DYNAMIC ENSEMBLE FOR PROBABILISTIC TIME- SERIES FORECASTING VIA DEEP REINFORCEMENT LEARNING
Keywords: Time series, ensemble, reinforcement learning
TL;DR: We develop a general dynamic ensemble framework for probabilistic multi-horizon time series forecasting using deep reinforcement learning.
Abstract: Ensembles from given base learners are known to be indispensable in improving accuracy for most of the prediction tasks, leading to numerous methods. However, the only ensembling strategies that have been considered for time series forecasting in the past have been static methods, ones that have access to the predictions of the base learners but not to the base learners themselves. In this paper, we propose a novel \textit{dynamic ensemble policy}, which, unlike static methods, uses the power of the ensemble to improve each of the base learners being ensembled by reducing the error accumulation of each base learner via consecutively feeding an ensembled sample to each base learner. To do so, we adopt a deep Reinforcement Learning (RL) framework with a Markov Decision Process (MDP) designed where the ensemble agent interacts with our environment (\textit{TS-GYM}) from offline data. The output of our ensemble strategy is a single autoregressive forecaster that supports several desirable properties of uncertainty quantification and sample path, along with notable performance gain. The effectiveness of the proposed framework is demonstrated in multiple synthetic and real-world experiments.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics
Submission Guidelines: Yes
Please Choose The Closest Area That Your Submission Falls Into: General Machine Learning (ie none of the above)
10 Replies
Loading