MIT and CMU's Ataraxos AI becomes the new Stratego champion

[Y2J_WCW]

For a lighter AI story, researchers from MIT, Carnegie Mellon, NYU and Stanford have built an AI called Ataraxos that has become the best Stratego player in the world. The name comes from Greek and roughly means unbothered or free from anxiety. That turns out to be a fitting description of how it plays

Stratego is a tough game for AI because it involves hidden information. You cannot see your opponent's pieces, so you have to guess, bluff and manage risk. Ataraxos combines self play reinforcement learning to build a strong base strategy with decision time planning, using a generative model that estimates where the opponent's pieces probably are and refines its moves on the fly. Unlike chess, there is no perfect information to calculate from. Good players rely on bluffing and reading their opponent

The results are impressive. It beat the world's strongest Stratego player by a record 15 wins, 1 loss and 4 draws, and went 39-2 against top players at the Stratego world championship. It also reached superhuman level in other hidden information games, including Barrage Stratego, Hanabi and the Chinese card game Dou dizhu. That is a dominant record by any standard

The efficiency is just as notable. It used less than one hundredth of the training examples of DeepMind's DeepNash and less than one thirtieth of the self play games, at a much lower computing cost. Senior author Gabriele Farina said it is good at calculating risk in a way humans are not, adding that a human might panic when their most valuable piece is exposed, but the bot stays composed. Calm under pressure is a fair description of most good poker players too

I love stories like this because they show AI progress in a way that is easy to understand. Does anyone here play Stratego? And would you want to play against something that never panics?

This game-playing AI is the new champ at Strategonews.mit.edu

[Driver]

Hanabi is a fascinating inclusion because it is cooperative. You have to work with partners while only seeing their cards, not your own. Doing well there means understanding what teammates are trying to tell you. That is closer to real world collaboration than most games

[DigitalNomad62]

Reasoning under uncertainty is the skill that transfers to real life. Negotiations, security and planning all involve guessing what the other side is hiding. That is why game research keeps getting funded

[BackpropMonk33]

Never panicking is a huge advantage in games of hidden information. Humans overreact to scary moments and make mistakes. An opponent with no fear just keeps making the best bet. Poker players will recognise that. Nerves cost more money than bad cards

[Dave_37]

Using a hundredth of the training data DeepNash needed is the real achievement. Making AI more efficient is more useful than throwing more compute at problems

It also makes this kind of research possible for university teams, not just giant labs. That is good for the field

[Kev96]

Fifteen wins to one against the world's best is crushing. Not even close. Games like this were thought to be safe from AI for a lot longer

The world champion must have found it humbling. Still, a draw or two is something

[Coder65]

Played Stratego endlessly as a kid. The bluffing was half the fun, moving a weak piece like it was your Marshal

Hard to imagine an AI that sees through all that. My brother still claims he never lost. He definitely did