AI-Fun & ELLIS Invited Speaker Series | Aurelien Bibaut

Online
Wednesday 11th November 2026
Our invited speaker on 11 November will be Aurelien Bibaut from Netflix.

Title: Safe Adaptive Policy Learning

Abstract: Offline policy learning can require less information than accurate estimation of the full environment. We provide a two-step offline policy learning procedure that constructs policy proposals on one data split and selects a proposal using an independent held-out split. Two natural proposal families are the offline natural policy-gradient path and a pessimistic-planning path. Pessimistic planning fixes the radius of a likelihood-based uncertainty set and maximizes worst-case improvement. Equivalently, one can fix a target improvement (v) and choose the policy that sustains it over the largest uncertainty set. We use this second description of the same Pareto frontier: for each (v), the proposal maximizes the minimum likelihood distance from the fitted distribution to models under which its improvement over the logger is at most (v). This construction is designed to make it easier for the second-stage held-out selector to certify a good policy.
 We establish safe minimax adaptation over a sieve of tabular contextual-bandit models with known design and optimal-action coverage. Pessimistic planning based on a global observed-data likelihood confidence set at the same prescribed confidence level, and maximin selection along the fitted offline NPG path, fail to adapt to this sieve. The resulting worst-case regret separation persists even when an oracle selects the best policy along the entire NPG path. Finally, a tabular reinforcement-learning example illustrates settings in which the proposal-and-selection framework improves on global pessimistic planning and selection along the offline NPG path.

Bio: Bibaut is a senior research scientist at Netflix. He received his PhD from Berkeley under the supervision of Mark van der Laan from whom he learned semiparametric statistics. His recent research interests include reinforcement learning, post-training, instrumental variable methods and inference after adaptive data collection.