Speed-up of posterior inference of highly-parameterized environmental models from a Kalman proposal distribution: DREAM(KZS)
Jiangjiang Zhang,Jasper A. Vrugt,Xiaoqing Shi,Guang Lin,Lingzao Zeng,Laosheng Wu
2017-01-01
Abstract:Author(s): Zhang, J; Vrugt, JA; Shi, X; Lin, G; Zeng, L; Wu, L | Abstract: Markov chain Monte Carlo (MCMC) simulation methods are widely used to generate samples from a target distribution. In posterior inference of highly-parameterized environmental models, the convergence speed of MCMC methods may be disturbingly low, even with the state-of-the-art algorithms, such as DREAM(ZS) (differential evolution adaptive Metropolis). At each iteration, DREAM(ZS) generates the proposal distributions with a mix of parallel direction jump and snooker jump that are only based on the information about the model parameters in the thinned chain history. In this study, to speed up the convergence of DREAM(ZS), we introduce a Kalman proposal distribution that utilizes the information contained in the covariance structure of the model parameters, the measurements and the model outputs. Compared with the parallel direction jump and the snooker jump, the Kalman jump can generate a more directional update of the model parameters. As the Kalman jump cannot maintain detailed balance, we restrict it only to the burn-in period and use the other two jumps with diminishing adaptation afterwards. The modified algorithm is called DREAM(KZS) as it uses the three jumps simultaneously with pre-defined probabilities. Numerical experiments demonstrate that DREAM(KZS) converges to the same posterior distribution as DREAM(ZS) but with much lower computational budget. Specifically, in problems with about 100 unknown model parameters, the saving can be as big as 20 times.