Multimodal information gain in Bayesian design of experiments

One of the well-known challenges in optimal experimental design is how to efficiently estimate nested integrations expected information gain. The Gaussian approximation and associated importance sampling have been shown be effective at reducing numerical costs. However, they may fail due non-negligible biases instabilities. A new approach developed compute gain, when posterior distribution multimodal—a situation previously ignored by methods aiming accelerating integrations. Specifically, approximated using a mixture constructed multiple runs global search for modes weighted local Laplace approximations. Under any given probability capturing all modes, we provide an estimation number searches, which dimension independent. It that novel global-local multimodal can significantly more accurate efficient than other existing approaches, especially large. applied designs experiments with both calibrated uncalibrated observation noises.