10阶多项式拟合10000个点。点的生成，使用bp思想拟合。Polynomial_fitting

用10阶多项式拟合10000个点

import matplotlib.pyplot as plt
import random

一、数据生成

#首先保证这10000个点x各不相同，从0-20000内随机生成不同的数,除以100，将x值变为0-200
num = random.sample(range(1,20000),10000)
num_x = []
for i in range(10000):
    num_x.append(num[i] / 100)
    
#y也一样,y的生成不需要不一样
num_y = [random.randint(0,100) for i in range(10000)]
print(len(num_y))

10000

#观察点
fig,ax = plt.subplots()
ax.set_ylabel('X')
ax.set_xlabel('Y')
ax.set_title('Polynomial Fitting')
ax.plot(num_x,num_y)

[<matplotlib.lines.Line2D at 0x25fd50f8048>]

• 这些点根本没办法拟合，重新弄数据，想到一种办法，可以先用10次多项式算10000个点，然后加上随机噪声

• 构造多项式

#构造十阶多项式, 该列表中从前到后对应10次，9次，8次等
polynomial = [random.random() for i in range(10)]
print(polynomial)
#计算多项式的值
def calculate(polynomial,x):
    n = len(polynomial)
    sum = polynomial[0]
    for i in range(1,n):
        sum *= x
        sum += polynomial[i]
    return sum

[0.053340626373557076, 0.2183335740410205, 0.9030944136075595, 0.7048816191727242, 0.18165101344430157, 0.9319508221357489, 0.4854520866475778, 0.5540870360221756, 0.5109581693655714, 0.5327543213626857]

• 重新生成数据

n = 10000
polynomial_true = [0.02,0.012,-0.06,0.03,-0.04,0.05,0.4,0.1,-0.2,0.3,10] #真实多项式
#依旧采用上面的方法作为x
num_x = []
num_x = set(num_x)
while len(num_x)<n:
    k = (random.random()-0.5)*5
    num_x.add(k)
    num_x = set(num_x)
print(len(num_x))
num_x = list(num_x)
num_y = []
for i in range(n):
#     temp = calculate(polynomial_true,num_x[i]) + random.randint(0,10) #真实值加上一个随机噪声
    temp = calculate(polynomial_true,num_x[i]) #想想其实不用随机噪声
    num_y.append(temp)
print(len(num_y))

10000
10000

#观察点
fig,ax = plt.subplots()
ax.set_ylabel('Y')
ax.set_xlabel('X')
ax.set_title('Polynomial Fitting')
ax.scatter(num_x,num_y)

<matplotlib.collections.PathCollection at 0x25fd8486320>

二、使用bp的思想做，构造预测多项式，减少多项式偏差

设预测值为o，真实值为y，那么误差 $loss=\frac{1}{2}(o-y{)}^2$,定义$l=(o-y)$其中y为

$$y=a_1{x}^{10}+a_2{x}^9+a_3{x}^8+a_4{x}^7+a_5{x}^6+a_6{x}^5+a_7{x}^4+a_8{x}^3+a_9{x}^2+a_{10}x+a_{11}$$

此时调整，需要向误差减小最明显的方向调整，用l分别对$a_1,a_2\dots \dots$求导得到，$-x^{10}l,-x^{9}l,-x^{8}l\dots \dots$

设学习率为lr,那么$a_1,a_2,\dots \dots$的调整值为$$-lr{x}^{10}l,-lr{x}^{9}l$$

归纳为,list中第i个元素每次的改变量为$$-lr{x}^{10-i}l,i=0,1,\dots \dots ,9$$

• 单单上面这种推导实现起来爆炸了，改进想法是，将所有的loss加起来除以10000，再去调整，以及使用绝对值代替二次函数防止溢出。
  调整学习率使每次学习到合适的量，同时再根据x次数不同项，乘上一个权值代替指数的$x^{10-i}$，即$lr\times x\times l\times (10-i)$。加上dropout。
  
  误差最好的时候降到35.拟合程度大体差不多。

random.seed(0)
polynomial_pre = [random.random()-0.5 for i in range(10)] #推测多项式，先随机赋值
print(polynomial_pre)
#将每个点带入拟合
epoch = 800
lr = 0.0000005
l = 0
for i in range(epoch):
    loss = []
    for j in range(n):
        loss.append(calculate(polynomial_pre,num_x[j]) - num_y[j])
        l += abs(calculate(polynomial_pre,num_x[j]) - num_y[j])
    l/=n
    if i%100 == 0:
        print('第',i,'轮误差',l)
    for k in range(10):
#         drop = random.randint(0,11)
#         if k == drop:
#             continue
        polynomial_pre[k] -= lr * (num_x[k]) * (loss[k] - num_y[j])*(10-k) 
print(polynomial_pre)

[0.3444218515250481, 0.2579544029403025, -0.079428419169155, -0.24108324970703665, 0.01127472136860852, -0.09506586254958571, 0.2837985890347726, -0.19668727392107255, -0.02340304584764419, 0.0833820394550312]
第 0 轮误差 125.56605977264955
第 100 轮误差 96.63328691501964
第 200 轮误差 76.40355209636932
第 300 轮误差 61.9613787759121
第 400 轮误差 51.27587641809833
第 500 轮误差 43.080682824405415
第 600 轮误差 37.97494996982369
第 700 轮误差 35.637086267932055
[0.3076666374487756, 0.4425229539123861, -1.095939035078886, -1.161408091549099, -1.335224244129828, -0.001029364227706257, 0.32750621887204023, -2.3130008991325552, 0.018007058085728295, -0.16715429895537884]

# 此时用新的曲线绘图，和原来比较
num_pre_y = [calculate(polynomial_pre,num_x[i]) for i in range(n)]
#观察点
fig,ax = plt.subplots()
ax.set_ylabel('Y')
ax.set_xlabel('X')
ax.set_title('Polynomial Fitting')
ax.scatter(num_x,num_pre_y)

<matplotlib.collections.PathCollection at 0x25fd8662e10>