A0608
Title: Local optimization for sequential design of experiments via sparse meta-model
Authors: Matteo Borrotti - University of Milan-Bicocca (Italy) [presenting]
Davide Ferrari - Free University of Bozen/Bolzano (Italy)
Abstract: High-dimensional experiments can be characterized by many input variables and, often, a limited number of observations. The final aim is to optimize the experimental response. A possible solution is local optimization, suitable for expensive, high-dimensional black-box experiments. One possible advantage of local optimization is that it does not need to explore the entire experimental space to reach the optimum. The proposal is to use a local probabilistic model to estimate the objective response surface. The local meta-model is used to guide the search on the objective response surface and to provide an automatic way of determining the step size from one iteration to the next. Furthermore, to handle a possible issue related to sparsity on the input variables, the proposed approach detects the most important variables at each step to move in significant directions of the objective response surface. The final solution is compared with a benchmark on different simulation data models.
