← Workbench

Faber Neuroevolution · live workbench · Programme 7

Neural coevolution: pursuers vs evaders

This is the real engine. Two populations of faber-tweann networks coevolve on a torus; every move is a live forward pass. It is the honest counterpart to the numerical Red Queen, which isolates the same measurement ideas where the truth is directly readable.

Pursuers want to catch, evaders want to survive. Given a small speed edge the pursuer learns to catch, then dominates: the evader's gradient dies. That is disengagement, one of the three coevolution outcomes, and it is why a genuine arms race is hard. Watch the evolved champions play, then read the curves that never became a race.

Champion match

step 0/12

pursuer (1.5× speed)   evader

Coevolution (no arms race)

1 0 generations → pursuer evader head-to-head

progress vs frozen gen-0 opponents; head-to-head pinned = disengagement

The curves never climb into a mutual arms race: the pursuer's benchmark progress rises as it learns to catch, and its head-to-head score pins at the top -- it dominates the current evader, which can never pull ahead. Tip the balance the other way -- equal speed -- and the evader escapes forever instead. A sustained two-sided arms race sits on a knife-edge between, and finding it is the open problem the embodied Programme-7 rung exists to tackle. That is the honest state of the art, measured with the instruments the numerical rungs proved correct.