Parallel Solution of DDDAS Variational Inference Problems

Inference problems in dynamically data-driven application systems use physical measurements along with a physical model to estimate the parameters or state of a physical system. Developing parallel algorithms to solve inference problems can improve the process of estimating and predicting the physical state of a system. Solution to inference problems using the variational approach require multiple evaluations of the associated cost function and gradient, where the gradient is defined as the increase/decrease inflection point of the variable between two points. In this paper we present a scalable algorithm based on augmented Lagrangian approach to solve the variational inference problem. The augmented Lagrangian framework facilitates parallel cost function and gradient computations. We show that the methodology is highly scalable with increasing problem size by applying it for the Lorenz-96 model