Paper, researchers, and primary source
Major lab research / agents
AlphaZero learns chess and shogi from self-play using one general reinforcement-learning and tree-search algorithm without human game data.
Contribution 1
Generalized self-play learning across multiple board games.
Contribution 2
Combined a neural policy-value function with Monte Carlo tree search.
Contribution 3
Removed handcrafted domain knowledge beyond game rules.
Research context
agents / 2017
Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm places alphazero inside the broader agents discussion at DeepMind, with self play supplying a second analytical lens. The paper's through-line contains three reported moves: Generalized self-play learning across multiple board games; Combined a neural policy-value function with Monte Carlo tree search; and Removed handcrafted domain knowledge beyond game rules. That sequence keeps tree search tied to the reported work instead of treating it as a free-standing promise. The practical stake is equally bounded: synthetic experience can produce strong policies in environments with exact rules and reliable outcome signals.
Methods and evidence reading
1 cataloged method notes
Method 1
The experimental design in Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm tests generalized self-play learning across multiple board games and combined a neural policy-value function with monte carlo tree search against the paper's documented baselines, datasets, model variants, or systems workloads.
How to read the evidence
Evidence for Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm is best read as a scoped argument rather than a universal verdict. The source record contains two linked moves: Generalized self-play learning across multiple board games; and Combined a neural policy-value function with Monte Carlo tree search. Its reported outcome is: Removed handcrafted domain knowledge beyond game rules. The defensible takeaway remains synthetic experience can produce strong policies in environments with exact rules and reliable outcome signals. That conclusion must travel with the recorded boundary that transfer from Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm must retain or retest compute budget, architecture choices, task distribution, documented data, comparison baselines, and evaluation protocol, because its alphazero finding is bounded by the reported study. A replication should preserve the disclosed setup and test whether alphazero still holds when self play conditions change. For a follow-on study of Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm, pair alphazero measurements with self play slices and preserve negative examples around Combined a neural policy-value function with Monte Carlo tree search as first-class evidence.
Findings in the source record
1 paper-specific findings
- The reported evidence in Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm supports removed handcrafted domain knowledge beyond game rules.
Practical implication for AI builders
DeepMind / 2017
Proposed BrokenGPT application
Research blueprint / proposed status
Proposed route placement / alphazero
For a proposed BrokenGPT experiment based on Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm, build sandboxed self-play test environments for tool-selection policies where outcomes are automatically scored before any policy is considered for production. Keep the alphazero path isolated, versioned, and attributable to this research record.
Proposed acceptance test / self play
Validate the proposed alphazero route against the paper's reported outcome: Removed handcrafted domain knowledge beyond game rules. Measure recovery behavior, task completion, trace quality, and intervention rate for the Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm candidate, then isolate self play regressions before judging the proposed alphazero route.
Proposed decision boundary / tree search
Balance autonomy, compute, and controllability before promoting the proposed tree search design. Because before adapting alphazero, a new evaluation should expose another user population, a later model revision, changed operating conditions, new hardware, and a different product rather than assuming Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm already covers them, adoption remains conditional on replay under BrokenGPT's selected model, runtime, and policy configuration.
Limitations, verification, and source
Boundaries recorded with the paper
Limitations
- Transfer from Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm must retain or retest compute budget, architecture choices, task distribution, documented data, comparison baselines, and evaluation protocol, because its alphazero finding is bounded by the reported study.
- Before adapting alphazero, a new evaluation should expose another user population, a later model revision, changed operating conditions, new hardware, and a different product rather than assuming Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm already covers them.
PRIMARY SOURCES
- 01Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm
DeepMind — Primary primary arXiv paper / 5 December 2017 / David Silver, Thomas Hubert, Julian Schrittwieser, and 10 more