Backup_Repos/jsketcher
1
0
Fork 0
mirror of https://github.com/xibyte/jsketcher synced 2025-12-15 12:53:52 +01:00
Code Issues Projects Releases Wiki Activity

cleanup/refactoring dog_leg

Browse source
This commit is contained in:
Val Erastov 2016-09-09 11:24:20 -07:00
parent c67afb1e71
commit c9c805a5fe
1 changed files with 46 additions and 47 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

93
web/app/math/optim.js
View file

@ -343,57 +343,60 @@ var dog_leg = function (subsys, rough) {
returnCode = 4;
} else if (err > divergingLim || err != err) {
returnCode = 6;
}
if (returnCode != 0) {
break;
}
// get the gauss-newton step
//h_gn = n.solve(Jx, n.mul(fx, -1));
h_gn = lsolve(Jx, n.mul(fx, -1));
//LU-Decomposition
//h_gn = lusolve(Jx, n.mul(fx, -1));
//Conjugate gradient method
//h_gn = cg(Jx, h_gn, n.mul(fx, -1), 1e-8, iterLimit);
//solve linear problem using svd formula to get the gauss-newton step
//h_gn = lls(Jx, n.mul(fx, -1));
var hitBoundary = false;
// compute the dogleg step
var gnorm = n.norm2(g);
var gnNorm = n.norm2(h_gn);
if (gnNorm < delta) {
h_dl = h_gn;
} else {
// get the gauss-newton step
//h_gn = n.solve(Jx, n.mul(fx, -1));
h_gn = lsolve(Jx, n.mul(fx, -1));
//LU-Decomposition
//h_gn = lusolve(Jx, n.mul(fx, -1));
//Conjugate gradient method
//h_gn = cg(Jx, h_gn, n.mul(fx, -1), 1e-8, iterLimit);
//solve linear problem using svd formula to get the gauss-newton step
//h_gn = lls(Jx, n.mul(fx, -1));
var hitBoundary = false;
// compute the dogleg step
var gnorm = n.norm2(g);
var gnNorm = n.norm2(h_gn);
if (gnNorm < delta) {
h_dl = h_gn;
var Jt = n.transpose(Jx);
var B = n.dot(Jt, Jx);
var gBg = n.dot(g, n.dot(B, g));
alpha = n.norm2Squared(g) / gBg;
if (alpha * gnorm >= delta) {
h_dl = n.mul(g, - delta / gnorm);
hitBoundary = true;
} else {
var Jt = n.transpose(Jx);
var B = n.dot(Jt, Jx);
var gBg = n.dot(g, n.dot(B, g));
alpha = n.norm2Squared(g) / gBg;
if (alpha * gnorm >= delta) {
h_dl = n.mul(g, - delta / gnorm);
hitBoundary = true;
var h_sd = n.mul(g, - alpha);
if (isNaN(gnNorm)) {
h_dl = h_sd;
} else {
var h_sd = n.mul(g, - alpha);
if (isNaN(gnNorm)) {
h_dl = h_sd;
} else {
var d = n.sub(h_gn, h_sd);
var d = n.sub(h_gn, h_sd);
var a = n.dot(d, d);
var b = 2 * n.dot(h_sd, d);
var c = n.dot(h_sd, h_sd) - delta * delta;
var a = n.dot(d, d);
var b = 2 * n.dot(h_sd, d);
var c = n.dot(h_sd, h_sd) - delta * delta;
var sqrt_discriminant = Math.sqrt(b * b - 4 * a * c);
var sqrt_discriminant = Math.sqrt(b * b - 4 * a * c);
var beta = (-b + sqrt_discriminant) / (2 * a);
var beta = (-b + sqrt_discriminant) / (2 * a);
// and update h_dl and dL with beta
h_dl = n.add(h_sd, n.mul(beta, d));
hitBoundary = true;
}
}
// and update h_dl and dL with beta
h_dl = n.add(h_sd, n.mul(beta, d));
hitBoundary = true;
}
}
}
@ -403,10 +406,6 @@ var dog_leg = function (subsys, rough) {
// returnCode = 5;
// break;
// }
if (returnCode != 0) {
break;
}
x_new = n.add(x, h_dl);
subsys.setParams(x_new);