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

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

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

Computer Science > Machine Learning

Title: Goal-Conditioned Reinforcement Learning with Imagined Subgoals

Abstract: Goal-conditioned reinforcement learning endows an agent with a large variety of skills, but it often struggles to solve tasks that require more temporally extended reasoning. In this work, we propose to incorporate imagined subgoals into policy learning to facilitate learning of complex tasks. Imagined subgoals are predicted by a separate high-level policy, which is trained simultaneously with the policy and its critic. This high-level policy predicts intermediate states halfway to the goal using the value function as a reachability metric. We don't require the policy to reach these subgoals explicitly. Instead, we use them to define a prior policy, and incorporate this prior into a KL-constrained policy iteration scheme to speed up and regularize learning. Imagined subgoals are used during policy learning, but not during test time, where we only apply the learned policy. We evaluate our approach on complex robotic navigation and manipulation tasks and show that it outperforms existing methods by a large margin.
Comments: ICML 2021. See the project webpage at this https URL
Subjects: Machine Learning (cs.LG); Robotics (cs.RO)
Cite as: arXiv:2107.00541 [cs.LG]
  (or arXiv:2107.00541v1 [cs.LG] for this version)

Submission history

From: Elliot Chane-Sane [view email]
[v1] Thu, 1 Jul 2021 15:30:59 GMT (1163kb,D)

Link back to: arXiv, form interface, contact.