Session: Applications of derivative-free optimization
Chair: Juliane Müller
Cluster: Derivative-free Optimization
Talk 1: Surrogate model guided search for optimizing metabolic models
Speaker: Juliane Müller
Abstract: The development and scaleup of novel biofuels require accurate prediction of microbial performance. While there are very sophisticated metabolic and expression (ME) models that simulate cellular metabolism and expression, their accuracy depends on a variety of parameters that must be tuned such that the model agrees with observation data. This yields a continuous optimization problem in which the quality of parameter sets must be assessed by running the compute intensive ME models. In this talk, we will present the challenge associated with this application and propose derivative-free surrogate models guided optimization approaches to tackle it.
Talk 2: A multilevel stochastic regularized first-order method with application to training
Speaker: Filippo Marini
Abstract: We present a new multilevel stochastic framework for the solution of optimization problems. The proposed approach uses random regularized first-order models that exploit an available hierarchical description of the problem, being either in the classical variable space or in the function space, meaning that different levels of accuracy for the objective function are available. We present the converge analysis of the method and show its numerical behavior on the solution of finite-sum minimization problems arising in binary classification problems.
Talk 3: Surrogate-based evolutionary optimization extended to categorical and dependent variables
Speaker: Charlotte Beauthier
Abstract: The objective of this work is the development of methods combining numerical simulation tools and advanced optimization algorithms, which play a crucial role in exploring new conceptual designs. Nowadays the exploitation of multi-disciplinary optimization using high-fidelity simulation models is common to many engineering design problems. A globally effective approach to optimization problems based on computationally expensive analysis lies in the exploitation of surrogate models. They act as cheap-to-evaluate alternatives to the original high-fidelity models reducing the computational cost, while still providing improved designs. Furthermore, in practice, various industrial design problems have continuous, discrete, and categorical variables. From an engineering point of view, the specific case of categorical variables is of great practical interest by their ability to represent the choice of a material, the type of engine architecture, the shape of a cross-section for a beam profile, etc. The contributions of this talk are focused on the management of mixed variables, in particular the categorical ones, in a surrogate-based evolutionary algorithm where dependency can also be defined between variables. More precisely, the dependency considered here is when the definition domain of a variable can be linked to another variable’s value. In order to deal with mixed and dependent variables, different methods are proposed to refine the notion of distance between mixed variables, using the notion of affinity, for instance, and to adapt the genetic operators in the evolutionary algorithm accordingly. These novel developments have been implemented in Minamo, Cenaero’s in-house design space exploration and multi-disciplinary optimization platform. Numerical results will be presented on specific test problems coming from structural and mechanical design frameworks to study the impact of the proposed strategies.
