Contextual bandit is a widely used model for sequential decision making.

We prove that CSP learns a near-optimal risk-free offline adaptation policy upon convergence.

Reward-free data is abundant and contains rich prior knowledge of human behaviors, but it is not well exploited by offline reinforcement learning (RL) algorithms. In this paper, we propose UBER, an unsupervised approach to extract useful behaviors from offline reward-free datasets via diversified rewards.

Both challenges involve modeling long-term dependencies.