Our research interests are focused on statistical approaches to the sequential decision problem. The multi-armed bandit (MAB) problem formulates the sequential decision problem in which a learner is sequentially faced with a set of available actions, chooses an action, and receives a random reward in response. The actions are often described as the arms of a bandit slot machine. The act
of choosing an action is characterized as pulling an arm of the bandit machine, where different arms give possibly different rewards. By repeating the process of pulling arms and receiving rewards, the learner accumulates information about the reward compensation
mechanism and learns from it, choosing the arm that is close to optimal as time elapses. In our lab, we integrate online learning and optimization techniques to develop algorithms that efficiently learn the reward model while maximizing the rewards. We also apply the developed algorithms to real tasks such as recommendation systems and mobile health apps. We also use causal inference to evaluate the performance of multi-armed bandit algorithms in a retrospective way.