-
Notifications
You must be signed in to change notification settings - Fork 1
/
NAFAgent.py
248 lines (187 loc) · 8.71 KB
/
NAFAgent.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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
import random
import gym
import numpy as np
import tensorflow as tf
from collections import deque
from tensorflow.contrib.layers import fully_connected
from tensorflow.contrib.framework import get_variables
class NAFAgent:
def __init__(self, env):
# Initialive discounts, networks, EVERYTHING!
self.gamma = 0.95
self.tau = 0.05
self.epsilon = 1.0
self.epsilon_decay = 0.98
self.alpha = 0.0001
self.update_samples = 200
self.update_steps = 10
self.hidden_layers = 2
self.hidden_nodes = 100
print(self.gamma, self.tau, self.epsilon, self.epsilon_decay, \
self.alpha, self.update_samples, self.update_steps, self.hidden_layers, \
self.hidden_nodes)
self.env = env
self.tf_sess = tf.InteractiveSession()
# Get state and action counts.
self.states = self.env.observation_space.shape[0]
if isinstance(self.env.action_space, gym.spaces.Discrete):
self.actions = self.env.action_space.n
else:
self.actions = self.env.action_space.shape[0]
self.network = {}
for name in ['nqn', 'nqn_p']:
self.create_network(name)
self.tf_sess.run(tf.global_variables_initializer())
if self.actions > 1:
self.one_hot = [self.tf_sess.run(tf.one_hot(i, self.actions)) for i in range(self.actions)]
# Buld update network:
self.update_vars = {}
for nqn_var, nqn_p_var in zip(self.network['nqn']['vars'], \
self.network['nqn_p']['vars']):
# Set initial values.
self.tf_sess.run(nqn_p_var.assign(nqn_var))
# Create tau update functions
self.update_vars[nqn_p_var.name] = nqn_p_var.assign\
(self.tau * nqn_var + (1. - self.tau) * nqn_p_var)
# Replay buffer
self.replay = deque([])
def create_network(self, name):
networks = {}
with tf.variable_scope(name):
# Input parameters
networks['x'] = tf.placeholder(tf.float32, shape=[None, self.states], \
name='states')
networks['u'] = tf.placeholder(tf.float32, shape=[None, self.actions], \
name='actions')
# hidden layers
init = 1./self.hidden_nodes/self.actions
hid = networks['x']
hid = fully_connected(hid, self.hidden_nodes, \
weights_initializer=tf.random_normal_initializer(init, init/5), \
biases_initializer=tf.random_normal_initializer(init, init/5), \
activation_fn=tf.tanh)
for i in range(self.hidden_layers-1):
hid = fully_connected(hid, self.hidden_nodes, \
weights_initializer=tf.random_normal_initializer(init, init/5), \
biases_initializer=tf.random_normal_initializer(init, init/5), \
activation_fn=tf.nn.relu)
#hid = tf.nn.softmax(hid)
# Output parameters
networks['V'] = fully_connected(hid, self.actions, \
weights_initializer=tf.random_normal_initializer(1., 0.1), \
biases_initializer=tf.random_normal_initializer(0., 0.1))
networks['mu'] = fully_connected(hid, self.actions, \
weights_initializer=tf.random_normal_initializer(1., 0.1), \
biases_initializer=tf.random_normal_initializer(0., 0.1))
networks['mu_out'] = tf.nn.softmax(networks['mu'])
# Linear output layer
l = fully_connected(hid, int((self.actions * (self.actions + 1))/2), \
weights_initializer=tf.random_normal_initializer(1., 0.1), \
biases_initializer=tf.random_normal_initializer(0., 0.1))
# Build A(x, u)
axis_T = 0
rows = []
# Identify diagonal
for i in range(self.actions):
count = self.actions - i
# Create a row with the diagonal elements exponentiated.
diag = tf.exp(tf.slice(l, (0, axis_T), (-1, 1)))
# Create the "other" elements of the row
others = tf.slice(l, (0, axis_T + 1), (-1, count - 1))
# Assemble them into a full row.
row = tf.pad(tf.concat((diag, others), axis=1), \
((0, 0), (i, 0)))
# Add each row to a list for L(x)
rows.append(row)
axis_T += count
# Assemble L(x) and matmul by its transpose.
networks['L'] = tf.transpose(tf.stack(rows, axis=1), (0, 2, 1))
networks['P'] = P = tf.matmul(networks['L'], \
tf.transpose(networks['L'], (0, 2, 1)))
mu_u = tf.expand_dims(networks['u'] - networks['mu'], -1)
# Combine the terms
p_mu_u = tf.matmul(P, mu_u, name='Pxmu_u')
p_mess = tf.matmul(tf.transpose(mu_u, [0, 2, 1]),
p_mu_u, name='mu_u_TxPxmu_u')
networks['A'] = tf.multiply(-1./2., p_mess, name='A')
networks['Q'] = tf.add(networks['A'], networks['V'], name='Q_func')
# Describe loss functions.
networks['y_'] = tf.placeholder(tf.float32, [None, 1], name='y_i')
networks['loss'] = tf.reduce_mean(tf.squared_difference(networks['y_'], \
tf.squeeze(networks['Q'])), name='loss')
# GradientDescent
networks['gdo'] = tf.train.AdamOptimizer(learning_rate=self.alpha,
epsilon=0.5).minimize(networks['loss'])
self.network[name] = networks
self.network[name]['vars'] = get_variables(name)
return
def update_target(self):
for variable in self.network['nqn_p']['vars']:
self.tf_sess.run(self.update_vars[variable.name])
return
def reset(self):
self.epsilon *= self.epsilon_decay
return
def get_action(self, state, report=False):
action = self.tf_sess.run(self.network['nqn']['mu'],\
feed_dict={self.network['nqn']['x']: [state]})
if self.actions == 1:
action = action[0][0] + np.random.normal(0, self.epsilon)
if report:
print('mu: {0}'.format(action))
# Bound the action
if action > self.env.action_space.high[0]:
action = self.env.action_space.high[0]
elif action < self.env.action_space.low[0]:
action = self.env.action_space.low[0]
return [action]
action_sm = self.tf_sess.run(self.network['nqn']['mu_out'],\
feed_dict={self.network['nqn']['x']: [state]})
if report:
print('mu: {0}'.format(action))
print('softmax: {0}'.format(action_sm))
action_sm += np.random.normal(0, self.epsilon, self.actions)
return np.argmax(action_sm)
def update(self, state, action, reward, state_prime, done):
#if done:
#reward = -reward
#print(reward)
self.replay.append((state, action, reward, state_prime))
if len(self.replay) > 1e6:
self.replay.popleft()
#print action, reward, done
m = self.update_samples
for _ in range(self.update_steps):
# Get m samples from self.replay
if m > len(self.replay):
m = len(self.replay)
replays = random.sample(self.replay, m)
x = []
u = []
x_p = []
y_ = []
for replay in replays:
x.append(replay[0])
x_p.append(replay[3])
if self.actions > 1:
u.append(self.one_hot[replay[1]])
else:
u.append(replay[1])
# self.nqn_y_ fed from r + self.gamma * V'(s_p)
y_.append(replay[2])
V_p = self.tf_sess.run(self.network['nqn_p']['V'], \
feed_dict={self.network['nqn_p']['x']: x_p})
#print y_, V_p
y_ = [[temp1 + temp2[0]] for temp1, temp2 in zip(y_, self.gamma * V_p)]
#print 'y_i: {0}'.format(y_)
#print 'u: {0}, y_:{1}'.format(u, y_)
# minimize loss function for y_i, x, u.
self.tf_sess.run(self.network['nqn']['gdo'], \
feed_dict={
self.network['nqn']['x']: x,
self.network['nqn']['u']: u,
self.network['nqn']['y_']: y_
})
# update target network.
self.update_target()
return