Title: Partial mixing and asymptotic expansion for batched bandits
Authors: Nakahiro Yoshida - University of Tokyo (Japan) [presenting]
Abstract: The asymptotic expansion method based on partial mixing was proposed previously and applied to the asymptotic expansion of the additive functional of a partially mixing epsilon-Markov process, such as a jump-diffusion process satisfying a stochastic differential equation in the random environment. We discuss an application of this scheme to the asymptotic expansion of an estimator appearing in the batched bandits. In the batched bandit, the environment of each stage is randomly set according to the random outcomes of the previous stage. We introduce a backwards asymptotic expansion formula and assess the backward propagation of errors.