Cross Validated Non parametric Bayesianism by Markov Chain Monte Carlo

Abstract

Completely automatic and adaptive non-parametric inference is a pie in the sky. The frequentist approach, best exemplified by the kernel estimators, has excellent asymptotic characteristics but it is very sensitive to the choice of smoothness parameters. On the other hand the Bayesian approach, best exemplified by the mixture of gaussians models, is optimal given the observed data but it is very sensitive to the choice of prior. In 1984 the author proposed to use the Cross-Validated gaussian kernel as the likelihood for the smoothness scale parameter h, and obtained a closed formula for the posterior mean of h based on Jeffreys's rule as the prior. The practical operational characteristics of this bayes' rule for the smoothness parameter remained unknown for all these years due to the combinatorial complexity of the formula. It is shown in this paper that a version of the metropolis algorithm can be used to approximate the value of h producing remarkably good completely automatic and adaptive kernel estimators. A close study of the form of the cross validated likelihood suggests a modification and a new approach to Bayesian Non-parametrics in general.