Synthesis of Stabilizing Recurrent Equilibrium Network Controllers

This is an implementation of the stabilizing REN controller presented at CDC 2022:

@INPROCEEDINGS{9992684,
  author={Junnarkar, Neelay and Yin, He and Gu, Fangda and Arcak, Murat and Seiler, Peter},
  booktitle={2022 IEEE 61st Conference on Decision and Control (CDC)}, 
  title={Synthesis of Stabilizing Recurrent Equilibrium Network Controllers}, 
  year={2022},
  volume={},
  number={},
  pages={7449-7454},
  doi={10.1109/CDC51059.2022.9992684}}

File Structure

• envs: models of plants.
• models: controller models and an implicit model for system identification. The controller presented in the paper is in ProjREN.py.
• learned_models: parameters trained with an implicit model for system identification.
• activations.py: activation functions with sector-bound information.
• trainers.py: trainers modified to include the projection step.

Runnable files

• train_controller.py: configure and train controllers.
• train_implicit_network.py: train an implicit model to learn plant dynamics.
• plots.py and rollout.py: plotting files.

Package Requirements

This code is tested with Python 3.9 and PyTorch 1.10.

Credits

• The fixed point solvers in the deq_lib folder and the basis for the PyTorch implicit model implementation are from the Deep Equilibrium Models repository.