Parallel Markov Chain Monte Carlo Methods for Large Scale Statistical Inverse Problems
The Bayesian method has proven to be a powerful way of modeling inverse problems. The solution to Bayesian inverse problems is the posterior distribution of estimated parameters which can provide not only estimates for the inferred parameters but also the uncertainty of these estimations. Markov chain Monte Carlo (MCMC) is a useful technique to sample the posterior distribution and information can be extracted from the sampled ensemble. However, MCMC is very expensive to compute, especially in inverse problems where the underlying forward problems involve solving differential equations. Even worse, MCMC is difficult to parallelize due to its sequential nature|that is, under the current framework, we can barely accelerate MCMC with parallel computing.
We develop a new framework of parallel MCMC algorithms-the Markov chain preconditioned Monte Carlo (MCPMC) method-for sampling Bayesian inverse problems. With the help of a fast auxiliary MCMC chain running on computationally cheaper approximate models, which serves as a stochastic preconditioner to the target distribution, the sampler randomly selects candidates from the preconditioning chain for further processing on the accurate model. As this accurate model processing can be executed in parallel, the algorithm is suitable for parallel systems. We implement it using a modified master-slave architecture, analyze its potential to accelerate sampling and apply it to three examples. A two dimensional Gaussian mixture example shows that the new sampler can bring statistical efficiency in addition to increasing sampling speed. Through a 2D inverse problem with an elliptic equation as the forward model, we demonstrate the use of an enhanced error model to build the preconditioner. With a 3D optical tomography problem we use adaptive finite element methods to build the approximate model. In both examples, the MCPMC successfully samples the posterior distributions with multiple processors, demonstrating efficient speedups comparing to traditional MCMC algorithms. In addition, the 3D optical tomography example shows the feasibility of applying MCPMC towards real world, large scale, statistical inverse problems.