FLARE MCMC Boosts Sampling Efficiency in Complex Models.

Markov chain Monte Carlo (MCMC) is a widely used technique for inference in intricate models, relying solely on the ability to evaluate likelihoods. However, its primary drawback is often a slow mixing rate, which necessitates generating a large number of samples to achieve accurate estimates, leading to high computational expense. A novel approach, FLARE MCMC, addresses this by introducing a multi-fidelity layered MCMC method. This technique leverages less precise, lower-fidelity approximations of the true likelihood calculation. By doing so, it enhances the mixing rate of the MCMC process, resulting in overall faster performance. FLARE MCMC is particularly effective in scientific and engineering fields where models frequently involve simulations with adjustable resolution or accuracy. The method employs recursive, layered chains with straightforward tuning and does not impose specific mathematical structures on the likelihood function. Experimental results across domains like hydrology and cosmology demonstrate that FLARE MCMC yields larger effective sample sizes for the same computational time.