sequince.steps.mutation.adaptive.de

Adapt a DE scale mixture only between outer SMC iterations.

Classes

AdaptDEScaleWeights

Replace immutable DE weights at the next outer iteration (eager only).

Module Contents

class sequince.steps.mutation.adaptive.de.AdaptDEScaleWeights(mutation, policy=None)

Bases: sequince.steps.mutation.adaptive.base.AdaptiveMutationStep

Replace immutable DE weights at the next outer iteration (eager only).

Parameters:

• mutation (RedBlueMetropolisHastingsMutation or RepeatedMutationStep) – DE mutation or a complete repeated block whose leaf uses a DifferentialEvolutionKernel. Wrap the complete block so weights stay fixed across all rounds. Place conditional execution outside.

• policy (DEScaleWeightPolicy or None) – Policy selecting weights from completed metrics. Defaults to ESJDScaleWeightPolicy. The multiplier bank and unit-move probability remain unchanged.

Notes

Compose as AdaptMutationSteps(AdaptDEScaleWeights(mutation), ...) to adapt both budgets and weights. Repeated calls at the same context iteration reuse frozen weights and consult the policy only once. The next weights are installed on entering the next iteration, never midway through a block. The wrapped kernel configuration is owned by this wrapper; do not share the mutation with another adapter or replace its kernel.

Required DEScaleMetrics are configured once on the leaf, preserving user metrics. An existing de_scales metric must match the configured scale bank.

Completed metrics gain a de_scale_adaptation entry containing weights used for this block and next_weights. Checkpoints retain both decisions and nested child/policy state. Reconstruct the same bank, kernel parameters and policy configuration before loading a checkpoint.

policy

property current_weights: tuple[float, ...] | None

Weights frozen for the current or most recent outer iteration.

Return type:

tuple[float, …] | None

property next_weights: tuple[float, ...] | None

Weights proposed for the next outer iteration.

Return type:

tuple[float, …] | None

apply_with_n_steps(state, context, rng, *, n_steps)

Run a complete block with a budget supplied by an outer adapter.

Parameters:

• state (sequince.state.ParticleState)

• context (sequince.steps.context.StepContext)

• rng (orng.RandomGenerator)

• n_steps (int)

Return type:

tuple[sequince.state.ParticleState, sequince.steps.context.StepContext, Any]

state_dict()

Return frozen/pending decisions and nested child/policy state.

Return type:

dict[str, Any]

load_state_dict(state)

Restore decisions without rerunning the policy.

Parameters:

state (collections.abc.Mapping[str, Any])

Return type:

None