TR2019-120

Quasi-Newton Trust Region Policy Optimization


We propose a trust region method for policy optimization that employs Quasi-Newton approximation for the Hessian, called Quasi-Newton Trust Region Policy Optimization (QNTRPO). Gradient descent is the de facto algorithm for reinforcement learning tasks with continuous controls. The algorithm has achieved state-of-the-art performance when used in reinforcement learning across a wide range of tasks. However, the algorithm suffers from a number of drawbacks including: lack of stepsize selection criterion, and slow convergence. We investigate the use of a trust region method using dogleg step and a Quasi-Newton approximation for the Hessian for policy optimization. We demonstrate through numerical experiments over a wide range of challenging continuous control tasks that our particular choice is efcient in terms of number of samples and improves performance.

 

  • Software Download

  • Related News & Events

  • Related Publications

  •  Jha, D., Raghunathan, A., Romeres, D., "Quasi-Newton Trust Region Policy Optimization", arXiv, December 2019.
    BibTeX arXiv
    • @article{Jha2019dec3,
    • author = {Jha, Devesh and Raghunathan, Arvind and Romeres, Diego},
    • title = {Quasi-Newton Trust Region Policy Optimization},
    • journal = {arXiv},
    • year = 2019,
    • month = dec,
    • url = {https://arxiv.org/abs/1912.11912}
    • }
  •  Jha, D., Raghunathan, A., Romeres, D., "QNTRPO: Including Curvature in TRPO", Optimization Foundations for Reinforcement Learning Workshop at NeurIPS, December 2019.
    BibTeX TR2019-154 PDF Software
    • @inproceedings{Jha2019dec,
    • author = {Jha, Devesh and Raghunathan, Arvind and Romeres, Diego},
    • title = {QNTRPO: Including Curvature in TRPO},
    • booktitle = {Optimization Foundations for Reinforcement Learning Workshop at NeurIPS},
    • year = 2019,
    • month = dec,
    • url = {https://www.merl.com/publications/TR2019-154}
    • }