contextual bandit algorithms are essential for solving many real-world
interactive machine learning problems. Despite multiple recent successes on
statistically and computationally efficient methods, the practical behavior of
these algorithms is still poorly understood. We leverage the