Acquisition of chess knowledge in AlphaZero

2021/11/17 by Thomas McGrath, Andrei Kapishnikov, Nenad Tomašev +6 · 5 voices · 13 citations
Computer Science · Economics, Econometrics and Finance · Mathematics · #Artificial Intelligence in Games #Explainable Artificial Intelligence (XAI) #Sports Analytics and Performance #cs.AI #stat.ML

paper · pdf · doi:10.1073/pnas.2206625119

openalex publication_date 2022/11/14 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

We analyze the knowledge acquired by AlphaZero, a neural network engine that learns chess solely by playing against itself yet becomes capable of outperforming human chess players. Although the system trains without access to human games or guidance, it appears to learn concepts analogous to those used by human chess players. We provide two lines of evidence. Linear probes applied to AlphaZero's internal state enable us to quantify when and where such concepts are represented in the network. We also describe a behavioral analysis of opening play, including qualitative commentary by a former world chess champion.

Citations

• Towards A Rigorous Science of Interpretable Machine Learning
• Axiomatic Attribution for Deep Networks
• A Unified Approach to Interpreting Model Predictions
• Hybrid Reward Architecture for Reinforcement Learning
• The (Un)reliability of saliency methods
• Relational Deep Reinforcement Learning
• On the Robustness of Interpretability Methods
• Towards Robust Interpretability with Self-Explaining Neural Networks
• Sanity Checks for Saliency Maps
• Deep Blue
• Explaining Classifiers with Causal Concept Effect (CaCE)
• DisCoRL: Continual Reinforcement Learning via Policy Distillation
• On a Method to Measure Supervised Multiclass Model’s Interpretability: Application to Degradation Diagnosis (Short Paper)
• SentiMATE: Learning to play Chess through Natural Language Processing
• A Theory of Usable Information Under Computational Constraints
• Explain Your Move: Understanding Agent Actions Using Specific and Relevant Feature Attribution
• Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess
• Towards falsifiable interpretability research
• Debiasing Concept-based Explanations with Causal Analysis

Cited by

• MindCraft: How Concept Trees Take Shape In Deep Models
• Concept activation vectors: a unifying view and adversarial attacks
• Model Science: getting serious about verification, explanation and control of AI systems
• Learning to Imitate with Less: Efficient Individual Behavior Modeling in Chess
• Spectral Principal Paths: A Spectral Perspective on Linear Representation Formation in LLMs
• Toward the Explainability of Protein Language Models
• Representation Consistency for Accurate and Coherent LLM Answer Aggregation
• Hidden Breakthroughs in Language Model Training
• Bridging the human–AI knowledge gap through concept discovery and transfer in AlphaZero
• Path Channels and Plan Extension Kernels: a Mechanistic Description of Planning in a Sokoban RNN
• Chess variation entropy and engine relevance for humans
• Understanding the learned look-ahead behavior of chess neural networks
• Partner Modelling Emerges in Recurrent Agents (But Only When It Matters)
• Interactive Explanations for Reinforcement-Learning Agents

Discussions

• Acquisition of Chess Knowledge in AlphaZero [hn, 4 points, 0 comments]
• Acquisition of Chess Knowledge in AlphaZero [hn, 2 points, 0 comments]
• Acquisition of Chess Knowledge in AlphaZero [pdf] [hn, 2 points, 0 comments]
• Acquisition of Chess Knowledge in AlphaZero (Deepmind 2021) [hn, 2 points, 0 comments]
• Acquisition of Chess Knowledge in AlphaZero [hn, 2 points, 1 comments]

Related

• Bridging the Human-AI Knowledge Gap: Concept Discovery and Transfer in AlphaZero
• Aligning Superhuman AI with Human Behavior: Chess as a Model System
• Bridging the human–AI knowledge gap through concept discovery and transfer in AlphaZero
• A Behavior-Based Knowledge Representation Improves Prediction of Players' Moves in Chess by 25%
• Neural Networks for Chess