We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

stat.ML

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Statistics > Machine Learning

Title: Imitation Learning by Reinforcement Learning

Authors: Kamil Ciosek
Abstract: Imitation learning algorithms learn a policy from demonstrations of expert behavior. We show that, for deterministic experts, imitation learning can be done by reduction to reinforcement learning with a stationary reward. Our theoretical analysis both certifies the recovery of expert reward and bounds the total variation distance between the expert and the imitation learner, showing a link to adversarial imitation learning. We conduct experiments which confirm that our reduction works well in practice for continuous control tasks.
Comments: Published in ICLR 2022
Subjects: Machine Learning (stat.ML); Machine Learning (cs.LG)
Cite as: arXiv:2108.04763 [stat.ML]
  (or arXiv:2108.04763v2 [stat.ML] for this version)

Submission history

From: Kamil Ciosek [view email]
[v1] Tue, 10 Aug 2021 16:14:41 GMT (191kb,D)
[v2] Tue, 15 Mar 2022 14:39:22 GMT (401kb,D)

Link back to: arXiv, form interface, contact.