大阪大学　数理・データ科学教育研究センター
Center for Mathematical Modeling and Data Science,Osaka University

Bayesian Inference for Intractable Likelihood Models

Krzysztof Łatuszyński (Warwick University)

1. ホーム
2. セミナー・集中講義
3. セミナー
4. Bayesian Inference for Intractable Likelihood Models

大阪大学 数理・データ科学セミナー データ科学セミナーシリーズ 第46回

Bayesian Inference for Intractable Likelihood Models

Krzysztof Łatuszyński (Warwick University)

Constructing MCMC algorithms for Bayesian inference in intractable likelihood models is problematic since difficulties in evaluating the likelihood make the application of the standard Metropolis-Hastings acceptance formula impossible. This enforces either (1) approximate methods that introduce bias of unknown magnitude, or (2) the pseudomarginal approach that is exact, but slows down MCMC convergence, sometimes dramatically. In this talk, I will present a new approach (3) based on unbiased estimators of the likelihood and the Barkers acceptance ratio. The approach is exact and retains the per iteration convergence rate comparable to that of the standard Metropolis-Hastings. I will illustrate the approach with examples of exact inference for stochastic differential equations.

This is joint work with Flavio Goncalves, Gareth Roberts, and Dootika Vats.

本発表はJST CREST JPMJCR14D7によってサポートされています．