A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play
Level 5 - mechanism / opinion, no new human data
Level 5 by design analogy, not clinical CEBM: computational benchmark experiment.
OpenAlex W2902907165 · doi:10.1126/science.aar6404
What was done
Researchers developed AlphaZero, a general reinforcement learning algorithm adapted from AlphaGo Zero, designed to learn two-player board games without human data or domain-specific heuristics beyond basic game rules. Starting from purely random play, the algorithm was evaluated against top programs in chess, shogi, and Go.
What was found
The abstract provides no quantitative win-loss statistics or exact performance metrics. It reports that AlphaZero convincingly defeated world-champion programs in chess, shogi, and Go starting from scratch without handcrafted evaluation functions.
Why it matters
It showed that a single generic reinforcement learning architecture can master multiple complex, highly distinct board games tabula rasa, replacing decades of hand-crafted domain-specific search heuristics.
Limits
The abstract reports no numerical match scores, compute requirements, or training durations. The approach was tested only on deterministic, discrete, perfect-information board games with fully known rules.
Cited by
- supports DeepMind's AlphaGo Zero utilized Sutton and Barto's reinforcement learning algorithm to train entirely from scratch without expert human gameplay input.