カンファレンス (国内) Adversarial federated contextual bandits with large actions in packs

Yaxiong Liu (Hokkaido University/RIKEN), Seng Pei Liew, Yang Cao (Hokkaido University), Atsuyoshi Nakamura (Hokkaido University), Tsubasa Takahashi

第26回情報論的学習理論ワークショップ (IBIS2023)


In this paper, we study the problem of large action space in the context of adversarial federated contextual bandits. First of all, we design a framework where all the actions are stored in packs on the central server side. The local device processes only a pack of actions allocated by the server at each round, reducing the running time and storage consumption of each device. Next, we propose an algorithm for our learning framework, where we make a trade-off between the communication cost and the approximation to the global optimum for all actions in hindsight. Our algorithm is tight with respect to the learning round $T$ as $\Theta(T^{2/3})$.