jsketcher/web/app/math/optim.js
Val Erastov (xibyte) da613a082e polynomial analysis
2020-01-21 00:35:04 -08:00

501 lines
No EOL
13 KiB
JavaScript

import numeric from 'numeric';
import {_vec, _matrix} from './math'
const SUCCESS = 1, ITER_LIMIT = 2, SMALL_DELTA = 3, SMALL_STEP = 4, DIVERGENCE = 5, INVALID_STATE = 6;
//Added strong wolfe condition to numeric's uncmin
export function fmin_bfgs(f,x0,tol,gradient,maxit,callback,options) {
var grad = numeric.gradient;
if(typeof options === "undefined") { options = {}; }
if(typeof tol === "undefined") { tol = 1e-8; }
if(typeof gradient === "undefined") { gradient = function(x) { return grad(f,x); }; }
if(typeof maxit === "undefined") maxit = 1000;
x0 = numeric.clone(x0);
var n = x0.length;
var f0 = f(x0),f1,df0;
if(isNaN(f0)) throw new Error('uncmin: f(x0) is a NaN!');
var max = Math.max, norm2 = numeric.norm2;
tol = max(tol,numeric.epsilon);
var step,g0,g1,H1 = options.Hinv || numeric.identity(n);
var dot = numeric.dot, inv = numeric.inv, sub = numeric.sub, add = numeric.add, ten = numeric.tensor, div = numeric.div, mul = numeric.mul;
var all = numeric.all, isfinite = numeric.isFinite, neg = numeric.neg;
var it=0,i,s,x1,y,Hy,Hs,ys,i0,t,nstep,t1,t2;
var msg = "";
g0 = gradient(x0);
while(it<maxit) {
if(typeof callback === "function") { if(callback(it,x0,f0,g0,H1)) { msg = "Callback returned true"; break; } }
if(!all(isfinite(g0))) { msg = "Gradient has Infinity or NaN"; break; }
step = neg(dot(H1,g0));
if(!all(isfinite(step))) { msg = "Search direction has Infinity or NaN"; break; }
nstep = norm2(step);
if(nstep < tol) { msg="Newton step smaller than tol"; break; }
t = 1;
df0 = dot(g0,step);
// line search
x1 = x0;
var tL = 0;
var tR = 100;
while(it < maxit) {
if(t*nstep < tol) { break; }
s = mul(step,t);
x1 = add(x0,s);
f1 = f(x1);
//Nocadel, 3.7(a,b)
if(f1-f0 >= 0.1*t*df0 || isNaN(f1)) {
tR = t;
t = (tL + tR) * 0.5;
++it;
} else {
var slope = dot(gradient(x1), step);
if (slope <= 0.9 * Math.abs(df0)){
break;
}else if ( slope >= 0.9 * df0) {
tR = t;
t = (tL+ tR) * 0.5;
}else{
tL = t;
t = (tL+ tR)*0.5;
}
}
}
if(t*nstep < tol) { msg = "Line search step size smaller than tol"; break; }
if(it === maxit) { msg = "maxit reached during line search"; break; }
g1 = gradient(x1);
y = sub(g1,g0);
ys = dot(y,s);
Hy = dot(H1,y);
// BFGS update on H1
H1 = sub(add(H1,
mul(
(ys+dot(y,Hy))/(ys*ys),
ten(s,s) )),
div(add(ten(Hy,s),ten(s,Hy)),ys));
x0 = x1;
f0 = f1;
g0 = g1;
++it;
}
return {solution: x0, f: f0, gradient: g0, invHessian: H1, iterations:it, message: msg};
};
var bfgs = function(f,x0,tol,gradient,maxit,callback,options) {
var grad = numeric.gradient;
if(typeof options === "undefined") { options = {}; }
if(typeof tol === "undefined") { tol = 1e-8; }
if(typeof gradient === "undefined") { gradient = function(x) { return grad(f,x); }; }
if(typeof maxit === "undefined") maxit = 1000;
x0 = numeric.clone(x0);
var n = x0.length;
var f0 = f(x0),f1,df0;
if(isNaN(f0)) throw new Error('uncmin: f(x0) is a NaN!');
var max = Math.max, norm2 = numeric.norm2;
tol = max(tol,numeric.epsilon);
var step,g0,g1,H1 = options.Hinv || numeric.identity(n);
var dot = numeric.dot, inv = numeric.inv, sub = numeric.sub, add = numeric.add, ten = numeric.tensor, div = numeric.div, mul = numeric.mul;
var all = numeric.all, isfinite = numeric.isFinite, neg = numeric.neg;
var it=0,i,s,x1,y,Hy,Hs,ys,i0,t,nstep,t1,t2;
var msg = "";
g0 = gradient(x0);
while(it<maxit) {
if(typeof callback === "function") { if(callback(it,x0,f0,g0,H1)) { msg = "Callback returned true"; break; } }
if(!all(isfinite(g0))) { msg = "Gradient has Infinity or NaN"; break; }
step = neg(dot(H1,g0));
if(!all(isfinite(step))) { msg = "Search direction has Infinity or NaN"; break; }
nstep = norm2(step);
if(nstep < tol) { msg="Newton step smaller than tol"; break; }
df0 = dot(g0,step);
// line search
t1 = 0.0;
f1 = f0;
t2 = 1.0;
s = mul(step,t2);
x1 = add(x0,s);
var f2 = f(x1);
var t3 = 2.0;
s = mul(step,t3);
x1 = add(x0,s);
var f3 = f(x1);
var tMax = 1e23;
while( (f2 > f1 || f2 > f3) && it < maxit) {
if(t*nstep < tol) { break; }
if (f2 > f1) {
//If f2 is greater than f1 then we shorten alpha2 and alpha3 closer to f1
//Effectively both are shortened by a factor of two.
t3 = t2;
f3 = f2;
t2 = t2 / 2;
s = mul(step,t2);
x1 = add(x0,s);
f2 = f(x1);
}
else if (f2 > f3) {
if (t3 >= tMax)
break;
//If f2 is greater than f3 then we increase alpha2 and alpha3 away from f1
//Effectively both are lengthened by a factor of two.
t2 = t3;
f2 = f3;
t3 = t3 * 2;
s = mul(step,t3);
x1 = add(x0,s);
f3 = f(x1);
}
it ++;
}
//Get the alpha for the minimum f of the quadratic approximation
var ts = t2 + ((t2-t1)*(f1-f3))/(3*(f1-2*f2+f3));
//Guarantee that the new alphaStar is within the bracket
if (ts >= t3 || ts <= t1)
ts = t2;
if (ts > tMax)
ts = tMax;
if (ts != ts)
ts = 0.;
//Take a final step to alphaStar
s = mul(step,ts);
x1 = add(x0,s);
f1 = f(x1);
if(t*nstep < tol) { msg = "Line search step size smaller than tol"; break; }
if(it === maxit) { msg = "maxit reached during line search"; break; }
g1 = gradient(x1);
y = sub(g1,g0);
ys = dot(y,s);
Hy = dot(H1,y);
// BFGS update on H1
H1 = sub(add(H1,
mul(
(ys+dot(y,Hy))/(ys*ys),
ten(s,s) )),
div(add(ten(Hy,s),ten(s,Hy)),ys));
x0 = x1;
f0 = f1;
g0 = g1;
++it;
}
return {solution: x0, f: f0, gradient: g0, invHessian: H1, iterations:it, message: msg};
};
var bfgs_updater = function(gradient, x0) {
var n = x0.length;
var max = Math.max, norm2 = numeric.norm2;
var g0,g1,H1 = numeric.identity(n);
var dot = numeric.dot, inv = numeric.inv, sub = numeric.sub, add = numeric.add, ten = numeric.tensor, div = numeric.div, mul = numeric.mul;
var all = numeric.all, isfinite = numeric.isFinite, neg = numeric.neg;
var y,Hy,Hs,ys;
var msg = "";
g0 = gradient(x0);
function step() {
return neg(dot(H1,g0));
}
function update(x, real_step) {
var s = real_step;
g1 = gradient(x);
y = sub(g1,g0);
ys = dot(y,s);
Hy = dot(H1,y);
// BFGS update on H1
H1 = sub(add(H1,
mul(
(ys+dot(y,Hy))/(ys*ys),
ten(s,s) )),
div(add(ten(Hy,s),ten(s,Hy)),ys));
g0 = g1;
}
return {step:step, update:update};
};
var inv = function inv(A) {
A = numeric.clone(A);
var s = numeric.dim(A), abs = Math.abs, m = s[0], n = s[1];
var Ai, Aj;
var I = numeric.identity(m), Ii, Ij;
var i,j,k,x;
for(j=0;j<n;++j) {
var i0 = -1;
var v0 = -1;
for(i=j;i!==m;++i) { k = abs(A[i][j]); if(k>v0) { i0 = i; v0 = k; } }
Aj = A[i0]; A[i0] = A[j]; A[j] = Aj;
Ij = I[i0]; I[i0] = I[j]; I[j] = Ij;
x = Aj[j];
if (x === 0) {
console.log("CAN' INVERSE MATRIX");
x = 1e-32
}
for(k=j;k!==n;++k) Aj[k] /= x;
for(k=n-1;k!==-1;--k) Ij[k] /= x;
for(i=m-1;i!==-1;--i) {
if(i!==j) {
Ai = A[i];
Ii = I[i];
x = Ai[j];
for(k=j+1;k!==n;++k) Ai[k] -= Aj[k]*x;
for(k=n-1;k>0;--k) { Ii[k] -= Ij[k]*x; --k; Ii[k] -= Ij[k]*x; }
if(k===0) Ii[0] -= Ij[0]*x;
}
}
}
return I;
};
var _result = function(evalCount, error, returnCode) {
return {
evalCount, error, returnCode,
success: returnCode === SUCCESS
};
};
var dog_leg = function (subsys, rough) {
//rough = true
//var tolg = rough ? 1e-3 : 1e-4;
var tolg, tolf;
if (rough) {
tolg = 1e-3;
tolf = 1e-3;
} else {
tolg = 1e-6;
tolf = 1e-6;
}
var tolx = 1e-80;
var xsize = subsys.params.length;
var csize = subsys.constraints.length;
if (xsize == 0) {
return _result(0, 0, 1);
}
var vec = _vec;
var mx = _matrix;
var n = numeric;
var x = vec(xsize);
var x_new = vec(xsize);
var fx = vec(csize);
var fx_new = vec(csize);
var J = mx(csize, xsize);
var J_new = mx(csize, xsize);
var gn_step = vec(xsize);
var dl_step = vec(xsize);
subsys.fillParams(x);
var err = subsys.calcResidual(fx);
subsys.fillJacobian(J);
function lsolve_slow(A, b) {
var At = n.transpose(A);
var res = n.dot(n.dot(At, inv(n.dot(A, At))), b);
return res;
}
function lsolve(A, b) {
if (csize < xsize) {
var At = n.transpose(A);
var sol = n.solve(n.dot(A, At), b, true);
return n.dot(At, sol);
} else {
return n.solve(A, b, false);
}
}
var g = n.dot(n.transpose(J), fx);
var g_inf = n.norminf(g);
var fx_inf = n.norminf(fx);
var iterLimit = rough ? 1000 : 10000;
var divergenceLimit = 1e6 * (err + 1e6);
var delta = 10;
var alpha = 0.;
var iter = 0, returnCode = 0;
//var log = [];
while (returnCode === 0) {
optim.DEBUG_HANDLER(iter, err);
if (fx_inf <= tolf) {
returnCode = SUCCESS;
} else if (g_inf <= tolg) {
returnCode = SUCCESS;
} else if (iter >= iterLimit) {
returnCode = ITER_LIMIT;
} else if (delta <= tolx * (tolx + n.norm2(x))) {
returnCode = SMALL_DELTA;
} else if (err > divergenceLimit) {
returnCode = DIVERGENCE;
} else if (isNaN(err)) {
returnCode = INVALID_STATE;
}
if (returnCode != 0) {
break;
}
// get the gauss-newton step
//gn_step = n.solve(J, n.mul(fx, -1));
gn_step = lsolve(J, n.mul(fx, -1));
//LU-Decomposition
//gn_step = lusolve(J, n.mul(fx, -1));
//Conjugate gradient method
//gn_step = cg(J, gn_step, n.mul(fx, -1), 1e-8, iterLimit);
//solve linear problem using svd formula to get the gauss-newton step
//gn_step = lls(J, n.mul(fx, -1));
var hitBoundary = false;
var gnorm = n.norm2(g);
var gnNorm = n.norm2(gn_step);
if (gnNorm < delta) {
dl_step = gn_step;
} else {
var Jt = n.transpose(J);
var B = n.dot(Jt, J);
var gBg = n.dot(g, n.dot(B, g));
alpha = n.norm2Squared(g) / gBg;
if (alpha * gnorm >= delta) {
dl_step = n.mul(g, - delta / gnorm);
hitBoundary = true;
} else {
var sd_step = n.mul(g, - alpha);
if (isNaN(gnNorm)) {
dl_step = sd_step;
} else {
var d = n.sub(gn_step, sd_step);
var a = n.dot(d, d);
var b = 2 * n.dot(sd_step, d);
var c = n.dot(sd_step, sd_step) - delta * delta;
var sqrt_discriminant = Math.sqrt(b * b - 4 * a * c);
var beta = (-b + sqrt_discriminant) / (2 * a);
dl_step = n.add(sd_step, n.mul(beta, d));
hitBoundary = true;
}
}
}
var dl_norm = n.norm2(dl_step);
// if (dl_norm <= tolx) {
// returnCode = SMALL_STEP;
// break;
// }
x_new = n.add(x, dl_step);
subsys.setParams(x_new);
var err_new = subsys.calcResidual(fx_new);
subsys.fillJacobian(J_new);
var fxNormSq = n.norm2Squared(fx);
var dF = fxNormSq - n.norm2Squared(fx_new);
var dL = fxNormSq - n.norm2Squared( n.add(fx, n.dot(J, dl_step)) );
var acceptCandidate;
if (dF == 0 || dL == 0) {
acceptCandidate = true;
} else {
var rho = dF / dL;
if (rho < 0.25) {
// if the model is a poor predictor reduce the size of the trust region
delta = 0.25 * dl_norm;
//delta *= 0.5;
} else {
// only increase the size of the trust region if it is taking a step of maximum size
// otherwise just assume it's doing good enough job
if (rho > 0.75 && hitBoundary) {
//delta = Math.max(delta,3*dl_norm);
delta *= 2;
}
}
acceptCandidate = rho > 0; // could be 0 .. 0.25
}
//log.push([stepKind,err, delta,rho]);
if (acceptCandidate) {
x = n.clone(x_new);
J = n.clone(J_new);
fx = n.clone(fx_new);
err = err_new;
g = n.dot(n.transpose(J), fx);
// get infinity norms
g_inf = n.norminf(g);
fx_inf = n.norminf(fx);
}
iter++;
}
//log.push(returnCode);
//window.___log(log);
return _result(iter, err, returnCode);
};
var cg = function(A, x, b, tol, maxIt) {
var _ = numeric;
var tr = _.transpose;
var At = tr(A);
if (A.length != A[0].length) {
A = _.dot(At, A);
b = _.dot(At, b);
}
var r = _.sub(_.dot(A, x), b);
var p = _.mul(r, -1);
var rr = _.dotVV(r, r);
var a;
var _rr;
var beta;
for (var i = 0; i < maxIt; ++i) {
if (_.norm2(r) <= tol) break;
var Axp =_.dot(A, p);
a = rr / _.dotVV(Axp, p);
x = _.add(x, _.mul(p, a));
r = _.add(r, _.mul(Axp, a));
_rr = rr;
rr = _.dotVV(r, r);
beta = rr / _rr;
p = _.add(_.mul(r, -1), _.mul(p, beta));
}
// console.log("liner problem solved in " + i);
return x;
};
var optim = {DEBUG_HANDLER : function() {}}; //backward compatibility
export {dog_leg, optim}