Super human Stratego with RL and test time search (opens in new tab)

(arxiv.org)

2 pointsalgo_trader7mo ago1 comments

1 comments

1 comments · 1 top-level

Only 2000 GPU hours Heavily customized network 95% win rate in recent human tournament sample Several training techniques for evaluation/learning rate

j / k navigate · click thread line to collapse