Hidden information
Player-specific 128-element observations exclude opponent hands, deck order, and the face-down offer. Illegal actions are masked across 74 policy outputs.
Open-source C++17 project
Agent Avenue AI
A self-play reinforcement-learning implementation for the hidden-information board game Agent Avenue. The repository includes the simulator, training pipeline, evaluation tools, checkpoints, and a Windows client.
17.3Mself-play games
85.28%win rate vs random
79.76%win rate vs heuristic
15,000reported evaluation games
A complete engineering loop
The project is not an API wrapper or isolated model notebook. Training, evaluation, automated tests, and the desktop app share the same game state machine.
Game simulator
Parallel self-play
Actor-critic training
Policy + value checkpoint
Evaluation + human vs AI
Reliable training
Deterministic CPU rollouts, optional CUDA/NVRTC updates, atomic saves, checksum validation, backup recovery, and exact training-state resume.
Reproducible evaluation
The frozen policy plays 5,000 games per opponent, balanced across first and second player, with methods and limitations documented publicly.
Playable result
Test the trained behavior yourself.
The release includes the 17M-game checkpoint and a native Windows client. Play in Chinese, English, or Spanish; inspect public cards; switch first player; and review why the game ended.
