Swarm intelligence & optimization

Most multi-agent design has agents pursuing goals. Swarm intelligence comes at it from a different direction: complex, useful global behavior that emerges from many simple agents following local rules — with no central controller. A flock of birds, an ant colony, a school of fish: none has a leader, yet the collective solves problems no individual could. Turned into algorithms, this gives a family of gradient-free optimizers.

Particle swarm optimization

In PSO, a swarm of particles drifts through the search space looking for the best solution. Each particle’s move is pulled by just two things: its own best position so far (cognitive) and the swarm’s best so far (social). Those simple local rules, plus sharing one global best, make the swarm converge on the optimum:

import numpy as np
rng = np.random.default_rng(0)

def f(x): return (x - 3.0)**2                          # minimize; optimum at x=3
pos = rng.uniform(-10, 10, 5); vel = np.zeros(5)        # 5 particles
pbest = pos.copy(); gbest = pbest[int(np.argmin([f(x) for x in pbest]))]

for step in range(1, 7):
    for i in range(5):
        vel[i] = 0.5*vel[i] + 0.8*(pbest[i]-pos[i]) + 0.8*(gbest-pos[i])   # inertia + toward pbest + gbest
        pos[i] += vel[i]
        if f(pos[i]) < f(pbest[i]): pbest[i] = pos[i]
    gbest = pbest[int(np.argmin([f(x) for x in pbest]))]
    print(f"step {step}: global best x={gbest:.2f}  f={f(gbest):.4f}")

step 1: global best x=2.74  f=0.0680
step 2: global best x=2.74  f=0.0680
step 3: global best x=3.08  f=0.0060
step 4: global best x=3.01  f=0.0001
step 5: global best x=3.01  f=0.0001
step 6: global best x=3.00  f=0.0000

No gradients, no central plan — just particles each nudged toward their own and the swarm’s best, and the collective homes in on x=3. That’s the whole trick: local rules plus shared information yield global optimization.

Ant colony optimization, and the principle

Ant colony optimization (ACO) solves a different shape of problem (routing, the traveling salesman) with a different mechanism: pheromone trails. Ants wander, laying pheromone on the paths they take; shorter paths accumulate more pheromone (they’re traversed faster and more often), which attracts more ants, which lay more pheromone — a positive feedback that converges on short routes. This is stigmergy: agents coordinate indirectly, through changes they make to the environment, never communicating directly.

In one breath

• Swarm intelligence: useful global behavior emerging from simple local rules with no central controller (flocks, ant colonies) — turned into gradient-free optimizers.
• Particle swarm optimization (PSO): particles move toward their own best and the swarm’s best; simple local rules + a shared global best converge on the optimum (the demo reaches x=3).
• Ant colony optimization (ACO): ants lay pheromone, shorter paths accumulate more and attract more ants — positive feedback that finds short routes via stigmergy (coordinating through the environment, not directly).
• They’re gradient-free metaheuristics (evaluate, don’t differentiate) for black-box/combinatorial problems, decentralized and robust (no single point of failure).
• They’re a clean model of emergence — intelligence in the interaction, the same idea behind generative agents and collective LLM-agent behavior.

Quick check

Next

Swarm optimization is collective behavior optimizing rather than learning; it’s a clean instance of the emergence behind generative agents. Modeling what other agents think — for coordination — is theory of mind.