stable-retro
Training engine where the Laya decision model learns to beat retro games: text game state, group-relative policy gradients, per-state checkpoints and a reverse curriculum.
