Cost-aware non-stationary online learning

Abstract

Online learning of an ensemble of classifiers or regressors is performed to predict a non-stationary time-varying parameter over a time series of episodes. For an episode, an ensemble action is selected from a set of ensemble actions based on ensemble action quality values (Q values) at an error state for the episode. The selected ensemble action is executed to update the ensemble. A cost of executing the selected ensemble action is computed or retrieved, and a reward is computed indicating how well the updated ensemble predicts the non-stationary time-varying parameter over the episode. The Q value for the selected ensemble action at the error state for the episode is updated based on both the reward and the cost of executing the selected ensemble action. The cost may be based on the ensemble action alone, or on both the ensemble action and the error state for the episode.