Koi-Koi AI

Learning based AI for playing multi-round Koi-Koi hanafuda card games.

Platform

• Python
• PyTorch
• PySimpleGUI (for the interface playing vs AI)

About Koi-Koi Hanafuda Card Games

Hanafuda is a kind of traditional Japanese playing cards. A hanafuda deck contains 48 cards divided by 12 suits corresponding to 12 months, which are also divided into four rank-like categories with different importance. Koi-Koi is a kind of two-player hanafuda card game. The goal of Koi-Koi is to collect cards by matching the cards by suit, and forming specific winning hands called Yaku from the acquired pile to earn points from the opponent.

Rules & Yaku List

Koi-Koi is consisted by multiple rounds and both players start with equal points. In every round, two players discard and draw to pair and collect cards by turn until someone forms Yakus successfully. Then, he can end this round to receive points from the opponent, or claim koi-koi and continues this round to earn more yakus and points. The detailed rules and Yaku list of this project is the same as PC game KoiKoi-Japan on Steam.

Network Model & Dataset

The model is a card-state self-attention based neural network with multiple Transformer encoder layers, and is trained by supervised learning. The dataset of 200 eight-round Koi-Koi game records is also provided.

Future Work

Try to improve the performance with reinforcement learning...