Work
Workflows¶
A reservoir computing project iterates through the same cycle: fit the readout (an algebraic solve that takes seconds), forecast against held-out validation data, adjust the hyperparameters that affect forecast quality, then scale the result with larger reservoirs, ensembles, GPU execution, and model persistence. These four pages cover that cycle in order.
-
Three training approaches — the one-pass algebraic solve, frozen features with a gradient head, and full BPTT — and when each is appropriate.
-
The two-phase forecast procedure, driver alignment during autoregression, coupled ensembles, and Lyapunov-time limits on forecast horizons.
-
Report forecast horizons in Lyapunov times, measure valid prediction time, and read a reservoir's own edge-of-chaos exponent.
-
The eight hyperparameters that drive forecast quality, intuition for spectral radius, leak rate, and ridge regularization, and complete hyperparameter studies with
run_hpo. -
When to move to the GPU, saving and loading models, and embedding frozen reservoirs in larger PyTorch pipelines.