-
Notifications
You must be signed in to change notification settings - Fork 3
/
Neural UCB v.2.py
133 lines (114 loc) · 5.54 KB
/
Neural UCB v.2.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
#import tensorflow as tf
import numpy as np
import torch
import math
from torch import nn
import torch.optim as optim
from torch.autograd import Variable
import pandas as pd
import matplotlib.pyplot as plt
net = nn.Sequential(
nn.Linear(2, 60),
#通常會加入一個 non-linear layer
nn.Linear(60, 60),
nn.ReLU(),
nn.Linear(60, 1),
nn.Sigmoid()
)
class Bandit:
def __init__(self, alpha = 10, regulation_parameter = 1 ,exporation_parameter = 0.01 , confidence_parameter = 1, norm_parameter = 1, step_size = 20, gradient_num = 100, network_width = 2, network_depth = 100 ):
self.T = alpha
self.rp = regulation_parameter
self.v = exporation_parameter
self.confi = confidence_parameter
self.s = norm_parameter
self.n = step_size
self.J = gradient_num
self.m = network_width
self.L = network_depth
self.gama = 0.1
self.optimizer = optim.Adam(net.parameters()) #?????
self.C1 = 0.1
self.C2 = 0.1
self.C3 = 0.1
z = (3780, 3780, 3780)
self.Z = np.zeros(z)
for i in range(3780):
for j in range(3780):
if i==j:
self.Z[0][i][j] = self.rp
self.NeuralUCB()
def TrainNN(self, J, t, f_output_result, r, m, rp):
MSE = 0
loss_MSE = []
loss_mrp = []
l2_reg = torch.tensor(0)
for j in range(J-1):
for param in net.parameters():
l2_reg += torch.norm(param).type_as(l2_reg)
for i in range(t+1):
MSE = np.power((f_output_result - r[i][0][0]), 2)/2
MSE = torch.from_numpy(MSE)
loss_MSE.append(MSE)
loss_mrp.append(m*rp*l2_reg/2)
#loss = Variable(loss)
self.optimizer.zero_grad()
loss = torch.stack(loss_MSE).sum() + torch.stack(loss_mrp).sum()
loss = Variable(loss, requires_grad=True)
loss.backward()
self.optimizer.step()
def NeuralUCB(self):
T = 2000
K = 10
U = [0]*K
self.regret = [0]*T
for t in range(T): #run t times
f_MAX = 0
MAX = 0
for a in range(K): #has K arms
X = [[0]*K for i in range(T)]
self.r = [[0]*K for i in range(T)]
#create context
for i in range(T):
for j in range(K):
k1 = np.random.normal(0, 0.1, 1)
k2 = np.random.normal(0, 0.1, 1)
k1 = list(k1)
k2 = list(k2)
X[i][j] = np.array([[k1[0], k2[0]]]) #context [1*2]
phi = np.array([[k1[0]], [k2[0]]]) #context [1*2]
self.r[i][j] = np.dot(X[i][j], phi) + np.random.normal(0, 0.1, 1)
self.f_output = [0]*K
self.f_output[a] = net(torch.from_numpy(X[t][a]).float())
print(self.f_output[a])
if (self.f_output[a] > f_MAX) :
f_MAX = self.f_output[a]
self.optimizer.zero_grad()
self.f_output[a].backward()
gradient = np.concatenate((np.array(net[0].weight.grad.numpy().flatten()), np.array(net[1].weight.grad.numpy().flatten()), np.array(net[3].weight.grad.numpy().flatten())))
gradient = np.array(gradient).reshape(1, -1)
gradient_t = gradient.transpose()
gradient_result = np.dot(gradient, np.linalg.inv(self.Z[t]))
haha = gradient * np.linalg.inv(self.Z[t]) * gradient_t
product1 = np.dot(gradient, np.linalg.inv(self.Z[t]))
product2 = np.dot(product1, gradient_t)
U[a] = self.f_output[a] + self.gama * math.pow(product2/self.m, 1/2)
if (U[a] > U[MAX]) or (a == 0): #choose best arm by choosing best UCB
MAX = a
self.r[t][0] = self.r[t][a]
self.regret[t] = f_MAX - self.f_output[MAX] #best(from now) minus our choice
self.f_output_result = self.f_output[MAX]
self.Z[t+1] = self.Z[t] + gradient * np.transpose(gradient)/(self.m)
self.TrainNN(self.J, t, self.f_output_result, self.r, self.m, self.rp)
self.gama = np.power( 1 + self.C1 * np.power(self.m, -1/6) * np.power(np.math.log(self.m, 10) * np.power((self.rp), -7/6), 1/2)\
* (self.v * np.power(np.math.log(np.linalg.det(self.Z[t])/np.power(self.rp, 990), 10) + (self.C2) * np.power(self.m,-1/6)\
* np.power(np.math.log(self.m, 10),1/2) * np.power(self.L, 4) * np.power(self.T, 5/3) * np.power(self.rp, -1/6)\
- 2*np.math.log(self.confi), 1/2) + np.power(self.rp, 1/2)*self.s)\
+ (self.rp + self.C3 * self.T * self.L) * (np.power(1-self.n*self.m*self.rp, self.J/2) * np.power(self.T/self.rp, 1/2) + np.power(self.m, -1/6)\
*np.power(np.math.log(self.m, 10),1/2)*np.power(self.L, 7/2)*np.power(self.T, 5/3)*np.power(self.rp, -5/3)*(1+np.power(self.T/self.rp, 1/2))), 1/2)
print('times: ', t)
cumsum = np.cumsum(self.regret)
print('cumsum: ', cumsum[i][0])
plt.plot(cumsum)
plt.show()
Bandit()