Acquisition of Chess Knowledge in AlphaZero

TL;DR

This paper investigates how AlphaZero acquires and represents human chess concepts during training, revealing that its neural network develops interpretable knowledge similar to human understanding, which aids interpretability.

Contribution

It provides evidence that AlphaZero learns human-like chess concepts, analyzes when and where these are represented, and offers behavioral and low-level representational insights.

Findings

AlphaZero encodes human chess concepts during training

Grandmaster Kramnik's qualitative analysis supports interpretability

Representational analysis shows alignment with human knowledge

Abstract

What is learned by sophisticated neural network agents such as AlphaZero? This question is of both scientific and practical interest. If the representations of strong neural networks bear no resemblance to human concepts, our ability to understand faithful explanations of their decisions will be restricted, ultimately limiting what we can achieve with neural network interpretability. In this work we provide evidence that human knowledge is acquired by the AlphaZero neural network as it trains on the game of chess. By probing for a broad range of human chess concepts we show when and where these concepts are represented in the AlphaZero network. We also provide a behavioural analysis focusing on opening play, including qualitative analysis from chess Grandmaster Vladimir Kramnik. Finally, we carry out a preliminary investigation looking at the low-level details of AlphaZero's representations, and make the resulting behavioural and representational analyses available online.