DI-sheep

Why we included this project

DI-sheep wraps deep reinforcement learning around a recognizable real game, the tile-matching puzzle 羊了个羊 ('3 Tiles'), and lays out a complete working pipeline. Rather than another toy CartPole demo, you get a Gym-style environment, an Actor-Critic model in PyTorch, and a DI-engine script that trains a PPO agent on the puzzle. Because the game is familiar, it is much easier to follow how the agent observes the board, picks moves, and improves from reward than it is with abstract textbook examples. A React front end with a Flask backend lets you watch a trained agent play in the browser or plug in your own model. The standalone environment file lifts out cleanly for adapting to similar matching puzzles, and the pretrained weights are still rough, which is handy if you want to see how tuning changes agent performance rather than judging a polished product。