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


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


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

Computer Science > Machine Learning

Title: Classifying Options for Deep Reinforcement Learning

Abstract: In this paper we combine one method for hierarchical reinforcement learning - the options framework - with deep Q-networks (DQNs) through the use of different "option heads" on the policy network, and a supervisory network for choosing between the different options. We utilise our setup to investigate the effects of architectural constraints in subtasks with positive and negative transfer, across a range of network capacities. We empirically show that our augmented DQN has lower sample complexity when simultaneously learning subtasks with negative transfer, without degrading performance when learning subtasks with positive transfer.
Comments: IJCAI 2016 Workshop on Deep Reinforcement Learning: Frontiers and Challenges
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Machine Learning (stat.ML)
Cite as: arXiv:1604.08153 [cs.LG]
  (or arXiv:1604.08153v3 [cs.LG] for this version)

Submission history

From: Kai Arulkumaran [view email]
[v1] Wed, 27 Apr 2016 17:48:39 GMT (204kb,D)
[v2] Wed, 8 Jun 2016 16:05:31 GMT (153kb,D)
[v3] Mon, 19 Jun 2017 15:34:58 GMT (153kb,D)

Link back to: arXiv, form interface, contact.