-
Notifications
You must be signed in to change notification settings - Fork 0
/
policy_gradient.py
191 lines (147 loc) · 5.99 KB
/
policy_gradient.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
"""
Policy Gradient Reinforcement Learning
Uses a 3 layer neural network as the policy network
"""
import numpy as np
import tensorflow as tf
class PolicyGradient:
def __init__(
self,
n_x,
n_y,
learning_rate=0.01,
reward_decay=0.95,
load_path=None,
save_path=None
):
self.n_x = n_x
self.n_y = n_y
self.lr = learning_rate
self.gamma = reward_decay
self.save_path = None
if save_path is not None:
self.save_path = save_path
self.episode_observations, self.episode_actions, self.episode_rewards = [], [], []
self.build_network()
self.cost_history = []
self.sess = tf.Session()
# $ tensorboard --logdir=logs
# http://0.0.0.0:6006/
tf.summary.FileWriter("logs/", self.sess.graph)
self.sess.run(tf.global_variables_initializer())
# 'Saver' op to save and restore all the variables
self.saver = tf.train.Saver()
# Restore model
if load_path is not None:
self.load_path = load_path
self.saver.restore(self.sess, self.load_path)
def store_transition(self, s, a, r):
"""
Store play memory for training
Arguments:
s: observation
a: action taken
r: reward after action
"""
self.episode_observations.append(s)
self.episode_rewards.append(r)
# Store actions as list of arrays
# e.g. for n_y = 2 -> [ array([ 1., 0.]), array([ 0., 1.]), array([ 0., 1.]), array([ 1., 0.]) ]
action = np.zeros(self.n_y)
action[a] = 1
self.episode_actions.append(action)
def choose_action(self, observation, moves):
"""
Choose action based on observation
Arguments:
observation: array of state, has shape (num_features)
Returns: index of action we want to choose
"""
# Reshape observation to (num_features, 1)
observation = observation[:, np.newaxis]
# Run forward propagation to get softmax probabilities
prob_weights = self.sess.run(self.outputs_softmax, feed_dict={self.X: observation})
for i in range(4):
if not moves[i]:
if prob_weights[0][i] ==1:
for j in range(4):
if i!=j:
prob_weights[0][j] = .33
prob_weights[0][i] = 0
else:
prob_weights[0][i] = 0
prob_weights /= prob_weights.sum()
# Select action using a biased sample
# this will return the index of the action we've sampled
action = np.random.choice(range(len(prob_weights.ravel())), p=prob_weights.ravel())
return action
def learn(self, isLearn):
# Discount and normalize episode reward
discounted_episode_rewards_norm = self.discount_and_norm_rewards()
# Train on episode
self.sess.run(self.train_op, feed_dict={
self.X: np.vstack(self.episode_observations).T,
self.Y: np.vstack(np.array(self.episode_actions)).T,
self.discounted_episode_rewards_norm: discounted_episode_rewards_norm,
})
# Reset the episode data
self.episode_observations, self.episode_actions, self.episode_rewards = [], [], []
# Save checkpoint
if self.save_path is not None and isLearn:
save_path = self.saver.save(self.sess, self.save_path)
print("Model saved in file: %s" % save_path)
return discounted_episode_rewards_norm
def discount_and_norm_rewards(self):
discounted_episode_rewards = np.zeros_like(self.episode_rewards, dtype=np.float64)
cumulative = 0
for t in reversed(range(len(self.episode_rewards))):
cumulative = cumulative * self.gamma + self.episode_rewards[t]
discounted_episode_rewards[t] = cumulative
discounted_episode_rewards -= np.mean(discounted_episode_rewards)
discounted_episode_rewards /= np.std(discounted_episode_rewards)
return discounted_episode_rewards
def build_network(self):
# Create placeholders
with tf.name_scope('inputs'):
self.X = tf.placeholder(tf.float32, shape=(self.n_x, None), name="X")
self.Y = tf.placeholder(tf.float32, shape=(self.n_y, None), name="Y")
self.discounted_episode_rewards_norm = tf.placeholder(tf.float32, [None, ], name="actions_value")
# Initialize parameters
units_layer_1 = 10
units_layer_2 = 10
units_output_layer = self.n_y
with tf.name_scope('parameters'):
W1 = tf.get_variable("W1", [units_layer_1, self.n_x], initializer=tf.contrib.layers.xavier_initializer(seed=1))
b1 = tf.get_variable("b1", [units_layer_1, 1], initializer=tf.contrib.layers.xavier_initializer(seed=1))
W2 = tf.get_variable("W2", [units_layer_2, units_layer_1],
initializer=tf.contrib.layers.xavier_initializer(seed=1))
b2 = tf.get_variable("b2", [units_layer_2, 1], initializer=tf.contrib.layers.xavier_initializer(seed=1))
W3 = tf.get_variable("W3", [self.n_y, units_layer_2], initializer=tf.contrib.layers.xavier_initializer(seed=1))
b3 = tf.get_variable("b3", [self.n_y, 1], initializer=tf.contrib.layers.xavier_initializer(seed=1))
# Forward prop
with tf.name_scope('layer_1'):
Z1 = tf.add(tf.matmul(W1, self.X), b1)
A1 = tf.nn.relu(Z1)
with tf.name_scope('layer_2'):
Z2 = tf.add(tf.matmul(W2, A1), b2)
A2 = tf.nn.relu(Z2)
with tf.name_scope('layer_3'):
Z3 = tf.add(tf.matmul(W3, A2), b3)
A3 = tf.nn.softmax(Z3)
# Softmax outputs, we need to transpose as tensorflow nn functions expects them in this shape
logits = tf.transpose(Z3)
labels = tf.transpose(self.Y)
self.outputs_softmax = tf.nn.softmax(logits, name='A3')
with tf.name_scope('loss'):
neg_log_prob = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels)
loss = tf.reduce_mean(neg_log_prob * self.discounted_episode_rewards_norm) # reward guided loss
with tf.name_scope('train'):
self.train_op = tf.train.AdamOptimizer(self.lr).minimize(loss)
def plot_cost(self):
import matplotlib
matplotlib.use("MacOSX")
import matplotlib.pyplot as plt
plt.plot(np.arange(len(self.cost_history)), self.cost_history)
plt.ylabel('Cost')
plt.xlabel('Training Steps')
plt.show()