Abstract
We introduce a fully stochastic gradient based approach to Bayesian optimalexperimental design (BOED). This is achieved through the use of variationallower bounds on the expected information gain (EIG) of an experiment that canbe simultaneously optimized with respect to both the variational and designparameters. This allows the design process to be carried out through a singleunified stochastic gradient ascent procedure, in contrast to existingapproaches that typically construct an EIG estimator on a pointwise basis,before passing this estimator to a separate optimizer. We show that this, inturn, leads to more efficient BOED schemes and provide a number of a differentvariational objectives suited to different settings. Furthermore, we show thatour gradient-based approaches are able to provide effective design optimizationin substantially higher dimensional settings than existing approaches.