A General Metric for Riemannian Manifold Hamiltonian Monte Carlo

摘要

Markov Chain Monte Carlo (MCMC) is an invaluable means of inference withcomplicated models, and Hamiltonian Monte Carlo, in particular RiemannianManifold Hamiltonian Monte Carlo (RMHMC), has demonstrated impressive successin many challenging problems. Current RMHMC implementations, however, rely on aRiemannian metric that limits their application to analytically-convenientmodels. In this paper I propose a new metric for RMHMC without theselimitations and verify its success on a distribution that emulates manyhierarchical and latent models.