Slides

Speaker: Tian Xu

Time: Mar 26 2pm-5pm

Short Abstract: We discuss the statistical limits of imitation learning (IL). In IL, the agent cannot observe the rewards but has an expert dataset in advance. We first introduce a famous IL algorithm, Behavioral Cloning (BC), and establish the sub-optimality upper bound of BC. Furthermore, we establish a suboptimality lower bound under the no-interaction setting, which suggests that BC is minimax optimal.

Reference:

Rajaraman, Nived, et al. “Toward the Fundamental Limits of Imitation Learning.” arXiv preprint arXiv:2009.05990 (2020).