GIVPConfig

givp.GIVPConfig dataclass

Hyper-parameters for the GRASP-ILS-VND-PR algorithm.

The optimization sense is controlled by minimize (boolean, preferred) or direction (string, SciPy/Optuna-style). Pass one of the two; if both are given, minimize wins and direction is rewritten to match. Internally the library always minimizes: when the user wants maximization the objective is wrapped with a sign flip and the returned value is restored to the original sign.

Attributes:

Name	Type	Description
max_iterations	int	Maximum number of GRASP outer iterations.
alpha	float	Initial randomization parameter for greedy construction (RCL).
vnd_iterations	int	Maximum iterations of the VND local search.
ils_iterations	int	Maximum iterations of the ILS loop.
perturbation_strength	int	Magnitude of the ILS perturbation.
use_elite_pool	bool	Whether to maintain an elite pool for path relinking.
elite_size	int	Maximum size of the elite pool.
path_relink_frequency	int	GRASP iteration period at which to run PR.
path_relink_strategy	PathRelinkStrategy	Explicit path-relinking direction control.
adaptive_alpha	bool	If True, alpha varies between alpha_min and alpha_max.
alpha_min	float	Lower bound used by adaptive alpha.
alpha_max	float	Upper bound used by adaptive alpha.
num_candidates_per_step	int	Candidates evaluated per construction step.
use_cache	bool	If True, evaluations are memoized via an LRU cache.
cache_size	int	Maximum entries kept by the LRU cache.
early_stop_threshold	int	Iterations without improvement to early-stop.
use_convergence_monitor	bool	Enable diversification/restart heuristics.
n_workers	int	Number of worker processes (or threads) used to evaluate candidates in parallel during the GRASP construction phase. Default is 1 (serial). Parallelism strategy (applied when n_workers > 1 and the evaluation cache is disabled): ProcessPoolExecutor — used when the objective is picklable (module-level functions, classes defined at import time). Runs in separate processes, completely bypassing the GIL and providing true multi-core speedup for any objective, including pure-Python ones. cloudpickle ProcessPoolExecutor — if the objective is a closure, lambda, or locally-defined function (not picklable with pickle), a second attempt uses cloudpickle to serialise it. Install the optional extra to enable this path:: pip install "givp[parallel]" ThreadPoolExecutor fallback — used only when both serialisation strategies fail (i.e. cloudpickle is not installed). Thread-based execution benefits objectives that release the GIL (NumPy-heavy, Cython, Numba-compiled code) but provides no speedup for pure-Python objectives due to the GIL.
time_limit	float	Wall-clock budget in seconds (0 = unlimited).
minimize	bool | None	Boolean convenience flag. True (default) means minimization, False means maximization. When set, it overrides direction.
direction	Direction	'minimize' (default) or 'maximize'. Kept for SciPy/Optuna-style API compatibility.
integer_split	int | None	Index where integer variables begin in the decision vector. None preserves the legacy SOG2 behavior of splitting in half. Set to num_vars for fully continuous problems or to 0 for fully integer problems.

as_core_config

as_core_config() -> GIVPConfig

Return self.

GIVPConfig structurally satisfies the internal _CoreConfigProto used by givp.core functions, so it can be passed directly without copying field values into a separate dataclass. The minimize and direction fields are ignored by core code.