-
Notifications
You must be signed in to change notification settings - Fork 3
/
deep_q_learning.py
200 lines (176 loc) · 7.03 KB
/
deep_q_learning.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
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
import os
os.environ['KMP_DUPLICATE_LIB_OK']='True'
import numpy as np
import matplotlib.pyplot as plt
from tqdm import tqdm
import torch
from torch import nn, optim
from torch.utils.data import DataLoader, Dataset
from mdp import *
class Qsa(nn.Module):
def __init__(self, input_size=7, num_classes=len(A)):
super().__init__()
self.fc_liner = nn.Sequential(
nn.Linear(input_size, 32),
nn.ReLU(),
# nn.Linear(32, 16),
# nn.ReLU(),
nn.Linear(32, num_classes)
)
def forward(self, x):
return self.fc_liner(x)
class StatesDataset(Dataset):
def __init__(self, states, rewards, actions):
self.states = torch.Tensor(states[:-1]).float()
self.states_next = torch.Tensor(states[1:]).float()
self.rewards = torch.Tensor(rewards).float()
self.actions = actions
def __len__(self):
return len(self.states)
def __getitem__(self, idx):
return {
'states': self.states[idx],
'states_next': self.states_next[idx],
'rewards': self.rewards[idx],
'actions': self.actions[idx]
}
def deep_q_learning(qsa,
series,
state_init,
pi,
optimizer,
loss_func,
epochs=10,
episode=100,
gamma=0.9,
lr=0.7,
eps=0.5,
min_eps=0.05,
decay=0.9,
greedy=False,
verbose=True,
sarsa=False
):
losses = list()
learning_curve = list()
# loop for each episode
for epi in tqdm(range(episode)):
# generate a trajectory
eps *= decay
states, rewards, actions = simulate(series, state_init, pi, greedy, eps=max(min_eps, eps))
# form dataset and data loader
dataset = StatesDataset(states, rewards, actions)
dataloader = DataLoader(dataset, batch_size=128, shuffle=False)
# fit the deep NN
for epo in range(epochs):
for data_pack in dataloader:
input_tensor = data_pack['states']
out = qsa(input_tensor)
output_tensor = out[[i for i in range(len(data_pack['rewards']))], [a+k for a in data_pack['actions']]]
with torch.no_grad():
max_qsa_out = qsa(data_pack['states_next'])
if not sarsa:
max_qsa = max_qsa_out[[i for i in range(len(data_pack['rewards']))], max_qsa_out.argmax(dim=1)]
max_qsa = torch.Tensor(data_pack['rewards']).float() + (gamma * max_qsa)
else:
max_qsa = max_qsa_out[[i for i in range(len(data_pack['rewards']))], [a+k for a in data_pack['actions'][1:]] + [0]]
target_tensor = (1 - lr) * output_tensor + lr * max_qsa
# update weights
loss = loss_func(output_tensor, target_tensor)
qsa.zero_grad()
loss.backward()
optimizer.step()
losses.append(loss.detach().item())
learning_curve.append(interact_test(pi, series_name='test', verbose=False))
# verbose
if verbose:
print("Last loss", losses[-1])
plt.plot(losses)
plt.xlabel('Iterations (Not Epochs)')
plt.ylabel('Loss')
plt.savefig('loss.pdf')
plt.show()
return learning_curve
def train_deep_q(verbose=False):
qsa = Qsa(input_size=7, num_classes=len(A))
state_init = [train_series['Close'][0], balance_init, 0] + list(train_series.iloc[0][5:])
series = train_series[1:]
optimizer = optim.Adam(
qsa.parameters(),
lr=1e-5, # 1e-4
# weight_decay=1e-6
)
# loss_func = nn.MSELoss()
loss_func = nn.HuberLoss()
def pi_deep(s, eps=0.2, greedy=False):
with torch.no_grad():
out_qsa = qsa(torch.Tensor(s).float()).squeeze()
action = out_qsa.argmax().item() - k
if not greedy:
r = np.random.rand()
# if it is on the less side, the explore other actions
if r > 1 - eps + (eps / len(A)):
a_ = np.random.choice(A)
while a_ == action:
a_ = np.random.choice(A)
action = a_
return action
learning_curve = deep_q_learning(qsa,
series,
state_init,
pi_deep,
optimizer,
loss_func,
epochs=10, # number of epochs for training NN in each episode 10
episode=30, # 30
gamma=0.6, # discount coefficient 0.618
lr=0.7, # learning rate for update q function
eps=0.8, # eps greedy policy
min_eps=0.2, # 0.2
decay=0.9, # 0.9
greedy=False,
verbose=verbose
)
return pi_deep, qsa, learning_curve
def train_deep_sarsa(verbose=False, sarsa=False):
qsa = Qsa(input_size=7, num_classes=len(A))
state_init = [train_series['Close'][0], balance_init, 0] + list(train_series.iloc[0][5:])
series = train_series[1:]
optimizer = optim.Adam(
qsa.parameters(),
lr=1e-5, # 1e-4
# weight_decay=1e-6
)
# loss_func = nn.MSELoss()
loss_func = nn.HuberLoss()
def pi_deep(s, eps=0.2, greedy=False):
with torch.no_grad():
out_qsa = qsa(torch.Tensor(s).float()).squeeze()
action = out_qsa.argmax().item() - k
if not greedy:
r = np.random.rand()
# if it is on the less side, the explore other actions
if r > 1 - eps + (eps / len(A)):
a_ = np.random.choice(A)
while a_ == action:
a_ = np.random.choice(A)
action = a_
return action
deep_q_learning(qsa,
series,
state_init,
pi_deep,
optimizer,
loss_func,
epochs=10, # number of epochs for training NN in each episode 10
episode=30, # 30
gamma=0.6, # discount coefficient 0.618
lr=0.7, # learning rate for update q function
eps=0.8, # eps greedy policy
min_eps=0.2, # 0.2
decay=0.9, # 0.9
greedy=False,
verbose=verbose,
sarsa=sarsa
)
return pi_deep, qsa