Non-Myopic Batch Bayesian Experimental Design for Quantifying Statistical Expectation
Keywords: Bayesian experimental design, uncertainty quantification
Abstract: In this work, we develop a non-myopic batch Bayesian experimental design for statistical expectation. The next batch of samples is selected to maximize the long-term information gain (as the acquisition) when they are added together. In addition, we formulate an analytic approximation of the acquisition to facilitate its optimization. The superior performance of the proposed algorithm, in terms of wall time saving and a faster or matched convergence rate than sequential sampling, is demonstrated in a case with arbitrary complex functions generated by RBF kernel and another case using a stochastic oscillator.
Submission Number: 101
Loading