riichienv 0.1.3

A High-Performance Research Environment for Riichi Mahjong

Accelerating Reproducible Mahjong Research

---

[!NOTE] This project is currently under active development. The API and specifications are subject to change before the stable release.

✨ Features

• High Performance: Core logic implemented in Rust for lightning-fast state transitions and rollouts.
• Gym-style API: Intuitive interface designed specifically for reinforcement learning.
• Mortal Compatibility: Seamlessly interface with the Mortal Bot using the standard MJAI protocol.
• Rule Flexibility: Support for diverse rule sets, including no-red-dragon variants and three-player mahjong.
• Game Visualization: Integrated replay viewer for Jupyter Notebooks.

📦 Installation

uv add riichienv
# Or
pip install riichienv

Currently, building from source requires the Rust toolchain.

uv sync --dev
uv run maturin develop --release

🚀 Usage

Gym-style API

from riichienv import RiichiEnv
from riichienv.agents import RandomAgent

agent = RandomAgent()
env = RiichiEnv()
obs_dict = env.reset()
while not env.done():
    actions = {player_id: agent.act(obs)
               for player_id, obs in obs_dict.items()}
    obs_dict = env.step(actions)

scores, points, ranks = env.scores(), env.points(), env.ranks()
print(scores, points, ranks)

env.reset() initializes the game state and returns the initial observations. The returned obs_dict maps each active player ID to their respective Observation object.

>>> from riichienv import RiichiEnv
>>> env = RiichiEnv()
>>> obs_dict = env.reset()
>>> obs_dict
{0: <riichienv._riichienv.Observation object at 0x7fae7e52b6e0>}

Use env.done() to check if the game has concluded.

>>> env.done()
False

By default, the environment runs a single round (kyoku). For game rules supporting sudden death or standard match formats like East-only or Half-round, the environment continues until the game-end conditions are met.

Observation

The Observation object provides all relevant information to a player, including the current game state and available legal actions.

obs.new_events() -> list[str] returns a list of new events since the last step, encoded as JSON strings in the MJAI protocol. The full history of events is accessible via obs.events.

>>> obs = obs_dict[0]
>>> obs.new_events()
['{"id":0,"type":"start_game"}', '{"bakaze":"E","dora_marker":"S", ...}', '{"actor":0,"pai":"6p","type":"tsumo"}']

obs.legal_actions() -> list[Action] provides the list of all valid moves the player can make.

>>> obs.legal_actions()
[Action(action_type=Discard, tile=Some(1), ...), ...]

If your agent communicates via the MJAI protocol, you can easily map an MJAI response to a valid Action object using obs.select_action_from_mjai().

>>> obs.select_action_from_mjai({"type":"dahai","pai":"1m","tsumogiri":False,"actor":0})
Action(action_type=Discard, tile=Some(1), consume_tiles=[])

Compatibility with Mortal

RiichiEnv is fully compatible with the Mortal MJAI bot processing flow. I have confirmed that MortalAgent can execute matches without errors in over 1,000,000+ hanchan games on RiichiEnv.

from riichienv import RiichiEnv, Action
from model import load_model

class MortalAgent:
    def __init__(self, player_id: int):
        self.player_id = player_id
        # Initialize your libriichi.mjai.Bot or equivalent
        self.model = load_model(player_id, "./mortal_v4.pth")

    def act(self, obs) -> Action:
        resp = None
        for event in obs.new_events():
            resp = self.model.react(event)

        action = obs.select_action_from_mjai(resp)
        assert action is not None, "Mortal must return a legal action"
        return action

env = RiichiEnv(game_mode="4p-red-half")
agents = {pid: MortalAgent(pid) for pid in range(4)}
obs_dict = env.reset()
while not env.done():
    actions = {pid: agents[pid].act(obs) for pid, obs in obs_dict.items()}
    obs_dict = env.step(actions)

print(env.scores(), env.points(), env.ranks())

Game Rules and Modes

RiichiEnv separates high-level game flow configuration (Mode) from detailed game mechanics (Rules).

• Game Mode (game_mode): Configuration for game length (e.g., East-only, Hanchan), player count, and termination conditions (e.g., Tobi/bust, sudden death).
• Game Rules (rule): Configuration for specific game mechanics (e.g., handling of Chankan (Robbing the Kan) for Kokushi Musou, Kuitan availability, etc.).

1. Game Mode Presets (game_mode)

You can select a standard game mode using the game_mode argument in the constructor. This configures the basic flow of the game.

game_mode	Players	Mode	Mechanics
4p-red-single	4	Single Round	No sudden death
4p-red-east	4	East-only (東風; Tonpuu)	Standard (Tenhou rule)
4p-red-half	4	Hanchan (半荘)	Standard (Tenhou rule)
3p-red-east	3	East-only (Tonpuu)	🚧 In progress

# Initialize a standard 4-player Hanchan game
env = RiichiEnv(game_mode="4p-red-half")

[!NOTE] We are also planning to implement "No-Red" rules (game modes without red 5 tiles), which are often adopted in professional leagues (e.g., M-League's team definitions or other competitive settings).

2. Customizing Game Rules (GameRule)

For detailed rule customization, you can pass a GameRule object to the RiichiEnv constructor. RiichiEnv provides presets for popular platforms (Tenhou, MJSoul) and allows granular configuration.

from riichienv import RiichiEnv, GameRule

# Example 1: Use MJSoul rules (allows Ron on Ankan for Kokushi Musou)
rule_mjsoul = GameRule.default_mjsoul()
env = RiichiEnv(game_mode="4p-red-half", rule=rule_mjsoul)

# Example 2: Fully custom rules based on Tenhou preset
rule_custom = GameRule.default_tenhou()
rule_custom.allows_ron_on_ankan_for_kokushi_musou = True  # Enable Kokushi Chankan
rule_custom.length_of_game_in_rounds = 8  # Force 8 rounds? (Note: Length is mainly controlled by game_mode logic usually)

env = RiichiEnv(game_mode="4p-red-half", rule=rule_custom)

Detailed mechanic flags (like allows_ron_on_ankan_for_kokushi_musou) are defined in the GameRule struct. See RULES.md for a full list of configurable options.

Tile Conversion & Hand Parsing

Standardize between various tile formats (136-tile, MPSZ, MJAI) and easily parse hand strings.

>>> import riichienv.convert as cvt
>>> cvt.mpsz_to_tid("1z")
108

>>> from riichienv import parse_hand
>>> parse_hand("123m406m789m777z")
([0, 4, 8, 12, 16, 20, 24, 28, 32, 132, 133, 134], [])

See DATA_REPRESENTATION.md for more details.

Agari Calculation

>>> from riichienv import AgariCalculator
>>> import riichienv.convert as cvt

>>> ac = AgariCalculator.hand_from_text("111m33p12s111666z")
>>> ac.is_tenpai()
True
>>> ac.calc(cvt.mpsz_to_tid("3s"))
Agari(agari=True, yakuman=False, ron_agari=12000, tsumo_agari_oya=0, tsumo_agari_ko=0, yaku=[8, 11, 10, 22], han=5, fu=60)

🛠 Development

For more architectural details and contribution guidelines, see CONTRIBUTING.md and DEVELOPMENT_GUIDE.md.

Check our Milestones for the future roadmap and development plans.

📄 License

Apache License 2.0