7. ENSEMBLING DISCOUNTED VAW EXPERTS WITH THE VAW META-LEARNER
How to Cite:
Rokhlin, Dmitry B., and Georgiy A. Karapetyants. "Ensembling Discounted VAW Experts with the VAW
Meta-Learner." Global and Stochastic Analysis, vol. 12, no. 5, 2025, pp. 45–58.
Meta-Learner." Global and Stochastic Analysis, vol. 12, no. 5, 2025, pp. 45–58.
Abstract
The Vovk-Azoury-Warmuth (VAW) forecaster is a powerful algorithm for online regression, but its standard form is
designed for stationary environments. Recently Jacobsen and Cutkosky (2024) introduced a discounting factor, γ, to
designed for stationary environments. Recently Jacobsen and Cutkosky (2024) introduced a discounting factor, γ, to
the VAW algorithm (DVAW), enabling it to track changing concepts by down-weighting old data. They also
proposed an ensemble method for learning γ on-the-fly. In this paper we use a simplified dynamic regret bound and
employ the standard VAW forecaster as a meta-learner to dynamically aggregate the predictions of DVAW experts.
The main result contains a bound for the dynamic regret of the proposed ensemble. Computer experiments on
synthetic data show that our ensembling approach significantly outperforms both the standard VAW and individual
DVAW experts in non-stationary settings, while remaining robust and competitive in stationary ones.
Keywords
Key words and phrases. Vovk-Azoury-Warmuth algorithm; discounting; dynamic regret.