Non-stationary Online Learning for Curved Losses: Improved Dynamic Regret via Mixability

Authors: Yu-Jie Zhang, Peng Zhao, Masashi Sugiyama

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical Let d denote the dimensionality and PT the path length of comparators that reflects the environmental non-stationarity. We demonstrate that an exponential-weight method with fixed-share updates achieves an O(d T 1/3P 2/3 T log T) dynamic regret for mixable losses, improving upon the bestknown O(d10/3T 1/3P 2/3 T log T) result (Baby & Wang, 2021) in d. More importantly, this improvement arises from a simple yet powerful analytical framework that exploits the mixability, which avoids the Karush-Kuhn-Tucker-based analysis required by existing work. Theorem 1. Under Assumption 1, 2 and 3, for any ut W, Algorithm 1 with µ = 1/T ensures D-REGT ({ut}T t=1) O d log T (1 + T 1/3P 2/3 T ). Theorem 3. Under Assumptions 1, 4 and 5, Algorithm 3 with µ = 1/T ensures D-REGT O d log T (1 + T 1/3P 2/3 T ) for any comparator sequence u1, . . . , u T W.
Researcher Affiliation Academia 1RIKEN AIP, Japan 2National Key Laboratory for Novel Software Technology, Nanjing University, China 3The University of Tokyo, Japan. Correspondence to: Yu-Jie Zhang <EMAIL>, Peng Zhao <EMAIL>.
Pseudocode Yes Algorithm 1 Fixed-share For Continuous Space Algorithm 2 Follow-the-Leading-History Algorithm 3 Projected Fixed-share with Surrogate Loss
Open Source Code No A future direction is to develop computationally more efficient methods for the logistic loss and the general OCO setting.
Open Datasets No No specific datasets are mentioned for empirical evaluation. The paper is theoretical in nature and analyzes different loss functions (squared loss, logistic loss, etc.) without performing experiments on specific datasets.
Dataset Splits No No information about dataset splits is provided as the paper does not conduct empirical experiments with data.
Hardware Specification No No specific hardware details are mentioned as the paper is theoretical and does not report experimental results.
Software Dependencies No No specific software dependencies with version numbers are mentioned, as the paper is theoretical and does not describe an implementation.
Experiment Setup No No specific experimental setup details, such as hyperparameter values or training configurations, are provided as the paper is theoretical.