在<<C# BP神经网络 类与实例>>这篇文章中，笔者转载了网友提供的一个类。非常感谢这位网友的无私奉献，使笔者得以快速的完成了软件的一个功能。
貌似我找到的是网友对原作改动过的版本，使得该类存在着一些不足，笔者对其进行了相应改进。
1. 未保存训练样本的结果，使得每次使用时都要重新训练数据，耗费时间：
虽然类中已经提供了保存w、v、b1、b2数组的方法，但其构造函数却有问题，同时没有保存in_rate inNum HideNum outNum 这些下面会用到的数据。
对此笔者进行了以下改进：
新建了一个无参的构造函数用来预测新样本时实例化对象；
提供saveParas、readParas方法用于保存和读取这些参数；
提供initial方法用来建立预测过程中用到的动态数组
2. 类中用伪随机数来初始化w、v数组，使得每次建立新对象时w、v会不同。笔者对该算法没有深入研究，不清楚为什么这样设计。
笔者把
R = new Random();
这行代码进行了修改
R = new Random(32); //加了一个参数，使产生的伪随机序列相同
下面是我修改后的类和示例：
using System;
using System.IO;
using System.Text;
namespace BpANNet
{
/// <summary>
/// BpNet 的摘要说明。
/// </summary>
public class BpNet
{
public int inNum;//输入节点数
int hideNum;//隐层节点数
public int outNum;//输出层节点数
public int sampleNum;//样本总数
Random R;
double[] x;//输入节点的输入数据
double[] x1;//隐层节点的输出
double[] x2;//输出节点的输出
double[] o1;//隐层的输入
double[] o2;//输出层的输入
public double[,] w;//权值矩阵w
public double[,] v;//权值矩阵V
public double[,] dw;//权值矩阵w
public double[,] dv;//权值矩阵V
public double rate;//学习率
public double[] b1;//隐层阈值矩阵
public double[] b2;//输出层阈值矩阵
public double[] db1;//隐层阈值矩阵
public double[] db2;//输出层阈值矩阵
double[] pp;//输出层的误差
double[] qq;//隐层的误差
double[] yd;//输出层的教师数据
public double e;//均方误差
double in_rate;//归一化比例系数
public int computeHideNum(int m, int n)
{
double s = Math.Sqrt(0.43 * m * n + 0.12 * n * n + 2.54 * m + 0.77 * n + 0.35) + 0.51;
int ss = Convert.ToInt32(s);
return ((s - (double)ss) > 0.5) ? ss + 1 : ss;
}
public BpNet(double[,] p, double[,] t)
{
// 构造函数逻辑
R = new Random(32); //加了一个参数，使产生的伪随机序列相同
this.inNum = p.GetLength(1); //数组第二维大小为 输入节点数
this.outNum = t.GetLength(1); //输出节点数
this.hideNum = computeHideNum(inNum, outNum); //隐藏节点数，不知其原理
// this.hideNum=18;
this.sampleNum = p.GetLength(0); //数组第一维大小 为
Console.WriteLine("输入节点数目： " + inNum);
Console.WriteLine("隐层节点数目：" + hideNum);
Console.WriteLine("输出层节点数目：" + outNum);
Console.ReadLine();
x = new double[inNum];
x1 = new double[hideNum];
x2 = new double[outNum];
o1 = new double[hideNum];
o2 = new double[outNum];
w = new double[inNum, hideNum];
v = new double[hideNum, outNum];
dw = new double[inNum, hideNum];
dv = new double[hideNum, outNum];
b1 = new double[hideNum];
b2 = new double[outNum];
db1 = new double[hideNum];
db2 = new double[outNum];
pp = new double[hideNum];
qq = new double[outNum];
yd = new double[outNum];
//初始化w
for (int i = 0; i < inNum; i++)
{
for (int j = 0; j < hideNum; j++)
{
w[i, j] = (R.NextDouble() * 2 - 1.0) / 2;
}
}
//初始化v
for (int i = 0; i < hideNum; i++)
{
for (int j = 0; j < outNum; j++)
{
v[i, j] = (R.NextDouble() * 2 - 1.0) / 2;
}
}
rate = 0.8;
e = 0.0;
in_rate = 1.0;
}
//训练函数
public void train(double[,] p, double[,] t)
{
e = 0.0;
//求p，t中的最大值
double pMax = 0.0;
for (int isamp = 0; isamp < sampleNum; isamp++)
{
for (int i = 0; i < inNum; i++)
{
if (Math.Abs(p[isamp, i]) > pMax)
{
pMax = Math.Abs(p[isamp, i]);
}
}
for (int j = 0; j < outNum; j++)
{
if (Math.Abs(t[isamp, j]) > pMax)
{
pMax = Math.Abs(t[isamp, j]);
}
}
in_rate = pMax;
}//end isamp
for (int isamp = 0; isamp < sampleNum; isamp++)
{
//数据归一化
for (int i = 0; i < inNum; i++)
{
x[i] = p[isamp, i] / in_rate;
}
for (int i = 0; i < outNum; i++)
{
yd[i] = t[isamp, i] / in_rate;
}
//计算隐层的输入和输出
for (int j = 0; j < hideNum; j++)
{
o1[j] = 0.0;
for (int i = 0; i < inNum; i++)
{
o1[j] += w[i, j] * x[i];
}
x1[j] = 1.0 / (1.0 + Math.Exp(-o1[j] - b1[j]));
}
//计算输出层的输入和输出
for (int k = 0; k < outNum; k++)
{
o2[k] = 0.0;
for (int j = 0; j < hideNum; j++)
{
o2[k] += v[j, k] * x1[j];
}
x2[k] = 1.0 / (1.0 + Math.Exp(-o2[k] - b2[k]));
}
//计算输出层误差和均方差
for (int k = 0; k < outNum; k++)
{
qq[k] = (yd[k] - x2[k]) * x2[k] * (1.0 - x2[k]);
e += (yd[k] - x2[k]) * (yd[k] - x2[k]);
//更新V
for (int j = 0; j < hideNum; j++)
{
v[j, k] += rate * qq[k] * x1[j];
}
}
//计算隐层误差
for (int j = 0; j < hideNum; j++)
{
pp[j] = 0.0;
for (int k = 0; k < outNum; k++)
{
pp[j] += qq[k] * v[j, k];
}
pp[j] = pp[j] * x1[j] * (1 - x1[j]);
//更新W
for (int i = 0; i < inNum; i++)
{
w[i, j] += rate * pp[j] * x[i];
}
}
//更新b2
for (int k = 0; k < outNum; k++)
{
b2[k] += rate * qq[k];
}
//更新b1
for (int j = 0; j < hideNum; j++)
{
b1[j] += rate * pp[j];
}
}//end isamp
e = Math.Sqrt(e);
// adjustWV(w,dw);
// adjustWV(v,dv);
}//end train
public void adjustWV(double[,] w, double[,] dw)
{
for (int i = 0; i < w.GetLength(0); i++)
{
for (int j = 0; j < w.GetLength(1); j++)
{
w[i, j] += dw[i, j];
}
}
}
public void adjustWV(double[] w, double[] dw)
{
for (int i = 0; i < w.Length; i++)
{
w[i] += dw[i];
}
}
public BpNet() //仿真函数 用的构造函数
{
}
//数据仿真函数
public double[] sim(double[] psim) //in_rate inNum HideNum outNum
{
for (int i = 0; i < inNum; i++)
x[i] = psim[i] / in_rate;
for (int j = 0; j < hideNum; j++)
{
o1[j] = 0.0;
for (int i = 0; i < inNum; i++)
o1[j] = o1[j] + w[i, j] * x[i];
x1[j] = 1.0 / (1.0 + Math.Exp(-o1[j] - b1[j]));
}
for (int k = 0; k < outNum; k++)
{
o2[k] = 0.0;
for (int j = 0; j < hideNum; j++)
o2[k] = o2[k] + v[j, k] * x1[j];
x2[k] = 1.0 / (1.0 + Math.Exp(-o2[k] - b2[k]));
x2[k] = in_rate * x2[k];
}
return x2;
} //end sim
//保存矩阵w,v
public void saveMatrix(double[,] w, string filename)
{
StreamWriter sw = new StreamWriter(filename);
for (int i = 0; i < w.GetLength(0); i++)
{
for (int j = 0; j < w.GetLength(1); j++)
{
sw.Write(w[i, j] + " ");
}
sw.WriteLine();
}
sw.Close();
}
//保存矩阵b1,b2
public void saveMatrix(double[] b, string filename)
{
StreamWriter sw = new StreamWriter(filename);
for (int i = 0; i < b.Length; i++)
{
sw.Write(b[i] + " ");
}
sw.Close();
}
//保存参数 in_rate inNum HideNum outNum
public void saveParas(string filename)
{
try
{
StreamWriter sw = new StreamWriter(filename);
string str = inNum.ToString() + " "
+ hideNum.ToString() + " "
+ outNum.ToString() + " "
+ in_rate.ToString();
sw.WriteLine(str);
sw.Close();
}
catch (Exception e)
{
// Let the user know what went wrong.
Console.WriteLine("The file could not be read:");
Console.WriteLine(e.Message);
}
}
//读回参数 in_rate inNum HideNum outNum, tjt 预测新数据
public void readParas(string filename)
{
StreamReader sr;
try
{
sr = new StreamReader(filename);
String line;
if((line = sr.ReadLine()) != null)
{
string[] strArr = line.Split(' ');
this.inNum = Convert.ToInt32(strArr[0]);
this.hideNum = Convert.ToInt32(strArr[1]);
this.outNum = Convert.ToInt32(strArr[2]);
this.in_rate = Convert.ToDouble(strArr[3]);
}
sr.Close();
}
catch (Exception e)
{
// Let the user know what went wrong.
Console.WriteLine("The file could not be read:");
Console.WriteLine(e.Message);
}
}
public void initial() // 建立一些中间数组 tjt 预测新数据
{
x = new double[inNum];
x1 = new double[hideNum];
x2 = new double[outNum];
o1 = new double[hideNum];
o2 = new double[outNum];
w = new double[inNum, hideNum];
v = new double[hideNum, outNum];
dw = new double[inNum, hideNum];
dv = new double[hideNum, outNum];
b1 = new double[hideNum];
b2 = new double[outNum];
db1 = new double[hideNum];
db2 = new double[outNum];
pp = new double[hideNum];
qq = new double[outNum];
yd = new double[outNum];
}
//读取矩阵W,V
public void readMatrixW(double[,] w, string filename)
{
StreamReader sr;
try
{
sr = new StreamReader(filename);
String line;
int i = 0;
while ((line = sr.ReadLine()) != null)
{
string[] s1 = line.Trim().Split(' ');
for (int j = 0; j < s1.Length; j++)
{
w[i, j] = Convert.ToDouble(s1[j]);
}
i++;
}
sr.Close();
}
catch (Exception e)
{
// Let the user know what went wrong.
Console.WriteLine("The file could not be read:");
Console.WriteLine(e.Message);
}
}
//读取矩阵b1,b2
public void readMatrixB(double[] b, string filename)
{
StreamReader sr;
try
{
sr = new StreamReader(filename);
String line;
while ((line = sr.ReadLine()) != null)
{
int i = 0;
string[] s1 = line.Trim().Split(' ');
for (int j = 0; j < s1.Length; j++)
{
b[i] = Convert.ToDouble(s1[j]);
i++;
}
}
sr.Close();
}
catch (Exception e)
{
// Let the user know what went wrong.
Console.WriteLine("The file could not be read:");
Console.WriteLine(e.Message);
}
}
}//end bpnet
} //end namespace
示例：
//主调用程序
using System;
namespace BpANNet
{
/// <summary>
/// Class1 的摘要说明。
/// </summary>
class Class1
{
/// <summary>
/// 应用程序的主入口点。
/// </summary>
[STAThread]
static void Main(string[] args)
{
// double [,] p1=new double[,]{{0.05,0.02},{0.09,0.11},{0.12,0.20},{0.15,0.22},{0.20,0.25},{0.75,0.75},{0.80,0.83},{0.82,0.80},{0.90,0.89},{0.95,0.89},{0.09,0.04},{0.1,0.1},{0.14,0.21},{0.18,0.24},{0.22,0.28},{0.77,0.78},{0.79,0.81},{0.84,0.82},{0.94,0.93},{0.98,0.99}};
// double [,] t1=new double[,]{{1,0},{1,0},{1,0},{1,0},{1,0},{0,1},{0,1},{0,1},{0,1},{0,1},{1,0},{1,0},{1,0},{1,0},{1,0},{0,1},{0,1},{0,1},{0,1},{0,1}};
double[,] p1 = new double[,] { { 0.1399, 0.1467, 0.1567, 0.1595, 0.1588, 0.1622 }, { 0.1467, 0.1567, 0.1595, 0.1588, 0.1622, 0.1611 }, { 0.1567, 0.1595, 0.1588, 0.1622, 0.1611, 0.1615 }, { 0.1595, 0.1588, 0.1622, 0.1611, 0.1615, 0.1685 }, { 0.1588, 0.1622, 0.1611, 0.1615, 0.1685, 0.1789 } };
double[,] t1 = new double[,] { { 0.1622 }, { 0.1611 }, { 0.1615 }, { 0.1685 }, { 0.1789 }, { 0.1790 } };
BpNet bp = new BpNet(p1, t1);
int study = 0;
do
{
study++;
bp.train(p1, t1);
// bp.rate=0.95-(0.95-0.3)*study/50000;
// Console.Write("第 "+ study+"次学习： ");
// Console.WriteLine(" 均方差为 "+bp.e);
} while (bp.e > 0.001 && study < 50000);
Console.Write("第 " + study + "次学习： ");
Console.WriteLine(" 均方差为 " + bp.e);
bp.saveMatrix(bp.w, "w.txt");
bp.saveMatrix(bp.v, "v.txt");
bp.saveMatrix(bp.b1, "b1.txt");
bp.saveMatrix(bp.b2, "b2.txt");
bp.saveParas("para.txt");
pretect(); //开始预测新样本
}
public static void pretect()
{
Console.WriteLine("预测开始...");
BpNet bp = new BpNet();
bp.readParas("para.txt");
bp.initial();
bp.readMatrixW(bp.w,"w.txt");
bp.readMatrixW(bp.v, "v.txt");
bp.readMatrixB(bp.b1, "b1.txt");
bp.readMatrixB(bp.b2, "b2.txt");
// double [,] p2=new double[,]{{0.05,0.02},{0.09,0.11},{0.12,0.20},{0.15,0.22},{0.20,0.25},{0.75,0.75},{0.80,0.83},{0.82,0.80},{0.90,0.89},{0.95,0.89},{0.09,0.04},{0.1,0.1},{0.14,0.21},{0.18,0.24},{0.22,0.28},{0.77,0.78},{0.79,0.81},{0.84,0.82},{0.94,0.93},{0.98,0.99}};
double[,] p2 = new double[,] { { 0.1399, 0.1467, 0.1567, 0.1595, 0.1588, 0.1622 }, { 0.1622, 0.1611, 0.1615, 0.1685, 0.1789, 0.1790 } };
int aa = bp.inNum;
int bb = bp.outNum;
int cc = p2.GetLength(0);
double[] p21 = new double[aa];
double[] t2 = new double[bb];
for (int n = 0; n < cc; n++)
{
for (int i = 0; i < aa; i++)
{
p21[i] = p2[n, i];
}
t2 = bp.sim(p21);
for (int i = 0; i < t2.Length; i++)
{
Console.WriteLine("预测数据" + n.ToString() + "： " + t2[i] + " ");
}
}
Console.ReadLine();
}
}
}
4楼 tfg1025 2013-07-01 15:08发表 [回复]
错误很多，估计也是从网上直接拷贝过来的，在进行权值更新的时候，dw dv都没有初始化，学习率也没有用到 怎么更新？
3楼 sunyixiao 2013-01-18 16:57发表 [回复]
哥们，我很抱歉的通知你，你的程序在逻辑上有一点问题！我不知道你是否熟悉神经网络。你的程序在更新权值的时候出现了问题。你不能先更新v，再计算w的误差，因为w的误差是在v的情形下产生的。所以你应该先把v的更新保存在一个地方，让后计算w的误差并更新。之后再更新v
2楼 liuhh1 2012-08-21 16:48发表 [回复]
pp应该为隐层误差；qq是输出层误差。注释稍微改一下就行。程序很好 类里面的dw dv没有用到 含dw dv的那段代码可以不要
1楼 yannian1990 2012-06-07 16:53发表 [回复] [引用] [举报]
不错，很好的修改