Parallel Metropolis chains with cooperative adaptation

TL;DR

This paper introduces a novel parallel MCMC scheme where chains interact and adapt cooperatively, improving efficiency by distributing computational effort and reactivating chains as needed, demonstrated through numerical simulations.

Contribution

The paper presents a new cooperative parallel MCMC algorithm with adaptive distribution of computational effort and chain reactivation, enhancing performance over traditional methods.

Findings

Improved sampling efficiency demonstrated in simulations

Adaptive effort distribution benefits convergence

Reactivation of chains enhances exploration

Abstract

Monte Carlo methods, such as Markov chain Monte Carlo (MCMC) algorithms, have become very popular in signal processing over the last years. In this work, we introduce a novel MCMC scheme where parallel MCMC chains interact, adapting cooperatively the parameters of their proposal functions. Furthermore, the novel algorithm distributes the computational effort adaptively, rewarding the chains which are providing better performance and, possibly even stopping other ones. These extinct chains can be reactivated if the algorithm considers necessary. Numerical simulations shows the benefits of the novel scheme.