| 1 : |
agomez |
1 |
/* Nonlinear Optimization using the algorithm of Hooke and Jeeves */
|
| 2 : |
|
|
/* 12 February 1994 author: Mark G. Johnson */
|
| 3 : |
ulcessvp |
11 |
//
|
| 4 : |
agomez |
1 |
/* Find a point X where the nonlinear function f(X) has a local */
|
| 5 : |
|
|
/* minimum. X is an n-vector and f(X) is a scalar. In mathe- */
|
| 6 : |
|
|
/* matical notation f: R^n -> R^1. The objective function f() */
|
| 7 : |
|
|
/* is not required to be continuous. Nor does f() need to be */
|
| 8 : |
|
|
/* differentiable. The program does not use or require */
|
| 9 : |
|
|
/* derivatives of f(). */
|
| 10 : |
ulcessvp |
11 |
//
|
| 11 : |
agomez |
1 |
/* The software user supplies three things: a subroutine that */
|
| 12 : |
|
|
/* computes f(X), an initial "starting guess" of the minimum point */
|
| 13 : |
|
|
/* X, and values for the algorithm convergence parameters. Then */
|
| 14 : |
|
|
/* the program searches for a local minimum, beginning from the */
|
| 15 : |
|
|
/* starting guess, using the Direct Search algorithm of Hooke and */
|
| 16 : |
|
|
/* Jeeves. */
|
| 17 : |
ulcessvp |
11 |
//
|
| 18 : |
agomez |
1 |
/* This C program is adapted from the Algol pseudocode found in */
|
| 19 : |
|
|
/* "Algorithm 178: Direct Search" by Arthur F. Kaupe Jr., Commun- */
|
| 20 : |
|
|
/* ications of the ACM, Vol 6. p.313 (June 1963). It includes the */
|
| 21 : |
|
|
/* improvements suggested by Bell and Pike (CACM v.9, p. 684, Sept */
|
| 22 : |
|
|
/* 1966) and those of Tomlin and Smith, "Remark on Algorithm 178" */
|
| 23 : |
|
|
/* (CACM v.12). The original paper, which I don't recommend as */
|
| 24 : |
|
|
/* highly as the one by A. Kaupe, is: R. Hooke and T. A. Jeeves, */
|
| 25 : |
|
|
/* "Direct Search Solution of Numerical and Statistical Problems", */
|
| 26 : |
|
|
/* Journal of the ACM, Vol. 8, April 1961, pp. 212-229. */
|
| 27 : |
ulcessvp |
11 |
//
|
| 28 : |
agomez |
1 |
/* Calling sequence: */
|
| 29 : |
|
|
/* int hooke(nvars, startpt, endpt, rho, epsilon, itermax) */
|
| 30 : |
|
|
/* */
|
| 31 : |
|
|
/* nvars {an integer} */
|
| 32 : |
|
|
/* This is the number of dimensions in the domain of f(). */
|
| 33 : |
|
|
/* It is the number of coordinates of the starting point */
|
| 34 : |
|
|
/* (and the minimum point.) */
|
| 35 : |
|
|
/* startpt {an array of doubles} */
|
| 36 : |
|
|
/* This is the user-supplied guess at the minimum. */
|
| 37 : |
|
|
/* endpt {an array of doubles} */
|
| 38 : |
|
|
/* This is the calculated location of the local minimum */
|
| 39 : |
|
|
/* rho {a double} */
|
| 40 : |
|
|
/* This is a user-supplied convergence parameter (more */
|
| 41 : |
|
|
/* detail below), which should be set to a value between */
|
| 42 : |
|
|
/* 0.0 and 1.0. Larger values of rho give greater */
|
| 43 : |
|
|
/* probability of convergence on highly nonlinear */
|
| 44 : |
|
|
/* functions, at a cost of more function evaluations. */
|
| 45 : |
|
|
/* Smaller values of rho reduces the number of evaluations */
|
| 46 : |
|
|
/* (and the program running time), but increases the risk */
|
| 47 : |
|
|
/* of nonconvergence. See below. */
|
| 48 : |
|
|
/* epsilon {a double} */
|
| 49 : |
|
|
/* This is the criterion for halting the search for a */
|
| 50 : |
|
|
/* minimum. When the algorithm begins to make less and */
|
| 51 : |
|
|
/* less progress on each iteration, it checks the halting */
|
| 52 : |
|
|
/* criterion: if the stepsize is below epsilon, terminate */
|
| 53 : |
|
|
/* the iteration and return the current best estimate of */
|
| 54 : |
|
|
/* the minimum. Larger values of epsilon (such as 1.0e-4) */
|
| 55 : |
|
|
/* give quicker running time, but a less accurate estimate */
|
| 56 : |
|
|
/* of the minimum. Smaller values of epsilon (such as */
|
| 57 : |
|
|
/* 1.0e-7) give longer running time, but a more accurate */
|
| 58 : |
|
|
/* estimate of the minimum. */
|
| 59 : |
|
|
/* itermax {an integer} A second, rarely used, halting */
|
| 60 : |
|
|
/* criterion. If the algorithm uses >= itermax */
|
| 61 : |
|
|
/* iterations, halt. */
|
| 62 : |
ulcessvp |
11 |
//
|
| 63 : |
agomez |
1 |
/* The user-supplied objective function f(x,n) should return a C */
|
| 64 : |
|
|
/* "double". Its arguments are x -- an array of doubles, and */
|
| 65 : |
|
|
/* n -- an integer. x is the point at which f(x) should be */
|
| 66 : |
|
|
/* evaluated, and n is the number of coordinates of x. That is, */
|
| 67 : |
|
|
/* n is the number of coefficients being fitted. */
|
| 68 : |
ulcessvp |
11 |
//
|
| 69 : |
agomez |
1 |
/* rho, the algorithm convergence control */
|
| 70 : |
ulcessvp |
11 |
//
|
| 71 : |
agomez |
1 |
/* The algorithm works by taking "steps" from one estimate of */
|
| 72 : |
|
|
/* a minimum, to another (hopefully better) estimate. Taking */
|
| 73 : |
|
|
/* big steps gets to the minimum more quickly, at the risk of */
|
| 74 : |
|
|
/* "stepping right over" an excellent point. The stepsize is */
|
| 75 : |
|
|
/* controlled by a user supplied parameter called rho. At each */
|
| 76 : |
|
|
/* iteration, the stepsize is multiplied by rho (0 < rho < 1), */
|
| 77 : |
|
|
/* so the stepsize is successively reduced. */
|
| 78 : |
|
|
/* Small values of rho correspond to big stepsize changes, */
|
| 79 : |
|
|
/* which make the algorithm run more quickly. However, there */
|
| 80 : |
|
|
/* is a chance (especially with highly nonlinear functions) */
|
| 81 : |
|
|
/* that these big changes will accidentally overlook a */
|
| 82 : |
|
|
/* promising search vector, leading to nonconvergence. */
|
| 83 : |
|
|
/* Large values of rho correspond to small stepsize changes, */
|
| 84 : |
|
|
/* which force the algorithm to carefully examine nearby points */
|
| 85 : |
|
|
/* instead of optimistically forging ahead. This improves the */
|
| 86 : |
|
|
/* probability of convergence. */
|
| 87 : |
|
|
/* The stepsize is reduced until it is equal to (or smaller */
|
| 88 : |
|
|
/* than) epsilon. So the number of iterations performed by */
|
| 89 : |
|
|
/* Hooke-Jeeves is determined by rho and epsilon: */
|
| 90 : |
|
|
/* rho**(number_of_iterations) = epsilon */
|
| 91 : |
|
|
/* In general it is a good idea to set rho to an aggressively */
|
| 92 : |
|
|
/* small value like 0.5 (hoping for fast convergence). Then, */
|
| 93 : |
|
|
/* if the user suspects that the reported minimum is incorrect */
|
| 94 : |
|
|
/* (or perhaps not accurate enough), the program can be run */
|
| 95 : |
|
|
/* again with a larger value of rho such as 0.85, using the */
|
| 96 : |
|
|
/* result of the first minimization as the starting guess to */
|
| 97 : |
|
|
/* begin the second minimization. */
|
| 98 : |
ulcessvp |
11 |
//
|
| 99 : |
agomez |
1 |
/* Normal use: */
|
| 100 : |
|
|
/* (1) Code your function f() in the C language */
|
| 101 : |
|
|
/* (2) Install your starting guess {or read it in} */
|
| 102 : |
|
|
/* (3) Run the program */
|
| 103 : |
|
|
/* (4) {for the skeptical}: Use the computed minimum */
|
| 104 : |
|
|
/* as the starting point for another run */
|
| 105 : |
ulcessvp |
11 |
//
|
| 106 : |
agomez |
1 |
/* Data Fitting: */
|
| 107 : |
|
|
/* Code your function f() to be the sum of the squares of the */
|
| 108 : |
|
|
/* errors (differences) between the computed values and the */
|
| 109 : |
|
|
/* measured values. Then minimize f() using Hooke-Jeeves. */
|
| 110 : |
|
|
/* EXAMPLE: you have 20 datapoints (ti, yi) and you want to */
|
| 111 : |
|
|
/* find A,B,C such that (A*t*t) + (B*exp(t)) + (C*tan(t)) */
|
| 112 : |
|
|
/* fits the data as closely as possible. Then f() is just */
|
| 113 : |
|
|
/* f(x) = SUM (measured_y[i] - ((A*t[i]*t[i]) + (B*exp(t[i])) */
|
| 114 : |
|
|
/* + (C*tan(t[i]))))^2 */
|
| 115 : |
|
|
/* where x[] is a 3-vector consisting of {A, B, C}. */
|
| 116 : |
ulcessvp |
11 |
//
|
| 117 : |
agomez |
1 |
/* The author of this software is M.G. Johnson. */
|
| 118 : |
|
|
/* Permission to use, copy, modify, and distribute this software */
|
| 119 : |
|
|
/* for any purpose without fee is hereby granted, provided that */
|
| 120 : |
|
|
/* this entire notice is included in all copies of any software */
|
| 121 : |
|
|
/* which is or includes a copy or modification of this software */
|
| 122 : |
|
|
/* and in all copies of the supporting documentation for such */
|
| 123 : |
|
|
/* software. THIS SOFTWARE IS BEING PROVIDED "AS IS", WITHOUT */
|
| 124 : |
|
|
/* ANY EXPRESS OR IMPLIED WARRANTY. IN PARTICULAR, NEITHER THE */
|
| 125 : |
|
|
/* AUTHOR NOR AT&T MAKE ANY REPRESENTATION OR WARRANTY OF ANY */
|
| 126 : |
|
|
/* KIND CONCERNING THE MERCHANTABILITY OF THIS SOFTWARE OR ITS */
|
| 127 : |
|
|
/* FITNESS FOR ANY PARTICULAR PURPOSE. */
|
| 128 : |
ulcessvp |
11 |
//
|
| 129 : |
agomez |
1 |
/* JMB this has been modified to work with the gadget object structure */
|
| 130 : |
|
|
/* This means that the function has been replaced by a call to ecosystem */
|
| 131 : |
|
|
/* object, and we can use the vector objects that have been defined */
|
| 132 : |
|
|
|
| 133 : |
|
|
#include "gadget.h" |
| 134 : |
|
|
#include "optinfo.h" |
| 135 : |
|
|
#include "mathfunc.h" |
| 136 : |
|
|
#include "doublevector.h" |
| 137 : |
|
|
#include "intvector.h" |
| 138 : |
|
|
#include "errorhandler.h" |
| 139 : |
|
|
#include "ecosystem.h" |
| 140 : |
|
|
#include "global.h" |
| 141 : |
|
|
|
| 142 : |
ulcessvp |
11 |
#ifndef NO_OPENMP
|
| 143 : |
|
|
#include "omp.h"
|
| 144 : |
|
|
#endif
|
| 145 : |
|
|
|
| 146 : |
agomez |
1 |
extern Ecosystem* EcoSystem;
|
| 147 : |
ulcessvp |
11 |
#ifndef NO_OPENMP
|
| 148 : |
|
|
extern Ecosystem** EcoSystems;
|
| 149 : |
|
|
#endif
|
| 150 : |
agomez |
1 |
|
| 151 : |
|
|
/* given a point, look for a better one nearby, one coord at a time */
|
| 152 : |
|
|
double OptInfoHooke::bestNearby(DoubleVector& delta, DoubleVector& point, double prevbest, IntVector& param) {
|
| 153 : |
|
|
|
| 154 : |
|
|
double minf, ftmp;
|
| 155 : |
|
|
int i;
|
| 156 : |
|
|
DoubleVector z(point);
|
| 157 : |
|
|
|
| 158 : |
|
|
minf = prevbest;
|
| 159 : |
ulcessvp |
11 |
// for (int k=0;k<point.Size(); k++)
|
| 160 : |
|
|
// cout << z[k] ;
|
| 161 : |
|
|
// cout << endl;
|
| 162 : |
agomez |
1 |
for (i = 0; i < point.Size(); i++) {
|
| 163 : |
ulcessvp |
11 |
|
| 164 : |
|
|
// for (int k=0;k<point.Size(); k++)
|
| 165 : |
|
|
// cout << z[k] << " " ;
|
| 166 : |
|
|
//cout << endl;
|
| 167 : |
agomez |
1 |
z[param[i]] = point[param[i]] + delta[param[i]];
|
| 168 : |
|
|
ftmp = EcoSystem->SimulateAndUpdate(z);
|
| 169 : |
ulcessvp |
11 |
// cout << i <<"-z["<< param[i]<<"]:" <<z[param[i]] << " - " << ftmp << endl;
|
| 170 : |
agomez |
1 |
if (ftmp < minf) {
|
| 171 : |
|
|
minf = ftmp;
|
| 172 : |
|
|
} else {
|
| 173 : |
|
|
delta[param[i]] = 0.0 - delta[param[i]];
|
| 174 : |
|
|
z[param[i]] = point[param[i]] + delta[param[i]];
|
| 175 : |
|
|
ftmp = EcoSystem->SimulateAndUpdate(z);
|
| 176 : |
|
|
if (ftmp < minf)
|
| 177 : |
|
|
minf = ftmp;
|
| 178 : |
|
|
else
|
| 179 : |
|
|
z[param[i]] = point[param[i]];
|
| 180 : |
|
|
}
|
| 181 : |
ulcessvp |
11 |
// cout << i <<"-z["<< param[i]<<"]:" <<z[param[i]] << " - " << ftmp <<" - " << prevbest << endl;
|
| 182 : |
agomez |
1 |
}
|
| 183 : |
|
|
|
| 184 : |
|
|
for (i = 0; i < point.Size(); i++)
|
| 185 : |
|
|
point[i] = z[i];
|
| 186 : |
|
|
return minf;
|
| 187 : |
|
|
}
|
| 188 : |
|
|
|
| 189 : |
ulcessvp |
11 |
/* given a point, look for a better one nearby, one coord at a time */
|
| 190 : |
|
|
#ifndef NO_OPENMP
|
| 191 : |
ulcessvp |
12 |
/*
|
| 192 : |
|
|
* function bestBeraby parallelized with OpenMP
|
| 193 : |
|
|
* · 2 threads per coord to parallelize the calculation of +delta/-delta
|
| 194 : |
|
|
* · parallelize the calculation of the best nearby of the coord
|
| 195 : |
|
|
*/
|
| 196 : |
ulcessvp |
11 |
double OptInfoHooke::bestNearbyOMP(DoubleVector& delta, DoubleVector& point, double prevbest, IntVector& param) {
|
| 197 : |
|
|
double minf;//, ftmp;
|
| 198 : |
|
|
int i, j, k;
|
| 199 : |
|
|
DoubleVector z(point);
|
| 200 : |
|
|
|
| 201 : |
|
|
struct Storage {
|
| 202 : |
|
|
DoubleVector z;
|
| 203 : |
|
|
DoubleVector delta;
|
| 204 : |
|
|
double ftmp;
|
| 205 : |
|
|
int iters;
|
| 206 : |
|
|
};
|
| 207 : |
|
|
|
| 208 : |
|
|
minf = prevbest;
|
| 209 : |
|
|
i = 0;
|
| 210 : |
|
|
|
| 211 : |
|
|
int paral_tokens, numThr, nvars = point.Size();
|
| 212 : |
|
|
numThr = omp_get_max_threads ( );
|
| 213 : |
|
|
|
| 214 : |
|
|
Storage* storage = new Storage[numThr];
|
| 215 : |
|
|
if ((numThr % 2) == 0)
|
| 216 : |
|
|
paral_tokens = numThr / 2;
|
| 217 : |
|
|
else {
|
| 218 : |
|
|
return -1;
|
| 219 : |
|
|
}
|
| 220 : |
|
|
|
| 221 : |
|
|
while ( i < nvars) {
|
| 222 : |
|
|
if ((i + paral_tokens -1) >= nvars)
|
| 223 : |
|
|
paral_tokens = nvars - i;
|
| 224 : |
|
|
omp_set_dynamic(0);
|
| 225 : |
|
|
omp_set_nested(1);
|
| 226 : |
|
|
#pragma omp parallel for num_threads(paral_tokens) private(k)
|
| 227 : |
|
|
for (j = 0; j < paral_tokens; ++j) {
|
| 228 : |
|
|
storage[j].z = z;
|
| 229 : |
|
|
storage[j].delta = delta;
|
| 230 : |
|
|
DoubleVector v1(z);
|
| 231 : |
|
|
DoubleVector v2(z);
|
| 232 : |
|
|
k = param[i+j];
|
| 233 : |
|
|
v1[k] += delta[k];
|
| 234 : |
|
|
v2[k] -= delta[k];
|
| 235 : |
|
|
|
| 236 : |
|
|
#pragma omp parallel sections num_threads(2)
|
| 237 : |
|
|
{
|
| 238 : |
ulcessvp |
12 |
#pragma omp section
|
| 239 : |
ulcessvp |
11 |
{
|
| 240 : |
|
|
storage[j].ftmp = EcoSystems[j]->SimulateAndUpdate(v1);
|
| 241 : |
|
|
}
|
| 242 : |
ulcessvp |
12 |
#pragma omp section
|
| 243 : |
ulcessvp |
11 |
{
|
| 244 : |
|
|
storage[j+paral_tokens].ftmp = EcoSystems[j+paral_tokens]->SimulateAndUpdate(v2);
|
| 245 : |
|
|
}
|
| 246 : |
|
|
}
|
| 247 : |
|
|
|
| 248 : |
|
|
if (storage[j].ftmp < minf) {
|
| 249 : |
|
|
storage[j].iters = 1;
|
| 250 : |
|
|
storage[j].z[k] = v1[k];
|
| 251 : |
|
|
} else {
|
| 252 : |
|
|
storage[j].iters = 2;
|
| 253 : |
|
|
storage[j].delta[k] = 0.0 - delta[k];
|
| 254 : |
|
|
if (storage[j+paral_tokens].ftmp < minf) {
|
| 255 : |
|
|
storage[j].ftmp = storage[j+paral_tokens].ftmp;
|
| 256 : |
|
|
storage[j].z[k] = v2[k];
|
| 257 : |
|
|
}
|
| 258 : |
|
|
}
|
| 259 : |
|
|
}
|
| 260 : |
|
|
|
| 261 : |
|
|
for (j = 0; j < paral_tokens; ++j) {
|
| 262 : |
|
|
i++;
|
| 263 : |
|
|
iters += storage[j].iters;
|
| 264 : |
|
|
if (storage[j].ftmp < minf) {
|
| 265 : |
|
|
minf = storage[j].ftmp;
|
| 266 : |
|
|
z = storage[j].z;
|
| 267 : |
|
|
delta = storage[j].delta;
|
| 268 : |
|
|
break;
|
| 269 : |
|
|
}
|
| 270 : |
|
|
}
|
| 271 : |
|
|
}
|
| 272 : |
|
|
|
| 273 : |
|
|
for (i = 0; i < nvars; ++i)
|
| 274 : |
|
|
point[i] = z[i];
|
| 275 : |
|
|
return minf;
|
| 276 : |
|
|
}
|
| 277 : |
|
|
#endif
|
| 278 : |
|
|
|
| 279 : |
agomez |
1 |
void OptInfoHooke::OptimiseLikelihood() {
|
| 280 : |
|
|
|
| 281 : |
|
|
double oldf, newf, bestf, steplength, tmp;
|
| 282 : |
|
|
int i, offset;
|
| 283 : |
|
|
int rchange, rcheck, rnumber; //Used to randomise the order of the parameters
|
| 284 : |
|
|
|
| 285 : |
|
|
handle.logMessage(LOGINFO, "\nStarting Hooke & Jeeves optimisation algorithm\n");
|
| 286 : |
|
|
int nvars = EcoSystem->numOptVariables();
|
| 287 : |
|
|
DoubleVector x(nvars);
|
| 288 : |
|
|
DoubleVector trialx(nvars);
|
| 289 : |
|
|
DoubleVector bestx(nvars);
|
| 290 : |
|
|
DoubleVector lowerb(nvars);
|
| 291 : |
|
|
DoubleVector upperb(nvars);
|
| 292 : |
|
|
DoubleVector init(nvars);
|
| 293 : |
|
|
DoubleVector initialstep(nvars, rho);
|
| 294 : |
|
|
DoubleVector delta(nvars);
|
| 295 : |
|
|
IntVector param(nvars, 0);
|
| 296 : |
|
|
IntVector lbound(nvars, 0);
|
| 297 : |
|
|
IntVector rbounds(nvars, 0);
|
| 298 : |
|
|
IntVector trapped(nvars, 0);
|
| 299 : |
|
|
|
| 300 : |
|
|
EcoSystem->scaleVariables();
|
| 301 : |
ulcessvp |
11 |
#ifndef NO_OPENMP
|
| 302 : |
|
|
int numThr = omp_get_max_threads ( );
|
| 303 : |
|
|
for (i = 0; i < numThr; i++)
|
| 304 : |
|
|
EcoSystems[i]->scaleVariables();
|
| 305 : |
|
|
#endif
|
| 306 : |
agomez |
1 |
EcoSystem->getOptScaledValues(x);
|
| 307 : |
|
|
EcoSystem->getOptLowerBounds(lowerb);
|
| 308 : |
|
|
EcoSystem->getOptUpperBounds(upperb);
|
| 309 : |
|
|
EcoSystem->getOptInitialValues(init);
|
| 310 : |
|
|
|
| 311 : |
|
|
for (i = 0; i < nvars; i++) {
|
| 312 : |
|
|
// Scaling the bounds, because the parameters are scaled
|
| 313 : |
|
|
lowerb[i] = lowerb[i] / init[i];
|
| 314 : |
|
|
upperb[i] = upperb[i] / init[i];
|
| 315 : |
|
|
if (lowerb[i] > upperb[i]) {
|
| 316 : |
|
|
tmp = lowerb[i];
|
| 317 : |
|
|
lowerb[i] = upperb[i];
|
| 318 : |
|
|
upperb[i] = tmp;
|
| 319 : |
|
|
}
|
| 320 : |
|
|
|
| 321 : |
|
|
bestx[i] = x[i];
|
| 322 : |
|
|
trialx[i] = x[i];
|
| 323 : |
|
|
param[i] = i;
|
| 324 : |
|
|
delta[i] = ((2 * (rand() % 2)) - 1) * rho; //JMB - randomise the sign
|
| 325 : |
|
|
}
|
| 326 : |
|
|
|
| 327 : |
|
|
bestf = EcoSystem->SimulateAndUpdate(trialx);
|
| 328 : |
|
|
if (bestf != bestf) { //check for NaN
|
| 329 : |
|
|
handle.logMessage(LOGINFO, "Error starting Hooke & Jeeves optimisation with f(x) = infinity");
|
| 330 : |
|
|
converge = -1;
|
| 331 : |
|
|
iters = 1;
|
| 332 : |
|
|
return;
|
| 333 : |
|
|
}
|
| 334 : |
|
|
|
| 335 : |
|
|
offset = EcoSystem->getFuncEval(); //number of function evaluations done before loop
|
| 336 : |
|
|
newf = bestf;
|
| 337 : |
|
|
oldf = bestf;
|
| 338 : |
|
|
steplength = lambda;
|
| 339 : |
|
|
if (isZero(steplength))
|
| 340 : |
|
|
steplength = rho;
|
| 341 : |
|
|
|
| 342 : |
ulcessvp |
11 |
iters = 0;
|
| 343 : |
|
|
|
| 344 : |
agomez |
1 |
while (1) {
|
| 345 : |
|
|
if (isZero(bestf)) {
|
| 346 : |
ulcessvp |
11 |
#ifdef NO_OPENMP
|
| 347 : |
agomez |
1 |
iters = EcoSystem->getFuncEval() - offset;
|
| 348 : |
ulcessvp |
11 |
#endif
|
| 349 : |
agomez |
1 |
handle.logMessage(LOGINFO, "Error in Hooke & Jeeves optimisation after", iters, "function evaluations, f(x) = 0");
|
| 350 : |
|
|
converge = -1;
|
| 351 : |
|
|
return;
|
| 352 : |
|
|
}
|
| 353 : |
|
|
|
| 354 : |
|
|
/* randomize the order of the parameters once in a while */
|
| 355 : |
|
|
rchange = 0;
|
| 356 : |
|
|
while (rchange < nvars) {
|
| 357 : |
|
|
rnumber = rand() % nvars;
|
| 358 : |
|
|
rcheck = 1;
|
| 359 : |
|
|
for (i = 0; i < rchange; i++)
|
| 360 : |
|
|
if (param[i] == rnumber)
|
| 361 : |
|
|
rcheck = 0;
|
| 362 : |
|
|
if (rcheck) {
|
| 363 : |
|
|
param[rchange] = rnumber;
|
| 364 : |
|
|
rchange++;
|
| 365 : |
|
|
}
|
| 366 : |
|
|
}
|
| 367 : |
|
|
|
| 368 : |
|
|
/* find best new point, one coord at a time */
|
| 369 : |
|
|
for (i = 0; i < nvars; i++)
|
| 370 : |
|
|
trialx[i] = x[i];
|
| 371 : |
ulcessvp |
11 |
#ifndef NO_OPENMP
|
| 372 : |
|
|
newf = this->bestNearbyOMP(delta, trialx, bestf, param);
|
| 373 : |
|
|
if (newf == -1) {
|
| 374 : |
|
|
handle.logMessage(LOGINFO, "\nStopping Hooke & Jeeves optimisation algorithm\n");
|
| 375 : |
|
|
handle.logMessage(LOGINFO, "\nThe number of threads must be a multiple of 2\n");
|
| 376 : |
|
|
return;
|
| 377 : |
|
|
}
|
| 378 : |
|
|
#else
|
| 379 : |
agomez |
1 |
newf = this->bestNearby(delta, trialx, bestf, param);
|
| 380 : |
ulcessvp |
11 |
#endif
|
| 381 : |
|
|
/* if too many function evaluations occur, terminate the algorithm */
|
| 382 : |
agomez |
1 |
|
| 383 : |
ulcessvp |
11 |
#ifdef NO_OPENMP
|
| 384 : |
agomez |
1 |
iters = EcoSystem->getFuncEval() - offset;
|
| 385 : |
ulcessvp |
11 |
#endif
|
| 386 : |
agomez |
1 |
if (iters > hookeiter) {
|
| 387 : |
|
|
handle.logMessage(LOGINFO, "\nStopping Hooke & Jeeves optimisation algorithm\n");
|
| 388 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped after", iters, "function evaluations");
|
| 389 : |
|
|
handle.logMessage(LOGINFO, "The steplength was reduced to", steplength);
|
| 390 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped because the maximum number of function evaluations");
|
| 391 : |
|
|
handle.logMessage(LOGINFO, "was reached and NOT because an optimum was found for this run");
|
| 392 : |
|
|
|
| 393 : |
|
|
score = EcoSystem->SimulateAndUpdate(trialx);
|
| 394 : |
|
|
handle.logMessage(LOGINFO, "\nHooke & Jeeves finished with a likelihood score of", score);
|
| 395 : |
|
|
for (i = 0; i < nvars; i++)
|
| 396 : |
|
|
bestx[i] = trialx[i] * init[i];
|
| 397 : |
|
|
EcoSystem->storeVariables(score, bestx);
|
| 398 : |
|
|
return;
|
| 399 : |
|
|
}
|
| 400 : |
|
|
|
| 401 : |
|
|
/* if we made some improvements, pursue that direction */
|
| 402 : |
|
|
while (newf < bestf) {
|
| 403 : |
|
|
for (i = 0; i < nvars; i++) {
|
| 404 : |
|
|
/* if it has been trapped but f has now gotten better (bndcheck) */
|
| 405 : |
|
|
/* we assume that we are out of the trap, reset the counters */
|
| 406 : |
|
|
/* and go back to the stepsize we had when we got trapped */
|
| 407 : |
|
|
if ((trapped[i]) && (newf < oldf * bndcheck)) {
|
| 408 : |
|
|
trapped[i] = 0;
|
| 409 : |
|
|
lbound[i] = 0;
|
| 410 : |
|
|
rbounds[i] = 0;
|
| 411 : |
|
|
delta[i] = initialstep[i];
|
| 412 : |
|
|
|
| 413 : |
|
|
} else if (trialx[i] < (lowerb[i] + verysmall)) {
|
| 414 : |
|
|
lbound[i]++;
|
| 415 : |
|
|
trialx[i] = lowerb[i];
|
| 416 : |
|
|
if (!trapped[i]) {
|
| 417 : |
|
|
initialstep[i] = delta[i];
|
| 418 : |
|
|
trapped[i] = 1;
|
| 419 : |
|
|
}
|
| 420 : |
|
|
/* if it has hit the bounds 2 times then increase the stepsize */
|
| 421 : |
|
|
if (lbound[i] >= 2)
|
| 422 : |
|
|
delta[i] /= rho;
|
| 423 : |
|
|
|
| 424 : |
|
|
} else if (trialx[i] > (upperb[i] - verysmall)) {
|
| 425 : |
|
|
rbounds[i]++;
|
| 426 : |
|
|
trialx[i] = upperb[i];
|
| 427 : |
|
|
if (!trapped[i]) {
|
| 428 : |
|
|
initialstep[i] = delta[i];
|
| 429 : |
|
|
trapped[i] = 1;
|
| 430 : |
|
|
}
|
| 431 : |
|
|
/* if it has hit the bounds 2 times then increase the stepsize */
|
| 432 : |
|
|
if (rbounds[i] >= 2)
|
| 433 : |
|
|
delta[i] /= rho;
|
| 434 : |
|
|
}
|
| 435 : |
|
|
}
|
| 436 : |
|
|
|
| 437 : |
|
|
for (i = 0; i < nvars; i++) {
|
| 438 : |
|
|
/* firstly, arrange the sign of delta[] */
|
| 439 : |
|
|
if (trialx[i] < x[i])
|
| 440 : |
|
|
delta[i] = 0.0 - fabs(delta[i]);
|
| 441 : |
|
|
else
|
| 442 : |
|
|
delta[i] = fabs(delta[i]);
|
| 443 : |
|
|
|
| 444 : |
|
|
/* now, move further in this direction */
|
| 445 : |
|
|
tmp = x[i];
|
| 446 : |
|
|
x[i] = trialx[i];
|
| 447 : |
|
|
trialx[i] = trialx[i] + trialx[i] - tmp;
|
| 448 : |
|
|
}
|
| 449 : |
|
|
|
| 450 : |
|
|
/* only move forward if this is really an improvement */
|
| 451 : |
|
|
oldf = newf;
|
| 452 : |
|
|
newf = EcoSystem->SimulateAndUpdate(trialx);
|
| 453 : |
ulcessvp |
11 |
#ifndef NO_OPENMP
|
| 454 : |
|
|
iters++;
|
| 455 : |
|
|
#endif
|
| 456 : |
agomez |
1 |
if ((isEqual(newf, oldf)) || (newf > oldf)) {
|
| 457 : |
|
|
newf = oldf; //JMB no improvement, so reset the value of newf
|
| 458 : |
|
|
break;
|
| 459 : |
|
|
}
|
| 460 : |
|
|
|
| 461 : |
|
|
/* OK, it's better, so update variables and look around */
|
| 462 : |
|
|
bestf = newf;
|
| 463 : |
|
|
for (i = 0; i < nvars; i++)
|
| 464 : |
|
|
x[i] = trialx[i];
|
| 465 : |
ulcessvp |
11 |
|
| 466 : |
|
|
#ifndef NO_OPENMP
|
| 467 : |
|
|
newf = this->bestNearbyOMP(delta, trialx, bestf, param);
|
| 468 : |
|
|
if (newf == -1) {
|
| 469 : |
|
|
handle.logMessage(LOGINFO, "\nStopping Hooke & Jeeves optimisation algorithm\n");
|
| 470 : |
|
|
handle.logMessage(LOGINFO, "\nThe number of threads must be a multiple of 2\n");
|
| 471 : |
|
|
return;
|
| 472 : |
|
|
}
|
| 473 : |
|
|
#else
|
| 474 : |
agomez |
1 |
newf = this->bestNearby(delta, trialx, bestf, param);
|
| 475 : |
ulcessvp |
11 |
#endif
|
| 476 : |
agomez |
1 |
if (isEqual(newf, bestf))
|
| 477 : |
|
|
break;
|
| 478 : |
|
|
|
| 479 : |
|
|
/* if too many function evaluations occur, terminate the algorithm */
|
| 480 : |
ulcessvp |
11 |
#ifdef NO_OPENMP
|
| 481 : |
agomez |
1 |
iters = EcoSystem->getFuncEval() - offset;
|
| 482 : |
ulcessvp |
11 |
#endif
|
| 483 : |
agomez |
1 |
if (iters > hookeiter) {
|
| 484 : |
|
|
handle.logMessage(LOGINFO, "\nStopping Hooke & Jeeves optimisation algorithm\n");
|
| 485 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped after", iters, "function evaluations");
|
| 486 : |
|
|
handle.logMessage(LOGINFO, "The steplength was reduced to", steplength);
|
| 487 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped because the maximum number of function evaluations");
|
| 488 : |
|
|
handle.logMessage(LOGINFO, "was reached and NOT because an optimum was found for this run");
|
| 489 : |
|
|
|
| 490 : |
|
|
score = EcoSystem->SimulateAndUpdate(trialx);
|
| 491 : |
|
|
handle.logMessage(LOGINFO, "\nHooke & Jeeves finished with a likelihood score of", score);
|
| 492 : |
|
|
for (i = 0; i < nvars; i++)
|
| 493 : |
|
|
bestx[i] = trialx[i] * init[i];
|
| 494 : |
|
|
EcoSystem->storeVariables(score, bestx);
|
| 495 : |
|
|
return;
|
| 496 : |
|
|
}
|
| 497 : |
|
|
}
|
| 498 : |
|
|
|
| 499 : |
ulcessvp |
11 |
#ifdef NO_OPENMP
|
| 500 : |
agomez |
1 |
iters = EcoSystem->getFuncEval() - offset;
|
| 501 : |
ulcessvp |
11 |
#endif
|
| 502 : |
agomez |
1 |
if (newf < bestf) {
|
| 503 : |
|
|
for (i = 0; i < nvars; i++)
|
| 504 : |
|
|
bestx[i] = x[i] * init[i];
|
| 505 : |
|
|
bestf = newf;
|
| 506 : |
|
|
handle.logMessage(LOGINFO, "\nNew optimum found after", iters, "function evaluations");
|
| 507 : |
|
|
handle.logMessage(LOGINFO, "The likelihood score is", bestf, "at the point");
|
| 508 : |
|
|
EcoSystem->storeVariables(bestf, bestx);
|
| 509 : |
|
|
EcoSystem->writeBestValues();
|
| 510 : |
|
|
|
| 511 : |
|
|
} else
|
| 512 : |
|
|
handle.logMessage(LOGINFO, "Checking convergence criteria after", iters, "function evaluations ...");
|
| 513 : |
|
|
|
| 514 : |
|
|
/* if the step length is less than hookeeps, terminate the algorithm */
|
| 515 : |
|
|
if (steplength < hookeeps) {
|
| 516 : |
|
|
handle.logMessage(LOGINFO, "\nStopping Hooke & Jeeves optimisation algorithm\n");
|
| 517 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped after", iters, "function evaluations");
|
| 518 : |
|
|
handle.logMessage(LOGINFO, "The steplength was reduced to", steplength);
|
| 519 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped because an optimum was found for this run");
|
| 520 : |
|
|
|
| 521 : |
|
|
converge = 1;
|
| 522 : |
|
|
score = bestf;
|
| 523 : |
|
|
handle.logMessage(LOGINFO, "\nHooke & Jeeves finished with a likelihood score of", score);
|
| 524 : |
|
|
EcoSystem->storeVariables(bestf, bestx);
|
| 525 : |
|
|
return;
|
| 526 : |
|
|
}
|
| 527 : |
|
|
|
| 528 : |
|
|
steplength *= rho;
|
| 529 : |
|
|
handle.logMessage(LOGINFO, "Reducing the steplength to", steplength);
|
| 530 : |
|
|
for (i = 0; i < nvars; i++)
|
| 531 : |
|
|
delta[i] *= rho;
|
| 532 : |
|
|
}
|
| 533 : |
|
|
}
|
| 534 : |
ulcessvp |
11 |
|
| 535 : |
ulcessvp |
12 |
/* Functions to perform the parallelization of the algorithm of HJ with OpenMP*/
|
| 536 : |
ulcessvp |
11 |
#ifdef GADGET_OPENMP
|
| 537 : |
|
|
double OptInfoHooke::bestNearbyOMP2(DoubleVector& delta, DoubleVector& point, double prevbest, IntVector& param) {
|
| 538 : |
ulcessvp |
12 |
double minf;
|
| 539 : |
ulcessvp |
11 |
int i, j, k, ii;
|
| 540 : |
|
|
DoubleVector z(point);
|
| 541 : |
|
|
int bestId = 0;
|
| 542 : |
|
|
struct Storage {
|
| 543 : |
|
|
DoubleVector z;
|
| 544 : |
|
|
DoubleVector delta;
|
| 545 : |
|
|
double ftmp;
|
| 546 : |
|
|
int iters;
|
| 547 : |
|
|
};
|
| 548 : |
|
|
|
| 549 : |
|
|
minf = prevbest;
|
| 550 : |
|
|
|
| 551 : |
|
|
int paral_tokens, numThr, nvars = point.Size();
|
| 552 : |
|
|
numThr = omp_get_max_threads ( );
|
| 553 : |
|
|
|
| 554 : |
|
|
Storage* storage = new Storage[numThr];
|
| 555 : |
|
|
if ((numThr % 2) == 0)
|
| 556 : |
|
|
paral_tokens = numThr / 2;
|
| 557 : |
|
|
else {
|
| 558 : |
|
|
return -1;
|
| 559 : |
|
|
}
|
| 560 : |
|
|
|
| 561 : |
|
|
for (ii=0; ii< paral_tokens; ii++) {
|
| 562 : |
|
|
i = 0;
|
| 563 : |
|
|
while ( i < nvars) {
|
| 564 : |
|
|
if ((i + paral_tokens -1) >= nvars)
|
| 565 : |
|
|
paral_tokens = nvars - i;
|
| 566 : |
|
|
omp_set_dynamic(0);
|
| 567 : |
|
|
omp_set_nested(1);
|
| 568 : |
|
|
#pragma omp parallel for num_threads(paral_tokens) private(k)
|
| 569 : |
|
|
for (j = 0; j < paral_tokens; ++j) {
|
| 570 : |
|
|
storage[j].z = z;
|
| 571 : |
|
|
storage[j].delta = delta;
|
| 572 : |
|
|
DoubleVector v1(z);
|
| 573 : |
|
|
DoubleVector v2(z);
|
| 574 : |
|
|
k = param[i+j];
|
| 575 : |
|
|
v1[k] += delta[k];
|
| 576 : |
|
|
v2[k] -= delta[k];
|
| 577 : |
|
|
|
| 578 : |
ulcessvp |
12 |
#pragma omp parallel sections num_threads(2)
|
| 579 : |
ulcessvp |
11 |
{
|
| 580 : |
|
|
#pragma omp section
|
| 581 : |
|
|
{
|
| 582 : |
|
|
storage[j].ftmp = EcoSystems[j]->SimulateAndUpdate(v1);
|
| 583 : |
|
|
}
|
| 584 : |
|
|
#pragma omp section
|
| 585 : |
|
|
{
|
| 586 : |
|
|
storage[j+paral_tokens].ftmp = EcoSystems[j+paral_tokens]->SimulateAndUpdate(v2);
|
| 587 : |
|
|
}
|
| 588 : |
|
|
}
|
| 589 : |
|
|
if (storage[j].ftmp < minf) {
|
| 590 : |
|
|
storage[j].iters = 1;
|
| 591 : |
|
|
storage[j].z[k] = v1[k];
|
| 592 : |
|
|
} else {
|
| 593 : |
|
|
storage[j].iters = 2;
|
| 594 : |
|
|
storage[j].delta[k] = 0.0 - delta[k];
|
| 595 : |
|
|
if (storage[j+paral_tokens].ftmp < minf) {
|
| 596 : |
|
|
storage[j].ftmp = storage[j+paral_tokens].ftmp;
|
| 597 : |
|
|
storage[j].z[k] = v2[k];
|
| 598 : |
|
|
}
|
| 599 : |
|
|
else iters += 2;
|
| 600 : |
|
|
}
|
| 601 : |
|
|
}
|
| 602 : |
|
|
|
| 603 : |
|
|
bestId = 0;
|
| 604 : |
|
|
for (j = 1; j < paral_tokens; ++j) {
|
| 605 : |
|
|
if (storage[j].ftmp < storage[bestId].ftmp)
|
| 606 : |
|
|
bestId = j;
|
| 607 : |
|
|
}
|
| 608 : |
|
|
if (storage[bestId].ftmp < minf) {
|
| 609 : |
|
|
iters += storage[bestId].iters;
|
| 610 : |
|
|
minf = storage[bestId].ftmp;
|
| 611 : |
|
|
z = storage[bestId].z;
|
| 612 : |
|
|
delta = storage[bestId].delta;
|
| 613 : |
|
|
}
|
| 614 : |
|
|
|
| 615 : |
|
|
i += paral_tokens;
|
| 616 : |
|
|
}
|
| 617 : |
|
|
}
|
| 618 : |
|
|
|
| 619 : |
|
|
delete[] storage;
|
| 620 : |
|
|
for (i = 0; i < nvars; ++i)
|
| 621 : |
|
|
point[i] = z[i];
|
| 622 : |
|
|
|
| 623 : |
|
|
return minf;
|
| 624 : |
|
|
}
|
| 625 : |
|
|
|
| 626 : |
|
|
void OptInfoHooke::OptimiseLikelihoodOMP() {
|
| 627 : |
|
|
double oldf, newf, bestf, steplength, tmp;
|
| 628 : |
|
|
int i, offset;
|
| 629 : |
|
|
int rchange, rcheck, rnumber; //Used to randomise the order of the parameters
|
| 630 : |
|
|
|
| 631 : |
|
|
handle.logMessage(LOGINFO, "\nStarting Hooke & Jeeves optimisation algorithm\n");
|
| 632 : |
|
|
int nvars = EcoSystem->numOptVariables();
|
| 633 : |
|
|
DoubleVector x(nvars);
|
| 634 : |
|
|
DoubleVector trialx(nvars);
|
| 635 : |
|
|
DoubleVector bestx(nvars);
|
| 636 : |
|
|
DoubleVector lowerb(nvars);
|
| 637 : |
|
|
DoubleVector upperb(nvars);
|
| 638 : |
|
|
DoubleVector init(nvars);
|
| 639 : |
|
|
DoubleVector initialstep(nvars, rho);
|
| 640 : |
|
|
DoubleVector delta(nvars);
|
| 641 : |
|
|
IntVector param(nvars, 0);
|
| 642 : |
|
|
IntVector lbound(nvars, 0);
|
| 643 : |
|
|
IntVector rbounds(nvars, 0);
|
| 644 : |
|
|
IntVector trapped(nvars, 0);
|
| 645 : |
|
|
|
| 646 : |
|
|
EcoSystem->scaleVariables();
|
| 647 : |
|
|
#ifndef NO_OPENMP
|
| 648 : |
|
|
int numThr = omp_get_max_threads ( );
|
| 649 : |
|
|
for (i = 0; i < numThr; i++)
|
| 650 : |
|
|
EcoSystems[i]->scaleVariables();
|
| 651 : |
|
|
#endif
|
| 652 : |
|
|
EcoSystem->getOptScaledValues(x);
|
| 653 : |
|
|
EcoSystem->getOptLowerBounds(lowerb);
|
| 654 : |
|
|
EcoSystem->getOptUpperBounds(upperb);
|
| 655 : |
|
|
EcoSystem->getOptInitialValues(init);
|
| 656 : |
|
|
|
| 657 : |
|
|
for (i = 0; i < nvars; i++) {
|
| 658 : |
|
|
// Scaling the bounds, because the parameters are scaled
|
| 659 : |
|
|
lowerb[i] = lowerb[i] / init[i];
|
| 660 : |
|
|
upperb[i] = upperb[i] / init[i];
|
| 661 : |
|
|
if (lowerb[i] > upperb[i]) {
|
| 662 : |
|
|
tmp = lowerb[i];
|
| 663 : |
|
|
lowerb[i] = upperb[i];
|
| 664 : |
|
|
upperb[i] = tmp;
|
| 665 : |
|
|
}
|
| 666 : |
|
|
|
| 667 : |
|
|
bestx[i] = x[i];
|
| 668 : |
|
|
trialx[i] = x[i];
|
| 669 : |
|
|
param[i] = i;
|
| 670 : |
|
|
delta[i] = ((2 * (rand() % 2)) - 1) * rho; //JMB - randomise the sign
|
| 671 : |
|
|
}
|
| 672 : |
|
|
|
| 673 : |
|
|
bestf = EcoSystem->SimulateAndUpdate(trialx);
|
| 674 : |
|
|
if (bestf != bestf) { //check for NaN
|
| 675 : |
|
|
handle.logMessage(LOGINFO, "Error starting Hooke & Jeeves optimisation with f(x) = infinity");
|
| 676 : |
|
|
converge = -1;
|
| 677 : |
|
|
iters = 1;
|
| 678 : |
|
|
return;
|
| 679 : |
|
|
}
|
| 680 : |
|
|
|
| 681 : |
|
|
offset = EcoSystem->getFuncEval(); //number of function evaluations done before loop
|
| 682 : |
|
|
newf = bestf;
|
| 683 : |
|
|
oldf = bestf;
|
| 684 : |
|
|
steplength = lambda;
|
| 685 : |
|
|
if (isZero(steplength))
|
| 686 : |
|
|
steplength = rho;
|
| 687 : |
|
|
|
| 688 : |
|
|
iters = 0;
|
| 689 : |
|
|
|
| 690 : |
|
|
while (1) {
|
| 691 : |
|
|
if (isZero(bestf)) {
|
| 692 : |
|
|
#ifdef NO_OPENMP
|
| 693 : |
|
|
iters = EcoSystem->getFuncEval() - offset;
|
| 694 : |
|
|
#endif
|
| 695 : |
|
|
handle.logMessage(LOGINFO, "Error in Hooke & Jeeves optimisation after", iters, "function evaluations, f(x) = 0");
|
| 696 : |
|
|
converge = -1;
|
| 697 : |
|
|
return;
|
| 698 : |
|
|
}
|
| 699 : |
|
|
|
| 700 : |
|
|
/* randomize the order of the parameters once in a while */
|
| 701 : |
|
|
rchange = 0;
|
| 702 : |
|
|
while (rchange < nvars) {
|
| 703 : |
|
|
rnumber = rand() % nvars;
|
| 704 : |
|
|
rcheck = 1;
|
| 705 : |
|
|
for (i = 0; i < rchange; i++)
|
| 706 : |
|
|
if (param[i] == rnumber)
|
| 707 : |
|
|
rcheck = 0;
|
| 708 : |
|
|
if (rcheck) {
|
| 709 : |
|
|
param[rchange] = rnumber;
|
| 710 : |
|
|
rchange++;
|
| 711 : |
|
|
}
|
| 712 : |
|
|
}
|
| 713 : |
|
|
|
| 714 : |
|
|
/* find best new point, one coord at a time */
|
| 715 : |
|
|
for (i = 0; i < nvars; i++)
|
| 716 : |
|
|
trialx[i] = x[i];
|
| 717 : |
|
|
#ifndef NO_OPENMP
|
| 718 : |
|
|
newf = this->bestNearbyOMP2(delta, trialx, bestf, param);
|
| 719 : |
|
|
if (newf == -1) {
|
| 720 : |
|
|
handle.logMessage(LOGINFO, "\nStopping Hooke & Jeeves optimisation algorithm\n");
|
| 721 : |
|
|
handle.logMessage(LOGINFO, "\nThe number of threads must be a multiple of 2\n");
|
| 722 : |
|
|
return;
|
| 723 : |
|
|
}
|
| 724 : |
|
|
#else
|
| 725 : |
|
|
newf = this->bestNearby(delta, trialx, bestf, param);
|
| 726 : |
|
|
#endif
|
| 727 : |
|
|
/* if too many function evaluations occur, terminate the algorithm */
|
| 728 : |
|
|
|
| 729 : |
|
|
#ifdef NO_OPENMP
|
| 730 : |
|
|
iters = EcoSystem->getFuncEval() - offset;
|
| 731 : |
|
|
#endif
|
| 732 : |
|
|
if (iters > hookeiter) {
|
| 733 : |
|
|
handle.logMessage(LOGINFO, "\nStopping Hooke & Jeeves optimisation algorithm\n");
|
| 734 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped after", iters, "function evaluations");
|
| 735 : |
|
|
handle.logMessage(LOGINFO, "The steplength was reduced to", steplength);
|
| 736 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped because the maximum number of function evaluations");
|
| 737 : |
|
|
handle.logMessage(LOGINFO, "was reached and NOT because an optimum was found for this run");
|
| 738 : |
|
|
|
| 739 : |
|
|
score = EcoSystem->SimulateAndUpdate(trialx);
|
| 740 : |
|
|
handle.logMessage(LOGINFO, "\nHooke & Jeeves finished with a likelihood score of", score);
|
| 741 : |
|
|
for (i = 0; i < nvars; i++)
|
| 742 : |
|
|
bestx[i] = trialx[i] * init[i];
|
| 743 : |
|
|
EcoSystem->storeVariables(score, bestx);
|
| 744 : |
|
|
return;
|
| 745 : |
|
|
}
|
| 746 : |
|
|
|
| 747 : |
|
|
/* if we made some improvements, pursue that direction */
|
| 748 : |
|
|
while (newf < bestf) {
|
| 749 : |
|
|
for (i = 0; i < nvars; i++) {
|
| 750 : |
|
|
/* if it has been trapped but f has now gotten better (bndcheck) */
|
| 751 : |
|
|
/* we assume that we are out of the trap, reset the counters */
|
| 752 : |
|
|
/* and go back to the stepsize we had when we got trapped */
|
| 753 : |
|
|
if ((trapped[i]) && (newf < oldf * bndcheck)) {
|
| 754 : |
|
|
trapped[i] = 0;
|
| 755 : |
|
|
lbound[i] = 0;
|
| 756 : |
|
|
rbounds[i] = 0;
|
| 757 : |
|
|
delta[i] = initialstep[i];
|
| 758 : |
|
|
|
| 759 : |
|
|
} else if (trialx[i] < (lowerb[i] + verysmall)) {
|
| 760 : |
|
|
lbound[i]++;
|
| 761 : |
|
|
trialx[i] = lowerb[i];
|
| 762 : |
|
|
if (!trapped[i]) {
|
| 763 : |
|
|
initialstep[i] = delta[i];
|
| 764 : |
|
|
trapped[i] = 1;
|
| 765 : |
|
|
}
|
| 766 : |
|
|
/* if it has hit the bounds 2 times then increase the stepsize */
|
| 767 : |
|
|
if (lbound[i] >= 2)
|
| 768 : |
|
|
delta[i] /= rho;
|
| 769 : |
|
|
|
| 770 : |
|
|
} else if (trialx[i] > (upperb[i] - verysmall)) {
|
| 771 : |
|
|
rbounds[i]++;
|
| 772 : |
|
|
trialx[i] = upperb[i];
|
| 773 : |
|
|
if (!trapped[i]) {
|
| 774 : |
|
|
initialstep[i] = delta[i];
|
| 775 : |
|
|
trapped[i] = 1;
|
| 776 : |
|
|
}
|
| 777 : |
|
|
/* if it has hit the bounds 2 times then increase the stepsize */
|
| 778 : |
|
|
if (rbounds[i] >= 2)
|
| 779 : |
|
|
delta[i] /= rho;
|
| 780 : |
|
|
}
|
| 781 : |
|
|
}
|
| 782 : |
|
|
|
| 783 : |
|
|
for (i = 0; i < nvars; i++) {
|
| 784 : |
|
|
/* firstly, arrange the sign of delta[] */
|
| 785 : |
|
|
if (trialx[i] < x[i])
|
| 786 : |
|
|
delta[i] = 0.0 - fabs(delta[i]);
|
| 787 : |
|
|
else
|
| 788 : |
|
|
delta[i] = fabs(delta[i]);
|
| 789 : |
|
|
|
| 790 : |
|
|
/* now, move further in this direction */
|
| 791 : |
|
|
tmp = x[i];
|
| 792 : |
|
|
x[i] = trialx[i];
|
| 793 : |
|
|
trialx[i] = trialx[i] + trialx[i] - tmp;
|
| 794 : |
|
|
}
|
| 795 : |
|
|
|
| 796 : |
|
|
/* only move forward if this is really an improvement */
|
| 797 : |
|
|
oldf = newf;
|
| 798 : |
|
|
newf = EcoSystem->SimulateAndUpdate(trialx);
|
| 799 : |
|
|
#ifndef NO_OPENMP
|
| 800 : |
|
|
iters++;
|
| 801 : |
|
|
#endif
|
| 802 : |
|
|
if ((isEqual(newf, oldf)) || (newf > oldf)) {
|
| 803 : |
|
|
newf = oldf; //JMB no improvement, so reset the value of newf
|
| 804 : |
|
|
break;
|
| 805 : |
|
|
}
|
| 806 : |
|
|
|
| 807 : |
|
|
/* OK, it's better, so update variables and look around */
|
| 808 : |
|
|
bestf = newf;
|
| 809 : |
|
|
for (i = 0; i < nvars; i++)
|
| 810 : |
|
|
x[i] = trialx[i];
|
| 811 : |
|
|
|
| 812 : |
|
|
#ifndef NO_OPENMP
|
| 813 : |
|
|
newf = this->bestNearbyOMP2(delta, trialx, bestf, param);
|
| 814 : |
|
|
if (newf == -1) {
|
| 815 : |
|
|
handle.logMessage(LOGINFO, "\nStopping Hooke & Jeeves optimisation algorithm\n");
|
| 816 : |
|
|
handle.logMessage(LOGINFO, "\nThe number of threads must be a multiple of 2\n");
|
| 817 : |
|
|
return;
|
| 818 : |
|
|
}
|
| 819 : |
|
|
#else
|
| 820 : |
|
|
newf = this->bestNearby(delta, trialx, bestf, param);
|
| 821 : |
|
|
#endif
|
| 822 : |
|
|
if (isEqual(newf, bestf))
|
| 823 : |
|
|
break;
|
| 824 : |
|
|
|
| 825 : |
|
|
/* if too many function evaluations occur, terminate the algorithm */
|
| 826 : |
|
|
#ifdef NO_OPENMP
|
| 827 : |
|
|
iters = EcoSystem->getFuncEval() - offset;
|
| 828 : |
|
|
#endif
|
| 829 : |
|
|
if (iters > hookeiter) {
|
| 830 : |
|
|
handle.logMessage(LOGINFO, "\nStopping Hooke & Jeeves optimisation algorithm\n");
|
| 831 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped after", iters, "function evaluations");
|
| 832 : |
|
|
handle.logMessage(LOGINFO, "The steplength was reduced to", steplength);
|
| 833 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped because the maximum number of function evaluations");
|
| 834 : |
|
|
handle.logMessage(LOGINFO, "was reached and NOT because an optimum was found for this run");
|
| 835 : |
|
|
|
| 836 : |
|
|
score = EcoSystem->SimulateAndUpdate(trialx);
|
| 837 : |
|
|
handle.logMessage(LOGINFO, "\nHooke & Jeeves finished with a likelihood score of", score);
|
| 838 : |
|
|
for (i = 0; i < nvars; i++)
|
| 839 : |
|
|
bestx[i] = trialx[i] * init[i];
|
| 840 : |
|
|
EcoSystem->storeVariables(score, bestx);
|
| 841 : |
|
|
return;
|
| 842 : |
|
|
}
|
| 843 : |
|
|
}
|
| 844 : |
|
|
|
| 845 : |
|
|
#ifdef NO_OPENMP
|
| 846 : |
|
|
iters = EcoSystem->getFuncEval() - offset;
|
| 847 : |
|
|
#endif
|
| 848 : |
|
|
if (newf < bestf) {
|
| 849 : |
|
|
for (i = 0; i < nvars; i++)
|
| 850 : |
|
|
bestx[i] = x[i] * init[i];
|
| 851 : |
|
|
bestf = newf;
|
| 852 : |
|
|
handle.logMessage(LOGINFO, "\nNew optimum found after", iters, "function evaluations");
|
| 853 : |
|
|
handle.logMessage(LOGINFO, "The likelihood score is", bestf, "at the point");
|
| 854 : |
|
|
EcoSystem->storeVariables(bestf, bestx);
|
| 855 : |
|
|
EcoSystem->writeBestValues();
|
| 856 : |
|
|
|
| 857 : |
|
|
} else
|
| 858 : |
|
|
handle.logMessage(LOGINFO, "Checking convergence criteria after", iters, "function evaluations ...");
|
| 859 : |
|
|
|
| 860 : |
|
|
/* if the step length is less than hookeeps, terminate the algorithm */
|
| 861 : |
|
|
if (steplength < hookeeps) {
|
| 862 : |
|
|
handle.logMessage(LOGINFO, "\nStopping Hooke & Jeeves optimisation algorithm\n");
|
| 863 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped after", iters, "function evaluations");
|
| 864 : |
|
|
handle.logMessage(LOGINFO, "The steplength was reduced to", steplength);
|
| 865 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped because an optimum was found for this run");
|
| 866 : |
|
|
|
| 867 : |
|
|
converge = 1;
|
| 868 : |
|
|
score = bestf;
|
| 869 : |
|
|
handle.logMessage(LOGINFO, "\nHooke & Jeeves finished with a likelihood score of", score);
|
| 870 : |
|
|
EcoSystem->storeVariables(bestf, bestx);
|
| 871 : |
|
|
return;
|
| 872 : |
|
|
}
|
| 873 : |
|
|
|
| 874 : |
|
|
steplength *= rho;
|
| 875 : |
|
|
handle.logMessage(LOGINFO, "Reducing the steplength to", steplength);
|
| 876 : |
|
|
for (i = 0; i < nvars; i++)
|
| 877 : |
|
|
delta[i] *= rho;
|
| 878 : |
|
|
}
|
| 879 : |
|
|
}
|
| 880 : |
|
|
#endif
|
Parent Directory
Revision Log