void C_piiA(int* m_ptr, int* p_ptr,
double* covariates,
int* distancesincluded,
double* transpar,
int* v,
double* pii,
double* A)
{
unsigned int p = *p_ptr, m = *m_ptr;
// unsigned int q = 1+2+p;
double p0, p1;
p0 = transpar(0,0);
p1 = transpar(1,0);
for (unsigned int j=0; j<p; j++){
p0 += transpar(0,3+j) * covariates(0,j);
p1 += transpar(1,3+j) * covariates(0,j);
}
pii[0] = exp(p0) / (exp(p0)+exp(p1));
pii[1] = 1 - pii[0];
double num11 = exp(transpar(0,1));
double num21 = exp(transpar(0,2));
for (unsigned int k=0; k<m-1; k++){
double denom12 = 0.;
for (unsigned int j=0; j<p; j++){
denom12 += transpar(1,3+j) * covariates(k+1,j);
}
double denom22 = denom12;
if (*distancesincluded){
denom22 -= transpar(1,3) * covariates(k+1,0); // undo plus
denom22 -= transpar(1,3) * covariates(k+1,0); // and minus
}
double num12 = exp(transpar(1,1) + denom12);
double num22 = exp(transpar(1,2) + denom22);
A(0,0,k) = num11/(num11 + num12);
A(0,1,k) = num12/(num11 + num12);
A(1,0,k) = num21/(num21 + num22);
A(1,1,k) = num22/(num21 + num22);
}
}
void C_ComputeGradient(int* m_ptr, int* p_ptr,
double* covariates,
int* distancesincluded,
double* dgammA,
double* gammA,
double* transpar,
int* v,
double* pii,
double* A,
double* gradient)
{
unsigned int p = *p_ptr, m = *m_ptr;
C_piiA(m_ptr, p_ptr, covariates, distancesincluded, transpar, v, pii, A);
gradient[0] = gammA(0,1) - pii[1];
for (unsigned int k=0; k<m-1; k++){
gradient[1] += dgammA(0,1,k) - gammA(k,0)*A(0,1,k);
gradient[2] += dgammA(1,1,k) - gammA(k,1)*A(1,1,k);
}
double* tmp = new double[m-1];
for (unsigned int k=0; k<m-1; k++){
tmp[k] = gammA(k+1,1) - (gammA(k,0)*A(0,1,k) + gammA(k,1)*A(1,1,k));
}
unsigned int first = 0;
if (*distancesincluded){
first = 1; // gradient[3] not concerned
}
for (unsigned int j=first; j<p; j++){
gradient[3+j] =(gammA(0,1) - pii[1])*covariates(0,j);
for (unsigned int k=0; k<m-1; k++){
gradient[3+j] += tmp[k]*covariates(k+1,j);
}
}
delete[] tmp;
if (*distancesincluded){
gradient[3] =(gammA(0,1) - pii[1])*covariates(0,0);
for (unsigned int k=0; k<m-1; k++){
gradient[3] += dgammA(0,1,k) * covariates(k+1,0)
- dgammA(1,1,k) * covariates(k+1,0)
- gammA(k,0)*A(0,1,k)*covariates(k+1,0)
+ gammA(k,1)*A(1,1,k)*covariates(k+1,0);
}
}
for (unsigned int k=0; k < 3+p; k++)
{
gradient[k] = -gradient[k];
}
}
void Kernel::createKernelFromDiscreteCovariates(std::string fn, int useColumn)
{
Covariate covariates(fn, "", std::vector<std::string>(), false);
std::vector<int> categories;
this->randomVarNames.clear();
if( communicator->mpiRoot == false)
{
if( covariates.rawCovars.size() < 1 )
{
misc.error("Error: File [ " + fn + " ] is empty.", 0);
}
if( useColumn >= covariates.rawCovars[0].size() )
{
misc.error("Error: Unable to get the column " + i2s(useColumn) + " from file [ " + fn + " ].", 0);
}
for(int i = 0; i < covariates.rawCovars.size(); i++)
{
std::string value = covariates.rawCovars[ i ][ useColumn ];
int idx = covariates.covarCategories[ useColumn ][ value ];
categories.push_back(idx);
}
for(std::map<std::string, int>::iterator it = covariates.covarCategories[ useColumn ].begin(); it != covariates.covarCategories[ useColumn ].end(); ++it)
{
this->randomVarNames.push_back(it->first);
}
}
if(this->kernel != NULL)
{
delete this->kernel;
}
if(this->N != NULL)
{
delete this->N;
}
this->kernel = new Matrix();
this->kernel->makeIntersectionMatrix(categories);
}
void C_LineSearch(int* m_ptr, int* p_ptr,
double* covariates,
int* distancesincluded,
double* dgammA,
double* gammA,
double* transpar,
double* phi,
int* iterCG,
double* ptol,
int* v,
int* allright,
double* nu,
double* pii,
double* A)
{
unsigned int p = *p_ptr, m = *m_ptr;
int i, j;
unsigned int n, k;
int niter = 0;
double difference = 1.0;
double dQ, dQ2, tmp, dQ2_tmp;
double* transparnew = new double[2*(3+p)];
for (j=0; j < 2; j++)
{
for (k=0; k < 3+p; k++)
{
transparnew(j,k) = transpar(j,k);
}
}
while (difference > *ptol && niter < *iterCG)
{
niter++;
C_piiA( m_ptr, p_ptr, covariates, distancesincluded, transparnew, v, pii, A);
// dQ
dQ = 0.0;
for (j=0; j < 2; j++)
{
tmp = 0.0;
for (k=3; k < p+3; k++)
{
if (*distancesincluded && k == 3)
tmp += 0;
else
tmp += phi(j,k) * covariates(0,k-3);
}
dQ += (phi(j,0) + tmp) * (gammA(0,j) - pii[j]);
}
//Rprintf("tmp : %f\n", dQ);
for (i=0; i < 2; i++)
{
for (j=0; j < 2; j++)
{
for (n = 0; n < m-1; n++)
{
tmp = 0.0;
for (k=3; k < p+3; k++)
{
if (*distancesincluded && k == 3 && i == 1)
tmp += (-phi(j,k)) * covariates(n+1, k-3);
else
tmp += phi(j,k) * covariates(n+1, k-3);
}
dQ += (phi(j,i+1) + tmp) * (dgammA(i,j,n) - (gammA(n,i) * A(i,j,n)));
}
}
}
//Rprintf("dQ : %f\n", dQ);
// dQ2
dQ2 = 0.0;
for (j=0; j < 2; j++)
{
tmp = 0.0;
for (k=3; k < p+3; k++)
{
if (*distancesincluded && k == 3)
tmp += 0;
else
tmp += phi(j,k) * covariates(0,k-3);
}
dQ2 += (phi(j,0) + tmp) * (phi(j,0) + tmp) * pii[j] * (1 - pii[j]);
}
dQ2 = -dQ2;
//Rprintf("dQ2_tmp : %f\n", dQ2);
for (i=0; i < 2; i++)
{
for (j=0; j < 2; j++)
{
dQ2_tmp = 0.0;
for (n = 0; n < m-1; n++)
{
tmp = 0.0;
for (k=3; k < p+3; k++)
{
if (*distancesincluded && k == 3 && i == 1)
tmp += (-phi(j,k)) * covariates(n+1, k-3);
else
tmp += phi(j,k) * covariates(n+1, k-3);
}
dQ2_tmp += ((phi(j,i+1) + tmp) * (phi(j,i+1) + tmp)) * gammA(n,i) * A(i,j,n) * (1 - A(i,j,n));
}
dQ2 -= dQ2_tmp;
}
}
//Rprintf("dQ2 : %f\n", dQ2);
if (dQ2 == 0)
{
difference = *ptol;
niter = *iterCG;
*nu = 0;
for (j=0; j < 2; j++)
{
for (k=0; k < 3+p; k++)
{
transparnew(j,k) = transpar(j,k);
}
}
*allright = false;
}
else
{
*nu = *nu - dQ / dQ2;
difference = fabs(dQ/ dQ2);
for (j=0; j < 2; j++)
{
for (k=0; k < 3+p; k++)
{
transparnew(j,k) = transpar(j,k) + *nu * phi(j,k);
}
}
}
}
for (j=0; j < 2; j++)
{
for (k=0; k < 3+p; k++)
{
transpar(j,k) = transparnew(j,k);
}
}
delete[] transparnew;
}
