What Stratego Reveals About the Cost of Good AI Decisions

Key takeaways

• Stratego reveals where opposing pieces are, but initially hides their identities.
• Belief models track possible identities and update those estimates as evidence arrives.
• Reinforcement learning can reward moves that gather useful information.
• Evaluating efficiency requires separating training costs from the cost of choosing each move.
• Real-world systems also need to recognize when their underlying assumptions fail.

Before a Stratego AI can choose a good move, it has to work out what it might be attacking. For Ataraxos, an AI built for the game, the challenge is making good decisions with hidden information while keeping computation affordable. That combination matters wherever AI must act before all the facts are in.

A visible board with invisible identities

In chess, you can tell an opposing queen from a pawn. In Stratego, you can see where an opponent’s pieces stand without knowing their ranks.

That makes Stratego an imperfect-information game. The difficulty goes beyond calculating what could happen next. Part of the current situation is already hidden.

An inviting target might be weak. It might also be a powerful piece positioned to lure you in. Searching further ahead helps only if your assessment accounts for what that piece could actually be.

Good play therefore requires two connected judgments: what you might be facing, and what to do about it.

Belief is a probability, not a conviction

A belief model estimates hidden information from observable evidence. Think of it as maintaining several possible explanations, each with a different probability.

Take bombs. In Stratego, they cannot move. If an opposing piece moves, you can rule out a bomb. If it stays still, you cannot conclude that it is one.

The same observation can eliminate one possibility while barely narrowing the others.

An opponent pushing a piece forward aggressively might suggest strength. It could also be a bluff. A useful model updates its estimates without treating its interpretation of the opponent’s intentions as established fact.

This describes how belief models can work. Whether a particular implementation handles uncertainty well is a separate performance question. Giving an AI a probability distribution does not make that distribution accurate.

Sometimes a lost piece buys a better decision

Once a system estimates the situation, reinforcement learning can help it learn which action to take. Actions receive feedback through rewards, allowing the system to improve its choices over time.

In Stratego, sacrificing a piece can reveal the identity of an important opponent. That knowledge may make later attacks or defenses more effective.

The sacrifice still needs to earn its keep. Information has strategic value only when it improves subsequent decisions enough to justify the immediate loss.

Learning that tradeoff is difficult. The system must connect an information-gathering move with a benefit that may arrive much later. Rewarding victory alone does not guarantee that useful probing strategies will emerge; the training process and experience available to the agent matter.

Part of playing well, then, is deciding what to find out next.

“Low compute” needs two separate bills

An AI can be inexpensive to run after being expensive to build.

Training cost covers the resources used to develop its playing ability. Inference cost covers the resources needed to choose a move once training is complete. Either can be high while the other is low.

A belief model does not automatically reduce either bill. It may help focus computation on more plausible situations, but running the model also costs resources. Incorrect estimates can steer attention away from the possibility that matters most.

For Ataraxos, an efficiency assessment would therefore need both performance results and resource measurements. How many games were played? Under which starting arrangements? How much decision time did each player receive?

Even a win against a champion would need that context. A single victory and a repeatable advantage across varied conditions carry very different evidential weight.

Outside the game, the assumptions can break

Stratego hides piece identities while keeping the types of pieces and combat rules fixed. The uncertainty has boundaries.

Real systems face a messier problem. Customer behavior changes. Sensors fail. Situations appear that were absent from training.

Consider equipment diagnosis. A system might assign probabilities to several possible faults, then choose another test to distinguish between them. That resembles the logic of a belief model.

But if the actual fault is missing from its list, confidently selecting the most likely listed cause still produces a wrong answer. The system must assess whether its existing explanations fit the situation at all. When they do not, requesting more information or deferring a decision may be the useful move.

A Stratego AI can play well without uncovering every hidden piece. The broader standard is making good choices at an acceptable cost while uncertainty remains. Outside the board, that also means recognizing when the possibilities you are considering are no longer enough.