Paper: arXiv 2603.28198
Authors: Hongkai Hu
Abstract
Static regret to a single expert is often the wrong target for strictly online prediction under non-stationarity, where the best expert may switch repeatedly over time. We study Policy-Controlled Generalized Share (PCGS), a general strictly online framework in which the generalized-share recursion is fixed while the post-loss update controls are allowed to vary adaptively. Its principal instantiation in this paper is PCGS-TF, which uses a causal Transformer as an update controller: after round t finishes and the loss vector is observed, the Transformer outputs the controls that map w_t to w_{t+1} without altering the already committed decision w_t. Under admissible post-loss update controls, we obtain a pathwise weighted regret guarantee for general time-varying learning rates, and a standard dynamic-regret guarantee against any expert path with at most S switches under the constant-learning-rate specialization. Empirically, on a controlled synthetic suite with exact dynamic-programming switching-oracle evaluation, PCGS-TF attains the lowest mean dynamic regret in all seven non-stationary families, with its advantage increasing for larger expert pools. On a reproduced household-electricity benchmark, PCGS-TF also achieves the lowest normalized dynamic regret for S = 5, 10, and 20.
Complexity vs Empirical Score
- Math Complexity: 8.0/10
- Empirical Rigor: 7.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper introduces an innovative hybrid framework (PCGS-TF) combining non-stationary online learning theory with causal Transformers for adaptive control. It provides strong theoretical guarantees with pathwise weighted regret proofs and evaluates performance rigorously against exact dynamic-programming switching-oracles on synthetic and benchmark datasets.
Research Flowchart
flowchart TD
A["Research Goal: Optimize strictly online prediction under non-stationarity via dynamic expert switching"] --> B["Framework & Model: Policy-Controlled Generalized Share with Causal Transformer (PCGS-TF)"]
subgraph Execution ["Input & Process"]
C["Inputs: Streaming Loss Vectors & Time-Varying Learning Rates"] --> D["Computation: Transformer outputs post-loss controls to map w_t to w_t+1 without altering w_t"]
End
B --> C
D --> E["Theoretical Outcome: Pathwise weighted & dynamic regret guarantees for S switches"]
D --> F["Empirical Outcomes: Lowest dynamic regret in 7 synthetic families & household electricity benchmark"]
Cited References (51)
- Prediction, Learning, and Games (2006)
- The Weighted Majority Algorithm (1994)
- A Decision-Theoretic Generalization of On-Line Learning and an Application to Boosting (1997)
- Tracking the Best Expert (1998)
- Online Convex Programming and Generalized Infinitesimal Gradient Ascent (2003)
- Non-stationary Stochastic Optimization (2015)
- Attention Is All You Need (2017)
- Empirical Properties of Asset Returns: Stylized Facts and Statistical Issues (2001)
and 43 more
- Robust Estimation of a Location Parameter (1964)
- Challenging the Empirical Mean and Empirical Variance: A Deviation Study (2012)
- Language Models are Few-Shot Learners (2020) — link
- Tracking the Best Expert (1998)
- How to Use Expert Advice (1997)
- Online Convex Programming and Generalized Infinitesimal Gradient Ascent (2003)
- Online Learning and Online Convex Optimization (2012)
- Prediction, Learning, and Games (2006)
- Introduction to Online Convex Optimization (2016)
- Stochastic Multi-Armed Bandit Problem with Non-stationary Rewards (2014)
- Online Learning in Changing Environments: A Survey (2020)
- Parameter-free Online Learning via Model Selection (2020)
- Worst-Case Optimal Dynamic Regret in Non-stationary Online Learning (2024)
- Efficient Methods for Non-stationary Online Learning (2025)
- Projection-free Non-stationary Online Learning (2024)
- Online Non-convex Learning in Dynamic Environments (2024)
- Regret Bounds for Stochastic Composite Mirror Descent with Non-smooth Losses (2022)
- Attention Is All You Need (2017)
- Decision Transformer: Reinforcement Learning via Sequence Modeling (2021)
- What Can Transformers Learn In-Context? A Case Study of Simple Function Classes (2022)
- Learning to Learn in Context: Gradient Descent with Transformers (2023)
- Transformers Learn In-Context by Gradient Descent (2023)
- Rethinking the Role of Demonstrations: What Makes In-Context Learning Work? (2022)
- A Survey on In-context Learning (2024)
- Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting (2021)
- Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting (2021)
- FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting (2022)
- Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting (2022)
- A Time Series is Worth 64 Words: Long-term Forecasting with Transformers (2023)
- iTransformer: Inverted Transformers Are Effective for Time Series Forecasting (2024)
- TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis (2023)
- Are Transformers Effective for Time Series Forecasting? (2023)
- Transformers in Time Series: A Survey (2023)
- RobustTSF: Towards Theory and Design of Robust Time Series Forecasting with Anomalies (2024)
- Parameter-free Regret in High Probability with Heavy Tails (2022)
- Challenging the Empirical Mean and Empirical Variance: A Deviations Study (2012)
- WILDS: A Benchmark of in-the-Wild Distribution Shifts (2021)
- In Search of Lost Domain Generalization (2021)
- WOODS: A Benchmark for Out-of-Distribution Generalization in Time Series (2025)
- Empirical Asset Pricing via Machine Learning (2020)
- A New Approach to the Economic Analysis of Nonstationary Time Series and the Business Cycle (1989)
- GARCH 101: The Use of ARCH/GARCH Models in Applied Econometrics (2001)
- The ARCH Model (2008)
0 of 51 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2518 linked papers.