Offline RL on 5TB of Human Play Learns to Recommend Better Chess Puzzles

allenainie · x · 2026-08-18

The authors trained an offline RL policy on nearly 5TB of human behavior data (a full year of play) to recommend chess puzzles with high pedagogical value. After many challenges, they found a recipe that worked, and blind evaluation by International Masters and Grandmasters showed the policy recommends puzzles players find more fun and more challenging.

Related event: Offline RL Mines 5TB of Chess Games to Find High-Value Puzzles(2 posts)→

Original post →

More from Research

Research channel →