Near-Term Forecasting of Nonstationary Dynamic Processes

In order to address the DoD’s most challenging prediction and decision making problems, we present Library Augmented Reservoir Computing—LARC.   LARC is a metalearning framework that vastly reduces data requirements for predicting a nonstationary time series of interest by learning the low dimensional manifold that contains Reservoir Computers (RC) trained on a library of related processes.  Because the manifold is low dimensional, it can be efficiently searched for an RC that produces a time series that most closely matches a short observed signal.  The identified RC can then be used as the basis of future predictions.  During Phase I, we showed that LARC achieves a two orders of magnitude reduction in reconstruction error compared to state-of-the-art on a standard time series metalearning benchmark.