ProjectAIResearchWIP
Slay the Spire solver | an RL agent for the hardest difficulty
An RL agent for the Watcher at Ascension 20, aiming past the best human win rate of about 90%. Worked on in bursts.
What I Built
- Game engine. A Python engine with full parity to the Java game: 13 RNG streams, 66 enemies, 51 events, 168 powers, 172 relics.
- Combat engine. A Rust combat engine with PyO3 bindings and a turn solver with multi-turn lookahead.
- Training. PPO with a transformer policy over a 260-dimension state and 512-dimension action encoding. A live React dashboard.
Stats
6,100passing tests across Python and Rust
3Mparameters in the policy transformer
30 msper combat-engine run
Floor 16reached, 200 saved trajectories
The Stack
EnginePython, Rust, PyO3
TrainingPyTorch, PPO
DashboardReact