Real
ex_update_models(fitinfo fit[], int num_models,
int weighted, int approximate_roughness,
int forced)
{
int n = 0;
Real sumsq = 0.;
int i;
// printf("calling update\n");
MODELS = num_models;
fit[0].penalty = 0.;
if (*constraints) (*constraints)(fit);
sumsq = fit[0].penalty;
if (!forced && sumsq >= FIT_REJECT_PENALTY) return sumsq;
for (i=0; i < num_models; i++) {
int n_i = 0;
Real sumsq_i = 0.;
fit_update(&fit[i], approximate_roughness);
if (weighted) fit_wsumsq(&fit[i],&n_i,&sumsq_i);
else fit_sumsq(&fit[i],&n_i,&sumsq_i);
fit[i].chisq_est = sumsq_i/n_i;
n += n_i;
sumsq += sumsq_i;
}
// printf("sumsq=%10g, n=%4d, pars=%d\n",sumsq,n,fit[0].pars.n);
return n < fit[0].pars.n ? sumsq : sumsq / (n - fit[0].pars.n) ;
}
void
fit_partial(fitinfo *fit, int approx, Real portion, Real best,
int weighted, int *totaldf, Real *totalsumsq)
{
int i, samples;
Real *work, *Q, *A, *B, *C, *D;
int nQ, worksize;
Real sumsq = 0.;
int df = 0;
/* FIXME kill partial for now --- we need to refactor the
* the other bits so they allow single point computations a little
* more cleanly.
*/
portion = 1.;
/* If all, update the model and calculate sumsq using all points. */
if (portion >= 1.) {
fit_update(fit, approx);
if (weighted) fit_wsumsq(fit,&df,&sumsq);
else fit_sumsq(fit,&df,&sumsq);
fit->chisq_est = sumsq/df;
return;
}
/* Generate the profile from the model */
model_profile(&fit->m, &fit->p);
/* Find the Q points at which we need to calculate the model */
find_target_Q(fit);
/* Clear work vector so resolution isn't calculated */
worksize = fit->worksize; fit->worksize = 0;
work = fit->work; fit->work = NULL;
/* Stash Q,R info */
nQ = fit->nQ;
Q = fit->fitQ;
A = fit->fitA;
B = fit->fitB;
C = fit->fitC;
D = fit->fitD;
/* For the rest of this function, assume a single point Q vector */
fit->nQ = 1;
/* Clear results vector */
for (i=0; i < nQ; i++) A[i] = NOVALUE;
/* First pass --- calculate a few randomly selected Q values */
/* We wat to take the minimum number of samples which will allow us to */
/* reject the hypothesis that the current individual comes from a */
/* chisq distribution as good or better than the one which gave rise */
/* to the best chisq. Any individual so bad is unlikely to propagate */
/* its genes to the next generation. For individuals as good as or */
/* better than the best, we want to keep sampling up to some portion. */
/* FIXME use a proper statistical test? */
for (i=0; i < 5; i++) {
int k= (int)floor(frandom()*nQ);
partial_point(fit,k,Q,A,B,C,D,weighted,&df,&sumsq);
}
samples = (int)ceil(portion*nQ);
while (i++ < samples) {
int k= (int)floor(frandom()*nQ);
partial_point(fit,k,Q,A,B,C,D,weighted,&df,&sumsq);
if (sumsq/df > 10.*best) break;
}
/* Accumulate the sumsq for this fit. */
/* printf("sample %d of %d with %g vs. %g\n",df,samples,sumsq/df,best); */
*totalsumsq += sumsq;
*totaldf += df;
fit->chisq_est = sumsq/df;
/* Restore work vector */
fit->worksize = worksize;
fit->work = work;
/* Restore Q,R info */
fit->nQ = nQ;
fit->fitQ = Q;
fit->fitA = A;
fit->fitB = B;
fit->fitC = C;
fit->fitD = D;
}
