🤖 AI Summary
Optimistic Online Learning relies on expert advice to improve predictive performance, yet experts—often constructed via approximate minimization—are prone to noise and incomplete information in dynamic environments; uncritical adoption thus risks overfitting and degraded robustness. To address this, we propose the Optimistic Temperance (OT) framework, the first systematic approach to question the unconditional validity of expert advice. OT introduces a gradient-driven adaptive weighting mechanism that dynamically calibrates expert credibility, preserving optimistic heuristics while enhancing robustness. Theoretically, OT achieves a tight dynamic regret bound of (O(sqrt{T(1+V_T)})), where (V_T) quantifies environmental non-stationarity. Empirically, OT consistently outperforms standard optimistic algorithms across diverse online learning tasks—including online convex optimization, prediction with expert advice, and online portfolio selection—demonstrating superior adaptability and stability under distributional shifts.
📝 Abstract
Optimistic Online Learning algorithms have been developed to exploit expert advices, assumed optimistically to be always useful. However, it is legitimate to question the relevance of such advices emph{w.r.t.} the learning information provided by gradient-based online algorithms. In this work, we challenge the confidence assumption on the expert and develop the emph{optimistically tempered} (OT) online learning framework as well as OT adaptations of online algorithms. Our algorithms come with sound theoretical guarantees in the form of dynamic regret bounds, and we eventually provide experimental validation of the usefulness of the OT approach.
