This is an un-official PyTorch implementation of a very interesting and groundbreaking paper from S. Grod and M. Zanon:

@article{ENMPC-RL,
  author={S. Gros and M. Zanon},
  journal={IEEE Transactions on Automatic Control}, 
  title={Data-Driven Economic NMPC Using Reinforcement Learning}, 
  year={2020},
  volume={65},
  number={2},
  pages={636-648},
  doi={10.1109/TAC.2019.2913768}
}

MPC-RL is currently implemented as:

• The MPC problem is discretized by the direct multiple shooting method.
• The KKT system to compute the Newton step is solved by Riccati recursion, in which a batch implementation is considered (this is inspired by mpc.pytorch).
• Q-learning for MPC function approximator
• Only the quadratic cost function is allowed.
• No economic cost can be considered (because we have not implemented the cost rotation yet).
• No constraints are considered.

Pendulum example (left: before Q-learning, right: after Q-learning, see examples/pendulum_inaccurate.py and examples/pendulum_q_learning.py)