On efficient black box optimization of systems defined by 100 or more parameters

Olivier Goury (1, 2) , Sven Ahlinder (3, 4)

1 INSERM - Institut National de la Santé et de la Recherche Médicale
2 DEFROST - Deformable Robots Simulation Team
3 Chalmers University of Technology [Göteborg]
4 Volvo Group

In product development, the optimization of hectovariate systems (that is, depending on more than a hundred variables) is hard to handle with usual black box methods, since they require a huge number of experiments. That is an issue since each experiment can be both expensive and time-consuming. Furthermore, these methods can't even handle such big problems within a reasonable period of time. The Lean optimization algorithm [1] was one approach to overcome these difficulties. In this paper, we will come up with a new lean algorithm, inspired from [1] and developing new ideas. It will then be tested on some classical mathematical problems of various sizes and a practical example of combustion simulation engine issued from Volvo Technology. We will see that this new algorithm has a faster and more robust convergence, and is able to significantly improve most systems within a number of experiments less than the number of variables.

Mots clés: Lean optimisation; Supersaturated design; Hectovariate problems; Engine

Domaines: Mathématiques [math]; Statistiques [stat]

Olivier Goury, Sven Ahlinder. On efficient black box optimization of systems defined by 100 or more parameters. 2010. ⟨hal-01598044⟩