ApogeeVGC
An AlphaZero-style AI that plays Pokémon doubles (VGC) live on the Showdown ladder.
- Two C++ battle engines, bit-for-bit equivalent
- Self-play training → ONNX → MCTS
- Plays live from partial information
ApogeeVGC trains a neural network by self-play against a fast, engine-accurate battle simulator, exports it to ONNX, and drives a Monte Carlo tree search that plays real games on the Pokémon Showdown ladder - where it only sees what a human player would.
How it fits together
- Two battle engines in C++. A reference engine is the correctness oracle; a fast engine built for self-play throughput must track it. A 2,000,000-draw equivalence gate keeps them bit-for-bit identical.
- Search and inference in C++. Decoupled-UCT tree search over simultaneous moves, with ONNX inference and a state encoder shared with training.
- Partial-information play. A live layer tracks a ladder game from Showdown’s protocol messages, reconstructs a full battle for the search, and fills in what it can’t see from a pool of real teams.
- A Python training pipeline. Self-play → training (value and policy heads) → ONNX export → an arena gate that only promotes a challenger that beats the champion.
The source is in a private repository.