Is Pessimism Provably Efficient for Offline RL?

                   Ying Jin * 1 Zhuoran Yang * 2 Zhaoran Wang * 3

              Abstract                 Vinyals et al., 2017) relies on two ingredients: (i) expressive
   We study offline reinforcement learning (RL),         function approximators, e.g., deep neural networks (LeCun
   which aims to learn an optimal policy based on        et al., 2015), which approximate policies and values, and
   a dataset collected a priori. Due to the lack of ...