Hyperparameter search schedules

Hyperband successive-halving schedules and Latin-hypercube search plans for hyperparameter optimization.

automl generates the budget-allocation math behind modern hyperparameter search — successive halving (Jamieson & Talwalkar 2016), Hyperband (Li et al. 2018), and Latin-hypercube random search — replacing sklearn.model_selection.HalvingRandomSearchCV's halving arithmetic and scipy.stats.qmc.LatinHypercube-style sampling with zero Python imports. It is callback-free by design: every function returns a schedule (plain data) that the caller executes however it likes — local train loops, remote workers, anything.

The modules

Module	What it does
automl	budget-allocation schedules for hyperparameter search — successive halving (Jamieson & Talwalkar 2016), Hyperband (Li et al. 2018), and Latin-hypercube random search — replacing sklearn.model_selection.HalvingRandomSearchCV's halving arithmetic and scipy.stats.qmc.LatinHypercube-style sampling

Functions

Function	What it does
successive_halving_schedule(n_configs, min_budget, eta)	SHA rounds of (n_survivors, budget): round i keeps floor(n_configs / eta^i) configs at budget min_budget * eta^i, until one survivor remains. Invalid input (n_configs < 1, min_budget < 1, eta < 2) → empty list.
hyperband_brackets(max_budget, eta)	Full Hyperband bracket list, ordered most-exploratory-first (s = s_max .. 0); each Bracket carries its (s, n_configs, min_budget) plus an embedded SHA schedule of s + 1 rounds.
promote_top_k(scores, k)	Indices of the k highest scores, ordered best-first; ties broken deterministically toward the lower index; k clamps to len(scores).
random_search_configs(n, dims, seed)	Latin-hypercube config matrix, n rows × dims columns, every value in [0, 1); per dimension the n samples occupy the n strata exactly once.

Determinism and integer-exact schedules

Budgets and config counts use exact integer math — never float log/pow — so schedules are bit-reproducible across platforms. The LHS sampler runs on a seeded xorshift64 generator: identical output for the same (n, dims, seed) (tests pin seed 42 reproducible, seed 43 different), and per dimension floor(v * n) hits each of the n strata exactly once.

Verification

The known-value tests pin the canonical schedules exactly, not approximately:

• SHA n_configs=27, eta=3 → (27,1), (9,3), (3,9), (1,27); min_budget=2 scales budgets to 2, 6, 18, 54; non-power-of-eta n=10 floors to (10,1), (3,3), (1,9).

• Hyperband max_budget=81, eta=3 → exactly 5 brackets with (n_s, r_s) = (81,1), (34,3), (15,9), (8,27), (5,81) — the canonical Hyperband example table. Every bracket's final round runs at budget 81 with ≥ 1 survivor, and the s=0 bracket degenerates to plain random search: a single (5, 81) round.

• promote_top_k([0.5, 0.9, 0.5, 0.7], 3) → [1, 3, 0]: the 0.5 tie resolves to the lower index, so promotion decisions never depend on evaluation order.

Next: promote what you trained with statistical model comparison, or browse the engine catalog.