4 présentations

• 
  15h30 - 15h55

  Nonlinear optimization with Artelys Knitro

  • Violette Berge, prés., Artelys Canada Inc.

  Nonlinear optimization is used in many applications in a broad range of industries such as economy, finance, energy, health, 3D modeling, and marketing. With four algorithms and great configuration capabilities, Artelys Knitro is the leading solver for nonlinear optimization and demonstrates high performance for large scale problems. This session will introduce you to Artelys Knitro, its algorithms (interior points and active sets methods for continuous problems and MIP Branch and Bounds), key features and modeling capabilities.
• 
  15h55 - 16h20

  Deterministic maintenance scheduling for large stochastic systems using blackbox optimization and a decomposition method

  • Thomas Bittar, prés., EDF / Ecole des Ponts ParisTech

  This work is motivated by the optimization of maintenance scheduling for components of hydroelectric power plants. We consider a system of several components (turbines, generators …) coupled by a common stock of spare parts and we seek the dates of preventive maintenance that minimize the expectation of the cost generated by the system. We use a decomposition method to tackle this high dimensional problem. The idea is to iteratively find the best maintenance policy on each component separately and then coordinate the components. The lower-dimensional subproblems on the individual components are solved using blackbox optimization.
  Keywords: maintenance scheduling, decomposition-coordination, blackbox optimization
• 
  16h20 - 16h45

  An upper trust bound feasibility criterion for constrained Bayesian optimization.

  • Rémy Priem, prés., ONERA, DTIS, Université de Toulouse, Toulouse, France
  • Nathalie Bartoli, ONERA, DTIS, Université de Toulouse, Toulouse, France
  • Youssef Diouane, ISAE-SUPAERO, Université de Toulouse, Toulouse, 31055 Cedex 4, France

  In this talk, we propose to address efficiently black box constrained optimization problems. Our approach combines sequential enrichment and adaptive surrogate models by means of three ingredients: the Bayesian optimization framework, Gaussian process models of the constraints, and a feasibility criterion built using the uncertainty estimation of the constraints. / Constrained Bayesian optimization / Gaussian process / Global Optimization
• 
  16h45 - 17h10

  Use of surrogate-based model search for parallel blackbox optimization

  • Stéphane Alarie, prés., Institut de recherche d'Hydro-Québec
  • Bastien Talgorn, Université McGill
  • Michael Kokkolaras, McGill University

  Blackbox problems are here solved with MADS on parallel computers. MADS generates $2n$ candidate solutions that can be simultaneously evaluated during the POLL step. The SEARCH step is however mostly sequential. Based on surrogates, proposition is made to also generate several candidates in the SEARCH. Results with NOMAD are presented.

  Blackbox Optimization ; Parallel Evaluations ; Surrogate-Based Models