Skip to content

Search Strategies

Particle Swarm Optimization

Each particle represents an encoded parameter position. Every iteration fits each particle, updates its personal best, then moves it using inertia (w), a cognitive coefficient (c1), and a social coefficient (c2).

Global topology follows the best particle in the full swarm. Ring-local topology follows the best of each particle and its immediate neighbours. A run plans particles x iterations candidates unless max_trials, timeout, cancellation, or early stopping ends it sooner.

Random search samples independent candidates from the same typed domains. A saved seed makes the sequence reproducible. It is the recommended baseline for judging whether PSO adds value for a dataset and budget.

Grid search evaluates every discrete grid value deterministically. Choice and IntRange produce exact values. Float domains produce values according to their precision or bounded log grid. Check SearchSpace.grid_size before launching a large grid.

runtime.trial_workers bounds concurrent candidates. Each candidate evaluates its folds sequentially. PSO waits for an iteration barrier before updating positions; random and grid search use bounded batches. Estimator n_jobs is one under candidate parallelism. Planned fits are candidates multiplied by fold count, plus one final refit when enabled. Multiple recorded runs can separately execute concurrently through PSPSO_MAX_WORKERS; one active run is the default.