forked from charleschen003/doudizhu-rl
/
game.py
290 lines (269 loc) · 12.4 KB
/
game.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
import os
import time
import json
import config as conf
import torch
BEGIN, logger, LOG_PATH = conf.get_logger()
class Game:
def __init__(self, env_cls, nets_dict, dqns_dict, reward_dict=None,
train_dict=None, preload=None, seed=None, debug=False):
if reward_dict is None:
reward_dict = {'lord': 100, 'down': 50, 'up': 50}
if train_dict is None:
train_dict = {'lord': True, 'down': True, 'up': True}
if preload is None:
preload = {}
assert not (nets_dict.keys() ^ dqns_dict.keys()), 'Net and DQN must match'
self.lord_wins, self.down_wins, self.up_wins = [], [], []
self.lord_total_loss = self.down_total_loss = self.up_total_loss = 0
self.lord_loss_count = self.down_loss_count = self.up_loss_count = 0
self.up_total_wins = self.lord_total_wins = self.down_total_wins = 0
self.up_recent_wins = self.lord_recent_wins = self.down_recent_wins = 0
self.lord_max_wins = self.farmer_max_wins = 0
self.env = env_cls(debug=debug, seed=seed)
self.lord = self.down = self.up = None
self.lord_train = self.down_train = self.up_train = False
for role in ['lord', 'down', 'up']:
if nets_dict.get(role):
setattr(self, role, dqns_dict[role](nets_dict[role]))
setattr(self, '{}_train'.format(role), train_dict[role])
if preload.get(role):
getattr(self, role).target_net.load(preload.get(role))
getattr(self, role).policy_net.load(preload.get(role))
self.lord_s0 = self.down_s0 = self.up_s0 = None
self.lord_a0 = self.down_a0 = self.up_a0 = None
self.reward_dict = reward_dict
self.preload = preload
self.train_dict = train_dict
def accumulate_loss(self, name, loss):
assert name in {'up', 'down', 'lord'}
if loss:
if name == 'lord':
self.lord_loss_count += 1
self.lord_total_loss += loss
elif name == 'down':
self.down_loss_count += 1
self.down_total_loss += loss
else:
self.up_loss_count += 1
self.up_total_loss += loss
def save_win_rates(self, episode):
self.lord_wins.append(self.lord_recent_wins)
self.up_wins.append(self.up_recent_wins)
self.down_wins.append(self.down_recent_wins)
# 是否高于最高胜率
if self.lord and self.up is None and self.down is None:
if self.lord_recent_wins > self.lord_max_wins:
self.lord_max_wins = self.lord_recent_wins
self.lord.policy_net.save(
'{}_lord_{}_{}'.format(BEGIN, episode, self.lord_max_wins))
if self.lord and not self.lord_train:
if self.up_recent_wins + self.down_recent_wins > self.farmer_max_wins:
self.farmer_max_wins = self.up_recent_wins + self.down_recent_wins
self.up.policy_net.save(
'{}_up_{}_{}'.format(BEGIN, episode, self.farmer_max_wins))
self.down.policy_net.save(
'{}_down_{}_{}'.format(BEGIN, episode, self.farmer_max_wins))
# 存一次胜率目录
data = {'lord': self.lord_wins, 'down': self.down_wins, 'up': self.up_wins}
path = os.path.join(conf.WIN_DIR, conf.name_dir(BEGIN))
dirname = os.path.dirname(path)
if not os.path.exists(dirname):
os.makedirs(dirname)
path = '{}.json'.format(path)
with open(path, 'w') as f:
json.dump(data, f)
def reset_recent(self):
self.lord_recent_wins = self.up_recent_wins = self.down_recent_wins = 0
self.lord_total_loss = self.down_total_loss = self.up_total_loss = 0
self.lord_loss_count = self.down_loss_count = self.up_loss_count = 0
def step(self, ai):
assert ai in {'lord', 'down', 'up'}
agent = getattr(self, ai)
continue_train = getattr(self, '{}_train'.format(ai))
if agent: # 不是使用规则
s0 = self.env.face
if continue_train: # 需要继续训练
setattr(self, '{}_s0'.format(ai), s0) # 更新状态s0
action_f = agent.e_greedy_action
else:
action_f = agent.greedy_action
a0 = action_f(s0, self.env.valid_actions())
if continue_train:
setattr(self, '{}_a0'.format(ai), a0) # 更新动作a0
_, done, _ = self.env.step_manual(a0)
else:
_, done, _ = self.env.step_auto()
return done
def feedback(self, ai, done, punish=False):
assert ai in {'lord', 'up', 'down'}
agent = getattr(self, ai)
if agent and getattr(self, '{}_train'.format(ai)): # 是需要继续训练的模型
if done:
reward = self.reward_dict[ai]
if punish:
reward = -reward
else:
reward = 0
s0 = getattr(self, '{}_s0'.format(ai))
a0 = getattr(self, '{}_a0'.format(ai))
s1 = self.env.face
if done:
a1 = torch.zeros((15, 4), dtype=torch.float).to(conf.DEVICE)
else:
a1 = agent.greedy_action(s1, self.env.valid_actions())
loss = agent.perceive(s0, a0, reward, s1, a1, done)
self.accumulate_loss(ai, loss)
def lord_turn(self):
done = self.step('lord')
if not done: # 本局未结束
if self.down_a0 is not None: # 如果下家曾经出过牌
self.feedback('down', done)
else: # 本局结束，地主胜利
if self.down_a0 is not None: # 如果下家曾经出过牌（不是一次性走完）
self.feedback('down', done, punish=True) # 下家负反馈
self.feedback('up', done, punish=True) # 上家负反馈
# 自己得到正反馈
self.feedback('lord', done)
self.lord_total_wins += 1
self.lord_recent_wins += 1
return done
def down_turn(self):
done = self.step('down')
if not done: # 本局未结束
if self.up_a0 is not None:
self.feedback('up', done)
else: # 本局结束，农民胜利
self.feedback('up', done)
self.feedback('lord', done, punish=True)
self.feedback('down', done)
self.down_recent_wins += 1
self.down_total_wins += 1
return done
def up_turn(self):
done = self.step('up')
if not done: # 本局未结束，地主得到0反馈
self.feedback('lord', done)
else: # 本局结束，农民胜利
self.feedback('lord', done, punish=True) # 地主得到负反馈
self.feedback('down', done) # 下家得到正反馈
self.feedback('up', done) # 自己得到正反馈
self.up_total_wins += 1
self.up_recent_wins += 1
return done
def play(self):
self.env.reset()
self.env.prepare()
while True: #
done = self.lord_turn()
if done:
break
done = self.down_turn()
if done:
break
done = self.up_turn()
if done:
break
def train(self, episodes, log_every=100, model_every=1000):
if not ((self.lord and self.lord_train)
or (self.up and self.up_train)
or (self.down and self.down_train)):
print('No agent need train.')
return
print('Logged at {}'.format(LOG_PATH))
messages = ''
for role in ['up', 'lord', 'down']:
m = '{}: {} based model.'.format(
role, 'AI' if getattr(self, role) else 'Rule')
if getattr(self, role):
preload = self.preload.get(role)
if preload:
m += ' With pretrained model {}.'.format(preload)
else:
m += ' Without pretrained model.'
if self.train_dict.get(role):
m += ' Continue training.'
messages += '\n{}'.format(m)
logger.info(messages + '\n------------------------------------')
print(messages)
start_time = time.time()
for episode in range(1, episodes + 1):
self.play()
if episode % log_every == 0:
end_time = time.time()
message = (
'Reach at round {}, recent {} rounds takes {:.2f}seconds\n'
'\tUp recent/total win: {:.2%}/{:.2%} [Mean loss: {:.2f}]\n'
'\tLord recent/total win: {:.2%}/{:.2%} [Mean loss: {:.2f}]\n'
'\tDown recent/total win: {:.2%}/{:.2%} [Mean loss: {:.2f}]\n'
).format(episode, log_every, end_time - start_time,
self.up_recent_wins / log_every, self.up_total_wins / episode,
self.up_total_loss / (self.up_loss_count + 1e-3),
self.lord_recent_wins / log_every, self.lord_total_wins / episode,
self.lord_total_loss / (self.lord_loss_count + 1e-3),
self.down_recent_wins / log_every, self.down_total_wins / episode,
self.down_total_loss / (self.down_loss_count + 1e-3))
logger.info(message)
self.save_win_rates(episode)
self.reset_recent()
start_time = time.time()
if episode % model_every == 0:
for role in ['lord', 'down', 'up']:
ai = getattr(self, role)
if ai:
ai.policy_net.save(
'{}_{}_{}'.format(BEGIN, role, episode))
for role in ['lord', 'down', 'up']:
ai = getattr(self, role)
if ai:
ai.update_epsilon(episode)
ai.update_target(episode)
@staticmethod
def compete(env_cls, nets_dict, dqns_dict, model_dict, total=1000,
print_every=100, debug=True):
import collections
assert not (nets_dict.keys() ^ dqns_dict.keys()), 'Net and DQN must match'
assert not (nets_dict.keys() ^ model_dict.keys()), 'Net and Model must match'
wins = collections.Counter()
total_wins = collections.Counter()
ai = {'up': None, 'lord': None, 'down': None}
for role in ['up', 'lord', 'down']:
if nets_dict.get(role) is not None:
print('AI based {}.'.format(role))
ai[role] = dqns_dict[role](nets_dict[role])
ai[role].policy_net.load(model_dict[role])
else:
print('Rule based {}.'.format(role))
env = env_cls(debug=debug)
start_time = time.time()
for episode in range(1, total + 1):
if debug:
print('\n-------------------------------------------')
env.reset()
env.prepare()
done = False
while not done:
for role in ['lord', 'down', 'up']:
if ai[role]:
action = ai[role].greedy_action(env.face, env.valid_actions())
_, done, _ = env.step_manual(action)
else:
_, done, _ = env.step_auto()
if done: # 地主结束本局，地主赢
wins[role] += 1
total_wins[role] += 1
break
if episode % print_every == 0:
end_time = time.time()
message = ('Reach at {}, Last {} rounds takes {:.2f}seconds\n'
'\tUp recent/total win rate: {:.2%}/{:.2%}\n'
'\tLord recent/total win rate: {:.2%}/{:.2%}\n'
'\tDown recent/total win rate: {:.2%}/{:.2%}\n')
args = (episode, print_every, end_time - start_time,
wins['up'] / print_every, total_wins['up'] / episode,
wins['lord'] / print_every, total_wins['lord'] / episode,
wins['down'] / print_every, total_wins['down'] / episode)
print(message.format(*args))
wins = collections.Counter()
start_time = time.time()
return total_wins