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: Automated curriculum generation for Policy Gradients from Demonstrations

Abstract: In this paper, we present a technique that improves the process of training an agent (using RL) for instruction following. We develop a training curriculum that uses a nominal number of expert demonstrations and trains the agent in a manner that draws parallels from one of the ways in which humans learn to perform complex tasks, i.e by starting from the goal and working backwards. We test our method on the BabyAI platform and show an improvement in sample efficiency for some of its tasks compared to a PPO (proximal policy optimization) baseline.
Comments: Accepted to Deep RL Workshop at NeurIPS 2019
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Machine Learning (stat.ML)
Cite as: arXiv:1912.00444 [cs.LG]
  (or arXiv:1912.00444v1 [cs.LG] for this version)

Submission history

From: Anirudh Srinivasan [view email]
[v1] Sun, 1 Dec 2019 17:08:34 GMT (1175kb,D)

Link back to: arXiv, form interface, contact.