← Workbench
Faber · coevolution methodology · insight 053
The Red Queen: running to stay in place
This is not neuroevolution.
The players here are plain numbers, not networks,
on purpose: it isolates the coevolution measurement trap
where the true progress
is a value you can read directly. It is the calibration of the ruler we use on the real
engine next door. For coevolving networks, see the neural coevolution demo.
A population competes with itself on a "bigger tends to win" game. Selection pushes the trait up without bound (real, observable progress). Yet the best individual's score against its current rivals never improves: it is always about as good relative to its peers, because they climb with it. Watch the two lines.
press Play to run the coevolution
The green trait climbs; the terracotta score stays flat. If you measured progress by "how am I doing against my current opponents", you would conclude nothing is happening. That is the Red Queen, and it is why coevolution must be measured against a fixed benchmark, not the moving opponent.