-
Notifications
You must be signed in to change notification settings - Fork 0
/
dqn.py
224 lines (192 loc) · 8.6 KB
/
dqn.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
import numpy as np
import tensorflow as tf
from statistics import mean
from collections import deque
class Network(tf.keras.Model):
'''
Base network
'''
def __init__(self, num_states, hidden_units, num_actions):
super(Network, self).__init__()
self.input_layer = tf.keras.layers.InputLayer(input_shape=(num_states,))
self.hidden_layers = []
for i in hidden_units:
self.hidden_layers.append(tf.keras.layers.Dense(
i, activation='relu', kernel_initializer='RandomNormal'))
self.output_layer = tf.keras.layers.Dense(
num_actions, activation='linear', kernel_initializer='RandomNormal'
)
@tf.function
def call(self, inputs):
z = self.input_layer(inputs)
for layer in self.hidden_layers:
z = layer(z)
output = self.output_layer(z)
return output
class Model:
'''
Model for the target network and local network
'''
def __init__(self, env, multistep, n_step):
# Environment info
self.env = env
self.state_size = self.env.observation_space.shape[0]
self.action_size = self.env.action_space.n
# Parameter for multi step
self.multistep = multistep
self.n_step = n_step
# Hyperparameter
self.batch_size = 32
self.learning_rate = 1e-4
self.gamma = 0.99
# Network
self.optimizer = tf.optimizers.Adam(self.learning_rate)
self.hidden_units = [200, 200]
self.network = Network(self.state_size, self.hidden_units, self.action_size)
# Experience replay buffer
self.experience = {'s': [], 'a': [], 'r': [], 's2': [], 'done': []}
self.max_experiences = 10000
self.min_experiences = 100
# Predict the value with state using network
def predict(self, state):
return self.network(np.atleast_2d(state.astype('float32')))
# Get action with state by epsilon-greedy algorithm
def get_action(self, states, epsilon):
if np.random.random() < epsilon:
return np.random.choice(self.action_size)
else:
return np.argmax(self.predict(np.atleast_2d(states))[0])
# Take mini batch and update the policy
def train_minibatch(self, TargetNet):
# If there is not enough experiences in the experience replay buffer, just return 0
if len(self.experience['s']) < self.min_experiences:
return 0
# Multistep DQN
if self.multistep:
# Sampling the mini batch index (maximum ids is smaller than full length, for preventing out of index error)
multi_ids = np.random.randint(low=0, high=max(len(self.experience['s'])-self.n_step+1, 0),\
size=self.batch_size)
# Lists of each sample's state, action, and reward
states = np.asarray([self.experience['s'][i] for i in multi_ids])
actions = np.asarray([self.experience['a'][i] for i in multi_ids])
rewards = np.asarray([self.experience['r'][i] for i in multi_ids])
# Information for n-th forward state from above state
end_step_states = np.asarray([self.experience['s2'][i+self.n_step-1] for i in multi_ids])
end_step_value = np.max(TargetNet.predict(end_step_states), axis=1)
dones = np.asarray([self.experience['done'][i+self.n_step-1] for i in multi_ids])
# Temp gamma for iteration
gamma = self.gamma
# Calculate multi step rewards
for i in range(1, self.n_step):
step_reward = np.asarray([self.experience['r'][j+i] for j in multi_ids]) * gamma
rewards += step_reward
gamma *= self.gamma
# Calculate q-target
q_target = np.where(dones, rewards, rewards + gamma * end_step_value)
# Original DQN (equivalent to n=1 step DQN)
else:
# Sampling the mini batch index
ids = np.random.randint(low=0, high=len(self.experience['s']), size=self.batch_size)
# Lists of each sample's state, action, reward, and done
states = np.asarray([self.experience['s'][i] for i in ids])
actions = np.asarray([self.experience['a'][i] for i in ids])
rewards = np.asarray([self.experience['r'][i] for i in ids])
dones = np.asarray([self.experience['done'][i] for i in ids])
# Information for next states
states_next = np.asarray([self.experience['s2'][i] for i in ids])
value_next = np.max(TargetNet.predict(states_next), axis=1)
# Calculate q-target
q_target = np.where(dones, rewards, rewards + self.gamma * value_next)
# Q-prediction get from wS_t
with tf.GradientTape() as tape:
selected_action_values = tf.math.reduce_sum(
self.predict(states) * tf.one_hot(actions, self.action_size), axis=1)
# Calculate loss
loss = tf.math.reduce_mean(tf.square(q_target - selected_action_values))
# Updating gradients
variables = self.network.trainable_variables
gradients = tape.gradient(loss, variables)
self.optimizer.apply_gradients((zip(gradients, variables)))
return loss
# Add new experience to the experience replay buffer
def add_experience(self, exp):
if len(self.experience['s']) >= self.max_experiences:
for key in self.experience.keys():
self.experience[key].pop(0)
for key, value in exp.items():
self.experience[key].append(value)
# Copy the network's weight to another network
def copy_weights(self, LocalNet):
variables1 = self.network.trainable_variables
variables2 = LocalNet.network.trainable_variables
for v1, v2 in zip(variables1, variables2):
v1.assign(v2.numpy())
class DQN:
'''
DQN consists of target network and local network
'''
def __init__(self, env, multistep=False):
# Environment info
self.env = env
self.state_size = self.env.observation_space.shape[0]
self.action_size = self.env.action_space.n
# Hyperparameter
self.epsilon = 0.1
self.min_epsilon = 0
self.epsilon_decay = 0.9999
self.copy_step = 3
# Parameter for multi step
self.multistep = multistep
self.n_steps = 3
# Build target network and local network
def _build_network(self):
self.TargetNet = Model(self.env, self.multistep, self.n_steps)
self.LocalNet = Model(self.env, self.multistep, self.n_steps)
# Play the one episode
def play_game(self):
rewards = 0
step_count = 0
done = False
observations = self.env.reset()
losses = list()
while not done:
action = self.LocalNet.get_action(observations, self.epsilon)
prev_observations = observations
observations, reward, done, _ = self.env.step(action)
rewards += reward
if done:
reward = -250
self.env.reset()
exp = {'s': prev_observations, 'a': action, 'r':reward, 's2': observations, 'done': done}
self.LocalNet.add_experience(exp)
loss = self.LocalNet.train_minibatch(self.TargetNet)
if isinstance(loss, int):
losses.append(loss)
else:
losses.append(loss.numpy())
step_count += 1
if step_count % self.copy_step == 0:
self.TargetNet.copy_weights(self.LocalNet)
return rewards, mean(losses), step_count
# Updating epsilon
def update_epsilon(self):
self.epsilon = max(self.min_epsilon, self.epsilon * self.epsilon_decay)
# Training the DQN
def learn(self, max_episode=1500):
self._build_network()
avg_step_count_list = []
last_100_episode_step_count = deque(maxlen=100)
total_rewards = np.empty(max_episode)
for episode in range(max_episode):
self.update_epsilon()
total_reward, losses, step_count = self.play_game()
total_rewards[episode] = total_reward
last_100_episode_step_count.append(step_count)
avg_step_count = np.mean(last_100_episode_step_count)
# Print progress
print("[Episode {:>5}] episode step_count: {:>5} avg step_count: {}".format(episode, step_count, avg_step_count))
avg_step_count_list.append(avg_step_count)
# If average step count exceed 475, stop early
if avg_step_count >= 475:
break
return avg_step_count_list