doudizhu-baseline

A multithreaded C++ Monte Carlo Tree Search player for Dou Di Zhu (斗地主), used as the search-based baseline and opponent in the related RL projects — doudizhu-C (the mcsearch import) and the doudizhu-tornado web game.

How It Works

Dou Di Zhu is imperfect-information, so plain MCTS doesn't apply directly. mcsearch uses determinized MCTS:

1. Determinization — the unseen cards are shuffled and dealt to the two hidden opponents (max_d random deals).
2. Parallel UCT — for each deal, a tree search runs max_iter iterations across n_threads threads on a shared, mutex-protected tree (UCB with c = √2), with uniformly random rollouts to the end of the game. Rewards are team-aware: +1 if the searching seat's side (landlord vs. farmers) wins, −1 otherwise.
3. Vote — root visit counts are summed over all determinizations and the most-visited action is played.

Supporting code:

• card.cpp / player.cpp — card/category model, enumeration of all legal card groups from a hand, and the follow/beat rules.
• env.cpp — a self-contained 3-player game environment whose seats are all MCTS players (step_auto plays one turn).
• minimax.cpp — an experimental exact alpha-beta solver with hashed states for two-player endgames.
• main.cpp — a self-play benchmark: plays N games, logs each deal to cards.bin, and reports landlord vs. farmer win counts. CLI: doudizhu [total_games] [n_threads] [max_d] [max_iter].

Building

Two targets:

• Python module (the way the other repos consume this): requires pybind11; cmake the doudizhu/ folder to build the mct extension exposing mcsearch, CEnv, CCardGroup, CCard, and CCategory.
• Standalone benchmark: open doudizhu.sln in Visual Studio (C++14).

from mct import mcsearch, CCard, CCardGroup, CCategory

intention = mcsearch(self_cards, unseen_cards, next_handcards_cnt,
                     last_cardgroup, current_idx, current_controller,
                     n_threads=10, max_d=50, max_iter=250)