have been around for more than 80 years, with applications ranging from medial trials in the 1930s to ad placement in the 2010s. In this mini-course I will focus on a groundbreaking model introduced in the 1990s which gets rid of the unrealistic i.i.d. assumption that is standard in statistics and learning theory. This paradigm shift leads to exciting new mathematical and algorithmic challenges. I will focus the lectures on the foundational results of this burgeoning field, as well as their connections with classical problems in mathematics such as the geometry of martingales and high-dimensional phenomena.

Lecture 1: Introduction to regret. Game theoretic viewpoint (duality, Bayesian version of the game) and derivation of the minimax regret via geometry of martingales (brief recall of type/cotype and entropic proof for ell_1).

Lecture 2: Introduction to the mirror descent algorithm. Connections with competitive analysis in online computations will also be discussed.

Lecture 3: Bandit Linear Optimization. Two proofs of optimal regret: one via low-rank decomposition in the information theoretic argument, and the other via mirror descent with self-concordant barriers.

Lecture 4 : Bandit Convex optimization 1. Kernel methods for online learning, Bernoulli convolution based kernel. 2. Gaussian approximation of Bernoulli convolutions, and restart type strategies.
Design of Experiments: New Challenges (30 avril - 5 mai)

Published on: Fri, 02 Mar 2018 13:39:00 +0100 by Julien Bect
Un message de la part des organisateurs du workshop "Design of Experiments: New Challenges", qui a lieu du 30 avril au 4 mai au CIRM :

There are still a few places available at our workshop on Design of Experiments at CIRM (Marseilles, France, April 30 - May 4, 2018):

In case you or some colleagues would be interested to participate, please send an email to

Poster presentations are welcome: send a tentative title if you are interested (the deadline for abstracts will be set later).

Acceptance will be made on a first come, first served basis.

There is no registration fee, rates for board and lodging at CIRM are available at
On behalf of the organizing committee, I am looking forward to hearing from you soon.

Best wishes,

Luc Pronzato

Old GdR newsletters

If you want to receive the newsletter by e-mail, please fill the register form.

Newsletter archives: