-
Notifications
You must be signed in to change notification settings - Fork 0
/
run_game_v2_battle_dqn.py
205 lines (166 loc) · 7.02 KB
/
run_game_v2_battle_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
import random
import logging
import argparse
import numpy as np
import sys
from collections import deque, OrderedDict
from keras.layers import Dense
from keras.optimizers import Adam
from keras.models import Sequential
from game_v2 import *
class DQNAgent:
def __init__(self, state_size, action_size, hidden_sizes,
discount_factor=0.99, learning_rate=0.001, # 0.99, 0.001 originally
batch_size=64, train_start=100,
epsilon=1.0, epsilon_min=0.005, epsilon_steps=1000, # 1.0, 0.005, 50000
memory_size=1000):
self.state_size = state_size
self.action_size = action_size
self.hidden_sizes = hidden_sizes
self.discount_factor = discount_factor
self.learning_rate = learning_rate
self.epsilon = epsilon
self.epsilon_min = epsilon_min
self.epsilon_decay = (self.epsilon - self.epsilon_min) / epsilon_steps
self.batch_size = batch_size
self.train_start = train_start
self.memory_size = memory_size
self.memory = deque(maxlen=memory_size)
self.model = self.build_model()
self.target_model = self.build_model()
self.update_target_model()
def build_model(self):
model = Sequential()
input_dim = self.state_size
for dim in self.hidden_sizes:
model.add(Dense(dim, input_dim=input_dim, activation='relu', kernel_initializer='he_uniform'))
input_dim = dim
model.add(Dense(self.action_size, input_dim=input_dim, activation='linear', kernel_initializer='he_uniform'))
model.compile(loss='mse', optimizer=Adam(lr=self.learning_rate))
return model
def update_target_model(self):
self.target_model.set_weights(self.model.get_weights())
def get_action(self, state):
if np.random.rand() <= self.epsilon:
return random.randrange(self.action_size)
else:
q_value = self.model.predict(state)
return np.argmax(q_value[0])
def replay_memory(self, state, action, reward, next_state, done):
self.memory.append((state, action, reward, next_state, done))
if self.epsilon > self.epsilon_min:
self.epsilon -= self.epsilon_decay
def train_replay(self):
memory_size = len(self.memory)
if memory_size < self.train_start:
return
batch_size = min(self.batch_size, memory_size)
mini_batch = random.sample(self.memory, batch_size)
update_input = np.zeros((batch_size, self.state_size))
update_target = np.zeros((batch_size, self.action_size))
for i in range(batch_size):
state, action, reward, next_state, done = mini_batch[i]
if not done:
target = reward + self.discount_factor * np.amax(self.target_model.predict(next_state)[0])
else:
target = reward
update_input[i] = state
update_target[i] = target
self.model.fit(update_input, update_target, batch_size=batch_size, epochs=1, verbose=0)
def load_weights(self, name):
self.model.load_weights(name)
def save_weights(self, name):
self.model.save_weights(name)
def save_model(self, name):
self.model.save(name)
if __name__ == "__main__":
# parse arguments
# example: `python3 run_game_v2_battle_dqn_local.py -n 64 64 -i 5`
parser = argparse.ArgumentParser()
parser.add_argument('--hidden_sizes', '-n', nargs='*', type=int, default=[])
parser.add_argument('--discount_factor', type=float, default=0.99)
parser.add_argument('--learning_rate', type=float, default=0.001)
parser.add_argument('--batch_size', type=int, default=256)
parser.add_argument('--train_start', type=int, default=100)
parser.add_argument('--epsilon', type=float, default=1.0)
parser.add_argument('--epsilon_min', type=float, default=0.005)
parser.add_argument('--epsilon_steps', type=int, default=5000)
parser.add_argument('--log_id', '-i', type=int, required=True)
parser.add_argument('--episodes', type=int, default=1000)
parser.add_argument('--state', type=str, default="1d_1")
parser.add_argument('--reward', type=str, default="score_1")
args = parser.parse_args()
kwargs = OrderedDict(sorted(args._get_kwargs(), key=lambda x: x[0]))
episodes = kwargs.pop('episodes')
# initialize logger
log_id = kwargs.pop('log_id')
assert log_id < 1000
filename = 'battle-dqn-{:0>3}-local.log'.format(log_id)
logger = logging.Logger("DQNLogger")
sh = logging.StreamHandler()
fh = logging.FileHandler(filename)
logger.addHandler(sh)
logger.addHandler(fh)
model_path_wr = 'battle-dqn-{:0>3}-best-wr-local.h5'.format(log_id)
model_path_ar = 'battle-dqn-{:0>3}-best-ar-local.h5'.format(log_id)
# log hyper-parameters
for k, v, in kwargs.items():
logger.info('{}={}'.format(k, v))
# initialize agent
state_version = kwargs.pop("state")
reward_version = kwargs.pop("reward")
env = BattleEnv(state_version=state_version, reward_version=reward_version)
state_size = env.state_space_dim
action_size = env.action_space_dim
agent = DQNAgent(state_size, action_size, **kwargs)
wins = []
rewards = []
max_window_wr = -sys.maxsize - 1
max_window_ar = -sys.maxsize - 1
state, done = env.start_round()
for i in range(episodes):
reward = 0
while not done:
action = agent.get_action(state)
next_state, reward, done = env.step(action)
agent.replay_memory(state, action, reward, next_state, done)
agent.train_replay()
state = next_state
agent.update_target_model()
win = env.ext_player.num_cards == 0
wins.append(int(win))
rewards.append(env.opp_player.loss if win else -env.ext_player.loss)
if i >= 100:
window_wr = sum(wins[-100:])
window_ar = sum(rewards[-100:]) / 100
else:
window_wr = sum(wins) / (i + 1) * 100
window_ar = sum(rewards) / (i + 1)
if agent.epsilon > agent.epsilon_min:
msg = " - epsilon={:.4f}".format(agent.epsilon)
else:
msg = ""
msg += " - win rate: {}%, avg reward: {}".format(
round(window_wr, 2),
round(window_ar, 2)
)
if win:
msg = "Round {}: win{}".format(i, msg)
else:
msg = "Round {}: lose - {} cards left{}".format(i, env.ext_player.num_cards, msg)
if window_wr > max_window_wr:
max_window_wr = window_wr
agent.save_model(model_path_wr)
msg += " (best wwr)"
if window_ar > max_window_ar:
max_window_ar = window_ar
agent.save_model(model_path_ar)
msg += " (best war)"
logger.info(msg)
if i < episodes - 1:
state, done = env.reset()
env.end_round()
logger.info("best window win rate: {}\nbest window avg reward: {}".format(
round(max_window_wr, 2),
round(max_window_ar, 2)
))