DeepMind trains AI to play Stratego at human-expert level

Researchers from DeepMind’s lab have created the AI agent DeepNash, capable of playing Stratego at a human-expert level. This is reported by Gizmodo.

\n\n

DeepNash learned to play by playing numerous games against itself. In the process, it was able to make complex decisions and consider compromises in ‘unconventional’ ways that were unavailable to previous AI systems.

\n\n

Researchers said that Stratego’s combination of long-term decision-making and the flow of imperfect information makes it a unique proving ground for AI.

\n\n

The game typically involves two players. It combines strategy with elements of deception. Each player has a set of pieces forming an “army,” each with its own value. Victory is achieved by capturing the opponent’s flag or by the opponent having no moves left.

\n\n

The presence of pieces with different values leads to an extremely large number of possible moves and outcomes. Researchers say that Stratego has far more ‘possible states’ than Texas Hold’em or Go.

\n\n

To win, DeepNash blended long-term strategy with short-term decision-making, such as bluffing and risk-taking. Historically, prior algorithms were not able to do this simultaneously.

\n\n

“DeepNash managed to find a nontrivial trade-off between information and material, bluffing and taking risks when necessary,” the researchers said.

\n\n

\n\n

Likely, the creators of DeepNash were inspired by American mathematician John Nash, who proposed the Nash equilibrium. In game theory, it is the set of actions for two or more players in which each participant has no incentive to change their strategy given the others’ choices.

\n\n

DeepNash seeks to find the Nash equilibrium in Stratego, using a combination of self-play and reinforcement learning called R-NaD. Employing both this algorithm and the deep neural network architecture, the researchers were able to create a model that outperformed opponents even in extremely challenging situations.

\n\n

The developers tested DeepNash by comparing it with other bots and ‘top players’ on the Gravon online platform. The AI agent defeated virtual opponents in 97% of cases. The win rate against humans stood at 84%.

\n\n

As a result, the AI ranked among the top three players this year as well as on the all-time leaderboard.

\n\n

“To our knowledge, this is the first time an AI algorithm has learned to play Stratego at a human-expert level,” the researchers said.

\n\n

In November, DeepMind created an artificial intelligence that interacts naturally with humans and learns from them.

\n\n

In the same month, Meta created an AI agent that plays Diplomacy at human level.

\n\n

Subscribe to ForkLog’s news on Telegram: ForkLog AI — all the news from the world of AI!