Startseite // SnT // News & E... // PhD Defense: Co-evolutionary hybrid bi-level optimization

PhD Defense: Co-evolutionary hybrid bi-level optimization

twitter linkedin facebook google+ email this page
Add to calendar
Sprecher: Emmanuel Kieffer
Veranstaltung: Freitag, den 18. Januar 2019 10:00 - 13:00
Ort: Room MNO 1.040
Campus Belval, Maison du nombre
6, avenue de la Fonte
L-4364 Esch-sur-Alzette

Members of the defence committee:

  • Chairman: Prof. Dr Ulrich Sorger, Université du Luxembourg
  • Vice-chairman: Prof. Dr Anass Nagih, Université de Lorraine, France
  • Supervisor: Prof. Dr Pascal Bouvry, Université du Luxembourg
  • Member: Prof. Dr Franciszek Seredyński, Cardinal Stefan Wyszynski University Warsaw, Poland
  • Member: Dr Grégoire Danoy, Université du Luxembourg

Abstract :

Multi-level optimization stems from the need to tackle complex problems involving multiple decision makers. Two-level optimization, referred to as “Bi-level optimization”, occurs when two decision makers only control part of the decision variables and impact each other decision. Bi-level problems are sequential by nature and can be represented by two nested optimization problems in which one problem is constrained by another one.

After a deep study of theoretical properties and a survey of the existing applications being bi-level by nature, this dissertation focuses first on problems that could benefit from bi-level modelling.  In this belief, a novel bi-level clustering model is proposed to combine distance metrics in a hierarchical manner. This new bi-level model has been successfully applied on large graph-based repositories.  Bi-level optimization takes its full meaning when describing adversarial situations such as market competition. As a result, a novel bi-level cloud pricing model has been developed as decision aid tool in the context of Cloud Pricing.

Although bi-level programming allows to model more accurately complex and adversarial situations, optimization of bi-level models remains extremely challenging when dealing with large-scale and combinatorial problems. This dissertation relies on co-evolution, a bio-inspired approach, to break the inherent nested structure that makes bi-level optimization so challenging. Throughout this work, the new concept of “learning to optimize” is also considered to automatically train dedicated instance solvers. Hybridizing co-evolution approaches and the “learning to optimize” concept, a novel hybrid bi-level co-evolutionary algorithm, i.e., CARBON has been developed to solve general and large-scale bi-level problems.