8 int mma_verbose = 0; /* > 0 for verbose output */
10 #define MIN(a,b) ((a) < (b) ? (a) : (b))
11 #define MAX(a,b) ((a) > (b) ? (a) : (b))
13 /* magic minimum value for rho in MMA ... the 2002 paper says it should
14 be a "fixed, strictly positive `small' number, e.g. 1e-5"
15 ... grrr, I hate these magic numbers, which seem like they
16 should depend on the objective function in some way ... in particular,
17 note that rho is dimensionful (= dimensions of objective function) */
18 #define MMA_RHOMIN 1e-5
20 /***********************************************************************/
21 /* function for MMA's dual solution of the approximate problem */
24 int n; /* must be set on input to dimension of x */
25 const double *x, *lb, *ub, *sigma, *dfdx; /* arrays of length n */
26 const double *dfcdx; /* m-by-n array of fc gradients */
27 double fval, rho; /* must be set on input */
28 const double *fcval, *rhoc; /* arrays of length m */
29 double *xcur; /* array of length n, output each time */
30 double gval, wval, *gcval; /* output each time (array length m) */
33 static double sqr(double x) { return x * x; }
35 static double dual_func(int m, const double *y, double *grad, void *d_)
37 dual_data *d = (dual_data *) d_;
39 const double *x = d->x, *lb = d->lb, *ub = d->ub, *sigma = d->sigma,
41 const double *dfcdx = d->dfcdx;
42 double rho = d->rho, fval = d->fval;
43 const double *rhoc = d->rhoc, *fcval = d->fcval;
44 double *xcur = d->xcur;
45 double *gcval = d->gcval;
51 for (i = 0; i < m; ++i) val += y[i] * (gcval[i] = fcval[i]);
53 for (j = 0; j < n; ++j) {
54 double u, v, dx, denominv, c, sigma2, dx2;
56 /* first, compute xcur[j] for y. Because this objective is
57 separable, we can minimize over x analytically, and the minimum
58 dx is given by the solution of a quadratic equation:
59 u dx^2 + 2 v sigma^2 dx + u sigma^2 = 0
60 where u and v are defined by the sums below. Because of
61 the definitions, it is guaranteed that |u/v| <= sigma,
62 and it follows that the only dx solution with |dx| <= sigma
64 (v/u) sigma^2 (-1 + sqrt(1 - (u / v sigma)^2))
65 (which goes to zero as u -> 0). */
68 v = fabs(dfdx[j]) * sigma[j] + 0.5 * rho;
69 for (i = 0; i < m; ++i) {
70 u += dfcdx[i*n + j] * y[i];
71 v += (fabs(dfcdx[i*n + j]) * sigma[j] + 0.5 * rhoc[i]) * y[i];
73 u *= (sigma2 = sqr(sigma[j]));
74 dx = u==0 ? 0 : (v/u)*sigma2 * (-1 + sqrt(1 - sqr(u/(v*sigma[j]))));
76 if (xcur[j] > ub[j]) xcur[j] = ub[j];
77 else if (xcur[j] < lb[j]) xcur[j] = lb[j];
78 if (xcur[j] > x[j]+0.9*sigma[j]) xcur[j] = x[j]+0.9*sigma[j];
79 else if (xcur[j] < x[j]-0.9*sigma[j]) xcur[j] = x[j]-0.9*sigma[j];
84 denominv = 1.0 / (sigma2 - dx2);
85 val += (u * dx + v * dx2) * denominv;
87 /* update gval, wval, gcval (approximant functions) */
89 d->gval += (dfdx[j] * c + (fabs(dfdx[j])*sigma[j] + 0.5*rho) * dx2)
91 d->wval += 0.5 * dx2 * denominv;
92 for (i = 0; i < m; ++i)
93 gcval[i] += (dfcdx[i*n+j] * c + (fabs(dfcdx[i*n+j])*sigma[j]
98 /* gradient is easy to compute: since we are at a minimum x (dval/dx=0),
99 we only need the partial derivative with respect to y, and
100 we negate because we are maximizing: */
101 if (grad) for (i = 0; i < m; ++i) grad[i] = -gcval[i];
105 /***********************************************************************/
107 nlopt_result mma_minimize(int n, nlopt_func f, void *f_data,
108 int m, nlopt_func fc,
109 void *fc_data_, ptrdiff_t fc_datum_size,
110 const double *lb, const double *ub, /* bounds */
111 double *x, /* in: initial guess, out: minimizer */
113 nlopt_stopping *stop,
114 nlopt_algorithm dual_alg,
115 double dual_tolrel, int dual_maxeval)
117 nlopt_result ret = NLOPT_SUCCESS;
118 double *xcur, rho, *sigma, *dfdx, *dfdx_cur, *xprev, *xprevprev, fcur;
119 double *dfcdx, *dfcdx_cur;
120 double *fcval, *fcval_cur, *rhoc, *gcval, *y, *dual_lb, *dual_ub;
122 char *fc_data = (char *) fc_data_;
126 sigma = (double *) malloc(sizeof(double) * (6*n + 2*m*n + m*7));
127 if (!sigma) return NLOPT_OUT_OF_MEMORY;
132 xprevprev = xprev + n;
133 fcval = xprevprev + n;
134 fcval_cur = fcval + m;
135 rhoc = fcval_cur + m;
138 dual_ub = dual_lb + m;
141 dfcdx_cur = dfcdx + m*n;
155 for (j = 0; j < n; ++j) {
156 if (nlopt_isinf(ub[j]) || nlopt_isinf(lb[j]))
157 sigma[j] = 1.0; /* arbitrary default */
159 sigma[j] = 0.5 * (ub[j] - lb[j]);
162 for (i = 0; i < m; ++i) {
164 dual_lb[i] = y[i] = 0.0;
165 dual_ub[i] = HUGE_VAL;
168 dd.fval = fcur = *minf = f(n, x, dfdx, f_data);
170 memcpy(xcur, x, sizeof(double) * n);
173 for (i = 0; i < m; ++i) {
174 fcval[i] = fc(n, x, dfcdx + i*n, fc_data + fc_datum_size * i);
175 feasible = feasible && (fcval[i] <= 0);
177 if (!feasible) { ret = NLOPT_FAILURE; goto done; } /* TODO: handle this */
179 while (1) { /* outer iterations */
181 if (nlopt_stop_evals(stop)) ret = NLOPT_MAXEVAL_REACHED;
182 else if (nlopt_stop_time(stop)) ret = NLOPT_MAXTIME_REACHED;
183 else if (*minf < stop->minf_max) ret = NLOPT_MINF_MAX_REACHED;
184 if (ret != NLOPT_SUCCESS) goto done;
185 if (++k > 1) memcpy(xprevprev, xprev, sizeof(double) * n);
186 memcpy(xprev, xcur, sizeof(double) * n);
188 while (1) { /* inner iterations */
190 int feasible_cur, inner_done;
193 /* solve dual problem */
195 reti = nlopt_minimize(
196 dual_alg, m, dual_func, &dd,
197 dual_lb, dual_ub, y, &min_dual,
198 -HUGE_VAL, dual_tolrel,0., 0.,NULL, dual_maxeval,
199 stop->maxtime - (nlopt_seconds() - stop->start));
200 if (reti < 0 || reti == NLOPT_MAXTIME_REACHED) {
205 dual_func(m, y, NULL, &dd); /* evaluate final xcur etc. */
207 fcur = f(n, xcur, dfdx_cur, f_data);
209 inner_done = dd.gval >= fcur;
211 for (i = 0; i < m; ++i) {
212 fcval_cur[i] = fc(n, xcur, dfcdx_cur + i*n,
213 fc_data + fc_datum_size * i);
214 feasible_cur = feasible_cur && (fcval_cur[i] <= 0);
215 inner_done = inner_done && (dd.gcval[i] >= fcval_cur[i]);
218 if (fcur < *minf && (feasible_cur || !feasible)) {
219 feasible = feasible_cur;
220 dd.fval = *minf = fcur;
221 memcpy(fcval, fcval_cur, sizeof(double)*m);
222 memcpy(x, xcur, sizeof(double)*n);
223 memcpy(dfdx, dfdx_cur, sizeof(double)*n);
224 memcpy(dfcdx, dfcdx_cur, sizeof(double)*n*m);
226 if (nlopt_stop_evals(stop)) ret = NLOPT_MAXEVAL_REACHED;
227 else if (nlopt_stop_time(stop)) ret = NLOPT_MAXTIME_REACHED;
228 else if (*minf < stop->minf_max) ret = NLOPT_MINF_MAX_REACHED;
229 if (ret != NLOPT_SUCCESS) goto done;
231 if (inner_done) break;
234 rho = MIN(10*rho, 1.1 * (rho + (fcur-dd.gval) / dd.wval));
235 for (i = 0; i < m; ++i)
236 if (fcval_cur[i] > dd.gcval[i])
239 1.1 * (rhoc[i] + (fcval_cur[i]-dd.gcval[i])
243 printf("MMA inner iteration: rho -> %g\n", rho);
244 for (i = 0; i < mma_verbose; ++i)
245 printf(" rhoc[%d] -> %g\n", i,rhoc[i]);
248 if (nlopt_stop_ftol(stop, fcur, fprev))
249 ret = NLOPT_FTOL_REACHED;
250 if (nlopt_stop_x(stop, xcur, xprev))
251 ret = NLOPT_XTOL_REACHED;
252 if (ret != NLOPT_SUCCESS) goto done;
254 /* update rho and sigma for iteration k+1 */
255 rho = MAX(0.1 * rho, MMA_RHOMIN);
257 printf("MMA outer iteration: rho -> %g\n", rho);
258 for (i = 0; i < m; ++i)
259 rhoc[i] = MAX(0.1 * rhoc[i], MMA_RHOMIN);
260 for (i = 0; i < mma_verbose; ++i)
261 printf(" rhoc[%d] -> %g\n", i, rhoc[i]);
263 for (j = 0; j < n; ++j) {
264 double dx2 = (xcur[j]-xprev[j]) * (xprev[j]-xprevprev[j]);
265 double gam = dx2 < 0 ? 0.7 : (dx2 > 0 ? 1.2 : 1);
267 if (!nlopt_isinf(ub[j]) && !nlopt_isinf(lb[j])) {
268 sigma[j] = MIN(sigma[j], 10*(ub[j]-lb[j]));
269 sigma[j] = MAX(sigma[j], 0.01*(ub[j]-lb[j]));
272 for (j = 0; j < mma_verbose; ++j)
273 printf(" sigma[%d] -> %g\n",