AI can now outmaneuver you in both Stratego and Diplomacy • TechCrunch

While artificial intelligence long ago surpassed human skills in chess, and recently Go – and let’s not forget punishment — other more complex board games still present a challenge to computer systems. Until recently, Stratego and Diplomacy were two of those games, but now the AI ​​has become good at the first table and passably human at the second.

On the surface, you might think this is just because these games require a certain level of long-term planning and strategy. But so do Go and Chess, just in a different way.

The essential difference is in fact that Stratego and Diplomacy are strategy based games imperfect information. In Chess and Go, you can see every piece on the board. Stratego hides the identity of pieces until they are encountered by another piece, and Diplomacy is mostly about establishing deals, alliances and, of course, vendettas that are kept secret but are essential to the game. No fair game of chess will involve a third party stepping in to protect your opponent’s bishop with a blue stone.

Both games don’t require crude calculations of paths to victory, but softer skills like guessing what your opponent is thinking and what they think the computer is thinking, and making moves that match and hopefully disrupt those guesses. In other words, one has to bluff and convince another player of something, not just beat them with the best moves possible.

The model that plays Stratego, from DeepMind, is called DeepNash, after the famous balance. It focuses less on clever moves and more on plays that cannot be exploited or predicted. In some cases this can be brave, like a game the team saw against a human player where the AI ​​sacrificed some high-level pieces, leaving it at a material disadvantage – but it was all a calculated risk for take out the other player’s big guns. , so she can strategize around them. (He won.)

DeepNash is good enough that it beats other Stratego systems almost every time, and 84% of the time against experienced people. Because the algorithms that work well in Go and chess don’t work well here, they invented a new algorithmic method called Regularized Nash Dynamics – but you have to read the letter if you want to understand it deeper than that. In the meantime, here’s a marked game:

From the side of diplomacy, we have an AI named Cicero (ah, hubris!) from Meta and CSAIL that manages to play the game on a human level – and if that sounds like damning with faint praise, remember Diplomacy is hard for most people to play on a human level. The level of trickery, backstabbing, false promises and general Machiavellian shenanigans that people get up to in the game is such that it is banned by many friendly gaming groups. Is a computer really capable of that level of trickery?

It seems so, and the advances that make it possible are interesting. After all, the interesting part of Diplomacy is not the map and parts of the world, which are fairly simple to read and appreciate, but the potential for hidden schemes in those deals. Is Venice being threatened on two fronts, or is it luring the western front into a loop through long thought? times-face?

Not only that, but in order to participate in the scam, you have to talk (or chat, online) with other players and convince them of your sincerity and intention. This takes more than CPU cycles!

Here’s how Cicero works:

1. Using the state of the board and the current dialogue, make an initial prediction of what everyone will do.
2. Refine that prediction using planning and then use those predictions to create a goal for yourself and your partner.
3. Generate several candidate messages based on the state of the board, its dialogue and goals.
4. Filter the candidate’s message to reduce nonsense, maximize value and ensure consistency with our goals.

Then, accept your chance and hope the other player doesn’t plan your death.

When released on webDiplomacy.net, Cicero played quite well against his opponents, ranking 2nd out of 19 in a league and generally outplaying the others.

It’s still a work in progress—it can lose track of what others are being told, or make other mistakes that people probably wouldn’t—but it’s too remarkable to be competitive at all.