Page Menu
Home
HEPForge
Search
Configure Global Search
Log In
Files
F7877580
No One
Temporary
Actions
View File
Edit File
Delete File
View Transforms
Subscribe
Mute Notifications
Award Token
Flag For Later
Size
25 KB
Subscribers
None
View Options
diff --git a/ACDC/ACDCGen.icc b/ACDC/ACDCGen.icc
--- a/ACDC/ACDCGen.icc
+++ b/ACDC/ACDCGen.icc
@@ -1,822 +1,822 @@
// -*- C++ -*-
//
// ACDCGen.icc is a part of ThePEG - Toolkit for HEP Event Generation
// Copyright (C) 1999-2011 Leif Lonnblad
//
// ThePEG is licenced under version 2 of the GPL, see COPYING for details.
// Please respect the MCnet academic guidelines, see GUIDELINES for details.
//
namespace ACDCGenerator {
template <typename Rnd, typename FncPtr>
inline ACDCGen<Rnd,FncPtr>::ACDCGen(Rnd * r)
: theRnd(r), theNAcc(0), theN(0), theNI(1, 0),
theSumW(1, 0.0), theSumW2(1, 0.0),
theEps(100*std::numeric_limits<double>::epsilon()), theMargin(1.1),
theNTry(100), theMaxTry(10000), useCheapRandom(false), theFunctions(1),
theDimensions(1, 0), thePrimaryCells(1), theSumMaxInts(1, 0.0), theLast(0),
theLastCell(0), theLastF(0.0) {
maxsize = 0;
}
template <typename Rnd, typename FncPtr>
inline ACDCGen<Rnd,FncPtr>::ACDCGen()
: theRnd(0), theNAcc(0), theN(0), theNI(1, 0),
theSumW(1, 0.0), theSumW2(1, 0.0),
theEps(100*std::numeric_limits<double>::epsilon()), theMargin(1.1),
theNTry(100), theMaxTry(10000), useCheapRandom(false), theFunctions(1),
theDimensions(1, 0), thePrimaryCells(1), theSumMaxInts(1, 0.0), theLast(0),
theLastCell(0), theLastF(0.0) {
maxsize = 0;
}
template <typename Rnd, typename FncPtr>
inline ACDCGen<Rnd,FncPtr>::~ACDCGen() {
clear();
}
template <typename Rnd, typename FncPtr>
inline void ACDCGen<Rnd,FncPtr>::setRnd(Rnd * r) {
theRnd = r;
}
template <typename Rnd, typename FncPtr>
inline void ACDCGen<Rnd,FncPtr>::clear() {
theNAcc = 0;
theN = 0;
theNI = vector<long>(1, 0);
theSumW = DVector(1, 0.0);
theSumW2 = DVector(1, 0.0);
theFunctions = FncVector(1);
theDimensions = DimVector(1, 0);
for ( int i = 0, N = thePrimaryCells.size(); i < N; ++i )
if ( thePrimaryCells[i] ) delete thePrimaryCells[i];
thePrimaryCells = CellVector(1);
theSumMaxInts = DVector(1, 0.0);
theLast = 0;
theLastCell = 0;
theLastPoint.clear();
theLastF = 0.0;
levels.clear();
}
template <typename Rnd, typename FncPtr>
inline bool ACDCGen<Rnd,FncPtr>::
addFunction(DimType dim, FncPtrType fnc, double maxrat) {
if ( maxrat < 0.0 ) maxrat = 1.0/nTry();
- typedef map<double,DVector> PointMap;
+ typedef multimap<double,DVector> PointMap;
theLast = theFunctions.size();
theFunctions.push_back(fnc);
theNI.push_back(0);
theSumW.push_back(0.0);
theSumW2.push_back(0.0);
theDimensions.push_back(dim);
// Generate nTry() points with non-zero function value
DVector x(dim);
PointMap pmap;
long itry = 0;
while ( pmap.size() < nTry() ) {
if ( ++itry > maxTry() ) {
thePrimaryCells.push_back(new ACDCGenCell(0.0));
theSumMaxInts.push_back(theSumMaxInts.back() + cells().back()->doMaxInt());
return false;
}
rnd(dim, x);
double val = FncTraits::value(fnc, x);
if ( val > 0.0 ) {
- pmap[val] = x;
+ pmap.insert(make_pair(val, x));
itry = 0;
}
}
// Create the root cell and set its overestimated function value to
// the smallest non-zero value found
double minf = pmap.begin()->first;
double maxf = (--pmap.end())->first;
minf = max(minf, maxrat*maxf);
// thePrimaryCells.push_back(new ACDCGenCell(pmap.begin()->first));
thePrimaryCells.push_back(new ACDCGenCell(minf));
theLastF = pmap.begin()->first;
pmap.erase(pmap.begin());
theSumMaxInts.push_back(theSumMaxInts.back() + cells().back()->doMaxInt());
// Start the divide-and-conquer procedure using the point with the
// highest function value found.
theLastF = (--pmap.end())->first;
theLastPoint = (--pmap.end())->second;
pmap.erase(--pmap.end());
DVector up(dim, 1.0);
DVector lo(dim, 0.0);
theLastCell = cells().back()->getCell(lo, lastPoint(), up);
if ( lastF() > lastCell()->g() ) {
compensate(lo, up);
levels.clear();
}
// For each other generated point check that it's below the
// overestimated value of the corresponding cell. If not start the
// divide-and-conquer using that point.
while ( !pmap.empty() ) {
theLastPoint = pmap.begin()->second;
theLastF = pmap.begin()->first;
pmap.erase(pmap.begin());
DVector up(dim, 1.0);
DVector lo(dim, 0.0);
theLastCell = cells().back()->getCell(lo, lastPoint(), up);
if ( lastF() > lastCell()->g() ) {
compensate(lo, up);
levels.clear();
}
}
// cells().back()->smooth(1.0/nTry());
return true;
}
template <typename Rnd, typename FncPtr>
inline void ACDCGen<Rnd,FncPtr>::chooseCell(DVector & lo, DVector & up) {
if ( compensating() ) {
// If we are compensating, we must choose the cell to be compensated.
up = levels.back().up;
lo = levels.back().lo;
theLastCell = levels.back().cell;
theLast = levels.back().index;
} else {
// Otherwise, first choose the function to be used and choose the
// corresponding root cell.
theLast = upper_bound(sumMaxInts().begin(), sumMaxInts().end(),
rnd()*sumMaxInts().back())
- sumMaxInts().begin();
if(theLast>=sumMaxInts().size()) {
throw ThePEG::Exception() << "Selected a function outside the allowed range"
<< " in ACDCGen::chooseCell(). This is usually due"
<< " to a floating point error (nan or inf) in the"
<< " calculation of the weight"
<< ThePEG::Exception::abortnow;
}
up = DVector(lastDimension(), 1.0);
lo = DVector(lastDimension(), 0.0);
theLastCell = lastPrimary();
}
// Now select randomly a sub-cell of the chosen cell
if ( cheapRandom() ) {
theLastCell = lastCell()->generate(lo, up, theRnd);
} else {
DVector rndv(lastDimension());
rnd(lastDimension(), rndv);
theLastCell = lastCell()->generate(lo, up, rndv);
}
}
template <typename Rnd, typename FncPtr>
inline typename ACDCGen<Rnd,FncPtr>::FncPtrType
ACDCGen<Rnd,FncPtr>::generate() {
long itry = 0;
while ( true ) {
if ( ++itry > maxTry() ) return FncPtrType();
++theN;
// First choose a function and a cell to generate in.
DVector up;
DVector lo;
chooseCell(lo, up);
// Now choose a point in that cell according to a flat distribution.
DimType D = lastDimension();
theLastPoint.resize(D);
for ( DimType d = 0; d < D; ++d ) theLastPoint[d] = rnd(lo[d], up[d]);
// Calculate the function value in this point
theLastF = FncTraits::value(lastFunction(), theLastPoint);
if ( theLastF <= 0.0 ) continue;
// If we are compensating we require the function value to be
// above the previous overestimate of the function.
if ( compensating() && lastF() < levels.back().g ) theLastF = 0.0;
double w = lastF()/lastCell()->g();
if ( w > 1.0 ) {
// If the value was above the overestimate then we must start
// the compensation procedure and the curren point is disregarded.
--theN;
compensate(lo, up);
continue;
}
// Accept the point according to the ratio of the true and
// overestimated function value.
theSumW[last()] += w;
theSumW2[last()] += w*w;
++theNI[last()];
if ( w > rnd() ) {
++theNAcc;
return lastFunction();
}
}
}
template <typename Rnd, typename FncPtr>
inline void ACDCGen<Rnd,FncPtr>::reject() {
theSumW[last()] -= 1.0;
theSumW2[last()] -= 1.0;
--theNAcc;
}
template <typename Rnd, typename FncPtr>
inline bool ACDCGen<Rnd,FncPtr>::compensating() {
while ( levels.size() && levels.back().lastN < N() ) levels.pop_back();
// Leave all levels which has reached there 'expiry date'.
return !levels.empty();
}
template <typename Rnd, typename FncPtr>
inline long ACDCGen<Rnd,FncPtr>::compleft() const {
if ( levels.empty() ) return 0;
long left = 1;
for ( int i = 0, Ni = levels.size(); i < Ni; ++i )
left = max(left, levels[i].lastN - N());
return left;
}
template <typename Rnd, typename FncPtr>
inline void ACDCGen<Rnd,FncPtr>::
compensate(const DVector & lo, const DVector & up) {
//Save the previous overestimated integral and create a new
//compensation level.
double i0 = maxInt();
Level level;
level.g = lastCell()->g();
// Start the divide-and-conquer algorithm slicing up the selected
// cell and specify it as the cell to compensate.
Slicer slicer(lastDimension(), *this, lo, up);
level.cell = slicer.first;
level.index = last();
level.up = slicer.firstup;
level.lo = slicer.firstlo;
// Now calculate the the new overestimated total integral and
// calculate the number of attempted points needed to compensate.
double rat = (doMaxInt())/i0;
level.lastN = long(N()*rat);
// If we are already compensating increase also the previous
// compensation levels.
for ( size_type i = 0; i < levels.size(); ++i )
levels[i].lastN = long(levels[i].lastN*rat);
levels.insert(levels.end(), level);
maxsize = std::max(maxsize, levels.size());
}
template <typename Rnd, typename FncPtr>
inline void ACDCGen<Rnd,FncPtr>::eps(double newEps) {
theEps = newEps;
}
template <typename Rnd, typename FncPtr>
inline void ACDCGen<Rnd,FncPtr>::margin(double newMargin) {
theMargin = newMargin;
}
template <typename Rnd, typename FncPtr>
inline long ACDCGen<Rnd,FncPtr>::N() const {
return theN;
}
template <typename Rnd, typename FncPtr>
inline long ACDCGen<Rnd,FncPtr>::n() const {
return theNAcc;
}
template <typename Rnd, typename FncPtr>
inline typename ACDCGen<Rnd,FncPtr>::size_type
ACDCGen<Rnd,FncPtr>::nTry() const {
return theNTry;
}
template <typename Rnd, typename FncPtr>
inline void ACDCGen<Rnd,FncPtr>::nTry(size_type newNTry) {
theNTry = newNTry;
}
template <typename Rnd, typename FncPtr>
inline long ACDCGen<Rnd,FncPtr>::maxTry() const {
return theMaxTry;
}
template <typename Rnd, typename FncPtr>
inline void ACDCGen<Rnd,FncPtr>::maxTry(long newMaxTry) {
theMaxTry = newMaxTry;
}
template <typename Rnd, typename FncPtr>
inline bool ACDCGen<Rnd,FncPtr>::cheapRandom() const {
return useCheapRandom;
}
template <typename Rnd, typename FncPtr>
inline void ACDCGen<Rnd,FncPtr>::cheapRandom(bool b) {
useCheapRandom = b;
}
template <typename Rnd, typename FncPtr>
inline double ACDCGen<Rnd,FncPtr>::maxInt() const {
return theSumMaxInts.back();
}
template <typename Rnd, typename FncPtr>
inline double ACDCGen<Rnd,FncPtr>::doMaxInt() {
for ( size_type i = 1, imax = functions().size(); i < imax; ++i )
theSumMaxInts[i] = sumMaxInts()[i - 1] + cells()[i]->doMaxInt();
return maxInt();
}
template <typename Rnd, typename FncPtr>
inline int ACDCGen<Rnd,FncPtr>::nBins() const {
int sum = 0;
for ( size_type i = 1; i < functions().size(); ++i )
sum += cell(i)->nBins();
return sum;
}
template <typename Rnd, typename FncPtr>
inline int ACDCGen<Rnd,FncPtr>::depth() const {
int mx = 0;
for ( size_type i = 1; i < functions().size(); ++i )
mx = max(mx, cell(i)->depth());
return mx;
}
template <typename Rnd, typename FncPtr>
inline double ACDCGen<Rnd,FncPtr>::efficiency() const {
return N() > 0? double(n())/double(N()): 0.0;
}
template <typename Rnd, typename FncPtr>
inline double ACDCGen<Rnd,FncPtr>::integral(FncPtrType f) const {
if ( N() <= 0 ) return maxInt();
double sumw = 0.0;
for ( size_type i = 1; i < functions().size(); ++i )
if ( functions()[i] == f || !f ) sumw += theSumW[i];
return maxInt()*sumw/N();
}
template <typename Rnd, typename FncPtr>
inline double ACDCGen<Rnd,FncPtr>::integralErr(FncPtrType f) const {
if ( N() <= 0 ) return maxInt();
double sumw2 = 0.0;
double sumw = 0.0;
for ( size_type i = 1; i < functions().size(); ++i )
if ( functions()[i] == f || !f ) {
sumw2 += theSumW2[i];
sumw += theSumW[i];
}
if ( f ) return maxInt()*sqrt(sumw2)/N();
return maxInt()*sqrt(max(0.,sumw2 - sumw*sumw/N()))/N();
}
template <typename Rnd, typename FncPtr>
inline double ACDCGen<Rnd,FncPtr>::eps() const {
return theEps;
}
template <typename Rnd, typename FncPtr>
inline double ACDCGen<Rnd,FncPtr>::margin() const {
return theMargin;
}
template <typename Rnd, typename FncPtr>
inline double ACDCGen<Rnd,FncPtr>::rnd() const {
return RndTraits::rnd(theRnd);
}
template <typename Rnd, typename FncPtr>
inline double ACDCGen<Rnd,FncPtr>::rnd(double lo, double up) const {
return RndTraits::rnd(theRnd, lo, up);
}
template <typename Rnd, typename FncPtr>
inline void ACDCGen<Rnd,FncPtr>::
rnd(const DVector & lo, const DVector & up, DVector & r) const {
RndTraits::rnd(theRnd, lo.begin(), lo.end(), up.begin(), r.begin());
}
template <typename Rnd, typename FncPtr>
inline void ACDCGen<Rnd,FncPtr>::
rnd(DimType D, DVector & r) const {
RndTraits::rnd(theRnd, D, r.begin());
}
template <typename Rnd, typename FncPtr>
inline long ACDCGen<Rnd,FncPtr>::rndInt(long x) const {
return RndTraits::rndInt(theRnd, x);
}
template <typename Rnd, typename FncPtr>
inline const typename ACDCGen<Rnd,FncPtr>::FncVector &
ACDCGen<Rnd,FncPtr>::functions() const {
return theFunctions;
}
template <typename Rnd, typename FncPtr>
inline typename ACDCGen<Rnd,FncPtr>::FncPtrType
ACDCGen<Rnd,FncPtr>::function(size_type i) const {
return functions()[i];
}
template <typename Rnd, typename FncPtr>
inline typename ACDCGen<Rnd,FncPtr>::FncPtrType
ACDCGen<Rnd,FncPtr>::lastFunction() const {
return function(last());
}
template <typename Rnd, typename FncPtr>
inline const typename ACDCGen<Rnd,FncPtr>::DimVector &
ACDCGen<Rnd,FncPtr>::dimensions() const {
return theDimensions;
}
template <typename Rnd, typename FncPtr>
inline DimType ACDCGen<Rnd,FncPtr>::dimension(size_type i) const {
return dimensions()[i];
}
template <typename Rnd, typename FncPtr>
inline DimType ACDCGen<Rnd,FncPtr>::lastDimension() const {
return dimension(last());
}
template <typename Rnd, typename FncPtr>
inline const typename ACDCGen<Rnd,FncPtr>::CellVector &
ACDCGen<Rnd,FncPtr>::cells() const {
return thePrimaryCells;
}
template <typename Rnd, typename FncPtr>
inline ACDCGenCell * ACDCGen<Rnd,FncPtr>::cell(size_type i) const {
return cells()[i];
}
template <typename Rnd, typename FncPtr>
inline ACDCGenCell * ACDCGen<Rnd,FncPtr>::lastPrimary() const {
return cell(last());
}
template <typename Rnd, typename FncPtr>
inline const DVector & ACDCGen<Rnd,FncPtr>::sumMaxInts() const {
return theSumMaxInts;
}
template <typename Rnd, typename FncPtr>
inline typename ACDCGen<Rnd,FncPtr>::size_type
ACDCGen<Rnd,FncPtr>::last() const {
return theLast;
}
template <typename Rnd, typename FncPtr>
inline typename ACDCGen<Rnd,FncPtr>::size_type
ACDCGen<Rnd,FncPtr>::size() const {
return cells().size() - 1;
}
template <typename Rnd, typename FncPtr>
inline ACDCGenCell * ACDCGen<Rnd,FncPtr>::lastCell() const {
return theLastCell;
}
template <typename Rnd, typename FncPtr>
inline const DVector & ACDCGen<Rnd,FncPtr>::lastPoint() const {
return theLastPoint;
}
template <typename Rnd, typename FncPtr>
inline double ACDCGen<Rnd,FncPtr>::lastF() const {
return theLastF;
}
template <typename Rnd, typename FncPtr>
typename ACDCGen<Rnd,FncPtr>::size_type ACDCGen<Rnd,FncPtr>::maxsize = 0;
template <typename Rnd, typename FncPtr>
vector<ACDCGenCellInfo> ACDCGen<Rnd,FncPtr>::extractCellInfo() const {
vector<ACDCGenCellInfo> ret;
for ( size_type i = 1; i < cells().size(); ++i ) {
DVector lo(dimension(i), 0.0);
DVector up(dimension(i), 1.0);
cell(i)->extract(lo, up, ret);
}
return ret;
}
template <typename Rnd, typename FncPtr>
ACDCGen<Rnd,FncPtr>::Slicer::
Slicer(DimType Din, ACDCGen & gen, const DVector & loin, const DVector & upin)
: D(Din), lo(loin), up(upin), xcl(loin), xcu(upin), xhl(loin), xhu(upin),
fhl(Din, 0.0), fhu(Din, 0.0), xsel(gen.lastPoint()), fsel(gen.lastF()),
current(gen.lastCell()), first(gen.lastCell()),
firstlo(loin), firstup(upin),f(gen.lastFunction()),
epsilon(gen.eps()), margin(gen.margin()), minf(0.0), wholecomp(false) {
divideandconquer();
}
template <typename Rnd, typename FncPtr>
ACDCGen<Rnd,FncPtr>::Slicer::~Slicer() {
// Added for debugging purposes.
}
template <typename Rnd, typename FncPtr>
void ACDCGen<Rnd,FncPtr>::Slicer::divideandconquer() {
// If the current function value was just a little above the
// overestimate, just increase the overestimate of this cell and
// we're done.
if ( fsel < current->g()*margin ) {
current->g(current->g()*margin);
return;
}
// First initialize and slice up the current cell and save the
// resulting for the compensation procedure.
init();
slice();
if ( !wholecomp ) {
first = current;
firstlo = lo;
firstup = up;
}
// Find the largest function value in the current cell and as long
// as it is above the current overestimate, increase the
// overestimate and repeat the slicing.
while ( shiftmaxmin() > current->g() ) {
current->g(minf*margin);
if ( current->g() > fsel ) return;
init();
slice();
}
}
template <typename Rnd, typename FncPtr>
double ACDCGen<Rnd,FncPtr>::Slicer::shiftmaxmin() {
// Find the largest diagonal
DVector test = xsel;
double scale = 0.0;
for ( DimType d = 0; d < D; ++d )
if ( fhl[d] > fsel || fhu[d] > fsel ) scale += 1.0;
scale = sqrt(scale);
for ( DimType d = 0; d < D; ++d ) {
if ( fhl[d] > fsel && fhl[d] > fhu[d] )
test[d] = test[d] + (xhl[d] - test[d])/scale;
if ( fhu[d] > fsel && fhu[d] > fhl[d] )
test[d] = test[d] + (xhu[d] - test[d])/scale;
}
// Find the largest value above overestimate
DimType dsel = -1;
double x = 0;
minf = fsel;
for ( DimType d = 0; d < D; ++d ) {
// Find the point with the function minimum value above the
// current overestimate.
minf = std::min(minf, fhl[d]);
minf = std::min(minf, fhu[d]);
// Find points with the maximum function value and shift the
// current point to it.
if ( fhu[d] > fsel ) {
fsel = fhu[d];
dsel = d;
x = xhu[d];
}
if ( fhl[d] > fsel ) {
fsel = fhl[d];
dsel = d;
x = xhl[d];
}
}
// Check also the largest diagonal
// double ftest = (*f)(test);
// if ( ftest > fsel ) {
// xsel = test;
// fsel = ftest;
// dsel = -1;
// }
if ( dsel >= 0 ) xsel[dsel] = x;
minf = std::max(minf, current->g());
return fsel;
}
template <typename Rnd, typename FncPtr>
void ACDCGen<Rnd,FncPtr>::Slicer::dohalf(DimType d) {
xcl[d] = lo[d];
// Find the point closest below the current point in the given
// dimension with a function value below the current overestimate,
// also find the one furthest away from the current point with a
// function value above the current overestimate. Use a crude
// iterative mid-point selection.
while ( true ) {
xhl[d] = (xsel[d] + xcl[d])*0.5;
std::swap(xsel[d], xhl[d]);
fhl[d] = FncTraits::value(f, xsel);
std::swap(xsel[d], xhl[d]);
if ( fhl[d] > current->g() ) break;
if ( xsel[d] - xcl[d] < epsilon ) break;
xcl[d] = xhl[d];
}
// Check if the current slicing is worth doing...
double cut = ( up[d] - xcl[d] )/( up[d] - lo[d] );
if ( cut < 1.0 - current->g()/fsel && cut > 0.0 )
rateslice.insert(std::make_pair(cut, -1-d));
// Find the point closest above the current point in the given
// dimension with a function value below the current overestimate,
// also find the one furthest away from the current point with a
// function value above the current overestimate. Use a crude
// iterative mid-point selection.
xcu[d] = up[d];
while ( true ) {
xhu[d] = (xsel[d] + xcu[d])*0.5;
std::swap(xsel[d], xhu[d]);
fhu[d] = FncTraits::value(f, xsel);
std::swap(xsel[d], xhu[d]);
if ( fhu[d] > current->g() ) break;
if ( xcu[d] - xsel[d] < epsilon ) break;
xcu[d] = xhu[d];
}
// Check if the current slicing is worth doing...
cut = ( xcu[d] - lo[d] )/( up[d] - lo[d] );
if ( cut < 1.0 - current->g()/fsel && cut > 0.0 )
rateslice.insert(std::make_pair(cut, 1+d));
}
template <typename Rnd, typename FncPtr>
void ACDCGen<Rnd,FncPtr>::Slicer::init() {
for ( DimType d = 0; d < D; ++d ) dohalf(d);
}
template <typename Rnd, typename FncPtr>
void ACDCGen<Rnd,FncPtr>::Slicer::slice() {
while ( !rateslice.empty() ) {
// Perform the slicing which reduces the volume of the cell the
// most first.
DimType d = rateslice.begin()->second;
rateslice.erase(rateslice.begin());
if ( d > 0 ) {
// Slice from above.
d = d - 1;
current->splitme(lo[d], xcu[d], up[d], d);
checkdiag(current->upper(), d, xcu[d], up[d]);
current = current->lower();
up[d] = xcu[d];
} else {
// Slice from below..
d = -d - 1;
current->splitme(lo[d], xcl[d], up[d], d);
checkdiag(current->lower(), d, lo[d], xcl[d]);
current = current->upper();
lo[d] = xcl[d];
}
dohalf(d);
}
}
template <typename Rnd, typename FncPtr>
void ACDCGen<Rnd,FncPtr>::Slicer::
checkdiag(ACDCGenCell * cell, DimType dc, double lod, double upd) {
return;
// Look at the midpoint in the dc direction in which a cell has been
// chopped off.
if ( upd - lod <= epsilon ) return;
DVector newlo = lo;
DVector newup = up;
newlo[dc] = lod;
newup[dc] = upd;
DVector newsel = xsel;
newsel[dc] = 0.5*(lod + upd);
double newfsel = FncTraits::value(f, newsel);
double newfh = newfsel;
DVector newxsel = newsel;
vector<int> dir(D, 0);
// For each other direction look at the mid point between the point
// chosen above and the borders of the cell. Save the point which
// gives the highest function value.
for ( DimType d = 0; d < D; ++d ) {
if ( d == dc ) continue;
double xdum = 0.5*(newlo[d] + newsel[d]);
swap(xdum, newsel[d]);
double fh1 = FncTraits::value(f, newsel);
if ( fh1 > newfsel ) {
newfsel = fh1;
newxsel = newsel;
}
if ( fh1 > newfh ) dir[d] = -1;
swap(xdum, newsel[d]);
xdum = 0.5*(newsel[d] + newup[d]);
swap(xdum, newsel[d]);
double fh2 = FncTraits::value(f, newsel);
if ( fh2 > newfsel ) {
newfsel = fh2;
newxsel = newsel;
}
if ( fh2 > newfh && fh2 > fh1 ) dir[d] = 1;
swap(xdum, newsel[d]);
}
// Now check along the diagonal where we found the highest values.
for ( DimType d = 0; d < D; ++d ) {
if ( dir[d] == 0 ) continue;
if ( dir[d] > 0 ) newsel[d] = 0.5*(newsel[d] + newup[d]);
else newsel[d] = 0.5*(newlo[d] + newsel[d]);
}
newfh = FncTraits::value(f, newsel);
if ( newfh > newfsel ) {
newfsel = newfh;
newxsel = newsel;
}
if ( newfsel < cell->g() ) return;
// If this the highest value is above the overestimate, also this
// cell needs to be divided up and conquered.
wholecomp = true;
Slicer dummy(D, *this, cell, newlo, newxsel, newup, newfsel);
}
template <typename Rnd, typename FncPtr>
ACDCGen<Rnd,FncPtr>::Slicer::
Slicer(DimType Din, const Slicer & s, ACDCGenCell * cellin,
const DVector & loin, const DVector & xselin, const DVector & upin,
double fselin)
: D(Din), lo(loin), up(upin), xcl(loin), xcu(upin), xhl(loin), xhu(upin),
fhl(Din, 0.0), fhu(Din, 0.0), xsel(xselin), fsel(fselin),
current(cellin), first(cellin),
firstlo(loin), firstup(upin),f(s.f),
epsilon(s.epsilon), margin(s.margin), minf(0.0), wholecomp(false) {
divideandconquer();
}
template <typename Rnd, typename FncPtr>
template <typename POStream>
void ACDCGen<Rnd,FncPtr>::output(POStream & os) const {
os << theNAcc << theN << theEps << theMargin << theNTry << theMaxTry
<< useCheapRandom << theLast << theLastPoint << theLastF
<< theFunctions.size() << levels.size();
for ( int i = 1, N = theFunctions.size(); i < N; ++i )
os << theFunctions[i] << theDimensions [i] << theSumMaxInts[i]
<< *thePrimaryCells[i] << theNI[i] << theSumW[i] << theSumW2[i];
os << thePrimaryCells[theLast]->getIndex(theLastCell);
for ( int i = 0, N = levels.size(); i < N; ++i )
os << levels[i].lastN << levels[i].g << levels[i].index
<< levels[i].up << levels[i].lo
<< thePrimaryCells[levels[i].index]->getIndex(levels[i].cell);
}
template <typename Rnd, typename FncPtr>
template <typename PIStream>
void ACDCGen<Rnd,FncPtr>::input(PIStream & is) {
clear();
long fsize = 0;
long lsize = 0;
is >> theNAcc >> theN >> theEps >> theMargin >> theNTry >> theMaxTry
>> useCheapRandom >> theLast >> theLastPoint >> theLastF >> fsize >> lsize;
while ( --fsize ) {
theFunctions.push_back(FncPtrType());
theDimensions.push_back(DimType());
theSumMaxInts.push_back(0.0);
theNI.push_back(0);
theSumW.push_back(0.0);
theSumW2.push_back(0.0);
thePrimaryCells.push_back(new ACDCGenCell(0.0));
is >> theFunctions.back() >> theDimensions.back() >> theSumMaxInts.back()
>> *thePrimaryCells.back() >> theNI.back()
>> theSumW.back() >> theSumW2.back();
}
long index = -1;
is >> index;
theLastCell = thePrimaryCells[theLast]->getCell(index);
while ( lsize-- ) {
levels.push_back(Level());
is >> levels.back().lastN >> levels.back().g >> levels.back().index
>> levels.back().up >> levels.back().lo >> index;
levels.back().cell = thePrimaryCells[levels.back().index]->getCell(index);
}
}
}
File Metadata
Details
Attached
Mime Type
text/x-diff
Expires
Tue, Nov 19, 3:50 PM (1 d, 19 h)
Storage Engine
blob
Storage Format
Raw Data
Storage Handle
3805025
Default Alt Text
(25 KB)
Attached To
rTHEPEGHG thepeghg
Event Timeline
Log In to Comment