Event

PhD Defence: Learning Optimisation Algorithms over Graphs

  • Conférencier  Gabriel Valentin DUFLO

  • Lieu

    Belval, MSA, room 2.380

    LU

You are cordially invited to join the PhD defence of Gabriel Valentin DUFLO on Friday, 31 March at 10:00 on Belval campus, MSA, room 2.380.

Members of the defense committee:

  • Prof. Dr Pascal BOUVRY, University of Luxembourg, Chairman
  • Prof. Dr El-Ghazali TALBI University of Lille – INRIA – CNRS/France, Deputy Chairman
  • Dr Grégoire DANOY, University of Luxembourg, Supervisor
  • Prof. Dr Ann NOWÉ, VRIJE UNIVERSITEIT BRUSSEL (VUB), Belgium, Member
  • Prof. Dr Roland BOUFFANAIS, University of Ottawa, Canada, Member

Abstract:

The paradigm of learning to optimise relies on the following principle: instead of designing an algorithm to solve a problem, we design an algorithm which will automate the design of such a solver. Hyper-heuristics constitute the main learning-to-optimise techniques. They are however designed to tackle a specific problem. Due to this lack of generality, existing works fully redesign hyper-heuristics when tackling a new problem, despite the fact that they may share a similar structure. In this dissertation, we tackle this challenge by proposing a generic way for learning to optimise any problem. To this end, this thesis introduces three main contributions: (i) an analysis of the formal functioning of learning-to-optimise techniques; (ii) a model of generic hyper-heuristic, named Algorithm Learner for Graph Optimisation problems (ALGO), constituting the central point of this work; (iii) a real-world use case where we use our generic hyper-heuristic to automate the design of behaviours within a swarm of drones.