EVOLVE International Conference

EVOLVE 2015 June 18-24
Iasi, Romania

HomeTutorials

Tutorials

The following tutorials will be offered (free of charge) during the conference (see below for detailed description and biographies), list to be updated:

Dr. Iryna Yevseyeva  

Multicriteria Decision-Aiding: Compensating and Non-compensating Methods

Dr. Michael T.M. Emmerich  

Set-Oriented Multicriteria Optimization: Deterministic and Stochastic Methods

Dr. Madalina Drugan   Evolutionary Reinforcement Learning or Reinforcement Evolutionary Algorithms?

 


 

Multicriteria Decision-Aiding: Compensating and Non-compensating Methods

Time: 105 Minutes (including break of 15 minutes) 

Tutorial Description:

This tutorial will present approaches for aiding a person or a group of people to make decision about selection, ranking or classification of alternatives, evaluated on several conflicting criteria. Multicriteria decision-aiding (MCDA) can be applied as a following step after multiobjective optimization, when a single solution should be selected among non-dominated ones, according to subjective preferences of decision maker(s). Two MCDA approaches are highlighted. Compensating approaches, such as utility or value-based approaches, which allow for trading-off some criteria at the expense of other(s), and non-compensating, such as outranking approaches, which are based on pairwise comparison of alternatives and not always allow for compensation. The tutorial covers foundations of the MCDA field and algorithms from each of the research approaches. The tutorial consists of three parts: foundations, existing MCDA approaches, and recommendations for choosing a method for problem solving. Working examples are used to present fundamental problems and illustrate decision-aiding process for coming up with solution in each of the discussed method. The foundation part will address formulations of selection, ranking and classification problems, parameters needed in these models, and modeling difficulties related to the subjective nature of decision-aiding process. Then, existing approaches to solve MCDA problems and the ways different methods use to overcome modeling difficulties will be presented. In particular, an utility/value-based compensatory method and an outranking non-compensatory method will be shown on the same working example. Final recommendations for selecting a method to be used depending on the initial available information and requirements to the final decision will be provided.

Dr. Iryna Yevseyeva

School of Computing Science, Newcastle University, UK

Biography: 

Dr. Iryna Yevseyeva is a research associate at the Choice Architecture for Information Security (ChAISe) project at Newcastle University, UK. She contributes to the project with her expertise in optimization and decision making, in particular, in multicriteria optimization and decision analysis. Before joining Newcastle University in 2013, she was a post-doctoral researcher on multiobjective optimization: in the Netherlands the Leiden Institute of Advanced Computer Science at the Leiden University in 2012-2013 for drug discovery; and in Portugal at the Polytechnic Institute of Leiria in 2011-2012 for spam filtering; at INESC Porto with “Ciencia 2008 program” in 2009-2011 for scheduling; at the University of Algarve with grants from Academy of Finland and European Commission (Erasmus Mundus) in 2008-2009 for algorithms design development. Iryna received her PhD degree in computer science and optimization from the University of Jyvaskyla, Finland, in 2007, for the research on multicriteria classification with applications in healthcare. Before joining the PhD program in 2004, she worked as a software engineer at the Niilo Maki Institute of Neuropsychology, Finland, in 2001-2003. She has a Master of Science degree in mobile computing from the University of Jyvaskyla, Finland (2001) and a Master degree with honors in information technology from the Kharkov National University of Radio-electronics, Ukraine (2000). Besides, Dr. Iryna Yevseyeva gave lectures on multicriteria decision analysis within courses on multicriteria optimization in Leiden in 2012 and on information security and trust in Newcastle in 2013, 2014.

 


 

Set-Oriented Multicriteria Optimization: Deterministic and Stochastic Methods

Time: 105 Minutes (including break of 15 minutes) 

Tutorial Description:

This tutorial will introduce deterministic and stochastic methods in set-oriented multicriteria optimization. The focus is on the a-posteriori approach to multicriteria optimization, i.e. the idea to produce a set of solutions. The tutorial is structured in three parts: Foundations, algorithms and application notes.

In the first part we will show different problem formulations in multicriteria and set-oriented optimization and geometrical aspects of orderings and result sets. The discussion will include novel aspects, such as level-set approximation and cone-orderings, which extend the Pareto orders.

On the algorithm side, we will cover concepts of mainstream stochastic algorithms, such as second generation evolutionary multicriteria optimization algorithms (EMOA) (NSGA-II and SPEA II), decomposition based EMOA (MOEA/D, NSGA-III, etc.) and Indicator Based Algorithms (HyPE, SMS-EMOA, R2-EMOA etc.). Besides the other stochastic search paradigms will be highlighted, such as multicriteria particle swarm optimization (MOPSO). Among deterministic paradigms for multicriteria optimization we will have a look at derivative based strategies such as direct identification of Karush Kuhn Tucker points, Set-oriented Newton’s method, Level Set Continuation, and efficient global optimization. The discussion will focus on continuous methods here.

Finally, a workflow of how to solve multicriteria optimization problems in practice will be discussed and exemplified for selected application domains in computational chemistry/biology and engineering.

Dr. Michael T.M. Emmerich

LIACS, Leiden University, The Netherlands

Biography: 

Dr. Michael Emmerich works as a professor for multicriteria analytics in the Leiden Institute for Advanced Computer Science, LIACS, Leiden University. He has received his Doctorate of the Natural Science from Dortmund University, Germany (promotor: Prof. H.-P. Schwefel and Prof. Peter Buchholz) on integrating Gaussian process surrogate modeling in multicriteria optimization. From 1999-2003 he worked in various projects with the German chemical industry (Center for Applied Systems Analysis/ICD e.V.; collaborations with Bayer, BASF, Degussa, and ACCESS Material Sciences), before in 2003 he joined the Collaborative Research Center Computational Intelligence, Dortmund. Since 2004 Michael Emmerich is appointed as assistant professor in LIACS, Leiden University, with a break in 2011, when he worked on a FCT project on Indicator Based Multicriteria Optimization with Prof. Carlos Fonseca (Faro, Portugal, IST Lisbon). His main topic is the design and application of multicriteria optimization algorithms, and he has contributed to the design of state-of-the-art techniques such as SMS-EMOA, Set-oriented Newton’s Method, and developed asymptotically optimal algorithms for computing hypervolume indicator gradients/contributions. Besides, Dr. Michael Emmerich teaches since 2011 a class on multicriteria optimization and decision analysis (MODA) and is leader of the MODA research group in this domain (http://moda.liacs.nl).

 


   

Evolutionary Reinforcement Learning or Reinforcement Evolutionary Algorithms?

Time: 105 Minutes (including break of 15 minutes) 

Tutorial Description:


A recent trend in machine learning (ML) is the transfer of knowledge from one area to another. We focus on potential synergies between multi-armed bandits (MAB) and evolutionary computation (EC). Evolutionary computation is a subfield of global optimization and is an incubator for a range of techniques (e.g., multi-objective evolutionary computation, estimation distribution algorithms) with a wide range of application. MAB addresses sequential decision problems in an initially unknown environment in a unitary framework for convergence guaranties to optimality. While MAB is a mathematical paradigm, the main strength of EC is its general applicability and computational efficiency. Although at first they seem very different, MAB and EC are two learning techniques that address basically the same problem: the maximization of the agent's reward in MAB and fitness function in EC in a potentially unknown environment. The goal of this tutorial is to give an overview of MAB techniques integrated in EC and vice-versa, EC techniques used in MABs. We discuss on resemblances and differences in learning and optimising with multi-armed bandits and evolutionary computation.

Dr. Madalina Drugan

Vrije Universiteit Brussel, Belgium

Biography: 

Madalina M. Drugan received the Diploma Engineer degree in Computer Science from the Technical University of Cluj-Napoca, Romania. In 2006, she received the PhD degree from the Computer Science Department, University of Utrecht, The Netherlands. She have carry out research in Machine Learning and Optimisation and related topics like Evolutionary Computation, Bioinformatics, Multi-objective optimisation, Meta-heuristics, Operational Research, and Reinforcement Learning. She has experience with research grants, reviewing services and a strong publication record (more than 50 publications, 10 journal publications) in international peer-reviewed journals and conferences. During last three years, she initiated and organized several special sessions, tutorials and workshops on Reinforcement Learning and Optimisation at IEEE WCCI, IEEE SSCI, PPSN, ESANN, CEC, IJCNN, SIMCO, EVOLVE. She co-supervised three PhD students and two Master students and some of their joint publications were awarded prices at international conferences.

Joomla templates by Joomlashine