Joakim Beck: Computationally efficient algorithms for computing expected information gain criteria

Abstract.

The talk is on computationally efficient algorithms for computing expected information gain criteria, and is divided into two parts:

a) A greedy algorithm based on Gaussian process models is proposed for estimating the expected information gain in the design of computer experiments. The greedy algorithm achieves near optimality and is demonstrate on a tsunami model.

b) A double loop Monte Carlo method using importance sampling based on Laplace approximations is proposed for Bayesian experimental design, where the expected information gain is used to determine the goodness of an experiment. A cost analysis is given, and the efficiency of the method is demonstrated on a few examples.