-
Notifications
You must be signed in to change notification settings - Fork 32
/
agent.py
172 lines (147 loc) · 7.7 KB
/
agent.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
import tensorflow as tf
import numpy as np
class Agent:
def __init__(self):
self.lane_len = 10
self.GAMMA = 0.9
self.states = []
self.rewards = []
self.actions = []
self.init_nn()
self.sess = tf.InteractiveSession()
self.sess.run(tf.global_variables_initializer())
return
def init_nn(self):
self.state_layer = tf.placeholder(tf.float32, [None, self.lane_len*8 + 4], 'state')
self.lane_layers = [tf.slice(self.state_layer, [0, i*self.lane_len], [-1, self.lane_len]) for i in range(0, 8)]
self.phase = tf.slice(self.state_layer, [0, 8*self.lane_len], [-1, 4])
self.subsize = 4
with tf.name_scope('subnet'):
dim = [self.lane_len, 16, 8, self.subsize]
nlayer = len(dim)-1
w = [tf.Variable(tf.truncated_normal(dim[i:i + 2]) / dim[i]) for i in range(nlayer)]
b = [tf.Variable(tf.constant(0.00, shape=[dim[i + 1]])) for i in range(nlayer)]
for iw in w: tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, iw)
for ib in b: tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, ib)
layer = []
for ilane in self.lane_layers:
for ilayer in range(nlayer):
ilane = tf.matmul(ilane, w[ilayer]) + b[ilayer]
layer.append(tf.nn.leaky_relu(ilane))
self.sub_layers = []
self.sub_layers.append(layer[0]+layer[4])
self.sub_layers.append(layer[2]+layer[6])
self.sub_layers.append(layer[1]+layer[5])
self.sub_layers.append(layer[3]+layer[7])
with tf.name_scope('actornet'):
dim = [self.subsize, 16, 8, 4, 1]
nlayer = len(dim)-1
for n in range(4):
w = [tf.Variable(tf.truncated_normal(dim[i:i + 2]) / dim[i]) for i in range(nlayer)]
b = [tf.Variable(tf.constant(0.00, shape=[dim[i + 1]])) for i in range(nlayer)]
for iw in w: tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, iw)
for ib in b: tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, ib)
layer4 = []
for k in range(4):
layer = self.sub_layers[(k+n) % 4]
for ilayer in range(nlayer-1):
layer = tf.matmul(layer, w[ilayer]) + b[ilayer]
layer = tf.nn.leaky_relu(layer)
layer = tf.matmul(layer, w[-1]) + b[-1]
layer4.append(layer)
if n < 1:
self.action_layer = tf.concat(layer4, 1)
else:
self.action_layer = self.action_layer + tf.concat(layer4, 1)
dim = [4, 16, 8, 4]
nlayer = len(dim) - 1
w = [tf.Variable(tf.truncated_normal(dim[i:i + 2]) / dim[i]) for i in range(nlayer)]
b = [tf.Variable(tf.constant(0.00, shape=[dim[i + 1]])) for i in range(nlayer)]
for iw in w: tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, iw)
for ib in b: tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, ib)
layer = self.phase
for ilayer in range(nlayer):
layer = tf.nn.leaky_relu(tf.matmul(layer, w[ilayer])+b[ilayer])
# self.action_layer += layer
self.action_layer = tf.nn.softmax(self.action_layer)
self.advantage_fb = tf.placeholder(tf.float32, [None])
self.action_fb = tf.placeholder(tf.float32, [None, 4])
p = tf.reduce_mean(tf.multiply(self.action_layer, self.action_fb), reduction_indices=1)
logp = tf.log(tf.clip_by_value(p, 1e-8, 1.))
cost = - tf.reduce_mean(tf.multiply(self.advantage_fb, logp))
regularizer = tf.contrib.layers.l2_regularizer(scale=1e-6)
reg_variables = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES, 'actornnet')
reg_variables.extend(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES, 'subnet'))
reg_term = tf.contrib.layers.apply_regularization(regularizer, reg_variables)
cost += reg_term
self.lr = tf.placeholder(tf.float32)
self.actor_opt = tf.train.AdamOptimizer(self.lr).minimize(cost)
with tf.name_scope('criticnet'):
dim = [self.subsize, 16, 8, 4, 1]
nlayer = len(dim) - 1
w = [tf.Variable(tf.truncated_normal(dim[i:i+2]) / dim[i]) for i in range(nlayer)]
b = [tf.Variable(tf.constant(0.00, shape=[dim[i+1]])) for i in range(nlayer)]
for iw in w: tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, iw)
for ib in b: tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, ib)
layer4 = []
for isub in self.sub_layers:
for ilayer in range(nlayer-1):
isub = tf.matmul(isub, w[ilayer]) + b[ilayer]
isub = tf.nn.leaky_relu(isub)
layer4.append(tf.matmul(isub, w[-1]) + b[-1])
self.value_layer = layer4[0] + layer4[1] + layer4[2] + layer4[3]
self.return_fb = tf.placeholder(tf.float32, [None, 1])
cost = tf.losses.mean_squared_error(self.return_fb, self.value_layer)
regularizer = tf.contrib.layers.l2_regularizer(scale=1e-6)
reg_variables = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES, 'criticnet')
reg_variables.extend(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES, 'subnet'))
reg_term = tf.contrib.layers.apply_regularization(regularizer, reg_variables)
cost += reg_term
self.critic_opt = tf.train.AdamOptimizer(self.lr).minimize(cost)
def _train(self, advantage, Return, action, state, lr, iter):
for _ in range(iter[0]):
self.actor_opt.run(feed_dict={
self.advantage_fb: advantage,
self.state_layer: state,
self.action_fb: action,
self.lr: lr
})
for _ in range(iter[1]):
self.critic_opt.run(feed_dict={
self.return_fb: Return,
self.state_layer: state,
self.lr: lr
})
def policy(self, state):
y = self.sess.run(self.action_layer, feed_dict={self.state_layer: [state]})
action = np.random.choice(4, p=y[0])
return action
def value(self, state):
y = self.sess.run(self.value_layer, feed_dict={self.state_layer: [state]})
return np.squeeze(y)
def train(self, state, action, reward, lr, para):
self.states.append(state)
self.rewards.append(reward)
a = np.eye(4)[action]
self.actions.append(a)
[TDn, batch, trig, iter0, iter1] = para
iter = [iter0, iter1]
length = len(self.rewards)
if trig or length > batch * TDn:
r = np.array(self.rewards)
returns = np.zeros_like(r)
values = np.zeros_like(r)
discounted_sum = self.value(state)
gammas = np.hstack([self.GAMMA**n for n in range(0, TDn)])
for t in reversed(range(0, length)):
if t > (length-TDn-2):
discounted_sum = discounted_sum * self.GAMMA + r[t]
returns[t] = discounted_sum
else:
returns[t] = np.sum(gammas * r[t:(t+TDn)]) + self.value(self.states[t+TDn])
values[t] = self.value(self.states[t])
advantage = returns - values
actions = np.array(self.actions)
returns = np.reshape(returns, [length, 1])
self._train(advantage, returns, actions, self.states, lr, iter)
self.rewards, self.actions, self.states = [], [], []