-
Notifications
You must be signed in to change notification settings - Fork 3
/
NFSP.py
305 lines (269 loc) · 16.7 KB
/
NFSP.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
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
import numpy as np
import tensorflow as tf
import tensorflow.contrib as tfc
from XFP import XFP, LeducRLEnv
class NFSP(object):
def __init__(self, flags):
self.flags = flags
self.env = LeducRLEnv(card_num=flags.card_num, seed=flags.seed)
np.random.seed(flags.seed)
self.iter = [0, 0]
self.epsilon = self.flags.epsilon
self.sl_replay = [ReservoirReplay(flags, self.env), ReservoirReplay(flags, self.env)]
self.rl_replay = [CircularReplay(flags, self.env), CircularReplay(flags, self.env)]
tf_config = tf.ConfigProto(
allow_soft_placement=True,
inter_op_parallelism_threads=flags.num_cpu,
intra_op_parallelism_threads=flags.num_cpu)
self.device = '/gpu:0' if self.flags.use_gpu else '/cpu:0'
self.graph = tf.Graph()
self.sess = tf.Session(graph=self.graph, config=tf_config)
self.sess.__enter__()
tf.set_random_seed(flags.seed)
with tf.device(self.device):
self.ops = [{}, {}]
for i in range(2):
with tf.variable_scope("player" + str(i)):
self.ops[i]['state_history_ph'] = tf.placeholder(tf.int8, [None] + self.env.state_history_space,
"state_history")
self.ops[i]['state_card_ph'] = tf.placeholder(tf.int8, [None] + self.env.state_card_space,
"state_card")
self.ops[i]['state_history_ph2'] = tf.placeholder(tf.int8, [None] + self.env.state_history_space,
"state_history2")
self.ops[i]['state_card_ph2'] = tf.placeholder(tf.int8, [None] + self.env.state_card_space,
"state_card2")
with tf.variable_scope('current_q'):
self.ops[i]['q_logits_s'] = self._build_inference(self.ops[i]['state_history_ph'], self.ops[i]['state_card_ph'])
with tf.variable_scope('old_q'):
self.ops[i]['q_logits_s2_old'] = self._build_inference(self.ops[i]['state_history_ph2'], self.ops[i]['state_card_ph2'])
with tf.variable_scope('average_pi'):
self.ops[i]['pi_logits_s'] = self._build_inference(self.ops[i]['state_history_ph'], self.ops[i]['state_card_ph'])
assign_ops = []
for (cur, old) in zip(tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope=tf.get_variable_scope().name + '.*current'),
tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope=tf.get_variable_scope().name + '.*old')):
assign_ops.append(tf.assign(old, cur))
self.ops[i]['copy'] = tf.group(*assign_ops, name="copy")
self.ops[i]['global_step'] = tf.get_variable("global_step", [], tf.int64,
tf.constant_initializer(0), trainable=False)
self.ops[i]['action_ph'] = tf.placeholder(tf.int8, [None])
self.ops[i]['reward_ph'] = tf.placeholder(tf.int8, [None])
self.ops[i]['terminal_ph'] = tf.placeholder(tf.int8, [None]) # 1.0 is terminal
self.ops[i]['apply_gradients_sl'], self.ops[i]['apply_gradients_rl'], self.ops[i]['sl_loss'], self.ops[i]['rl_loss'] = \
self._build_train(self.ops[i]['action_ph'],
self.ops[i]['pi_logits_s'],
self.ops[i]['reward_ph'],
self.ops[i]['terminal_ph'],
self.ops[i]['q_logits_s'],
self.ops[i]['q_logits_s2_old'],
self.ops[i]['global_step'])
self.init = tf.global_variables_initializer()
self.sess.run(self.init)
self.sess.run([self.ops[0]['copy'], self.ops[1]['copy']])
def _build_inference(self, state_history_ph, state_card_ph, reuse=False):
state_history = tf.reshape(tf.cast(state_history_ph, tf.float32),
[-1] + [reduce(lambda x, y: x * y, self.env.state_history_space)])
state_card = tf.cast(state_card_ph, tf.float32)
state = tf.concat([state_card, state_history], axis=1)
logits = self._network(state, reuse)
return logits
def _build_train(self, action, pi_logits_s, reward, terminal, q_logits_s, q_logits_s2_old, global_step):
action = tf.cast(action, tf.int32)
one_hot_actions = tf.one_hot(action, self.env.action_space, dtype=tf.float32)
neglog_pi = tf.nn.softmax_cross_entropy_with_logits(logits=pi_logits_s, labels=one_hot_actions)
reward = tf.cast(reward, tf.float32)
terminal = tf.cast(terminal, tf.float32)
target = reward + (1.0 - terminal) * tf.reduce_max(q_logits_s2_old, axis=1)
q_s_a = tf.reduce_sum(q_logits_s * one_hot_actions, axis=1)
loss = tf.reduce_mean(tf.square(q_s_a - tf.stop_gradient(target)))
optimizer_sl = tf.train.GradientDescentOptimizer(self.flags.lr_sl)
optimizer_rl = tf.train.GradientDescentOptimizer(self.flags.lr_rl)
grad_var_list_sl = optimizer_sl.compute_gradients(neglog_pi)
grad_var_list_rl = optimizer_rl.compute_gradients(loss)
apply_gradients_sl = optimizer_sl.apply_gradients(grad_var_list_sl, global_step)
apply_gradients_rl = optimizer_rl.apply_gradients(grad_var_list_rl)
return apply_gradients_sl, apply_gradients_rl, neglog_pi, loss
def _linear_layer(self, linear_in, dim, hiddens):
weights = tf.get_variable('weights', [dim, hiddens], tf.float32,
initializer=tfc.layers.variance_scaling_initializer(mode='FAN_AVG', uniform=True))
bias = tf.get_variable('bias', [hiddens], tf.float32,
initializer=tf.constant_initializer(0.1))
pre_activations = tf.add(tf.matmul(linear_in, weights), bias)
linear_out = tf.nn.relu(pre_activations)
return linear_out
def _network(self, state, reuse=False):
if reuse:
tf.get_variable_scope().reuse_variables()
with tf.variable_scope('linear1'):
dim = state.get_shape().as_list()[1]; hiddens = 64
linear1 = self._linear_layer(state, dim, hiddens)
with tf.variable_scope('linear2'):
dim = hiddens; hiddens = self.env.action_space
logits = self._linear_layer(linear1, dim, hiddens)
return logits
def choose_action(self, position, state_history, state_card):
if np.random.rand() < self.flags.anticipatory: # epsilon greedy
if np.random.rand() < self.epsilon: # explore
action = np.random.randint(0, self.env.action_space)
else:
q_logits_s = self.sess.run(self.ops[position]['q_logits_s'],
feed_dict={self.ops[position]['state_history_ph']: [state_history],
self.ops[position]['state_card_ph']: [state_card]})
action = np.argmax(q_logits_s[0])
return action, 'br' # best response
else:
pi_logits_s = self.sess.run(self.ops[position]['pi_logits_s'],
feed_dict={self.ops[position]['state_history_ph']: [state_history],
self.ops[position]['state_card_ph']: [state_card]})
prob = np.exp(pi_logits_s[0])
prob = prob / np.sum(prob)
action = np.random.choice(self.env.action_space, p=prob)
return action, 'avg' # average
def play_game(self):
ob = self.env.reset()
while True:
position = ob['turn']
if ob['turn'] == -1:
self.rl_replay[0].add_terminal(ob['payoff'][0])
self.rl_replay[1].add_terminal(ob['payoff'][1])
break
state_history = ob['state']
state_card = ob['card']
action, tag = self.choose_action(position, state_history, state_card)
if tag == 'br': # greedy best response
self.sl_replay[position].add(state_history, state_card, action)
self.rl_replay[position].add(state_history, state_card, action, 0, False)
self.iter[position] += 1
if self.iter[position] % self.flags.train_frequency == 0 and self.iter[position] > self.flags.train_start:
self.train(position)
ob = self.env.act(action)
def train(self, position):
batch_state_history_buffer, batch_state_card_buffer, batch_action_buffer = self.sl_replay[position].get_random_batch()
global_step, _ = self.sess.run([self.ops[position]['global_step'], self.ops[position]['apply_gradients_sl']],
feed_dict={self.ops[position]['state_history_ph']: batch_state_history_buffer,
self.ops[position]['state_card_ph']: batch_state_card_buffer,
self.ops[position]['action_ph']: batch_action_buffer})
batch_state_history_buffer, batch_state_card_buffer, batch_action_buffer, \
batch_reward_buffer, batch_terminal_buffer, batch_state_history_buffer2, \
batch_state_card_buffer2 = self.rl_replay[position].get_random_batch()
_ = self.sess.run(self.ops[position]['apply_gradients_rl'],
feed_dict={self.ops[position]['state_history_ph']: batch_state_history_buffer,
self.ops[position]['state_card_ph']: batch_state_card_buffer,
self.ops[position]['action_ph']: batch_action_buffer,
self.ops[position]['reward_ph']: batch_reward_buffer,
self.ops[position]['terminal_ph']: batch_terminal_buffer,
self.ops[position]['state_history_ph2']: batch_state_history_buffer2,
self.ops[position]['state_card_ph2']: batch_state_card_buffer2})
if global_step % self.flags.refit == 0:
self.sess.run(self.ops[position]['copy'])
print 'train policy {:d} at step {:d}, global_step={:d}, epsilon={:.4f}'.format(position, self.iter[position], global_step, self.epsilon)
def compute_self_policy(self):
policy = [{}, {}]
for cards in XFP.possible_cards_list:
for history in LeducRLEnv.history_string2vector:
state_card = np.zeros([self.env.card_num], np.int32)
state_history = LeducRLEnv.history_string2vector[history]
pround = 1 if history in XFP.round1_states_set else 2
if history in XFP.player1_states_set:
state_card[int(cards[0])] = 1
card = cards[0]
if pround == 2:
state_card[int(self.env.card_num) / 2 + int(cards[1])] = 1
card = cards[:2]
if card + history in policy[0]: continue
q_p, pi = self.sess.run([self.ops[0]['q_logits_s'], self.ops[0]['pi_logits_s']],
feed_dict={self.ops[0]['state_history_ph']: [state_history],
self.ops[0]['state_card_ph']: [state_card]})
q_max_a = np.argmax(q_p[0])
prob = np.exp(pi[0])
prob = prob / np.sum(prob)
prob = prob * (1 - self.flags.anticipatory)
prob[q_max_a] += self.flags.anticipatory * 1.0
policy[0][card + history] = prob
else:
state_card[int(cards[2])] = 1
card = cards[2]
if pround == 2:
state_card[int(self.env.card_num) / 2 + int(cards[1])] = 1
card = cards[1:]
if card + history in policy[1]: continue
q_p, pi = self.sess.run([self.ops[1]['q_logits_s'], self.ops[1]['pi_logits_s']],
feed_dict={self.ops[1]['state_history_ph']: [state_history],
self.ops[1]['state_card_ph']: [state_card]})
q_max_a = np.argmax(q_p[0])
prob = np.exp(pi[0])
prob = prob / np.sum(prob)
prob = prob * (1 - self.flags.anticipatory)
prob[q_max_a] += self.flags.anticipatory * 1.0
policy[1][card + history] = prob
return policy
class ReservoirReplay(object): # sl
def __init__(self, flags, env):
self.flags = flags
self.env = env
self.state_history_buffer = np.zeros([self.flags.sl_len] + self.env.state_history_space, np.int8)
self.state_card_buffer = np.zeros([self.flags.sl_len] + self.env.state_card_space, np.int8)
self.action_buffer = np.zeros([self.flags.sl_len], np.int8)
self.size = 0
self.top = 0
def add(self, state_history, state_card, action):
if self.size < self.flags.sl_len:
self.state_history_buffer[self.top] = state_history
self.state_card_buffer[self.top] = state_card
self.action_buffer[self.top] = action
self.top += 1
self.size += 1
else:
prob_add = float(self.flags.sl_len) / float(self.top + 1)
if np.random.rand() < prob_add:
index = np.random.randint(0, self.flags.sl_len)
self.state_history_buffer[index] = state_history
self.state_card_buffer[index] = state_card
self.action_buffer[index] = action
self.top += 1
def get_random_batch(self):
indices = np.random.randint(0, self.size, self.flags.batch)
batch_state_history_buffer = np.take(self.state_history_buffer, indices, axis=0)
batch_state_card_buffer = np.take(self.state_card_buffer, indices, axis=0)
batch_action_buffer = np.take(self.action_buffer, indices)
return batch_state_history_buffer, batch_state_card_buffer, batch_action_buffer
class CircularReplay(object): # rl
def __init__(self, flags, env):
self.flags = flags
self.env = env
self.state_history_buffer = np.zeros([self.flags.rl_len] + self.env.state_history_space, np.int8)
self.state_card_buffer = np.zeros([self.flags.rl_len] + self.env.state_card_space, np.int8)
self.action_buffer = np.zeros([self.flags.rl_len], np.int8)
self.reward_buffer = np.zeros([self.flags.rl_len], np.int8)
self.terminal_buffer = np.zeros([self.flags.rl_len], np.int8)
self.size = 0
self.top = 0
self.bottom = 0
def add(self, state_history, state_card, action, reward, terminal):
self.state_history_buffer[self.top] = state_history
self.state_card_buffer[self.top] = state_card
self.action_buffer[self.top] = action
self.reward_buffer[self.top] = reward
self.terminal_buffer[self.top] = terminal
if self.size == self.flags.rl_len:
self.bottom = (self.bottom + 1) % self.flags.rl_len
else:
self.size += 1
self.top = (self.top + 1) % self.flags.rl_len
def add_terminal(self, reward):
last_top = (self.top - 1) % self.flags.rl_len
self.reward_buffer[last_top] = reward
self.terminal_buffer[last_top] = True
def get_random_batch(self):
indices = np.random.randint(0, self.size, self.flags.batch)
indices2 = indices + 1
batch_state_history_buffer = np.take(self.state_history_buffer, indices, axis=0)
batch_state_card_buffer = np.take(self.state_card_buffer, indices, axis=0)
batch_state_history_buffer2 = np.take(self.state_history_buffer, indices2, axis=0, mode='wrap')
batch_state_card_buffer2 = np.take(self.state_card_buffer, indices2, axis=0, mode='wrap')
batch_action_buffer = np.take(self.action_buffer, indices)
batch_reward_buffer = np.take(self.reward_buffer, indices)
batch_terminal_buffer = np.take(self.terminal_buffer, indices)
return batch_state_history_buffer, batch_state_card_buffer, batch_action_buffer, batch_reward_buffer, \
batch_terminal_buffer, batch_state_history_buffer2, batch_state_card_buffer2