/*! \brief Compute y = a1 exp(-x/|a0|) + (1-a1) exp(-x/|a2|) */
static double lmc_exp_exp(double x, const double *a)
{
double e1, e2;
e1 = safe_exp(-x/fabs(a[0]));
e2 = safe_exp(-x/(fabs(a[0])+fabs(a[2])));
return a[1]*e1 + (1-a[1])*e2;
}
/*! \brief Compute y = a0 exp(-x/|a1|) + a2 exp(-x/(|a1|+|a3|)) + a4 */
static double lmc_exp_5_parm(double x, const double *a)
{
double e1, e2;
e1 = safe_exp(-x/fabs(a[1]));
e2 = safe_exp(-x/(fabs(a[1])+fabs(a[3])));
return a[0]*e1 + a[2]*e2 + a[4];
}
/*! \brief Compute 7 parameter exponential function value.
*
* Compute y = a0 exp(-x/|a1|) + a2 exp(-x/(|a1|+|a3|)) +
* a4 exp(-x/(|a1|+|a3|+|a5|)) + a6
*/
static double lmc_exp_7_parm(double x, const double *a)
{
double e1, e2, e3;
double fa1, fa3, fa5;
fa1 = fabs(a[1]);
fa3 = fa1 + fabs(a[3]);
fa5 = fa3 + fabs(a[5]);
e1 = safe_exp(-x/fa1);
e2 = safe_exp(-x/fa3);
e3 = safe_exp(-x/fa5);
return a[0]*e1 + a[2]*e2 + a[4]*e3 + a[6];
}
// recompute data likelihoods using marginals from the combos
// assumes that the genotype combos are in the same order as the likelihoods
// assumes that the genotype combos are the same size as the number of samples in the likelihoods
// returns the delta from the previous marginals, informative in the case of EM
long double marginalGenotypeLikelihoods(list<GenotypeCombo>& genotypeCombos, SampleDataLikelihoods& likelihoods) {
long double delta = 0;
vector< map<Genotype*, long double> > rawMarginals;
rawMarginals.resize(likelihoods.size());
vector< map<Genotype*, long double> >::iterator rawMarginalsItr;
// push the marginal likelihoods into the rawMarginals maps
for (list<GenotypeCombo>::iterator gc = genotypeCombos.begin(); gc != genotypeCombos.end(); ++gc) {
rawMarginalsItr = rawMarginals.begin();
for (GenotypeCombo::const_iterator i = gc->begin(); i != gc->end(); ++i) {
const SampleDataLikelihood& sdl = **i;
map<Genotype*, long double>& rmgs = *rawMarginalsItr++;
map<Genotype*, long double>::iterator rmgsItr = rmgs.find(sdl.genotype);
if (rmgsItr == rmgs.end()) {
rmgs[sdl.genotype] = gc->posteriorProb;
} else {
//vector<long double> x;
//x.push_back(rmgsItr->second); x.push_back(gc->posteriorProb);
//rmgs[sdl.genotype] = logsumexp_probs(x);
rmgs[sdl.genotype] = log(safe_exp(rmgsItr->second) + safe_exp(gc->posteriorProb));
}
}
}
// safely add the raw marginal vectors using logsumexp
// and use to update the sample data likelihoods
rawMarginalsItr = rawMarginals.begin();
for (SampleDataLikelihoods::iterator s = likelihoods.begin(); s != likelihoods.end(); ++s) {
vector<SampleDataLikelihood>& sdls = *s;
const map<Genotype*, long double>& rawmgs = *rawMarginalsItr++;
map<Genotype*, long double> marginals;
vector<long double> rawprobs;
for (map<Genotype*, long double>::const_iterator m = rawmgs.begin(); m != rawmgs.end(); ++m) {
long double p = m->second;
marginals[m->first] = p;
rawprobs.push_back(p);
}
long double normalizer = logsumexp_probs(rawprobs);
for (vector<SampleDataLikelihood>::iterator sdl = sdls.begin(); sdl != sdls.end(); ++sdl) {
long double newmarginal = marginals[sdl->genotype] - normalizer;
delta += newmarginal - sdl->marginal;
sdl->marginal = newmarginal;
}
}
return delta;
}
/*! \brief Compute 9 parameter exponential function value.
*
* Compute y = a0 exp(-x/|a1|) + a2 exp(-x/(|a1|+|a3|)) +
* a4 exp(-x/(|a1|+|a3|+|a5|)) + a6 exp(-x/(|a1|+|a3|+|a5|+|a7|)) + a8
*/
static double lmc_exp_9_parm(double x, const double *a)
{
double e1, e2, e3, e4;
double fa1, fa3, fa5, fa7;
fa1 = fabs(a[1]);
fa3 = fa1 + fabs(a[3]);
fa5 = fa3 + fabs(a[5]);
fa7 = fa5 + fabs(a[7]);
e1 = safe_exp(-x/fa1);
e2 = safe_exp(-x/fa3);
e3 = safe_exp(-x/fa5);
e4 = safe_exp(-x/fa7);
return a[0]*e1 + a[2]*e2 + a[4]*e3 + a[6]*e4 + a[8];
}
/*! \brief Compute vac function */
static double lmc_vac_2_parm(double x, const double *a)
{
/* Fit to function
*
* y = 1/2 (1 - 1/w) exp(-(1+w)v) + 1/2 (1 + 1/w) exp(-(1-w)v)
*
* = exp(-v) (cosh(wv) + 1/w sinh(wv))
*
* v = x/(2 a0)
* w = sqrt(1 - a1)
*
* For tranverse current autocorrelation functions:
* a0 = tau
* a1 = 4 tau (eta/rho) k^2
*
*/
double y, v, det, omega, wv, em, ec, es;
double wv_max = 100;
v = x/(2*fabs(a[0]));
det = 1 - a[1];
em = safe_exp(-v);
if (det != 0)
{
omega = sqrt(fabs(det));
wv = std::min(omega*v, wv_max);
if (det > 0)
{
ec = em*0.5*(safe_exp(wv)+safe_exp(-wv));
es = em*0.5*(safe_exp(wv)-safe_exp(-wv))/omega;
}
else
{
ec = em*cos(wv);
es = em*sin(wv)/omega;
}
y = ec + es;
}
else
{
y = (1+v)*em;
}
return y;
}
/*! \brief Compute y = (1-a0)*exp(-(x/|a2|)^|a3|)*cos(x*|a1|) + a0*exp(-(x/|a4|)^|a5|) */
static double lmc_pres_6_parm(double x, const double *a)
{
double term1, term2, term3;
double pow_max = 10;
term3 = 0;
if ((a[4] != 0) && (a[0] != 0))
{
double power = std::min(fabs(a[5]), pow_max);
term3 = a[0] * safe_exp(-pow((x/fabs(a[4])), power));
}
term1 = 1-a[0];
term2 = 0;
if ((term1 != 0) && (a[2] != 0))
{
double power = std::min(fabs(a[3]), pow_max);
term2 = safe_exp(-pow((x/fabs(a[2])), power)) * cos(x*fabs(a[1]));
}
return term1*term2 + term3;
}
double Logistic::execute ()
{
int icase ;
double *tptr, term, sum1, sum2 ;
sum1 = sum2 = 0.0 ;
for (icase=0 ; icase<ncases ; icase++) {
tptr = tset + (ninputs + 1) * icase ; // This case
predict ( tptr , &term ) ; // Log odds ratio
sum1 += term * tptr[ninputs] ; // Output stored after inputs
sum2 += log ( 1.0 + safe_exp ( term )) ;
} // For all training cases
return sum1 - sum2 ;
}
static double logit_crit ( double *x )
{
int i ;
double x1, y1, x2, y2, x3, y3 ;
for (i=0 ; i<local_logistic->ninputs ; i++)
local_logistic->coefs[i] = safe_exp ( x[i] ) ;
glob_min ( -20.0 , 20.0 , 5 , 0 , -1.e160 , logit_unicrit ,
&x1 , &y1 , &x2 , &y2 , &x3 , &y3 ) ;
y2 = brentmin ( 50 , -1.e160 , 1.e-10 , 1.e-10 , logit_unicrit ,
&x1 , &x2 , &x3 , y2 ) ;
local_logistic->coefs[local_logistic->ninputs] = x2 ;
return -y2 ;
}
/*
多変量正規分布 N(Μ,Σ)の X に対する値を計算する。
単変量の正規分布は上の関数を使うこと。
*/
double
get_multi_norm(double *mean, double **covmat, double *X, int dim)
{
double *tmp1 = NULL, *tmp2 = NULL;
double det, **invmat = NULL;
double val;
tmp1 = del_vec(X, mean, dim);
tmp2 = tmp1;
invmat = matinv(dim, covmat, &det);
tmp1 = vecxmat(tmp1, dim, invmat, dim, dim);
val = dvid(safe_exp(-0.5 * innerprod(dim, tmp1, tmp2)),
(pow(2.0 * M_PI, dim / 2.0) * sqrt(det)));
free_double_matrix(invmat);
free_double_vector(tmp1);
free_double_vector(tmp2);
return val;
}
/*! \brief Compute y = a1 exp(-x/|a0|) */
static double lmc_exp_two_parm(double x, const double *a)
{
return a[1]*safe_exp(-x/fabs(a[0]));
}
/*! \brief Compute y = exp(-x/|a0|) */
static double lmc_exp_one_parm(double x, const double *a)
{
return safe_exp(-x/fabs(a[0]));
}
/*
単変量正規分布 N(μ,σ)の X における値を計算する。
多変量の正規分布には対応していないので注意。
*/
double
get_val_norm(double mean, double sigma, double X)
{
return mult(dvid(1.0, mult(sqrt(2.0 * M_PI), sigma)),
safe_exp(mult((-1.0 / 2.0), pow2(dvid((X - mean), sigma)))));
}
