1 : |
agomez |
1 |
/* ABSTRACT: */
|
2 : |
|
|
/* Simulated annealing is a global optimization method that distinguishes */
|
3 : |
|
|
/* different local optima. Starting from an initial point, the algorithm */
|
4 : |
|
|
/* takes a step and the function is evaluated. When minimizing a function,*/
|
5 : |
|
|
/* any downhill step is accepted and the process repeats from this new */
|
6 : |
|
|
/* point. An uphill step may be accepted (thus, it can escape from local */
|
7 : |
|
|
/* optima). This uphill decision is made by the Metropolis criteria. As */
|
8 : |
|
|
/* the optimization process proceeds, the length of the steps decline and */
|
9 : |
|
|
/* the algorithm closes in on the global optimum. Since the algorithm */
|
10 : |
|
|
/* makes very few assumptions regarding the function to be optimized, it */
|
11 : |
|
|
/* is quite robust with respect to non-quadratic surfaces. The degree of */
|
12 : |
|
|
/* robustness can be adjusted by the user. In fact, simulated annealing */
|
13 : |
|
|
/* can be used as a local optimizer for difficult functions. */
|
14 : |
|
|
/* */
|
15 : |
|
|
/* The author can be contacted at h2zr1001@vm.cis.smu.edu */
|
16 : |
|
|
|
17 : |
|
|
/* This file is a translation of a fortran code, which is an example of the*/
|
18 : |
|
|
/* Corana et al. simulated annealing algorithm for multimodal and robust */
|
19 : |
|
|
/* optimization as implemented and modified by by Goffe et al. */
|
20 : |
|
|
/* */
|
21 : |
|
|
/* Use the sample function from Judge with the following suggestions */
|
22 : |
|
|
/* to get a feel for how SA works. When you've done this, you should be */
|
23 : |
|
|
/* ready to use it on most any function with a fair amount of expertise. */
|
24 : |
|
|
/* 1. Run the program as is to make sure it runs okay. Take a look at */
|
25 : |
|
|
/* the intermediate output and see how it optimizes as temperature */
|
26 : |
|
|
/* (T) falls. Notice how the optimal point is reached and how */
|
27 : |
|
|
/* falling T reduces VM. */
|
28 : |
|
|
/* 2. Look through the documentation to SA so the following makes a */
|
29 : |
|
|
/* bit of sense. In line with the paper, it shouldn't be that hard */
|
30 : |
|
|
/* to figure out. The core of the algorithm is described on pp. 4-6 */
|
31 : |
|
|
/* and on pp. 28. Also see Corana et al. pp. 264-9. */
|
32 : |
|
|
/* 3. To see the importance of different temperatures, try starting */
|
33 : |
|
|
/* with a very low one (say T = 10E-5). You'll see (i) it never */
|
34 : |
|
|
/* escapes from the local optima (in annealing terminology, it */
|
35 : |
|
|
/* quenches) & (ii) the step length (VM) will be quite small. This */
|
36 : |
|
|
/* is a key part of the algorithm: as temperature (T) falls, step */
|
37 : |
|
|
/* length does too. In a minor point here, note how VM is quickly */
|
38 : |
|
|
/* reset from its initial value. Thus, the input VM is not very */
|
39 : |
|
|
/* important. This is all the more reason to examine VM once the */
|
40 : |
|
|
/* algorithm is underway. */
|
41 : |
|
|
/* 4. To see the effect of different parameters and their effect on */
|
42 : |
|
|
/* the speed of the algorithm, try RT = .95 & RT = .1. Notice the */
|
43 : |
|
|
/* vastly different speed for optimization. Also try NT = 20. Note */
|
44 : |
|
|
/* that this sample function is quite easy to optimize, so it will */
|
45 : |
|
|
/* tolerate big changes in these parameters. RT and NT are the */
|
46 : |
|
|
/* parameters one should adjust to modify the runtime of the */
|
47 : |
|
|
/* algorithm and its robustness. */
|
48 : |
|
|
/* 5. Try constraining the algorithm with either LB or UB. */
|
49 : |
|
|
|
50 : |
|
|
/* Synopsis: */
|
51 : |
|
|
/* This routine implements the continuous simulated annealing global */
|
52 : |
|
|
/* optimization algorithm described in Corana et al.'s article */
|
53 : |
|
|
/* "Minimizing Multimodal Functions of Continuous Variables with the */
|
54 : |
|
|
/* "Simulated Annealing" Algorithm" in the September 1987 (vol. 13, */
|
55 : |
|
|
/* no. 3, pp. 262-280) issue of the ACM Transactions on Mathematical */
|
56 : |
|
|
/* Software. */
|
57 : |
|
|
|
58 : |
|
|
/* A very quick (perhaps too quick) overview of SA: */
|
59 : |
|
|
/* SA tries to find the global optimum of an N dimensional function. */
|
60 : |
|
|
/* It moves both up and downhill and as the optimization process */
|
61 : |
|
|
/* proceeds, it focuses on the most promising area. */
|
62 : |
|
|
/* To start, it randomly chooses a trial point within the step length */
|
63 : |
|
|
/* VM (a vector of length N) of the user selected starting point. The */
|
64 : |
|
|
/* function is evaluated at this trial point and its value is compared */
|
65 : |
|
|
/* to its value at the initial point. */
|
66 : |
|
|
/* In a maximization problem, all uphill moves are accepted and the */
|
67 : |
|
|
/* algorithm continues from that trial point. Downhill moves may be */
|
68 : |
|
|
/* accepted; the decision is made by the Metropolis criteria. It uses T */
|
69 : |
|
|
/* (temperature) and the size of the downhill move in a probabilistic */
|
70 : |
|
|
/* manner. The smaller T and the size of the downhill move are, the more */
|
71 : |
|
|
/* likely that move will be accepted. If the trial is accepted, the */
|
72 : |
|
|
/* algorithm moves on from that point. If it is rejected, another point */
|
73 : |
|
|
/* is chosen instead for a trial evaluation. */
|
74 : |
|
|
/* Each element of VM periodically adjusted so that half of all */
|
75 : |
|
|
/* function evaluations in that direction are accepted. */
|
76 : |
|
|
/* A fall in T is imposed upon the system with the RT variable by */
|
77 : |
|
|
/* T(i+1) = RT*T(i) where i is the ith iteration. Thus, as T declines, */
|
78 : |
|
|
/* downhill moves are less likely to be accepted and the percentage of */
|
79 : |
|
|
/* rejections rise. Given the scheme for the selection for VM, VM falls. */
|
80 : |
|
|
/* Thus, as T declines, VM falls and SA focuses upon the most promising */
|
81 : |
|
|
/* area for optimization. */
|
82 : |
|
|
|
83 : |
|
|
/* The importance of the parameter T: */
|
84 : |
|
|
/* The parameter T is crucial in using SA successfully. It influences */
|
85 : |
|
|
/* VM, the step length over which the algorithm searches for optima. For */
|
86 : |
|
|
/* a small intial T, the step length may be too small; thus not enough */
|
87 : |
|
|
/* of the function might be evaluated to find the global optima. The user */
|
88 : |
|
|
/* should carefully examine VM in the intermediate output (set IPRINT = */
|
89 : |
|
|
/* 1) to make sure that VM is appropriate. The relationship between the */
|
90 : |
|
|
/* initial temperature and the resulting step length is function */
|
91 : |
|
|
/* dependent. */
|
92 : |
|
|
/* To determine the starting temperature that is consistent with */
|
93 : |
|
|
/* optimizing a function, it is worthwhile to run a trial run first. Set */
|
94 : |
|
|
/* RT = 1.5 and T = 1.0. With RT > 1.0, the temperature increases and VM */
|
95 : |
|
|
/* rises as well. Then select the T that produces a large enough VM. */
|
96 : |
|
|
|
97 : |
|
|
/* For modifications to the algorithm and many details on its use, */
|
98 : |
|
|
/* (particularly for econometric applications) see Goffe, Ferrier */
|
99 : |
|
|
/* and Rogers, "Global Optimization of Statistical Functions with */
|
100 : |
|
|
/* the Simulated Annealing," Journal of Econometrics (forthcoming) */
|
101 : |
|
|
/* For a pre-publication copy, contact */
|
102 : |
|
|
/* Bill Goffe */
|
103 : |
|
|
/* Department of Economics */
|
104 : |
|
|
/* Southern Methodist University */
|
105 : |
|
|
/* Dallas, TX 75275 */
|
106 : |
|
|
/* h2zr1001 @ smuvm1 (Bitnet) */
|
107 : |
|
|
/* h2zr1001 @ vm.cis.smu.edu (Internet) */
|
108 : |
|
|
|
109 : |
|
|
/* As far as possible, the parameters here have the same name as in */
|
110 : |
|
|
/* the description of the algorithm on pp. 266-8 of Corana et al. */
|
111 : |
|
|
|
112 : |
|
|
/* Input Parameters: */
|
113 : |
|
|
/* Note: The suggested values generally come from Corana et al. To */
|
114 : |
|
|
/* drastically reduce runtime, see Goffe et al., pp. 17-8 for */
|
115 : |
|
|
/* suggestions on choosing the appropriate RT and NT. */
|
116 : |
|
|
/* n - Number of variables in the function to be optimized. (INT) */
|
117 : |
|
|
/* x - The starting values for the variables of the function to be */
|
118 : |
|
|
/* optimized. (DP(N)) */
|
119 : |
|
|
/* max - Denotes whether the function should be maximized or */
|
120 : |
|
|
/* minimized. A true value denotes maximization while a false */
|
121 : |
|
|
/* value denotes minimization. */
|
122 : |
|
|
/* RT - The temperature reduction factor. The value suggested by */
|
123 : |
|
|
/* Corana et al. is .85. See Goffe et al. for more advice. (DP) */
|
124 : |
|
|
/* EPS - Error tolerance for termination. If the final function */
|
125 : |
|
|
/* values from the last neps temperatures differ from the */
|
126 : |
|
|
/* corresponding value at the current temperature by less than */
|
127 : |
|
|
/* EPS and the final function value at the current temperature */
|
128 : |
|
|
/* differs from the current optimal function value by less than */
|
129 : |
|
|
/* EPS, execution terminates and IER = 0 is returned. (EP) */
|
130 : |
|
|
/* NS - Number of cycles. After NS*N function evaluations, each element */
|
131 : |
|
|
/* of VM is adjusted so that approximately half of all function */
|
132 : |
|
|
/* evaluations are accepted. The suggested value is 20. (INT) */
|
133 : |
|
|
/* nt - Number of iterations before temperature reduction. After */
|
134 : |
|
|
/* NT*NS*N function evaluations, temperature (T) is changed */
|
135 : |
|
|
/* by the factor RT. Value suggested by Corana et al. is */
|
136 : |
|
|
/* MAX(100, 5*N). See Goffe et al. for further advice. (INT) */
|
137 : |
|
|
/* NEPS - Number of final function values used to decide upon termi- */
|
138 : |
|
|
/* nation. See EPS. Suggested value is 4. (INT) */
|
139 : |
|
|
/* maxevl - The maximum number of function evaluations. If it is */
|
140 : |
|
|
/* exceeded, IER = 1. (INT) */
|
141 : |
|
|
/* lb - The lower bound for the allowable solution variables. (DP(N)) */
|
142 : |
|
|
/* ub - The upper bound for the allowable solution variables. (DP(N)) */
|
143 : |
|
|
/* If the algorithm chooses X(I) .LT. LB(I) or X(I) .GT. UB(I), */
|
144 : |
|
|
/* I = 1, N, a point is from inside is randomly selected. This */
|
145 : |
|
|
/* This focuses the algorithm on the region inside UB and LB. */
|
146 : |
|
|
/* Unless the user wishes to concentrate the search to a par- */
|
147 : |
|
|
/* ticular region, UB and LB should be set to very large positive */
|
148 : |
|
|
/* and negative values, respectively. Note that the starting */
|
149 : |
|
|
/* vector X should be inside this region. Also note that LB and */
|
150 : |
|
|
/* UB are fixed in position, while VM is centered on the last */
|
151 : |
|
|
/* accepted trial set of variables that optimizes the function. */
|
152 : |
|
|
/* c - Vector that controls the step length adjustment. The suggested */
|
153 : |
|
|
/* value for all elements is 2.0. (DP(N)) */
|
154 : |
|
|
/* t - On input, the initial temperature. See Goffe et al. for advice. */
|
155 : |
|
|
/* On output, the final temperature. (DP) */
|
156 : |
|
|
/* vm - The step length vector. On input it should encompass the */
|
157 : |
|
|
/* region of interest given the starting value X. For point */
|
158 : |
|
|
/* X(I), the next trial point is selected is from X(I) - VM(I) */
|
159 : |
|
|
/* to X(I) + VM(I). Since VM is adjusted so that about half */
|
160 : |
|
|
/* of all points are accepted, the input value is not very */
|
161 : |
|
|
/* important (i.e. is the value is off, SA adjusts VM to the */
|
162 : |
|
|
/* correct value). (DP(N)) */
|
163 : |
|
|
|
164 : |
|
|
/* Output Parameters: */
|
165 : |
|
|
/* xopt - The variables that optimize the function. (DP(N)) */
|
166 : |
|
|
/* fopt - The optimal value of the function. (DP) */
|
167 : |
|
|
|
168 : |
|
|
/* JMB this has been modified to work with the gadget object structure */
|
169 : |
|
|
/* This means that the function has been replaced by a call to ecosystem */
|
170 : |
|
|
/* object, and we can use the vector objects that have been defined */
|
171 : |
|
|
|
172 : |
|
|
#include "gadget.h" |
173 : |
|
|
#include "optinfo.h" |
174 : |
|
|
#include "mathfunc.h" |
175 : |
|
|
#include "doublevector.h" |
176 : |
|
|
#include "intvector.h" |
177 : |
|
|
#include "errorhandler.h" |
178 : |
|
|
#include "ecosystem.h" |
179 : |
|
|
#include "global.h" |
180 : |
|
|
|
181 : |
|
|
extern Ecosystem* EcoSystem;
|
182 : |
|
|
|
183 : |
|
|
|
184 : |
|
|
void OptInfoSimann::OptimiseLikelihood() {
|
185 : |
|
|
|
186 : |
|
|
//set initial values
|
187 : |
|
|
int nacc = 0; //The number of accepted function evaluations
|
188 : |
|
|
int nrej = 0; //The number of rejected function evaluations
|
189 : |
|
|
int naccmet = 0; //The number of metropolis accepted function evaluations
|
190 : |
|
|
|
191 : |
|
|
double tmp, p, pp, ratio, nsdiv;
|
192 : |
|
|
double fopt, funcval, trialf;
|
193 : |
|
|
int a, i, j, k, l, offset, quit;
|
194 : |
|
|
int rchange, rcheck, rnumber; //Used to randomise the order of the parameters
|
195 : |
|
|
|
196 : |
|
|
handle.logMessage(LOGINFO, "\nStarting Simulated Annealing optimisation algorithm\n");
|
197 : |
|
|
int nvars = EcoSystem->numOptVariables();
|
198 : |
|
|
DoubleVector x(nvars);
|
199 : |
|
|
DoubleVector init(nvars);
|
200 : |
|
|
DoubleVector trialx(nvars, 0.0);
|
201 : |
|
|
DoubleVector bestx(nvars);
|
202 : |
|
|
DoubleVector scalex(nvars);
|
203 : |
|
|
DoubleVector lowerb(nvars);
|
204 : |
|
|
DoubleVector upperb(nvars);
|
205 : |
|
|
DoubleVector fstar(tempcheck);
|
206 : |
|
|
DoubleVector vm(nvars, vminit);
|
207 : |
|
|
IntVector param(nvars, 0);
|
208 : |
|
|
IntVector nacp(nvars, 0);
|
209 : |
|
|
|
210 : |
|
|
EcoSystem->resetVariables(); //JMB need to reset variables in case they have been scaled
|
211 : |
|
|
if (scale)
|
212 : |
|
|
EcoSystem->scaleVariables();
|
213 : |
|
|
EcoSystem->getOptScaledValues(x);
|
214 : |
|
|
EcoSystem->getOptLowerBounds(lowerb);
|
215 : |
|
|
EcoSystem->getOptUpperBounds(upperb);
|
216 : |
|
|
EcoSystem->getOptInitialValues(init);
|
217 : |
|
|
|
218 : |
|
|
for (i = 0; i < nvars; i++) {
|
219 : |
|
|
bestx[i] = x[i];
|
220 : |
|
|
param[i] = i;
|
221 : |
|
|
}
|
222 : |
|
|
|
223 : |
|
|
if (scale) {
|
224 : |
|
|
for (i = 0; i < nvars; i++) {
|
225 : |
|
|
scalex[i] = x[i];
|
226 : |
|
|
// Scaling the bounds, because the parameters are scaled
|
227 : |
|
|
lowerb[i] = lowerb[i] / init[i];
|
228 : |
|
|
upperb[i] = upperb[i] / init[i];
|
229 : |
|
|
if (lowerb[i] > upperb[i]) {
|
230 : |
|
|
tmp = lowerb[i];
|
231 : |
|
|
lowerb[i] = upperb[i];
|
232 : |
|
|
upperb[i] = tmp;
|
233 : |
|
|
}
|
234 : |
|
|
}
|
235 : |
|
|
}
|
236 : |
|
|
|
237 : |
|
|
//funcval is the function value at x
|
238 : |
|
|
funcval = EcoSystem->SimulateAndUpdate(x);
|
239 : |
|
|
if (funcval != funcval) { //check for NaN
|
240 : |
|
|
handle.logMessage(LOGINFO, "Error starting Simulated Annealing optimisation with f(x) = infinity");
|
241 : |
|
|
converge = -1;
|
242 : |
|
|
iters = 1;
|
243 : |
|
|
return;
|
244 : |
|
|
}
|
245 : |
|
|
|
246 : |
|
|
//the function is to be minimised so switch the sign of funcval (and trialf)
|
247 : |
|
|
funcval = -funcval;
|
248 : |
|
|
offset = EcoSystem->getFuncEval(); //number of function evaluations done before loop
|
249 : |
|
|
nacc++;
|
250 : |
|
|
cs /= lratio; //JMB save processing time
|
251 : |
|
|
nsdiv = 1.0 / ns;
|
252 : |
|
|
fopt = funcval;
|
253 : |
|
|
for (i = 0; i < tempcheck; i++)
|
254 : |
|
|
fstar[i] = funcval;
|
255 : |
|
|
|
256 : |
|
|
//Start the main loop. Note that it terminates if
|
257 : |
|
|
//(i) the algorithm succesfully optimises the function or
|
258 : |
|
|
//(ii) there are too many function evaluations
|
259 : |
|
|
while (1) {
|
260 : |
|
|
for (a = 0; a < nt; a++) {
|
261 : |
|
|
//Randomize the order of the parameters once in a while, to avoid
|
262 : |
|
|
//the order having an influence on which changes are accepted
|
263 : |
|
|
rchange = 0;
|
264 : |
|
|
while (rchange < nvars) {
|
265 : |
|
|
rnumber = rand() % nvars;
|
266 : |
|
|
rcheck = 1;
|
267 : |
|
|
for (i = 0; i < rchange; i++)
|
268 : |
|
|
if (param[i] == rnumber)
|
269 : |
|
|
rcheck = 0;
|
270 : |
|
|
if (rcheck) {
|
271 : |
|
|
param[rchange] = rnumber;
|
272 : |
|
|
rchange++;
|
273 : |
|
|
}
|
274 : |
|
|
}
|
275 : |
|
|
|
276 : |
|
|
for (j = 0; j < ns; j++) {
|
277 : |
|
|
for (l = 0; l < nvars; l++) {
|
278 : |
|
|
//Generate trialx, the trial value of x
|
279 : |
|
|
for (i = 0; i < nvars; i++) {
|
280 : |
|
|
if (i == param[l]) {
|
281 : |
|
|
trialx[i] = x[i] + ((randomNumber() * 2.0) - 1.0) * vm[i];
|
282 : |
|
|
|
283 : |
|
|
//If trialx is out of bounds, try again until we find a point that is OK
|
284 : |
|
|
if ((trialx[i] < lowerb[i]) || (trialx[i] > upperb[i])) {
|
285 : |
|
|
//JMB - this used to just select a random point between the bounds
|
286 : |
|
|
k = 0;
|
287 : |
|
|
while ((trialx[i] < lowerb[i]) || (trialx[i] > upperb[i])) {
|
288 : |
|
|
trialx[i] = x[i] + ((randomNumber() * 2.0) - 1.0) * vm[i];
|
289 : |
|
|
k++;
|
290 : |
|
|
if (k > 10) //we've had 10 tries to find a point neatly, so give up
|
291 : |
|
|
trialx[i] = lowerb[i] + (upperb[i] - lowerb[i]) * randomNumber();
|
292 : |
|
|
}
|
293 : |
|
|
}
|
294 : |
|
|
|
295 : |
|
|
} else
|
296 : |
|
|
trialx[i] = x[i];
|
297 : |
|
|
}
|
298 : |
|
|
|
299 : |
|
|
//Evaluate the function with the trial point trialx and return as -trialf
|
300 : |
|
|
trialf = EcoSystem->SimulateAndUpdate(trialx);
|
301 : |
|
|
trialf = -trialf;
|
302 : |
|
|
|
303 : |
|
|
//If too many function evaluations occur, terminate the algorithm
|
304 : |
|
|
iters = EcoSystem->getFuncEval() - offset;
|
305 : |
|
|
if (iters > simanniter) {
|
306 : |
|
|
handle.logMessage(LOGINFO, "\nStopping Simulated Annealing optimisation algorithm\n");
|
307 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped after", iters, "function evaluations");
|
308 : |
|
|
handle.logMessage(LOGINFO, "The temperature was reduced to", t);
|
309 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped because the maximum number of function evaluations");
|
310 : |
|
|
handle.logMessage(LOGINFO, "was reached and NOT because an optimum was found for this run");
|
311 : |
|
|
handle.logMessage(LOGINFO, "Number of directly accepted points", nacc);
|
312 : |
|
|
handle.logMessage(LOGINFO, "Number of metropolis accepted points", naccmet);
|
313 : |
|
|
handle.logMessage(LOGINFO, "Number of rejected points", nrej);
|
314 : |
|
|
|
315 : |
|
|
score = EcoSystem->SimulateAndUpdate(bestx);
|
316 : |
|
|
handle.logMessage(LOGINFO, "\nSimulated Annealing finished with a likelihood score of", score);
|
317 : |
|
|
return;
|
318 : |
|
|
}
|
319 : |
|
|
|
320 : |
|
|
//Accept the new point if the new function value better
|
321 : |
|
|
if ((trialf - funcval) > verysmall) {
|
322 : |
|
|
for (i = 0; i < nvars; i++)
|
323 : |
|
|
x[i] = trialx[i];
|
324 : |
|
|
funcval = trialf;
|
325 : |
|
|
nacc++;
|
326 : |
|
|
nacp[param[l]]++; //JMB - not sure about this ...
|
327 : |
|
|
|
328 : |
|
|
} else {
|
329 : |
|
|
//Accept according to metropolis condition
|
330 : |
|
|
p = expRep((trialf - funcval) / t);
|
331 : |
|
|
pp = randomNumber();
|
332 : |
|
|
if (pp < p) {
|
333 : |
|
|
//Accept point
|
334 : |
|
|
for (i = 0; i < nvars; i++)
|
335 : |
|
|
x[i] = trialx[i];
|
336 : |
|
|
funcval = trialf;
|
337 : |
|
|
naccmet++;
|
338 : |
|
|
nacp[param[l]]++;
|
339 : |
|
|
} else {
|
340 : |
|
|
//Reject point
|
341 : |
|
|
nrej++;
|
342 : |
|
|
}
|
343 : |
|
|
}
|
344 : |
|
|
|
345 : |
|
|
// JMB added check for really silly values
|
346 : |
|
|
if (isZero(trialf)) {
|
347 : |
|
|
handle.logMessage(LOGINFO, "Error in Simulated Annealing optimisation after", iters, "function evaluations, f(x) = 0");
|
348 : |
|
|
converge = -1;
|
349 : |
|
|
return;
|
350 : |
|
|
}
|
351 : |
|
|
|
352 : |
|
|
//If greater than any other point, record as new optimum
|
353 : |
|
|
if ((trialf > fopt) && (trialf == trialf)) {
|
354 : |
|
|
for (i = 0; i < nvars; i++)
|
355 : |
|
|
bestx[i] = trialx[i];
|
356 : |
|
|
fopt = trialf;
|
357 : |
|
|
|
358 : |
|
|
if (scale) {
|
359 : |
|
|
for (i = 0; i < nvars; i++)
|
360 : |
|
|
scalex[i] = bestx[i] * init[i];
|
361 : |
|
|
EcoSystem->storeVariables(-fopt, scalex);
|
362 : |
|
|
} else
|
363 : |
|
|
EcoSystem->storeVariables(-fopt, bestx);
|
364 : |
|
|
|
365 : |
|
|
handle.logMessage(LOGINFO, "\nNew optimum found after", iters, "function evaluations");
|
366 : |
|
|
handle.logMessage(LOGINFO, "The likelihood score is", -fopt, "at the point");
|
367 : |
|
|
EcoSystem->writeBestValues();
|
368 : |
|
|
}
|
369 : |
|
|
}
|
370 : |
|
|
}
|
371 : |
|
|
|
372 : |
|
|
//Adjust vm so that approximately half of all evaluations are accepted
|
373 : |
|
|
for (i = 0; i < nvars; i++) {
|
374 : |
|
|
ratio = nsdiv * nacp[i];
|
375 : |
|
|
nacp[i] = 0;
|
376 : |
|
|
if (ratio > uratio) {
|
377 : |
|
|
vm[i] = vm[i] * (1.0 + cs * (ratio - uratio));
|
378 : |
|
|
} else if (ratio < lratio) {
|
379 : |
|
|
vm[i] = vm[i] / (1.0 + cs * (lratio - ratio));
|
380 : |
|
|
}
|
381 : |
|
|
|
382 : |
|
|
if (vm[i] < rathersmall)
|
383 : |
|
|
vm[i] = rathersmall;
|
384 : |
|
|
if (vm[i] > (upperb[i] - lowerb[i]))
|
385 : |
|
|
vm[i] = upperb[i] - lowerb[i];
|
386 : |
|
|
}
|
387 : |
|
|
}
|
388 : |
|
|
|
389 : |
|
|
//Check termination criteria
|
390 : |
|
|
for (i = tempcheck - 1; i > 0; i--)
|
391 : |
|
|
fstar[i] = fstar[i - 1];
|
392 : |
|
|
fstar[0] = funcval;
|
393 : |
|
|
|
394 : |
|
|
quit = 0;
|
395 : |
|
|
if (fabs(fopt - funcval) < simanneps) {
|
396 : |
|
|
quit = 1;
|
397 : |
|
|
for (i = 0; i < tempcheck - 1; i++)
|
398 : |
|
|
if (fabs(fstar[i + 1] - fstar[i]) > simanneps)
|
399 : |
|
|
quit = 0;
|
400 : |
|
|
}
|
401 : |
|
|
|
402 : |
|
|
handle.logMessage(LOGINFO, "Checking convergence criteria after", iters, "function evaluations ...");
|
403 : |
|
|
|
404 : |
|
|
//Terminate SA if appropriate
|
405 : |
|
|
if (quit) {
|
406 : |
|
|
handle.logMessage(LOGINFO, "\nStopping Simulated Annealing optimisation algorithm\n");
|
407 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped after", iters, "function evaluations");
|
408 : |
|
|
handle.logMessage(LOGINFO, "The temperature was reduced to", t);
|
409 : |
|
|
handle.logMessage(LOGINFO, "The optimisation stopped because an optimum was found for this run");
|
410 : |
|
|
handle.logMessage(LOGINFO, "Number of directly accepted points", nacc);
|
411 : |
|
|
handle.logMessage(LOGINFO, "Number of metropolis accepted points", naccmet);
|
412 : |
|
|
handle.logMessage(LOGINFO, "Number of rejected points", nrej);
|
413 : |
|
|
|
414 : |
|
|
converge = 1;
|
415 : |
|
|
score = EcoSystem->SimulateAndUpdate(bestx);
|
416 : |
|
|
handle.logMessage(LOGINFO, "\nSimulated Annealing finished with a likelihood score of", score);
|
417 : |
|
|
return;
|
418 : |
|
|
}
|
419 : |
|
|
|
420 : |
|
|
//If termination criteria is not met, prepare for another loop.
|
421 : |
|
|
t *= rt;
|
422 : |
|
|
if (t < rathersmall)
|
423 : |
|
|
t = rathersmall; //JMB make sure temperature doesnt get too small
|
424 : |
|
|
|
425 : |
|
|
handle.logMessage(LOGINFO, "Reducing the temperature to", t);
|
426 : |
|
|
funcval = fopt;
|
427 : |
|
|
for (i = 0; i < nvars; i++)
|
428 : |
|
|
x[i] = bestx[i];
|
429 : |
|
|
}
|
430 : |
|
|
}
|