Deep Learning Mahjong

Reinforcement learning (RL) implementation of imperfect information game Mahjong using markov decision processes to predict future game states
Alternatives To Deep Learning Mahjong
Project NameStarsDownloadsRepos Using ThisPackages Using ThisMost Recent CommitTotal ReleasesLatest ReleaseOpen IssuesLicenseLanguage
Pysc27,745206a month ago8September 27, 201947apache-2.0Python
StarCraft II Learning Environment
Open_spiel3,68612 days ago15May 28, 202237apache-2.0C++
OpenSpiel is a collection of environments and algorithms for research in general reinforcement learning and search/planning in games.
Alpha Zero General3,248
13 days ago65mitJupyter Notebook
A clean implementation based on AlphaZero for any game in any framework + tutorial + Othello/Gobang/TicTacToe/Connect4 and more
Football2,98428 months ago27January 25, 202226apache-2.0Python
Check out the new game server:
2 days ago61otherPython
High-quality single file implementation of Deep Reinforcement Learning algorithms with research-friendly features (PPO, DQN, C51, DDPG, TD3, SAC, PPG)
Rlcard2,181118 days ago38March 23, 202253mitPython
Reinforcement Learning / AI Bots in Card (Poker) Games - Blackjack, Leduc, Texas, DouDizhu, Mahjong, UNO.
Muzero General2,074
2 months ago49mitPython
9 days agoapache-2.0Python
Artificial intelligence for the Snake game.
Awesome Deep Rl1,225
4 months agomitHTML
For deep RL and the future of AI.
17 days ago30February 08, 202214otherJupyter Notebook
​TextWorld is a sandbox learning environment for the training and evaluation of reinforcement learning (RL) agents on text-based games.
Alternatives To Deep Learning Mahjong
Select To Compare

Alternative Project Comparisons

Deep Learning Mahjong [机器学习深度学习麻將] 🔴

Mahjong [麻將] 🔴

Status GitHub Issues GitHub Pull Requests License

Motivation 🔻

Watching my grandma play mahjong online and I was curious how the NPCs made their decisions.

Game Play 🔻

  • 3 or 4 player draw-and-discard game with 144 tiles based on Chinese characters and symbols.
  • Match open pairs of identical tiles, remove from board, exposing the tiles under them for play.
  • Game ends when all pairs of tiles have been removed from the board or no more exposed pairs remaining.
  • Players get more realistic experiences and playable content in a game that involves skill, strategy, calculation, and chance.

Types [Make Configurable] 🔴

  • Old Hong Kong / Cantonese Mahjong [DEFAULT MODE]
  • Competitive Mahjong International Standard
  • Three-Player Mahjong [3-ka]
  • Battle Mahjong [Player vs Cartoon NPC]

Game Tile Count Per Set [Total 144] 🔻

  • Simples [108]
    • Dots 36
    • Bamboo 36
    • Characters 36
  • Honors [28]
    • Winds - [North, West, South, East] 16
    • Dragons - [Red, Green, White] 12
  • Bonus [8]
    • Flower - [Plum Blossom, Orchid, Chrysanthemum, Bamboo] 4
    • Seasons - [Spring, Summer, Autumn, Winter] 4
  • Mahjong Combos
    • Heavenly Hand [天糊]
    • Great Winds [大四喜]
    • Great Dragons [大三元]
    • All Kongs [十八羅漢]
    • All Honor Tiles [字一色]
    • Thirteen Orphans [十三幺]
    • Nine Gates Hand [九蓮宝燈]
    • Self Triplets [四暗刻]
    • All in Triplets [對對糊]
    • Mixed One Suit [混一色]
    • All One Suit [清一色]
    • Common Hand [平糊]
    • Small Dragons [小三元]
    • Small Winds [小四喜]

Image source Wikipedia

Technical Mahjong Game Documentation [技术文档] 🔴

Mahjong 🔻

  • ML Algorithms allows game to react and respond more dynamically and in more imaginative ways.
  • Deep Neural Network with reinforcement learning implemented.
  • Learn from its own game and top human players (via Classic Supervised Learning) where computations are made for every move or position.


Machine Learning [机器学习] 🔻

  • Non-Player Characters (NPCs)

    • Algorithms playing as NPCs (with adjustable difficulties) respond to player’s actions in unique, unexpected ways.
    • NPCs are non hard-coded.
    • Train NPCs by imitating Top Mahjong Players to learn dynamic movements and actions.
    • Natural Language Processing [NLP] to build realistic interactions in conversations. Key for Battle Mahjong [Player vs Cartoon NPC] style.
  • Computational Modelling

    • Complex game states modelled such that game can predict and alter downstream effects:
    • Ex1: Team chemistry score calculated based on personalities of each gamer.
    • Ex2: Morale of each player’s abilities as game is played in real-time.
  • Game Aesthetics

    • Ex: Computer Vision Algorithms used for mahjong textures and objects to render dynamically as player moves tiles on the board.

Deep Learning [深度学习] 🔻

  • DL Game Play
    • AI will win through intelligence rather than faster mechanicals speed.
    • Computers can programatically issue commands instantly whereas humans must physically move a mouse or hit the keyboard.
    • Knowledge based hierarchy foundation with Goals, Strategies, Tactics, and Chains.
    • Each objective inspects current game state and decides which lower level objective will be best to achieve it.

  • Reinforcement Learning

    • Markov Decision process to make decisions involving chain of if-then statements.
    • Positive or Negative Reward.
    • Algorithm will learn what actions will maximize the reward and which to be avoided.
  • Deep Neural Network

    • 3 Hidden layers of 120 neutrons.
    • 3 Dropout layers to optimize generalization and reduce over-fitting.
      • Input - State
      • Output - Values related to Mahjong Actions
    • Last layer uses Softmax Function to return probabilities.
  • Deep Q-Learning

    • Q-table matrix that updates Q-table based on the Prediction of Future Mahjong States.

    • Q-values updated according to the Bellman Equation.

      Deep Q action-value function

Search Gaming Optimization Algorithm [搜索优化] 🔻

  • Alpha-Beta Prunning
    • AI weeds out bad moves.
  • “Lookahead” Search Algorithms
  • Open World Games
    • Typically require thousands of hours of developer and artist time to render.
    • Become more efficient using ML Path-Finding Algorithms.
    • Have the potential to be unlimited in size

Database [数据库] 🔴

  • Optimize game data with databases.
  • Pre-Computed Moves for the beginning/end phrases of the game.
  • Two Databases
    • Opening DB
    • Endgame DB

References [图书] 🔴

Popular Video Game Projects
Popular Reinforcement Learning Projects
Popular Games Categories
Related Searches

Get A Weekly Email With Trending Projects For These Categories
No Spam. Unsubscribe easily at any time.
Computer Vision
Game Development
Reinforcement Learning
Deep Neural Networks
Supervised Learning
Deep Q Network