A reduction of imitation learning and structured prediction to no-regret online learning

S Ross, G Gordon, D Bagnell - Proceedings of the fourteenth …, 2011 - proceedings.mlr.press
Proceedings of the fourteenth international conference on …, 2011proceedings.mlr.press
Sequential prediction problems such as imitation learning, where future observations
depend on previous predictions (actions), violate the common iid assumptions made in
statistical learning. This leads to poor performance in theory and often in practice. Some
recent approaches provide stronger guarantees in this setting, but remain somewhat
unsatisfactory as they train either non-stationary or stochastic policies and require a large
number of iterations. In this paper, we propose a new iterative algorithm, which trains a …
Abstract
Sequential prediction problems such as imitation learning, where future observations depend on previous predictions (actions), violate the common iid assumptions made in statistical learning. This leads to poor performance in theory and often in practice. Some recent approaches provide stronger guarantees in this setting, but remain somewhat unsatisfactory as they train either non-stationary or 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 demonstrate that this new approach outperforms previous approaches on two challenging imitation learning problems and a benchmark sequence labeling problem.
proceedings.mlr.press
Showing the best result for this search. See all results