Train

train!(rc, train_data, target_data, ps, st,
       train_method=StandardRidge(0.0);
       washout=0, return_states=false)

Trains a given reservoir computing by creating the reservoir states from train_data, and then fiting the readout layer using target_data as target. The learned weights/layer are written into ps.

Arguments

• rc: A reservoir computing model, either provided by ReservoirComputing.jl or built with ReservoirChain. Must contain a trainable layer (for example LinearReadout), and a collection point Collect.
• train_data: input sequence where columns are time steps.
• target_data: targets aligned with train_data.
• ps: model parameters.
• st: model states.
• train_method: training algorithm. Default is StandardRidge.

Keyword arguments

• washout: number of initial time steps to discard (applied equally to features and targets). Default 0.
• return_states: if true, also returns the feature matrix used for the fit.
• kwargs...: additional keyword arguments for the training algorithm, if needed. Defaults vary according to the different training method.

Returns

• (ps, st): updated model parameters and states.
• (ps, st), states: If return_states=true.

Notes

• Features are produced by collectstates(rc, train_data, ps, st). If you rely on the implicit collection of a LinearReadout, make sure that readout was created with include_collect=true, or insert an explicit Collect() earlier in the ReservoirChain.

train(train_method, states, target_data; kwargs...)

Lower level training hook to fit a readout from precomputed reservoir features and given targets.

Dispatching on this method with different training methods allows one to hook directly into train! without additional changes.

Arguments

• train_method: An object describing the training algorithm and its hyperparameters (e.g. regularization strength, solver choice, constraints).
• states: Feature matrix with reservoir states (ie. obtained with collectstates). Shape (n_features, T), where T is the number of samples (e.g. time steps).
• target_data: Target matrix aligned with states. Shape (n_outputs, T).

Returns

• output_weights: Trained readout. Should be a forward method to be hooked into a layer. For instance, in case of linear regression output_weights is a mtrix consumable by LinearReadout.

Notes

• Any sequence pre-processing (e.g. washout) should be handled by the caller before invoking train. See train! for an end-to-end workflow.
• For very long T, consider chunked or iterative solvers to reduce memory usage.
• If your approach returns additional artifacts (e.g. diagnostics), prefer storing them inside train_method or exposing a separate API; keep train’s return value as the forward method only.

StandardRidge([Type], [reg])

Ridge regression method.

Equations

\[\mathbf{w} = (\mathbf{X}^\top \mathbf{X} + \lambda \mathbf{I})^{-1} \mathbf{X}^\top \mathbf{y}\]

Arguments

• Type: type of the regularization argument. Default is inferred internally, there's usually no need to tweak this
• reg: regularization coefficient. Default is set to 0.0 (linear regression).