Adaptive Model-Switch Thermodynamic Integration: A Bayes Factor Estimator

概要

Bayes factors are a widely used model comparison criterion, but are often quite challenging and computationally expensive to estimate well. We propose a variation of the model-switch thermodynamic integration method for estimating Bayes factors that decreases the variance of the estimator and essentially eliminates integration error. We choose samples whose density in $β$ is proportional to the standard deviation of previously drawn samples. Rather than averaging many samples at each $β$, we use…