PureTD: Reinforcement Learning for Backgammon Money Games with No Evaluation-time Search

Name: PureTD: Reinforcement Learning for Backgammon Money Games with No Evaluation-time Search
Uploaded: 2026-10-17T07:41:25.062238+00:00
Duration: 241 s
Description: This work introduces PureTD, a reinforcement learning method for backgammon money games that learns both checker play and cube actions from scratch without any search time. The model uses pure self-play training to achieve near state-of-the-art strength, proving that simple RL can outperform engines using limited lookahead searches.

This work introduces PureTD, a reinforcement learning method for backgammon money games that learns both checker play and cube actions from scratch without any search time. The model uses pure self-play training to achieve near state-of-the-art strength, proving that simple RL can outperform engines using limited lookahead searches.

Open the full episode