/** \brief Apply Hessian approximation to vector.
This function applies the Hessian of the Moreau-Yosida penalty to the vector \f$v\f$.
@param[out] hv is the the action of the Hessian on \f$v\f$.
@param[in] v is the direction vector.
@param[in] x is the current iterate.
@param[in] tol is a tolerance for inexact Moreau-Yosida penalty computation.
*/
void hessVec( Vector<Real> &hv, const Vector<Real> &v, const Vector<Real> &x, Real &tol ) {
// Apply objective Hessian to a vector
obj_->hessVec(hv,v,x,tol);
// Add Hessian of the Moreau-Yosida penalty
if ( con_->isActivated() ) {
Real one = 1.0;
computePenalty(x);
v_->set(v);
con_->pruneLowerActive(*v_,*xlam_);
v_->scale(-one);
v_->plus(v);
dv_->set(v_->dual());
dv2_->set(*dv_);
con_->pruneLowerActive(*dv_,*xlam_);
dv_->scale(-one);
dv_->plus(*dv2_);
hv.axpy(mu_,*dv_);
v_->set(v);
con_->pruneUpperActive(*v_,*xlam_);
v_->scale(-one);
v_->plus(v);
dv_->set(v_->dual());
dv2_->set(*dv_);
con_->pruneUpperActive(*dv_,*xlam_);
dv_->scale(-one);
dv_->plus(*dv2_);
hv.axpy(mu_,*dv_);
}
}
/** \brief Compute gradient.
This function returns the Moreau-Yosida penalty gradient.
@param[out] g is the gradient.
@param[in] x is the current iterate.
@param[in] tol is a tolerance for inexact Moreau-Yosida penalty computation.
*/
void gradient( Vector<Real> &g, const Vector<Real> &x, Real &tol ) {
// Compute gradient of objective function
obj_->gradient(*g_,x,tol);
ngval_++;
g.set(*g_);
// Add gradient of the Moreau-Yosida penalty
if ( con_->isActivated() ) {
computePenalty(x);
g.axpy(-mu_,*dl1_);
g.axpy(mu_,*du1_);
}
}
/** \brief Compute value.
This function returns the Moreau-Yosida penalty value.
@param[in] x is the current iterate.
@param[in] tol is a tolerance for inexact Moreau-Yosida penalty computation.
*/
Real value( const Vector<Real> &x, Real &tol ) {
Real half = 0.5;
// Compute objective function value
fval_ = obj_->value(x,tol);
nfval_++;
// Add value of the Moreau-Yosida penalty
Real fval = fval_;
if ( con_->isActivated() ) {
computePenalty(x);
fval += half*mu_*(l1_->dot(*l1_) + u1_->dot(*u1_));
}
return fval;
}
void updateMultipliers(Real mu, const ROL::Vector<Real> &x) {
if ( con_->isActivated() ) {
Real one = 1.0;
computePenalty(x);
lam_->set(*u1_);
lam_->axpy(-one,*l1_);
lam_->scale(mu_);
mu_ = mu;
}
nfval_ = 0; ngval_ = 0;
isConEvaluated_ = false;
}
vector<int> discreteEmbed(const PtGraph &graph, const vector<Sphere> &spheres,
const Skeleton &skeleton, const vector<vector<int> > &possibilities)
{
int i, j;
FP fp(graph, skeleton, spheres);
fp.footBase = 1.;
for(i = 0; i < (int)graph.verts.size(); ++i)
fp.footBase = min(fp.footBase, graph.verts[i][1]);
vector<PenaltyFunction *> penaltyFunctions = getPenaltyFunctions(&fp);
int toMatch = skeleton.cGraph().verts.size();
std::cout << "Matching!" << endl;
priority_queue<PartialMatch> todo;
PartialMatch output(graph.verts.size());
todo.push(output);
int maxSz = 0;
while(!todo.empty()) {
PartialMatch cur = todo.top();
todo.pop();
int idx = cur.match.size();
int curSz = (int)log((double)todo.size());
if(curSz > maxSz) {
maxSz = curSz;
if(maxSz > 3)
std::cout << "Reached " << todo.size() << endl;
}
if(idx == toMatch) {
output = cur;
std::cout << "Found: residual = " << cur.penalty << endl;
break;
}
for(i = 0; i < (int)possibilities[idx].size(); ++i) {
int candidate = possibilities[idx][i];
int k;
double extraPenalty = computePenalty(penaltyFunctions, cur, candidate);
if(extraPenalty < 0)
std::cout << "ERR = " << extraPenalty << endl;
if(cur.penalty + extraPenalty < 1.) {
PartialMatch next = cur;
next.match.push_back(candidate);
next.penalty += extraPenalty;
next.heuristic = next.penalty;
//compute taken vertices and edges
if(idx > 0) {
vector<int> path = fp.paths.path(candidate, next.match[skeleton.cPrev()[idx]]);
for(j = 0; j < (int)path.size(); ++j)
next.vTaken[path[j]] = true;
}
//compute heuristic
for(j = idx + 1; j < toMatch; ++j) {
if(skeleton.cPrev()[j] > idx)
continue;
double minP = 1e37;
for(k = 0; k < (int)possibilities[j].size(); ++k) {
minP = min(minP, computePenalty(penaltyFunctions, next, possibilities[j][k], j));
}
next.heuristic += minP;
if(next.heuristic > 1.)
break;
}
if(next.heuristic > 1.)
continue;
todo.push(next);
}
}
}
if(output.match.size() == 0)
{
std::cout << "No Match" << endl;
}
for(i = 0; i < (int)penaltyFunctions.size(); ++i)
delete penaltyFunctions[i];
return output.match;
}
