Distributed reinforcement learning for Slay the Spire

Project Snapshot

The current public snapshot focuses on making the existing ML systems work legible: training infrastructure, deterministic evaluation, dashboarding, architecture documentation, and honest early results.

What To Review First

System In Motion

AscensionAI runs as a local distributed training loop: live game workers write checkpoint-tagged rollout files, the offline trainer consumes fresh batches, saves an updated checkpoint, and workers reload the policy for the next collection cycle.

Docs hub all project docs Demo assets public-safe media Portfolio page project framing Experiment registry machine-readable

Current Results

The latest PPO snapshot still improves over the BC checkpoint on the 150-game evaluation, but the wider sample shows it remains behind the heuristic baseline. The value here is the complete training system and evaluation pipeline, with clear metrics for future improvement.