Doctoral Seminar

The following talks will be presented in the SPOTSeven Lab Doctoral Seminar (Summer Term 2015).

Date Speaker 1 Title 1 Speaker 2 Title 2
27.3.15: Patryk Filipiak Proactive Evolutionary Algorithms Beate Breiderhoff Optimierung unter Nebenbedingungen
3.4.15: Karfreitag
10.4.15: Blockwoche
15.4.15: Martin Zaefferer Vortrag TU Dortmund Christian Jung Vortrag TU Dortmund
17.4.15: Steffen Moritz Heating, Ventilation and Air Conditioning Jörg Stork Experimental Design and Analysis of Experiments
24.4.15: Martin Zaefferer Kombinatorische Optimierung Martina Friese Ensemble-based Optimization
1.5.15: Maifeiertag
6.5.15: Oliver Flasch Kolloquium TU Dortmund
8.5.15: Christian Jung Prozessoptimierung in der Stahlindustrie Steffen Moritz Heating, Ventilation and Air Conditioning
15.5.15: kein Doktorandenseminar
22.5.15: Christian Jung Prozessoptimierung in der Stahlindustrie Beate Breiderhoff Optimierung unter Nebenbedingungen
29.5.15: Andreas Fischbach Design of Experiments Martin Zaefferer Kombinatorische Optimierung
5.6.15: Martina Friese Ensemble-based Optimization
12.6.15: Christian Jung Prozessoptimierung in der Stahlindustrie Steffen Moritz Heating, Ventilation and Air Conditioning
19.6.15: Tag der exzellenten Lehre
26.6.15: Jörg Stork Experimental Design and Analysis of Experiments Andreas Fischbach Design of Experiments
3.7.15: Martin Zaefferer Kombinatorische Optimierung Martina Friese Ensemble-based Optimization
10.7.15: Christian Jung Prozessoptimierung in der Stahlindustrie Steffen Moritz Heating, Ventilation and Air Conditioning

National and International Guests

The SPOTSeven team is well connected to leading research groups and companies. Guests from all around the word are invited to present their cutting edge research results.
The following researchers were invited to the SPOTSeven Seminar.


Dr. Jörn Mehnen, Cranfield University UK

Evolutionäre Optimisation aus Sicht der Praxis

Tuesday, April. 9th, 1:00 p.m., Ziegelbau

Abstract. Die evolutionäre Optimierung nimmt eine sehr wichtige Stellung in der aktuellen Forschung ein. Von einem praktischen Standpunkt aus scheint die evolutionäre Optimierung und speziell die mehrkriterielle Optimierung jedoch eine nur relativ untergeordnete Rolle zu spielen. In der Literatur finden sich viele theoretische Untersuchungen, jedoch nur wenige aussagekräftige industrielle Fallbeispiele. Wenn man sich die internationale Industrielandschaft genauer ansieht, stellt man fest, dass es weltweit nur ein paar kommerzielle Firmen gibt, die sich auf die Optimierung spezialisiert haben. Es ist interessant zu sehen, welche Anwendungsbereiche durch die evolutionäre Optimierung abgedeckt werden und in welche Richtung aktuelle Unternehmungen zeigen. Dieser Vortrag versucht, den oben genannten Komplex durch diverse Übersichten und ein Praxisbeispiel zu veranschaulichen.

M. Sc. Sebastian Zareba, Institut für Produktentwicklung und Konstruktionstechnik (Prof. Dr.-Ing. Mohieddine Jelali) , FH Köln

Optimierte modellprädiktive Regelung in Kaltwalzanlagen

Thursday, Feb. 21th, 11:30 a.m., Ziegelbau

Abstract: Walzwerke dienen der Herstellung von Metallband, -blech oder -folie. Metall (z.B. Stahl, Aluminium, Kupfer) wird dabei schrittweise durch Walzen verformt bis die gewünschte Enddicke erreicht ist. Die Walzung erfolgt entweder in mehreren hintereinander geschalteten Walzgerüsten bei Tandemwalzstraßen oder sogenannten „Stichen“ bei Reversiergerüsten. Wichtige Qualitätsparamter sind dabei neben der Banddicke die Bandplanheit, die Bandoberfläche und die Materialeigenschaften. Kontinuierliche Produktions- und Produktinnovationen stellen immer höhere Anforderungen an die Prozessführung moderner Walzanlagen. Es müssen enge Toleranzen hinsichtlich Abmessungen, Oberflächenqualität und Materialeigenschaften, eine maximale Anlagenverfügbarkeit sowie ein hoher Durchsatz sichergestellt werden. Darüber hinaus gewinnt die Forderung nach einer effizienten Prozessführung bzgl. Material- und Energieverbrauch eine immer größere Bedeutung. Dies erfordert ein eingehendes Verständnis, eine genaue Beherrschung der Vorgänge innerhalb der Prozesse und eine Koordinierung der Produktionsstufen entlang der Prozessketten.

Prof. Dr. Jack P.C. Kleijnen, Department of Information Management / CentER, Tilburg University, the Netherlands

Simulation optimization via bootstrapped Kriging: survey

Friday, Nov. 16th, 11:55 a.m., Kienbaumsaal, R 1.122

Abstract: In this talk both deterministic and random (stochastic, discrete-event) simulation models are covered, focusing on simulation-optimization via Kriging metamodels (also called “Gaussian Process” or “spatial correlation” “surrogates” or “emulators”). We may analyze these Kriging metamodels through bootstrapping, which is a versatile statistical method; bootstrapping, however, must be adapted to the specific problem being analyzed. More precisely, we should run a random simulation model several times for the same scenario (combination of values for the simulation inputs). The resulting replicated responses may be resampled with replacement, which is called “distribution-free bootstrapping”. A deterministic simulation model, however, is run only once for the same scenario, so we may use parametric bootstrapping assuming a multivariate Gaussian distribution. This distribution is sampled after its (hyper)parameters are estimated from the simulation input/output (I/O) data. More specifically, I shall survey: (1) Kriging for Efficient Global Optimization (EGO) via Expected Improvement (EI) using parametric bootstrapping to obtain an unbiased estimator of the Kriging predictor’s variance; this bootstrap accounts for the randomness resulting from estimating the Kriging (hyper)parameters. (2) Kriging with distribution-free bootstrapping, to solve constrained optimization via Mathematical Programming. (3) Kriging with distribution-free bootstrapping to improve the assumed convexity of the I/O function of the underlying simulation model.

Dr. Dirk Deschrijver and Prof. Dr. Tom Dhaene, Department of Information Technology, Ghent University, Belgium 

Surrogate Modeling of Computationally Expensive Black-Box Systems

Friday, Nov. 16th, 1:30 p.m., Kienbaumsaal, R 1.122

Abstract: For many problems in science and engineering it is impractical to perform experiments on the physical world directly. Instead, complex, physics-based simulation codes are used to run experiments on computer hardware. While allowing scientists more flexibility to study phenomena under controlled conditions, such experiments require a substantial investment of computation time. This places a serious computational burden on associated optimization problems. Surrogate-based optimization becomes standard practice in analyzing such expensive black-box problems. This talk discusses some approaches that make use of surrogate models for optimization and its application to several examples.

Ihsan Yanikoglu, Prof. Dick den Hertog, Prof. Jack P.C. Kleijnen, CentER, Tilburg University, the Netherlands

Adjustable Robust Parameter Design using Metamodels

Friday, Nov. 16th, 2:15 p.m., Kienbaumsaal, R 1.122

Abstract: We present a novel combination of Robust Optimization developed in Mathematical Programming, and Robust Parameter Design (RPD). RPD methods use metamodels estimated from experiments with controllable and environmental inputs (factors). These experiments may be performed with either real or simulated systems; we focus on simulation experiments. RPD methods assume that the mean and covariance, and sometimes even the distribution of the environmental inputs, are known. We develop a Robust Optimization (RO) approach that only uses experimental data, and does not need such assumptions. Moreover, we develop an adjustable RO approach, in which the values of (some of) the controllable factors are adjusted after observing the values of (some of) the environmental inputs. We illustrate our novel methods through several numerical examples.

Dr. Guido Smits (Dow Benelux B.V.)

On the Integration of Data-Driven Modeling into High Throughput Research Workflows

Friday, July 20th 2012, 1:15 p.m., Institut für Informatik, R 2.113

Abstract: In order to accelerate the speed at which Dow Coating Materials brings innovative technologies and products to market, significant resources have been committed to develop High Throughput Research technology. These methods facilitate formulation development by integrating materials development, coatings applications and formulations, high throughput hardware and software, statistics expertise and market knowledge. In this talk we describe the high-level data mining and modeling philosophy and more specifically the role of data-driven modeling.

Prof. Michael Emmerich (Natural Computing Group LIACS, Leiden University)

Robust and Multiobjective Optimization with Expensive Objective Function Evaluations

Friday, July 1st 2012, 11 a.m., Institut für Informatik, R 3.112

Abstract: The need to optimize with computationally expensive objective functions in industrial applications motivates techniques that use knowledge from past evaluations to (approximately) compare new solutions. Transductive learning and metamodelling can be used for accomplishing this. This talk highlights some recent results on how to accelerate robust and hypervolume-based multiobjective optimization in the presence of costly objective functions. Among others, archive-based robust evolutionary optimization and efficient global optimization using the hypervolume-based expected improvement will be discussed. Apart from this, some recent applications in building engineering and quantum chemistry will be briefly presented.

Prof. Dr. Tobias Glasmachers (Institut für Neuroinformatik, Ruhr-Universität Bochum)

Natural Evolution Strategies

Friday, April 27th 2012, 2 p.m., Institut für Informatik, R 3.112

Abstract: Evolution strategies are evolutionary direct search methods suitable for “black box” optimization of continuous variables. As evolutionary algorithms, their most prominent particularity is the ability to adapt the search strategy online to the problem at hand. This results in fast convergence rates and allows to approximate optima in the continuous domain efficiently with arbitrary precision. So-called “natural evolution strategies” are a relatively recent development. They offer a novel, principled strategy adaptation mechanism based on natural gradient ascent in the space of search distributions. This general scheme has lead to the development of efficient algorithms, and maybe more important, to a greatly improved understanding and a unification of the mechanisms driving existing algorithms, like the highly efficient covariance matrix adaptation evolution strategy (CMA-ES).

Dr. Volker Kraft (JMP Academic Ambassador, SAS Institute GmbH, Germany – Cologne Area)

JMP Workshop

Friday, February 24th 2012, 11 a.m., Institut für Informatik, R 2213

Agenda

  1. JMP Intro und Basic Analytics
  2. Advanced Analytics
  3. Ausblick JMP 10
  4. JMP Academic Program
  5. Diskussion

 

Dr. Luis Marti (Group of Applied Artificial Intelligence (GIAA), Dept. of Computer Science. University Carlos III of Madrid)

Model-building issues in multi-objective estimation of distribution algorithms

Friday, February 3rd 2012, 11 a.m. Institut für Informatik, R.3.109

Abstract: There are some issues with multi–objective estimation of distribution algorithms (MOEDAs) that have been undermining their performance when dealing with problems with many objectives. In this talk we examine the model–building issue related to estimation of distribution algorithms (EDAs) and show that some of their, as yet overlooked, characteristics render most current MOEDAs unviable in the presence of many objectives. We will present model–building as a problem with particular requirements and explain why some current approaches cannot properly deal with some of these conditions. Then, we will discuss the strategies proposed for adapting EDAs to this problem.