|Département Génie mathématique et industriel |
Titre : Direct Model Predective Control: A Theoretical and Numerical Analysis
Jeudi 7 juin à 14h – amphi 104 de l’Espace Fauriel.
Présentatrice : Marie-Liesse Cauwet.
Résumé : The power system problem is formulated as a stochastic decision process with large constrained action space, high stochasticity and dozens of state variables. Direct Model Predictive Control has previously been proposed to encompass a large class of stochastic decision making problems. It is a hybrid model which merges the properties of two different dynamic optimization methods, Model Predictive Control and Stochastic Dual Dynamic Programming. We prove that Direct Model Predictive Control reaches an optimal policy for a wider class of decision processes than those solved by Model Predictive Control (suboptimal by nature), Stochastic Dynamic Programming (which needs a moderate size of state space) or Stochastic Dual Dynamic Programming (which requires convexity of Bellman values and a moderate complexity of the random value state). The algorithm is tested on a multiple-battery management problem and two hydroelectric problems. Direct Model Predictive Control clearly outperforms Model Predictive Control on the tested problems.
Titres : Activité en Science des Données à Total
Jeudi 26 avril à 14h – Salle 528 de l’Espace Fauriel à Saint-Étienne
Présentateur : Michel LUTZ (Total)
Titre : Optimization of computer experiments with multiple responses
Mardi 20 mars, 9h, Espace Fauriel, Amphi 104
Présentateurs : Sonja KUHNT et Dominik KIRCHHOFF (Université de Dortmund)
Résumé : In this talk we present results and on-going research from two different projects. In engineering and natural sciences it has become common practice to replace physical experiments with computer simulations (black-box experiments). As these simulations can be very time-consuming and complex, often a surrogate or meta-model of the computer experiment is built first and analysis and optimization are then based on this model. Gaussian process models, better known as Kriging models, are widely used for this purpose.
Our first project is set in the field of turbo machines, which play an important role for the whole process chain of energy transformation. A specific impeller geometry is simulated over the entire operating range with regard to its efficiency. Computer experiments following a space-filling design are run and Gaussian process models are fitted to multiple responses, some of which are to be optimized whereas others need to be kept within certain bounds. The popular efficient global optimization (EGO) procedure provides a sequential Kriging-based optimization with the expected improvement as criterion. We look at extensions of the EGO algorithm to multiple responses as well as constraints.
In the second part we deal with computer simulations with mixed continuous and categorical input variables. The original Kriging model can only cope with purely continuous inputs. First extensions exist to incorporate also categorical variables. Here, we consider three approaches called Exchangeable Correlation, Multiplicative Correlation and Unrestrictive Hypersphere-based Correlation that are different in terms of their flexibility and computational effort. We introduce a new approach, where the number of unknown parameters can be chosen flexibly by the practitioner. We present first results from a computer simulation of a distribution warehouse.