Choose the role you teach

Hider: survive 120 decisions. Seeker: reach the hider’s tile before that limit. The hider gets six extra opening decisions. Both roles have their own weights.

Each bot receives its own position, the fixed cover map, visits, and only a visible rival or a sighting from the last eight decisions. Detection is a geometric line-of-sight sensor with a five-tile Manhattan range, not camera pixels.

Each network is a single linear neuron with eight learned weights. Inputs describe path distance to a known rival or patrol corner, predicted occlusion, adjacent cover, visits, waiting, sight and memory availability, plus a constant bias input.

Gradient descent imitates authored action scores. Rules supply the map, path distances, legal actions, sensor filtering and memory. The networks learn action scoring; there is no rescue policy, pixel recognition or adversarial self-play.

48 fitting seeds, 8 validation seeds, 16 final seeds. Validation chooses weights before final testing. Each role plays against the reference opponent. Up to 60 epochs per role, 30 seconds and two attempts per training round.

Ready · active weights stay unchanged while training.

Activation requires ≥80% near-best reference-score agreement for both scorers and at least 8/16 complete-round wins for your selected role, without regression against its initial or checked incumbent wins. Fit success and goal success are separate. The reference can lose, so accurate imitation alone can fail the win goal. Failed-round retries label the actual failed role observations and keep fresh final seeds separate.