PROJECT TITLE :

Online Multi-Task Learning Framework for Ensemble Forecasting - 2017

ABSTRACT:

Ensemble forecasting may be a widely-used numerical prediction technique for modeling the evolution of nonlinear dynamic systems. To predict the longer term state of such systems, a set of ensemble member forecasts is generated from multiple runs of laptop models, where each run is obtained by perturbing the starting condition or employing a totally different model representation of the system. The ensemble mean or median is typically chosen as a point estimate for the ensemble member forecasts. These approaches are restricted in that they assume every ensemble member is equally skillful and might not preserve the temporal autocorrelation of the anticipated time series. To overcome these limitations, we tend to gift an on-line multi-task learning framework referred to as ORION to estimate the optimal weights for combining the ensemble member forecasts. Unlike alternative existing formulations, the proposed framework is novel in that its learning algorithm should backtrack and revise its previous forecasts before creating future predictions if the sooner forecasts were incorrect when verified against new observation information. We tend to termed this strategy as online learning with restart. Our proposed framework employs a graph Laplacian regularizer to confirm consistency of the anticipated time series. It will also accommodate completely different types of loss functions, together with ?-insensitive and quantile loss functions, the latter of that is particularly helpful for extreme value prediction. A theoretical proof demonstrating the convergence of our algorithm is additionally given. Experimental results on seasonal soil moisture forecasts from 12 major river basins in North America demonstrate the superiority of ORION compared to different baseline algorithms.