No-Regret Reductions for Imitation Learning and Structured Prediction

Sequential prediction problems such as imitation learning, where future observations depend on previous predictions (actions), violate the common i.i.d. assumptions made in statistical learning. This leads to poor performance in both theory and often in practice. Some recent approaches provide stronger performance guarantees in this setting, but remain somewhat unsatisfactory as they train either non-stationary or a stochastic policies and require a large number of iterations. In this paper, we propose a new iterative algorithm, which trains a stationary deterministic policy, that can be seen as a no regret algorithm in an online learning setting. We show that any such no regret algorithm, combined with additional reduction assumptions, must find a policy with good performance under the distribution of observations it induces in such sequential settings. We additionally show that this new approach outperforms previous approaches on two challenging imitation learning problem and a benchmark sequence labeling problem.