/*** do cyclic coordinate descent optimization for fixed lambda1
* repeat coordinate descent algorithm until solution is converged ***/
bool
cdescent_do_update_one_cycle (cdescent *cd)
{
int ccd_iter = 0;
bool converged = false;
update_one_cycle update_func;
update_func = (cd->rule == CDESCENT_SELECTION_RULE_STOCHASTIC) ?
cdescent_do_update_once_cycle_stochastic : cdescent_do_update_once_cycle_cyclic;
if (!cd) error_and_exit ("cdescent_do_cyclic_update", "cdescent *cd is empty.", __FILE__, __LINE__);
while (!converged) {
converged = update_func (cd);
if (++ccd_iter >= cd->maxiter) {
printf_warning ("cdescent_do_cyclic_update", "reaching max number of iterations.", __FILE__, __LINE__);
break;
}
}
cd->total_iter += ccd_iter;
return converged;
}
/*** do reweighted coordinate descent optimization.
* The weight for L1 norm is updated by calling reweighting function
* cd->path->func using beta of the previous iteration. ***/
bool
cdescent_do_reweighting (cdescent *cd)
{
int iter;
double dnrm1;
bool converged = true;
if (!cd->rwt) error_and_exit ("cdescent_do_reweighting", "cd->rwt is empty.", __FILE__, __LINE__);
if (!cd->rwt->func) error_and_exit ("cdescent_do_reweighting", "cd->rwt->func is empty.", __FILE__, __LINE__);
iter = 0;
do {
double nrm1_prev = cd->nrm1;
mm_dense *w = cd->rwt->func->function (cd, cd->rwt->func->data);
cdescent_set_penalty_factor (cd, w, cd->rwt->func->tau);
mm_real_free (w);
if (!(converged = cdescent_do_update_one_cycle (cd))) break;
if (++iter > cd->rwt->maxiter) {
printf_warning ("cdescent_do_reweighting", "reaching max number of iterations.", __FILE__, __LINE__);
break;
}
dnrm1 = fabs (cd->nrm1 - nrm1_prev);
} while (dnrm1 > cd->rwt->tolerance);
return converged;
}
void team_list_init(void)
{
char *name;
if (g_server.team_list != NULL)
return ;
g_server.team_list = list_create();
name = strtok(g_server.param.team_names, "\n");
while (name)
{
team_create(++g_server.param.team_max_id, name);
name = strtok(NULL, "\n");
}
free(g_server.param.team_names);
if (g_server.team_list->size < 1)
{
printf_warning("No team names specified, defaulting to \"%s\"",
DEFAULT_N);
team_create(++g_server.param.team_max_id, DEFAULT_N);
}
}
/*** do pathwise cyclic coordinate descent optimization.
* The regression is starting at the smallest value lambda1_max for which
* the entire vector beta = 0, and decreasing sequence of values for lambda1
* while log10(lambda1) >= path->log10_lambda1_lower.
* log10(lambda1_max) is identical with log10 ( max ( abs(X' * y) ) ), where this
* value is stored in cd->lreg->log10camax.
* The interval of decreasing sequence on the log10 scale is path->dlog10_lambda1. ***/
bool
cdescent_do_pathwise_optimization (cdescent *cd)
{
int iter = 0;
double logt;
bool stop_flag = false;
bool converged;
FILE *fp_path = NULL;
FILE *fp_bic = NULL;
if (!cd) error_and_exit ("cdescent_do_pathwise_optimization", "cdescent *cd is empty.", __FILE__, __LINE__);
if (!cd->path) error_and_exit ("cdescent_do_pathwise_optimization", "cd->path is empty.", __FILE__, __LINE__);
// reset cdescent object if need
if (cd->was_modified) cdescent_reset (cd);
// reset pathwise object if need
if (cd->path->was_modified) pathwise_reset (cd->path);
/* warm start */
stop_flag = set_logt (cd->path->log10_lambda_lower, cd->path->log10_lambda_upper, &logt);
if (cd->path->output_fullpath) {
if (!(fp_path = fopen (cd->path->fn_path, "w"))) {
char msg[80];
sprintf (msg, "cannot open file %s.", cd->path->fn_path);
printf_warning ("cdescent_do_pathwise_optimization", msg, __FILE__, __LINE__);
}
if (fp_path) fprintf_solutionpath (fp_path, iter, cd->beta);
}
if (cd->path->output_bic_info) {
if (!(fp_bic = fopen (cd->path->fn_bic, "w"))) {
char msg[80];
sprintf (msg, "cannot open file %s.", cd->path->fn_bic);
printf_warning ("cdescent_do_pathwise_optimization", msg, __FILE__, __LINE__);
}
// output BIC info headers
if (fp_bic) fprintf (fp_bic, "# nrm1\t\tBIC\t\tRSS\t\tdf\t\tnrm2\t\tlambda1\tlambda2\n");
}
if (cd->path->verbos) fprintf (stderr, "starting pathwise optimization.\n");
while (1) {
bic_info *info;
iter++;
cdescent_set_log10_lambda (cd, logt);
if (cd->path->verbos) fprintf (stderr, "%d-th iteration lambda1 = %.4e, lamba2 = %.4e ", iter, cd->lambda1, cd->lambda2);
if (!(converged = cdescent_do_update_one_cycle (cd))) break;
// reweighting
if (cd->rwt && !(converged = cdescent_do_reweighting (cd))) break;
// output solution path
if (fp_path) fprintf_solutionpath (fp_path, iter, cd->beta);
info = cdescent_eval_bic (cd);
// if bic_val < min_bic_val, update min_bic_val, lambda1_opt, nrm1_opt and beta_opt
if (info->bic_val < cd->path->min_bic_val) {
store_optimal (cd, iter);
cd->path->min_bic_val = info->bic_val;
if (!cd->path->was_modified) cd->path->was_modified = true;
}
if (cd->path->verbos) fprintf (stderr, "bic = %.4e ... ", info->bic_val);
// output BIC info
if (fp_bic) {
// |beta| BIC RSS df ||beta||^2 lambda2
fprintf (fp_bic, "%.8e\t%.8e\t%.8e\t%.8e\t%.8e\t%.8e\t%.8e\n", cd->nrm1, info->bic_val, info->rss, info->df, mm_real_xj_ssq (cd->beta, 0), cd->lambda1, cd->lambda2);
fflush (fp_bic);
}
free (info);
if (stop_flag) break;
/* if logt - dlog10_lambda1 < log10_lambda1, logt = log10_lambda1 and stop_flag is set to true
* else logt -= dlog10_lambda1 */
stop_flag = set_logt (cd->path->log10_lambda_lower, logt - cd->path->dlog10_lambda, &logt);
if (cd->path->verbos) fprintf (stderr, "done.\n");
}
if (fp_path) fclose (fp_path);
if (fp_bic) fclose (fp_bic);
return converged;
}
