Summary of Amortized Planning with Large-scale Transformers: a Case Study on Chess, by Anian Ruoss et al.
Amortized Planning with Large-Scale Transformers: A Case Study on Chess
by Anian Ruoss, Grégoire Delétang, Sourabh Medapati, Jordi Grau-Moya, Li Kevin Wenliang, Elliot Catt, John Reid, Cannada A. Lewis, Joel Veness, Tim Genewein
First submitted to arxiv on: 7 Feb 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (stat.ML)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper investigates the performance of transformers on planning tasks where memorization is futile by using chess as a landmark problem in AI. To assess their capabilities, the authors release ChessBench, a large-scale benchmark dataset containing 10 million chess games with legal move and value annotations. The authors then train transformers with up to 270 million parameters on this dataset via supervised learning and perform extensive ablations to understand the impact of various factors on performance. Notably, the largest models learn to predict action-values for novel boards accurately, demonstrating highly non-trivial generalization capabilities. When applied to chess puzzles, these policies achieve a strong Lichess blitz Elo rating of 2895 against human opponents (grandmaster level). The authors also compare their approach to Leela Chess Zero and AlphaZero, trained via self-play with and without search. While transformers can approximate Stockfish’s search-based algorithm through supervised learning, perfect distillation remains elusive, making ChessBench a valuable benchmark for future research. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper uses chess to test how well AI models called transformers do on planning tasks where memorization doesn’t help. They create a big dataset of 10 million chess games with information about moves and values. Then they train the transformers using this data and try different things to see what works best. The really good models can even predict what moves to make in new situations! When applied to tricky chess puzzles, these models do surprisingly well against human players. The authors also compare their approach to other AI systems that play chess differently. While the transformers are very good, they’re not perfect and still have room for improvement. |
Keywords
* Artificial intelligence * Distillation * Generalization * Supervised
