def sum_data_influence_direct():
parameter = {
"region": 8,
"start_point": 5,
"n": 100, # 生成方差的次数
"sum_data_range": list(range(1000, 23000, 2000)),
"exp_times": 100, # 测量的重复次数
}
# 保存参数
func_name = sys._getframe().f_code.co_name
print(func_name, " is running at ", time.strftime("%Y-%m-%d_%H-%M-%S"))
data_saver = DataSaver(func_name)
data_saver.add_item("parameter", parameter)
# 均匀分布的参数
region = parameter["region"]
start_point = parameter["start_point"]
# 确定数据量变化的范围
x_axis = sum_data_range = parameter["sum_data_range"]
# 确定产生方差的次数
n = parameter["n"]
# 确定测量重复的次数
exp_times = parameter["exp_times"]
y_axis = []
for sum_data in sum_data_range:
print(sum_data, time.strftime("%Y-%m-%d_%H-%M-%S"))
MSE = 0
for i in range(n):
var_vector = np.random.randint(start_point,
start_point + region + 1, 5)
# 计算最优比例
topo = Topology(var_vector=var_vector, way="Direct")
# 测量
mse = 0
for j in range(exp_times):
cov = topo.gen_delay(topo.reduced_matrix, topo.Phi, sum_data)
measured_X = topo.cal_measured_link_parameter(cov)
mse += np.dot((np.array(measured_X) - np.array(var_vector))**2,
topo.H_x) / exp_times
MSE += mse / n
y_axis.append(MSE)
# 画图
print("x_axis:", x_axis)
print("y_axis:", y_axis)
plt.figure()
plt.plot(x_axis, y_axis, marker="o", color="blue", label="Direct")
plt.xlabel("#probe")
plt.ylabel("MSE")
# plt.title("the influence of #probe")
plt.legend()
dir = os.path.join(os.getcwd(), "data", func_name)
fig_name = dir + "/" + func_name + "_" + time.strftime(
"%Y-%m-%d_%H-%M-%S") + ".png"
if not os.path.exists(dir):
os.makedirs(dir)
plt.savefig(fig_name)
plt.show()
plt.close()
data_saver.add_item("x_axis", x_axis)
data_saver.add_item("y_axis", y_axis)
data_saver.to_file()
def uniform_direct():
"""
按照 Fig 1.a 选定合适的 链路时延方差
Fig 1.b x-axis: k 步逼近，y-axis： NT的估计误差 （所有 链路的 平均估计误差）；重复实验k次
最直接的一个方案：分成 k 步来逼近，每步按照 (最优比例-均匀比例)/k 来修正上一步的比例
是的哈，按照上面的式子，你肯定最终 是收敛到最优方案的
"""
parameter = {
"var_vector": [5, 8, 16, 12, 10], # 由图1 a中
"average_proportion": [0.2, 0.2, 0.2, 0.2, 0.2],
"k_step": 10,
"sum_data": 10000,
"exp_times": 500
}
# 保存参数
func_name = sys._getframe().f_code.co_name
print(func_name, "is running at ", time.strftime("%Y-%m-%d_%H-%M-%S"))
data_saver = DataSaver(func_name)
data_saver.add_item("parameter", parameter)
# 计算最优比例,直接测量
var_vector = parameter["var_vector"]
direct_topo = Topology(var_vector=var_vector, way="Direct")
direct_proportion = direct_topo.Phi
data_saver.add_item("direct_proportion", direct_proportion)
# 平均比列
average_proportion = parameter["average_proportion"]
k_step = parameter["k_step"] # 分多少次接近最优比列
x_axis = list(range(k_step + 1))
all_proportion = []
for i in range(k_step):
proportion = np.array(average_proportion) + (np.array(
direct_proportion) - np.array(average_proportion)) / k_step * i
all_proportion.append(proportion.tolist())
all_proportion.append(direct_proportion)
# 这里各个链路占的比例应该也从平均比例变化到H_x
H_x_proportions = []
for i in range(k_step):
proportion = np.array(average_proportion) + (np.array(
direct_topo.H_x) - np.array(average_proportion)) / k_step * i
H_x_proportions.append(proportion.tolist())
H_x_proportions.append(direct_topo.H_x.tolist())
data_saver.add_item("data_proportion", all_proportion)
data_saver.add_item("H_x_proportions", H_x_proportions)
y_axis = []
# 数据总量
sum_data = parameter["sum_data"]
exp_times = parameter["exp_times"] # 实验重复次数
for index, proportion in enumerate(all_proportion):
print(index, time.strftime("%Y-%m-%d_%H-%M-%S"))
mse = 0
for _ in range(exp_times):
cov = direct_topo.gen_delay(direct_topo.reduced_matrix, proportion,
sum_data)
measured_X = direct_topo.cal_measured_link_parameter(cov)
# mse += np.mean((np.array(measured_X) - np.array(var_vector))**2)/exp_times
mse += np.dot((np.array(measured_X) - np.array(var_vector))**2,
H_x_proportions[index]) / exp_times
y_axis.append(mse)
print("x_axis:", x_axis)
print("y_axis:", y_axis)
# 画图
plt.figure()
plt.plot(x_axis, y_axis, marker="o", color="blue")
plt.xlabel("step")
plt.ylabel("MSE")
plt.legend()
plt.title("average proportion to FIM proportion")
dir = os.path.join(os.getcwd(), "data", func_name)
fig_name = dir + "/" + func_name + "_" + time.strftime(
"%Y-%m-%d_%H-%M-%S") + ".png"
if not os.path.exists(dir):
os.makedirs(dir)
plt.savefig(fig_name)
plt.show()
plt.close()
data_saver.add_item("x_axis", x_axis)
data_saver.add_item("y_axis", y_axis)
data_saver.to_file()
def region_error_direct():
"""
画出均匀分布区间和估计误差
Fig 1.a 所有的链路的时延方差 服从均匀分布（区间长度从 0 开始；0 区间长度表示所有链路具有相同的时延方差），x-axis：区间的长度，y-axis： NT的估计误差 （所有 链路的 平均估计误差）；重复实验k次
:return:
"""
# 参数设置
parameter = {
"regions": list(range(0, 20 + 2, 2)),
"start_point": 5,
"exp_times": 100, # 测量的重复次数
"n": 100, # 生成方差的次数
"sum_data": 10000,
}
# 保存参数
func_name = sys._getframe().f_code.co_name
print(func_name, " is running at ", time.strftime("%Y-%m-%d_%H-%M-%S"))
data_saver = DataSaver(func_name)
data_saver.add_item("parameter", parameter)
x_axis = regions = parameter["regions"] # 设定均匀分布的区间长度值
start_point = parameter["start_point"] # 设置均匀分布的中间值
exp_times = parameter["exp_times"]
n = parameter["n"]
# 数据总量
sum_data = parameter["sum_data"]
y_axis = []
for region in regions:
print(region, time.strftime("%Y-%m-%d_%H-%M-%S"))
Direct_mse = 0
for i in range(n):
# 1.均匀分布生成方差
var_vector = np.random.randint(start_point,
start_point + region + 1, 5)
# 2.构建topo
topo = Topology(var_vector=var_vector, way="Direct")
direct_mse = 0
for j in range(exp_times):
# 3.生成数据，并且推断
# FIM值
direct_cov = topo.gen_delay(topo.reduced_matrix, topo.Phi,
sum_data)
direct_measured_X = topo.cal_measured_link_parameter(
direct_cov)
# 计算mse
direct_mse += np.dot((np.array(direct_measured_X) - var_vector)
**2, topo.H_x) / exp_times
Direct_mse += direct_mse / n
y_axis.append(Direct_mse)
data_saver.add_item("Direct_mse", y_axis)
plt.figure()
plt.plot(regions, y_axis, label="Direct", marker=".", color="black")
plt.xlabel("scale")
plt.ylabel("MSE")
# plt.ylim([0,1])
plt.legend()
# plt.title("the variation of scale vs MSE")
dir = os.path.join(os.getcwd(), "data", func_name)
fig_name = dir + "/" + func_name + "_" + time.strftime(
"%Y-%m-%d_%H-%M-%S") + ".png"
if not os.path.exists(dir):
os.makedirs(dir)
plt.savefig(fig_name)
plt.show()
plt.close()
data_saver.to_file()
def cdf_mse():
"""
选出Fig 1 中 区间长度为0，中间，最大共3个值：
Fig 2,3,4 给出对应的 3副 CDF 统计图 ，x-axis: 估计误差（注意 这里是统计每条链路每一次得到的 估计误差），y-axis:估计误差的累积概率
:return:
"""
parameter = {
# "regions": [0,10,20],
"regions": [15],
"start_point": 5,
"n": 2000, # 生成方差的次数
"sum_data": 10000,
"exp_times": 100, # 测量的重复次数
}
# 保存参数
func_name = sys._getframe().f_code.co_name
data_saver = DataSaver(func_name)
data_saver.add_item("parameter", parameter)
# 均匀分布的参数,0，中间，最大
regions = parameter["regions"]
start_point = parameter["start_point"]
# 确定产生方差的次数
n = parameter["n"]
# 确定测量重复的次数
exp_times = parameter["exp_times"]
# 产生的总数据量
sum_data = parameter["sum_data"]
for region in regions:
print(region)
MSE = np.empty(shape=[0, 5])
for i in range(n):
var_vector = np.random.randint(start_point,
start_point + region + 1, 5)
# 计算最优比例
topo = Topology(var_vector)
optimal_proportion = topo.Phi
# 测量
mse = np.array([0, 0, 0, 0, 0], dtype=float)
for j in range(exp_times):
cov = topo.gen_delay(topo.reduced_matrix, optimal_proportion,
sum_data)
measured_X = topo.cal_measured_link_parameter(cov)
mse += (np.array(measured_X) -
np.array(var_vector))**2 / exp_times
MSE = np.row_stack((MSE, mse))
MSE_data = MSE.flatten()
# 画CDF图
plt.figure()
plt.hist(MSE_data, cumulative=True, normed=True)
plt.xlabel("MSE")
plt.ylabel("CDF")
# plt.title("the cdf ("+str(region)+")")
plt.legend()
dir = os.path.join(os.getcwd(), "data", func_name)
fig_name = dir + "/" + func_name + "_" + str(
region) + "_" + time.strftime("%Y-%m-%d_%H-%M-%S") + ".png"
if not os.path.exists(dir):
os.makedirs(dir)
plt.savefig(fig_name)
plt.show()
plt.close()
# 对MSE数据进行保存
data_saver.add_item(str(region), MSE)
data_saver.to_file()
def test(times):
data_saver = DataSaver("qlearning_test")
data_saver.add_item("start_time", time.strftime("%Y-%m-%d_%H-%M-%S"))
from src.envs.ec.ec_env import ECMA
env = ECMA()
RL = QLearningTable()
data_saver.add_item("qlearning_param",RL.param)
data_saver.add_item("episodes",times)
for episode in range(0, times):
env.reset()
observation = env.get_obs()
observation = list(np.round(np.array(observation,dtype=float).flatten(), 3))
while True:
action = RL.choose_action(str(observation))
refined_action = RL.refine_action(action)
reward, done ,_ = env.step(refined_action)
observation_ = env.get_obs()
observation_ = list(np.round(np.array(observation_,dtype=float).flatten(), 3))
observation = copy.deepcopy(observation_)
data_saver.append(str(episode),[observation,refined_action,reward])
if done:
if episode / 1000 > 0:
print(episode, time.strftime("%Y-%m-%d_%H-%M-%S"))
break
data_saver.add_item("end_time", time.strftime("%Y-%m-%d_%H-%M-%S"))
data_saver.to_file()
def train(times=10000):
data_saver = DataSaver("qlearning_training")
data_saver.add_item("start_time", time.strftime("%Y-%m-%d_%H-%M-%S"))
#创建环境
env = create_env()
RL = QLearningTable()
data_saver.add_item("RL_param",RL.param)
data_saver.add_item("episodes",times)
for episode in range(0, times):
env.reset()
observation = env.get_obs()
observation = list(np.round(np.array(observation,dtype=float).flatten(), 3))
while True:
action = RL.choose_action(str(observation))
refined_action = RL.refine_action(action)
reward, done ,_ = env.step(refined_action)
observation_ = env.get_obs()
observation_ = list(np.round(np.array(observation_,dtype=float).flatten(), 3))
RL.learn(str(observation), action, reward, str(observation_),done)
observation = copy.deepcopy(observation_)
data_saver.append(str(episode),[observation,refined_action,reward])
if done:
if episode / 1000 > 0:
print(episode, time.strftime("%Y-%m-%d_%H-%M-%S"))
if episode / 10000:
RL.save_model()
break
data_saver.add_item("end_time", time.strftime("%Y-%m-%d_%H-%M-%S"))
data_saver.to_file()