/**
* emulate the model at the ith entry in results->new_x
*/
void emulate_ith_location(modelstruct *the_model, optstruct *options, resultstruct *results,int i, gsl_matrix* h_matrix, gsl_matrix* cinverse, gsl_vector *beta_vector){
double kappa;
double temp_mean, temp_var;
gsl_vector_view new_x_row;
gsl_vector *kplus = gsl_vector_alloc(options->nmodel_points);
gsl_vector *h_vector = gsl_vector_alloc(options->nregression_fns);
// read the new x location
new_x_row = gsl_matrix_row(results->new_x, i);
//fprintf(stderr, "i(%d) new_x_row: ", i);
//print_vector_quiet(&new_x_row.vector, options->nparams);
makeKVector(kplus, the_model->xmodel, &new_x_row.vector, the_model->thetas, options->nmodel_points, options->nthetas, options->nparams);
makeHVector(h_vector, &new_x_row.vector, options->nparams);
temp_mean = makeEmulatedMean(cinverse, the_model->training_vector, kplus, h_vector, h_matrix, beta_vector, options->nmodel_points);
kappa = covariance_fn(&new_x_row.vector, &new_x_row.vector, the_model->thetas, options->nthetas, options->nparams);
temp_var = makeEmulatedVariance(cinverse, kplus, h_vector, h_matrix, kappa, options->nmodel_points, options->nregression_fns);
//fprintf(stderr, "temp_mean %lf\ttemp_var %lf\n", temp_mean, temp_var);
gsl_vector_set(results->emulated_mean, i, temp_mean);
gsl_vector_set(results->emulated_var, i, temp_var);
gsl_vector_free(kplus);
gsl_vector_free(h_vector);
}
void emulateQuick( modelstruct *the_model, gsl_vector* the_point, optstruct* options,
double* mean_out, double* var_out, gsl_matrix *h_matrix, gsl_matrix *cinverse, gsl_vector* beta_vector ){
double kappa;
double temp_mean, temp_var;
gsl_vector_view new_x_row;
gsl_vector *kplus = gsl_vector_alloc(options->nmodel_points);
gsl_vector *h_vector = gsl_vector_alloc(options->nregression_fns);
makeKVector(kplus, the_model->xmodel, the_point, the_model->thetas, options->nmodel_points, options->nthetas, options->nparams);
makeHVector(h_vector, the_point, options->nparams);
temp_mean = makeEmulatedMean(cinverse, the_model->training_vector, kplus, h_vector, h_matrix, beta_vector, options->nmodel_points);
kappa = covariance_fn(the_point, the_point, the_model->thetas, options->nthetas, options->nparams);
temp_var = makeEmulatedVariance(cinverse, kplus, h_vector, h_matrix, kappa, options->nmodel_points, options->nregression_fns);
//fprintf(stderr, "temp_mean %lf\ttemp_var %lf\n", temp_mean, temp_var);
// save the results
*mean_out = temp_mean;
*var_out = temp_var;
gsl_vector_free(kplus);
gsl_vector_free(h_vector);
}
/**
* emulate the model at the location of the_point
* @return *the_variance is set to the variance at this location
* @return *the_mean is set to the emulated mean at this loc
*/
void emulateAtPoint(modelstruct *the_model, gsl_vector* the_point, optstruct* options, double* the_mean, double* the_variance){
int i;
double determinant_c = 0.0;
gsl_matrix *c_matrix = gsl_matrix_alloc(options->nmodel_points, options->nmodel_points);
gsl_matrix *cinverse = gsl_matrix_alloc(options->nmodel_points, options->nmodel_points);
gsl_vector *beta_vector = gsl_vector_alloc(options->nregression_fns);
gsl_matrix *h_matrix = gsl_matrix_alloc(options->nmodel_points, options->nregression_fns);
gsl_matrix *temp_matrix = gsl_matrix_alloc(options->nmodel_points, options->nmodel_points);
double kappa;
double temp_mean, temp_var;
gsl_vector_view new_x_row;
gsl_vector *kplus = gsl_vector_alloc(options->nmodel_points);
gsl_vector *h_vector = gsl_vector_alloc(options->nregression_fns);
// create the covariance matrix and save a copy in temp_matrix
makeCovMatrix(c_matrix, the_model->xmodel, the_model->thetas,options->nmodel_points, options->nthetas, options->nparams);
gsl_matrix_memcpy(temp_matrix, c_matrix);
chol_inverse_cov_matrix(options, temp_matrix, cinverse, &determinant_c);
// regression cpts
makeHMatrix(h_matrix, the_model->xmodel, options->nmodel_points, options->nparams, options->nregression_fns);
estimateBeta(beta_vector, h_matrix, cinverse, the_model->training_vector, options->nmodel_points, options->nregression_fns);
makeKVector(kplus, the_model->xmodel, the_point, the_model->thetas, options->nmodel_points, options->nthetas, options->nparams);
makeHVector(h_vector, the_point, options->nparams);
temp_mean = makeEmulatedMean(cinverse, the_model->training_vector, kplus, h_vector, h_matrix, beta_vector, options->nmodel_points);
kappa = covariance_fn(the_point, the_point, the_model->thetas, options->nthetas, options->nparams);
temp_var = makeEmulatedVariance(cinverse, kplus, h_vector, h_matrix, kappa, options->nmodel_points, options->nregression_fns);
if(temp_mean > 100){
printf("EF:mean:%lf\tvar:%lf\n", temp_mean, temp_var);
print_vector_quiet(the_point,options->nparams);
printf("kplus:\n");
print_vector_quiet(kplus, options->nmodel_points);
printf("hvector:\n");
print_vector_quiet(h_vector, options->nregression_fns);
printf("beta:\n");
print_vector_quiet(beta_vector, options->nregression_fns);
//exit(1);
}
*the_mean = temp_mean;
*the_variance = temp_var;
gsl_vector_free(kplus);
gsl_vector_free(h_vector);
gsl_matrix_free(c_matrix);
gsl_matrix_free(cinverse);
gsl_matrix_free(h_matrix);
gsl_matrix_free(temp_matrix);
gsl_vector_free(beta_vector);
}
double getLogLikelyhood(gsl_matrix *cinverse, double det_cmatrix, gsl_matrix *xmodel,
gsl_vector *trainingvector, gsl_vector *thetas, gsl_matrix *h_matrix, int nmodel_points,
int nthetas, int nparams, int nregression_fns,
void(*makeHVector)(gsl_vector *, gsl_vector*, int))
{
(void) thetas;
(void) nthetas;
int i;
double the_likelyhood = 0.0;
double vector_matrix_vector_product = 0.0;
double log_2_pi = 1.83788;
gsl_vector_view xmodel_row;
gsl_vector *result_holder = gsl_vector_alloc(nmodel_points);
gsl_vector *h_vector = gsl_vector_alloc(nregression_fns);
gsl_vector *beta_vector = gsl_vector_alloc(nregression_fns);
gsl_vector *estimated_mean = gsl_vector_alloc(nmodel_points);
gsl_vector *train_sub_mean = gsl_vector_alloc(nmodel_points);
double estimated_mean_val = 0.0;
double log_det_c = log(det_cmatrix);
//log_det_c = fabs(log_det_c);
/* we need to calculate the mean vector for this set of thetas
* estMean[i] = hvector(training[i]).betavector
*/
estimateBeta(beta_vector, h_matrix, cinverse, trainingvector, nmodel_points, nregression_fns);
for(i = 0; i < nmodel_points; i++){
xmodel_row = gsl_matrix_row(xmodel, i);
// can't call this...
makeHVector(h_vector, &xmodel_row.vector, nparams);
//print_vector_quiet(h_vector, nregression_fns);
gsl_blas_ddot(beta_vector, h_vector, &estimated_mean_val);
gsl_vector_set(estimated_mean, i, estimated_mean_val);
}
/* train_sub_mean = trainingvector- estimated_mean */
gsl_vector_memcpy(train_sub_mean, trainingvector);
gsl_vector_sub(train_sub_mean, estimated_mean);
//print_vector_quiet(beta_vector, nregression_fns); <- verified
//print_vector_quiet(trainingvector, nmodel_points); <- verified
//print_vector_quiet(estimated_mean, nmodel_points);
/* print_vector_quiet(train_sub_mean, nmodel_points); */
// DEBUG
// the log likelyhood is a given by
// L = (-1/2)*Log[Det[cinverse]] - (1/2)*trainingvector.cinverse.trainingvector - (nmodel_points/2)*Log[2*Pi]
the_likelyhood = -(1.0/2.0)*log_det_c - (nmodel_points/2.0)*log_2_pi;
gsl_blas_dgemv(CblasNoTrans, 1.0, cinverse, train_sub_mean, 0.0, result_holder);
gsl_blas_ddot(train_sub_mean, result_holder, &vector_matrix_vector_product);
/* printf("%g\n", (log_2_pi)*(nmodel_points/2.0)); */
/* printf("parta:(-1/2)*log_det_c = %g\n", (-0.5)*log_det_c); */
/* printf("partb:(training-mean).cinverse.(training-mean) = %g\n", (-0.5)*vector_matrix_vector_product); */
/* printf("partc:(-nmodel_points/2.0)*log_2_pi = %g\n", -(nmodel_points/2.0)*log_2_pi); */
the_likelyhood += vector_matrix_vector_product*(-1.0/2.0);
gsl_vector_free(result_holder);
gsl_vector_free(h_vector);
gsl_vector_free(beta_vector);
gsl_vector_free(estimated_mean);
gsl_vector_free(train_sub_mean);
return(the_likelyhood);
}
