13個の手牌を1014個の情報にして聴牌かどうかを判断するニューラルネットワーク

※この記事は2020年に作成しました

#概要
13個の手牌を286通りにして3つのニューラルネットワークにかけてその結果から聴牌かどうかを判断するニューラルネットワークにさらに、13個から2個取ってくる78通りの2つの牌が、対子かどうかを判断するYNと、塔子がどうかを判断するYNを付けたもの。
3つ牌を選んだ286通りを3つのYNで、2つの牌を選んだ78通りを2つのYNでの結果を聴牌YNの学習データとしている。
データ数は4942個の聴牌データと4942個のノーテンデータを使用している。
YNのモデルで判断する回数は、「データ数×（（286通り×3つのYN）＋（78通り×2つのYN））」なので、1002万2376回である。

 _{13}C_2=78

#結果
テストデータに対する正解率は、98.6％ほどになった。

#作成したプログラム

# coding: UTF-8
import numpy as np
import itertools
import time
from keras.models import load_model
from keras.utils import to_categorical
from keras.models import Sequential
from keras.layers import Dense
from keras.optimizers import Adam
np.set_printoptions(threshold=10000000,linewidth=200)
t1 = time.time() 

#x_test = np.loadtxt("C:/Users/p-user/Desktop/csvdata/old/tehai_xtest.csv",delimiter=",")
#y_test = np.loadtxt("C:/Users/p-user/Desktop/csvdata/old/tehai_ytest.csv",delimiter=",")

data = 4942
x_train = np.zeros((data,34), dtype='float32')# x_trainは聴牌の手牌
with open('C:/sqlite/tenpai.csv', 'r') as fr:
    for i,row in enumerate(fr.readlines(),start=0):
        if i <data:
            x_train[i] += np.array(list(map(np.float,row[:34])))

y_train = np.zeros((data,34), dtype='float32')# y_trainはノーテンの手牌
with open('C:/sqlite/noten.csv', 'r') as fr:
    for i,row in enumerate(fr.readlines(),start=0):
        if i <data:
            y_train[i] += np.array(list(map(np.float,row[:34])))


tehai = np.zeros((data,13))#x_trainを13個の手牌tehaiへ
for i in range(data):
    cnt = 0
    for j in range(34):
        if x_train[i,j] >= 1:
            tehai[i,cnt] = j
            cnt += 1
            x_train[i,j] -=1
        if x_train[i,j] >= 1:
            tehai[i,cnt] = j
            cnt += 1
            x_train[i,j] -=1
        if x_train[i,j] >= 1:
            tehai[i,cnt] = j
            cnt += 1
            x_train[i,j] -=1
        if x_train[i,j] >= 1:
            tehai[i,cnt] = j
            cnt += 1
            x_train[i,j] -=1

tehaiN = np.zeros((data,13))#y_trainを13個の手牌tehaiNへ
for i in range(data):
    cnt = 0
    for j in range(34):
        if y_train[i,j] >= 1:
            tehaiN[i,cnt] = j
            cnt += 1
            y_train[i,j] -=1
        if y_train[i,j] >= 1:
            tehaiN[i,cnt] = j
            cnt += 1
            y_train[i,j] -=1
        if y_train[i,j] >= 1:
            tehaiN[i,cnt] = j
            cnt += 1
            y_train[i,j] -=1
        if y_train[i,j] >= 1:
            tehaiN[i,cnt] = j
            cnt += 1
            y_train[i,j] -=1

label=[0,1]
tehaiC = np.zeros((data,286,3))
for i in range(data):#tehaiを286通りへ
    tehaiC[i,] = list(itertools.combinations(tehai[i,], 3))
    #print(tehaiC[i,])

tehaiCN = np.zeros((data,286,3))
for i in range(data):#tehaiNを286通りへ
    tehaiCN[i,] = list(itertools.combinations(tehaiN[i,], 3))
    #print(tehaiC[i,])

tehaiC78 = np.zeros((data,78,2))
for i in range(data):#tehaiを78通りへ
    tehaiC78[i,] = list(itertools.combinations(tehai[i,], 2))

tehaiCN78 = np.zeros((data,78,2))
for i in range(data):#tehaiNを78通りへ
    tehaiCN78[i,] = list(itertools.combinations(tehaiN[i,], 2))

modelToitsu = load_model('toitsuYN.h5')   
ToitsuYN = np.zeros((data*2,78))#対子YNの結果を入れる
for i in range(data):#tehaiC78を対子YNにかける
    for j in range(78):
        x_train = np.zeros((1,2))
        x_train[0,] = tehaiC78[i,j]
        x_train = x_train/33.
        pred = modelToitsu.predict(x_train, batch_size=1, verbose=1)
        pred_label = label[np.argmax(pred[0])]
        ToitsuYN[i,j] = np.argmax(pred[0])
for i in range(data):#tehaiCN78を対子YNにかける
    for j in range(78):
        x_train = np.zeros((1,2))
        x_train[0,] = tehaiCN78[i,j]
        x_train = x_train/33.
        pred = modelToitsu.predict(x_train, batch_size=1, verbose=1)
        pred_label = label[np.argmax(pred[0])]
        ToitsuYN[i+data,j] = np.argmax(pred[0])

modelTatsu = load_model('tatsuYN.h5')  
TatsuYN = np.zeros((data*2,78))#塔子YNの結果を入れる
for i in range(data):#tehaiC78を塔子YNにかける
    for j in range(78):
        x_train = np.zeros((1,2))
        x_train[0,] = tehaiC78[i,j]
        x_train = x_train/33.
        pred = modelTatsu.predict(x_train, batch_size=1, verbose=1)
        pred_label = label[np.argmax(pred[0])]
        TatsuYN[i,j] = np.argmax(pred[0])
for i in range(data):#tehaiCN78を塔子YNにかける
    for j in range(78):
        x_train = np.zeros((1,2))
        x_train[0,] = tehaiCN78[i,j]
        x_train = x_train/33.
        pred = modelTatsu.predict(x_train, batch_size=1, verbose=1)
        pred_label = label[np.argmax(pred[0])]
        TatsuYN[i+data,j] = np.argmax(pred[0])

modelShunts = load_model('shuntsYN.h5')   
ShuntsYN = np.zeros((data*2,286))#順子YNの結果を入れる
for i in range(data):#tehaiCを順子YNにかける
    for j in range(286):
        x_train = np.zeros((1,3))
        x_train[0,] = tehaiC[i,j]
        x_train = x_train/33.
        pred = modelShunts.predict(x_train, batch_size=1, verbose=1)
        pred_label = label[np.argmax(pred[0])]
        ShuntsYN[i,j] = np.argmax(pred[0])
for i in range(data):#tehaiCNを順子YNにかける
    for j in range(286):
        x_train = np.zeros((1,3))
        x_train[0,] = tehaiCN[i,j]
        x_train = x_train/33.
        pred = modelShunts.predict(x_train, batch_size=1, verbose=1)
        pred_label = label[np.argmax(pred[0])]
        ShuntsYN[i+data,j] = np.argmax(pred[0])

modelkotsu = load_model('kotsuYN.h5')
kotsuYN = np.zeros((data*2,286))#刻子YNの結果を入れる
for i in range(data):#tehaiCを刻子YNにかける
    for j in range(286):
        x_train = np.zeros((1,3))
        x_train[0,] = tehaiC[i,j]
        x_train = x_train/33.
        pred = modelkotsu.predict(x_train, batch_size=1, verbose=1)
        pred_label = label[np.argmax(pred[0])]
        kotsuYN[i,j] = np.argmax(pred[0])
for i in range(data):#tehaiCNを刻子YNにかける
    for j in range(286):
        x_train = np.zeros((1,3))
        x_train[0,] = tehaiCN[i,j]
        x_train = x_train/33.
        pred = modelkotsu.predict(x_train, batch_size=1, verbose=1)
        pred_label = label[np.argmax(pred[0])]
        kotsuYN[i+data,j] = np.argmax(pred[0])

modelkokushi = load_model('kokushiYN.h5')
kokushiYN = np.zeros((data*2,286))#国士YNの結果を入れる
for i in range(data):#tehaiCを国士YNにかける
    for j in range(286):
        x_train = np.zeros((1,3))
        x_train[0,] = tehaiC[i,j]
        x_train = x_train/33.
        pred = modelkokushi.predict(x_train, batch_size=1, verbose=1)
        pred_label = label[np.argmax(pred[0])]
        kokushiYN[i,j] = np.argmax(pred[0])
for i in range(data):
    for j in range(286):#tehaiCNを国士YNにかける
        x_train = np.zeros((1,3))
        x_train[0,] = tehaiCN[i,j]
        x_train = x_train/33.
        pred = modelkokushi.predict(x_train, batch_size=1, verbose=1)
        pred_label = label[np.argmax(pred[0])]
        kokushiYN[i+data,j] = np.argmax(pred[0])

YN3 = np.zeros((data*2,1014))
for i in range(data*2):
    ShunKotYN = np.hstack((ShuntsYN[i,],kotsuYN[i,]))
    SKKYN = np.hstack((ShunKotYN,kokushiYN[i,]))
    SKKYN_toitsu = np.hstack((SKKYN,ToitsuYN[i,]))
    SKKYN_tatsu = np.hstack((SKKYN_toitsu,TatsuYN[i,]))
    YN3[i,] = SKKYN_tatsu
#print(YN3[0,])
#print(YN3.shape)

train = np.zeros((data*2,1015))
for i in range(data):
    x = np.append(YN3[i,],1)
    train[i,] = x

for i in range(data):
    x = np.append(YN3[i+data,],0)
    train[i+data,] = x

np.random.shuffle(train)
lavel = np.zeros((data*2,1))

for i in range(data*2):
    if train[i,1014] == 1:
        lavel[i,0] = 1

train=np.delete(train,1014,1)
lavel = to_categorical(lavel,2)
#trainが学習データ
#lavelが正解ラベル

model = Sequential()

#中間層、入力層、活性化関数ReLU関数
model.add(Dense(units=507,input_shape=(1014,),activation='relu'))
#出力層2、活性化関数ソフトマックス関数
model.add(Dense(units=2,activation='softmax'))

#最適化アルゴリズムAdam、損失関数クロスエントロピー
model.compile(
    optimizer=Adam(lr=0.005),
    loss='categorical_crossentropy',
    metrics=['accuracy'],
)

history_adam=model.fit(
    train,
    lavel,
    batch_size=100,
    epochs=50,
    verbose=1,
    shuffle = True,
    validation_split=0.2
)
#model.save('tenpaiYN.h5')

t2 = time.time()
elapsed_time = t2-t1
print(f"経過時間：{elapsed_time}")