強化学習 x ニューラルネットワーク 8 (A2C)

パラメータを持った関数で戦略を実装します。攻略する環境はCartPoleです。

まずは親クラスとなるフレームワークを作成します。(前回のソースと同じです。)

fn_framework.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
import os
import io
import re
from collections import namedtuple
from collections import deque
import numpy as np
import tensorflow as tf
from tensorflow.python import keras as K
from PIL import Image
import matplotlib.pyplot as plt

# s:状態
# a:行動
# r:報酬
# n_s:遷移先の状態
# d:エピソード終了フラグ
Experience = namedtuple("Experience",
["s", "a", "r", "n_s", "d"])

# ニューラルネットワークを使い状態から評価を行う。
class FNAgent():

def __init__(self, epsilon, actions):
self.epsilon = epsilon
self.actions = actions
self.model = None
self.estimate_probs = False
self.initialized = False

# 学習したエージェントを保存
def save(self, model_path):
self.model.save(model_path, overwrite=True, include_optimizer=False)

# 学習したエージェントを読み込み
@classmethod
def load(cls, env, model_path, epsilon=0.0001):
actions = list(range(env.action_space.n))
agent = cls(epsilon, actions)
agent.model = K.models.load_model(model_path)
agent.initialized = True
return agent

# 初期化
# experiences:エージェントの経験
def initialize(self, experiences):
raise Exception("You have to implements estimate method.")

# 関数による予測
def estimate(self, s):
raise Exception("You have to implements estimate method.")

# パラメータの更新
def update(self, experiences, gamma):
raise Exception("You have to implements update method.")

def policy(self, s):
if np.random.random() < self.epsilon or not self.initialized:
return np.random.randint(len(self.actions))
else:
estimates = self.estimate(s)
if self.estimate_probs:
action = np.random.choice(self.actions,
size=1, p=estimates)[0]
return action
else:
return np.argmax(estimates)

def play(self, env, episode_count=5, render=True):
for e in range(episode_count):
s = env.reset()
done = False
episode_reward = 0
while not done:
if render:
env.render()
a = self.policy(s)
n_state, reward, done, info = env.step(a)
episode_reward += reward
s = n_state
else:
print("Get reward {}.".format(episode_reward))

# エージェントの学習を行う
class Trainer():

def __init__(self, buffer_size=1024, batch_size=32,
gamma=0.9, report_interval=10, log_dir=""):
self.buffer_size = buffer_size
self.batch_size = batch_size
self.gamma = gamma
self.report_interval = report_interval
self.logger = Logger(log_dir, self.trainer_name)
# エージェントの行動履歴(古い行動からすてる)
self.experiences = deque(maxlen=buffer_size)
self.training = False
self.training_count = 0
self.reward_log = []

@property
def trainer_name(self):
class_name = self.__class__.__name__
snaked = re.sub("(.)([A-Z][a-z]+)", r"\1_\2", class_name)
snaked = re.sub("([a-z0-9])([A-Z])", r"\1_\2", snaked).lower()
snaked = snaked.replace("_trainer", "")
return snaked

def train_loop(self, env, agent, episode=200, initial_count=-1,
render=False, observe_interval=0):
self.experiences = deque(maxlen=self.buffer_size)
self.training = False
self.training_count = 0
self.reward_log = []
frames = []

for i in range(episode):
s = env.reset()
done = False
step_count = 0
self.episode_begin(i, agent)
while not done:
if render:
env.render()
if self.training and observe_interval > 0 and\
(self.training_count == 1 or
self.training_count % observe_interval == 0):
frames.append(s)

a = agent.policy(s)
n_state, reward, done, info = env.step(a)
e = Experience(s, a, reward, n_state, done)
self.experiences.append(e)
if not self.training and \
len(self.experiences) == self.buffer_size:
self.begin_train(i, agent)
self.training = True

self.step(i, step_count, agent, e)

s = n_state
step_count += 1
else:
self.episode_end(i, step_count, agent)

if not self.training and \
initial_count > 0 and i >= initial_count:
self.begin_train(i, agent)
self.training = True

if self.training:
if len(frames) > 0:
self.logger.write_image(self.training_count,
frames)
frames = []
self.training_count += 1

def episode_begin(self, episode, agent):
pass

def begin_train(self, episode, agent):
pass

def step(self, episode, step_count, agent, experience):
pass

def episode_end(self, episode, step_count, agent):
pass

def is_event(self, count, interval):
return True if count != 0 and count % interval == 0 else False

def get_recent(self, count):
recent = range(len(self.experiences) - count, len(self.experiences))
return [self.experiences[i] for i in recent]

# 環境から取得される「状態」の前処理を行う
class Observer():

def __init__(self, env):
self._env = env

@property
def action_space(self):
return self._env.action_space

@property
def observation_space(self):
return self._env.observation_space

def reset(self):
return self.transform(self._env.reset())

def render(self):
self._env.render()

def step(self, action):
n_state, reward, done, info = self._env.step(action)
return self.transform(n_state), reward, done, info

def transform(self, state):
raise Exception("You have to implements transform method.")

# 学習経過の記録を行う
class Logger():

def __init__(self, log_dir="", dir_name=""):
self.log_dir = log_dir
if not log_dir:
self.log_dir = os.path.join(os.path.dirname(__file__), "logs")
if not os.path.exists(self.log_dir):
os.mkdir(self.log_dir)

if dir_name:
self.log_dir = os.path.join(self.log_dir, dir_name)
if not os.path.exists(self.log_dir):
os.mkdir(self.log_dir)

self._callback = K.callbacks.TensorBoard(self.log_dir)

@property
def writer(self):
return self._callback.writer

def set_model(self, model):
self._callback.set_model(model)

def path_of(self, file_name):
return os.path.join(self.log_dir, file_name)

def describe(self, name, values, episode=-1, step=-1):
mean = np.round(np.mean(values), 3)
std = np.round(np.std(values), 3)
desc = "{} is {} (+/-{})".format(name, mean, std)
if episode > 0:
print("At episode {}, {}".format(episode, desc))
elif step > 0:
print("At step {}, {}".format(step, desc))

def plot(self, name, values, interval=10):
indices = list(range(0, len(values), interval))
means = []
stds = []
for i in indices:
_values = values[i:(i + interval)]
means.append(np.mean(_values))
stds.append(np.std(_values))
means = np.array(means)
stds = np.array(stds)
plt.figure()
plt.title("{} History".format(name))
plt.grid()
plt.fill_between(indices, means - stds, means + stds,
alpha=0.1, color="g")
plt.plot(indices, means, "o-", color="g",
label="{} per {} episode".format(name.lower(), interval))
plt.legend(loc="best")
plt.show()

def write(self, index, name, value):
summary = tf.Summary()
summary_value = summary.value.add()
summary_value.tag = name
summary_value.simple_value = value
self.writer.add_summary(summary, index)
self.writer.flush()

def write_image(self, index, frames):
# Deal with a 'frames' as a list of sequential gray scaled image.
last_frames = [f[:, :, -1] for f in frames]
if np.min(last_frames[-1]) < 0:
scale = 127 / np.abs(last_frames[-1]).max()
offset = 128
else:
scale = 255 / np.max(last_frames[-1])
offset = 0
channel = 1 # gray scale
tag = "frames_at_training_{}".format(index)
values = []

for f in last_frames:
height, width = f.shape
array = np.asarray(f * scale + offset, dtype=np.uint8)
image = Image.fromarray(array)
output = io.BytesIO()
image.save(output, format="PNG")
image_string = output.getvalue()
output.close()
image = tf.Summary.Image(
height=height, width=width, colorspace=channel,
encoded_image_string=image_string)
value = tf.Summary.Value(tag=tag, image=image)
values.append(value)

summary = tf.Summary(value=values)
self.writer.add_summary(summary, index)
self.writer.flush()

戦略に深層学習を適用したAdvantage Actor Critic(A2C)で実装します。

a2c_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
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
304
305
306
307
308
309
310
311
312
313
import random
import argparse
from collections import deque
import numpy as np
import tensorflow as tf
from tensorflow.python import keras as K
from PIL import Image
import gym
import gym_ple
from fn_framework import FNAgent, Trainer, Observer, Experience

class ActorCriticAgent(FNAgent):

def __init__(self, epsilon, actions):
super().__init__(epsilon, actions)
self._updater = None

@classmethod
def load(cls, env, model_path, epsilon=0.0001):
actions = list(range(env.action_space.n))
agent = cls(epsilon, actions)
agent.model = K.models.load_model(model_path, custom_objects={
"SampleLayer": SampleLayer})
agent.initialized = True
return agent

def initialize(self, experiences, optimizer):
feature_shape = experiences[0].s.shape
self.make_model(feature_shape)
self.set_updater(optimizer)
self.initialized = True

def make_model(self, feature_shape):
normal = K.initializers.glorot_normal()
model = K.Sequential()
model.add(K.layers.Conv2D(
32, kernel_size=8, strides=4, padding="same",
input_shape=feature_shape,
kernel_initializer=normal, activation="relu"))
model.add(K.layers.Conv2D(
64, kernel_size=4, strides=2, padding="same",
kernel_initializer=normal, activation="relu"))
model.add(K.layers.Conv2D(
64, kernel_size=3, strides=1, padding="same",
kernel_initializer=normal, activation="relu"))
model.add(K.layers.Flatten())
model.add(K.layers.Dense(256, kernel_initializer=normal,
activation="relu"))

actor_layer = K.layers.Dense(len(self.actions),
kernel_initializer=normal)
action_evals = actor_layer(model.output)
actions = SampleLayer()(action_evals)

critic_layer = K.layers.Dense(1, kernel_initializer=normal)
values = critic_layer(model.output)

self.model = K.Model(inputs=model.input,
outputs=[actions, action_evals, values])

def set_updater(self, optimizer,
value_loss_weight=1.0, entropy_weight=0.1):
actions = tf.placeholder(shape=(None), dtype="int32")
rewards = tf.placeholder(shape=(None), dtype="float32")

_, action_evals, values = self.model.output

neg_logs = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=action_evals, labels=actions)
advantages = rewards - values

policy_loss = tf.reduce_mean(neg_logs * tf.nn.softplus(advantages))
value_loss = tf.losses.mean_squared_error(rewards, values)
action_entropy = tf.reduce_mean(self.categorical_entropy(action_evals))

loss = policy_loss + value_loss_weight * value_loss
loss -= entropy_weight * action_entropy

updates = optimizer.get_updates(loss=loss,
params=self.model.trainable_weights)

self._updater = K.backend.function(
inputs=[self.model.input,
actions, rewards],
outputs=[loss,
policy_loss,
tf.reduce_mean(neg_logs),
tf.reduce_mean(advantages),
value_loss,
action_entropy],
updates=updates)

def categorical_entropy(self, logits):
"""
From OpenAI baseline implementation
https://github.com/openai/baselines/blob/master/baselines/common/distributions.py#L192
"""
a0 = logits - tf.reduce_max(logits, axis=-1, keepdims=True)
ea0 = tf.exp(a0)
z0 = tf.reduce_sum(ea0, axis=-1, keepdims=True)
p0 = ea0 / z0
return tf.reduce_sum(p0 * (tf.log(z0) - a0), axis=-1)

def policy(self, s):
if np.random.random() < self.epsilon or not self.initialized:
return np.random.randint(len(self.actions))
else:
action, action_evals, values = self.model.predict(np.array([s]))
return action[0]

def estimate(self, s):
action, action_evals, values = self.model.predict(np.array([s]))
return values[0][0]

def update(self, states, actions, rewards):
return self._updater([states, actions, rewards])

class SampleLayer(K.layers.Layer):

def __init__(self, **kwargs):
self.output_dim = 1 # sample one action from evaluations
super(SampleLayer, self).__init__(**kwargs)

def build(self, input_shape):
super(SampleLayer, self).build(input_shape)

def call(self, x):
noise = tf.random_uniform(tf.shape(x))
return tf.argmax(x - tf.log(-tf.log(noise)), axis=1)

def compute_output_shape(self, input_shape):
return (input_shape[0], self.output_dim)

class ActorCriticAgentTest(ActorCriticAgent):

def make_model(self, feature_shape):
normal = K.initializers.glorot_normal()
model = K.Sequential()
model.add(K.layers.Dense(64, input_shape=feature_shape,
kernel_initializer=normal, activation="relu"))
model.add(K.layers.Dense(64, kernel_initializer=normal,
activation="relu"))

actor_layer = K.layers.Dense(len(self.actions),
kernel_initializer=normal)

action_evals = actor_layer(model.output)
actions = SampleLayer()(action_evals)

critic_layer = K.layers.Dense(1, kernel_initializer=normal)
values = critic_layer(model.output)

self.model = K.Model(inputs=model.input,
outputs=[actions, action_evals, values])

class CatcherObserver(Observer):

def __init__(self, env, width, height, frame_count):
super().__init__(env)
self.width = width
self.height = height
self.frame_count = frame_count
self._frames = deque(maxlen=frame_count)

def transform(self, state):
grayed = Image.fromarray(state).convert("L")
resized = grayed.resize((self.width, self.height))
resized = np.array(resized).astype("float")
normalized = resized / 255.0 # scale to 0~1
if len(self._frames) == 0:
for i in range(self.frame_count):
self._frames.append(normalized)
else:
self._frames.append(normalized)
feature = np.array(self._frames)
# Convert the feature shape (f, w, h) => (w, h, f).
feature = np.transpose(feature, (1, 2, 0))
return feature

class ActorCriticTrainer(Trainer):

def __init__(self, buffer_size=50000, batch_size=32,
gamma=0.99, initial_epsilon=0.1, final_epsilon=1e-3,
learning_rate=1e-3, report_interval=10,
log_dir="", file_name=""):
super().__init__(buffer_size, batch_size, gamma,
report_interval, log_dir)
self.file_name = file_name if file_name else "a2c_agent.h5"
self.initial_epsilon = initial_epsilon
self.final_epsilon = final_epsilon
self.learning_rate = learning_rate
self.d_experiences = deque(maxlen=self.buffer_size)
self.training_episode = 0
self.losses = {}
self._max_reward = -10

def train(self, env, episode_count=900, initial_count=10,
test_mode=False, render=False, observe_interval=100):
actions = list(range(env.action_space.n))
if not test_mode:
agent = ActorCriticAgent(1.0, actions)
else:
agent = ActorCriticAgentTest(1.0, actions)
observe_interval = 0
self.training_episode = episode_count

self.train_loop(env, agent, episode_count, initial_count, render,
observe_interval)
return agent

def episode_begin(self, episode, agent):
self.losses = {}
for key in ["loss", "loss_policy", "loss_action", "loss_advantage",
"loss_value", "entropy"]:
self.losses[key] = []
self.experiences = []

def step(self, episode, step_count, agent, experience):
if self.training:
loss, lp, ac, ad, vl, en = agent.update(*self.make_batch())
self.losses["loss"].append(loss)
self.losses["loss_policy"].append(lp)
self.losses["loss_action"].append(ac)
self.losses["loss_advantage"].append(ad)
self.losses["loss_value"].append(vl)
self.losses["entropy"].append(en)

def make_batch(self):
batch = random.sample(self.d_experiences, self.batch_size)
states = [e.s for e in batch]
actions = [e.a for e in batch]
rewards = [e.r for e in batch]
return states, actions, rewards

def begin_train(self, episode, agent):
self.logger.set_model(agent.model)
agent.epsilon = self.initial_epsilon
self.training_episode -= episode
print("Done initialization. From now, begin training!")

def episode_end(self, episode, step_count, agent):
rewards = [e.r for e in self.experiences]
self.reward_log.append(sum(rewards))

if not agent.initialized:
optimizer = K.optimizers.Adam(lr=self.learning_rate, clipnorm=5.0)
agent.initialize(self.experiences, optimizer)

discounteds = []
for t, r in enumerate(rewards):
future_r = [_r * (self.gamma ** i) for i, _r in
enumerate(rewards[t:])]
_r = sum(future_r)
discounteds.append(_r)

for i, e in enumerate(self.experiences):
s, a, r, n_s, d = e
d_r = discounteds[i]
d_e = Experience(s, a, d_r, n_s, d)
self.d_experiences.append(d_e)

if not self.training and len(self.d_experiences) == self.buffer_size:
self.begin_train(i, agent)
self.training = True

if self.training:
reward = sum(rewards)
self.logger.write(self.training_count, "reward", reward)
self.logger.write(self.training_count, "reward_max", max(rewards))
self.logger.write(self.training_count, "epsilon", agent.epsilon)
for k in self.losses:
loss = sum(self.losses[k]) / step_count
self.logger.write(self.training_count, "loss/" + k, loss)
if reward > self._max_reward:
agent.save(self.logger.path_of(self.file_name))
self._max_reward = reward

diff = (self.initial_epsilon - self.final_epsilon)
decay = diff / self.training_episode
agent.epsilon = max(agent.epsilon - decay, self.final_epsilon)

if self.is_event(episode, self.report_interval):
recent_rewards = self.reward_log[-self.report_interval:]
self.logger.describe("reward", recent_rewards, episode=episode)

def main(play, is_test):
file_name = "a2c_agent.h5" if not is_test else "a2c_agent_test.h5"
trainer = ActorCriticTrainer(file_name=file_name)
path = trainer.logger.path_of(trainer.file_name)
agent_class = ActorCriticAgent

if is_test:
print("Train on test mode")
obs = gym.make("CartPole-v0")
agent_class = ActorCriticAgentTest
else:
env = gym.make("Catcher-v0")
obs = CatcherObserver(env, 80, 80, 4)
trainer.learning_rate = 7e-5

if play:
agent = agent_class.load(obs, path)
agent.play(obs, episode_count=10, render=True)
else:
trainer.train(obs, test_mode=is_test)

if __name__ == "__main__":
parser = argparse.ArgumentParser(description="A2C Agent")
parser.add_argument("--play", action="store_true", help="play with trained model")
parser.add_argument("--test", action="store_true", help="train by test mode")

args = parser.parse_args()
main(args.play, args.test)

テスト用のCartPoleを実行してみます。
まずは学習学習です。

1
python a2c_agent.py --test

結果(後半部)
あまり報酬が増えてないような気もしますが、この学習データを使ってゲームをプレイします。

1
python a2c_agent.py --test --play

結果(コンソール)
プレイしてる様子は下記の動画で確認できます。

・・・もうちょっと頑張ってほしいところです。

強化学習 x ニューラルネットワーク 7 (Policy Gradient)

パラメータを持った関数で戦略を実装します。攻略する環境はCartPoleです。

まずは親クラスとなるフレームワークを作成します。(前回のCatcherと同じです。)

fn_framework.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
import os
import io
import re
from collections import namedtuple
from collections import deque
import numpy as np
import tensorflow as tf
from tensorflow.python import keras as K
from PIL import Image
import matplotlib.pyplot as plt

# s:状態
# a:行動
# r:報酬
# n_s:遷移先の状態
# d:エピソード終了フラグ
Experience = namedtuple("Experience",
["s", "a", "r", "n_s", "d"])

# ニューラルネットワークを使い状態から評価を行う。
class FNAgent():

def __init__(self, epsilon, actions):
self.epsilon = epsilon
self.actions = actions
self.model = None
self.estimate_probs = False
self.initialized = False

# 学習したエージェントを保存
def save(self, model_path):
self.model.save(model_path, overwrite=True, include_optimizer=False)

# 学習したエージェントを読み込み
@classmethod
def load(cls, env, model_path, epsilon=0.0001):
actions = list(range(env.action_space.n))
agent = cls(epsilon, actions)
agent.model = K.models.load_model(model_path)
agent.initialized = True
return agent

# 初期化
# experiences:エージェントの経験
def initialize(self, experiences):
raise Exception("You have to implements estimate method.")

# 関数による予測
def estimate(self, s):
raise Exception("You have to implements estimate method.")

# パラメータの更新
def update(self, experiences, gamma):
raise Exception("You have to implements update method.")

def policy(self, s):
if np.random.random() < self.epsilon or not self.initialized:
return np.random.randint(len(self.actions))
else:
estimates = self.estimate(s)
if self.estimate_probs:
action = np.random.choice(self.actions,
size=1, p=estimates)[0]
return action
else:
return np.argmax(estimates)

def play(self, env, episode_count=5, render=True):
for e in range(episode_count):
s = env.reset()
done = False
episode_reward = 0
while not done:
if render:
env.render()
a = self.policy(s)
n_state, reward, done, info = env.step(a)
episode_reward += reward
s = n_state
else:
print("Get reward {}.".format(episode_reward))

# エージェントの学習を行う
class Trainer():

def __init__(self, buffer_size=1024, batch_size=32,
gamma=0.9, report_interval=10, log_dir=""):
self.buffer_size = buffer_size
self.batch_size = batch_size
self.gamma = gamma
self.report_interval = report_interval
self.logger = Logger(log_dir, self.trainer_name)
# エージェントの行動履歴(古い行動からすてる)
self.experiences = deque(maxlen=buffer_size)
self.training = False
self.training_count = 0
self.reward_log = []

@property
def trainer_name(self):
class_name = self.__class__.__name__
snaked = re.sub("(.)([A-Z][a-z]+)", r"\1_\2", class_name)
snaked = re.sub("([a-z0-9])([A-Z])", r"\1_\2", snaked).lower()
snaked = snaked.replace("_trainer", "")
return snaked

def train_loop(self, env, agent, episode=200, initial_count=-1,
render=False, observe_interval=0):
self.experiences = deque(maxlen=self.buffer_size)
self.training = False
self.training_count = 0
self.reward_log = []
frames = []

for i in range(episode):
s = env.reset()
done = False
step_count = 0
self.episode_begin(i, agent)
while not done:
if render:
env.render()
if self.training and observe_interval > 0 and\
(self.training_count == 1 or
self.training_count % observe_interval == 0):
frames.append(s)

a = agent.policy(s)
n_state, reward, done, info = env.step(a)
e = Experience(s, a, reward, n_state, done)
self.experiences.append(e)
if not self.training and \
len(self.experiences) == self.buffer_size:
self.begin_train(i, agent)
self.training = True

self.step(i, step_count, agent, e)

s = n_state
step_count += 1
else:
self.episode_end(i, step_count, agent)

if not self.training and \
initial_count > 0 and i >= initial_count:
self.begin_train(i, agent)
self.training = True

if self.training:
if len(frames) > 0:
self.logger.write_image(self.training_count,
frames)
frames = []
self.training_count += 1

def episode_begin(self, episode, agent):
pass

def begin_train(self, episode, agent):
pass

def step(self, episode, step_count, agent, experience):
pass

def episode_end(self, episode, step_count, agent):
pass

def is_event(self, count, interval):
return True if count != 0 and count % interval == 0 else False

def get_recent(self, count):
recent = range(len(self.experiences) - count, len(self.experiences))
return [self.experiences[i] for i in recent]

# 環境から取得される「状態」の前処理を行う
class Observer():

def __init__(self, env):
self._env = env

@property
def action_space(self):
return self._env.action_space

@property
def observation_space(self):
return self._env.observation_space

def reset(self):
return self.transform(self._env.reset())

def render(self):
self._env.render()

def step(self, action):
n_state, reward, done, info = self._env.step(action)
return self.transform(n_state), reward, done, info

def transform(self, state):
raise Exception("You have to implements transform method.")

# 学習経過の記録を行う
class Logger():

def __init__(self, log_dir="", dir_name=""):
self.log_dir = log_dir
if not log_dir:
self.log_dir = os.path.join(os.path.dirname(__file__), "logs")
if not os.path.exists(self.log_dir):
os.mkdir(self.log_dir)

if dir_name:
self.log_dir = os.path.join(self.log_dir, dir_name)
if not os.path.exists(self.log_dir):
os.mkdir(self.log_dir)

self._callback = K.callbacks.TensorBoard(self.log_dir)

@property
def writer(self):
return self._callback.writer

def set_model(self, model):
self._callback.set_model(model)

def path_of(self, file_name):
return os.path.join(self.log_dir, file_name)

def describe(self, name, values, episode=-1, step=-1):
mean = np.round(np.mean(values), 3)
std = np.round(np.std(values), 3)
desc = "{} is {} (+/-{})".format(name, mean, std)
if episode > 0:
print("At episode {}, {}".format(episode, desc))
elif step > 0:
print("At step {}, {}".format(step, desc))

def plot(self, name, values, interval=10):
indices = list(range(0, len(values), interval))
means = []
stds = []
for i in indices:
_values = values[i:(i + interval)]
means.append(np.mean(_values))
stds.append(np.std(_values))
means = np.array(means)
stds = np.array(stds)
plt.figure()
plt.title("{} History".format(name))
plt.grid()
plt.fill_between(indices, means - stds, means + stds,
alpha=0.1, color="g")
plt.plot(indices, means, "o-", color="g",
label="{} per {} episode".format(name.lower(), interval))
plt.legend(loc="best")
plt.show()

def write(self, index, name, value):
summary = tf.Summary()
summary_value = summary.value.add()
summary_value.tag = name
summary_value.simple_value = value
self.writer.add_summary(summary, index)
self.writer.flush()

def write_image(self, index, frames):
# Deal with a 'frames' as a list of sequential gray scaled image.
last_frames = [f[:, :, -1] for f in frames]
if np.min(last_frames[-1]) < 0:
scale = 127 / np.abs(last_frames[-1]).max()
offset = 128
else:
scale = 255 / np.max(last_frames[-1])
offset = 0
channel = 1 # gray scale
tag = "frames_at_training_{}".format(index)
values = []

for f in last_frames:
height, width = f.shape
array = np.asarray(f * scale + offset, dtype=np.uint8)
image = Image.fromarray(array)
output = io.BytesIO()
image.save(output, format="PNG")
image_string = output.getvalue()
output.close()
image = tf.Summary.Image(
height=height, width=width, colorspace=channel,
encoded_image_string=image_string)
value = tf.Summary.Value(tag=tag, image=image)
values.append(value)

summary = tf.Summary(value=values)
self.writer.add_summary(summary, index)
self.writer.flush()

上記の親クラスを継承し、パラメータを持った関数として戦略を実装していきます。

policy_gradient_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
import os
import argparse
import random
from collections import deque
import numpy as np
from sklearn.preprocessing import StandardScaler
from sklearn.externals import joblib
import tensorflow as tf
from tensorflow.python import keras as K
import gym
from fn_framework import FNAgent, Trainer, Observer, Experience

class PolicyGradientAgent(FNAgent):

def __init__(self, epsilon, actions):
super().__init__(epsilon, actions)
self.estimate_probs = True
self.scaler = None
self._updater = None

def save(self, model_path):
super().save(model_path)
joblib.dump(self.scaler, self.scaler_path(model_path))

@classmethod
def load(cls, env, model_path, epsilon=0.0001):
agent = super().load(env, model_path, epsilon)
agent.scaler = joblib.load(agent.scaler_path(model_path))
return agent

def scaler_path(self, model_path):
fname, _ = os.path.splitext(model_path)
fname += "_scaler.pkl"
return fname

def initialize(self, experiences, optimizer):
self.scaler = StandardScaler()
states = np.vstack([e.s for e in experiences])
self.scaler.fit(states)

feature_size = states.shape[1]
self.model = K.models.Sequential([
K.layers.Dense(10, activation="relu", input_shape=(feature_size,)),
K.layers.Dense(10, activation="relu"),
K.layers.Dense(len(self.actions), activation="softmax")
])
self.set_updater(optimizer)
self.initialized = True
print("Done initialization. From now, begin training!")

def set_updater(self, optimizer):
actions = tf.placeholder(shape=(None), dtype="int32")
rewards = tf.placeholder(shape=(None), dtype="float32")
one_hot_actions = tf.one_hot(actions, len(self.actions), axis=1)
action_probs = self.model.output
selected_action_probs = tf.reduce_sum(one_hot_actions * action_probs,
axis=1)
clipped = tf.clip_by_value(selected_action_probs, 1e-10, 1.0)
loss = - tf.log(clipped) * rewards
loss = tf.reduce_mean(loss)

updates = optimizer.get_updates(loss=loss,
params=self.model.trainable_weights)
self._updater = K.backend.function(
inputs=[self.model.input,
actions, rewards],
outputs=[loss],
updates=updates)

def estimate(self, s):
normalized = self.scaler.transform(s)
action_probs = self.model.predict(normalized)[0]
return action_probs

def update(self, states, actions, rewards):
normalizeds = self.scaler.transform(states)
actions = np.array(actions)
rewards = np.array(rewards)
self._updater([normalizeds, actions, rewards])

class CartPoleObserver(Observer):

def transform(self, state):
return np.array(state).reshape((1, -1))

class PolicyGradientTrainer(Trainer):

def __init__(self, buffer_size=1024, batch_size=32,
gamma=0.9, report_interval=10, log_dir=""):
super().__init__(buffer_size, batch_size, gamma,
report_interval, log_dir)
self._reward_scaler = None
self.d_experiences = deque(maxlen=buffer_size)

def train(self, env, episode_count=220, epsilon=0.1, initial_count=-1,
render=False):
actions = list(range(env.action_space.n))
agent = PolicyGradientAgent(epsilon, actions)

self.train_loop(env, agent, episode_count, initial_count, render)
return agent

def episode_begin(self, episode, agent):
self.experiences = []

def step(self, episode, step_count, agent, experience):
if agent.initialized:
agent.update(*self.make_batch())

def make_batch(self):
batch = random.sample(self.d_experiences, self.batch_size)
states = np.vstack([e.s for e in batch])
actions = [e.a for e in batch]
rewards = [e.r for e in batch]
rewards = np.array(rewards).reshape((-1, 1))
rewards = self._reward_scaler.transform(rewards).flatten()
return states, actions, rewards

def begin_train(self, episode, agent):
optimizer = K.optimizers.Adam(clipnorm=1.0)
agent.initialize(self.d_experiences, optimizer)
self._reward_scaler = StandardScaler(with_mean=False)
rewards = np.array([[e.r] for e in self.d_experiences])
self._reward_scaler.fit(rewards)

def episode_end(self, episode, step_count, agent):
rewards = [e.r for e in self.experiences]
self.reward_log.append(sum(rewards))

discounteds = []
for t, r in enumerate(rewards):
d_r = [_r * (self.gamma ** i) for i, _r in
enumerate(rewards[t:])]
d_r = sum(d_r)
discounteds.append(d_r)

for i, e in enumerate(self.experiences):
s, a, r, n_s, d = e
d_r = discounteds[i]
d_e = Experience(s, a, d_r, n_s, d)
self.d_experiences.append(d_e)

if not self.training and len(self.d_experiences) == self.buffer_size:
self.begin_train(i, agent)
self.training = True

if self.is_event(episode, self.report_interval):
recent_rewards = self.reward_log[-self.report_interval:]
self.logger.describe("reward", recent_rewards, episode=episode)

def main(play):
env = CartPoleObserver(gym.make("CartPole-v0"))
trainer = PolicyGradientTrainer()
path = trainer.logger.path_of("policy_gradient_agent.h5")

if play:
agent = PolicyGradientAgent.load(env, path)
agent.play(env)
else:
trained = trainer.train(env, episode_count=250)
trainer.logger.plot("Rewards", trainer.reward_log,
trainer.report_interval)
trained.save(path)

if __name__ == "__main__":
parser = argparse.ArgumentParser(description="PG Agent")
parser.add_argument("--play", action="store_true", help="play with trained model")

args = parser.parse_args()
main(args.play)

結果は次の通りです。
結果(コンソール)

結果(グラフ)
前回実装した価値関数の場合より、報酬が増えるようになるには時間がかかるようです。

参考

Pythonで学ぶ強化学習 -入門から実践まで- サンプルコード

強化学習 x ニューラルネットワーク 6 (Catcher)

ニューラルネットワークでCatcherという環境を攻略してみます。
Catcherは、ボールキャッチを行うゲームです。

前回のCartPoleでは4つの入力パラメータを使って学習を行いましたが、今回は画面を入力パラメータとしています。

まずは親クラスとなるフレームワークを作成します。(前回のCartPoleと同じです。)

fn_framework.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
import os
import io
import re
from collections import namedtuple
from collections import deque
import numpy as np
import tensorflow as tf
from tensorflow.python import keras as K
from PIL import Image
import matplotlib.pyplot as plt

# s:状態
# a:行動
# r:報酬
# n_s:遷移先の状態
# d:エピソード終了フラグ
Experience = namedtuple("Experience",
["s", "a", "r", "n_s", "d"])

# ニューラルネットワークを使い状態から評価を行う。
class FNAgent():

def __init__(self, epsilon, actions):
self.epsilon = epsilon
self.actions = actions
self.model = None
self.estimate_probs = False
self.initialized = False

# 学習したエージェントを保存
def save(self, model_path):
self.model.save(model_path, overwrite=True, include_optimizer=False)

# 学習したエージェントを読み込み
@classmethod
def load(cls, env, model_path, epsilon=0.0001):
actions = list(range(env.action_space.n))
agent = cls(epsilon, actions)
agent.model = K.models.load_model(model_path)
agent.initialized = True
return agent

# 初期化
# experiences:エージェントの経験
def initialize(self, experiences):
raise Exception("You have to implements estimate method.")

# 関数による予測
def estimate(self, s):
raise Exception("You have to implements estimate method.")

# パラメータの更新
def update(self, experiences, gamma):
raise Exception("You have to implements update method.")

def policy(self, s):
if np.random.random() < self.epsilon or not self.initialized:
return np.random.randint(len(self.actions))
else:
estimates = self.estimate(s)
if self.estimate_probs:
action = np.random.choice(self.actions,
size=1, p=estimates)[0]
return action
else:
return np.argmax(estimates)

def play(self, env, episode_count=5, render=True):
for e in range(episode_count):
s = env.reset()
done = False
episode_reward = 0
while not done:
if render:
env.render()
a = self.policy(s)
n_state, reward, done, info = env.step(a)
episode_reward += reward
s = n_state
else:
print("Get reward {}.".format(episode_reward))

# エージェントの学習を行う
class Trainer():

def __init__(self, buffer_size=1024, batch_size=32,
gamma=0.9, report_interval=10, log_dir=""):
self.buffer_size = buffer_size
self.batch_size = batch_size
self.gamma = gamma
self.report_interval = report_interval
self.logger = Logger(log_dir, self.trainer_name)
# エージェントの行動履歴(古い行動からすてる)
self.experiences = deque(maxlen=buffer_size)
self.training = False
self.training_count = 0
self.reward_log = []

@property
def trainer_name(self):
class_name = self.__class__.__name__
snaked = re.sub("(.)([A-Z][a-z]+)", r"\1_\2", class_name)
snaked = re.sub("([a-z0-9])([A-Z])", r"\1_\2", snaked).lower()
snaked = snaked.replace("_trainer", "")
return snaked

def train_loop(self, env, agent, episode=200, initial_count=-1,
render=False, observe_interval=0):
self.experiences = deque(maxlen=self.buffer_size)
self.training = False
self.training_count = 0
self.reward_log = []
frames = []

for i in range(episode):
s = env.reset()
done = False
step_count = 0
self.episode_begin(i, agent)
while not done:
if render:
env.render()
if self.training and observe_interval > 0 and\
(self.training_count == 1 or
self.training_count % observe_interval == 0):
frames.append(s)

a = agent.policy(s)
n_state, reward, done, info = env.step(a)
e = Experience(s, a, reward, n_state, done)
self.experiences.append(e)
if not self.training and \
len(self.experiences) == self.buffer_size:
self.begin_train(i, agent)
self.training = True

self.step(i, step_count, agent, e)

s = n_state
step_count += 1
else:
self.episode_end(i, step_count, agent)

if not self.training and \
initial_count > 0 and i >= initial_count:
self.begin_train(i, agent)
self.training = True

if self.training:
if len(frames) > 0:
self.logger.write_image(self.training_count,
frames)
frames = []
self.training_count += 1

def episode_begin(self, episode, agent):
pass

def begin_train(self, episode, agent):
pass

def step(self, episode, step_count, agent, experience):
pass

def episode_end(self, episode, step_count, agent):
pass

def is_event(self, count, interval):
return True if count != 0 and count % interval == 0 else False

def get_recent(self, count):
recent = range(len(self.experiences) - count, len(self.experiences))
return [self.experiences[i] for i in recent]

# 環境から取得される「状態」の前処理を行う
class Observer():

def __init__(self, env):
self._env = env

@property
def action_space(self):
return self._env.action_space

@property
def observation_space(self):
return self._env.observation_space

def reset(self):
return self.transform(self._env.reset())

def render(self):
self._env.render()

def step(self, action):
n_state, reward, done, info = self._env.step(action)
return self.transform(n_state), reward, done, info

def transform(self, state):
raise Exception("You have to implements transform method.")

# 学習経過の記録を行う
class Logger():

def __init__(self, log_dir="", dir_name=""):
self.log_dir = log_dir
if not log_dir:
self.log_dir = os.path.join(os.path.dirname(__file__), "logs")
if not os.path.exists(self.log_dir):
os.mkdir(self.log_dir)

if dir_name:
self.log_dir = os.path.join(self.log_dir, dir_name)
if not os.path.exists(self.log_dir):
os.mkdir(self.log_dir)

self._callback = K.callbacks.TensorBoard(self.log_dir)

@property
def writer(self):
return self._callback.writer

def set_model(self, model):
self._callback.set_model(model)

def path_of(self, file_name):
return os.path.join(self.log_dir, file_name)

def describe(self, name, values, episode=-1, step=-1):
mean = np.round(np.mean(values), 3)
std = np.round(np.std(values), 3)
desc = "{} is {} (+/-{})".format(name, mean, std)
if episode > 0:
print("At episode {}, {}".format(episode, desc))
elif step > 0:
print("At step {}, {}".format(step, desc))

def plot(self, name, values, interval=10):
indices = list(range(0, len(values), interval))
means = []
stds = []
for i in indices:
_values = values[i:(i + interval)]
means.append(np.mean(_values))
stds.append(np.std(_values))
means = np.array(means)
stds = np.array(stds)
plt.figure()
plt.title("{} History".format(name))
plt.grid()
plt.fill_between(indices, means - stds, means + stds,
alpha=0.1, color="g")
plt.plot(indices, means, "o-", color="g",
label="{} per {} episode".format(name.lower(), interval))
plt.legend(loc="best")
plt.show()

def write(self, index, name, value):
summary = tf.Summary()
summary_value = summary.value.add()
summary_value.tag = name
summary_value.simple_value = value
self.writer.add_summary(summary, index)
self.writer.flush()

def write_image(self, index, frames):
# Deal with a 'frames' as a list of sequential gray scaled image.
last_frames = [f[:, :, -1] for f in frames]
if np.min(last_frames[-1]) < 0:
scale = 127 / np.abs(last_frames[-1]).max()
offset = 128
else:
scale = 255 / np.max(last_frames[-1])
offset = 0
channel = 1 # gray scale
tag = "frames_at_training_{}".format(index)
values = []

for f in last_frames:
height, width = f.shape
array = np.asarray(f * scale + offset, dtype=np.uint8)
image = Image.fromarray(array)
output = io.BytesIO()
image.save(output, format="PNG")
image_string = output.getvalue()
output.close()
image = tf.Summary.Image(
height=height, width=width, colorspace=channel,
encoded_image_string=image_string)
value = tf.Summary.Value(tag=tag, image=image)
values.append(value)

summary = tf.Summary(value=values)
self.writer.add_summary(summary, index)
self.writer.flush()

Catcherの画面を入力とし、各行動の価値を出力します。
このネットワークをCNNで構築し、学習していきます。

Catcherの行動は次の3つです。

  • 左に移動
  • 右に移動
  • 停止

ボールをキャッチできれば報酬1、キャッチできなければ報酬-1となります。

dqn_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
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
import random
import argparse
from collections import deque
import numpy as np
from tensorflow.python import keras as K
from PIL import Image
import gym
import gym_ple
from fn_framework import FNAgent, Trainer, Observer

class DeepQNetworkAgent(FNAgent):

def __init__(self, epsilon, actions):
super().__init__(epsilon, actions)
self._scaler = None
self._teacher_model = None

def initialize(self, experiences, optimizer):
feature_shape = experiences[0].s.shape
self.make_model(feature_shape)
self.model.compile(optimizer, loss="mse")
self.initialized = True
print("Done initialization. From now, begin training!")

def make_model(self, feature_shape):
normal = K.initializers.glorot_normal()
model = K.Sequential()
model.add(K.layers.Conv2D(
32, kernel_size=8, strides=4, padding="same",
input_shape=feature_shape, kernel_initializer=normal,
activation="relu"))
model.add(K.layers.Conv2D(
64, kernel_size=4, strides=2, padding="same",
kernel_initializer=normal,
activation="relu"))
model.add(K.layers.Conv2D(
64, kernel_size=3, strides=1, padding="same",
kernel_initializer=normal,
activation="relu"))
model.add(K.layers.Flatten())
model.add(K.layers.Dense(256, kernel_initializer=normal,
activation="relu"))
model.add(K.layers.Dense(len(self.actions),
kernel_initializer=normal))
self.model = model
self._teacher_model = K.models.clone_model(self.model)

def estimate(self, state):
return self.model.predict(np.array([state]))[0]

def update(self, experiences, gamma):
states = np.array([e.s for e in experiences])
n_states = np.array([e.n_s for e in experiences])

estimateds = self.model.predict(states)
future = self._teacher_model.predict(n_states)

for i, e in enumerate(experiences):
reward = e.r
if not e.d:
reward += gamma * np.max(future[i])
estimateds[i][e.a] = reward

loss = self.model.train_on_batch(states, estimateds)
return loss

def update_teacher(self):
self._teacher_model.set_weights(self.model.get_weights())

class DeepQNetworkAgentTest(DeepQNetworkAgent):

def __init__(self, epsilon, actions):
super().__init__(epsilon, actions)

def make_model(self, feature_shape):
normal = K.initializers.glorot_normal()
model = K.Sequential()
model.add(K.layers.Dense(64, input_shape=feature_shape,
kernel_initializer=normal, activation="relu"))
model.add(K.layers.Dense(len(self.actions), kernel_initializer=normal,
activation="relu"))
self.model = model
self._teacher_model = K.models.clone_model(self.model)

class CatcherObserver(Observer):

def __init__(self, env, width, height, frame_count):
super().__init__(env)
self.width = width
self.height = height
self.frame_count = frame_count
self._frames = deque(maxlen=frame_count)

def transform(self, state):
grayed = Image.fromarray(state).convert("L")
resized = grayed.resize((self.width, self.height))
resized = np.array(resized).astype("float")
normalized = resized / 255.0 # scale to 0~1
if len(self._frames) == 0:
for i in range(self.frame_count):
self._frames.append(normalized)
else:
self._frames.append(normalized)
feature = np.array(self._frames)
# Convert the feature shape (f, w, h) => (w, h, f).
feature = np.transpose(feature, (1, 2, 0))

return feature

class DeepQNetworkTrainer(Trainer):

def __init__(self, buffer_size=50000, batch_size=32,
gamma=0.99, initial_epsilon=0.5, final_epsilon=1e-3,
learning_rate=1e-3, teacher_update_freq=3, report_interval=10,
log_dir="", file_name=""):
super().__init__(buffer_size, batch_size, gamma,
report_interval, log_dir)
self.file_name = file_name if file_name else "dqn_agent.h5"
self.initial_epsilon = initial_epsilon
self.final_epsilon = final_epsilon
self.learning_rate = learning_rate
self.teacher_update_freq = teacher_update_freq
self.loss = 0
self.training_episode = 0
self._max_reward = -10

def train(self, env, episode_count=1200, initial_count=200,
test_mode=False, render=False, observe_interval=100):
actions = list(range(env.action_space.n))
if not test_mode:
agent = DeepQNetworkAgent(1.0, actions)
else:
agent = DeepQNetworkAgentTest(1.0, actions)
observe_interval = 0
self.training_episode = episode_count

self.train_loop(env, agent, episode_count, initial_count, render,
observe_interval)
return agent

def episode_begin(self, episode, agent):
self.loss = 0

def begin_train(self, episode, agent):
optimizer = K.optimizers.Adam(lr=self.learning_rate, clipvalue=1.0)
agent.initialize(self.experiences, optimizer)
self.logger.set_model(agent.model)
agent.epsilon = self.initial_epsilon
self.training_episode -= episode

def step(self, episode, step_count, agent, experience):
if self.training:
batch = random.sample(self.experiences, self.batch_size)
self.loss += agent.update(batch, self.gamma)

def episode_end(self, episode, step_count, agent):
reward = sum([e.r for e in self.get_recent(step_count)])
self.loss = self.loss / step_count
self.reward_log.append(reward)
if self.training:
self.logger.write(self.training_count, "loss", self.loss)
self.logger.write(self.training_count, "reward", reward)
self.logger.write(self.training_count, "epsilon", agent.epsilon)
if reward > self._max_reward:
agent.save(self.logger.path_of(self.file_name))
self._max_reward = reward
if self.is_event(self.training_count, self.teacher_update_freq):
agent.update_teacher()

diff = (self.initial_epsilon - self.final_epsilon)
decay = diff / self.training_episode
agent.epsilon = max(agent.epsilon - decay, self.final_epsilon)

if self.is_event(episode, self.report_interval):
recent_rewards = self.reward_log[-self.report_interval:]
self.logger.describe("reward", recent_rewards, episode=episode)

def main(play, is_test):
file_name = "dqn_agent.h5" if not is_test else "dqn_agent_test.h5"
trainer = DeepQNetworkTrainer(file_name=file_name)
path = trainer.logger.path_of(trainer.file_name)
print('path', path)
agent_class = DeepQNetworkAgent

if is_test:
print("Train on test mode")
obs = gym.make("CartPole-v0")
agent_class = DeepQNetworkAgentTest
else:
env = gym.make("Catcher-v0")
obs = CatcherObserver(env, 80, 80, 4)
trainer.learning_rate = 1e-4

if play:
agent = agent_class.load(obs, path)
agent.play(obs, render=True)
else:
trainer.train(obs, test_mode=is_test)

if __name__ == "__main__":
parser = argparse.ArgumentParser(description="DQN Agent")
parser.add_argument("--play", action="store_true", help="play with trained model")
parser.add_argument("--test", action="store_true", help="train by test mode")

args = parser.parse_args()
main(args.play, args.test)

実行するために下記のコマンドを実行し、必要な環境をインストールしておきます。

1
2
3
4
pip install gym_ple
git clone https://github.com/ntasfi/PyGame-Learning-Environment.git
cd PyGame-Learning-Environment/
pip install -e .

学習を場合は下記のコマンドを実行します。

1
python dqn_agent.py

学習データを使ってゲームを攻略するには下記コマンドを実行します。

1
python dqn_agent.py --play

上記コマンドでCatcherを実行している様子は下記の動画で確認できます。

なかなかのスピードでボールが落ちてきますが正確にキャッチできていることがわかります。
簡単なゲームながら画面から学習して、攻略している様子をみると応用範囲も期待もふくらみます。

参考

Pythonで学ぶ強化学習 -入門から実践まで- サンプルコード

強化学習 x ニューラルネットワーク 5 (CartPole)

ニューラルネットワークでCartPoleという環境を攻略してみます。
CartPoleは、棒が倒れないようにカートの位置を調整する環境です。

まずは親クラスとなるフレームワークを作成します。
フレームワークは下記の4種類のクラスで構成されています。

  • Agent:パラメータを持った関数(ニューラルネットワーク)で実装されたエージェント。
  • Trainer:エージェントの学習を行う。
  • Observer:環境から取得される「状態」の前処理を行う。
  • Logger:学習経過の記録を行う。
fn_framework.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
import os
import io
import re
from collections import namedtuple
from collections import deque
import numpy as np
import tensorflow as tf
from tensorflow.python import keras as K
from PIL import Image
import matplotlib.pyplot as plt

# s:状態
# a:行動
# r:報酬
# n_s:遷移先の状態
# d:エピソード終了フラグ
Experience = namedtuple("Experience",
["s", "a", "r", "n_s", "d"])

# ニューラルネットワークを使い状態から評価を行う。
class FNAgent():

def __init__(self, epsilon, actions):
self.epsilon = epsilon
self.actions = actions
self.model = None
self.estimate_probs = False
self.initialized = False

# 学習したエージェントを保存
def save(self, model_path):
self.model.save(model_path, overwrite=True, include_optimizer=False)

# 学習したエージェントを読み込み
@classmethod
def load(cls, env, model_path, epsilon=0.0001):
actions = list(range(env.action_space.n))
agent = cls(epsilon, actions)
agent.model = K.models.load_model(model_path)
agent.initialized = True
return agent

# 初期化
# experiences:エージェントの経験
def initialize(self, experiences):
raise Exception("You have to implements estimate method.")

# 関数による予測
def estimate(self, s):
raise Exception("You have to implements estimate method.")

# パラメータの更新
def update(self, experiences, gamma):
raise Exception("You have to implements update method.")

def policy(self, s):
if np.random.random() < self.epsilon or not self.initialized:
return np.random.randint(len(self.actions))
else:
estimates = self.estimate(s)
if self.estimate_probs:
action = np.random.choice(self.actions,
size=1, p=estimates)[0]
return action
else:
return np.argmax(estimates)

def play(self, env, episode_count=5, render=True):
for e in range(episode_count):
s = env.reset()
done = False
episode_reward = 0
while not done:
if render:
env.render()
a = self.policy(s)
n_state, reward, done, info = env.step(a)
episode_reward += reward
s = n_state
else:
print("Get reward {}.".format(episode_reward))

# エージェントの学習を行う
class Trainer():

def __init__(self, buffer_size=1024, batch_size=32,
gamma=0.9, report_interval=10, log_dir=""):
self.buffer_size = buffer_size
self.batch_size = batch_size
self.gamma = gamma
self.report_interval = report_interval
self.logger = Logger(log_dir, self.trainer_name)
# エージェントの行動履歴(古い行動からすてる)
self.experiences = deque(maxlen=buffer_size)
self.training = False
self.training_count = 0
self.reward_log = []

@property
def trainer_name(self):
class_name = self.__class__.__name__
snaked = re.sub("(.)([A-Z][a-z]+)", r"\1_\2", class_name)
snaked = re.sub("([a-z0-9])([A-Z])", r"\1_\2", snaked).lower()
snaked = snaked.replace("_trainer", "")
return snaked

def train_loop(self, env, agent, episode=200, initial_count=-1,
render=False, observe_interval=0):
self.experiences = deque(maxlen=self.buffer_size)
self.training = False
self.training_count = 0
self.reward_log = []
frames = []

for i in range(episode):
s = env.reset()
done = False
step_count = 0
self.episode_begin(i, agent)
while not done:
if render:
env.render()
if self.training and observe_interval > 0 and\
(self.training_count == 1 or
self.training_count % observe_interval == 0):
frames.append(s)

a = agent.policy(s)
n_state, reward, done, info = env.step(a)
e = Experience(s, a, reward, n_state, done)
self.experiences.append(e)
if not self.training and \
len(self.experiences) == self.buffer_size:
self.begin_train(i, agent)
self.training = True

self.step(i, step_count, agent, e)

s = n_state
step_count += 1
else:
self.episode_end(i, step_count, agent)

if not self.training and \
initial_count > 0 and i >= initial_count:
self.begin_train(i, agent)
self.training = True

if self.training:
if len(frames) > 0:
self.logger.write_image(self.training_count,
frames)
frames = []
self.training_count += 1

def episode_begin(self, episode, agent):
pass

def begin_train(self, episode, agent):
pass

def step(self, episode, step_count, agent, experience):
pass

def episode_end(self, episode, step_count, agent):
pass

def is_event(self, count, interval):
return True if count != 0 and count % interval == 0 else False

def get_recent(self, count):
recent = range(len(self.experiences) - count, len(self.experiences))
return [self.experiences[i] for i in recent]

# 環境から取得される「状態」の前処理を行う
class Observer():

def __init__(self, env):
self._env = env

@property
def action_space(self):
return self._env.action_space

@property
def observation_space(self):
return self._env.observation_space

def reset(self):
return self.transform(self._env.reset())

def render(self):
self._env.render()

def step(self, action):
n_state, reward, done, info = self._env.step(action)
return self.transform(n_state), reward, done, info

def transform(self, state):
raise Exception("You have to implements transform method.")

# 学習経過の記録を行う
class Logger():

def __init__(self, log_dir="", dir_name=""):
self.log_dir = log_dir
if not log_dir:
self.log_dir = os.path.join(os.path.dirname(__file__), "logs")
if not os.path.exists(self.log_dir):
os.mkdir(self.log_dir)

if dir_name:
self.log_dir = os.path.join(self.log_dir, dir_name)
if not os.path.exists(self.log_dir):
os.mkdir(self.log_dir)

self._callback = K.callbacks.TensorBoard(self.log_dir)

@property
def writer(self):
return self._callback.writer

def set_model(self, model):
self._callback.set_model(model)

def path_of(self, file_name):
return os.path.join(self.log_dir, file_name)

def describe(self, name, values, episode=-1, step=-1):
mean = np.round(np.mean(values), 3)
std = np.round(np.std(values), 3)
desc = "{} is {} (+/-{})".format(name, mean, std)
if episode > 0:
print("At episode {}, {}".format(episode, desc))
elif step > 0:
print("At step {}, {}".format(step, desc))

def plot(self, name, values, interval=10):
indices = list(range(0, len(values), interval))
means = []
stds = []
for i in indices:
_values = values[i:(i + interval)]
means.append(np.mean(_values))
stds.append(np.std(_values))
means = np.array(means)
stds = np.array(stds)
plt.figure()
plt.title("{} History".format(name))
plt.grid()
plt.fill_between(indices, means - stds, means + stds,
alpha=0.1, color="g")
plt.plot(indices, means, "o-", color="g",
label="{} per {} episode".format(name.lower(), interval))
plt.legend(loc="best")
plt.show()

def write(self, index, name, value):
summary = tf.Summary()
summary_value = summary.value.add()
summary_value.tag = name
summary_value.simple_value = value
self.writer.add_summary(summary, index)
self.writer.flush()

def write_image(self, index, frames):
# Deal with a 'frames' as a list of sequential gray scaled image.
last_frames = [f[:, :, -1] for f in frames]
if np.min(last_frames[-1]) < 0:
scale = 127 / np.abs(last_frames[-1]).max()
offset = 128
else:
scale = 255 / np.max(last_frames[-1])
offset = 0
channel = 1 # gray scale
tag = "frames_at_training_{}".format(index)
values = []

for f in last_frames:
height, width = f.shape
array = np.asarray(f * scale + offset, dtype=np.uint8)
image = Image.fromarray(array)
output = io.BytesIO()
image.save(output, format="PNG")
image_string = output.getvalue()
output.close()
image = tf.Summary.Image(
height=height, width=width, colorspace=channel,
encoded_image_string=image_string)
value = tf.Summary.Value(tag=tag, image=image)
values.append(value)

summary = tf.Summary(value=values)
self.writer.add_summary(summary, index)
self.writer.flush()

次に価値関数により行動を決定するエージェントを作成し、実際にCartPoleを攻略してみます。
CartPoleにおける「状態」は次の4つです。

  • 位置
  • 加速度
  • ポールの角度
  • ポールの倒れる速度(角速度)

「行動」はカートの左右への移動です。

「報酬」は常に1で、ポールが倒れたらエピソード終了になります。つまりポールが倒れない時間が長いほど報酬が手に入ります。

value_function_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
import random
import argparse
import numpy as np
from sklearn.neural_network import MLPRegressor
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
from sklearn.externals import joblib
import gym
from fn_framework import FNAgent, Trainer, Observer

class ValueFunctionAgent(FNAgent): # 親クラス(フレームワーク)を継承

def save(self, model_path):
joblib.dump(self.model, model_path)

@classmethod
def load(cls, env, model_path, epsilon=0.0001):
actions = list(range(env.action_space.n))
agent = cls(epsilon, actions)
agent.model = joblib.load(model_path)
agent.initialized = True
return agent

def initialize(self, experiences):
scaler = StandardScaler()
estimator = MLPRegressor(hidden_layer_sizes=(10, 10), max_iter=1)
self.model = Pipeline([("scaler", scaler), ("estimator", estimator)])

states = np.vstack([e.s for e in experiences])
self.model.named_steps["scaler"].fit(states)

# Avoid the predict before fit.
self.update([experiences[0]], gamma=0)
self.initialized = True
print("Done initialization. From now, begin training!")

def estimate(self, s):
estimated = self.model.predict(s)[0]
return estimated

def _predict(self, states):
if self.initialized:
predicteds = self.model.predict(states)
else:
size = len(self.actions) * len(states)
predicteds = np.random.uniform(size=size)
predicteds = predicteds.reshape((-1, len(self.actions)))
return predicteds

def update(self, experiences, gamma):
states = np.vstack([e.s for e in experiences])
n_states = np.vstack([e.n_s for e in experiences])

estimateds = self._predict(states)
future = self._predict(n_states)

for i, e in enumerate(experiences):
reward = e.r
if not e.d:
reward += gamma * np.max(future[i])
estimateds[i][e.a] = reward

estimateds = np.array(estimateds)
states = self.model.named_steps["scaler"].transform(states)
self.model.named_steps["estimator"].partial_fit(states, estimateds)

class CartPoleObserver(Observer):

def transform(self, state):
return np.array(state).reshape((1, -1))

class ValueFunctionTrainer(Trainer):

def train(self, env, episode_count=220, epsilon=0.1, initial_count=-1,
render=False):
actions = list(range(env.action_space.n))
agent = ValueFunctionAgent(epsilon, actions)
self.train_loop(env, agent, episode_count, initial_count, render)
return agent

def begin_train(self, episode, agent):
agent.initialize(self.experiences)

def step(self, episode, step_count, agent, experience):
if self.training:
batch = random.sample(self.experiences, self.batch_size)
# 学習を行う。
agent.update(batch, self.gamma)

def episode_end(self, episode, step_count, agent):
rewards = [e.r for e in self.get_recent(step_count)]
self.reward_log.append(sum(rewards))

if self.is_event(episode, self.report_interval):
recent_rewards = self.reward_log[-self.report_interval:]
self.logger.describe("reward", recent_rewards, episode=episode)

def main(play):
env = CartPoleObserver(gym.make("CartPole-v0"))
trainer = ValueFunctionTrainer()
path = trainer.logger.path_of("value_function_agent.pkl")

if play:
agent = ValueFunctionAgent.load(env, path)
agent.play(env)
else:
trained = trainer.train(env)
trainer.logger.plot("Rewards", trainer.reward_log, trainer.report_interval)
trained.save(path)

if __name__ == "__main__":
parser = argparse.ArgumentParser(description="VF Agent")
parser.add_argument("--play", action="store_true", help="play with trained model")

args = parser.parse_args()
main(args.play)

結果は下記のとおりです。
結果(コンソール)

結果(グラフ)

エピソードをこなすほど獲得報酬が増加していて、うまくカートを動かす方法を学習していることがわかります。

参考

Pythonで学ぶ強化学習 -入門から実践まで- サンプルコード

強化学習 x ニューラルネットワーク 4 (CNN)

CNN(畳み込みニューラルネットワーク)を実装してみます。おなじみの手書き数字の判定です。(MNIST)

CNNは強化学習にとって「画面入力による行動獲得」を可能にしたという点でとても重要な手法となります。

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
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.datasets import load_digits
from sklearn.metrics import classification_report
from sklearn.metrics import accuracy_score
from tensorflow.python import keras as K

dataset = load_digits()
image_shape = (8, 8, 1)
num_class = 10 # 各数字に対する確率

y = dataset.target # 画像に対する数字(0~9)
y = K.utils.to_categorical(y, num_class)
X = dataset.data
# 8 x 8 x 1のサイズに変更
X = np.array([data.reshape(image_shape) for data in X])

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33)

model = K.Sequential([
# CNNレイヤー
# フィルターの枚数 5
# フィルタサイズ kernel_size=3
# ストライド幅 strides=1
# フィルタタイズを補うようにパディング padding="same"
K.layers.Conv2D(
5, kernel_size=3, strides=1, padding="same", input_shape=image_shape, activation="relu"),
K.layers.Conv2D(
3, kernel_size=2, strides=1, padding="same", activation="relu"),
# 3次元の特徴マップを1次元のベクトルに変換
K.layers.Flatten(),
# 出力 10 units=num_class
K.layers.Dense(units=num_class, activation="softmax")
])
# 損失関数(出力された確率値と実際のクラスを比較) loss="categorical_crossentropy"
# 確率的勾配降下法 optimizer="sgd"
model.compile(loss="categorical_crossentropy", optimizer="sgd")
# 学習を行う
model.fit(X_train, y_train, epochs=10)

# 予測を行う
predicts = model.predict(X_test)
# argmax 配列で一番大きい要素のインデックスを返す => 予測された数字を意味する
predicts = np.argmax(predicts, axis=1)
# argmax 配列で一番大きい要素のインデックスを返す => 正解の数字を意味する
actual = np.argmax(y_test, axis=1)
print(classification_report(actual, predicts)) # 適合率・検出率・F値をまとめて表示
print('正解率 {:.2f}%'.format(accuracy_score(actual, predicts) * 100)) # 正解率

結果は次の通りです。
結果

正解率は93.43%とまずまずの結果となりました。

出力結果の見方は下記の用語集を参照してください。
機械学習に関する用語集 精度に関する用語

参考

Pythonで学ぶ強化学習 -入門から実践まで- サンプルコード

強化学習 x ニューラルネットワーク 3 (ボストン市の住宅価格予測)

ボストン市の住宅価格をニューラルネットワークで予想してみます。
住宅価格のデータセットは13の特徴量(入力)と住宅価格(出力)のセットとなっています。

ニューラルネットワークは13の変数 x から1つの値 y を出力することになります。
学習は予測した価格と実際の住宅価格の差異が小さくなるようにパラメータを調整することで行います。

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
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.datasets import load_boston
import pandas as pd
import matplotlib.pyplot as plt
from tensorflow.python import keras as K

# ボストン市の住宅価格
dataset = load_boston()

# 入力と出力に分ける
y = dataset.target
X = dataset.data

# 訓練データとテストデータに分ける
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33)

model = K.Sequential([
# データの正規化(入力は13の特徴量)
K.layers.BatchNormalization(input_shape=(13,)),
# 1層目のニューラルネットワーク
# 活性化関数はsoftplus
# kernel_regularizer正則化=>重みに制限をかける=>過学習防止
K.layers.Dense(units=13, activation="softplus", kernel_regularizer="l1"),
# 2層目のニューラルネットワーク
K.layers.Dense(units=1)
])
# loss=最小二乗法 optimizer=最適化に確率的勾配降下法
model.compile(loss="mean_squared_error", optimizer="sgd")

# 学習を行う(学習回数 epochs は8回)
model.fit(X_train, y_train, epochs=8)

# 予測を行う
predicts = model.predict(X_test)

result = pd.DataFrame({
"predict": np.reshape(predicts, (-1,)), # 2次元データを1次元データに変換
"actual": y_test
})
limit = np.max(y_test) # 最大値の取得

# 結果をグラフ表示する。
result.plot.scatter(x="actual", y="predict", xlim=(0, limit), ylim=(0, limit))
plt.show()

結果(コンソール)
8回の学習で誤差が 165.8451 から 18.0005 まで減っていることがわかります。

次に予測結果をグラフで確認します。
結果(グラフ)
横軸 x が実際の住宅価格で、縦軸 y が予測した住宅価格となります。
完全に一致していれば対角線上にプロットされることになります。今回の予測はだいたいあっているようです。

参考

Pythonで学ぶ強化学習 -入門から実践まで- サンプルコード

強化学習 x ニューラルネットワーク 2 (2層のニューラルネットワーク)

TensorFlowで2層のニューラルネットワークを実装してみます。

np.random.rand(3, 2)で3件の座標データbatchを作成しています。

また1層目から2層目にデータを送るときには、活性化関数(シグモイド)を適用しています。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
import numpy as np
from tensorflow.python import keras as K

# 2層のニューラルネットワーク
model = K.Sequential([
# 1層目 出力サイズ4、入力1行2列、活性化関数はシグモイド
K.layers.Dense(units=4, input_shape=((2, )), activation="sigmoid"),
# 2層目 出力サイズ4
K.layers.Dense(units=4),
])

print('-------------------------')
# 3件の座標をまとめたバッチ (2次元).
batch = np.random.rand(3, 2)
print('batchの形状', batch.shape)
print('batch', batch)
print('-------------------------')
y = model.predict(batch)
print('出力yの形状', y.shape)
print('出力y', y)
print('-------------------------')

結果

入力データbatchが3行2列(3件の座標データ)となり、出力データが3行4列(3件の4次元データ)となっていることがわかります。
このようにしてみると1行目に1件目の入力データと1件目の出力データが表示され、2行目に次の入力とその出力がされていて対応がわかりやすくなってます。

参考

Pythonで学ぶ強化学習 -入門から実践まで- サンプルコード

強化学習 x ニューラルネットワーク 1 (1層のニューラルネットワーク)

これまではQ[s][a]というテーブル内の値を更新することで学習していましたが、今回からは関数のパラメータを調整することで学習していきます。

まずはTensorFlowで1層のニューラルネットワークを実装してみます。

入力(x)は2行1列の座標、出力(y)は行動価値4行1列を想定しています。
対応する重み(weight)は4行2列でバイアス(bias)は4行1列となります。

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
import numpy as np
from tensorflow.python import keras as K

model = K.Sequential([ # 複数の層をまとめるためのモジュール
# K.layers.Dense => ニューラルネットワークを表す(重みとバイアスを持つ層)
# units=4 => 出力サイズ
# input_shape => 入力サイズ
K.layers.Dense(units=4, input_shape=((2, ))),
])

weight, bias = model.layers[0].get_weights() # 第1層の重みとバイアスを取得
print('---------------------')
print('重みの形状 {}.'.format(weight.shape))
print('重み', weight)
print('---------------------')
print('バイアスの形状 {}.'.format(bias.shape))
print('バイアス', bias)
x = np.random.rand(1, 2)
y = model.predict(x)
print('---------------------')
print('x(入力)の形状', x.shape)
print('x(入力)', x)
print('---------------------')
print('y(出力・結果)の形状', y.shape)
print('y(出力)', y)
print('---------------------')

結果

想定した行列とは全て逆の結果となりました。

これは座標を1行2列で入力したためです。=> np.random.rand(1, 2)
このため重み、バイアス、出力の全てが行列が反対になってしまっていますが、本質的な結果は変わりません。

多くの深層学習フレームワークでは行をデータ数(バッチサイズ)を表すのに使うため、このような仕様となっていますので慣れてしまいましょう。

参考

Pythonで学ぶ強化学習 -入門から実践まで- サンプルコード


Your browser is out-of-date!

Update your browser to view this website correctly. Update my browser now

×