Classical (values only)¶
Boltzmann(t_init, sigma)Logarithmic cooling + Gaussian + Metropolis.
Fast(t_init, gamma)Reciprocal cooling + Cauchy + Metropolis.
Gsa(t_init, =q_v, q_a)=Tsalliscooling +Tsallisvisit +Tsallisaccept.
All called via run(obj, low, high, preset, n_epochs, steps_per_epoch, seed).
Pilot and low-discrepancy¶
pilot_draws_qmc(n, seed)-> (n,3) array of (T0, sigma,q_v).low_discrepancy_points(low, high, n, skip=1).
Polish¶
These drivers require gradients.
Gradient support is provided by eindir; see the eindir gradients guide.
=polish(obj, grad, low, high, x0, max_fevals=200, …)
qmc_polish(obj, grad, low, high, n_starts, max_fevals_per_start, ...)=shifted_qmc_polish(…, n_replicates=1, …)
Advanced drivers¶
additive_independence(obj, low, high, max_fevals, seed=0, degree=8, ... n_pilot=0)(values only, rank-1 surrogate).gle_langevin(obj, grad, low, high, max_fevals, seed=0, omega0=0.2, dt=0.2, n_epochs=40)(gradient, colored noise).
Device and ensemble (same kernel)¶
run_device(single chain, NumPy or CuPy bounds)run_ensemble(batched,n_chains, same kernel)
Hamiltonian Monte Carlo (HMC) / quasi-Monte Carlo (QMC) variants (also exposed)¶
run_hmc,run_qmc(seepython/anneal/__init__.pyand tests for current signatures).
All return dicts with best_pos, best_val, n_evals (or the rich History / DeviceHistory / EnsembleHistory objects from run / run_device).