Sengin Technologies

Learning to Play Rozz

Neural Evaluation for a Simultaneous-Move Economic Strategy Game

Matt Goddard — Sengin Technologies

We present a neural position evaluator for Rozz, a simultaneous sealed-bid strategy game in which all players commit moves secretly before any are revealed. Unlike sequential games, Rozz admits no deterministic lookahead: the optimal move depends on the opponent’s simultaneously-chosen move. We demonstrate that a pure state evaluator — trained without game-tree search — can achieve top-of-field play in a 22-player evaluation arena, matching or exceeding all hand-crafted heuristic opponents. The training methodology, built around a novel anticipatory “foil” mechanism, select-play against a bench roster, and exogenous exploration injection, surfaces 35+ publishable findings on the structure of simultaneous-move learning.

Selected Novel Contributions

• Foil starvation & endogenous training support — In self-play with a greedy foil, the support of the training distribution is determined by the foil policy; actions whose refutation lies outside the foil’s predicted support are permanently misevaluated. Remedy: exogenous exploration injection (forced opening pass).
• Feature leakage is a “when,” not “whether,” problem — Positional features (victory points, defeat status) are legitimate information at and after specific game rounds; the correct fix is a round-based gate, not feature removal.
• Oscillation is not one phenomenon — A trainable “level” (moved sharply by data/domain fixes) is separable from an intrinsic “variance” arising from the game’s simultaneous-bid randomness.
• Option-set completeness matters independently of evaluation quality — Providing the model with the full response option set (including multi-move combinations) independently improved play quality with no change to the evaluator weights.
• Endogenous support recurs at the opponent-selection level — A model cannot learn to consistently beat an opponent it is not trained against, independent of foil calibration — the same structural finding, instantiated one level up the training loop.