Grandmaster-Level Chess Without Search

[pavs]

From Google Deepmind: https://github.com/google-deepmind/searchless_chess

This repository provides an implementation of our paper Grandmaster-Level Chess Without Search.

The recent breakthrough successes in machine learning are mainly attributed to scale: namely large-scale attention-based architectures and datasets of unprecedented scale. This paper investigates the impact of training at scale for chess. Unlike traditional chess engines that rely on complex heuristics, explicit search, or a combination of both, we train a 270M parameter transformer model with supervised learning on a dataset of 10 million chess games. We annotate each board in the dataset with action-values provided by the powerful Stockfish 16 engine, leading to roughly 15 billion data points. Our largest model reaches a Lichess blitz Elo of 2895 against humans, and successfully solves a series of challenging chess puzzles, without any domain-specific tweaks or explicit search algorithms. We also show that our model outperforms AlphaZero's policy and value networks (without MCTS) and GPT-3.5-turbo-instruct. A systematic investigation of model and dataset size shows that strong chess performance only arises at sufficient scale. To validate our results, we perform an extensive series of ablations of design choices and hyperparameters.

[towforce]

This has been there 6 months - why has nobody mentioned this before?

[smatovic]

The paper was posted:

Grandmaster-Level Chess Without Search. Large-scale attention-based architectures and datasets of unprecedented scale.
viewtopic.php?p=958355#p958355

--
Srdja

[Viz]

Because leela nets did this long time before google and with smaller nets (and also having a better playing strength).
Mind you that leela also switched to transformer nets almost 2 years ago.

[Rebel]

You can make a reasonable opening book using Leela at depth=1.

Did this 2-3 years ago, long before the article.

[chrisw]

Arguably, the sorts of tactically complex positions with lots of material on board that no-search-nets might tactically fail on, are quite likely to be the sorts of positions that GMs could tactically fail on also.

[chrisw]

You can make a reasonable opening book using Leela at depth=1.

Did this 2-3 years ago, long before the article.

Yes but. If you have such an opening book with 1% blunders in it, it will blunder 1% of games. If you used the same net to give node evals for an AB or minimax search, you can guarantee the search will track down and play the erroneous blunder positions at well above 1% rate. Search finds bugs.

[towforce]

The paper was posted:

Grandmaster-Level Chess Without Search. Large-scale attention-based architectures and datasets of unprecedented scale.
viewtopic.php?p=958355#p958355

--
Srdja

Thanks - very useful reminder, and a very good thread in which I learned some valuable lessons: Google will have had important reasons for doing this work - but LC0's NN:

1. Got to GM at ply 1 first
2. Is smaller
3. Is better

[towforce]

People have been looking to search and to tablebases to solve chess.

Wrong place: the solution will come from knowledge. Human skill will as well: later this century, using the new knowledge and the new techniques for teaching it, the children will go from beginner to well into the 2000s elo level in just a few months.

[BrendanJNorman]

People have been looking to search and to tablebases to solve chess.

Wrong place: the solution will come from knowledge. Human skill will as well: later this century, using the new knowledge and the new techniques for teaching it, the children will go from beginner to well into the 2000s elo level in just a few months.

Agree.

The same principle applies when a master gives a simul to amateurs.

He is not outcalculating them as he tears all 24 boards to pieces, despite his opponent's vastly larger thinking time.

It is his vast mental storehouse of knowledge, of tactical patterns, or structures etc that make him superior.