← News

ResearchSources: arXiv · 18 Nov 2026

Researchers propose cancellation-free approach for constrained bandit decisions

An arXiv paper describes a modified SquareCB algorithm that keeps expected costs within budget in every round, reporting lower violations than primal-dual and virtual-queue baselines in control tests.

In plain words

The paper addresses contextual bandits, decision-making settings where a system repeatedly picks one of several actions while keeping an average cost, such as energy use, below a limit. The authors modify the SquareCB algorithm so that whenever estimated cost exceeds the budget, the cost penalty is raised just enough to pull it back. They report fewer budget violations than existing methods on two control simulation tasks.

What was announced

In constrained contextual bandits, a learner chooses one of K actions each round to minimize loss while keeping expected cost within a budget, the paper says. Existing reductions to online regression weight the cost with a Lagrange multiplier learned from past constraint residuals, meaning slack in some rounds can offset violations in others. The authors instead run SquareCB with a rectified cost weight: zero when estimated cost is within budget, otherwise just large enough to bring it back to the budget.

Under realizability and Slater's condition with margin δ, the authors report a regret bound of O(√(KT·Reg_Sq(T))/δ) and cancellation-free violation of O(√(KT·Reg_Sq(T))), where Reg_Sq is the regression oracle's error. The same bounds hold when the budget must be met in every context, including adversarially chosen contexts. In controller selection experiments on a safe-control-gym quadrotor and a BOPTEST building, they report violation 6 to 9 times lower than primal-dual and virtual-queue baselines, with lower or comparable regret.

What it does not show

  • The evidence is the paper's abstract only; the full article, proofs and experimental setup were not available for review.
  • The reported regret and violation bounds depend on assumptions (realizability, Slater's condition with margin δ) whose validity in practice is not established in the excerpt.
  • The comparison figures come from the authors' own experiments and have not been independently confirmed.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.