PRIMORDIUM — a coevolution experiment

▲ PREDATOR · gen 0

◈ PREY · gen 0

Live encounteri

Two evolved champions, each steered only by its own neural network. Drag the dials to pit a predator from one generation against prey from another — the heart of the Red Queen test.

Predator generation ⓘ0

Prey generation ⓘ0

lock to same generation

—

Pre-registered testi

In one line: do prey strategies keep cycling when the predator evolves back, but settle down when it can't? We run both cases and compare.

H1. Under live coevolution, the prey's evasion-strategy trait oscillates (cyclicity) more than under a control where prey adapt to a frozen predator panel.

H0. No difference: coevolution yields no more cycling than directed convergence.

Statistic: depth of the deepest negative auto-correlation trough of the detrended trait series (lags 1–10). Test: two-sided permutation test, 20 000 shuffles. Decide before running; the verdict is computed live.

Generations per run ⓘ

Replicates per condition ⓘ

—

Heavier settings give tighter confidence intervals but take longer. Everything runs in your browser; nothing is sent anywhere.

No experiment run yet. Set the parameters and press Run experiment.
A typical 45×8 run takes roughly half a minute.

Generation-by-generation series for one example lineage per condition — open it anywhere and study it yourself.

How the prey's strategy movedi

coevolutioncontrol (frozen target)

The prey champion's escape behaviour, generation by generation. Green wavering = the strategy keeps changing (cycling); amber settling toward a line = it found one answer and stuck. The wobble is the thing we're testing for.

Is the wobble real, or luck?i

coevocontrol— median

Every lineage gets one "how cyclic was it" score (a dot); the bar is the group's middle. Evolution is noisy, so we ask: does the green cloud really sit higher than the amber one, or could chance alone produce this gap? That's the p-value.

The arms race — why the scoreboard liesi

coevo: predator fitnesscontrol: predator catch-rate

Watch the amber line fall: against frozen predators, evolving prey get away more and more — clear, visible progress. The green line stays flat near the middle: under live coevolution both sides improve at once, so the raw score barely moves. The prey are getting better — the scoreboard just can't see it. This is the single most important reason a coevolution study needs a control.

The brains grow themselvesi

coevolutioncontrol

Active connections in the prey champion's brain, generation by generation. Nobody designed these networks — they start minimal and add wiring only when the extra complexity earns its keep. Rising lines mean evolution is discovering it needs memory or finer control to keep up.

Every era's predator vs every era's prey (CIAO)i

A round-robin tournament across time: each predator generation (rows, top→bottom) plays each prey generation (columns, left→right); brighter = the predator won. A clean bright-to-dark diagonal means newer always beats older — steady, ladder-like progress. Patchy, mottled bands mean a newer predator sometimes loses to an ancestral prey (intransitivity) — like rock-paper-scissors, the fingerprint of Red Queen cycling. Intransitivity index: —

Evolved controlleri

An actual genome from the current lineage. Topology, recurrence and plasticity all arose by mutation and selection — none were designed.

Generation ⓘ0

▬ excitatory ▬ inhibitory
┄ dashed = recurrent / self-loop (memory)
◦ ring = plastic synapse (lifetime learning)
node brightness = its live activation.

A field monograph · methods & honest limits

Does predator–prey coevolution cycle?

A self-contained, reproducible experiment — not a discovery. It implements established mechanisms to test one clear prediction, with a control, replicates, and a verdict computed from the data rather than asserted — including a verdict of "no effect."

Run the experiment in the Laboratory tab to populate the live result here.

i. The three mechanisms (and that they really emerge)

Recurrence / memory. Each controller is updated synchronously in discrete time; any connection that points "backward" (or a self-loop) carries the previous step's activation forward, giving genuine internal state. Within-lifetime learning. Every synapse owns an evolvable learning-rate gene eta; weights are nudged during the encounter by reward-modulated Hebbian plasticity, where the neuromodulatory signal is the agent's instantaneous fitness gradient (closing distance for the predator, opening it for the prey). These lifetime changes are not inherited — evolution acts on the rule, not its results. Evolvable topology. NEAT-style structural mutation adds nodes and connections under a global innovation registry, with compatibility-distance speciation and explicit fitness sharing. Across ten runs these were not hypothetical: hidden nodes appeared in 9/10, recurrence in 10/10, plastic synapses in 10/10.

ii. The task and the fitness

A toroidal pursuit–evasion arena. The predator is slightly faster; the prey slightly more agile. Payoff is continuous and zero-sum — a faster capture scores the predator higher, and merely staying close earns partial credit — so there is always a selection gradient and no saturating ceiling. Each individual is scored against a random sample of the opposing population over several encounters with randomized starting geometry, to discourage overfitting to one opponent.

iii. Why a control, and which one

Cliff & Miller's central warning is that under the Red Queen, naïve fitness curves mislead: a population can improve while its measured score stays flat, because the opponent improves too. So a single coevolving run proves nothing about cycling. The control here holds the predators frozen (a competent panel harvested from a warm-up run) while the prey still evolve. Same observable, same selection strength — but a static target. The clean prediction: prey should converge against a fixed challenge and cycle only when the target moves. The difference between conditions is the evidence.

iv. Two lenses, which can disagree

The primary test tracks a one-dimensional prey trait (evasion alignment) and asks whether it oscillates more under coevolution. The secondary lens is a CIAO matrix pitting every predator generation against every prey generation; non-monotone bands reveal intransitive dominance directly. These can disagree — a rich intransitivity in the matrix that a single scalar trait fails to capture is itself a finding about how much a low-dimensional behavioural summary can hide.

What this is not. It is not a novel result. Cliff & Miller coevolved recurrent pursuit/evasion networks and built CIAO plots three decades ago; this reproduces that lineage in miniature. Population sizes, mutation rates and arena constants are hand-tuned for tractability in a browser, not derived. With a handful of replicates the permutation p-value is noisy; treat a single run as suggestive, not decisive, and raise the replicate count to tighten it.

On honest reporting. The verdict is whatever the data say — including "no effect." Coevolution frequently fails to produce clean cycling, collapsing instead into disengagement or "mediocre stable states" (Ficici & Pollack). A null here is a real outcome, not a failure of the instrument, and the code does not tune itself toward significance.

v. What a "null result" really means

If a run reports no effect, it is tempting to feel the experiment failed. It didn't. A null result is a genuine measurement: under these conditions, we could not distinguish coevolution from the control. That is information. It rules things out, it bounds how big any effect can plausibly be, and — crucially — it is the outcome an honest test must be willing to return, or the test was never really a test.

Three different things can hide behind a single null here, and they are worth telling apart. (1) There is genuinely no extra cycling — the prey found one good escape and the predator couldn't force them off it. (2) There is cycling, but our ruler is too blunt — a single number (one prey trait) can't see motion that's spread across many dimensions of behaviour; this is exactly when the CIAO grid lights up with intransitivity while the trait line looks calm. (3) The run is just underpowered — too few replicates or generations, so real signal is buried in noise. More replicates and longer runs separate (3) from (1) and (2); comparing the trait verdict against the CIAO grid separates (2) from (1).

vi. What happens in real biology when the race doesn't run

Coevolution failing to produce a clean, cycling arms race is not an artifact of toy simulations — it is common in nature, and biologists have names for the ways it happens.

Stalemates and "mediocre stable states." Two species can settle into a stand-off where neither can cheaply improve against the other, and selection for further change goes slack (Ficici & Pollack named this failure mode in coevolutionary systems). Disengagement. One side can get so far ahead that the other effectively drops out of the contest — the pressure that drove the race disappears. Evolutionary stasis. Much of the fossil record shows long stretches where lineages barely change ("living fossils" like horseshoe crabs, coelacanths, and crocodilians have held a body plan for tens or hundreds of millions of years), because what they have already works and nothing keeps pushing. Extinction. The bleakest resolution of Van Valen's original observation: if you cannot keep running, you fall behind and disappear — which is why extinction risk stays stubbornly present no matter how long a lineage has already survived.

And where the race does run, biology shows it directly: cycling host and parasite genotypes in Daphnia revived from pond sediment, sexual snails out-competing clones precisely where parasites are thick, garter snakes and their toxic newt prey escalating poison and resistance in lockstep. The point of a control and an honest null is to tell these apart — to know whether you are watching a real Red Queen or a stalemate wearing its costume.

vii. Lineage of the methods

Van Valen, L. A New Evolutionary Law. Evolutionary Theory 1: 1–30, 1973. (The Red Queen.)

Cliff, D. & Miller, G. F. Tracking the Red Queen: Measurements of Adaptive Progress in Co-Evolutionary Simulations. ECAL 1995, pp. 200–218. (CIAO plots; coevolved recurrent pursuit/evasion nets.)

Stanley, K. O. & Miikkulainen, R. Evolving Neural Networks through Augmenting Topologies (NEAT). Evolutionary Computation 10(2): 99–127, 2002.

Soltoggio, A., Bullinaria, J. A., Mattiussi, C., Dürr, P. & Floreano, D. Evolutionary Advantages of Neuromodulated Plasticity in Dynamic, Reward-based Scenarios. ALIFE XI, 2008. See also Soltoggio, Stanley & Risi, Born to Learn (EPANNs), Neural Networks, 2018.

Ficici, S. G. & Pollack, J. B. Challenges in Coevolutionary Learning: arms-race dynamics, open-endedness, and mediocre stable states. Artificial Life VI, 1998.

Yaeger, L. PolyWorld: Life in a New Context. Artificial Life III, 1994. (Neural-net agents in an ecology — the spiritual ancestor.)

Citations verified against primary/reliable sources. My knowledge has a cutoff; newer work past it would not be reflected here, and exact page references should be double-checked against the originals before formal use.