echostatenetwork

Echo state networks / reservoir computing in pure numpy. One class, EchoStateNetwork: ridge-regression training with chaotic recycle validation (contiguous runs or ragged dwell segments), parametric inputs, optional Bayesian hyperparameter search (scikit-optimize), closed-loop prediction, and Jacobians for data assimilation.

This is the single shared reservoir core behind qlrom (quantized-local ESN families) and romda (real-time ESN forecasting and bias-aware data assimilation).

Tutorials

• The EchoStateNetwork class — train an ESN on the Lorenz 63 system, with full and partial observability.
• Parametric ESN — one reservoir conditioned on a physical parameter, forecasting at unseen parameter values.
• Leaky-integrator ESN — the leak-rate hyperparameter and when it helps.
• Validation strategies — SSV/WFV/KFV/recycle validation compared live on Lorenz 63, including seeded ensembles with train(n_seeds=...); see also Validation strategies.

Acknowledgements

This repository is based on alberacca/Echo-State-Networks, the reference implementation of the recycle-validation ESN:

Racca, A., & Magri, L. (2021). Robust optimization and validation of echo state networks for learning chaotic dynamics. Neural Networks, 142, 252-268. doi:10.1016/j.neunet.2021.05.004