Stochastic Optimization for Design of Experiments

Abstract

We present a stochastic optimization method to converge to local maxima of the Shannon's expected information gain of experiments using a Bayesian framework. We avoid the high cost of evaluating several double loop Monte Carlo simulation (DLMC) each iteration by employing the stochastic gradient. Our method has proven to converge to local maxima with a fraction of the cost of the classical approach, making possible to optimize experiments with more expensive models. We confirm the convergence of our method optimizing experimental design problems with analytical ODE models and with finite elements approximations of PDEs.