Monte Carlo methods for Bayesian Uncertainty Quantification

Many data-centric engineering applications often lack a critical stage, which is quantifying the uncertainty in the diagnostics. The origin of uncertainties include errors in the data, uncertainty in the adopted process models etc. Bayesian uncertainty quantification (BUQ) is, statistically speaking, the gold-standard for quantifying risk but for large structured models with many thousands of variables and terabytes of data, a BUQ computational suite is still impractical and largely an open problem in Academia.

A scalable BUQ computational suite will be designed using state of the art practice in Monte Carlo methods. The challenge posed by this type of variable and data intensive problems will be addressed using a combination of ideas: (i) devise localised inference algorithms, with limited interaction, that can largely run in parallel. (ii) Allowing for a degree of asynchrony, or a freedom to update variables independently of each other, is key to reducing idle times in a Cloud implementation. (iii) Develop cloud based BUQ methods for constrained by real-time computing budgets to deliver timely outputs.

These are related with several EPSRC research areas, namely: AI technologies, Statistics and Applied Probability, Digital Signal Processing and Theoretical Computer Science.