GRNN（广义回归神经网络）-RBF的一种

GRNN网路结构如图所示，包括四层：输入层、模式层、求和层和输出层。

1.输入层

节点数等于训练样本的个数

2.模式层

一般使用高斯函数对输入数据进行处理，节点数为训练样本的个数，具体的计算公式如下：

其中x_i为训练样本，x_j为学习样本，σ为平滑因子，

3.求和层

节点数为输出样本数+1,

其中一个节点输出S_D为模式层输出的算术和，其余节点输出S_Ni均为模式层输出的加权和，具体的计算公式如下，其中ω_ij为加权系数，即第i个输入对应的输出。

4.输出层

该层节点数为输出样本维度，主要根据求和层求出的算术和与加权和进行计算输出，计算公式如下：

注意：平滑因子对网络性能影响较大，需要使用优化算法优化。

实例：Generalized-regression-neural-networks-library-from-scratch

python实现：https://www.codetd.com/article/14610679

#库的导入
import numpy as np
import pandas as pd

#输入层
#读取训练数据
print('------------------------1. Load train data------------------------')
df = pd.read_csv("train.csv")
df.columns = ["Co", "Cr", "Mg", "Pb", "Ti"]
Co = df["Co"]
Co = np.array(Co)
Cr = df["Cr"]
Cr = np.array(Cr)
Mg=df["Mg"]
Mg=np.array(Mg)
Pb = df["Pb"]
Pb =np.array(Pb)
Ti = df["Ti"]
Ti = np.array(Ti)
inputX = np.mat([Co,Cr,Mg,Pb])
inputX = inputX.transpose()
inputY = np.mat(Ti)
inputY = inputY.transpose()


#读取测试数据
print('------------------------2. Load test data-------------------------')
df = pd.read_csv("test.csv")
df = pd.read_csv("test.csv")
df.columns = ["Co", "Cr", "Mg", "Pb", "Ti"]
Co = df["Co"]
Co = np.array(Co)
Cr = df["Cr"]
Cr = np.array(Cr)
Mg=df["Mg"]
Mg=np.array(Mg)
Pb = df["Pb"]
Pb =np.array(Pb)
Ti = df["Ti"]
Ti = np.array(Ti)
testX = np.mat([Co,Cr,Mg,Pb])
testX = testX.transpose()
testY = Ti

#模式层
#计算样本欧式距离
print('----------------3. Calculate euclidean distance-------------------')
m, n = np.shape(inputX)
p = np.shape(testX)[0]
distance = np.mat(np.zeros((p, m)))
for i in range(p):
    for j in range(m):
        distance[i,j] = np.linalg.norm(testX[i,:]-inputX[j,:])

#计算高斯矩阵
print('------------------4. Calculate gaussian matrix--------------------')
sigma = 2 #平滑因子
Gauss = np.mat(np.zeros((p,m)))
for i in range(p):
    for j in range(m):
        Gauss[i,j] = np.exp(- distance[i,j] / (2 * (sigma ** 2)))

#求和层
print('----------------------5. Output of sum Layer----------------------')
n = np.shape(inputY)[1]
sum_mat = np.mat(np.zeros((p,n+1)))
#计算算术和
for i in range(p):
    sum_mat[i, 0] = np.sum(Gauss[i, :], axis=1)
#计算加权和
for i in range(p):
    for j in range(n):
        total = 0.0
        for s in range(m):
            total += Gauss[i,s] * inputY[s,j]
        sum_mat[i,j+1] = total

#输出层
print('--------------------6. Output of output Layer---------------------')
#计算预测值
predict = np.mat(np.zeros((p,n)))
for i in range(n):
    predict[:,i] = sum_mat[:,i+1] / sum_mat[:,0]
predict = predict.transpose()
predict = np.array(predict)
output1=predict.flatten()#降成一维数组
predict=output1.tolist()

#预测效果评估
print('----------------7. Forecasting effect evaluation------------------')
#预测差值err
err = predict - testY
#MAE等评价指标的计算
mae = np.sum(np.abs(predict-testY))/p
average_loss1=np.sum(np.abs((predict-testY)/testY))/p
mape="%.2f%%"%(average_loss1*100)
f1 = 0
for m in range(p):
    f1 = f1 + np.abs(testY[m]-predict[m])/((np.abs(testY[m])+np.abs(predict[m]))/2)
f2 = f1 / p
smape="%.2f%%"%(f2*100)

#计算预测值与真实值误差与真实值之比的分布
A=0
B=0
C=0
D=0
E=0
for m in range(p):
    y1 = np.abs(testY[m]-predict[m])/np.abs(testY[m])
    if y1 <= 0.1:
        A = A + 1
    elif y1 > 0.1 and y1 <= 0.2:
        B = B + 1
    elif y1 > 0.2 and y1 <= 0.3:
        C = C + 1
    elif y1 > 0.3 and y1 <= 0.4:
        D = D + 1
    else:
        E = E + 1
print("The distribution of the predicted difference ratio in different intervals is as follows:")
print("Ratio <= 0.1 :",A)
print("0.1< Ratio <= 0.2 :",B)
print("0.2< Ratio <= 0.3 :",C)
print("0.3< Ratio <= 0.4 :",D)
print("Ratio > 0.4 :",E)
print("The different error index values are as follows：")
print("the MAE is :",mae)
print("the MAPE is :",mape)
print("the SMAPE is :",smape)

#保存误差和真实值
np.save("GRNN-err.npy",err)
np.save("GRNN-output.npy",predict)

print("The prediction value and real value comparison figure has been generated ！")

个人感觉GRNN就是计算测试样本和训练样本之间的相似度，所有的现有数据都会对预测起作用，是另一种形式的归一化，但是这种方法应该更适合小样本，或者写好了优化算法的。不然调起来很麻烦。