A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Game Theory

Background:

• Chess has been a long-standing benchmark for artificial intelligence (AI) research.
• Current AI chess programs rely on human expertise and handcrafted evaluation functions.
• Recent advances in reinforcement learning (RL) have shown promise in complex games like Go.

Purpose of the Study:

• To develop a generalized AI algorithm capable of achieving superhuman performance across multiple challenging games.
• To demonstrate the efficacy of reinforcement learning from self-play without domain-specific knowledge.

Main Methods:

• The study introduces the AlphaZero algorithm, a unified approach generalizing the AlphaGo Zero methodology.
• AlphaZero utilizes deep neural networks and Monte Carlo tree search, trained via reinforcement learning from self-play.
• The algorithm is provided only with the rules of the game, starting from random play.

Main Results:

• AlphaZero achieved superhuman performance in chess, shogi, and Go.
• The algorithm convincingly defeated world-champion level AI programs in these games.
• This demonstrates a significant advancement in generalized game-playing AI.

Conclusions:

• Reinforcement learning from self-play is a powerful paradigm for developing superhuman AI in complex strategic domains.
• AlphaZero represents a significant step towards more general artificial intelligence systems.
• The approach removes the need for extensive domain-specific engineering in AI game development.