An AI coach for Happy DouDiZhu that reads your screen

What it does The project wraps the open-source DouZero card-playing model in a PyQt desktop tool that watches your 1920×1080 screen, recognizes the cards in the official Happy DouDiZhu Windows client, and tells you what to play next. You still click the cards yourself; the AI merely whispers suggestions after analyzing the board state through pixel-level screenshot regions.

The interesting bit Instead of building a fully autonomous bot, it acts as a human-in-the-loop advisor—likely a pragmatic dodge around anti-cheat, but also a neat way to see how a deep RL agent reasons in real time against human opponents. It borrows template-matching logic from another card-recognition project to parse the live game UI.

Key highlights

• Ships with three model flavors: a supervised-learning baseline, an ADP (average difference points) agent, and a WP (winning percentage) agent set as the default.
• Uses hardcoded screenshot coordinates inside MyPyQT_Form to grab hand history, visible cards, and role indicators from the maximized game window.
• Includes a pos_debug.py utility for recalibrating those screen regions when layouts shift.
• Recognizes cards via template images borrowed from the cardRecorder project.

Caveats

• Tightly coupled to 1920×1080 resolution and a maximized, bottom-right window position; any UI drift breaks recognition.
• Not fully automated: you must manually start each round, execute every suggested move yourself, and finish the hand if the tracker bugs out.
• The README warns that the “rocket” (both Jokers) animation is long enough to occasionally trick the recognizer into seeing a single Joker.

Verdict Worth a look if you want to watch a research-grade RL agent think through a commercial card game in real time. Skip it if you were hoping for a fire-and-forget bot or if your monitor isn’t 1080p.