Reference

Glossary

Arbiter state. The full game state including hidden pieces, held only by the engine and the server.

Observation. What one seat may know: its own pieces, revealed opponent pieces, which hidden opponent pieces have moved, the counts of lost pieces, the counters and the public anti-chase trackers.

Information state. The paper's term for the observation plus the history features fed to the networks.

Determinization. A full state sampled to be consistent with an observation, used by expectimax and rollout search.

Belief. A distribution over the types of hidden opponent pieces. The uniform belief comes from counting; the learned belief comes from the belief network.

Threat, evade, chase. A threat moves a piece adjacent to an opponent's movable piece without attacking. An evade moves the threatened piece away from the threatening piece. A chase is an unbroken alternation of threats and evades.

Two-square rule. No piece may cross the same square boundary on more than three consecutive turns of its owner.

Battleless-move rule. The game is drawn after k consecutive moves without a battle, k being 100 in training and 200 in competitive play.

Effective win rate. Wins plus half the draws, divided by games.

Paired seeds. Each seed is played twice with the colours swapped.

Damping. The paper's coordination of regularisation strength and update size over training: strong regularisation and large steps early, weak regularisation and small steps late.

Magnet policy. The move policy that picks a movable piece uniformly and then one of its legal moves uniformly; the target of the reverse KL regulariser.

Update equivalence. Treating test-time search as one more policy update, so the same regularisers and step control apply.

Magnetic mirror descent. The tabular update that maximises expected value minus two reverse KL penalties, giving a closed-form policy proportional to (exp(q) rho^alpha pi^beta)^(1/(alpha+beta)).

Regularised Nash Dynamics (R-NaD). DeepNash's learning rule: transform the rewards with a log-ratio penalty toward a regularisation policy, run replicator dynamics to the fixed point, and make that fixed point the next regularisation policy.

Rung. One player on the research ladder; a rung is promoted when it beats the previous rung under a preregistered decision rule.