Thompson Sampling (TS) for combinatorial semi-bandits is sub-optimal in the sense that its regret scales exponentially in the ambient dimension for some well chosen exemples.
Dec 7, 2021
