www.pudn.com > RBPNN.rar > CRBPNN.CPP, change:2006-06-19,size:12551b
#include
#include
#include"CRBPNN.H"
#include
#include
#include
CRBPNN::CRBPNN()
{
this->m_pN=0;
this->m_pD=0;
this->m_pC=0;
this->m_in=0;
this->m_out=0;
this->m_mid1=0;
this->m_mid2=0;
this->m_kernel=NULL;
this->m_learningalgorithm=NULL;
this->v_pNC=NULL;
this->v_pSort=NULL;
this->v_pClass=NULL;
this->v_pX=NULL;
this->v_expectD=NULL;
this->v_realY=NULL;
this->v_conS=NULL;
this->v_midZ1=NULL;
this->v_midZ2=NULL;
this->v_weight=NULL;
}
CRBPNN::~CRBPNN()
{
int i;
for(i=0;im_mid1;i++)
delete[] this->v_conS[i];
for(i=0;im_out;i++)
delete[] this->v_weight[i];
delete[] this->v_conS,this->v_weight,this->v_pNC;
if(this->v_realY!=NULL)
{
for(i=0;im_pN;i++)
delete[] this->v_realY[i];
delete[] this->v_realY;
}
}
void CRBPNN::Training(int pn,
int pd,
int pc,
char* kernel,
char* learning,
char* sampledata)
{
if(pn>0&&pd>0&&pc>0&&kernel!=NULL&&learning!=NULL&&sampledata!=NULL)
{
//////////////////////////////////////////
//初始化网络，分配空间
//////////////////////////////////////////
int i,j;
this->m_pN=pn;
this->m_pD=pd;
this->m_pC=pc;
this->m_in=pd;
this->m_kernel=kernel;
this->m_learningalgorithm=learning;
this->m_mid1=pn;
this->m_mid2=pc;
this->m_out=pc;
this->v_pNC=new int[this->m_pC];
this->v_pSort=new int[this->m_pN];
this->v_pClass=new int[this->m_pN];
this->v_pX=new double*[this->m_pN];
this->v_expectD=new double*[this->m_pN];
this->v_conS=new double*[this->m_mid1];
this->v_midZ1=new double*[this->m_pN];
this->v_midZ2=new double*[this->m_pN];
this->v_weight=new double*[this->m_out];
for(j=0;jm_pN;j++)
{
this->v_pX[j]=new double[this->m_in];
this->v_expectD[j]=new double[this->m_out];
this->v_midZ1[j]=new double[this->m_mid1];
this->v_midZ2[j]=new double[this->m_mid1];
}
for(j=0;jm_mid1;j++)
this->v_conS[j]=new double[this->m_in];
for(j=0;jm_out;j++)
this->v_weight[j]=new double[this->m_mid2];
ifstream fin(sampledata,ios::nocreate);
if(fin.fail())
{
cerr<<"Error opening the file:"<m_pN;i++)
{
int c;
fin>>c;
this->v_pClass[i]=c; //类别标志
for(j=0;jm_out;j++)
{
if(j==(c-1)) this->v_expectD[i][j]=1;
else this->v_expectD[i][j]=0;
}
for(j=0;jm_in;j++)
fin>>this->v_pX[i][j];
}
fin.close();
for(i=0;im_out;i++)
for(j=0;jm_mid2;j++)
this->v_weight[i][j]=1;
//////////////////////////////////////////
//训练网络，更新并记录权值
//////////////////////////////////////////
this->CalcNumberperC();
this->ChooseandSortconS(); //选取径向基函数中心矢量
this->CalcMidZ1(); //选择变换函数，计算隐层输出
this->CalcMidZ2();
if(strcmp(this->m_learningalgorithm,"RLSA")==0)
RLSA();
//其他学习算法
//else if
//else if
//...
//////////////////////////////////////////
//释放空间，仅保留权值、中心控制矢量和各个变量
//////////////////////////////////////////
for(j=0;jm_pN;j++)
delete[] this->v_pX[j],this->v_expectD[j],this->v_midZ1[j],this->v_midZ2[j];
delete[] this->v_pX,this->v_expectD,this->v_midZ1,this->v_midZ2;
delete[] this->v_pClass,this->v_pSort;
}
}
/*计算每个类别对应的模式个数*/
void CRBPNN::CalcNumberperC()
{
int c,i,j;
j=0;
for(c=1;c<=this->m_pC;c++)
{
this->v_pNC[c-1]=0;
for(i=0;im_pN;i++)
{
if(this->v_pClass[i]==c)
{
this->v_pSort[j]=i;
j++;
this->v_pNC[c-1]++;
}
}
}
}
/*设置中心控制矢量，并将其按类别排好*/
void CRBPNN::ChooseandSortconS()
{
int i,j;
for(i=0;im_mid1;i++)
for(j=0;jm_in;j++)
this->v_conS[i][j]=this->v_pX[this->v_pSort[i]][j];
}
/*计算第一隐层输出*/
void CRBPNN::CalcMidZ1()
{
if(strcmp(this->m_kernel,"Gaussian")==0)
{
double* v_alpha; //高斯形状控制参数
v_alpha=new double[this->m_mid1];
for(int m=0;mm_mid1;m++)
v_alpha[m]=0.3; //将v_alpha[i]设为固定值0.5
/*
//将v_alpha[i]设为0、1之间的随机值
int m;
srand((unsigned)time(NULL));
for(m=0;mm_pN;t++)
{
for(i=0;im_mid1;i++)
{
sum=0;
for(j=0;jm_in;j++)
sum+=pow(this->v_pX[t][j]-this->v_conS[i][j],2);
sum=sum*1.0/(2.0*(pow(v_alpha[i],2)));
this->v_midZ1[t][i]=exp((-1)*sum);
}
}
delete[] v_alpha;
}
//采用其他核计算第一层隐层输出
//if
//if
//if
}
/*计算第二隐层输出*/
void CRBPNN::CalcMidZ2()
{
int i,j,k,t;
double sum;
for(i=0;i>m_lamda;
m_lamda=0.99;
//cout<<"请输入终止误差:"<>m_enderror;
m_enderror=0.01;
//cout<<"请输入终止次数:"<>m_endcyc;
m_endcyc=1000;
double *v_gain; //卡尔曼增益
v_gain=new double[m_mid2];
double **v_icorrP; //逆相关矩阵
v_icorrP=new double*[m_mid2];
for(i=0;im_mid2;i++)
{
v_icorrP[i]=new double[m_mid2];
for(j=0;jm_mid2;j++)
if(j==i) v_icorrP[i][j]=1;
else v_icorrP[i][j]=0;
}
//迭代算法，计算权值和累计误差能量
do
{
pk=int(k%this->m_pN); //选择输入模式
this->CalcGain(pk,m_lamda,v_gain,v_icorrP); //计算卡尔曼增益
this->CalcIcorrP(pk,m_lamda,v_gain,v_icorrP); //计算逆相关矩阵
m_energyJ=this->CalcEnergyJ(pk,m_lamda,m_energyJ); //计算累计误差能量
this->CalcWeight(pk,v_gain); //更新连接权值
k++;
fout<m_enderror&&km_out;j++)
{
sum=0;
for(k=0;km_mid2;k++)
sum+=this->v_midZ2[t][k]*this->v_weight[j][k];
sum=this->v_expectD[t][j]-sum;
for(i=0;iv_weight[j][i]=this->v_weight[j][i]+gain[i]*sum;
}
}
/*计算卡尔曼增益
t:迭代的次序
lamda:加权遗忘因子
gain:卡尔曼增益(输出)
icorrP:逆相关矩阵*/
void CRBPNN::CalcGain(int t,double lamda,double* gain,double **icorrP)
{
int i,j;
double g=0.0;
for(i=0;im_mid2;i++)
{
gain[i]=0.0;
for(j=0;jm_mid2;j++)
gain[i]+=icorrP[i][j]*this->v_midZ2[t][j];
}
for(j=0;jm_mid2;j++)
g+=this->v_midZ2[t][j]*gain[j];
g+=lamda;
for(i=0;im_mid2;i++)
gain[i]/=g;
}
/*迭代计算逆相关矩阵
t:迭代的次序
lamda:加权遗忘因子
gain:卡尔曼增益
icorrP:逆相关矩阵(输出)*/
void CRBPNN::CalcIcorrP(int t,double lamda,double *gain,double **icorrP)
{
int i,j,k;
double** temp1;
double** temp2;
temp1=new double*[this->m_mid2];
temp2=new double*[this->m_mid2];
for(i=0;im_mid2;i++)
{
temp1[i]=new double[this->m_mid2];
for(j=0;jm_mid2;j++)
temp1[i][j]=gain[i]*this->v_midZ2[t][j];
}
for(i=0;im_mid2;i++)
{
temp2[i]=new double[this->m_mid2];
for(j=0;jm_mid2;j++)
{
temp2[i][j]=0;
for(k=0;km_mid2;k++)
temp2[i][j]+=temp1[i][k]*icorrP[k][j];
}
}
for(i=0;im_mid2;i++)
{
for(j=0;jm_mid2;j++)
icorrP[i][j]=(icorrP[i][j]-temp2[i][j])/lamda;
}
for(i=0;im_mid2;i++)
delete[] temp1[i],temp2[i];
delete[] temp1,temp2;
}
/*迭代计算累积误差能量函数
t:迭代的次序
lamda:加权遗忘因子
energyJ:累积误差能量*/
double CRBPNN::CalcEnergyJ(int t,double lamda,double energyJ)
{
int k,j;
double sum1,sum2;
sum1=0;
for(j=0;jm_out;j++)
{
sum2=0;
for(k=0;km_mid2;k++)
sum2+=this->v_midZ2[t][k]*this->v_weight[j][k];
sum2=this->v_expectD[t][j]-sum2;
sum1+=pow(sum2,2);
}
return lamda*energyJ+sum1/2.0;
}
/*模式分类
n:需要分类的模式个数
testdata:分类数据*/
void CRBPNN::Classifying(int n,char* testdata)
{
if(n>0&&testdata!=NULL&&this->v_weight!=NULL)
{
int i,j;
this->m_pN=n;
this->v_pX=new double*[this->m_pN];
this->v_midZ1=new double*[this->m_pN];
this->v_midZ2=new double*[this->m_pN];
for(j=0;jm_pN;j++)
{
this->v_pX[j]=new double[this->m_in];
this->v_midZ1[j]=new double[this->m_mid1];
this->v_midZ2[j]=new double[this->m_mid1];
}
ifstream fin(testdata,ios::nocreate);
if(fin.fail())
{
cerr<<"Error opening the file:"<m_in;j++)
fin>>this->v_pX[i][j];
i++;
}while(!fin.eof());
fin.close();
this->CalcMidZ1();
this->CalcMidZ2();
this->CalcOutput();
for(j=0;jm_pN;j++)
delete[] this->v_pX[j],this->v_midZ1[j],this->v_midZ2[j];
delete[] this->v_pX,this->v_midZ1,this->v_midZ2;
}
}
/*计算网络输出*/
void CRBPNN::CalcOutput()
{
int i,j,t;
this->v_realY=new double*[this->m_pN];
for(i=0;im_pN;i++)
this->v_realY[i]=new double[this->m_out];
for(t=0;tm_pN;t++)
{
for(j=0;jm_out;j++)
{
v_realY[t][j]=0.0;
for(i=0;iv_realY[t][j]+=this->v_midZ2[t][i]*this->v_weight[j][i];
}
}
}
/*网络输出
pFileName:用于保存输出结果的文件名称*/
void CRBPNN::GetOutput(char *pFileName)
{
ofstream fout(pFileName);
if(fout.fail())
{
cerr<<"Error opening file:"<m_pN;i++)
{
int t=0;
double y=this->v_realY[i][0];
for(j=0;jm_out;j++)
{
if(yv_realY[i][j]) {y=this->v_realY[i][j];t=j;}
fout<m_pN-1) fout<v_weight!=NULL)
{
ofstream fout(pFileName);
if(fout.fail())
{
cerr<<"Error opening file:"<m_out;i++)
{
for(j=0;jm_mid2;j++)
{
fout<v_weight[i][j];
}
fout<v_weight==NULL)
{
ofstream fout(pFileName);
if(fout.fail())
{
cerr<<"Error opening file:"<m_out;i++)
{
for(j=0;jm_mid2;j++)
{
fout.precision(5);
fout<v_weight[i][j];
}
fout<v_realY!=NULL)
{
ifstream fin(ptestclass);
if(fin.fail())
{
cerr<<"Error opening file:"<m_pN;i++)
{
fin>>d;
y=this->v_realY[i][0];
t=0;
for(j=1;jm_out;j++)
{
if(yv_realY[i][j]) {y=this->v_realY[i][j];t=j;}
}
if(d==t+1) count++;
}
return count*100.0/m_pN;
}
return -1;
}