Title: Adaptive online model selection for linear and non-linear AR
Authors: Stephane Chretien - NPL (United Kingdom) [presenting]
Abstract: Model selection is an important task which may have a tremendous impact on time series forecasting. Unfortunately, model selection can be computationaly expensive, especially in the case of non-linear models such as models governed by complex deep networks. We present a new approach to model selection using the Hedge algorithm. We will analyse the theoretical performance of the method in the linear case and in some non-linear cases. We will show computational experiments that demonstrate the effectiveness of the proposed approach in various settings. In particular, experimental results using a recent approach will be presented.