B0655
Title: MCMC-driven importance samplers using partial posteriors
Authors: Luca Martino - Universidad Rey Juan Carlos (Spain) [presenting]
Abstract: Many applications require the approximation of intractable integrals involving complex posterior distributions. Monte Carlo methods such as Markov Chain Monte Carlo (MCMC) and Importance Sampling (IS) are often employed to approximate these integrals. We propose adaptive IS schemes driven by MCMC chains each one addressing a partial posterior, i.e., a posterior of subsets of data. Partition of the data is commonly used in distributed frameworks, but we do not consider this setting. The goal is to leverage the use of partial posteriors for constructing more efficient importance samplers. Several schemes are discussed: they are improved versions of the so-called Layered Adaptive Importance Sampling (LAIS) algorithm. We also show an application to minibatch selection. Our schemes are validated in very challenging real-world problems, such as exoplanet detection.
