Timezone: »
While meta reinforcement learning (Meta-RL) methods have achieved remarkable success, obtaining correct and low variance estimates for policy gradients remains a significant challenge. In particular, estimating a large Hessian, poor sample efficiency and unstable training continue to make Meta-RL difficult. We propose a surrogate objective function named, Tamed MAML (TMAML), that adds control variates into gradient estimation via automatic differentiation. TMAML improves the quality of gradient estimation by reducing variance without introducing bias. We further propose a version of our method that extends the meta-learning framework to learning the control variates themselves, enabling efficient learning from a distribution of MDPs. We empirically compare our approach with MAML and other variance-bias trade-off methods including DICE, LVC, and action-dependent control variates. Our approach is easy to implement and outperforms existing methods in terms of the variance and accuracy of gradient estimation, ultimately yielding higher performance across a variety of challenging Meta-RL environments.
Author Information
Hao Liu (Salesforce Research, UC Berkeley)
Richard Socher (Salesforce)
Caiming Xiong (Salesforce)
Related Events (a corresponding poster, oral, or spotlight)
-
2019 Poster: Taming MAML: Efficient unbiased meta-reinforcement learning »
Thu Jun 13th 01:30 -- 04:00 AM Room Pacific Ballroom
More from the Same Authors
-
2020 Poster: Explore, Discover and Learn: Unsupervised Discovery of State-Covering Skills »
Victor Campos · Alexander Trott · Caiming Xiong · Richard Socher · Xavier Giro-i-Nieto · Jordi Torres -
2019 Poster: Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting »
Xilai Li · Yingbo Zhou · Tianfu Wu · Richard Socher · Caiming Xiong -
2019 Poster: On the Generalization Gap in Reparameterizable Reinforcement Learning »
Huan Wang · Stephan Zheng · Caiming Xiong · Richard Socher -
2019 Oral: Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting »
Xilai Li · Yingbo Zhou · Tianfu Wu · Richard Socher · Caiming Xiong -
2019 Oral: On the Generalization Gap in Reparameterizable Reinforcement Learning »
Huan Wang · Stephan Zheng · Caiming Xiong · Richard Socher