A0
Science Communication DeepMind 2017

Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm

David Silver*, Thomas Hubert*, Julian Schrittwieser*, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, Timothy Lillicrap, Karen Simonyan, Demis Hassabis

"AlphaZero generalizes the tabula rasa reinforcement learning approach of AlphaGo Zero into a single, generic algorithm. Starting from absolute zero human knowledge except basic game rules, it achieved superhuman mastery in Chess, Shogi, and Go within 24 hours, convincingly defeating world-champion engines."

Tabula Rasa Learning Monte-Carlo Tree Search (MCTS) Dual-Policy-Value Networks General Game-Playing AI

01. Why This Matters

For decades, computer chess engines represented the pinnacle of artificial intelligence. These systems, such as the legendary Deep Blue or the modern open-source champion Stockfish, rely on highly optimized search algorithms paired with handcrafted evaluation functions. These functions utilize hundreds of arbitrary rules and weights tuned by human Grandmasters over generations.

While spectacularly strong, these traditional engines suffer from a fatal flaw: extreme domain specificity. A chess engine cannot play Shogi; a Shogi engine cannot play Go. They are brittle structures of human engineering. AlphaZero shattered this paradigm by demonstrating that a single, general-purpose reinforcement learning algorithm could master all three highly complex board games from scratch, without human intervention or pre-calculated opening books.

02. The Big Idea

The core breakthrough of AlphaZero is replacing human-designed heuristics with a single, deep neural network $f_\theta(s)$ that outputs both move probabilities (the policy) and a position evaluation (the value). Instead of evaluating millions of moves using brute-force search trees, AlphaZero uses a highly selective Monte-Carlo Tree Search (MCTS) guided by this neural network.

"Instead of a handcrafted evaluation function and move ordering heuristics, AlphaZero utilizes a deep neural network $(p, v) = f_\theta(s)$... trained solely by reinforcement learning from games of self-play."

By playing millions of games against itself, the algorithm continually updates its network parameters $\theta$ to match the actual outcomes of games and the search distributions of its MCTS. It learns the game's intrinsic geometric structure natively.

03. How It Works

1. Input Representation & Action Planes

AlphaZero represents the board state as a spatial tensor of size $N \times N \times (MT + L)$. For Chess, this is an $8 \times 8$ grid with $119$ planes representing piece positions over a history of $T=8$ steps, alongside scalar parameters like castling rights and move counts.

Interactive Widget: Input Plane Stack Explorer

Select a game to visualize how the board state is structured into a 3D tensor before being fed into AlphaZero's deep convolutional network.

Chess Input Stack

Total of 119 feature planes representing an 8-step history of the game board.

2. Search Strategy: Selective MCTS vs. Alpha-Beta

Traditional engines use Alpha-Beta minimax search, evaluating up to 70 million positions per second (Stockfish). AlphaZero evaluates only 80,000 positions per second, but uses its policy network to focus search efforts purely on high-probability variations. It acts more "human-like", prioritizing quality over quantity.

Interactive Widget: Search Tree Topology

Compare the search strategies. Toggle between Alpha-Beta (dense, wide, uniform depth) and AlphaZero's MCTS (highly focused, asymmetric, deep down promising paths).

Showing typical tree expansion. Green nodes represent active evaluations; greyed nodes are pruned or ignored.

3. Self-Play & Unified Loss Function

The parameters $\theta$ of AlphaZero's neural network are updated continually through gradient descent on a joint loss function $l$. This loss combines value error (mean-squared error) and policy error (cross-entropy), regularized by $L_2$ weight decay:

$$l = (z - v)^2 - \boldsymbol{\pi}^T \log \mathbf{p} + c||\theta||^2$$

Interactive Widget: Loss Function Simulator

Adjust the sliders to see how discrepancies between search probabilities $\pi$ and network predictions $p$, or game outcome $z$ and value prediction $v$, impact the total loss.

Value Prediction Error ($|z - v|$) 0.50

Difference between predicted value $v \in [-1, 1]$ and actual game outcome $z \in \{-1, 0, 1\}$.

Policy Cross-Entropy Divergence 0.80

Measures how well the network's move probabilities $p$ align with the MCTS search distributions $\pi$.

Calculated Total Loss
1.050
Formula: $\text{Loss} = \text{MSE} + \text{Cross-Entropy} + 0.01 \text{ (Regularization)}$

04. Key Results

AlphaZero achieved superhuman performance with astonishing speed. In head-to-head match tournaments of 100 games against the world champions, AlphaZero did not lose a single game as White to Stockfish in Chess, and dominated Shogi's Elmo.

Tournament Performance (100-Game Matches)

Chess: AlphaZero vs Stockfish 28 Wins / 72 Draws / 0 Losses
28% Win
72% Draw
0% Loss
Shogi: AlphaZero vs Elmo 90 Wins / 2 Draws / 8 Losses
90% Win
2%
8% Loss
Go: AlphaZero vs AlphaGo Zero (3-Day) 60 Wins / 40 Losses
60% Win
40% Loss
Note: Tournament matches were played at tournament time controls of 1 minute per move. Stockfish and Elmo played at their absolute strongest skill level using 64 threads and a hash size of 1GB.

05. Limitations & Open Questions

While AlphaZero represents an unprecedented leap in general artificial intelligence, several critical limitations and open questions remain:

  • Enormous Compute Cost: Training required massive computational infrastructure (5,000 first-generation TPUs to generate self-play games and 64 second-generation TPUs to train the networks over several hours). This is far beyond the reach of standard consumer hardware.
  • Asymmetry and Symmetries: Unlike Go, which features rotational and reflectional symmetries, Chess and Shogi are highly asymmetric. AlphaZero cannot use data augmentation techniques, which makes learning inherently slower per step.
  • The Draw Problem: In high-level chess, the prevalence of draws compresses the Elo scale, making it difficult to measure absolute progress near perfect play.

06. Glossary

Tabula Rasa

Latin for "blank slate." In machine learning, it refers to training an agent from scratch without any pre-existing human knowledge, guidelines, or database of expert games.

Monte-Carlo Tree Search (MCTS)

A search algorithm used for decision-making processes, notably in game-playing AI. It builds a search tree by simulating games step-by-step, evaluating nodes based on statistical outcomes of random or guided rollouts.

Alpha-Beta Pruning

An optimization technique for the minimax search algorithm that reduces the number of nodes evaluated in the search tree by ignoring branches that are mathematically proven to be worse than previously evaluated moves.

07. Citation

@article{silver2017mastering,
  title={Mastering chess and shogi by self-play with a general reinforcement learning algorithm},
  author={Silver, David and Hubert, Thomas and Schrittwieser, Julian and Antonoglou, Ioannis and Lai, Matthew and Guez, Arthur and Lanctot, Marc and Sifre, Laurent and Kumaran, Dharshan and Graepel, Thore and others},
  journal={arXiv preprint arXiv:1712.01815},
  year={2017}
}
Made with Flash Papers — make your own