5 Real-Time Prediction Using Kernel Methods and Data Assimilation

Creating new knowledge through analysis of massive data sets brings about a profound positive impact on society. Data streams are created from a multitude of sources (e.g., sensors, models) and then are compiled into heterogeneous sets of information. The users of these data may be modelers who have specific requirements to update numerical models dynamically, through assimilation techniques. Assimilation is problematic because linear techniques, such as Kalman filters, are applied to nonlinear dynamics. We propose an innovative approach to ameliorate these problems and provide scalable algorithms whose computational complexity is much lower than with traditional methods. Our research uses support vector machines and other kernel methods for data mining (e.g., data thinning) to accomplish these tasks. Computational results on a free fall model were highly encouraging.