Cycle-cutset sampling for Bayesian networks

摘要: The paper presents a new sampling methodology for Bayesian networks called cutset that samples only subset of the variables and applies exact inference others. We show this approach can be implemented efficiently when sampled constitute cycle-cutset network otherwise it is exponential in induced-width network's graph, whose are removed. Cutset an instance well known Rao-Blakwellisation technique variance reduction investigated [5, 2, 16]. Moreover, proposed scheme extends standard methods to non-ergodic with ergodic subspaces. Our empirical results confirm those expectations cycle superior Gibbs variety benchmarks, yielding simple, yet powerful scheme.