Abstract
Current nonlinear design theory and computational methods cannot (with a few exceptions) be correctly applied problems where posterior uncertainties will not be approximately Gaussian, due to either a non-Gaussian prior or the model nonlinear.
Some work on MCMC methods for evaluating expected gains in Shannon information has been done following Lindley's 1956 paper, which is valid in the fully nonlinear case, but too computationally heavy to be applied to cases where the expected information gains are larger than a few bits.
We present a novel approach to the problem, based on minimizing the the expected variance of the experiment outcome, valid for the fully nonlinear problem along with efficient computational methods for estimating and optimising this experiment quality criterion.