Computational framework for fast uncertainty quantification and decision analytics

点击翻译按钮获取中文摘要

We have developed a novel deep Bayesian learning architecture coupled to a Bayesian analytics engine which together form a powerful and highly flexible computational framework for decision support in complex engineering applications. We call our system the Deep Bayesian Analytics Engine (DBAE). At the fundamental level, our invention allows the discovery of actionable relationships between decision variables and performance outcomes by overcoming the curse of dimensionality in challenging, high-dimensional decision making problems. A distinguishing feature of our technology is its ability to infer the intrinsic/effective dimensionality of the problem via a statistically rigorous nonlinear decomposition method that facilitates the exploitation of the underlying problem structure for accelerated analytics computations. Specifically, our framework enables dramatic speedups in tasks that routinely arise in the computation of decision analytics such as the evaluation of high-dimensional integrals to estimate predictive statistics, search/optimization in high-dimensional spaces, decision robustness calculations, statistical calibration of computer models using data from physical experiments and onboard sensors, and the solution of inverse problems. We have evaluated a prototype software tool on challenging benchmarks and industrial-scale problems from the aircraft engine design sector. In the proposed project, we will enhance our prototype software implementation in order to begin field testing with potential partners and end users. Work will focus on scaling and accelerating our core algorithms in conjunction with the development of an improved user interface with immersive visualization capabilities. To broaden our potential market, additional testing will be carried out on novel applications in the environmental sciences and online system health monitoring sectors where there is significant demand for such advanced statistical tools.