iwopy.interfaces.scipy¶

Interface to the scipy package

Optimizer_scipy solves continuous single-objective problems with any algorithm exposed by scipy.optimize.minimize(). For gradient-capable methods, it supplies the complete objective and constraint Jacobian through Problem.get_jacobian. Objective and constraint callbacks at the same point share one cached CSR Jacobian calculation.

trust-constr receives CSR constraint Jacobians without a dense intermediate. Its native sparse_jacobian option defaults to True; explicit values are preserved. Set it to False when selecting a dense-only QRFactorization or SVDFactorization. Other methods, including SLSQP, receive dense constraint arrays as required by their native interface. Scalar objective gradients remain dense vectors for every SciPy method. Sparse callbacks do not remove dense Hessian approximations or prevent every native factorization fallback to dense storage.

By default, Jacobians are requested with pop=True. A wrapper such as LocalFD can therefore evaluate all finite-difference points together through the problem’s population interface. Set vectorize=False to use serial iwopy Jacobian evaluation. The interface does not fall back to SciPy’s scalar finite differences, so every required derivative must be analytical or provided by a problem wrapper.

The derivative path applies to CG, BFGS, Newton-CG, L-BFGS-B, TNC, SLSQP, trust-constr, dogleg, trust-ncg, trust-exact, and trust-krylov, as well as SciPy’s automatically chosen default and custom callable methods. Methods that require a Hessian or Hessian-vector product still need explicit hess or hessp parameters.

Select a method with optimizer on the class, or with create_optimizer(family="scipy", optimizer=..., problem=...). Supply tol, memory_size, vectorize, and supported max_iterations / max_evaluations explicitly; flat native method keywords such as ftol become SciPy’s method options. See Termination for method units and unsupported limits. Custom callables own their stopping settings.

Nelder-Mead, Powell, COBYLA, and COBYQA are derivative-free; SciPy offers no population-evaluation hook for their trial points, so no Jacobian is supplied. Constraint components are grouped by equality, lower, and upper bounds for every constrained method.

The parameters bounds, callback, constraints, and jac are managed by the interface. Integer variables are rejected because scipy.optimize.minimize operates on continuous variables.

Optimizer_scipy_de provides genuine population-vectorized differential evolution for bounded single-objective floating, integer, and mixed problems. It uses deferred updating, one worker, and no scalar polishing. See SciPy Differential Evolution and its explicit setup in the catalog. Its explicit controls include max_iterations, population_multiplier, seed, tol, atol, and vectorize; remaining native options are flat keywords. population_multiplier is per free dimension, not an absolute population size. DE has no portable native evaluation cap, so max_evaluations must be None.