Approaches

World models for a cheap environment

Does a model-based learner have anything to offer when the environment runs at 84 million plies per second and every wall-clock second goes to network inference?

The mechanism

NashDreamer (Holeček, 2026) trains a centralised multi-agent recurrent state-space model on both players' observations, with per-player infoset models that reconstruct each player's own history, and learns policies by regularised Nash dynamics inside imagined rollouts. Its results are on Goofspiel and Battleship and its claim is fewer real environment steps for the same exploitability. Its stated limits are full unrolls from the first move and a posterior collapse in stochastic games.

Two of those facts matter here. Stratego games run for hundreds of plies, so imagination from the first move is impractical. And the training loop's cost is the network, not the environment: at 4,100 steps per second on the small preset the engine's share of wall time is small (the throughput study measures it). The part that may transfer is the infoset model, a belief state learned by the centralised objective rather than by the supervised hidden-piece decoder of the Ataraxos belief network.

The plan

One study, infoset-belief, in two stages with a stop between them. First, take the environment's share of training wall time from the throughput study; if it is under 20%, imagination is ruled out and the page says so. Second, train a NashDreamer-style infoset model on the same self-play data as the belief network at equal parameters and compare the held-out hidden-piece log-likelihood. The budget is one GPU-day.

Decision rule

The infoset model becomes a second sampler for the belief study if its held-out log-likelihood is within 5% of the belief decoder's; otherwise the direction closes with the numbers filed.

What could disprove it

An environment share above 20% would reopen imagination as a lever and send the study back for a new plan. An infoset model far below the decoder closes the direction.

Evidence

None yet.