Of Moments and Matching: A Game-Theoretic Framework for Closing the Imitation Gap

Gokul Swamy, Sanjiban Choudhury, J. Andrew Bagnell, Steven Wu
Proceedings of the 38th International Conference on Machine Learning, PMLR 139:10022-10032, 2021.

Abstract

We provide a unifying view of a large family of previous imitation learning algorithms through the lens of moment matching. At its core, our classification scheme is based on whether the learner attempts to match (1) reward or (2) action-value moments of the expert’s behavior, with each option leading to differing algorithmic approaches. By considering adversarially chosen divergences between learner and expert behavior, we are able to derive bounds on policy performance that apply for all algorithms in each of these classes, the first to our knowledge. We also introduce the notion of moment recoverability, implicit in many previous analyses of imitation learning, which allows us to cleanly delineate how well each algorithmic family is able to mitigate compounding errors. We derive three novel algorithm templates (AdVIL, AdRIL, and DAeQuIL) with strong guarantees, simple implementation, and competitive empirical performance.

Cite this Paper


BibTeX
@InProceedings{pmlr-v139-swamy21a, title = {Of Moments and Matching: A Game-Theoretic Framework for Closing the Imitation Gap}, author = {Swamy, Gokul and Choudhury, Sanjiban and Bagnell, J. Andrew and Wu, Steven}, booktitle = {Proceedings of the 38th International Conference on Machine Learning}, pages = {10022--10032}, year = {2021}, editor = {Meila, Marina and Zhang, Tong}, volume = {139}, series = {Proceedings of Machine Learning Research}, month = {18--24 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v139/swamy21a/swamy21a.pdf}, url = {https://proceedings.mlr.press/v139/swamy21a.html}, abstract = {We provide a unifying view of a large family of previous imitation learning algorithms through the lens of moment matching. At its core, our classification scheme is based on whether the learner attempts to match (1) reward or (2) action-value moments of the expert’s behavior, with each option leading to differing algorithmic approaches. By considering adversarially chosen divergences between learner and expert behavior, we are able to derive bounds on policy performance that apply for all algorithms in each of these classes, the first to our knowledge. We also introduce the notion of moment recoverability, implicit in many previous analyses of imitation learning, which allows us to cleanly delineate how well each algorithmic family is able to mitigate compounding errors. We derive three novel algorithm templates (AdVIL, AdRIL, and DAeQuIL) with strong guarantees, simple implementation, and competitive empirical performance.} }
Endnote
%0 Conference Paper %T Of Moments and Matching: A Game-Theoretic Framework for Closing the Imitation Gap %A Gokul Swamy %A Sanjiban Choudhury %A J. Andrew Bagnell %A Steven Wu %B Proceedings of the 38th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2021 %E Marina Meila %E Tong Zhang %F pmlr-v139-swamy21a %I PMLR %P 10022--10032 %U https://proceedings.mlr.press/v139/swamy21a.html %V 139 %X We provide a unifying view of a large family of previous imitation learning algorithms through the lens of moment matching. At its core, our classification scheme is based on whether the learner attempts to match (1) reward or (2) action-value moments of the expert’s behavior, with each option leading to differing algorithmic approaches. By considering adversarially chosen divergences between learner and expert behavior, we are able to derive bounds on policy performance that apply for all algorithms in each of these classes, the first to our knowledge. We also introduce the notion of moment recoverability, implicit in many previous analyses of imitation learning, which allows us to cleanly delineate how well each algorithmic family is able to mitigate compounding errors. We derive three novel algorithm templates (AdVIL, AdRIL, and DAeQuIL) with strong guarantees, simple implementation, and competitive empirical performance.
APA
Swamy, G., Choudhury, S., Bagnell, J.A. & Wu, S.. (2021). Of Moments and Matching: A Game-Theoretic Framework for Closing the Imitation Gap. Proceedings of the 38th International Conference on Machine Learning, in Proceedings of Machine Learning Research 139:10022-10032 Available from https://proceedings.mlr.press/v139/swamy21a.html.

Related Material