pso — particle swarm optimization (sampler)

Role: optimization · kind: n4m_sampler_kind_t = N4M_SAMPLER_PSO · since: ABI 2.1 (F3)

Particle Swarm Optimization over the unit hypercube (Kennedy & Eberhart). A swarm of swarm_size (= 16 in F3) particles — each with a position u ∈ [0,1)^P, a velocity, and a remembered personal best — is asked out per iteration. Once the iteration’s trials are terminal, completed scores update personal and global bests, then velocities and positions advance:

v ← w·v + c1·r1·(pbest − x) + c2·r2·(gbest − x)      x ← clamp01(x + v)

with the Clerc & Kennedy (2002) convergence constants w = 0.729, c1 = c2 = 1.494. Candidates are decoded with the shared Optimizer::decode_candidate, so mixed continuous / discrete / categorical spaces work. Conditional activation is honored. Particle position and velocity contain one coordinate per declared axis, but the coordinate for a sorted_tuple is unused by decoding; tuple components are drawn independently by the base RNG.

Synchronous update (F3): the swarm advances only once its whole iteration is terminal (liar = none), so ask_batch returns a partial batch at an iteration boundary. Only completed scores can update personal/global bests; pruned and failed particles receive worst fitness for that iteration. Any non-none liar value is rejected at optimizer creation with N4M_ERR_NOT_IMPLEMENTED. Velocities are capped at vmax = 0.5 of the unit range. Warm-start (n4m_optimizer_enqueue) is not supported for population samplers (returns N4M_ERR_UNSUPPORTED).

Search spaces containing hard mutex_group, requires or exclude constraints are rejected by n4m_optimizer_create with N4M_ERR_UNSUPPORTED; PSO does not turn infeasibility into a host fitness penalty. Hard constraints that reference a sorted_tuple root are rejected for every sampler.

PSO is intended for smooth-ish continuous and mixed surfaces; it complements the more disruptive ga. It uses the same terminal-state generation guard as GA, keyed on trial-id ranges.

HPO-sampler PSO over the typed space — distinct from the feature-selection n4m_feature_selection_pso_select (binary PSO over feature masks).

Usage (C ABI)

n4m_optimizer_options_t opts;
n4m_optimizer_options_init(&opts);
opts.sampler = N4M_SAMPLER_PSO;
opts.seed = 42;

Parity

  • Tier B-state target: the swarm trajectory is deterministic for a fixed seed, ask/tell order and score tape. Track-Q commits the native pso_sphere2 trace as a target for future bindings; it does not yet exercise PSO with every pruner or search-space feature. Convergence on a continuous objective is verified in the C++ tests.

References

  • Kennedy & Eberhart, A discrete binary version of the particle swarm algorithm, IEEE SMC (1997); Clerc & Kennedy (2002) convergence constants. See _finetuning_bibliography.bib.