A Reduction of Imitation Learning and Structured Prediction to No-Regret\n Online Learning

Abstract

Sequential prediction problems such as imitation learning, where future\nobservations depend on previous predictions (actions), violate the common\ni.i.d. assumptions made in statistical learning. This leads to poor performance\nin theory and often in practice. Some recent approaches provide stronger\nguarantees in this setting, but remain somewhat unsatisfactory as they train\neither non-stationary or stochastic policies and require a large number of\niterations. In this paper, we propose a new iterative algorithm, which trains a\nstationary deterministic policy, that can be seen as a no regret algorithm in\nan online learning setting. We show that any such no regret algorithm, combined\nwith additional reduction assumptions, must find a policy with good performance\nunder the distribution of observations it induces in such sequential settings.\nWe demonstrate that this new approach outperforms previous approaches on two\nchallenging imitation learning problems and a benchmark sequence labeling\nproblem.\n