With most information hidden, the game Stratego had stumped AI—until now

Summary

AI has finally conquered Stratego, a game previously resistant to AI due to its hidden information. Researchers developed Ataraxos, an AI that beat the best human player by incorporating a belief model to guess hidden pieces and employing a novel training strategy.

Key Takeaways

1. Ataraxos, an AI developed by researchers from Carnegie Mellon, MIT, NYU, and Stanford, defeated the best Stratego player, Pim Niemeijer, 15 games to one.
2. The AI's success was attributed to a second neural network, a belief model, trained to guess the opponent's hidden pieces based on their movements.
3. Ataraxos was trained efficiently on 16 GPUs for a week, costing a few thousand dollars, a significant cost reduction compared to previous AI attempts.

Statement Breakdown

• Claimed Facts: 70% of statements the article presents as facts
• Opinions: 20% of statements classified as editorial or subjective
• Claims: 10% of statements surfaced for additional reader evaluation

Credibility & Bias Reasoning

Credibility assessment: The article cites research from multiple universities and a peer-reviewed journal, providing specific details about the AI's performance and training. It contrasts the new AI with previous attempts, offering a balanced perspective on the challenges of Stratego.

Bias assessment: Technological Optimism. The article highlights the success of AI in complex games, framing it as a significant advancement. While objective in reporting the facts, there's an underlying positive tone regarding AI's capabilities and future applications.

Note: This article provides a detailed account of AI advancements in game theory. While factual, consider the inherent optimism in reporting technological breakthroughs.

Credibility flag: Informative, Tech-Focused

Claimed Facts (8)

• This provides historical context for AI achievements in games.
• This states the primary outcome of the AI's performance against a human champion.
• This describes the basic components of the game of Stratego.
• This states the objective of the game.
• This explains a fundamental rule of Stratego regarding hidden information.
• This details the mechanism by which piece identities are revealed.
• This describes the training methodology used for the AI.
• This indicates the versatility of the AI's underlying architecture.

Opinions (8)

• This is an interpretation of the game's characteristics and its comparison to poker.
• This is an explanation of why previous AIs failed, based on the team's analysis.
• This is an explanation of the origin and meaning of the AI's name.
• This is a speculative interpretation of the AI's behavior and its advantages over humans.
• This is an interpretation of how the AI's capabilities translate to game strategy.
• This is a subjective assessment of the AI's impact on the game's competitive landscape.
• This expresses a practical limitation faced by academic researchers.
• This indicates the future aspirations and potential applications of the AI technology.

Claims (5)

• This statement implies a definitive failure of DeepMind without providing specific details or context on their attempts for Stratego.
• This is a defensive claim by the researchers to downplay a loss without providing independent verification.
• Attributing a win solely to luck is speculative and subjective.
• The use of 'bluff' for an AI is anthropomorphic and the description of 'very, very casually' is subjective and lacks quantifiable evidence.
• This is a statement about future goals and current limitations, implying a subjective assessment of 'explainable' AI.

Key Sources

• Jacek Krywko — Author
• Eugene Vinitsky — Researcher at NYU
• Gabriele Farina — MIT computer scientist
• Pim Niemeijer — Stratego Player
• Samuel Sokota — Lead Author
• Carnegie Mellon — University
• MIT — University
• New York University — University
• Stanford University — University
• DeepMind — AI Research Lab
• Google — Technology Company