2019/03/03 by Kelvin Hsu, Hsu, Kelvin, Fábio Ramos +1
Computer Science · Mathematics · #Machine Learning and Algorithms #Gaussian Processes and Bayesian Inference #Markov Chains and Monte Carlo Methods
paper · pdf · doi:10.48550/arxiv.1903.00863
In likelihood-free settings where likelihood evaluations are intractable,\napproximate Bayesian computation (ABC) addresses the formidable inference task\nto discover plausible parameters of simulation programs that explain the\nobservations. However, they demand large quantities of simulation calls.\nCritically, hyperparameters that determine measures of simulation discrepancy\ncrucially balance inference accuracy and sample efficiency, yet are difficult\nto tune. In this paper, we present kernel embedding likelihood-free inference\n(KELFI), a holistic framework that automatically learns model hyperparameters\nto improve inference accuracy given limited simulation budget. By leveraging\nlikelihood smoothness with conditional mean embeddings, we nonparametrically\napproximate likelihoods and posteriors as surrogate densities and sample from\nclosed-form posterior mean embeddings, whose hyperparameters are learned under\nits approximate marginal likelihood. Our modular framework demonstrates\nimproved accuracy and efficiency on challenging inference problems in ecology.\n