An ant colony that forages well has committed to a trail. A colony that can change its mind has not committed too hard. Eric Bonabeau’s argument is that both are possible at once only if the colony sits just past a bifurcation: ordered enough to build a trail, close enough to the instability to abandon it.
Interactive companion to Bonabeau E., Marginally Stable Swarms Are Flexible and Efficient (1996) and its follow-up Flexibility at the Edge of Chaos (1997). Every curve below is simulated in your browser from the paper’s model and parameter values. Nothing is pre-rendered.
Each dot is one ant in one of four states: at the nest, foraging at source 1, foraging at source 2, or lost. Ants at the nest are recruited to a source in proportion to how many ants are already there and how strong the trail is. The knob that matters most is N, the number of active foragers. Slide it and watch the colony go from random to committed.
Transition rules follow equations (2) to (5) of the paper with b = 0.017, c = 0.018, g = 25, p = 0.15, a2 = 0.03 and a1 = (1 − Δ) a2. With χ below 1 the trail is a decaying pheromone field (equation 7) instead of the instantaneous head count. One addition: an ant at the nest leaves to explore on its own with probability 0.003 per step. As printed, the rules only let ants leave the nest when someone is already foraging, so an idle colony would stay idle forever; the small exploration term gives the colony the noise floor the paper’s figures show below the bifurcation. Runs differ from one another, and near the critical point they differ a lot, which is the point.
Run the colony to its steady state for every N from 10 to 400 and plot the order parameter M = N2 − N1. Below a critical N the colony splits evenly. Above it, the symmetry breaks. In experiment E3 the interesting part is the window where the colony, already committed to source 1, still switches to the richer source 2 when it appears.
Below the bifurcation. Both sources used equally; no structured trail, nothing to exploit.
Just past the bifurcation. A trail forms, and a richer source introduced later still pulls the colony over.
Deep in the ordered phase. The first trail is self-reinforcing and the colony cannot leave it.
Each point averages the last 3000 of 6000 steps over 4 independent runs of the same rules as the live colony, computed when you change a control. For E3 the colony first settles with source 1 alone, then source 2 is switched on and the run continues for another 6000 steps. The paper used 100 runs of 10⁵ steps, so its curves are smoother and its thresholds (Nc = 185; switching for 115 < N < 295 at Δ = 0.03) sit lower than what the exploration term produces here. The shape is what matters: flat, then a window, then lock-in.
The window where switching succeeds grows with the quality gap. The paper reports the width S scaling as Δ1/δ with δ = 0.85 ± 0.02. Compute it here for the current χ.
Strip the colony down to one number M and one control μ. Near a symmetry-breaking bifurcation the dynamics reduce to a particle relaxing in a double well. A quality difference between the sources is an external field h that tilts the well.
Two colonies sit in the same (left) well with h = 0. Colony A is barely past μc; colony B is far into the ordered phase. Raise h and see which one rolls over.
The left well disappears when |h| exceeds 2((μ − μc)/3)3/2. For A that threshold is tiny, so a small advantage is enough to move it; for B it is large, so B stays put. A below-threshold h does not move A either, which is the efficiency side of the argument: the colony does not thrash on every small fluctuation.
In the 1997 follow-up (Acta Biotheoretica 45:29–50) the model is extended with a parameter β for the recruitment strategy. Trail-following precision becomes F = (1 − β)k + β (g + X²)/(h + X²): at β = 0 a recruiter simply leads a group to the source with fixed reliability k, at β = 1 the colony relies entirely on an anonymous pheromone trail whose reliability grows with the number of ants on it. Real species mix the two.
Experiment E3 integrated from the mean-field equations with the Beckers et al. (1990) values: a1 = 0.028, a2 = 0.03, b = 0.017, c = 0.018, p = 0.15, g = 25, h1 = 37, h2 = 35, k1 = 0.58, k2 = 0.6. Because group recruitment keeps F well above zero even with few ants on the trail, the null state is unstable here and no exploration term is needed. Pure mass recruitment (β = 1) never switches at any N. The richer source is exploited most heavily just before the cliff near β ≈ 0.8, which is the paper’s Figure 13: flexibility is assured if the mix of strategies puts the colony on that edge.
Three things are true at the same time in the marginally stable band:
Efficient. The colony is past the bifurcation, so recruitment is autocatalytic and a strong trail to one source is built and maintained.
Flexible. The colony is close to the bifurcation, so the well it sits in is shallow and a modest field, a richer source, tips it into the other state.
Not twitchy. Flexibility is bounded: a gain smaller than the tilt threshold does not move the colony. The ordered phase filters noise.
The paper maps these to real species: mass recruiters with large colonies (behavior B1) commit and cannot switch; tandem recruiters with tiny colonies (B3) never commit; species that mix group and mass recruitment (B2) do both, and the β model above says why: the mix itself places them at the edge. The last section of the paper asks how a colony stays in such a narrow band when its parameters fluctuate, and sketches candidate answers: a parameter that drifts with hunger, periodic bursts of synchronized activity, feedback in the style of self-organized criticality, and stochastic resonance.