-
Notifications
You must be signed in to change notification settings - Fork 1
/
DQN_ER_solution.py
177 lines (145 loc) · 6.32 KB
/
DQN_ER_solution.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
#Guy Zaks and Maor Reuven
import random
import gym
import math
import numpy as np
from collections import deque
from keras.models import Sequential
from keras.layers import Dense
from keras.optimizers import Adam
import matplotlib.pyplot as plt
score_log = []
loss_log = []
class FitForwardBuilder:
@staticmethod
def build_3_layers(alpha, alpha_decay):
model = Sequential()
model.add(Dense(24, input_dim=4, activation='tanh'))
model.add(Dense(48, activation='tanh'))
model.add(Dense(48, activation='tanh'))
model.add(Dense(2, activation='linear'))
model.compile(loss='mse', optimizer=Adam(lr=alpha, decay=alpha_decay))
return model
@staticmethod
def build_5_layers(alpha, alpha_decay):
model = Sequential()
model.add(Dense(100, input_dim=4, activation='relu'))
model.add(Dense(100, activation='relu'))
model.add(Dense(100, activation='relu'))
model.add(Dense(100, activation='relu'))
model.add(Dense(100, activation='relu'))
model.add(Dense(2, activation='linear'))
model.compile(loss='mse', optimizer=Adam(lr=alpha, decay=alpha_decay))
return model
class Memory:
def __init__(self, max_size=100000):
self._memory = deque(maxlen=max_size)
def store_transition(self, transition):
self._memory.append(transition)
def get_sample(self, sample_size):
return random.sample(self._memory, min(len(self._memory), sample_size))
class DQNSolverUsingER():
def __init__(self, max_env_steps=None, gamma=0.95, batch_size=64, update_steps=65, layer_type=3):
#default hyperparameters
self.memory = Memory()
self.cart_pole = gym.make('CartPole-v1')
self.epsilon = 1.0
self.epsilon_min = 0.01
self.epsilon_decay = 0.995
self.alpha = 0.01
self.alpha_decay = 0.01
self.n_episodes = 100000000
self.steps_target = 475
#passed hyperparameters
self.cart_pole._max_episode_steps = max_env_steps
self.gamma = gamma
self.batch_size = batch_size
self.update_steps = update_steps
#network structure type
if (layer_type == 3):
self.current_model = FitForwardBuilder.build_3_layers(self.alpha, self.alpha_decay)
self.target_model = FitForwardBuilder.build_3_layers(self.alpha, self.alpha_decay)
else:
self.current_model = FitForwardBuilder.build_5_layers(self.alpha, self.alpha_decay)
self.target_model = FitForwardBuilder.build_5_layers(self.alpha, self.alpha_decay)
# graphic properties
plt.rcParams['image.cmap'] = 'RdYlGn'
plt.rcParams['figure.figsize'] = [15.0, 6.0]
plt.rcParams['figure.dpi'] = 80
plt.rcParams['savefig.dpi'] = 30
def choose_action(self, state, epsilon):
return self.cart_pole.action_space.sample() if (np.random.random() <= epsilon) else np.argmax(
self.current_model.predict(state))
def get_epsilon(self, t):
return max(self.epsilon_min, min(self.epsilon, 1.0 - math.log10((t + 1) * self.epsilon_decay)))
def preprocess_state(self, state):
return np.reshape(state, [1, 4])
def replay(self, batch_size):
states, target_states = [], []
minibatch = self.memory.get_sample(batch_size)
for state, action, reward, next_state, done in minibatch:
target_state = self.current_model.predict(state)
target_state[0][action] = reward
if not done:
target_state[0][action] += self.gamma * np.max(self.target_model.predict(next_state)[0])
states.append(state[0])
target_states.append(target_state[0])
history = self.current_model.fit(np.array(states), np.array(target_states), batch_size=len(states), verbose=0)
loss = history.history['loss'][0]
if self.epsilon > self.epsilon_min:
self.epsilon *= self.epsilon_decay
return loss
def update_target_model(self):
self.target_model.set_weights(self.current_model.get_weights())
def print_graphs(self):
plt.plot(range(len(score_log)), score_log[0:(len(score_log))], 'o')
plt.title("Total reward per episode")
plt.show()
plt.plot(range(len(loss_log)), loss_log, 'o')
plt.title("Total loss per episode")
plt.show()
avg_list = []
for r in range(len(score_log) + 1):
if r > 99:
avg_list.append(np.mean(score_log[(r - 100): r]))
else:
avg_list.append(np.mean(score_log[(0): r]))
plt.plot(range(len(avg_list)), avg_list, '-')
plt.title("AVG reward per 100 episode")
plt.show()
def run(self):
scores = deque(maxlen=100)
steps = 0
for episode in range(self.n_episodes):
state = self.preprocess_state(self.cart_pole.reset())
done = False
total_reward = 0
while not done:
steps += 1
if steps % self.update_steps == 0:
self.update_target_model()
action = self.choose_action(state, self.get_epsilon(episode))
next_state, reward, done, _ = self.cart_pole.step(action)
next_state = self.preprocess_state(next_state)
self.memory.store_transition((state, action, reward, next_state, done))
loss = self.replay(self.batch_size)
loss_log.append(loss)
state = next_state
total_reward += reward
scores.append(total_reward)
score_log.append(total_reward)
mean_score = np.mean(scores)
if mean_score >= self.steps_target and episode >= 100:
print('Done after {} episodes'.format(episode))
return episode - 100
if episode % 10 == 0:
print('[Episode {}] - Average reward over last 100 episodes was {}.'.format(episode, mean_score))
if total_reward > 500:
print('Episode: {} got score of {} to the current avg of {}'.format(episode, total_reward, mean_score))
if episode % 1000 == 0 and episode > 0:
self.print_graphs()
return episode
if __name__ == '__main__':
solver = DQNSolverUsingER(max_env_steps=10000, gamma=0.95, batch_size=64, update_steps=64, layer_type=3)
solver.run()
solver.print_graphs()