|
Abstract:
|
Several adaptive sampling strategies have been proved to be ergodic and tested with various benchmark examples. However, applications in real modeling cases may still provide challenges. Typical sources for the difficulties are high dimensionality, strongly correlated or non--identifiable parameters, bad initial proposal distribution or bad initial values for the sampling parameters. As a results, the adaptation might get properly started only after some hand tuning, while the very idea of adaptive schemes should be to make sampling more automatic. Here we show how adaptation may be combined with other approaches---Delayed Rejection, Reversible Jump MCMC, sampling from Gaussian mixtures---to overcome such difficulties. The real-life modeling examples come from chemical kinetics, biology, and remote sensing
|
