Meta-Learning for Simple Regret Minimization

We develop a meta-learning framework for simple regret minimization in bandits. In this framework, learning agent interacts with sequence of bandit tasks, which are sampled i.i.d. from an unknown prior distribution, and learns its meta-parameters to perform better on future tasks. propose the first Bayesian frequentist algorithms setting. The algorithm has access distribution over meta m tasks horizon n is mere O(m / √n). On other hand, O(n√m + m/ While worse, more general because it does not need meta-parameters. It can also be analyzed settings. instantiate our several classes problems. Our we complement theory by evaluating them empirically environments.