hcpe_score_value_fit.py

import argparse
from cshogi import *
import numpy as np
import pandas as pd
from scipy.optimize import curve_fit

parser = argparse.ArgumentParser()
parser.add_argument('hcpe')
args = parser.parse_args()

# 評価値から勝率への変換
def score_to_value(score, a):
    return 1.0 / (1.0 + np.exp(-score / a))

hcpes = np.fromfile(args.hcpe, HuffmanCodedPosAndEval)

turns = hcpes['hcp'][:,0] & 1 # hcpの1ビット目はturnを表す
signs = 1 - turns.astype(np.int8) * 2 # 後手の符号を反転
df = pd.DataFrame({'score': hcpes['eval'] * signs, 'result': 2 - hcpes['gameResult']})

print(df['score'].describe())
print(df['result'].describe())

X = df['score']
Y = df['result']

popt, pcov = curve_fit(score_to_value, X, Y, p0=[600.0])
print(popt)

print('score < 1000')
df1 = df[df['score'].abs() < 1000]
X = df1['score']
Y = df1['result']

popt, pcov = curve_fit(score_to_value, X, Y, p0=[600.0])
print(popt)

print('score >= 1000')
df2 = df[df['score'].abs() >= 1000]
X = df2['score']
Y = df2['result']

popt, pcov = curve_fit(score_to_value, X, Y, p0=[600.0])
print(popt)