Migrating to ReservoirComputing.jl v1

Changes from the last 0.12 releases to v1. Deprecated names still emit warnings on 0.12.x and are removed at the v1.0 cut.

Initializer sign patterns

Initializer sign handling now uses typed sign-pattern objects instead of resolving a function from the sampling_type symbol and forwarding sampler-specific keywords.

The available patterns are:

• RandomSigns(positive_probability): independently preserves each existing sign with the given probability and flips it otherwise. The default probability is 0.5.
• RegularSigns(strides): flips signs at positions selected by an integer stride or a repeating tuple of strides. Vectors of strides are also accepted.
• IrrationalDigitSigns(irrational; start): flips signs where the corresponding decimal digit of an irrational number is odd.
• nothing: leaves every sign unchanged. This is the default.

Pass the pattern through the signs keyword:

minimal_init(100, 3; signs = RandomSigns(0.5))
minimal_init(100, 3; signs = RegularSigns(2))
minimal_init(100, 3; signs = RegularSigns((2, 3)))
minimal_init(100, 3; signs = IrrationalDigitSigns(pi; start = 1))
minimal_init(100, 3) # signs = nothing

Migrating sampling_type

sampling_type and its sampler-specific forwarded keywords are removed in v1.0.

Do not pass signs together with sampling_type. Sampler-specific keywords such as positive_prob, strides, and irrational now belong to the corresponding pattern constructor instead of the initializer.

This migration applies to the component-building functions and to initializers that forward their keyword arguments to those components, including:

• delay_line! and delay_line;
• backward_connection! and reservoir initializers containing backward connections;
• simple_cycle!, reverse_simple_cycle!, and cycle-based reservoir initializers;
• add_jumps! and cycle_jumps;
• self_loop! and self-loop reservoir initializers;
• minimal_init and weighted_minimal.

Nested keyword bundles accept sign patterns directly:

cycle_jumps(100, 100; jump_kwargs = (; signs = RegularSigns((2, 3))))

minimal_init default

minimal_init previously applied Bernoulli sign flipping by default. Its new default is signs = nothing, so generated weights retain their original signs. Pass signs = RandomSigns() to preserve the previous default explicitly:

minimal_init(100, 3; signs = RandomSigns())

Initializer behavior corrections

informed_init now extracts scalar random values correctly when assigning informed input connections. Earlier versions could fail on zero-dimensional array arithmetic. Call sites do not need changes.

Input-extended reservoir states

Extend can be passed in state_modifiers to prepend the current model input to the wrapped modifier's output:

model = ESN(3, 100, 3; state_modifiers = (Extend(Collect()),))

High-level constructors size their linear readout automatically for Extend when its wrapped operation preserves the feature width. For a custom modifier that changes the feature width, pass the resulting width explicitly through readout_in_dims.

Training API

ps, st = train(model, train_data, target_data, ps, st;
    objective = RidgeRegression(1e-3))

Generic train! → train

The model-level bang API is removed in v1.0. Map the positional training method to objective:

washout, return_states, and solver keep the same names. Conceptor train!(rng, concept::Conceptor, ...) is unchanged.

Objective-level train

train(objective, states, targets; solver=...) is removed in v1.0. Fit through the model instead (returns (ps, st), not a bare weight matrix):

# before
W = train(RidgeRegression(1e-3), states, targets)

# after
ps, st = train(model, train_data, target_data, ps, st;
    objective = RidgeRegression(1e-3))

Renames

Conceptor tolerance

conceptor_and accepts optional tolerance for its relative rank cutoff. Default nothing preserves the previous behavior:

conceptor_and(C1, C2; tolerance = 1e-10)