-
Notifications
You must be signed in to change notification settings - Fork 12
/
Copy pathppo.py
202 lines (159 loc) · 8.21 KB
/
ppo.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
from keras.models import Model, model_from_json, load_model
from keras.optimizers import Adam, RMSprop
import os
from keras.layers import Input, Dense
import keras.backend as K
import time
from copy import deepcopy
import numpy as np
class Memory:
def __init__(self):
self.batch_s = []
self.batch_a = []
self.batch_r = []
self.batch_s_ = []
self.batch_done = []
def store(self, s, a, s_, r, done):
self.batch_s.append(s)
self.batch_a.append(a)
self.batch_r.append(r)
self.batch_s_.append(s_)
self.batch_done.append(done)
def clear(self):
self.batch_s.clear()
self.batch_a.clear()
self.batch_r.clear()
self.batch_s_.clear()
self.batch_done.clear()
@property
def cnt_samples(self):
return len(self.batch_s)
class Agent:
def __init__(self, dic_agent_conf, dic_path, dic_env_conf):
self.dic_agent_conf = dic_agent_conf
self.dic_path = dic_path
self.dic_env_conf = dic_env_conf
self.n_actions = self.dic_agent_conf["ACTION_DIM"]
self.actor_network = self._build_actor_network()
self.actor_old_network = self.build_network_from_copy(self.actor_network)
self.critic_network = self._build_critic_network()
self.dummy_advantage = np.zeros((1, 1))
self.dummy_old_prediction = np.zeros((1, self.n_actions))
self.memory = Memory()
def choose_action(self, state):
assert isinstance(state, np.ndarray), "state must be numpy.ndarry"
state = np.reshape(state, [-1, self.dic_agent_conf["STATE_DIM"][0]])
prob = self.actor_network.predict_on_batch([state, self.dummy_advantage, self.dummy_old_prediction]).flatten()
action = np.random.choice(self.n_actions, p=prob)
return action
def train_network(self):
n = self.memory.cnt_samples
discounted_r = []
if self.memory.batch_done[-1]:
v = 0
else:
v = self.get_v(self.memory.batch_s_[-1])
for r in self.memory.batch_r[::-1]:
v = r + self.dic_agent_conf["GAMMA"] * v
discounted_r.append(v)
discounted_r.reverse()
batch_s, batch_a, batch_discounted_r = np.vstack(self.memory.batch_s), \
np.vstack(self.memory.batch_a), \
np.vstack(discounted_r)
batch_v = self.get_v(batch_s)
batch_advantage = batch_discounted_r - batch_v
batch_old_prediction = self.get_old_prediction(batch_s)
batch_a_final = np.zeros(shape=(len(batch_a), self.n_actions))
batch_a_final[:, batch_a.flatten()] = 1
# print(batch_s.shape, batch_advantage.shape, batch_old_prediction.shape, batch_a_final.shape)
self.actor_network.fit(x=[batch_s, batch_advantage, batch_old_prediction], y=batch_a_final, verbose=0)
self.critic_network.fit(x=batch_s, y=batch_discounted_r, epochs=2, verbose=0)
self.memory.clear()
self.update_target_network()
def get_old_prediction(self, s):
s = np.reshape(s, (-1, self.dic_agent_conf["STATE_DIM"][0]))
return self.actor_old_network.predict_on_batch(s)
def store_transition(self, s, a, s_, r, done):
self.memory.store(s, a, s_, r, done)
def get_v(self, s):
s = np.reshape(s, (-1, self.dic_agent_conf["STATE_DIM"][0]))
v = self.critic_network.predict_on_batch(s)
return v
def save_model(self, file_name):
self.actor_network.save(os.path.join(self.dic_path["PATH_TO_MODEL"], "%s_actor_network.h5" % file_name))
self.critic_network.save(os.path.join(self.dic_path["PATH_TO_MODEL"], "%s_critic_network.h5" % file_name))
def load_model(self):
self.actor_network = load_model(self.dic_path["PATH_TO_MODEL"], "%s_actor_network.h5")
self.critic_network = load_model(self.dic_path["PATH_TO_MODEL"], "%s_critic_network.h5")
self.actor_old_network = deepcopy(self.actor_network)
def _build_actor_network(self):
state = Input(shape=self.dic_agent_conf["STATE_DIM"], name="state")
advantage = Input(shape=(1, ), name="Advantage")
old_prediction = Input(shape=(self.n_actions,), name="Old_Prediction")
shared_hidden = self._shared_network_structure(state)
action_dim = self.dic_agent_conf["ACTION_DIM"]
policy = Dense(action_dim, activation="softmax", name="actor_output_layer")(shared_hidden)
actor_network = Model(inputs=[state, advantage, old_prediction], outputs=policy)
if self.dic_agent_conf["OPTIMIZER"] is "Adam":
actor_network.compile(optimizer=Adam(lr=self.dic_agent_conf["ACTOR_LEARNING_RATE"]),
loss=self.proximal_policy_optimization_loss(
advantage=advantage, old_prediction=old_prediction,
))
elif self.dic_agent_conf["OPTIMIZER"] is "RMSProp":
actor_network.compile(optimizer=RMSprop(lr=self.dic_agent_conf["ACTOR_LEARNING_RATE"]))
else:
print("Not such optimizer for actor network. Instead, we use adam optimizer")
actor_network.compile(optimizer=Adam(lr=self.dic_agent_conf["ACTOR_LEARNING_RATE"]))
print("=== Build Actor Network ===")
actor_network.summary()
time.sleep(1.0)
return actor_network
def update_target_network(self):
alpha = self.dic_agent_conf["TARGET_UPDATE_ALPHA"]
self.actor_old_network.set_weights(alpha*np.array(self.actor_network.get_weights())
+ (1-alpha)*np.array(self.actor_old_network.get_weights()))
def _build_critic_network(self):
state = Input(shape=self.dic_agent_conf["STATE_DIM"], name="state")
shared_hidden = self._shared_network_structure(state)
if self.dic_env_conf["POSITIVE_REWARD"]:
q = Dense(1, activation="relu", name="critic_output_layer")(shared_hidden)
else:
q = Dense(1, name="critic_output_layer")(shared_hidden)
critic_network = Model(inputs=state, outputs=q)
if self.dic_agent_conf["OPTIMIZER"] is "Adam":
critic_network.compile(optimizer=Adam(lr=self.dic_agent_conf["ACTOR_LEARNING_RATE"]),
loss=self.dic_agent_conf["CRITIC_LOSS"])
elif self.dic_agent_conf["OPTIMIZER"] is "RMSProp":
critic_network.compile(optimizer=RMSprop(lr=self.dic_agent_conf["ACTOR_LEARNING_RATE"]),
loss=self.dic_agent_conf["CRITIC_LOSS"])
else:
print("Not such optimizer for actor network. Instead, we use adam optimizer")
critic_network.compile(optimizer=Adam(lr=self.dic_agent_conf["ACTOR_LEARNING_RATE"]),
loss=self.dic_agent_conf["CRITIC_LOSS"])
print("=== Build Critic Network ===")
critic_network.summary()
time.sleep(1.0)
return critic_network
def build_network_from_copy(self, actor_network):
network_structure = actor_network.to_json()
network_weights = actor_network.get_weights()
network = model_from_json(network_structure)
network.set_weights(network_weights)
network.compile(optimizer=Adam(lr=self.dic_agent_conf["ACTOR_LEARNING_RATE"]), loss="mse")
return network
def _shared_network_structure(self, state_features):
dense_d = self.dic_agent_conf["D_DENSE"]
hidden1 = Dense(dense_d, activation="relu", name="hidden_shared_1")(state_features)
hidden2 = Dense(dense_d, activation="relu", name="hidden_shared_2")(hidden1)
return hidden2
def proximal_policy_optimization_loss(self, advantage, old_prediction):
loss_clipping = self.dic_agent_conf["CLIPPING_LOSS_RATIO"]
entropy_loss = self.dic_agent_conf["ENTROPY_LOSS_RATIO"]
def loss(y_true, y_pred):
prob = y_true * y_pred
old_prob = y_true * old_prediction
r = prob / (old_prob + 1e-10)
return -K.mean(K.minimum(r * advantage, K.clip(r, min_value=1 - loss_clipping,
max_value=1 + loss_clipping) * advantage) + entropy_loss * (
prob * K.log(prob + 1e-10)))
return loss