Trained transformer-based chess models to play like humans (including thinking time) [P]

2026-05-13 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Advanced, short

Summary

A set of transformer-based deep learning models has been trained to emulate human chess play, including decision-making time. These models are segmented by 100-point rating buckets, ranging from approximately 800 to 2500+ ELO. Initial training for a mid-strength model utilized an 8xH100 cluster, followed by fine-tuning for other rating ranges on a local 5090 GPU. The training dataset comprised nearly a year of Lichess data, totaling about 1 billion games. Each rating range includes three distinct models: a move model, a thinking time model, and a win/draw/loss prediction model. Despite their small size (9 million parameters), the move models achieve accuracy comparable to MAIA-3. The models also incorporate player ratings and clock times, influencing blunder rates under time pressure and win probabilities.

Key takeaway

For machine learning engineers developing human-like AI agents, consider segmenting your model by skill level and incorporating auxiliary models for non-action behaviors like thinking time. Your training pipeline should prioritize efficient data loading, such as pre-shuffling datasets and sequential reading, to prevent I/O bottlenecks and maximize GPU utilization, especially with large datasets like 1 billion games.

Key insights

Transformer models can emulate human chess play and thinking times using large-scale Lichess data.

Principles

• Small models can achieve high accuracy.
• Player ratings and clock times influence game outcomes.

Method

Train separate transformer models for move prediction, thinking time, and win probability, conditioned on player ratings and clock times, using a pre-shuffled, sequentially read dataset to maximize GPU utilization.

In practice

• Pre-shuffle large datasets to optimize GPU I/O.
• Condition models on dynamic game factors like clock time.

Topics

• Transformer Models
• Chess AI
• Human-like Play
• Thinking Time Modeling
• Lichess Dataset

Code references

• thomasj02/1e4_ai

Best for: AI Scientist, Machine Learning Engineer, AI Student

Related on AIssential

• [P] Vibecoded on a home PC: building a ~2700 Elo browser-playable neural chess engine with a Karpathy-inspired AI-assisted research loop — An AI-assisted research loop can rapidly develop high-performance neural chess engines on consumer hardware.
• Using Cognitive Models to Improve Language Model Simulation of Human Persuasion Games — Cognitive models and RL can guide LLMs to simulate diverse human decision-making, improving training environments.
• Building a Chess Coach — Anant Dole and Asbjorn Steinskog, Take Take Take — Combining classical chess engines with LLMs and human-like move prediction enhances AI chess coaching.
• Thinking into the Future: Latent Lookahead Training for Transformers — Latent lookahead training enables Transformers to "think" ahead by performing multi-step latent space predictions.
• PAWN: Piece Value Analysis with Neural Networks — Full board state context, via CNN autoencoders, significantly enhances chess piece value prediction accuracy.
• Assessing World Models In Machines | ARC Prize @ MIT — Games offer a robust benchmark for AI, revealing hierarchical abstraction and problem-making capabilities beyond mere problem-solving.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.