| Summary: | We outline a general purpose framework for black-box marginal maximum a pos- teriori estimation of probabilistic program variables using Bayesian optimization with Gaussian processes. We introduce the concept of an optimization query, whereby a probabilistic program returns an infinite lazy sequence of increasingly optimal estimates, and explain how a general purpose program transformation would allow the evidence of any probabilistic program, and therefore any graphical model, to be optimized with respect to an arbitrary subset of its variables.
|
|---|