Transformations in semi-parametric Bayesian synthetic likelihood

Bayesian synthetic likelihood (BSL) is an established method for performing approximate inference when the function intractable. In methods, approximated parametrically via model simulations, and then standard likelihood-based techniques are used to perform inference. The Gaussian estimator has become ubiquitous in BSL literature, primarily its simplicity ease of implementation. However, it often too restrictive may lead poor posterior approximations. Recently, a more flexible semi-parametric (semiBSL) been introduced, which significantly robust irregularly distributed summary statistics. A number extensions semiBSL proposed. First, even estimators marginal distributions considered, using transformation kernel density estimation. Second, whitening (wsemiBSL) proposed – improve computational efficiency semiBSL. wsemiBSL uses decorrelate statistics at each algorithm iteration. methods developed herein versatility algorithms.