線形分類不可能のXor問題を解いてみました。重み更新ははバックプロパゲーションによって行っています。中間層のユニットは2です
# -*- coding: utf-8 -*-
import random
import math
studyFlg = True
weight =[[0.1,0.1,0.06],[-0.04,0.01,-0.1]]
deltaWeight = [[0,0,0],[0,0,0]]
vWeight = [0.02,-0.009,-0.051]
deltaVWeight = [0,0,0]
hidden = [0,0,1]
thetaz = 0
thetay = [0,0]
n = 0.3#学習係数
a = 0.9#慣性係数
def out(parameter,w):
s = 0
for index,value in enumerate(parameter):
s += value*w[index]
e = math.exp(-s)
return 1/(1+e)
def output(x):
for index,value in enumerate(weight):
hidden[index] = out(x,value)
return out(hidden,vWeight)
def study(x,t):
z = output(x)
thetaz = z*(1-z)*(t-z)
thetay[0] = hidden[0]*(1-hidden[0])*thetaz
thetay[1] = hidden[1]*(1-hidden[1])*thetaz
for index,value in enumerate(vWeight):
deltaVWeight[index] = n*thetaz*hidden[index] + deltaVWeight[index]*a
vWeight[index] = vWeight[index] + deltaVWeight[index]
for i in range(2):
for j in range(3):
deltaWeight[i][j] = n*thetay[i]*x[j] + deltaWeight[i][j]*a
weight[i][j] = weight[i][j] + deltaWeight[i][j]
#print weight
#print deltaWeight
return z
arrayX = [[0,0,1],[1,0,1],[0,1,1],[1,1,1]]
arrayT = [0,1,1,0]
for i,v in enumerate(arrayX):
print output(v)
while studyFlg:
flg = False
s = 0
for x,t in zip(arrayX,arrayT):
answer = study(x,t)
answer = (answer - t)*(answer - t)
s += answer
flg = s > 0.0001 or flg
studyFlg = flg
#print output([0,0,1])
for i,v in enumerate(arrayX):
print output(v)
出力結果
0.488758075607 0.488905017624 0.48887700069 0.489022967732 0.00499997290096 0.995327822571 0.995327066708 0.00559726403334


