Drivers

Reference drivers.

Two deliberately simple strategies that already span the design space CADAQUES cares about:

• RandomDriver — the universal, budget-oblivious baseline against which every intelligent driver must justify its cost.

• AnnealedLocalDriver — a budget-aware strategy: it explores globally while funds are plentiful and contracts around the incumbent as the budget runs out, reading the remaining fraction from the BudgetView it receives each iteration.

Bayesian optimization, gradient-based and LLM-agent drivers plug in through the same two-method protocol.

class cadaques.drivers.reference.RandomDriver(space, fidelity=<factory>, seed=None)

Uniform random search over a box-bounded space.

Parameters:

• space (Mapping[str, tuple[float, float]])

• fidelity (Mapping[str, int])

• seed (int | None)

class cadaques.drivers.reference.AnnealedLocalDriver(space, fidelity=<factory>, sigma_max=0.5, sigma_min=0.02, maximize=True, seed=None)

Gaussian local search whose step size anneals with the budget.

The proposal is a perturbation of the best point seen so far, with standard deviation sigma_max at the start of the campaign shrinking linearly to sigma_min as budget.fraction_used approaches 1 — spend on exploration while rich, on exploitation while poor.

Parameters:

• space (Mapping[str, tuple[float, float]])

• fidelity (Mapping[str, int])

• sigma_max (float)

• sigma_min (float)

• maximize (bool)

• seed (int | None)