Particle Swarm Optimization: Swarm Intelligence Explained

A flock's memory, turned into an algorithm

Kennedy and Eberhart's 1995 insight was deceptively simple: good solutions attract neighbours. Just as a bird adjusts its flight to keep up with its flock while remembering where it personally found the best food, each particle in PSO carries a social memory — the swarm's best-known position, gbest — and a cognitive memory, its own best-known position, pbest. PSO belongs to the broader family of swarm intelligence algorithms (alongside Ant Colony Optimisation and Artificial Bee Colony), all sharing the property of emergence: near-optimal global behaviour from simple local rules, with no centralised control.

The velocity update equation

A swarm of N particles is initialised at random positions in the search space. At each iteration, every particle updates its velocity from three forces — inertia (keep moving in the current direction), a cognitive pull toward its own best-known position, and a social pull toward the swarm's global best — then updates its position by simply adding the new velocity. The random scalars r1 and r2 are resampled per particle, per dimension, per iteration, which is essential: without that randomness, particles in a symmetric configuration could lock into identical movement and the swarm would collapse to a deterministic pattern.

v(t+1) = ω·v(t) + c1·r1·(pbest − x(t)) + c2·r2·(gbest − x(t))
x(t+1) = x(t) + v(t+1)

Typical: ω 0.4→0.9 (linearly decayed), c1 = c2 = 1.5–2.0
Clerc's constriction values: ω = 0.729, c1 + c2 ≤ 4

Convergence, stagnation, and the ring topology fix

Clerc and Kennedy (2002) formalised PSO's convergence conditions: without parameter control the swarm can oscillate and never settle, but a region of (ω, c1, c2) space — the convergence triangle — guarantees a stable trajectory. The most common failure mode is premature convergence, where the whole swarm collapses onto a local optimum before exploring the space, typically because the social coefficient c2 dominates or diversity has already been lost. Mitigations include random restarts for stagnated particles or switching from a single shared gbest to a local (lbest) ring topology, where each particle only sees its nearest neighbours — slower to converge, but far more resistant to getting trapped in one basin.

PSO versus genetic algorithms and simulated annealing

PSO, genetic algorithms and simulated annealing are all gradient-free, but they differ in mechanism: PSO moves a population continuously via velocity and persistent memory (pbest, gbest); a genetic algorithm applies crossover and mutation with selection pressure and no per-individual memory; simulated annealing perturbs a single solution under a cooling temperature schedule. In practice PSO is often preferred for continuous, low-to-medium dimensional problems (≤100 dimensions) because it's simple to implement and converges in a few dozen iterations — from engineering design and PID tuning to wind-farm layout optimisation and hyperparameter search.

Frequently asked questions

Try it live

Everything above runs in your browser — open Particle Swarm Optimization and watch a swarm of particles converge on the global minimum of several test functions, tuning inertia, cognitive and social weights live. Nothing is installed, nothing is uploaded.