We derive a new analysis of Follow The Regularized Leader (FTRL) for online learning with delayed bandit feedback. By separating the cost of delayed feedback from that of bandit feedback, our analysis allows us to obtain new results in three important settings. On the one hand, we derive the first optimal (up to logarithmic factors) regret bounds for combinatorial semi-bandits with delay and adversarial Markov decision processes with delay (and known transition functions). On the other hand, we use our analysis to derive an efficient algorithm for linear bandits with delay achieving near-optimal regret bounds. Our novel regret decomposition shows that FTRL remains stable across multiple rounds under mild assumptions on the Hessian of the regularizer.
A Unified Analysis of Nonstochastic Delayed Feedback for Combinatorial Semi-Bandits, Linear Bandits, and MDPs / D. van der Hoeven, L. Zierahn, T. Lancewicki, A. Rosenberg, N. Cesa Bianchi (PROCEEDINGS OF MACHINE LEARNING RESEARCH). - In: Proceedings of Thirty Sixth Conference on Learning Theory / [a cura di] G. Neu, L. Rosasco. - [s.l] : PMLR, 2023. - pp. 1285-1321 (( Intervento presentato al 36. convegno Annual Conference on Learning Theory tenutosi a Bangalore nel 2023.
A Unified Analysis of Nonstochastic Delayed Feedback for Combinatorial Semi-Bandits, Linear Bandits, and MDPs
L. Zierahn;N. Cesa Bianchi
2023
Abstract
We derive a new analysis of Follow The Regularized Leader (FTRL) for online learning with delayed bandit feedback. By separating the cost of delayed feedback from that of bandit feedback, our analysis allows us to obtain new results in three important settings. On the one hand, we derive the first optimal (up to logarithmic factors) regret bounds for combinatorial semi-bandits with delay and adversarial Markov decision processes with delay (and known transition functions). On the other hand, we use our analysis to derive an efficient algorithm for linear bandits with delay achieving near-optimal regret bounds. Our novel regret decomposition shows that FTRL remains stable across multiple rounds under mild assumptions on the Hessian of the regularizer.File | Dimensione | Formato | |
---|---|---|---|
hoeven23a.pdf
accesso aperto
Tipologia:
Publisher's version/PDF
Dimensione
473.54 kB
Formato
Adobe PDF
|
473.54 kB | Adobe PDF | Visualizza/Apri |
Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.