Full text loading...
-
Improving Markov chain Monte Carlo efficieny with resolution constraints
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, 79th EAGE Conference and Exhibition 2017, Jun 2017, Volume 2017, p.1 - 3
Abstract
Probabilistic inverse methods for large non-linear inverse problems, such as seismic traveltime tomography, provide a full overview of the posterior distribution. They require smart and computationally expensive methods, like Markov-Chain-Monte-Carlo, short McMC, to sample the relevant parts of the model space. Each new model in the chain is created by adding one random perturbation to the last model. Any random perturbation will inevitably introduce a bias on the global traveltimes of the current model. However, if the current model is already good, which is reasonable for chain members, it is unlikely to have a global bias of traveltimes. Introducing one will probably degrade the model’s likelihood. In this study compensations for these perturbations are introduced in order to increase acceptance rates, step length and prevent or reduce this bias. These compensations are based on the resolution matrix and are done in slowness. The resolution matrix describes the relationship between true and calculated model. It also shows the trade-off between model parameters. If parameters are not uniquely determined, like the resolution matrix indicates, random perturbations should affect all correlating parameters. Hence the information about the trade-off is used to modify the McMC perturbations with compensations based on the resolution matrix.