Acquisition of chess knowledge in AlphaZero

By Thomas McGrath, Andrei Kapishnikov, Nenad Tomašev, Adam Pearce, Demis Hassabis, Been Kim, Ulrich Paquet, Vladimir Kramnik. 2021

204 citations recorded by the source corpus.

Significance Seventy years ago, Alan Turing conjectured that a chess-playing machine could be built that would self-learn and continuously profit from its own experience. The AlphaZero system—a neural network-powered reinforcement learner—passed this milestone. In this paper, we ask the following questions. How did it do it? What did it learn from its experience, and how did it encode it? Did it learn anything like a human understanding of chess, in spite of having never seen a human game? Remarkably, we find many strong correspondences between human concepts and AlphaZero’s representations that emerge during training, even though none of these concepts were initially present in the network.

Subfields: Artificial Intelligence, Economics and Econometrics, Reinforcement Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Economics and Econometrics

In Reinforcement Learning

Citations in the corpus

Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling (2023, background): “Of the research along these lines that does exist (McGrath et al., 2021; Tirumala et al., 2022; Xia et al., 2022), it is overwhelmingly done on non-public models or model checkpoints, further emphasizing the importance of having publicly available model suites for scientific research.”
The alignment problem from a deep learning perspective (2022, background): “…policies learned, one example of a model-based policy learning robust outcome representations comes from AlphaZero, which learned a range of human chess concepts, including concepts used in top chess engine Stockﬁsh’s hand-crafted evaluation function (e.g. “king safety”) [McGrath et al., 2021].”
On scientific understanding with artificial intelligence (2022, background): “I AlphaZero has learned about chess and how human-like knowledge is encoded in the internal representation [94].”
Symbols as a Lingua Franca for Bridging Human-AI Chasm for Explainable and Advisable AI Systems (2021, unknown): “An interesting ﬁrst step taken in this direction is the analysis performed by McGrath et al. (2021), in analyzing the concepts learned by AlphaZero.”
Post-hoc Concept Bottleneck Models (2022, background): “Related work focuses on understanding if neural networks encode and use concepts (Lucieri et al., 2020; Kim et al., 2018; McGrath et al., 2021), or generate counterfactual explanations to understand model behavior (Ghandeharioun et al., 2021; Abid et al., 2022; Akula et al., 2020).”
Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task (2022, background): “Even one of the most general-purpose game-playing engines, AlphaZero (Silver et al., 2018), has built-in knowledge of basic board structure and game rules (although, intriguingly, it seems to develop interpretable models of various strategic concepts (McGrath et al., 2021; Forde et al., 2022)).”
Progress measures for grokking via mechanistic interpretability (2023, background): “McGrath et al. (2021) find that AlphaZero quickly learns many human chess concepts between 10k and 30k training steps and reinvents human opening theory between 25k and 60k training steps.”
Mechanistic Interpretability for AI Safety - A Review (2024, methodology): “…al., 2020), and causal ablation (Wang et al., 2023), share the common goal of modifying a neural model’s internal state by replacing specific activations with alternative values, such as zeros, mean activations across samples, random noise, or activations from a different forward pass (Figure 9a).”