TR2021-108

Model-Based Reinforcement Learning Using Monte Carlo Gradient Estimation

    •  Amadio, F., Dalla Libera, A., Carli, R., Romeres, D., "Model-Based Reinforcement Learning Using Monte Carlo Gradient Estimation", Automatica.it, September 2021.
      BibTeX TR2021-108 PDF
      • @inproceedings{Amadio2021sep,
      • author = {Amadio, Fabio and Dalla Libera, Alberto and Carli, Ruggero and Romeres, Diego},
      • title = {Model-Based Reinforcement Learning Using Monte Carlo Gradient Estimation},
      • booktitle = {Automatica.it},
      • year = 2021,
      • month = sep,
      • url = {https://www.merl.com/publications/TR2021-108}
      • }
  • MERL Contact:
  • Research Areas:

    Control, Robotics

Abstract:

We propose an MBRL algorithm named Monte Carlo Probabilistic Inference for Learning COntrol (MC-PILCO). MC-PILCO is a policy gradient algorithm, which uses GPs to model the system dynamics, but it overcomes PILCO’s limitations by relying on a particle-based method to compute long-term predictions, instead of using moment matching.