output_file = "/home/hugh/connect_comp/ProgrammingAssignment/Task1.txt"
mlp = MultiLayerPerceptron((2, 2, 1), hidden_activation="sigmoid", max_iters=10000, linear_factor=0.2, learning_rate=0.7,
verbose=(True, 1000,), output_activation="linear", output_file=output_file)
x = np.array([[1, 0], [0, 1], [0, 0], [1, 1]])
y = np.array([1, 1, 0, 0])
errors = mlp.fit(x, y)
plt.figure()
plt.plot(errors, color="blue", label="Sigmoid + Linear")
plt.title("Mean Squared Error Over Time")
plt.xlabel("$Epochs$")
plt.ylabel("$Error$")
plt.show()
#plt.savefig("/home/hugh/connect_comp/ProgrammingAssignment/Task1.png")
print("1, 1")
print(mlp.predict(np.array([1, 1])))
print("0, 0")
print(mlp.predict(np.array([0, 0])))
print("0, 1")
print(mlp.predict(np.array([0, 1])))
print("1, 0")
print(mlp.predict(np.array([1, 0])))
#mlp.to_str()
class QNeuralNework(object):
"""
action: パターンの数だけ保持
学習アルゴリズム: Q学習
a = getNextAction(s)
lean(S,a,r,S_next)
※ ゲームルールによらない汎用性を持たす
"""
ALPHA = 0.5
GAMMA = 0.9
DATASET_NUMBER = 100
LEAN_EPOCHS = 500
LEAN_RATE = 0.2
GREEDY_RATIO = 0.5
def __init__(self, _numInput=3, numAction=3):
self.action_paturn = range(numAction)
self.learningObj = MultiLayerPerceptron(numInput=_numInput, numHidden=5, numOutput=numAction, activate1="tanh",
activate2="sigmoid")
self.X = []
self.Y = []
self.learnFlg = True
def learn(self, o, a, r, o_next):
"""Q学習 or NeuralNetworkを使って,Q値を学習"""
dQs = self.learningObj.predict(o)
qk = dQs[a]
maxQ = np.max(dQs)
dQs[a] = qk + self.ALPHA * (r + self.GAMMA * maxQ - qk)
self.X.append(np.asarray(o))
self.Y.append(np.asarray(dQs))
if len(self.X) > self.DATASET_NUMBER:
self.X.pop(0)
self.Y.pop(0)
err = self.learningObj.fit(np.asarray(self.X), np.asarray(self.Y), learning_rate=self.LEAN_RATE,
epochs=self.LEAN_EPOCHS)
return err
def getNextAction(self, o):
a = None
# 最大Q値の行動選択
# 観測(observe)から、NNでQ値(配列)を取得
Q_t = self.learningObj.predict(o)
maxQt_idx = np.argmax(Q_t)
best_actions = maxQt_idx
# Q値最大の中からランダム選択
a = np.random.choice(np.atleast_1d(best_actions))
# if self.learnFlg:
# return a
# greedyの行動選択
import random
if self.GREEDY_RATIO < random.random():
return a
else:
return np.random.choice(self.action_paturn)
errors = mlp.fit(x_train, y_train)
plt.plot(errors, color="green", label="Tanh + Linear")
plt.legend()
#plt.show()
plt.savefig("/home/hugh/connect_comp/ProgrammingAssignment/Task3a.png")
correct = 0
total = 0
#file = open("/home/hugh/connect_comp/ProgrammingAssignment/Task3.txt", 'a+')
#file.write("\n\n")
#file.close()
for i in range(len(y_test)):
with open("/home/hugh/connect_comp/ProgrammingAssignment/Task3.txt",
'a+') as fh:
prediction = mlp.predict(x_test[i])
#fh.write("Prediction {0}\n".format(prediction))
#fh.write("Actual {0}\n".format(y_test[i]))
if prediction == y_test[i]:
correct += 1
total += 1
#with open("/home/hugh/connect_comp/ProgrammingAssignment/Task3.txt", 'a+') as fh:
#fh.write("Percentage accuracy on test set: {0}".format((correct/total)*100))
#mlp.to_str()
# -*- coding: utf-8 -*-
"""
MLP_test.py
多層パーセプトロン
"""
import numpy as np
import sys
from MultiLayerPerceptron import MultiLayerPerceptron
from ndprint import ndprint, ndprints
if __name__ == '__main__':
# データセットの生成
# ------------------
X = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
y = np.array([0, 1, 1, 0])
# 多層パーセプトロンの初期化
mlp = MultiLayerPerceptron(numInput=2, numHidden=5, numOutput=1, activate1="tanh", activate2="identity")
# パーセプトロンの学習
mlp.fit(X, y,learning_rate=0.2, epochs=10000)
# パーセプトロンを実行
y0 = mlp.predict(X[0])
for x, y in zip(X, y):
print 'X:%s, y:%0.2f, pred:%0.2f' % (ndprint(x), y, mlp.predict(x))
