Adaptive Metropolis Sampling with Product DistributionsThe Metropolis-Hastings (MH) algorithm is a way to sample a provided target distribution pi(z). It works by repeatedly sampling a separate proposal distribution T(x,x') to generate a random walk {x(t)}. We consider a modification of the MH algorithm in which T is dynamically updated during the walk. The update at time t uses the {x(t' less than t)} to estimate the product distribution that has the least Kullback-Leibler distance to pi. That estimate is the information-theoretically optimal mean-field approximation to pi. We demonstrate through computer experiments that our algorithm produces samples that are superior to those of the conventional MH algorithm.

Document ID

20050082137

Document Type

Other

Authors

Wolpert, David H. (NASA Ames Research Center Moffett Field, CA, United States)

Lee, Chiu Fan (Oxford Univ. Oxford, United Kingdom)