-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
179 lines (156 loc) · 9.52 KB
/
main.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
#coding: utf-8
'''主调度函数'''
import logging
import pandas as pd
import itertools
import random
import conf
from network.agent_nn import AgentDeepNetwork
from system.system_simul_v2 import AHPSystemSimulator
from utils_global import *
# 配置日志输出
logging.basicConfig(level=10) #DEBUG level
class MaintainWorker:
def initialization(self): #debug DONE
'''系统仿真模块、神经网络模块初始化；参数和中间变量初始化'''
# 部件信息
self.agents = conf.AGENT_COMPONENTS
self.sys_comp_info, self.comp_agent_mapping, self.flatten_colname_list, self.flatten_colname_list_agents = construct_comp_info(conf.SYSTEM_COMPONENT, self.agents)
self.num_comp = len(self.sys_comp_info)
self.length_of_comp_action = len(self.sys_comp_info[0]['comp_action_fc'])
self.num_agents = len(self.agents)
# 系统模块初始化
self.system_simulator = AHPSystemSimulator(self.sys_comp_info)
self.system_simulator.init_system()
# 神经网络模块初始化
agents = conf.AGENT_COMPONENTS
num_agent = len(agents)
self.agent_nns = []
for agent_index in range(num_agent):
agent_comp_info = []
for comp_index in agents[agent_index]:
agent_comp_info.append(self.sys_comp_info[comp_index])
self.agent_nns.append(AgentDeepNetwork(agent_index, agents[agent_index], agent_comp_info))
self.agent_nns[agent_index].init_network()
# 行动选择
self.action_select_strategy = conf.ACTION_SELECT_STRATEGY
self.action_select_params = conf.ACTION_SELECT_PARAMS
# 训练过程
self.agent_nn_batch_size = conf.BATCH_SIZE
self.termination_criterion = conf.TERMINATION_CRITERION
self.max_termination_num_iteration = conf.MAX_TERMINATION_NUM_ITERATION
# 历史样本存储和回放
self.current_sample_collection_org = []
self.history_memory_df = pd.DataFrame(columns=self.flatten_colname_list.extend(['total_epoch', 'train_epoch']))
self.replay_one_batch_size = conf.REPLAY_ONE_BATCH_SIZE
self.replay_batch_num = conf.REPLAY_BATCH_NUM
self.replay_after_epoch = conf.REPLAY_AFTER_EPOCH
def simulation(self, selected_action = None): #默认selected_action是None即不采取行动，用于initialization阶段自然转移
'''系统采取action并进行仿真，直至遇到决策时间点，返回上次产生的成本，以及本次待干预的系统状态向量'''
last_reward = self.system_simulator._progress_one_action(selected_action)
system_states = self.system_simulator._progress_one_epoch()
feasible_actions = self.system_simulator._update_feasible_action()
return last_reward, system_states, feasible_actions
def decision(self, system_states, feasible_actions):
'''将系统状态传给神经网络作为样本，返回最优决策'''
def gen_test_sample_agents(system_states, feasible_actions):
# 将state和action按照agent分割
agent_test_sample_state = []
agent_feasible_actions = [[]] * self.num_agent
for comp_index in range(self.num_comp):
agent_test_sample_state[comp_index].append(system_states[comp_index])
agent_index = self.comp_agent_mapping(str(comp_index))
agent_feasible_actions[agent_index].append(feasible_actions[comp_index])
agent_test_sample_actions = [[]] * self.num_agents
agent_test_sample = [[]] * self.num_agents
flatten_test_sample_agents_df_list = []
for agent_index in range(self.num_agents):
# 遍历形成所有可行的action样本
feasible_actions_agent = agent_feasible_actions[agent_index]
feasible_iter = [value_list for comp_value_list in feasible_actions_agent for value_list in comp_value_list]
p_res = itertools.product(*feasible_iter)
len_action = self.length_of_comp_action
for s in p_res:
test_sample_action_org = [s[i: i+len_action] for i in range(0, len(s), len_action)] #CHECK: 是否有重叠
agent_test_sample_actions[agent_index].append(test_sample_action_org)
# 将state和action进行拼接（都是organize格式）
for action in agent_test_sample_actions[agent_index]:
agent_test_sample[agent_index].append([agent_test_sample_state[agent_index], action]) #在前面拼接state
# 利用colname_agents转换成flatten格式的sample。可以直接复用sparse_flatten那个函数，需要额外传进一个对应的colname_list
colname_list = self.flatten_colname_list_agents[agent_index]
test_sample = agent_test_sample[agent_index]
flattened_test_sample_df = sample_parse_flatten(test_sample, colname_list) #CHECK: 函数兼容性
flatten_test_sample_agents_df_list.append(flattened_test_sample_df)
return flatten_test_sample_agents_df_list
flatten_test_sample_agents_df_list = gen_test_sample_agents(system_states, feasible_actions)
# 对每个agent nn送入样本进行test
pred_value = []
agent_test_index = 0
for an in self.agent_nns:
pred_value.append(an._predict(flatten_test_sample_agents_df_list[agent_test_index]))
agent_test_index += 1
# 根据value选择每个agent的策略
selected_action_dict = {}
for agent_index in range(len(pred_value)):
agent_best_action_index = select_best_action_agent(pred_value[agent_index], params = self.action_select_params, strategy = self.action_select_strategy)
agent_best_action = agent_test_sample_actions[agent_index][agent_best_action_index]
# 合并入dict结果
agent_comps = self.agents[agent_index]
for i in range(len(agent_comps)):
comp_index = agent_comps[i]
selected_action_dict[comp_index] = agent_best_action[i] #CHECK: 两个i是否对齐（正确的部件对应正确的行动）
return selected_action_dict
# TODO:10.21调试
def sample_collect(self, system_states, selected_action_dict, reward):
'''将系统产生的样本添加到当前样本集合中'''
system_action_org = []
for comp_index in range(self.num_comp):
system_action_org.append(selected_action_dict[comp_index])
self.current_sample_collection_org.append([system_states, system_action_org, reward])
def memory_replay(self):
'''从历史样本池中拿一部分连续样本返回'''
replay_train_sample_df = pd.DataFrame(columns=self.flatten_colname_list.extend(['total_epoch', 'train_epoch']))
current_memory_size = self.history_memory_df.shape[0]
for i in range(self.replay_batch_num):
index = random.randrange(0, current_memory_size)
replay_batch_df = self.history_memory_df.iloc[index: min(index+self.replay_one_batch_size, current_memory_size), :] #CHECK:左右开闭情况
replay_train_sample_df = pd.concat([replay_train_sample_df, replay_batch_df], axis=0)
return replay_train_sample_df
def network_train(self):
'''样本凑够一个batch后，送入神经网络进行训练'''
flatten_train_sample_df = sample_parse_flatten(self.current_sample_collection_org)
# memory replay
if self.total_epoch > self.replay_after_epoch:
flatten_train_sample_df = pd.concat([flatten_train_sample_df, self.memory_replay()], axis=0)
flatten_train_sample_agents_df_list = split_agent_flatten_sample(flatten_train_sample_df, self.flatten_colname_list, self.agents, self.comp_agent_mapping)
agent_train_index = 0
for an in self.agent_nns:
an._train(flatten_train_sample_agents_df_list[agent_train_index])
agent_train_index += 1
flatten_train_sample_df['total_epoch'] = self.total_epoch
flatten_train_sample_df['train_epoch'] = self.train_epoch
self.history_memory_df = pd.concat([self.history_memory_df, flatten_train_sample_df], axis=0)
def _go_to_work(self):
#初始化
self.initialization()
#执行系统仿真、神经网络训练
self.total_epoch = 0
self.train_epoch = 0
_, system_states, feasible_actions = self.simulation() #第一步冷启动
while not conf.is_terminate(self.total_epoch, self.termination_criterion, self.max_termination_num_iteration):
self.total_epoch += 1 #第一步索引是epoch=1
selected_action_dict = self.decision(system_states, feasible_actions)
reward, system_states, feasible_actions = self.simulation(selected_action_dict)
self.sample_collect(system_states, selected_action_dict, reward)
if self.total_epoch % self.agent_nn_batch_size == 0:
self.train_epoch += 1
self.network_train()
def _report_kpi(self):
# TODO: 对关键指标做埋点，输出结果数值和指标
return 0
# TODO: 组织起来之后，继续做一些抽象，把高频功能函数集中到一起。目前能想到的是：如何组织system层面信息和agent层面信息，最好能有一个转换机制
# TODO: 调试方法。先局部验证，然后进行关键结果做日志埋点，再调试。
if __name__ == "__main__":
maintain_worker = MaintainWorker()
maintain_worker._go_to_work()
# maintain_worker._report_kpi()