Concepts

Why Stratego is hard

Stratego is chess with the identities hidden. Each player arranges forty pieces in secret, so before the first move the opponent's army can be any of more than arrangements, and the game tree has about states (Perolat et al., 2022). Poker, the other classical imperfect-information benchmark, hides a two-card hand: 1,326 possibilities, few enough to enumerate. The methods that solved poker by transforming the game into its public tree scale with the amount of hidden information, so they never reached Stratego. That is why the game stayed a human stronghold after chess, Go and poker had fallen, and why DeepMind's DeepNash reached the top three on an online platform but not the level of top human players.

B23F510B23F510B23F510B23F510B23F510B23F510What the observer knows· which squares hold an opponent piece· which of those have moved (not a Bomb or the Flag)· which were revealed by a battle or a long Scout move· how many of each type have been lostSetups per side: 40! / (8! 5! 4!³ 3! 2! 6!) ≈ 1.4 × 10³³Texas hold'em, for comparison: 1,326 hands.Bar length: probability under a uniform prior over the remaining pieces. The belief network learns a sharper one.
A hidden opponent row could be any arrangement of the remaining types, and each battle collapses part of that uncertainty.

Value depends on frequencies

In a perfect-information game the right move is the one leading to the position of highest value. Under hidden information the value of a move depends on how often you would have made it in other situations, because the opponent's beliefs, and therefore their replies, are shaped by those frequencies. A Marshal that always charges a hidden piece gets baited onto a Bomb. A Marshal that never does gets ignored. The right frequency is a property of the whole strategy, not of the position.

The Ataraxos paper draws these dependencies as a graph: nodes are decisions, and an edge says whether making one decision more often raises or lowers the value of another. Even among decisions made at the same moment the graph is dense, and the dependencies also run across time, since earlier decisions determine the posterior over hidden pieces at later ones.

bluff with 4attack with 10probe with 2hold the lanesacrifice a 3more of one raises the othermore of one lowers the other
Changing how often one decision is made raises the value of some decisions and lowers the value of others.

What this rules out

Search that simply maximises value at each node, the recipe behind chess and Go engines, is not enough, because a pure best response is predictable and therefore exploitable. Enumerating the hidden information, the recipe behind poker solvers, is impossible at . What remains is learning a stochastic strategy directly, and finding a way to make search respect that strategy's frequencies. DeepNash did the first with game-theoretic dynamics; Ataraxos did both with regularised self-play and a search that mimics one more learning step.

The size of the hidden information

Forty pieces of twelve types give distinct arrangements, about , before a single move. Each battle reveals two types and each long Scout move one, and a moved piece rules out Bomb and Flag, so the posterior narrows through the game; but the number of configurations consistent with a mid-game observation is still astronomically beyond enumeration. Every player on the ladder has to answer the same question with sampling: how many configurations to draw, from which distribution, and how to act on values averaged across them.

What this program studies

The engine here plays the competitive rules at 84 million random moves per second across this workstation's 32 threads, which makes it possible to try each of these ideas at a small scale. The ladder starts with search over sampled hidden configurations, moves to learned evaluations, and ends with the two learning rules from the papers compared under the same compute budget. The question is which ideas carry the strength and which only carry the cost.

Continue with the rules the engine enforces or skip to self-play reinforcement learning.