Reproduce “Doubly-Bounded Queue for Constrained Online Learning: Keeping Pace with Dynamics of Both Loss and Constraint” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Juncheng Wang, Bingjie Yan, Yituo Liu
An online convex optimization algorithm, COLDQ, introduces a novel doubly-bounded virtual queue to manage time-varying losses and constraints, achieving dynamic regret bounds of O(T^(1+Vx)/2) and hard constraint violation bounds of O(T^Vg). It demonstrates performance that smoothly approaches optimal bounds as problem dynamics decrease and empirically reduces hard constraint violation by over 40% in some time-varying scenarios without requiring the Slater condition.
Key insights
- Enforcing both a lower and an upper bound on virtual queues provides tighter control over hard constraint violations in dynamic online learning.
- An algorithm can achieve performance bounds for dynamic regret and hard constraint violation that smoothly approach static and fixed-constraint optimal bounds as problem dynamics decrease, a property termed 'keeping pace'.
- Effective constrained online learning can be achieved without requiring the Slater condition, broadening the applicability of OCO algorithms to more complex and realistic scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.