Proposal and Discussion on Adaptive Exchange Monte Carlo Method

Abstract

This paper proposes an adaptive Monte Carlo method based on Exchange Monte Carlo and gives a discussion with respect to it through experiments. Exchange Monte Carlo is used widely for both purposes of optimization and estimation of a probability distribution. There have been studies to propose an improvement on the method, however, adaptive method has not been studied well. The paper proposes generalized joint probability of parameters’ distribution, and target distribution that is desired to be estimated. Using Hamiltonian Dynamics in addition, the proposed method approximates the original distribution. Experiments show the adaptive functionality can be observed, and original distribution is approximated well.