
Emilie KAUFFMANN (CNRS- INRIA Lille) – Multi-objective bandits revisited
Statistical Seminar: Every Monday at 2:00 pm.
Time: 2:00 pm – 3:00 pm
Date: 26th May
Place: 3001
Emilie KAUFFMANN (CNRS- INRIA Lille) – Multi-objective bandits revisited
Abstract:
In multi-objective bandit models, arms are multi-variate distributions. Their expectations are thus vectors in Rd and there is no obvious notion of “best arm” as some arms may be better for some objective but worse for others. In this talk, I will present algorithms for the adaptive identification of the Pareto set of the arms means, in a fixed-confidence setting. These algorithms adaptively sample the arms and also adaptively stop the data collection process so as to guarantee an error at most δ for their guess of the Pareto set. I will present a first algorithm based on confidence regions that achieves a near-optimal sample complexity bound featuring some appropriate notions of “sub-optimality gaps”. Then we will discuss asymptotically optimal algorithm, i.e. algorithms whose sample complexity is matching a lower bound in the regime in which the error probability is small.
This talk is based on joint works with Cyrille Koné, Laura Richert and Marc Jourdan. https://arxiv.org/abs/2307.00424 https://arxiv.org/abs/2411.04939
Organizers:
Anna KORBA (CREST), Karim LOUNICI (CMAP) , Jaouad MOURTADA (CREST)
Sponsors:
CREST-CMAP