Online Learning and Bandits with Queried Hints
summary
Algorithm queries an ML model with two candidate actions (or arms) and plays the chosen one, coupling differentially private Laplace noise with a regularized-leader coupling to prove robustness: adversarial online linear/convex optimization gets O(sqrt(B+1)) regret when at most B hints are wrong (and constant regret if hints are perfect). For stochastic bandits the method runs UCB-V on meta-arms (pairs), yielding O(n^2 log T) regret; k‑probe variants remove T or change dependence on n.
AI-written summary · about
chapters (5)
related talks
32mMulti-Player Multi-Armed Bandit: Can We Still Collaborate at Homes Without "Zoom"?
01h 21mAdversarial Bandits: Theory and Algorithms
01h 01mStochastic Bandits: Foundations and Current Perspectives
44mParsimonious Learning-Augmented Algorithms
39mBandits and Agents: How to Incentivize Exploration?
01h 15mBridging Stochastic and Adversarial Bandits
04mOn gradient-based methods for finding game-theoretic equilibria
43mOnline Learning via Offline Greedy Algorithms: Applications in Market Design and Optimization