InitializationStrategy

InitializationStrategy(name: str | None = None)

Bases: StrategyBase, ABC

Base class for strategies that produce an initial selection of k vectors.

Use the factory methods (fast, random_one_shot, random_batched, eager) to create instances.

get_next_samples `abstractmethod`

get_next_samples(
    state: SolverState, k_remaining: int | int32
) -> NDArray[int32]

Return next batch of samples to be added to the initial selection.

This method is called repeatedly by the Solver, until enough samples have been selected to reach the desired selection size.

Parameters:

Name	Type	Description	Default
`state`	`SolverState`	(SolverState) The current solver state, to fetch problem size, constraints, distances, etc..., so initial selection can be made in an informed way.	required
`k_remaining`	`int \| int32`	(int) number of samples that remain to be selected.	required

Returns:

Type	Description
`NDArray[int32]`	np.array of unique np.int32 values, shape=(b,), with indices of samples to be added to the selection. b can be any value in range [1, k_remaining]. Samples should be unique and not yet selected.

fast `classmethod`

fast() -> Self

Deterministic greedy initialization. Selects vectors one-by-one, always picking the one that maximizes separation from the current selection. Fast but seed-independent.

random_one_shot `classmethod`

random_one_shot(
    uniform: bool = False, ignore_constraints: bool = False
) -> Self

Random initialization that selects all k vectors in a single batch, with probabilities biased by global separation (unless uniform=True).

Parameters:

Name	Type	Description	Default
`uniform`	`bool`	If True, sample uniformly instead of using separation-based probabilities.	`False`
`ignore_constraints`	`bool`	If True, ignore constraints during sampling.	`False`

random_batched `classmethod`

random_batched(b: int, ignore_constraints: bool = False) -> Self

Random initialization that selects vectors in batches of b, re-evaluating separations between batches.

Parameters:

Name	Type	Description	Default
`b`	`int`	Batch size (number of vectors to select per batch).	required
`ignore_constraints`	`bool`	If True, ignore constraints during sampling.	`False`

eager `classmethod`

eager(nc: int, ignore_constraints: bool = False) -> Self

Greedy initialization that evaluates nc random candidates per step and picks the best one. Higher nc gives better quality but is slower.

Parameters:

Name	Type	Description	Default
`nc`	`int`	Number of candidates to evaluate at each step.	required
`ignore_constraints`	`bool`	If True, ignore constraints during sampling.	`False`

InitializationStrategy