Revamped the rootfinder algorithm.

Behavior is substantially different.
A summary table is now written out at the end of the solve for
some print levels.
This commit is contained in:
Harry Moffat 2011-09-23 20:31:17 +00:00
parent 445adce8fc
commit d251da7a84
2 changed files with 510 additions and 136 deletions

View file

@ -18,7 +18,7 @@
#ifdef DEBUG_MODE
#include "mdp_allo.h"
#endif
#include "stringUtils.h"
/* Standard include files */
#include <cstdio>
@ -48,6 +48,13 @@ namespace Cantera {
#define DSIGN(x) (( (x) == (0.0) ) ? (0.0) : ( ((x) > 0.0) ? 1.0 : -1.0 ))
#endif
#ifdef SWAP
#undef SWAP
#endif
#ifndef SWAP
#define SWAP(x1, x2, tmp) ((tmp) = (x2), (x2) = (x1), (x1) = (tmp))
#endif
// turn on debugging for now
#ifndef DEBUG_MODE
#define DEBUG_MODE
@ -275,17 +282,20 @@ namespace Cantera {
// Function to decide whether two real numbers are the same or not
/*
* A comparison is made between the two numbers to decide whether they
* are close to one another. This is defined as being within delXMeaningful() of each other
* are close to one another. This is defined as being within factor * delXMeaningful() of each other.
*
* @param x2 First number
* @param x2 second number
* @param x1 second number
* @param factor Multiplicative factor for delta X. defaults to 1
*
* @return Returns a boolean indicating whether the two numbers are the same or not.
*/
bool RootFind::theSame(doublereal x2, doublereal x1) const {
bool RootFind::theSame(doublereal x2, doublereal x1, doublereal factor) const {
doublereal x = fabs(x2) + fabs(x1);
doublereal deltaX = delXMeaningful(x);
if (fabs(x2 - x1) < deltaX) {
doublereal deltaXSmall = factor * deltaX;
deltaXSmall = MAX(deltaXSmall , x * 1.0E-15);
if (fabs(x2 - x1) < deltaXSmall) {
return true;
}
return false;
@ -322,9 +332,10 @@ namespace Cantera {
char fileName[80];
FILE *fp = 0;
#endif
int doFinalFuncCall = 0;
doublereal x1, x2, xnew, f1, f2, fnew, slope;
doublereal deltaX1 = 0.0, deltaX2 = 0.0, deltaXnew = 0.0;
int its = 0;
int posStraddle = 0;
int retn = ROOTFIND_FAILEDCONVERGENCE;
int foundPosF = 0;
@ -335,11 +346,15 @@ namespace Cantera {
doublereal xNegF = 0.0;
doublereal fNegF = -1.0E300;
doublereal fnorm; /* A valid norm for the making the function value dimensionless */
doublereal c[9], f[3], xn1, xn2, x0 = 0.0, f0 = 0.0, root, theta, xquad, xDelMin;
doublereal CR0, CR1, CR2, CRnew, CRdenom;
doublereal x0 = 0.0, f0 = 0.0, xDelMin;
doublereal sgn;
doublereal dtmp;
doublereal fnoise = 0.0;
rfHistory_.clear();
rfTable rfT;
rfT.clear();
rfT.reasoning = "First Point: ";
callNum++;
#ifdef DEBUG_MODE
if (printLvl >= 3 && writeLogAllowed_) {
@ -397,10 +412,14 @@ namespace Cantera {
x1 = *xbest;
if (x1 < xmin || x1 > xmax) {
x1 = (xmin + xmax) / 2.0;
rfT.reasoning += " x1 set middle between xmin and xmax because entrance is outside bounds.";
} else {
rfT.reasoning += " x1 set to entrance x.";
}
x_maxTried_ = x1;
x_minTried_ = x1;
int its = 1;
f1 = func(x1);
#ifdef DEBUG_MODE
@ -410,45 +429,62 @@ namespace Cantera {
}
#endif
if (f1 == 0.0) {
*xbest = x1;
return 0;
} else if (f1 > 0.0) {
} else if (f1 > fnoise) {
foundPosF = 1;
xPosF = x1;
fPosF = f1;
} else {
} else if (f1 < -fnoise) {
foundNegF = 1;
xNegF = x1;
fNegF = x1;
fNegF = f1;
}
rfT.its = its;
rfT.TP_its = 0;
rfT.xval = x1;
rfT.fval = f1;
rfT.foundPos = foundPosF;
rfT.foundNeg = foundNegF;
rfT.deltaXConverged = m_rtolx * (fabs(x1) + 0.001);
rfT.deltaFConverged = fabs(f1) * m_rtolf;
rfT.delX = xmax - xmin;
rfHistory_.push_back(rfT);
rfT.clear();
/*
* Now, this is actually a tricky part of the algorithm - Find the x value for
* the second point. It's tricky because we don't have a valid idea of the scale of x yet
*
*/
rfT.reasoning = "Second Point: ";
if (x1 == 0.0) {
x2 = x1 + 0.01 * DeltaXnorm_;
rfT.reasoning += "Set by DeltaXnorm_";
} else {
x2 = x1 * 1.0001;
rfT.reasoning += "Set slightly higher.";
}
if (x2 > xmax) {
x2 = x1 - 0.01 * DeltaXnorm_;
rfT.reasoning += " - But adjusted to be within bounds";
}
/*
* Find the second function value f2 = func(x2), Process it
*/
deltaX2 = x2 - x1;
f2 = func(x2);
its++;
f2 = func(x2);
#ifdef DEBUG_MODE
if (printLvl >= 3 && writeLogAllowed_) {
print_funcEval(fp, x2, f2, its);
fprintf(fp, "%-5d %-5d %-15.5E %-15.5E", -1, 0, x2, f2);
}
#endif
/*
* Calculate the norm of the function, this is the nominal value of f. We try
* to reduce the nominal value of f by rtolf, this is the main convergence requirement.
@ -458,23 +494,33 @@ namespace Cantera {
} else {
fnorm = 0.5*(fabs(f1) + fabs(f2)) + fabs(m_funcTargetValue) + m_atolf;
}
fnoise = 1.0E-100;
if (f2 == 0.0) {
*xbest = x2;
return ROOTFIND_SUCCESS;
} else if (f2 > 0.0) {
if (f2 > fnoise) {
if (!foundPosF) {
foundPosF = 1;
xPosF = x2;
fPosF = x2;
fPosF = f2;
}
} else {
} else if (f2 < - fnoise) {
if (!foundNegF) {
foundNegF = 1;
xNegF = x2;
fNegF = f2;
}
} else if (f2 == 0.0) {
*xbest = x2;
return ROOTFIND_SUCCESS;
}
rfT.its = its;
rfT.TP_its = 0;
rfT.xval = x2;
rfT.fval = f2;
rfT.foundPos = foundPosF;
rfT.foundNeg = foundNegF;
/*
* See if we have already achieved a straddle
*/
@ -483,7 +529,7 @@ namespace Cantera {
if (xPosF > xNegF) posStraddle = 1;
else posStraddle = 0;
}
bool doQuad = false;
bool useNextStrat = false;
bool slopePointingToHigher = true;
// ---------------------------------------------------------------------------------------------
@ -499,8 +545,13 @@ namespace Cantera {
printf(" RootFind: we are here x2 = %g x1 = %g\n", x2, x1);
}
#endif
doublereal delXtmp = deltaXControlled(x2, x1);
slope = (f2 - f1) / delXtmp;
rfT.slope = slope;
rfHistory_.push_back(rfT);
rfT.clear();
rfT.reasoning = "";
if (fabs(slope) <= 1.0E-100) {
if (printLvl >= 2) {
writelogf("%s functions evals produced the same result, %g, at %g and %g\n",
@ -509,6 +560,7 @@ namespace Cantera {
xnew = x2 + DeltaXnorm_;
slopePointingToHigher = true;
useNextStrat = true;
rfT.reasoning += "Slope is close to zero. ";
} else {
useNextStrat = false;
xnew = x2 - f2 / slope;
@ -517,6 +569,7 @@ namespace Cantera {
} else {
slopePointingToHigher = false;
}
rfT.reasoning += "Slope is good. ";
}
#ifdef DEBUG_MODE
if (printLvl >= 3 && writeLogAllowed_) {
@ -530,13 +583,18 @@ namespace Cantera {
if (!foundStraddle) {
if (fabs(xnew - x2) > DeltaXMax_) {
useNextStrat = true;
rfT.reasoning += "Too large change in xnew from slope. ";
}
if (fabs(deltaXnew) < fabs(deltaX2)) {
deltaXnew = 1.2 * deltaXnew;
xnew = x2 + deltaXnew;
}
}
/*
* If the slope can't be trusted using a different strategy for picking the next point
*/
if (useNextStrat) {
rfT.reasoning += "Using DeltaXnorm, " + fp2str(DeltaXnorm_) + " and FuncIsGenerallyIncreasing hints. ";
if (f2 < 0.0) {
if (FuncIsGenerallyIncreasing_) {
if (slopePointingToHigher) {
@ -580,53 +638,24 @@ namespace Cantera {
}
}
/*
* Do a quadratic fit -> Note this algorithm seems
* to work OK. The quadratic approximation doesn't kick in until
* the end of the run, when it becomes reliable.
* Here, if we have a straddle, we purposefully overshoot the smaller side by 5%. Yes it does lead to
* more iterations. However, we're interested in bounding x, and not just doing Newton's method.
*/
if (its > 0 && doQuad) {
c[0] = 1.; c[1] = 1.; c[2] = 1.;
c[3] = x0; c[4] = x1; c[5] = x2;
c[6] = SQUARE(x0); c[7] = SQUARE(x1); c[8] = SQUARE(x2);
f[0] = - f0; f[1] = - f1; f[2] = - f2;
int rrr = smlequ(c, 3, 3, f, 1);
if (rrr == 1) goto QUAD_BAIL;
root = f[1]* f[1] - 4.0 * f[0] * f[2];
if (root >= 0.0) {
xn1 = (- f[1] + sqrt(root)) / (2.0 * f[2]);
xn2 = (- f[1] - sqrt(root)) / (2.0 * f[2]);
if (fabs(xn2 - x2) < fabs(xn1 - x2) && xn2 > 0.0 ) xquad = xn2;
else xquad = xn1;
theta = fabs(xquad - xnew) / fabs(xnew - x2);
theta = MIN(1.0, theta);
xnew = theta * xnew + (1.0 - theta) * xquad;
#ifdef DEBUG_MODE
if (printLvl >= 3 && writeLogAllowed_) {
if (theta != 1.0) {
fprintf(fp, " | xquad = %-11.5E", xnew);
}
}
#endif
} else {
/*
* Pick out situations where the convergence may be
* accelerated.
*/
if ((DSIGN(xnew - x2) == DSIGN(x2 - x1)) &&
(DSIGN(x2 - x1) == DSIGN(x1 - x0)) ) {
xnew += xnew - x2;
#ifdef DEBUG_MODE
if (printLvl >= 3 && writeLogAllowed_) {
fprintf(fp, " | xquada = %-11.5E", xnew);
}
#endif
if (foundStraddle) {
double delta = fabs(x2 - x1);
if (fabs(xnew - x1) < .01 * delta) {
xnew = x1 + 0.01 * (x2 - x1);
} else if (fabs(xnew - x2) < .01 * delta) {
xnew = x1 + 0.01 * (x2 - x1);
} else if ((xnew > x1 && xnew < x2) || (xnew < x1 && xnew > x2)) {
if (fabs(xnew - x1) < fabs(x2 - xnew)) {
xnew = x1 + 20./19. * (xnew - x1);
} else {
xnew = x2 + 20./19. * (xnew - x2);
}
}
}
QUAD_BAIL: ;
/*
* OK, we have an estimate xnew.
*
@ -637,7 +666,7 @@ namespace Cantera {
/*
* If we are doing a jump in between the two previous points, make sure
* the new trial is no closer that 10% of the distances between x2-x1 to
* any of the original points.
* any of the original points. This is an important part of finding a good bound.
*/
xDelMin = fabs(x2 - x1) / 10.;
if (fabs(xnew - x1) < xDelMin) {
@ -674,13 +703,37 @@ namespace Cantera {
if (printLvl >= 3 && writeLogAllowed_) {
fprintf(fp, " | xlimitsize = %-11.5E", xnew);
}
#endif
}
/*
* If we are doing a jump outside the two previous points, make sure
* the new trial is no closer that 10% of the distances between x2-x1 to
* any of the original points. This is an important part of finding a good bound.
*/
xDelMin = 0.1 * fabs(x2 - x1);
if (fabs(xnew - x2) < xDelMin) {
xnew = x2 + DSIGN(xnew - x2) * xDelMin;
#ifdef DEBUG_MODE
if (printLvl >= 3 && writeLogAllowed_) {
fprintf(fp, " | x10%% = %-11.5E", xnew);
}
#endif
}
if (fabs(xnew - x1) < xDelMin) {
xnew = x1 + DSIGN(xnew - x1) * xDelMin;
#ifdef DEBUG_MODE
if (printLvl >= 3 && writeLogAllowed_) {
fprintf(fp, " | x10%% = %-11.5E", xnew);
}
#endif
}
}
/*
* HKM -> Not sure this section is needed
*/
if (foundStraddle) {
#ifdef DEBUG_MODE
slope = xnew;
double xorig = xnew;
#endif
if (posStraddle) {
if (f2 > 0.0) {
@ -717,7 +770,7 @@ namespace Cantera {
}
#ifdef DEBUG_MODE
if (printLvl >= 3 && writeLogAllowed_) {
if (slope != xnew) {
if (xorig != xnew) {
fprintf(fp, " | xstraddle = %-11.5E", xnew);
}
}
@ -735,6 +788,8 @@ namespace Cantera {
sgn = -1.0;
}
deltaXnew = 1.2 * delXMeaningful(xnew) * sgn;
rfT.reasoning += "Enforcing minimum stepsize from " + fp2str(xnew - x2) +
" to " + fp2str(deltaXnew);
xnew = x2 + deltaXnew;
}
}
@ -746,20 +801,18 @@ namespace Cantera {
topBump++;
if (topBump < 3) {
xnew = x2 + (xmax - x2) / 2.0;
rfT.reasoning += ("xval reduced to " + fp2str(xnew) + " because predicted xnew was above max value of " + fp2str(xmax));
} else {
if (x2 == xmax || x1 == xmax) {
// we are here when we are bumping against the top limit.
// No further action is possible
if (xnew > xmax) {
slope = (f2 - f1) / delXtmp;
xnew = x2 - f2 / slope;
if (xnew > xmax) {
retn = ROOTFIND_SOLNHIGHERTHANXMAX;
*xbest = xnew;
}
}
retn = ROOTFIND_SOLNHIGHERTHANXMAX;
*xbest = xnew;
rfT.slope = slope;
rfT.reasoning += "Giving up because we're at xmax and xnew point higher: " + fp2str(xnew);
goto done;
} else {
rfT.reasoning += "xval reduced from " + fp2str(xnew) + " to the max value, " + fp2str(xmax);
xnew = xmax;
}
}
@ -772,21 +825,20 @@ namespace Cantera {
if (xnew < xmin) {
bottomBump++;
if (bottomBump < 3) {
rfT.reasoning += ("xnew increased from " + fp2str(xnew) +" to " + fp2str(x2 - (x2 - xmin) / 2.0) +
" because above min value of " + fp2str(xmin));
xnew = x2 - (x2 - xmin) / 2.0;
} else {
if (x2 == xmin || x1 == xmin) {
// we are here when we are bumping against the bottom limit.
// No further action is possible
if (xnew < xmin) {
slope = (f2 - f1) / delXtmp;
xnew = x2 - f2 / slope;
if (xnew < xmin) {
retn = ROOTFIND_SOLNLOWERTHANXMIN;
*xbest = xnew;
}
}
retn = ROOTFIND_SOLNLOWERTHANXMIN;
*xbest = xnew;
rfT.slope = slope;
rfT.reasoning = "Giving up because we're already at xmin and xnew points lower: " + fp2str(xnew);
goto done;
} else {
} else {
rfT.reasoning += "xval increased from " + fp2str(xnew) + " to the min value, " + fp2str(xmin);
xnew = xmin;
}
}
@ -796,10 +848,10 @@ namespace Cantera {
}
#endif
}
its++;
fnew = func(xnew);
CRdenom = MAX(fabs(fnew), MAX(fabs(f2), MAX(fabs(f1), fnorm)));
CRnew = sqrt(fabs(fnew) / CRdenom);
#ifdef DEBUG_MODE
if (printLvl >= 3 && writeLogAllowed_) {
fprintf(fp,"\n");
@ -807,7 +859,9 @@ namespace Cantera {
fprintf(fp, "%-5d %-5d %-15.5E %-15.5E", its, 0, xnew, fnew);
}
#endif
rfT.xval = xnew;
rfT.fval = fnew;
rfT.its = its;
if (foundStraddle) {
if (posStraddle) {
if (fnew > 0.0) {
@ -837,18 +891,20 @@ namespace Cantera {
}
if (! foundStraddle) {
if (fnew > 0.0) {
if (fnew > fnoise) {
if (!foundPosF) {
foundPosF = 1;
rfT.foundPos = 1;
xPosF = xnew;
fPosF = fnew;
foundStraddle = 1;
if (xPosF > xNegF) posStraddle = 1;
else posStraddle = 0;
}
} else {
} else if (fnew < - fnoise) {
if (!foundNegF) {
foundNegF = 1;
rfT.foundNeg = 1;
xNegF = xnew;
fNegF = fnew;
foundStraddle = 1;
@ -860,15 +916,14 @@ namespace Cantera {
x0 = x1;
f0 = f1;
CR0 = CR1;
x1 = x2;
f1 = f2;
CR1 = CR2;
x2 = xnew;
f2 = fnew;
/*
* As we go on to new data points, we make sure that
* As we go on to new data points, we make sure that
* we have the best straddle of the solution with the choice of F1 and F2 when
* we do have a straddle to work with.
*/
@ -947,64 +1002,190 @@ namespace Cantera {
}
}
}
}
}
AssertThrow((f1 * f2 <= 0.0), "F1 and F2 aren't bounding");
}
deltaX1 = deltaX2;
deltaX2 = deltaXnew;
CR2 = CRnew;
deltaXnew = x2 - x1;
deltaXConverged_ = 0.5 * deltaXConverged_ + 0.5 * (m_rtolx * 0.5 * (fabs(x2) + fabs(x1)) + m_atolx);
if (fabs(fnew / fnorm) < m_rtolf) {
if (deltaX2 < deltaXConverged_ && deltaXnew < deltaXConverged_) {
converged = 1;
rfT.deltaXConverged = deltaXConverged_;
rfT.deltaFConverged = fnorm * m_rtolf;
if (foundStraddle) {
rfT.delX = MAX(fabs(deltaX2), fabs(deltaXnew));
} else {
rfT.delX = MAX(fabs(deltaX2), fabs(deltaXnew));
if (x2 < x1) {
rfT.delX = MAX(rfT.delX, x2 - xmin);
} else {
rfT.delX = MAX(rfT.delX, xmax - x2);
}
if (fabs(slope) > 1.0E-100) {
double xdels = fabs(fnew / slope);
if (xdels < deltaXConverged_ * 0.5) {
converged = 1;
}
}
}
/*
* Check for excess convergence in the x coordinate
* Section To Determine CONVERGENCE criteria
*/
if (foundStraddle) {
doublereal denom = fabs(x1) + fabs(x2);
if (denom < 1.0E-200) {
retn = ROOTFIND_FAILEDCONVERGENCE;
converged = true;
doFinalFuncCall = 0;
if ((fabs(fnew / fnorm) < m_rtolf) && foundStraddle) {
if (fabs(deltaX2) < deltaXConverged_ && fabs(deltaXnew) < deltaXConverged_) {
converged = 1;
rfT.reasoning += "NormalConvergence";
retn = ROOTFIND_SUCCESS;
}
else if (fabs(slope) > 1.0E-100) {
double xdels = fabs(fnew / slope);
if (xdels < deltaXConverged_ * 0.3) {
converged = 1;
rfT.reasoning += "NormalConvergence-SlopelimitsDelX";
doFinalFuncCall = 1;
retn = ROOTFIND_SUCCESS;
}
}
if (theSame(x2, x1)) {
converged = true;
/*
* Check for excess convergence in the x coordinate
*/
if (!converged) {
if (foundStraddle) {
doublereal denom = fabs(x1 - x2);
if (denom < 1.0E-200) {
retn = ROOTFIND_FAILEDCONVERGENCE;
converged = true;
rfT.reasoning += "ConvergenceFZero but X1X2Identical";
}
if (theSame(x2, x1, 1.0E-2)) {
converged = true;
rfT.reasoning += " ConvergenceF and XSame";
retn = ROOTFIND_SUCCESS;
}
}
}
} else {
/*
* We are here when F is not converged, but we may want to end anyway
*/
if (!converged) {
if (foundStraddle) {
doublereal denom = fabs(x1 - x2);
if (denom < 1.0E-200) {
retn = ROOTFIND_FAILEDCONVERGENCE;
converged = true;
rfT.reasoning += "FNotConverged but X1X2Identical";
}
/*
* The premise here is that if x1 and x2 get close to one another,
* then the accuracy of the calculation gets destroyed.
*/
if (theSame(x2, x1, 1.0E-5)) {
converged = true;
retn = ROOTFIND_SUCCESS_XCONVERGENCEONLY;
rfT.reasoning += "FNotConverged but XSame";
}
}
}
}
its++;
} while (! converged && its < itmax);
done:
if (converged) {
retn = ROOTFIND_SUCCESS;
if (fabs(f1) < 2.0 * fabs(f2)) {
slope = (f2 - f1) / (x2 - x1);
xnew = x2 - f2 / slope;
rfT.slope = slope;
rfHistory_.push_back(rfT);
rfT.clear();
rfT.its = its;
AssertThrow((f1 * f2 <= 0.0), "F1 and F2 aren't bounding");
double x_fpos = x2;
double x_fneg = x1;
if (f2 < 0.0) {
x_fpos = x1;
x_fneg = x2;
}
rfT.delX = fabs(x_fpos - x_fneg);
if (doFinalFuncCall || (fabs(f1) < 2.0 * fabs(f2))) {
double delXtmp = deltaXControlled(x2, x1);
slope = (f2 - f1) / delXtmp;
xnew = x2 - f2 / slope;
its++;
fnew = func(xnew);
if (fabs(fnew) < fabs(f2)) {
if (fnew > 0.0) {
if (fabs(xnew - x_fneg) < fabs(x_fpos - x_fneg)) {
x_fpos = xnew;
rfT.delX = fabs(xnew - x_fneg);
}
} else {
if (fabs(xnew - x_fpos) < fabs(x_fpos - x_fneg)) {
x_fneg = xnew;
rfT.delX = fabs(xnew - x_fpos);
}
}
rfT.its = its;
if (fabs(fnew) < fabs(f2) && (fabs(fnew) < fabs(f1))) {
*xbest = xnew;
if (doFinalFuncCall) {
rfT.reasoning += "CONVERGENCE: Another Evaluation Requested";
rfT.delX = fabs(xnew - x2);
} else {
rfT.reasoning += "CONVERGENCE: Another Evaluation done because f1 < f2";
rfT.delX = fabs(xnew - x1);
}
rfT.fval = fnew;
rfT.xval = xnew;
x2 = xnew;
f2 = fnew;
} else if (fabs(f1) < fabs(f2)) {
rfT.its = its;
rfT.xval = xnew;
rfT.fval = fnew;
rfT.slope = slope;
rfT.reasoning += "CONVERGENCE: Another Evaluation not as good as Second Point ";
rfHistory_.push_back(rfT);
rfT.clear();
rfT.its = its;
SWAP(f1, f2, dtmp);
SWAP(x1, x2, dtmp);
*xbest = x2;
if (fabs(fnew) < fabs(f1)) {
if (f1 * fnew > 0.0) {
SWAP(f1, fnew, dtmp);
SWAP(x1, xnew, dtmp);
}
}
rfT.its = its;
rfT.xval = *xbest;
rfT.fval = f2;
rfT.delX = fabs(x_fpos - x_fneg);
rfT.reasoning += "CONVERGENCE: NormalEnding -> Second point used";
} else {
rfT.its = its;
rfT.xval = xnew;
rfT.fval = fnew;
rfT.slope = slope;
rfT.reasoning += "CONVERGENCE: Another Evaluation not as good as First Point ";
rfHistory_.push_back(rfT);
rfT.clear();
rfT.its = its;
*xbest = x2;
rfT.xval = *xbest;
rfT.fval = f2;
rfT.delX = fabs(x_fpos - x_fneg);
rfT.reasoning += "CONVERGENCE: NormalEnding -> Last point used";
}
if (fabs(f1) < fabs(f2)) {
x2 = x1;
f2 = f1;
*xbest = x1;
fnew = func(x2);
}
} else {
*xbest = x2;
rfT.xval = *xbest;
rfT.fval = f2;
rfT.delX = fabs(x2 - x1);
rfT.reasoning += "CONVERGENCE: NormalEnding -> Last point used";
}
funcTargetValue = f2 + m_funcTargetValue;
rfT.slope = slope;
if (printLvl >= 1) {
writelogf("RootFind success: convergence achieved\n");
@ -1013,35 +1194,50 @@ namespace Cantera {
if (printLvl >= 3 && writeLogAllowed_) {
fprintf(fp, " | RootFind success in %d its, fnorm = %g\n", its, fnorm);
}
#endif
#endif
rfHistory_.push_back(rfT);
} else {
rfT.reasoning = "FAILED CONVERGENCE ";
rfT.slope = slope;
rfT.its = its;
if (retn == ROOTFIND_SOLNHIGHERTHANXMAX) {
if (printLvl >= 1) {
writelogf("RootFind ERROR: Soln probably lies higher than xmax, %g: best guess = %g\n", xmax, *xbest);
}
rfT.reasoning += "Soln probably lies higher than xmax, " + fp2str(xmax) + ": best guess = " + fp2str(*xbest);
} else if (retn == ROOTFIND_SOLNLOWERTHANXMIN) {
if (printLvl >= 1) {
writelogf("RootFind ERROR: Soln probably lies lower than xmin, %g: best guess = %g\n", xmin, *xbest);
}
rfT.reasoning += "Soln probably lies lower than xmin, " + fp2str(xmin) + ": best guess = " + fp2str(*xbest);
} else {
retn = ROOTFIND_FAILEDCONVERGENCE;
if (printLvl >= 1) {
writelogf("RootFind ERROR: maximum iterations exceeded without convergence, cause unknown\n");
}
rfT.reasoning += "Maximum iterations exceeded without convergence, cause unknown";
}
#ifdef DEBUG_MODE
if (printLvl >= 3 && writeLogAllowed_) {
fprintf(fp, "\nRootFind failure in %d its\n", its);
}
#endif
*xbest = x2;
funcTargetValue = f2 + m_funcTargetValue;
rfT.xval = *xbest;
rfT.fval = f2;
rfHistory_.push_back(rfT);
}
*xbest = x2;
funcTargetValue = f2 + m_funcTargetValue;
#ifdef DEBUG_MODE
if (printLvl >= 3 && writeLogAllowed_) {
fclose(fp);
}
#endif
if (printLvl >= 2) {
printTable();
}
return retn;
}
@ -1173,4 +1369,22 @@ namespace Cantera {
specifiedDeltaXMax_ = 1;
}
//====================================================================================================================
//====================================================================================================================
void RootFind::printTable() {
printf("\t----------------------------------------------------------------------------------------------------------------------------------------\n");
printf("\t RootFinder Summary table: \n");
printf("\t FTarget = %g\n", m_funcTargetValue);
printf("\t Iter | xval delX deltaXConv | slope | foundP foundN| F - F_targ deltaFConv | Reasoning\n");
printf("\t----------------------------------------------------------------------------------------------------------------------------------------\n");
for (int i = 0; i < (int) rfHistory_.size(); i++) {
struct rfTable rfT = rfHistory_[i];
printf("\t %3d |%- 17.11E %- 13.7E %- 13.7E |%- 13.5E| %3d %3d | %- 12.5E %- 12.5E | %s \n",
rfT.its, rfT.xval, rfT.delX, rfT.deltaXConverged, rfT.slope, rfT.foundPos, rfT.foundNeg, rfT.fval,
rfT.deltaFConverged, (rfT.reasoning).c_str());
}
printf("\t----------------------------------------------------------------------------------------------------------------------------------------\n");
}
//====================================================================================================================
}

View file

@ -29,6 +29,14 @@ namespace Cantera {
//! This means that the root solver was a success
#define ROOTFIND_SUCCESS 0
//! This return value means that the root finder resolved a solution in the x coordinate
//! However, convergence in F was not achieved.
/*!
* A common situation for this to happen is that f(x) is discontinuous about f(x) = f_0,
* where we seek the x where the function is equal to f_0. f(x) spans the
* f_0 while not being equal to f_0 anywhere.
*/
#define ROOTFIND_SUCCESS_XCONVERGENCEONLY 1
//! This means that the root solver failed to achieve convergence
#define ROOTFIND_FAILEDCONVERGENCE -1
//! This means that the input to the root solver was defective
@ -40,12 +48,96 @@ namespace Cantera {
*
*/
#define ROOTFIND_SOLNHIGHERTHANXMAX -4
//@{
//@}
//! Root finder for 1D problems
/*!
*
* The root finder solves a single nonlinear equation described below.
*
* \f[
* f(x) = f_0
* \f]
*
* \f$ f(x) \f$ is assumed to be single valued as a function of x.\f$ f(x) \f$ is not assumed to be continuous nor is
* its derivative assumed to be well formed.
*
* Root finders are significantly different in the sense that do not have to rely
* solely on Newton's method to find the answer to the problem. Instead they use a method to bound
* the solution between high and low values and then use a method to refine that bound. The eventual
* solution to the problem is presented as x_best and as a bound, delta_X, on the solution
* component. Because of this, they are far more stable for functions and Jacobians that have discontinuities
* or noise associated with them.
*
* The algorithm is a convolution of a local Secant method with an approach of finding a straddle in x.
* The Jacobian is never required.
*
* There is a general breakdown of the algorithm into stages. The first stage seeks to find a straddle of the
* function. The second stage seeks to reduce the bounds in x and f in order to satisfy the specification of the
* stopping criteria. In the last stage the algorithm seeks to find the base value of x that satisfies the
* original equation given what it current knows about the function.
*
* Globalization strategy
*
* Specifying the General Changes in x
*
* Supplying Hints with General Function Behavior Flags
*
*
*
* Stopping Criteria
*
* Specification of the Stopping Criteria
*
*
* Additional constraints
*
* Bounds Criteria For the Routine
*
* Example
*
* @code
* // Define a residual. The definition of a residual involves a lot more work than is shown here.
* ResidEval * ec;
* // Instantiate the root finder with the residual to be solved, ec.
* RootFind rf(&ec);
* // Set the relative and absolute tolerancess for f and x.
* rf.setTol(1.0E-5, 1.0E-10, 1.0E-5, 1.0E-11);
* // Give a hint about the function's dependence on x. This is needed, for example, if the function has
* // flat regions.
* rf.setFuncIsGenerallyIncreasing(true);
* rf.setDeltaX(0.01);
* // Supply an initial guess for the solution
* double xbest = phiM;
* double oldP = printLvl_;
* // Set the print level for the solver. Zero produces no output. Two produces a summary table of each iteration.
* rf.setPrintLvl(2);
* // Define a minimum and maximum for the independent variable.
* double phimin = 1.3;
* double phimax = 2.2;
* // Define a maximum iteration number
* int itmax = 100;
* // Define the f_0 value, and on return will contain the actual value of f(x) obtained
* double currentObtained;
* // Call the solver
* status = rf.solve(phimin, phimax, 100, currentObtained, &xbest);
* if (status == 0) {
* if (printLvl_ > 1) {
* printf("Electrode::integrateConstantCurrent(): Volts (%g amps) = %g\n", currentObtained, xbest);
* }
* } else {
* if (printLvl_) {
* printf("Electrode::integrateConstantCurrent(): bad status = %d Volts (%g amps) = %g\n",
* status, currentObtained, xbest);
* }
* }
* @endcode
*
* @todo Noise
* @todo General Search to be done when all else fails
*
*/
class RootFind {
@ -110,14 +202,18 @@ namespace Cantera {
//! Function to decide whether two real numbers are the same or not
/*!
* A comparison is made between the two numbers to decide whether they
* are close to one another. This is defined as being within delXMeaningful() of each other
* are close to one another. This is defined as being within factor * delXMeaningful() of each other.
*
* The basic premise here is that if the two numbers are too close, the noise
* will prevent an accurate calculation of the function and its slope.
*
* @param x1 First number
* @param x2 second number
* @param factor Multiplicative factor to multiple deltaX with
*
* @return Returns a boolean indicating whether the two numbers are the same or not.
*/
bool theSame(doublereal x2, doublereal x1) const;
bool theSame(doublereal x2, doublereal x1, doublereal factor = 1.0) const;
public:
@ -233,6 +329,9 @@ namespace Cantera {
*/
void setDeltaXMax(doublereal deltaX);
//! Print the iteration history table
void printTable();
public:
//! Pointer to the residual function evaluator
@ -249,6 +348,7 @@ namespace Cantera {
//! Relative tolerance for the value of f and x
doublereal m_rtolf;
//! Relative tolerance for the value of x
doublereal m_rtolx;
@ -256,10 +356,19 @@ namespace Cantera {
doublereal m_maxstep;
protected:
//! Print level
/*!
* 0 No printing of any kind
* 1 Single print line indicating success or failure of the routine.
* 2 Summary table printed at the end of the routine, with a convergence history
* 3 Printouts during the iteration are added. Summary table is printed out at the end.
* if writeLogAllowed_ is turned on, a file is written out with the convergence history.
*/
int printLvl;
public:
//! Boolean to turn on the possibility of writing a log file.
bool writeLogAllowed_;
@ -304,6 +413,57 @@ namespace Cantera {
//! Internal variable tracking f(x) of smallest x tried.
doublereal fx_minTried_;
//! Structure containing the iteration history
struct rfTable {
//@{
int its;
int TP_its;
double slope;
double xval;
double fval;
int foundPos;
int foundNeg;
double deltaXConverged;
double deltaFConverged;
double delX;
std::string reasoning;
void clear() {
its = 0;
TP_its = 0;
slope = -1.0E300;
xval = -1.0E300;
fval = -1.0E300;
reasoning = "";
};
rfTable() :
its(-2),
TP_its(0),
slope(-1.0E300),
xval(-1.0E300),
fval(-1.0E300),
foundPos(0),
foundNeg(0),
deltaXConverged(-1.0E300),
deltaFConverged(-1.0E300),
delX(-1.0E300),
reasoning("")
{
};
//@}
};
//! Vector of iteration histories
std::vector<struct rfTable> rfHistory_;
};
}
#endif