Reinforcement Learning based dynamic weighing of Ensemble Models for Time Series Forecasting
Authors:
Satheesh K. Perepu,
Bala Shyamala Balaji,
Hemanth Kumar Tanneru,
Sudhakar Kathari,
Vivek Shankar Pinnamaraju
Abstract:
Ensemble models are powerful model building tools that are developed with a focus to improve the accuracy of model predictions. They find applications in time series forecasting in varied scenarios including but not limited to process industries, health care, and economics where a single model might not provide optimal performance. It is known that if models selected for data modelling are distinc…
▽ More
Ensemble models are powerful model building tools that are developed with a focus to improve the accuracy of model predictions. They find applications in time series forecasting in varied scenarios including but not limited to process industries, health care, and economics where a single model might not provide optimal performance. It is known that if models selected for data modelling are distinct (linear/non-linear, static/dynamic) and independent (minimally correlated models), the accuracy of the predictions is improved. Various approaches suggested in the literature to weigh the ensemble models use a static set of weights. Due to this limitation, approaches using a static set of weights for weighing ensemble models cannot capture the dynamic changes or local features of the data effectively. To address this issue, a Reinforcement Learning (RL) approach to dynamically assign and update weights of each of the models at different time instants depending on the nature of data and the individual model predictions is proposed in this work. The RL method implemented online, essentially learns to update the weights and reduce the errors as the time progresses. Simulation studies on time series data showed that the dynamic weighted approach using RL learns the weight better than existing approaches. The accuracy of the proposed method is compared with an existing approach of online Neural Network tuning quantitatively through normalized mean square error(NMSE) values.
△ Less
Submitted 20 August, 2020;
originally announced August 2020.