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

This paper generalizes the self-play reinforcement learning approach of AlphaGo Zero into a single algorithm called AlphaZero that can master multiple challenging board games. Whereas the strongest traditional programs, especially in chess, depend on sophisticated search techniques, domain-specific adaptations, and handcrafted evaluation functions refined by human experts over decades, AlphaZero begins from random play and is given no domain knowledge beyond the rules of each game. It learns entirely through reinforcement learning from self-play.

Starting tabula rasa, AlphaZero achieved superhuman performance and convincingly defeated world-champion programs in chess and shogi as well as Go. Demonstrating that one algorithm could adapt to three different games without game-specific engineering, the work was a notable step toward a general game-playing system and showed the power of self-play reinforcement learning beyond a single domain.