random — uniform random search

Role: optimization · kind: n4m_sampler_kind_t = N4M_SAMPLER_RANDOM · since: ABI 2.1 (F0)

Uniform random sampling over the typed search space. Each parameter is drawn independently from its declared distribution: uniform for int/float, log-uniform for log_int/log_float, uniform-over-choices for categorical/ordinal, and (for sorted_tuple) a sorted vector of uniform draws. Conditional parameters are activated per their condition_in/_not_in constraint; mutex/requires/exclude constraints are enforced by bounded rejection sampling.

Random search is the default non-adaptive baseline and the startup-seeding strategy for the adaptive samplers. It is competitive with grid search at a fraction of the cost when only a few hyperparameters matter, per Bergstra & Bengio (2012).

Usage (C ABI)

n4m_search_space_t* space = NULL;
n4m_search_space_create(&space);
n4m_search_space_add_int(space, "n_components", 1, 30, 1, /*log=*/0);
n4m_search_space_add_float(space, "alpha", 1e-4, 1e0, 0.0, /*log=*/1);

n4m_optimizer_options_t opts;
n4m_optimizer_options_init(&opts);          /* sampler=random, pruner=none */
opts.seed = 42;

n4m_optimizer_t* opt = NULL;
n4m_optimizer_create(ctx, space, &opts, &opt);
for (int i = 0; i < n_trials; ++i) {
    n4m_trial_t* t = NULL;
    n4m_optimizer_ask(opt, &t);
    int64_t nc; n4m_trial_get_int(t, "n_components", &nc);
    /* ... host evaluates, obtains `score` ... */
    int64_t id; n4m_trial_get_id(t, &id);
    n4m_optimizer_tell(opt, id, score);
}
n4m_trial_t* best; double best_score;
n4m_optimizer_best(opt, &best, &best_score);

Parity

• Tier C (self-consistency + regret): a random study replays bit-identically for a fixed seed, and reaches a known optimum within the expected budget on benchmark spaces.

• Cross-binding: identical seed → identical trial sequence across all bindings (1e-12), via the shared n4m_rng (splitmix64).

References

• Bergstra & Bengio, Random Search for Hyper-Parameter Optimization, JMLR 13 (2012), 281–305. bergstra2012random