harwiltz / stable-dynamics-models

PyTorch implementation of "Learning Stable Deep Dynamics Models" (https://papers.nips.cc/paper/9292-learning-stable-deep-dynamics-models), with extensions to controlled dynamical systems.

Stable Dynamics Models

This repo contains an implementation of the method introduced in "Learning Stable Deep Dynamics Models" with PyTorch. This includes a PyTorch implementation of an Input Convex Neural Network.

The goal with this repo is to extend the results to account for controlled systems.

To get started with this library, install like so:

git clone https://github.com/harwiltz/stable-dynamics-models.git
cd stable-dynamics-models
pip install -r --user requirements.txt
python setup.py install --user

You can then instantiate an instance of a stable dynamics model with

from stablemodels import StableDynamicsModel
model = StableDynamicsModel((4,),                 # input shape
                            control_size=2,       # action size
                            alpha=0.9,            # lyapunov constant
                            layer_sizes=[64, 64], # NN layer sizes for lyapunov
                            lr=3e-4,              # learning rate for dynamics model
                            lyapunov_lr=3e-4,     # learning rate for lyapunov function
                            lyapunov_eps=1e-3)    # penalty for equilibrium away from 0

prediction = model(state, action)