<html>
<head>
<meta http-equiv="content-type" content="text/html; charset=UTF-8">
</head>
<body text="#000000" bgcolor="#FFFFFF">
<font face="Times New Roman, Times, serif">V torek, 23. aprila 2019
bo ob 9h v Oranzni predavalnici (drugo nadstropje glavne stavbe)
289. Solomonov seminar. Oranzna predavalnica je v drugem
nadstropju glavne stavbe IJS na Jamovi 39. Posnetki preteklih
seminarjev so na <a class="moz-txt-link-freetext"
href="http://videolectures.net/solomon/">http://videolectures.net/solomon/</a><br>
<br>
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~<span
style="color:windowtext" lang="EN-AU"></span><br>
<br>
Title: Quality Indicator Maximization in Multiobjective
Optimization Via Single-Objective Solvers: Unflattened Hypervolume
Improvement in the Sofomore Framework<br>
<br>
Lecturer: Dimo Brockhoff, Inria and École Polytechnique,
Palaiseau, France <br>
<br>
Multiobjective Optimization problems appear frequently in practice
when multiple objective functions need to be optimized
simultaneously. Often, a multiobjective problem is approached by
aiming to find a set of p solutions that maximizes a given
quality, for example as defined by the hypervolume indicator.<br>
<br>
In this talk, I will present a new multiobjective framework which
attacks the optimization of p solutions in a search space of
dimension n towards the maximum of a quality indicator by
successive dynamic (single-objective) subspace optimization of an
n times p dimensional problem. When instantiated as the
COMO-CMA-ES with an "unflattened" version of the hypervolume
improvement and the well-known CMA-ES as single-objective solver,
we observe linear convergence to the optimal placement of p
solutions with respect to the hypervolume indicator on various
bi-objective convex-quadratic problems. In addition to the general
idea of the framework and details on the concrete COMO-CMA-ES, I
will present in particular the intuition why the choice of the
"unflattened" hypervolume is crucial to the performance of the
algorithm. The presentation of benchmarking data from comparisons
with other well-known multiobjective algorithms on the bbob-biobj
suite of the COCO platform will top off the presentation.<br>
<br>
This presentation is based on work with Cheikh Touré, Anne Auger,
and Nikolaus Hansen: "Unflattened Hypervolume Improvement for
Multiobjective Problems: COMO-CMA-ES and the Sofomore framework",
accepted at GECCO-2019<br>
<br>
<span style="color:windowtext" lang="EN-AU">Short bio of the
lecturer</span>: <br>
<br>
Dimo Brockhoff received his diploma in computer science from
University of Dortmund, Germany in 2005 and his PhD (Dr. sc. ETH)
from ETH Zurich, Switzerland in 2009. After two postdocs at Inria
Saclay Ile-de-France (2009-2010) and at Ecole Polytechnique
(2010-2011), he joined Inria in November 2011 as a permanent
researcher (first in its Lille - Nord Europe research center and
since October 2016 in the Saclay - Ile-de-France one). His
research interests are focused on evolutionary multiobjective
optimization (EMO), in particular on theoretical aspects of
indicator-based search and on the benchmarking of blackbox
algorithms in general. </font>
<pre class="moz-signature" cols="72">
</pre>
</body>
</html>