Cyclical Variational Bayes Monte Carlo for Efficient Multi-Modal Posterior Distributions Evaluation
share
Statistical model updating is frequently used in engineering to calculate the uncertainty of some unknown latent parameters when a set of measurements on observable quantities is given. Variational inference is an alternative approach to sampling methods that has been developed by the machine learning community to estimate posterior approximations through an optimization approach. In this paper, the Variational Bayesian Monte Carlo (VBMC) method is investigated with the purpose of dealing with statistical model updating problems in engineering involving expensive-to-run models. This method combines the active-sampling Bayesian quadrature with a Gaussian-process based variational inference to yield a non-parametric estimation of the posterior distribution of the identified parameters involving few runs of the expensive-to-run model. VBMC can also be used for model selection as it produces an estimation of the model's evidence lower bound. In this paper, a variant of the VBMC algorithm is developed through the introduction of a cyclical annealing schedule into the algorithm. The proposed cyclical VBMC algorithm allows to deal effectively with multi-modal posteriors by having multiple cycles of exploration and exploitation phases. Four numerical examples are used to compare the standard VBMC algorithm, the monotonic VBMC, the cyclical VBMC and the Transitional Ensemble Markov Chain Monte Carlo (TEMCMC). Overall, it is found that the proposed cyclical VBMC approach yields accurate results with a very reduced number of model runs compared to the state of the art sampling technique TEMCMC. In the presence of potential multi-modal problems, the proposed cyclical VBMC algorithm outperforms all the other approaches in terms of accuracy of the resulting posterior.
READ FULL TEXT