Do neural networks learn to implement algorithms such as look-ahead or search
"in the wild"? Or do they rely purely on collections of simple heuristics? We
present evidence of learned look-ahead in the policy network of Leela Chess
Zero, the currently strongest neural chess engine. We find that Leela
internally represents future optimal moves and that these representations are
crucial for its final output in certain board states. Concretely, we exploit
the fact that Leela is a transformer that treats every chessboard square like a
token in language models, and give three lines of evidence (1) activations on
certain squares of future moves are unusually important causally; (2) we find
attention heads that move important information "forward and backward in time,"
e.g., from squares of future moves to squares of earlier ones; and (3) we train
a simple probe that can predict the optimal move 2 turns ahead with 92%
accuracy (in board states where Leela finds a single best line). These findings
are an existence proof of learned look-ahead in neural networks and might be a
step towards a better understanding of their capabilities.

神经网络在野外是否能够学习实现诸如预测或搜索等算法？还是纯粹依赖简单启发式规则的集合？我们通过 Leela Chess Zero 中感知网络的学到的预测来证明神经网络的学到的预测能力，并发现这些表示对其在特定局面下的最终输出至关重要。这些发现证明了神经网络学到的预测存在，并可能对我们对其能力的理解产生影响。

国际象棋博弈神经网络中已学习到的预测能力的证据

Evidence of Learned Look-Ahead in a Chess-Playing Neural Network

Past, Present and Future are considered to be three temporal and logical
concepts which are well defined by human beings for their existence and growth.
We, as human beings, have the privilege of using our intelligence to mentally
execute an activity before physical occurrence of the same in the real world.
Knowledge of the past, aplomb of present and visualisation for the future
correspond to three concepts such as look-back, look-at and look-ahead
respectively in real life as well as in diversified domains of computing.
Look-Ahead(LA) deals with the future prediction of information and processing
of input to produce the output in advance. In this article, our main objective
is to learn, understand and explore the concept of LA and design novel models
as solution for real world problems. We present three well known algorithmic
frameworks used in practice based on availability of input information such as
offline, online and semi-online. We introduce interesting real life
applications and well known computing problems where LA plays a significant
role for making a process, system or algorithm efficient. We define new types
of LA and propose a taxonomy for LA based on literature review for designing
novel LA models in future. Using the concept of LA, We identify and present
many interesting and non-trivial research challenges as future potential
research directions. Intuitively, we observe that LA can be used as a powerful
tool and framework for future researchers in design of efficient computational
models and algorithms for solving non-trivial and challenging optimization
problems.

通过对未来信息的预测和处理提前生成输出，Look-Ahead（LA）在实际中具有重要作用。本文介绍了三种实践中使用的算法框架，包括线下、线上和半在线方法，并提出了基于文献综述的 LA 分类。此外，本文还识别和提出了未来潜在研究方向中的许多有趣且非平凡的研究挑战。