§ 01What "optimal" means in a hidden game

In chess, "optimal" has a fixed meaning: best move for the position, both sides see everything. Stratego is different. Your deployment happens before you know anything about your opponent's deployment — there are more than 1066 legal arrangements of the 40 pieces, and the full game tree contains on the order of 10535 states, far beyond Go or no-limit hold'em. No computer has "solved" Stratego, and no lookup table of optimal setups exists.

So when players ask for "the game-theory-optimal setup", the honest answer is: no single formation is optimal, because any fixed formation can be read and countered. What game theory offers instead is something better — a way to be unexploitable.

§ 02What game theory actually says

The core concept is the Nash equilibrium: a strategy (or mix of strategies) where neither player can improve by changing their own approach while the other stays put. In rock-paper-scissors, the equilibrium strategy is well known: play each option exactly one-third of the time at random. Not because any single throw is "best", but because a random mix cannot be exploited.

Stratego deployment works the same way, scaled up enormously:

  • A fixed setup is a pattern. Patterns get read. A read pattern gets countered — Miners steered to your corner, Scouts aimed at your flag row.
  • A randomized mix of structurally different setups is unexploitable in the opening. Your opponent's best counter to an unknown distribution is... nothing targeted. They must play the board, not your habits.
  • Equilibrium play guarantees a floor. In a perfectly played symmetric game, unexploitable play ensures you can't be pushed below an even result — the exploitation has to come from their mistakes.

This is why asking "what's the best Stratego setup?" is like asking "what's the best poker hand to always raise?" The correct answer is a range, not a card.

§ 03The DeepNash evidence

This isn't just theory. In 2022, DeepMind published Mastering the game of Stratego with model-free multiagent reinforcement learning in Science, introducing DeepNash — an agent trained with a game-theoretic method called Regularised Nash Dynamics (R-NaD), which steers learning toward a Nash equilibrium. Its scoreboard, verified on the Gravon games platform:

  • 97%+ win rate (often 100%) against the strongest existing Stratego bots, including Computer Stratego World Championship winners;
  • 84% win rate against human expert players, earning an all-time top-three ranking on Gravon.

DeepMind's own analysis of how it played is a masterclass in deployment theory:

To become hard to exploit, DeepNash developed an unpredictable strategy — creating initial deployments varied enough to prevent its opponent spotting patterns over a series of games, and randomising between seemingly equivalent actions to prevent exploitable tendencies.

Two behaviours stand out for setup design:

  • It priced information above material. In a published game it deliberately sacrificed a Major (7) and a Colonel (8) early — falling behind on the board — purely to locate the opponent's Marshal (10), General (9), another Colonel and two Majors. Material behind, map ahead, it evaluated its winning chances at 70% and converted the win.
  • It bluffed with structure. It chased a known enemy Colonel with a mere Scout (2), behaving like a Marshal. The human, convinced the pursuer was the 10, tried to ambush it with their Spy — and DeepNash traded a nearly worthless piece for the enemy Spy, the single most dangerous piece to its Marshal. The bluff wasn't theatre; it was an information-forcing mechanism baked into how it deployed and moved.
THE LESSON

The strongest Stratego player ever measured wins not with a secret super-formation, but with varied deployments, randomized choices, and aggressive information trading. Copy the process, not a picture.

§ 04Five principles of unexploitable deployment

1. Rotate archetypes, not squares

Don't vary one setup slightly — vary its structure. If your flag was cornered last game, vault it mid-back next game, then run a decoy corner the game after. The goal is that an opponent who has played you five times still can't describe "your setup".

2. Do the Bomb-and-Miner arithmetic

Each side has 6 Bombs and 5 Miners. A sealed corner flag costs you 2 Bombs; a full three-Bomb vault costs 3. For the enemy to reach a sealed corner flag, they must defuse 2 Bombs — spending 2 of their 5 Miners, or trading pieces into them. When you choose where the Bombs go, you are literally writing the price list the enemy must pay. Spreading the cost across the board (flag vault + lane traps + Marshal shield) forces them to spend Miners on uncertainty, not on your actual flag.

3. The flag is a message — make it a lie

Opponents read your back row. A corner stuffed with Bombs says "flag here". That message is only true as often as you make it true. A cheap decoy cluster (2 Bombs in a corner, flag elsewhere) taxes their Miners and their attention. Just remember the tax only stays profitable if the flag sometimes is in the corner.

4. Screen the Marshal, escort the Spy

Your Marshal beats everything except an attacking Spy — its one death sentence. So: keep a General or cheap screen in front of it so it never fights blind; keep your own Spy within counter-attack range of the corridors the enemy Marshal will use; and track the enemy Spy's location every single turn. The moment it dies, your Marshal is a lawless king.

5. Buy information with small denominations

DeepNash traded a 7 and an 8 for a map of the enemy's high command — and won. You can afford the same in miniature: spend Sergeants and Scouts deliberately to force reveals, especially around pieces that might be Bombs or the flag. A setup that never intends to trade is a setup that never learns.

§ 05Two deployable archetypes

Below are two structurally different formations (your four setup rows, back row at the top, viewed from your side of the board). They are teaching examples — the optimal strategy is rotating between this kind of structure and others, never planting one of them as "your setup".

ARCHETYPE A — "CORNER FORTRESS" · BACK ROW ON TOP
Flag Bomb Piece (Mr Marshal · Ge General · Co Colonel · Ma Major · Ca Captain · Lt Lieutenant · Sg Sergeant · Mi Miner · Sc Scout · Sp Spy)

What it does: the flag sits in the back-left corner sealed by exactly 2 Bombs — the minimum fortress. The right-side Bombs and the mid-board Bomb pair act as lane traps and a slow-down screen. The Marshal starts one row behind a General screen, and the Spy lurks where it can slide toward the left corridor to hunt an overextended enemy Marshal. This is the classic "tax them to reach the corner" school.

ARCHETYPE B — "DECOY CORNER, MID-BACK VAULT" · BACK ROW ON TOP
Flag Bomb Piece (abbreviations as above)

What it does: the left corner screams "flag fortress" — but it's a decoy with no flag behind it. The real flag sits mid-back inside a 3-Bomb vault. Enemy Miners committed to the fake corner are Miners not defusing the vault. The lone front-row Bomb is a mid-lane trap that forces the opponent to spend a Miner (or a piece) before your flag row is even threatened. Costlier than Archetype A, but it plays the opponent, not just the board.

CAVEAT

These two archetypes are deliberately different in structure — corner flag vs. vault flag, 2-Bomb seal vs. 3-Bomb vault, decoy present vs. absent. That contrast is the point: your opening strategy should be a rotation across such structures, with small random variations each time.

Theory needs live fire

Test your rotations against the built-in AI — full combat rules, instant battles.

Play the Blitz Skirmish →

§ 06How to stress-test a setup

Before a setup enters your rotation, interrogate it:

  1. The Miner audit. Which squares must an enemy Miner defuse to reach your flag? Fewer than 2 means your flag is cheap; more than 4 means you've spent too many Bombs in one place.
  2. The first-contact test. Play ten quick games. Does the setup survive its first revealed piece without collapsing a whole flank?
  3. The Scout-lane test. Can an enemy Scout reach your back row in one move down any open file? If yes, is the piece sitting there one you're happy to reveal or lose?
  4. The Marshal path. Trace how your Marshal gets into the fight. If it must cross a narrow, spy-friendly corridor, redesign.
  5. The pattern check. After five games with this setup family, ask a regular opponent to describe "your deployment". If they can, the rotation is too slow.

§ 07Frequently asked questions

Is there a single best Stratego setup?
No. Since deployments are secret and simultaneous, every fixed formation has counters. Optimal play is a mixed strategy — a rotation of structurally different setups with random variation, so opponents can't exploit patterns.
Where is the safest place for the flag?
No square is safe by itself. A 2-Bomb sealed corner is the cheapest fortress; a 3-Bomb mid-back vault is costlier but deflects Miner raids; a decoy corner punishes pattern-reading. Safety = unpredictability × Bomb-and-Miner arithmetic.
How many Bombs should guard the flag?
Most strong setups spend 2–3 Bombs on the flag and keep 3–4 as lane traps or a Marshal shield. All six around the flag both telegraphs its location and leaves the rest of the board unmined.
What's the strongest evidence about optimal play?
DeepMind's DeepNash, published in Science in 2022. It reached 97%+ wins against championship-level bots and an 84% win rate against human experts (all-time top 3 on Gravon) — while deliberately randomizing deployments and trading material for information.
Should my Marshal start up front?
Generally no. Screen it with a General or cheap piece, keep it out of blind fights, and commit it once the enemy Spy is located — or dead.

Sources: "Mastering the game of Stratego with model-free multiagent reinforcement learning", Science, 2022 (DeepNash / R-NaD); DeepMind's public research blog; Gravon platform results as reported by DeepMind. Win-rate figures are DeepMind's published measurements. Stratego® is a registered trademark of Jumbo / Hasbro; this is an independent fan site.