Data-driven cold starting of good reservoirs

INSTALLATION

Clone repository and install dependencies

git clone https://github.com/Learning-of-Dynamic-Processes/coldstart.git
cd lpde
pip install -r requirements.txt

In addition, a running version of pycuda is necessary for the data generation.

USAGE

This python package contains functions to initialize reservoir computers for two examples, the Brusselator and the Lorenz system.

• Run brusselator_optuna.py to find a good set of reservoir hyperparamerters for the Brusselator system. Paste the best set of parameters into brusselator/config.py.
• Run brusselator_train.py to create a reservoir using the parameters in brusselator/config.py.
• Use brusselator.ipynb to create the cold start map and the Brusselator figures shown in the paper.

Likewise, for the Lorenz system:

• Run lorenz_optuna.py to find a good set of reservoir hyperparamerters for the Lorenz system. Paste the best set of parameters into lorenz/config.py.
• Run lorenz_train.py to create a reservoir using the parameters in lorenz/config.py.
• Use lorenz.ipynb to create the cold start map and the Lorenz figures shown in the paper.

To recreate results from section 4.1 of the article, use the code in coldstart_fig_9/.

LICENCE

This work is licenced under MIT License. Please cite

"Data-driven cold starting of good reservoirs" L. Grigoryeva et al., (2023). ArXiv arXiv:2403.10325 (https://arxiv.org/abs/2403.10325)

if you use this code for publications.

In addition, if you use the diffusion maps or geometric harmonics functions, cite

" datafold: data-driven models for point clouds and time series on manifolds" Lehmberg et al., (2020). Journal of Open Source Software, 5(51), 2283, (https://doi.org/10.21105/joss.02283)