1 | /* $Id: ClpSolve.cpp 1793 2011-09-10 07:35:23Z forrest $ */ |
2 | // Copyright (C) 2003, International Business Machines |
3 | // Corporation and others. All Rights Reserved. |
4 | // This code is licensed under the terms of the Eclipse Public License (EPL). |
5 | |
6 | // This file has higher level solve functions |
7 | |
8 | #include "CoinPragma.hpp" |
9 | #include "ClpConfig.h" |
10 | |
11 | // check already here if COIN_HAS_GLPK is defined, since we do not want to get confused by a COIN_HAS_GLPK in config_coinutils.h |
12 | #if defined(COIN_HAS_AMD) || defined(COIN_HAS_CHOLMOD) || defined(COIN_HAS_GLPK) |
13 | #define UFL_BARRIER |
14 | #endif |
15 | |
16 | #include <math.h> |
17 | |
18 | #include "CoinHelperFunctions.hpp" |
19 | #include "ClpHelperFunctions.hpp" |
20 | #include "CoinSort.hpp" |
21 | #include "ClpFactorization.hpp" |
22 | #include "ClpSimplex.hpp" |
23 | #include "ClpSimplexOther.hpp" |
24 | #include "ClpSimplexDual.hpp" |
25 | #ifndef SLIM_CLP |
26 | #include "ClpQuadraticObjective.hpp" |
27 | #include "ClpInterior.hpp" |
28 | #include "ClpCholeskyDense.hpp" |
29 | #include "ClpCholeskyBase.hpp" |
30 | #include "ClpPlusMinusOneMatrix.hpp" |
31 | #include "ClpNetworkMatrix.hpp" |
32 | #endif |
33 | #include "ClpEventHandler.hpp" |
34 | #include "ClpLinearObjective.hpp" |
35 | #include "ClpSolve.hpp" |
36 | #include "ClpPackedMatrix.hpp" |
37 | #include "ClpMessage.hpp" |
38 | #include "CoinTime.hpp" |
39 | |
40 | #include "ClpPresolve.hpp" |
41 | #ifndef SLIM_CLP |
42 | #include "Idiot.hpp" |
43 | #ifdef COIN_HAS_WSMP |
44 | #include "ClpCholeskyWssmp.hpp" |
45 | #include "ClpCholeskyWssmpKKT.hpp" |
46 | #endif |
47 | #include "ClpCholeskyUfl.hpp" |
48 | #ifdef TAUCS_BARRIER |
49 | #include "ClpCholeskyTaucs.hpp" |
50 | #endif |
51 | #include "ClpCholeskyMumps.hpp" |
52 | #ifdef COIN_HAS_VOL |
53 | #include "VolVolume.hpp" |
54 | #include "CoinHelperFunctions.hpp" |
55 | #include "CoinPackedMatrix.hpp" |
56 | #include "CoinMpsIO.hpp" |
57 | |
58 | //############################################################################# |
59 | |
60 | class lpHook : public VOL_user_hooks { |
61 | private: |
62 | lpHook(const lpHook&); |
63 | lpHook& operator=(const lpHook&); |
64 | private: |
65 | /// Pointer to dense vector of structural variable upper bounds |
66 | double *colupper_; |
67 | /// Pointer to dense vector of structural variable lower bounds |
68 | double *collower_; |
69 | /// Pointer to dense vector of objective coefficients |
70 | double *objcoeffs_; |
71 | /// Pointer to dense vector of right hand sides |
72 | double *rhs_; |
73 | /// Pointer to dense vector of senses |
74 | char *sense_; |
75 | |
76 | /// The problem matrix in a row ordered form |
77 | CoinPackedMatrix rowMatrix_; |
78 | /// The problem matrix in a column ordered form |
79 | CoinPackedMatrix colMatrix_; |
80 | |
81 | public: |
82 | lpHook(double* clb, double* cub, double* obj, |
83 | double* rhs, char* sense, const CoinPackedMatrix& mat); |
84 | virtual ~lpHook(); |
85 | |
86 | public: |
87 | // for all hooks: return value of -1 means that volume should quit |
88 | /** compute reduced costs |
89 | @param u (IN) the dual variables |
90 | @param rc (OUT) the reduced cost with respect to the dual values |
91 | */ |
92 | virtual int compute_rc(const VOL_dvector& u, VOL_dvector& rc); |
93 | |
94 | /** Solve the subproblem for the subgradient step. |
95 | @param dual (IN) the dual variables |
96 | @param rc (IN) the reduced cost with respect to the dual values |
97 | @param lcost (OUT) the lagrangean cost with respect to the dual values |
98 | @param x (OUT) the primal result of solving the subproblem |
99 | @param v (OUT) b-Ax for the relaxed constraints |
100 | @param pcost (OUT) the primal objective value of <code>x</code> |
101 | */ |
102 | virtual int solve_subproblem(const VOL_dvector& dual, const VOL_dvector& rc, |
103 | double& lcost, VOL_dvector& x, VOL_dvector& v, |
104 | double& pcost); |
105 | /** Starting from the primal vector x, run a heuristic to produce |
106 | an integer solution |
107 | @param x (IN) the primal vector |
108 | @param heur_val (OUT) the value of the integer solution (return |
109 | <code>DBL_MAX</code> here if no feas sol was found |
110 | */ |
111 | virtual int heuristics(const VOL_problem& p, |
112 | const VOL_dvector& x, double& heur_val) { |
113 | return 0; |
114 | } |
115 | }; |
116 | |
117 | //############################################################################# |
118 | |
119 | lpHook::lpHook(double* clb, double* cub, double* obj, |
120 | double* rhs, char* sense, |
121 | const CoinPackedMatrix& mat) |
122 | { |
123 | colupper_ = cub; |
124 | collower_ = clb; |
125 | objcoeffs_ = obj; |
126 | rhs_ = rhs; |
127 | sense_ = sense; |
128 | assert (mat.isColOrdered()); |
129 | colMatrix_.copyOf(mat); |
130 | rowMatrix_.reverseOrderedCopyOf(mat); |
131 | } |
132 | |
133 | //----------------------------------------------------------------------------- |
134 | |
135 | lpHook::~lpHook() |
136 | { |
137 | } |
138 | |
139 | //############################################################################# |
140 | |
141 | int |
142 | lpHook::compute_rc(const VOL_dvector& u, VOL_dvector& rc) |
143 | { |
144 | rowMatrix_.transposeTimes(u.v, rc.v); |
145 | const int psize = rowMatrix_.getNumCols(); |
146 | |
147 | for (int i = 0; i < psize; ++i) |
148 | rc[i] = objcoeffs_[i] - rc[i]; |
149 | return 0; |
150 | } |
151 | |
152 | //----------------------------------------------------------------------------- |
153 | |
154 | int |
155 | lpHook::solve_subproblem(const VOL_dvector& dual, const VOL_dvector& rc, |
156 | double& lcost, VOL_dvector& x, VOL_dvector& v, |
157 | double& pcost) |
158 | { |
159 | int i; |
160 | const int psize = x.size(); |
161 | const int dsize = v.size(); |
162 | |
163 | // compute the lagrangean solution corresponding to the reduced costs |
164 | for (i = 0; i < psize; ++i) |
165 | x[i] = (rc[i] >= 0.0) ? collower_[i] : colupper_[i]; |
166 | |
167 | // compute the lagrangean value (rhs*dual + primal*rc) |
168 | lcost = 0; |
169 | for (i = 0; i < dsize; ++i) |
170 | lcost += rhs_[i] * dual[i]; |
171 | for (i = 0; i < psize; ++i) |
172 | lcost += x[i] * rc[i]; |
173 | |
174 | // compute the rhs - lhs |
175 | colMatrix_.times(x.v, v.v); |
176 | for (i = 0; i < dsize; ++i) |
177 | v[i] = rhs_[i] - v[i]; |
178 | |
179 | // compute the lagrangean primal objective |
180 | pcost = 0; |
181 | for (i = 0; i < psize; ++i) |
182 | pcost += x[i] * objcoeffs_[i]; |
183 | |
184 | return 0; |
185 | } |
186 | |
187 | //############################################################################# |
188 | /** A quick inlined function to convert from lb/ub style constraint |
189 | definition to sense/rhs/range style */ |
190 | inline void |
191 | convertBoundToSense(const double lower, const double upper, |
192 | char& sense, double& right, |
193 | double& range) |
194 | { |
195 | range = 0.0; |
196 | if (lower > -1.0e20) { |
197 | if (upper < 1.0e20) { |
198 | right = upper; |
199 | if (upper == lower) { |
200 | sense = 'E'; |
201 | } else { |
202 | sense = 'R'; |
203 | range = upper - lower; |
204 | } |
205 | } else { |
206 | sense = 'G'; |
207 | right = lower; |
208 | } |
209 | } else { |
210 | if (upper < 1.0e20) { |
211 | sense = 'L'; |
212 | right = upper; |
213 | } else { |
214 | sense = 'N'; |
215 | right = 0.0; |
216 | } |
217 | } |
218 | } |
219 | |
220 | static int |
221 | solveWithVolume(ClpSimplex * model, int numberPasses, int doIdiot) |
222 | { |
223 | VOL_problem volprob; |
224 | volprob.parm.gap_rel_precision = 0.00001; |
225 | volprob.parm.maxsgriters = 3000; |
226 | if(numberPasses > 3000) { |
227 | volprob.parm.maxsgriters = numberPasses; |
228 | volprob.parm.primal_abs_precision = 0.0; |
229 | volprob.parm.minimum_rel_ascent = 0.00001; |
230 | } else if (doIdiot > 0) { |
231 | volprob.parm.maxsgriters = doIdiot; |
232 | } |
233 | if (model->logLevel() < 2) |
234 | volprob.parm.printflag = 0; |
235 | else |
236 | volprob.parm.printflag = 3; |
237 | const CoinPackedMatrix* mat = model->matrix(); |
238 | int psize = model->numberColumns(); |
239 | int dsize = model->numberRows(); |
240 | char * sense = new char[dsize]; |
241 | double * rhs = new double [dsize]; |
242 | |
243 | // Set the lb/ub on the duals |
244 | volprob.dsize = dsize; |
245 | volprob.psize = psize; |
246 | volprob.dual_lb.allocate(dsize); |
247 | volprob.dual_ub.allocate(dsize); |
248 | int i; |
249 | const double * rowLower = model->rowLower(); |
250 | const double * rowUpper = model->rowUpper(); |
251 | for (i = 0; i < dsize; ++i) { |
252 | double range; |
253 | convertBoundToSense(rowLower[i], rowUpper[i], |
254 | sense[i], rhs[i], range); |
255 | switch (sense[i]) { |
256 | case 'E': |
257 | volprob.dual_lb[i] = -1.0e31; |
258 | volprob.dual_ub[i] = 1.0e31; |
259 | break; |
260 | case 'L': |
261 | volprob.dual_lb[i] = -1.0e31; |
262 | volprob.dual_ub[i] = 0.0; |
263 | break; |
264 | case 'G': |
265 | volprob.dual_lb[i] = 0.0; |
266 | volprob.dual_ub[i] = 1.0e31; |
267 | break; |
268 | default: |
269 | printf("Volume Algorithm can't work if there is a non ELG row\n" ); |
270 | return 1; |
271 | } |
272 | } |
273 | // Check bounds |
274 | double * saveLower = model->columnLower(); |
275 | double * saveUpper = model->columnUpper(); |
276 | bool good = true; |
277 | for (i = 0; i < psize; i++) { |
278 | if (saveLower[i] < -1.0e20 || saveUpper[i] > 1.0e20) { |
279 | good = false; |
280 | break; |
281 | } |
282 | } |
283 | if (!good) { |
284 | saveLower = CoinCopyOfArray(model->columnLower(), psize); |
285 | saveUpper = CoinCopyOfArray(model->columnUpper(), psize); |
286 | for (i = 0; i < psize; i++) { |
287 | if (saveLower[i] < -1.0e20) |
288 | saveLower[i] = -1.0e20; |
289 | if(saveUpper[i] > 1.0e20) |
290 | saveUpper[i] = 1.0e20; |
291 | } |
292 | } |
293 | lpHook myHook(saveLower, saveUpper, |
294 | model->objective(), |
295 | rhs, sense, *mat); |
296 | |
297 | volprob.solve(myHook, false /* no warmstart */); |
298 | |
299 | if (saveLower != model->columnLower()) { |
300 | delete [] saveLower; |
301 | delete [] saveUpper; |
302 | } |
303 | //------------- extract the solution --------------------------- |
304 | |
305 | //printf("Best lagrangean value: %f\n", volprob.value); |
306 | |
307 | double avg = 0; |
308 | for (i = 0; i < dsize; ++i) { |
309 | switch (sense[i]) { |
310 | case 'E': |
311 | avg += CoinAbs(volprob.viol[i]); |
312 | break; |
313 | case 'L': |
314 | if (volprob.viol[i] < 0) |
315 | avg += (-volprob.viol[i]); |
316 | break; |
317 | case 'G': |
318 | if (volprob.viol[i] > 0) |
319 | avg += volprob.viol[i]; |
320 | break; |
321 | } |
322 | } |
323 | |
324 | //printf("Average primal constraint violation: %f\n", avg/dsize); |
325 | |
326 | // volprob.dsol contains the dual solution (dual feasible) |
327 | // volprob.psol contains the primal solution |
328 | // (NOT necessarily primal feasible) |
329 | CoinMemcpyN(volprob.dsol.v, dsize, model->dualRowSolution()); |
330 | CoinMemcpyN(volprob.psol.v, psize, model->primalColumnSolution()); |
331 | return 0; |
332 | } |
333 | #endif |
334 | static ClpInterior * currentModel2 = NULL; |
335 | #endif |
336 | //############################################################################# |
337 | // Allow for interrupts |
338 | // But is this threadsafe ? (so switched off by option) |
339 | |
340 | #include "CoinSignal.hpp" |
341 | static ClpSimplex * currentModel = NULL; |
342 | |
343 | extern "C" { |
344 | static void |
345 | #if defined(_MSC_VER) |
346 | __cdecl |
347 | #endif // _MSC_VER |
348 | signal_handler(int /*whichSignal*/) |
349 | { |
350 | if (currentModel != NULL) |
351 | currentModel->setMaximumIterations(0); // stop at next iterations |
352 | #ifndef SLIM_CLP |
353 | if (currentModel2 != NULL) |
354 | currentModel2->setMaximumBarrierIterations(0); // stop at next iterations |
355 | #endif |
356 | return; |
357 | } |
358 | } |
359 | |
360 | /** General solve algorithm which can do presolve |
361 | special options (bits) |
362 | 1 - do not perturb |
363 | 2 - do not scale |
364 | 4 - use crash (default allslack in dual, idiot in primal) |
365 | 8 - all slack basis in primal |
366 | 16 - switch off interrupt handling |
367 | 32 - do not try and make plus minus one matrix |
368 | 64 - do not use sprint even if problem looks good |
369 | */ |
370 | int |
371 | ClpSimplex::initialSolve(ClpSolve & options) |
372 | { |
373 | ClpSolve::SolveType method = options.getSolveType(); |
374 | //ClpSolve::SolveType originalMethod=method; |
375 | ClpSolve::PresolveType presolve = options.getPresolveType(); |
376 | int saveMaxIterations = maximumIterations(); |
377 | int finalStatus = -1; |
378 | int numberIterations = 0; |
379 | double time1 = CoinCpuTime(); |
380 | double timeX = time1; |
381 | double time2; |
382 | ClpMatrixBase * saveMatrix = NULL; |
383 | ClpObjective * savedObjective = NULL; |
384 | if (!objective_ || !matrix_) { |
385 | // totally empty |
386 | handler_->message(CLP_EMPTY_PROBLEM, messages_) |
387 | << 0 |
388 | << 0 |
389 | << 0 |
390 | << CoinMessageEol; |
391 | return -1; |
392 | } else if (!numberRows_ || !numberColumns_ || !getNumElements()) { |
393 | presolve = ClpSolve::presolveOff; |
394 | } |
395 | if (objective_->type() >= 2 && optimizationDirection_ == 0) { |
396 | // pretend linear |
397 | savedObjective = objective_; |
398 | // make up objective |
399 | double * obj = new double[numberColumns_]; |
400 | for (int i = 0; i < numberColumns_; i++) { |
401 | double l = fabs(columnLower_[i]); |
402 | double u = fabs(columnUpper_[i]); |
403 | obj[i] = 0.0; |
404 | if (CoinMin(l, u) < 1.0e20) { |
405 | if (l < u) |
406 | obj[i] = 1.0 + randomNumberGenerator_.randomDouble() * 1.0e-2; |
407 | else |
408 | obj[i] = -1.0 - randomNumberGenerator_.randomDouble() * 1.0e-2; |
409 | } |
410 | } |
411 | objective_ = new ClpLinearObjective(obj, numberColumns_); |
412 | delete [] obj; |
413 | } |
414 | ClpSimplex * model2 = this; |
415 | bool interrupt = (options.getSpecialOption(2) == 0); |
416 | CoinSighandler_t saveSignal = static_cast<CoinSighandler_t> (0); |
417 | if (interrupt) { |
418 | currentModel = model2; |
419 | // register signal handler |
420 | saveSignal = signal(SIGINT, signal_handler); |
421 | } |
422 | // If no status array - set up basis |
423 | if (!status_) |
424 | allSlackBasis(); |
425 | ClpPresolve * pinfo = new ClpPresolve(); |
426 | pinfo->setSubstitution(options.substitution()); |
427 | int presolveOptions = options.presolveActions(); |
428 | bool presolveToFile = (presolveOptions & 0x40000000) != 0; |
429 | presolveOptions &= ~0x40000000; |
430 | if ((presolveOptions & 0xffff) != 0) |
431 | pinfo->setPresolveActions(presolveOptions); |
432 | // switch off singletons to slacks |
433 | //pinfo->setDoSingletonColumn(false); // done by bits |
434 | int printOptions = options.getSpecialOption(5); |
435 | if ((printOptions & 1) != 0) |
436 | pinfo->statistics(); |
437 | double timePresolve = 0.0; |
438 | double timeIdiot = 0.0; |
439 | double timeCore = 0.0; |
440 | eventHandler()->event(ClpEventHandler::presolveStart); |
441 | int savePerturbation = perturbation_; |
442 | int saveScaling = scalingFlag_; |
443 | #ifndef SLIM_CLP |
444 | #ifndef NO_RTTI |
445 | if (dynamic_cast< ClpNetworkMatrix*>(matrix_)) { |
446 | // network - switch off stuff |
447 | presolve = ClpSolve::presolveOff; |
448 | } |
449 | #else |
450 | if (matrix_->type() == 11) { |
451 | // network - switch off stuff |
452 | presolve = ClpSolve::presolveOff; |
453 | } |
454 | #endif |
455 | #endif |
456 | if (presolve != ClpSolve::presolveOff) { |
457 | bool costedSlacks = false; |
458 | int numberPasses = 5; |
459 | if (presolve == ClpSolve::presolveNumber) { |
460 | numberPasses = options.getPresolvePasses(); |
461 | presolve = ClpSolve::presolveOn; |
462 | } else if (presolve == ClpSolve::presolveNumberCost) { |
463 | numberPasses = options.getPresolvePasses(); |
464 | presolve = ClpSolve::presolveOn; |
465 | costedSlacks = true; |
466 | // switch on singletons to slacks |
467 | pinfo->setDoSingletonColumn(true); |
468 | // gub stuff for testing |
469 | //pinfo->setDoGubrow(true); |
470 | } |
471 | #ifndef CLP_NO_STD |
472 | if (presolveToFile) { |
473 | // PreSolve to file - not fully tested |
474 | printf("Presolving to file - presolve.save\n" ); |
475 | pinfo->presolvedModelToFile(*this, "presolve.save" , dblParam_[ClpPresolveTolerance], |
476 | false, numberPasses); |
477 | model2 = this; |
478 | } else { |
479 | #endif |
480 | model2 = pinfo->presolvedModel(*this, dblParam_[ClpPresolveTolerance], |
481 | false, numberPasses, true, costedSlacks); |
482 | #ifndef CLP_NO_STD |
483 | } |
484 | #endif |
485 | time2 = CoinCpuTime(); |
486 | timePresolve = time2 - timeX; |
487 | handler_->message(CLP_INTERVAL_TIMING, messages_) |
488 | << "Presolve" << timePresolve << time2 - time1 |
489 | << CoinMessageEol; |
490 | timeX = time2; |
491 | if (!model2) { |
492 | handler_->message(CLP_INFEASIBLE, messages_) |
493 | << CoinMessageEol; |
494 | model2 = this; |
495 | eventHandler()->event(ClpEventHandler::presolveInfeasible); |
496 | problemStatus_ = pinfo->presolveStatus(); |
497 | if (options.infeasibleReturn() || (moreSpecialOptions_ & 1) != 0) { |
498 | delete pinfo; |
499 | return -1; |
500 | } |
501 | presolve = ClpSolve::presolveOff; |
502 | } else { |
503 | model2->eventHandler()->setSimplex(model2); |
504 | int rcode=model2->eventHandler()->event(ClpEventHandler::presolveSize); |
505 | // see if too big or small |
506 | if (rcode==2) { |
507 | delete model2; |
508 | delete pinfo; |
509 | return -2; |
510 | } else if (rcode==3) { |
511 | delete model2; |
512 | delete pinfo; |
513 | return -3; |
514 | } |
515 | } |
516 | model2->setMoreSpecialOptions(model2->moreSpecialOptions()&(~1024)); |
517 | model2->eventHandler()->setSimplex(model2); |
518 | // We may be better off using original (but if dual leave because of bounds) |
519 | if (presolve != ClpSolve::presolveOff && |
520 | numberRows_ < 1.01 * model2->numberRows_ && numberColumns_ < 1.01 * model2->numberColumns_ |
521 | && model2 != this) { |
522 | if(method != ClpSolve::useDual || |
523 | (numberRows_ == model2->numberRows_ && numberColumns_ == model2->numberColumns_)) { |
524 | delete model2; |
525 | model2 = this; |
526 | presolve = ClpSolve::presolveOff; |
527 | } |
528 | } |
529 | } |
530 | if (interrupt) |
531 | currentModel = model2; |
532 | // For below >0 overrides |
533 | // 0 means no, -1 means maybe |
534 | int doIdiot = 0; |
535 | int doCrash = 0; |
536 | int doSprint = 0; |
537 | int doSlp = 0; |
538 | int primalStartup = 1; |
539 | model2->eventHandler()->event(ClpEventHandler::presolveBeforeSolve); |
540 | bool tryItSave = false; |
541 | // switch to primal from automatic if just one cost entry |
542 | if (method == ClpSolve::automatic && model2->numberColumns() > 5000 && |
543 | (specialOptions_ & 1024) != 0) { |
544 | int numberColumns = model2->numberColumns(); |
545 | int numberRows = model2->numberRows(); |
546 | const double * obj = model2->objective(); |
547 | int nNon = 0; |
548 | for (int i = 0; i < numberColumns; i++) { |
549 | if (obj[i]) |
550 | nNon++; |
551 | } |
552 | if (nNon == 1) { |
553 | #ifdef COIN_DEVELOP |
554 | printf("Forcing primal\n" ); |
555 | #endif |
556 | method = ClpSolve::usePrimal; |
557 | tryItSave = numberRows > 200 && numberColumns > 2000 && |
558 | (numberColumns > 2 * numberRows || (specialOptions_ & 1024) != 0); |
559 | } |
560 | } |
561 | if (method != ClpSolve::useDual && method != ClpSolve::useBarrier |
562 | && method != ClpSolve::useBarrierNoCross) { |
563 | switch (options.getSpecialOption(1)) { |
564 | case 0: |
565 | doIdiot = -1; |
566 | doCrash = -1; |
567 | doSprint = -1; |
568 | break; |
569 | case 1: |
570 | doIdiot = 0; |
571 | doCrash = 1; |
572 | if (options.getExtraInfo(1) > 0) |
573 | doCrash = options.getExtraInfo(1); |
574 | doSprint = 0; |
575 | break; |
576 | case 2: |
577 | doIdiot = 1; |
578 | if (options.getExtraInfo(1) > 0) |
579 | doIdiot = options.getExtraInfo(1); |
580 | doCrash = 0; |
581 | doSprint = 0; |
582 | break; |
583 | case 3: |
584 | doIdiot = 0; |
585 | doCrash = 0; |
586 | doSprint = 1; |
587 | break; |
588 | case 4: |
589 | doIdiot = 0; |
590 | doCrash = 0; |
591 | doSprint = 0; |
592 | break; |
593 | case 5: |
594 | doIdiot = 0; |
595 | doCrash = -1; |
596 | doSprint = -1; |
597 | break; |
598 | case 6: |
599 | doIdiot = -1; |
600 | doCrash = -1; |
601 | doSprint = 0; |
602 | break; |
603 | case 7: |
604 | doIdiot = -1; |
605 | doCrash = 0; |
606 | doSprint = -1; |
607 | break; |
608 | case 8: |
609 | doIdiot = -1; |
610 | doCrash = 0; |
611 | doSprint = 0; |
612 | break; |
613 | case 9: |
614 | doIdiot = 0; |
615 | doCrash = 0; |
616 | doSprint = -1; |
617 | break; |
618 | case 10: |
619 | doIdiot = 0; |
620 | doCrash = 0; |
621 | doSprint = 0; |
622 | if (options.getExtraInfo(1) > 0) |
623 | doSlp = options.getExtraInfo(1); |
624 | break; |
625 | case 11: |
626 | doIdiot = 0; |
627 | doCrash = 0; |
628 | doSprint = 0; |
629 | primalStartup = 0; |
630 | break; |
631 | default: |
632 | abort(); |
633 | } |
634 | } else { |
635 | // Dual |
636 | switch (options.getSpecialOption(0)) { |
637 | case 0: |
638 | doIdiot = 0; |
639 | doCrash = 0; |
640 | doSprint = 0; |
641 | break; |
642 | case 1: |
643 | doIdiot = 0; |
644 | doCrash = 1; |
645 | if (options.getExtraInfo(0) > 0) |
646 | doCrash = options.getExtraInfo(0); |
647 | doSprint = 0; |
648 | break; |
649 | case 2: |
650 | doIdiot = -1; |
651 | if (options.getExtraInfo(0) > 0) |
652 | doIdiot = options.getExtraInfo(0); |
653 | doCrash = 0; |
654 | doSprint = 0; |
655 | break; |
656 | default: |
657 | abort(); |
658 | } |
659 | } |
660 | #ifndef NO_RTTI |
661 | ClpQuadraticObjective * quadraticObj = (dynamic_cast< ClpQuadraticObjective*>(objectiveAsObject())); |
662 | #else |
663 | ClpQuadraticObjective * quadraticObj = NULL; |
664 | if (objective_->type() == 2) |
665 | quadraticObj = (static_cast< ClpQuadraticObjective*>(objective_)); |
666 | #endif |
667 | // If quadratic then primal or barrier or slp |
668 | if (quadraticObj) { |
669 | doSprint = 0; |
670 | doIdiot = 0; |
671 | // off |
672 | if (method == ClpSolve::useBarrier) |
673 | method = ClpSolve::useBarrierNoCross; |
674 | else if (method != ClpSolve::useBarrierNoCross) |
675 | method = ClpSolve::usePrimal; |
676 | } |
677 | #ifdef COIN_HAS_VOL |
678 | // Save number of idiot |
679 | int saveDoIdiot = doIdiot; |
680 | #endif |
681 | // Just do this number of passes in Sprint |
682 | int maxSprintPass = 100; |
683 | // See if worth trying +- one matrix |
684 | bool plusMinus = false; |
685 | int numberElements = model2->getNumElements(); |
686 | #ifndef SLIM_CLP |
687 | #ifndef NO_RTTI |
688 | if (dynamic_cast< ClpNetworkMatrix*>(matrix_)) { |
689 | // network - switch off stuff |
690 | doIdiot = 0; |
691 | if (doSprint < 0) |
692 | doSprint = 0; |
693 | } |
694 | #else |
695 | if (matrix_->type() == 11) { |
696 | // network - switch off stuff |
697 | doIdiot = 0; |
698 | //doSprint=0; |
699 | } |
700 | #endif |
701 | #endif |
702 | int numberColumns = model2->numberColumns(); |
703 | int numberRows = model2->numberRows(); |
704 | // If not all slack basis - switch off all except sprint |
705 | int numberRowsBasic = 0; |
706 | int iRow; |
707 | for (iRow = 0; iRow < numberRows; iRow++) |
708 | if (model2->getRowStatus(iRow) == basic) |
709 | numberRowsBasic++; |
710 | if (numberRowsBasic < numberRows) { |
711 | doIdiot = 0; |
712 | doCrash = 0; |
713 | //doSprint=0; |
714 | } |
715 | if (options.getSpecialOption(3) == 0) { |
716 | if(numberElements > 100000) |
717 | plusMinus = true; |
718 | if(numberElements > 10000 && (doIdiot || doSprint)) |
719 | plusMinus = true; |
720 | } else if ((specialOptions_ & 1024) != 0) { |
721 | plusMinus = true; |
722 | } |
723 | #ifndef SLIM_CLP |
724 | // Statistics (+1,-1, other) - used to decide on strategy if not +-1 |
725 | CoinBigIndex statistics[3] = { -1, 0, 0}; |
726 | if (plusMinus) { |
727 | saveMatrix = model2->clpMatrix(); |
728 | #ifndef NO_RTTI |
729 | ClpPackedMatrix* clpMatrix = |
730 | dynamic_cast< ClpPackedMatrix*>(saveMatrix); |
731 | #else |
732 | ClpPackedMatrix* clpMatrix = NULL; |
733 | if (saveMatrix->type() == 1) |
734 | clpMatrix = |
735 | static_cast< ClpPackedMatrix*>(saveMatrix); |
736 | #endif |
737 | if (clpMatrix) { |
738 | ClpPlusMinusOneMatrix * newMatrix = new ClpPlusMinusOneMatrix(*(clpMatrix->matrix())); |
739 | if (newMatrix->getIndices()) { |
740 | // CHECKME This test of specialOptions and the one above |
741 | // don't seem compatible. |
742 | if ((specialOptions_ & 1024) == 0) { |
743 | model2->replaceMatrix(newMatrix); |
744 | } else { |
745 | // in integer - just use for sprint/idiot |
746 | saveMatrix = NULL; |
747 | delete newMatrix; |
748 | } |
749 | } else { |
750 | handler_->message(CLP_MATRIX_CHANGE, messages_) |
751 | << "+- 1" |
752 | << CoinMessageEol; |
753 | CoinMemcpyN(newMatrix->startPositive(), 3, statistics); |
754 | saveMatrix = NULL; |
755 | plusMinus = false; |
756 | delete newMatrix; |
757 | } |
758 | } else { |
759 | saveMatrix = NULL; |
760 | plusMinus = false; |
761 | } |
762 | } |
763 | #endif |
764 | if (this->factorizationFrequency() == 200) { |
765 | // User did not touch preset |
766 | model2->defaultFactorizationFrequency(); |
767 | } else if (model2 != this) { |
768 | // make sure model2 has correct value |
769 | model2->setFactorizationFrequency(this->factorizationFrequency()); |
770 | } |
771 | if (method == ClpSolve::automatic) { |
772 | if (doSprint == 0 && doIdiot == 0) { |
773 | // off |
774 | method = ClpSolve::useDual; |
775 | } else { |
776 | // only do primal if sprint or idiot |
777 | if (doSprint > 0) { |
778 | method = ClpSolve::usePrimalorSprint; |
779 | } else if (doIdiot > 0) { |
780 | method = ClpSolve::usePrimal; |
781 | } else { |
782 | if (numberElements < 500000) { |
783 | // Small problem |
784 | if(numberRows * 10 > numberColumns || numberColumns < 6000 |
785 | || (numberRows * 20 > numberColumns && !plusMinus)) |
786 | doSprint = 0; // switch off sprint |
787 | } else { |
788 | // larger problem |
789 | if(numberRows * 8 > numberColumns) |
790 | doSprint = 0; // switch off sprint |
791 | } |
792 | // switch off sprint or idiot if any free variable |
793 | int iColumn; |
794 | double * columnLower = model2->columnLower(); |
795 | double * columnUpper = model2->columnUpper(); |
796 | for (iColumn = 0; iColumn < numberColumns; iColumn++) { |
797 | if (columnLower[iColumn] < -1.0e10 && columnUpper[iColumn] > 1.0e10) { |
798 | doSprint = 0; |
799 | doIdiot = 0; |
800 | break; |
801 | } |
802 | } |
803 | int nPasses = 0; |
804 | // look at rhs |
805 | int iRow; |
806 | double largest = 0.0; |
807 | double smallest = 1.0e30; |
808 | double largestGap = 0.0; |
809 | int numberNotE = 0; |
810 | bool notInteger = false; |
811 | for (iRow = 0; iRow < numberRows; iRow++) { |
812 | double value1 = model2->rowLower_[iRow]; |
813 | if (value1 && value1 > -1.0e31) { |
814 | largest = CoinMax(largest, fabs(value1)); |
815 | smallest = CoinMin(smallest, fabs(value1)); |
816 | if (fabs(value1 - floor(value1 + 0.5)) > 1.0e-8) { |
817 | notInteger = true; |
818 | break; |
819 | } |
820 | } |
821 | double value2 = model2->rowUpper_[iRow]; |
822 | if (value2 && value2 < 1.0e31) { |
823 | largest = CoinMax(largest, fabs(value2)); |
824 | smallest = CoinMin(smallest, fabs(value2)); |
825 | if (fabs(value2 - floor(value2 + 0.5)) > 1.0e-8) { |
826 | notInteger = true; |
827 | break; |
828 | } |
829 | } |
830 | // CHECKME This next bit can't be right... |
831 | if (value2 > value1) { |
832 | numberNotE++; |
833 | if (value2 > 1.0e31 || value1 < -1.0e31) |
834 | largestGap = COIN_DBL_MAX; |
835 | else |
836 | largestGap = value2 - value1; |
837 | } |
838 | } |
839 | bool tryIt = numberRows > 200 && numberColumns > 2000 && |
840 | (numberColumns > 2 * numberRows || (method != ClpSolve::useDual && (specialOptions_ & 1024) != 0)); |
841 | tryItSave = tryIt; |
842 | if (numberRows < 1000 && numberColumns < 3000) |
843 | tryIt = false; |
844 | if (notInteger) |
845 | tryIt = false; |
846 | if (largest / smallest > 10 || (largest / smallest > 2.0 && largest > 50)) |
847 | tryIt = false; |
848 | if (tryIt) { |
849 | if (largest / smallest > 2.0) { |
850 | nPasses = 10 + numberColumns / 100000; |
851 | nPasses = CoinMin(nPasses, 50); |
852 | nPasses = CoinMax(nPasses, 15); |
853 | if (numberRows > 20000 && nPasses > 5) { |
854 | // Might as well go for it |
855 | nPasses = CoinMax(nPasses, 71); |
856 | } else if (numberRows > 2000 && nPasses > 5) { |
857 | nPasses = CoinMax(nPasses, 50); |
858 | } else if (numberElements < 3 * numberColumns) { |
859 | nPasses = CoinMin(nPasses, 10); // probably not worh it |
860 | } |
861 | } else if (largest / smallest > 1.01 || numberElements <= 3 * numberColumns) { |
862 | nPasses = 10 + numberColumns / 1000; |
863 | nPasses = CoinMin(nPasses, 100); |
864 | nPasses = CoinMax(nPasses, 30); |
865 | if (numberRows > 25000) { |
866 | // Might as well go for it |
867 | nPasses = CoinMax(nPasses, 71); |
868 | } |
869 | if (!largestGap) |
870 | nPasses *= 2; |
871 | } else { |
872 | nPasses = 10 + numberColumns / 1000; |
873 | nPasses = CoinMin(nPasses, 200); |
874 | nPasses = CoinMax(nPasses, 100); |
875 | if (!largestGap) |
876 | nPasses *= 2; |
877 | } |
878 | } |
879 | //printf("%d rows %d cols plus %c tryIt %c largest %g smallest %g largestGap %g npasses %d sprint %c\n", |
880 | // numberRows,numberColumns,plusMinus ? 'Y' : 'N', |
881 | // tryIt ? 'Y' :'N',largest,smallest,largestGap,nPasses,doSprint ? 'Y' :'N'); |
882 | //exit(0); |
883 | if (!tryIt || nPasses <= 5) |
884 | doIdiot = 0; |
885 | if (doSprint) { |
886 | method = ClpSolve::usePrimalorSprint; |
887 | } else if (doIdiot) { |
888 | method = ClpSolve::usePrimal; |
889 | } else { |
890 | method = ClpSolve::useDual; |
891 | } |
892 | } |
893 | } |
894 | } |
895 | if (method == ClpSolve::usePrimalorSprint) { |
896 | if (doSprint < 0) { |
897 | if (numberElements < 500000) { |
898 | // Small problem |
899 | if(numberRows * 10 > numberColumns || numberColumns < 6000 |
900 | || (numberRows * 20 > numberColumns && !plusMinus)) |
901 | method = ClpSolve::usePrimal; // switch off sprint |
902 | } else { |
903 | // larger problem |
904 | if(numberRows * 8 > numberColumns) |
905 | method = ClpSolve::usePrimal; // switch off sprint |
906 | // but make lightweight |
907 | if(numberRows * 10 > numberColumns || numberColumns < 6000 |
908 | || (numberRows * 20 > numberColumns && !plusMinus)) |
909 | maxSprintPass = 10; |
910 | } |
911 | } else if (doSprint == 0) { |
912 | method = ClpSolve::usePrimal; // switch off sprint |
913 | } |
914 | } |
915 | if (method == ClpSolve::useDual) { |
916 | double * saveLower = NULL; |
917 | double * saveUpper = NULL; |
918 | if (presolve == ClpSolve::presolveOn) { |
919 | int numberInfeasibilities = model2->tightenPrimalBounds(0.0, 0); |
920 | if (numberInfeasibilities) { |
921 | handler_->message(CLP_INFEASIBLE, messages_) |
922 | << CoinMessageEol; |
923 | delete model2; |
924 | model2 = this; |
925 | presolve = ClpSolve::presolveOff; |
926 | } |
927 | } else if (numberRows_ + numberColumns_ > 5000) { |
928 | // do anyway |
929 | saveLower = new double[numberRows_+numberColumns_]; |
930 | CoinMemcpyN(model2->columnLower(), numberColumns_, saveLower); |
931 | CoinMemcpyN(model2->rowLower(), numberRows_, saveLower + numberColumns_); |
932 | saveUpper = new double[numberRows_+numberColumns_]; |
933 | CoinMemcpyN(model2->columnUpper(), numberColumns_, saveUpper); |
934 | CoinMemcpyN(model2->rowUpper(), numberRows_, saveUpper + numberColumns_); |
935 | int numberInfeasibilities = model2->tightenPrimalBounds(); |
936 | if (numberInfeasibilities) { |
937 | handler_->message(CLP_INFEASIBLE, messages_) |
938 | << CoinMessageEol; |
939 | CoinMemcpyN(saveLower, numberColumns_, model2->columnLower()); |
940 | CoinMemcpyN(saveLower + numberColumns_, numberRows_, model2->rowLower()); |
941 | delete [] saveLower; |
942 | saveLower = NULL; |
943 | CoinMemcpyN(saveUpper, numberColumns_, model2->columnUpper()); |
944 | CoinMemcpyN(saveUpper + numberColumns_, numberRows_, model2->rowUpper()); |
945 | delete [] saveUpper; |
946 | saveUpper = NULL; |
947 | } |
948 | } |
949 | #ifndef COIN_HAS_VOL |
950 | // switch off idiot and volume for now |
951 | doIdiot = 0; |
952 | #endif |
953 | // pick up number passes |
954 | int nPasses = 0; |
955 | int numberNotE = 0; |
956 | #ifndef SLIM_CLP |
957 | if ((doIdiot < 0 && plusMinus) || doIdiot > 0) { |
958 | // See if candidate for idiot |
959 | nPasses = 0; |
960 | Idiot info(*model2); |
961 | // Get average number of elements per column |
962 | double ratio = static_cast<double> (numberElements) / static_cast<double> (numberColumns); |
963 | // look at rhs |
964 | int iRow; |
965 | double largest = 0.0; |
966 | double smallest = 1.0e30; |
967 | double largestGap = 0.0; |
968 | for (iRow = 0; iRow < numberRows; iRow++) { |
969 | double value1 = model2->rowLower_[iRow]; |
970 | if (value1 && value1 > -1.0e31) { |
971 | largest = CoinMax(largest, fabs(value1)); |
972 | smallest = CoinMin(smallest, fabs(value1)); |
973 | } |
974 | double value2 = model2->rowUpper_[iRow]; |
975 | if (value2 && value2 < 1.0e31) { |
976 | largest = CoinMax(largest, fabs(value2)); |
977 | smallest = CoinMin(smallest, fabs(value2)); |
978 | } |
979 | if (value2 > value1) { |
980 | numberNotE++; |
981 | if (value2 > 1.0e31 || value1 < -1.0e31) |
982 | largestGap = COIN_DBL_MAX; |
983 | else |
984 | largestGap = value2 - value1; |
985 | } |
986 | } |
987 | if (doIdiot < 0) { |
988 | if (numberRows > 200 && numberColumns > 5000 && ratio >= 3.0 && |
989 | largest / smallest < 1.1 && !numberNotE) { |
990 | nPasses = 71; |
991 | } |
992 | } |
993 | if (doIdiot > 0) { |
994 | nPasses = CoinMax(nPasses, doIdiot); |
995 | if (nPasses > 70) { |
996 | info.setStartingWeight(1.0e3); |
997 | info.setDropEnoughFeasibility(0.01); |
998 | } |
999 | } |
1000 | if (nPasses > 20) { |
1001 | #ifdef COIN_HAS_VOL |
1002 | int returnCode = solveWithVolume(model2, nPasses, saveDoIdiot); |
1003 | if (!returnCode) { |
1004 | time2 = CoinCpuTime(); |
1005 | timeIdiot = time2 - timeX; |
1006 | handler_->message(CLP_INTERVAL_TIMING, messages_) |
1007 | << "Idiot Crash" << timeIdiot << time2 - time1 |
1008 | << CoinMessageEol; |
1009 | timeX = time2; |
1010 | } else { |
1011 | nPasses = 0; |
1012 | } |
1013 | #else |
1014 | nPasses = 0; |
1015 | #endif |
1016 | } else { |
1017 | nPasses = 0; |
1018 | } |
1019 | } |
1020 | #endif |
1021 | if (doCrash) { |
1022 | switch(doCrash) { |
1023 | // standard |
1024 | case 1: |
1025 | model2->crash(1000, 1); |
1026 | break; |
1027 | // As in paper by Solow and Halim (approx) |
1028 | case 2: |
1029 | case 3: |
1030 | model2->crash(model2->dualBound(), 0); |
1031 | break; |
1032 | // Just put free in basis |
1033 | case 4: |
1034 | model2->crash(0.0, 3); |
1035 | break; |
1036 | } |
1037 | } |
1038 | if (!nPasses) { |
1039 | int saveOptions = model2->specialOptions(); |
1040 | if (model2->numberRows() > 100) |
1041 | model2->setSpecialOptions(saveOptions | 64); // go as far as possible |
1042 | //int numberRows = model2->numberRows(); |
1043 | //int numberColumns = model2->numberColumns(); |
1044 | if (dynamic_cast< ClpPackedMatrix*>(matrix_)) { |
1045 | // See if original wanted vector |
1046 | ClpPackedMatrix * clpMatrixO = dynamic_cast< ClpPackedMatrix*>(matrix_); |
1047 | ClpMatrixBase * matrix = model2->clpMatrix(); |
1048 | if (dynamic_cast< ClpPackedMatrix*>(matrix) && clpMatrixO->wantsSpecialColumnCopy()) { |
1049 | ClpPackedMatrix * clpMatrix = dynamic_cast< ClpPackedMatrix*>(matrix); |
1050 | clpMatrix->makeSpecialColumnCopy(); |
1051 | //model2->setSpecialOptions(model2->specialOptions()|256); // to say no row copy for comparisons |
1052 | model2->dual(0); |
1053 | clpMatrix->releaseSpecialColumnCopy(); |
1054 | } else { |
1055 | model2->dual(0); |
1056 | } |
1057 | } else { |
1058 | model2->dual(0); |
1059 | } |
1060 | } else if (!numberNotE && 0) { |
1061 | // E so we can do in another way |
1062 | double * pi = model2->dualRowSolution(); |
1063 | int i; |
1064 | int numberColumns = model2->numberColumns(); |
1065 | int numberRows = model2->numberRows(); |
1066 | double * saveObj = new double[numberColumns]; |
1067 | CoinMemcpyN(model2->objective(), numberColumns, saveObj); |
1068 | CoinMemcpyN(model2->objective(), |
1069 | numberColumns, model2->dualColumnSolution()); |
1070 | model2->clpMatrix()->transposeTimes(-1.0, pi, model2->dualColumnSolution()); |
1071 | CoinMemcpyN(model2->dualColumnSolution(), |
1072 | numberColumns, model2->objective()); |
1073 | const double * rowsol = model2->primalRowSolution(); |
1074 | double offset = 0.0; |
1075 | for (i = 0; i < numberRows; i++) { |
1076 | offset += pi[i] * rowsol[i]; |
1077 | } |
1078 | double value2; |
1079 | model2->getDblParam(ClpObjOffset, value2); |
1080 | //printf("Offset %g %g\n",offset,value2); |
1081 | model2->setDblParam(ClpObjOffset, value2 - offset); |
1082 | model2->setPerturbation(51); |
1083 | //model2->setRowObjective(pi); |
1084 | // zero out pi |
1085 | //memset(pi,0,numberRows*sizeof(double)); |
1086 | // Could put some in basis - only partially tested |
1087 | model2->allSlackBasis(); |
1088 | //model2->factorization()->maximumPivots(200); |
1089 | //model2->setLogLevel(63); |
1090 | // solve |
1091 | model2->dual(0); |
1092 | model2->setDblParam(ClpObjOffset, value2); |
1093 | CoinMemcpyN(saveObj, numberColumns, model2->objective()); |
1094 | // zero out pi |
1095 | //memset(pi,0,numberRows*sizeof(double)); |
1096 | //model2->setRowObjective(pi); |
1097 | delete [] saveObj; |
1098 | //model2->dual(0); |
1099 | model2->setPerturbation(50); |
1100 | model2->primal(); |
1101 | } else { |
1102 | // solve |
1103 | model2->setPerturbation(100); |
1104 | model2->dual(2); |
1105 | model2->setPerturbation(50); |
1106 | model2->dual(0); |
1107 | } |
1108 | if (saveLower) { |
1109 | CoinMemcpyN(saveLower, numberColumns_, model2->columnLower()); |
1110 | CoinMemcpyN(saveLower + numberColumns_, numberRows_, model2->rowLower()); |
1111 | delete [] saveLower; |
1112 | saveLower = NULL; |
1113 | CoinMemcpyN(saveUpper, numberColumns_, model2->columnUpper()); |
1114 | CoinMemcpyN(saveUpper + numberColumns_, numberRows_, model2->rowUpper()); |
1115 | delete [] saveUpper; |
1116 | saveUpper = NULL; |
1117 | } |
1118 | time2 = CoinCpuTime(); |
1119 | timeCore = time2 - timeX; |
1120 | handler_->message(CLP_INTERVAL_TIMING, messages_) |
1121 | << "Dual" << timeCore << time2 - time1 |
1122 | << CoinMessageEol; |
1123 | timeX = time2; |
1124 | } else if (method == ClpSolve::usePrimal) { |
1125 | #ifndef SLIM_CLP |
1126 | if (doIdiot) { |
1127 | int nPasses = 0; |
1128 | Idiot info(*model2); |
1129 | // Get average number of elements per column |
1130 | double ratio = static_cast<double> (numberElements) / static_cast<double> (numberColumns); |
1131 | // look at rhs |
1132 | int iRow; |
1133 | double largest = 0.0; |
1134 | double smallest = 1.0e30; |
1135 | double largestGap = 0.0; |
1136 | int numberNotE = 0; |
1137 | for (iRow = 0; iRow < numberRows; iRow++) { |
1138 | double value1 = model2->rowLower_[iRow]; |
1139 | if (value1 && value1 > -1.0e31) { |
1140 | largest = CoinMax(largest, fabs(value1)); |
1141 | smallest = CoinMin(smallest, fabs(value1)); |
1142 | } |
1143 | double value2 = model2->rowUpper_[iRow]; |
1144 | if (value2 && value2 < 1.0e31) { |
1145 | largest = CoinMax(largest, fabs(value2)); |
1146 | smallest = CoinMin(smallest, fabs(value2)); |
1147 | } |
1148 | if (value2 > value1) { |
1149 | numberNotE++; |
1150 | if (value2 > 1.0e31 || value1 < -1.0e31) |
1151 | largestGap = COIN_DBL_MAX; |
1152 | else |
1153 | largestGap = value2 - value1; |
1154 | } |
1155 | } |
1156 | bool increaseSprint = plusMinus; |
1157 | if ((specialOptions_ & 1024) != 0) |
1158 | increaseSprint = false; |
1159 | if (!plusMinus) { |
1160 | // If 90% +- 1 then go for sprint |
1161 | if (statistics[0] >= 0 && 10 * statistics[2] < statistics[0] + statistics[1]) |
1162 | increaseSprint = true; |
1163 | } |
1164 | bool tryIt = tryItSave; |
1165 | if (numberRows < 1000 && numberColumns < 3000) |
1166 | tryIt = false; |
1167 | if (tryIt) { |
1168 | if (increaseSprint) { |
1169 | info.setStartingWeight(1.0e3); |
1170 | info.setReduceIterations(6); |
1171 | // also be more lenient on infeasibilities |
1172 | info.setDropEnoughFeasibility(0.5 * info.getDropEnoughFeasibility()); |
1173 | info.setDropEnoughWeighted(-2.0); |
1174 | if (largest / smallest > 2.0) { |
1175 | nPasses = 10 + numberColumns / 100000; |
1176 | nPasses = CoinMin(nPasses, 50); |
1177 | nPasses = CoinMax(nPasses, 15); |
1178 | if (numberRows > 20000 && nPasses > 5) { |
1179 | // Might as well go for it |
1180 | nPasses = CoinMax(nPasses, 71); |
1181 | } else if (numberRows > 2000 && nPasses > 5) { |
1182 | nPasses = CoinMax(nPasses, 50); |
1183 | } else if (numberElements < 3 * numberColumns) { |
1184 | nPasses = CoinMin(nPasses, 10); // probably not worh it |
1185 | if (doIdiot < 0) |
1186 | info.setLightweight(1); // say lightweight idiot |
1187 | } else { |
1188 | if (doIdiot < 0) |
1189 | info.setLightweight(1); // say lightweight idiot |
1190 | } |
1191 | } else if (largest / smallest > 1.01 || numberElements <= 3 * numberColumns) { |
1192 | nPasses = 10 + numberColumns / 1000; |
1193 | nPasses = CoinMin(nPasses, 100); |
1194 | nPasses = CoinMax(nPasses, 30); |
1195 | if (numberRows > 25000) { |
1196 | // Might as well go for it |
1197 | nPasses = CoinMax(nPasses, 71); |
1198 | } |
1199 | if (!largestGap) |
1200 | nPasses *= 2; |
1201 | } else { |
1202 | nPasses = 10 + numberColumns / 1000; |
1203 | nPasses = CoinMin(nPasses, 200); |
1204 | nPasses = CoinMax(nPasses, 100); |
1205 | info.setStartingWeight(1.0e-1); |
1206 | info.setReduceIterations(6); |
1207 | if (!largestGap) |
1208 | nPasses *= 2; |
1209 | //info.setFeasibilityTolerance(1.0e-7); |
1210 | } |
1211 | // If few passes - don't bother |
1212 | if (nPasses <= 5 && !plusMinus) |
1213 | nPasses = 0; |
1214 | } else { |
1215 | if (doIdiot < 0) |
1216 | info.setLightweight(1); // say lightweight idiot |
1217 | if (largest / smallest > 1.01 || numberNotE || statistics[2] > statistics[0] + statistics[1]) { |
1218 | if (numberRows > 25000 || numberColumns > 5 * numberRows) { |
1219 | nPasses = 50; |
1220 | } else if (numberColumns > 4 * numberRows) { |
1221 | nPasses = 20; |
1222 | } else { |
1223 | nPasses = 5; |
1224 | } |
1225 | } else { |
1226 | if (numberRows > 25000 || numberColumns > 5 * numberRows) { |
1227 | nPasses = 50; |
1228 | info.setLightweight(0); // say not lightweight idiot |
1229 | } else if (numberColumns > 4 * numberRows) { |
1230 | nPasses = 20; |
1231 | } else { |
1232 | nPasses = 15; |
1233 | } |
1234 | } |
1235 | if (ratio < 3.0) { |
1236 | nPasses = static_cast<int> (ratio * static_cast<double> (nPasses) / 4.0); // probably not worth it |
1237 | } else { |
1238 | nPasses = CoinMax(nPasses, 5); |
1239 | } |
1240 | if (numberRows > 25000 && nPasses > 5) { |
1241 | // Might as well go for it |
1242 | nPasses = CoinMax(nPasses, 71); |
1243 | } else if (increaseSprint) { |
1244 | nPasses *= 2; |
1245 | nPasses = CoinMin(nPasses, 71); |
1246 | } else if (nPasses == 5 && ratio > 5.0) { |
1247 | nPasses = static_cast<int> (static_cast<double> (nPasses) * (ratio / 5.0)); // increase if lots of elements per column |
1248 | } |
1249 | if (nPasses <= 5 && !plusMinus) |
1250 | nPasses = 0; |
1251 | //info.setStartingWeight(1.0e-1); |
1252 | } |
1253 | } |
1254 | if (doIdiot > 0) { |
1255 | // pick up number passes |
1256 | nPasses = options.getExtraInfo(1) % 1000000; |
1257 | if (nPasses > 70) { |
1258 | info.setStartingWeight(1.0e3); |
1259 | info.setReduceIterations(6); |
1260 | if (nPasses >= 5000) { |
1261 | int k = nPasses % 100; |
1262 | nPasses /= 100; |
1263 | info.setReduceIterations(3); |
1264 | if (k) |
1265 | info.setStartingWeight(1.0e2); |
1266 | } |
1267 | // also be more lenient on infeasibilities |
1268 | info.setDropEnoughFeasibility(0.5 * info.getDropEnoughFeasibility()); |
1269 | info.setDropEnoughWeighted(-2.0); |
1270 | } else if (nPasses >= 50) { |
1271 | info.setStartingWeight(1.0e3); |
1272 | //info.setReduceIterations(6); |
1273 | } |
1274 | // For experimenting |
1275 | if (nPasses < 70 && (nPasses % 10) > 0 && (nPasses % 10) < 4) { |
1276 | info.setStartingWeight(1.0e3); |
1277 | info.setLightweight(nPasses % 10); // special testing |
1278 | #ifdef COIN_DEVELOP |
1279 | printf("warning - odd lightweight %d\n" , nPasses % 10); |
1280 | //info.setReduceIterations(6); |
1281 | #endif |
1282 | } |
1283 | } |
1284 | if (options.getExtraInfo(1) > 1000000) |
1285 | nPasses += 1000000; |
1286 | if (nPasses) { |
1287 | doCrash = 0; |
1288 | #if 0 |
1289 | double * solution = model2->primalColumnSolution(); |
1290 | int iColumn; |
1291 | double * saveLower = new double[numberColumns]; |
1292 | CoinMemcpyN(model2->columnLower(), numberColumns, saveLower); |
1293 | double * saveUpper = new double[numberColumns]; |
1294 | CoinMemcpyN(model2->columnUpper(), numberColumns, saveUpper); |
1295 | printf("doing tighten before idiot\n" ); |
1296 | model2->tightenPrimalBounds(); |
1297 | // Move solution |
1298 | double * columnLower = model2->columnLower(); |
1299 | double * columnUpper = model2->columnUpper(); |
1300 | for (iColumn = 0; iColumn < numberColumns; iColumn++) { |
1301 | if (columnLower[iColumn] > 0.0) |
1302 | solution[iColumn] = columnLower[iColumn]; |
1303 | else if (columnUpper[iColumn] < 0.0) |
1304 | solution[iColumn] = columnUpper[iColumn]; |
1305 | else |
1306 | solution[iColumn] = 0.0; |
1307 | } |
1308 | CoinMemcpyN(saveLower, numberColumns, columnLower); |
1309 | CoinMemcpyN(saveUpper, numberColumns, columnUpper); |
1310 | delete [] saveLower; |
1311 | delete [] saveUpper; |
1312 | #else |
1313 | // Allow for crossover |
1314 | //#define LACI_TRY |
1315 | #ifndef LACI_TRY |
1316 | //if (doIdiot>0) |
1317 | info.setStrategy(512 | info.getStrategy()); |
1318 | #endif |
1319 | // Allow for scaling |
1320 | info.setStrategy(32 | info.getStrategy()); |
1321 | info.crash(nPasses, model2->messageHandler(), model2->messagesPointer()); |
1322 | #endif |
1323 | time2 = CoinCpuTime(); |
1324 | timeIdiot = time2 - timeX; |
1325 | handler_->message(CLP_INTERVAL_TIMING, messages_) |
1326 | << "Idiot Crash" << timeIdiot << time2 - time1 |
1327 | << CoinMessageEol; |
1328 | timeX = time2; |
1329 | } |
1330 | } |
1331 | #endif |
1332 | // ? |
1333 | if (doCrash) { |
1334 | switch(doCrash) { |
1335 | // standard |
1336 | case 1: |
1337 | model2->crash(1000, 1); |
1338 | break; |
1339 | // As in paper by Solow and Halim (approx) |
1340 | case 2: |
1341 | model2->crash(model2->dualBound(), 0); |
1342 | break; |
1343 | // My take on it |
1344 | case 3: |
1345 | model2->crash(model2->dualBound(), -1); |
1346 | break; |
1347 | // Just put free in basis |
1348 | case 4: |
1349 | model2->crash(0.0, 3); |
1350 | break; |
1351 | } |
1352 | } |
1353 | #ifndef SLIM_CLP |
1354 | if (doSlp > 0 && objective_->type() == 2) { |
1355 | model2->nonlinearSLP(doSlp, 1.0e-5); |
1356 | } |
1357 | #endif |
1358 | #ifndef LACI_TRY |
1359 | if (options.getSpecialOption(1) != 2 || |
1360 | options.getExtraInfo(1) < 1000000) { |
1361 | if (dynamic_cast< ClpPackedMatrix*>(matrix_)) { |
1362 | // See if original wanted vector |
1363 | ClpPackedMatrix * clpMatrixO = dynamic_cast< ClpPackedMatrix*>(matrix_); |
1364 | ClpMatrixBase * matrix = model2->clpMatrix(); |
1365 | if (dynamic_cast< ClpPackedMatrix*>(matrix) && clpMatrixO->wantsSpecialColumnCopy()) { |
1366 | ClpPackedMatrix * clpMatrix = dynamic_cast< ClpPackedMatrix*>(matrix); |
1367 | clpMatrix->makeSpecialColumnCopy(); |
1368 | //model2->setSpecialOptions(model2->specialOptions()|256); // to say no row copy for comparisons |
1369 | model2->primal(primalStartup); |
1370 | clpMatrix->releaseSpecialColumnCopy(); |
1371 | } else { |
1372 | model2->primal(primalStartup); |
1373 | } |
1374 | } else { |
1375 | model2->primal(primalStartup); |
1376 | } |
1377 | } |
1378 | #endif |
1379 | time2 = CoinCpuTime(); |
1380 | timeCore = time2 - timeX; |
1381 | handler_->message(CLP_INTERVAL_TIMING, messages_) |
1382 | << "Primal" << timeCore << time2 - time1 |
1383 | << CoinMessageEol; |
1384 | timeX = time2; |
1385 | } else if (method == ClpSolve::usePrimalorSprint) { |
1386 | // Sprint |
1387 | /* |
1388 | This driver implements what I called Sprint when I introduced the idea |
1389 | many years ago. Cplex calls it "sifting" which I think is just as silly. |
1390 | When I thought of this trivial idea |
1391 | it reminded me of an LP code of the 60's called sprint which after |
1392 | every factorization took a subset of the matrix into memory (all |
1393 | 64K words!) and then iterated very fast on that subset. On the |
1394 | problems of those days it did not work very well, but it worked very |
1395 | well on aircrew scheduling problems where there were very large numbers |
1396 | of columns all with the same flavor. |
1397 | */ |
1398 | |
1399 | /* The idea works best if you can get feasible easily. To make it |
1400 | more general we can add in costed slacks */ |
1401 | |
1402 | int originalNumberColumns = model2->numberColumns(); |
1403 | int numberRows = model2->numberRows(); |
1404 | ClpSimplex * originalModel2 = model2; |
1405 | |
1406 | // We will need arrays to choose variables. These are too big but .. |
1407 | double * weight = new double [numberRows+originalNumberColumns]; |
1408 | int * sort = new int [numberRows+originalNumberColumns]; |
1409 | int numberSort = 0; |
1410 | // We are going to add slacks to get feasible. |
1411 | // initial list will just be artificials |
1412 | int iColumn; |
1413 | const double * columnLower = model2->columnLower(); |
1414 | const double * columnUpper = model2->columnUpper(); |
1415 | double * columnSolution = model2->primalColumnSolution(); |
1416 | |
1417 | // See if we have costed slacks |
1418 | int * negSlack = new int[numberRows]; |
1419 | int * posSlack = new int[numberRows]; |
1420 | int iRow; |
1421 | for (iRow = 0; iRow < numberRows; iRow++) { |
1422 | negSlack[iRow] = -1; |
1423 | posSlack[iRow] = -1; |
1424 | } |
1425 | const double * element = model2->matrix()->getElements(); |
1426 | const int * row = model2->matrix()->getIndices(); |
1427 | const CoinBigIndex * columnStart = model2->matrix()->getVectorStarts(); |
1428 | const int * columnLength = model2->matrix()->getVectorLengths(); |
1429 | //bool allSlack = (numberRowsBasic==numberRows); |
1430 | for (iColumn = 0; iColumn < originalNumberColumns; iColumn++) { |
1431 | if (!columnSolution[iColumn] || fabs(columnSolution[iColumn]) > 1.0e20) { |
1432 | double value = 0.0; |
1433 | if (columnLower[iColumn] > 0.0) |
1434 | value = columnLower[iColumn]; |
1435 | else if (columnUpper[iColumn] < 0.0) |
1436 | value = columnUpper[iColumn]; |
1437 | columnSolution[iColumn] = value; |
1438 | } |
1439 | if (columnLength[iColumn] == 1) { |
1440 | int jRow = row[columnStart[iColumn]]; |
1441 | if (!columnLower[iColumn]) { |
1442 | if (element[columnStart[iColumn]] > 0.0 && posSlack[jRow] < 0) |
1443 | posSlack[jRow] = iColumn; |
1444 | else if (element[columnStart[iColumn]] < 0.0 && negSlack[jRow] < 0) |
1445 | negSlack[jRow] = iColumn; |
1446 | } else if (!columnUpper[iColumn]) { |
1447 | if (element[columnStart[iColumn]] < 0.0 && posSlack[jRow] < 0) |
1448 | posSlack[jRow] = iColumn; |
1449 | else if (element[columnStart[iColumn]] > 0.0 && negSlack[jRow] < 0) |
1450 | negSlack[jRow] = iColumn; |
1451 | } |
1452 | } |
1453 | } |
1454 | // now see what that does to row solution |
1455 | double * rowSolution = model2->primalRowSolution(); |
1456 | CoinZeroN (rowSolution, numberRows); |
1457 | model2->clpMatrix()->times(1.0, columnSolution, rowSolution); |
1458 | // See if we can adjust using costed slacks |
1459 | double penalty = CoinMax(1.0e5, CoinMin(infeasibilityCost_ * 0.01, 1.0e10)) * optimizationDirection_; |
1460 | const double * lower = model2->rowLower(); |
1461 | const double * upper = model2->rowUpper(); |
1462 | for (iRow = 0; iRow < numberRows; iRow++) { |
1463 | if (lower[iRow] > rowSolution[iRow] + 1.0e-8) { |
1464 | int jColumn = posSlack[iRow]; |
1465 | if (jColumn >= 0) { |
1466 | if (columnSolution[jColumn]) |
1467 | continue; |
1468 | double difference = lower[iRow] - rowSolution[iRow]; |
1469 | double elementValue = element[columnStart[jColumn]]; |
1470 | if (elementValue > 0.0) { |
1471 | double movement = CoinMin(difference / elementValue, columnUpper[jColumn]); |
1472 | columnSolution[jColumn] = movement; |
1473 | rowSolution[iRow] += movement * elementValue; |
1474 | } else { |
1475 | double movement = CoinMax(difference / elementValue, columnLower[jColumn]); |
1476 | columnSolution[jColumn] = movement; |
1477 | rowSolution[iRow] += movement * elementValue; |
1478 | } |
1479 | } |
1480 | } else if (upper[iRow] < rowSolution[iRow] - 1.0e-8) { |
1481 | int jColumn = negSlack[iRow]; |
1482 | if (jColumn >= 0) { |
1483 | if (columnSolution[jColumn]) |
1484 | continue; |
1485 | double difference = upper[iRow] - rowSolution[iRow]; |
1486 | double elementValue = element[columnStart[jColumn]]; |
1487 | if (elementValue < 0.0) { |
1488 | double movement = CoinMin(difference / elementValue, columnUpper[jColumn]); |
1489 | columnSolution[jColumn] = movement; |
1490 | rowSolution[iRow] += movement * elementValue; |
1491 | } else { |
1492 | double movement = CoinMax(difference / elementValue, columnLower[jColumn]); |
1493 | columnSolution[jColumn] = movement; |
1494 | rowSolution[iRow] += movement * elementValue; |
1495 | } |
1496 | } |
1497 | } |
1498 | } |
1499 | delete [] negSlack; |
1500 | delete [] posSlack; |
1501 | int nRow = numberRows; |
1502 | bool network = false; |
1503 | if (dynamic_cast< ClpNetworkMatrix*>(matrix_)) { |
1504 | network = true; |
1505 | nRow *= 2; |
1506 | } |
1507 | int * addStarts = new int [nRow+1]; |
1508 | int * addRow = new int[nRow]; |
1509 | double * addElement = new double[nRow]; |
1510 | addStarts[0] = 0; |
1511 | int numberArtificials = 0; |
1512 | int numberAdd = 0; |
1513 | double * addCost = new double [numberRows]; |
1514 | for (iRow = 0; iRow < numberRows; iRow++) { |
1515 | if (lower[iRow] > rowSolution[iRow] + 1.0e-8) { |
1516 | addRow[numberAdd] = iRow; |
1517 | addElement[numberAdd++] = 1.0; |
1518 | if (network) { |
1519 | addRow[numberAdd] = numberRows; |
1520 | addElement[numberAdd++] = -1.0; |
1521 | } |
1522 | addCost[numberArtificials] = penalty; |
1523 | numberArtificials++; |
1524 | addStarts[numberArtificials] = numberAdd; |
1525 | } else if (upper[iRow] < rowSolution[iRow] - 1.0e-8) { |
1526 | addRow[numberAdd] = iRow; |
1527 | addElement[numberAdd++] = -1.0; |
1528 | if (network) { |
1529 | addRow[numberAdd] = numberRows; |
1530 | addElement[numberAdd++] = 1.0; |
1531 | } |
1532 | addCost[numberArtificials] = penalty; |
1533 | numberArtificials++; |
1534 | addStarts[numberArtificials] = numberAdd; |
1535 | } |
1536 | } |
1537 | if (numberArtificials) { |
1538 | // need copy so as not to disturb original |
1539 | model2 = new ClpSimplex(*model2); |
1540 | if (network) { |
1541 | // network - add a null row |
1542 | model2->addRow(0, NULL, NULL, -COIN_DBL_MAX, COIN_DBL_MAX); |
1543 | numberRows++; |
1544 | } |
1545 | model2->addColumns(numberArtificials, NULL, NULL, addCost, |
1546 | addStarts, addRow, addElement); |
1547 | } |
1548 | delete [] addStarts; |
1549 | delete [] addRow; |
1550 | delete [] addElement; |
1551 | delete [] addCost; |
1552 | // look at rhs to see if to perturb |
1553 | double largest = 0.0; |
1554 | double smallest = 1.0e30; |
1555 | for (iRow = 0; iRow < numberRows; iRow++) { |
1556 | double value; |
1557 | value = fabs(model2->rowLower_[iRow]); |
1558 | if (value && value < 1.0e30) { |
1559 | largest = CoinMax(largest, value); |
1560 | smallest = CoinMin(smallest, value); |
1561 | } |
1562 | value = fabs(model2->rowUpper_[iRow]); |
1563 | if (value && value < 1.0e30) { |
1564 | largest = CoinMax(largest, value); |
1565 | smallest = CoinMin(smallest, value); |
1566 | } |
1567 | } |
1568 | double * saveLower = NULL; |
1569 | double * saveUpper = NULL; |
1570 | if (largest < 2.01 * smallest) { |
1571 | // perturb - so switch off standard |
1572 | model2->setPerturbation(100); |
1573 | saveLower = new double[numberRows]; |
1574 | CoinMemcpyN(model2->rowLower_, numberRows, saveLower); |
1575 | saveUpper = new double[numberRows]; |
1576 | CoinMemcpyN(model2->rowUpper_, numberRows, saveUpper); |
1577 | double * lower = model2->rowLower(); |
1578 | double * upper = model2->rowUpper(); |
1579 | for (iRow = 0; iRow < numberRows; iRow++) { |
1580 | double lowerValue = lower[iRow], upperValue = upper[iRow]; |
1581 | double value = randomNumberGenerator_.randomDouble(); |
1582 | if (upperValue > lowerValue + primalTolerance_) { |
1583 | if (lowerValue > -1.0e20 && lowerValue) |
1584 | lowerValue -= value * 1.0e-4 * fabs(lowerValue); |
1585 | if (upperValue < 1.0e20 && upperValue) |
1586 | upperValue += value * 1.0e-4 * fabs(upperValue); |
1587 | } else if (upperValue > 0.0) { |
1588 | upperValue -= value * 1.0e-4 * fabs(lowerValue); |
1589 | lowerValue -= value * 1.0e-4 * fabs(lowerValue); |
1590 | } else if (upperValue < 0.0) { |
1591 | upperValue += value * 1.0e-4 * fabs(lowerValue); |
1592 | lowerValue += value * 1.0e-4 * fabs(lowerValue); |
1593 | } else { |
1594 | } |
1595 | lower[iRow] = lowerValue; |
1596 | upper[iRow] = upperValue; |
1597 | } |
1598 | } |
1599 | int i; |
1600 | // Just do this number of passes in Sprint |
1601 | if (doSprint > 0) |
1602 | maxSprintPass = options.getExtraInfo(1); |
1603 | // but if big use to get ratio |
1604 | double ratio = 3; |
1605 | if (maxSprintPass > 1000) { |
1606 | ratio = static_cast<double> (maxSprintPass) * 0.0001; |
1607 | ratio = CoinMax(ratio, 1.1); |
1608 | maxSprintPass = maxSprintPass % 1000; |
1609 | #ifdef COIN_DEVELOP |
1610 | printf("%d passes wanted with ratio of %g\n" , maxSprintPass, ratio); |
1611 | #endif |
1612 | } |
1613 | // Just take this number of columns in small problem |
1614 | int smallNumberColumns = static_cast<int> (CoinMin(ratio * numberRows, static_cast<double> (numberColumns))); |
1615 | smallNumberColumns = CoinMax(smallNumberColumns, 3000); |
1616 | smallNumberColumns = CoinMin(smallNumberColumns, numberColumns); |
1617 | //int smallNumberColumns = CoinMin(12*numberRows/10,numberColumns); |
1618 | //smallNumberColumns = CoinMax(smallNumberColumns,3000); |
1619 | //smallNumberColumns = CoinMax(smallNumberColumns,numberRows+1000); |
1620 | // redo as may have changed |
1621 | columnLower = model2->columnLower(); |
1622 | columnUpper = model2->columnUpper(); |
1623 | columnSolution = model2->primalColumnSolution(); |
1624 | // Set up initial list |
1625 | numberSort = 0; |
1626 | if (numberArtificials) { |
1627 | numberSort = numberArtificials; |
1628 | for (i = 0; i < numberSort; i++) |
1629 | sort[i] = i + originalNumberColumns; |
1630 | } |
1631 | // maybe a solution there already |
1632 | for (iColumn = 0; iColumn < originalNumberColumns; iColumn++) { |
1633 | if (model2->getColumnStatus(iColumn) == basic) |
1634 | sort[numberSort++] = iColumn; |
1635 | } |
1636 | for (iColumn = 0; iColumn < originalNumberColumns; iColumn++) { |
1637 | if (model2->getColumnStatus(iColumn) != basic) { |
1638 | if (columnSolution[iColumn] > columnLower[iColumn] && |
1639 | columnSolution[iColumn] < columnUpper[iColumn] && |
1640 | columnSolution[iColumn]) |
1641 | sort[numberSort++] = iColumn; |
1642 | } |
1643 | } |
1644 | numberSort = CoinMin(numberSort, smallNumberColumns); |
1645 | |
1646 | int numberColumns = model2->numberColumns(); |
1647 | double * fullSolution = model2->primalColumnSolution(); |
1648 | |
1649 | |
1650 | int iPass; |
1651 | double lastObjective = 1.0e31; |
1652 | // It will be safe to allow dense |
1653 | model2->setInitialDenseFactorization(true); |
1654 | |
1655 | // We will be using all rows |
1656 | int * whichRows = new int [numberRows]; |
1657 | for (iRow = 0; iRow < numberRows; iRow++) |
1658 | whichRows[iRow] = iRow; |
1659 | double originalOffset; |
1660 | model2->getDblParam(ClpObjOffset, originalOffset); |
1661 | int totalIterations = 0; |
1662 | double lastSumArtificials = COIN_DBL_MAX; |
1663 | int originalMaxSprintPass = maxSprintPass; |
1664 | maxSprintPass = 20; // so we do that many if infeasible |
1665 | for (iPass = 0; iPass < maxSprintPass; iPass++) { |
1666 | //printf("Bug until submodel new version\n"); |
1667 | //CoinSort_2(sort,sort+numberSort,weight); |
1668 | // Create small problem |
1669 | ClpSimplex small(model2, numberRows, whichRows, numberSort, sort); |
1670 | small.setPerturbation(model2->perturbation()); |
1671 | small.setInfeasibilityCost(model2->infeasibilityCost()); |
1672 | if (model2->factorizationFrequency() == 200) { |
1673 | // User did not touch preset |
1674 | small.defaultFactorizationFrequency(); |
1675 | } |
1676 | // now see what variables left out do to row solution |
1677 | double * rowSolution = model2->primalRowSolution(); |
1678 | double * sumFixed = new double[numberRows]; |
1679 | CoinZeroN (sumFixed, numberRows); |
1680 | int iRow, iColumn; |
1681 | // zero out ones in small problem |
1682 | for (iColumn = 0; iColumn < numberSort; iColumn++) { |
1683 | int kColumn = sort[iColumn]; |
1684 | fullSolution[kColumn] = 0.0; |
1685 | } |
1686 | // Get objective offset |
1687 | double offset = 0.0; |
1688 | const double * objective = model2->objective(); |
1689 | for (iColumn = 0; iColumn < numberColumns; iColumn++) |
1690 | offset += fullSolution[iColumn] * objective[iColumn]; |
1691 | small.setDblParam(ClpObjOffset, originalOffset - offset); |
1692 | model2->clpMatrix()->times(1.0, fullSolution, sumFixed); |
1693 | |
1694 | double * lower = small.rowLower(); |
1695 | double * upper = small.rowUpper(); |
1696 | for (iRow = 0; iRow < numberRows; iRow++) { |
1697 | if (lower[iRow] > -1.0e50) |
1698 | lower[iRow] -= sumFixed[iRow]; |
1699 | if (upper[iRow] < 1.0e50) |
1700 | upper[iRow] -= sumFixed[iRow]; |
1701 | rowSolution[iRow] -= sumFixed[iRow]; |
1702 | } |
1703 | delete [] sumFixed; |
1704 | // Solve |
1705 | if (interrupt) |
1706 | currentModel = &small; |
1707 | small.defaultFactorizationFrequency(); |
1708 | if (dynamic_cast< ClpPackedMatrix*>(matrix_)) { |
1709 | // See if original wanted vector |
1710 | ClpPackedMatrix * clpMatrixO = dynamic_cast< ClpPackedMatrix*>(matrix_); |
1711 | ClpMatrixBase * matrix = small.clpMatrix(); |
1712 | if (dynamic_cast< ClpPackedMatrix*>(matrix) && clpMatrixO->wantsSpecialColumnCopy()) { |
1713 | ClpPackedMatrix * clpMatrix = dynamic_cast< ClpPackedMatrix*>(matrix); |
1714 | clpMatrix->makeSpecialColumnCopy(); |
1715 | small.primal(1); |
1716 | clpMatrix->releaseSpecialColumnCopy(); |
1717 | } else { |
1718 | #if 1 |
1719 | small.primal(1); |
1720 | #else |
1721 | int numberColumns = small.numberColumns(); |
1722 | int numberRows = small.numberRows(); |
1723 | // Use dual region |
1724 | double * rhs = small.dualRowSolution(); |
1725 | int * whichRow = new int[3*numberRows]; |
1726 | int * whichColumn = new int[2*numberColumns]; |
1727 | int nBound; |
1728 | ClpSimplex * small2 = ((ClpSimplexOther *) (&small))->crunch(rhs, whichRow, whichColumn, |
1729 | nBound, false, false); |
1730 | if (small2) { |
1731 | small2->primal(1); |
1732 | if (small2->problemStatus() == 0) { |
1733 | small.setProblemStatus(0); |
1734 | ((ClpSimplexOther *) (&small))->afterCrunch(*small2, whichRow, whichColumn, nBound); |
1735 | } else { |
1736 | small2->primal(1); |
1737 | if (small2->problemStatus()) |
1738 | small.primal(1); |
1739 | } |
1740 | delete small2; |
1741 | } else { |
1742 | small.primal(1); |
1743 | } |
1744 | delete [] whichRow; |
1745 | delete [] whichColumn; |
1746 | #endif |
1747 | } |
1748 | } else { |
1749 | small.primal(1); |
1750 | } |
1751 | totalIterations += small.numberIterations(); |
1752 | // move solution back |
1753 | const double * solution = small.primalColumnSolution(); |
1754 | for (iColumn = 0; iColumn < numberSort; iColumn++) { |
1755 | int kColumn = sort[iColumn]; |
1756 | model2->setColumnStatus(kColumn, small.getColumnStatus(iColumn)); |
1757 | fullSolution[kColumn] = solution[iColumn]; |
1758 | } |
1759 | for (iRow = 0; iRow < numberRows; iRow++) |
1760 | model2->setRowStatus(iRow, small.getRowStatus(iRow)); |
1761 | CoinMemcpyN(small.primalRowSolution(), |
1762 | numberRows, model2->primalRowSolution()); |
1763 | double sumArtificials = 0.0; |
1764 | for (i = 0; i < numberArtificials; i++) |
1765 | sumArtificials += fullSolution[i + originalNumberColumns]; |
1766 | if (sumArtificials && iPass > 5 && sumArtificials >= lastSumArtificials) { |
1767 | // increase costs |
1768 | double * cost = model2->objective() + originalNumberColumns; |
1769 | double newCost = CoinMin(1.0e10, cost[0] * 1.5); |
1770 | for (i = 0; i < numberArtificials; i++) |
1771 | cost[i] = newCost; |
1772 | } |
1773 | lastSumArtificials = sumArtificials; |
1774 | // get reduced cost for large problem |
1775 | double * djs = model2->dualColumnSolution(); |
1776 | CoinMemcpyN(model2->objective(), numberColumns, djs); |
1777 | model2->clpMatrix()->transposeTimes(-1.0, small.dualRowSolution(), djs); |
1778 | int numberNegative = 0; |
1779 | double sumNegative = 0.0; |
1780 | // now massage weight so all basic in plus good djs |
1781 | // first count and do basic |
1782 | numberSort = 0; |
1783 | for (iColumn = 0; iColumn < numberColumns; iColumn++) { |
1784 | double dj = djs[iColumn] * optimizationDirection_; |
1785 | double value = fullSolution[iColumn]; |
1786 | if (model2->getColumnStatus(iColumn) == ClpSimplex::basic) { |
1787 | sort[numberSort++] = iColumn; |
1788 | } else if (dj < -dualTolerance_ && value < columnUpper[iColumn]) { |
1789 | numberNegative++; |
1790 | sumNegative -= dj; |
1791 | } else if (dj > dualTolerance_ && value > columnLower[iColumn]) { |
1792 | numberNegative++; |
1793 | sumNegative += dj; |
1794 | } |
1795 | } |
1796 | handler_->message(CLP_SPRINT, messages_) |
1797 | << iPass + 1 << small.numberIterations() << small.objectiveValue() << sumNegative |
1798 | << numberNegative |
1799 | << CoinMessageEol; |
1800 | if (sumArtificials < 1.0e-8 && originalMaxSprintPass >= 0) { |
1801 | maxSprintPass = iPass + originalMaxSprintPass; |
1802 | originalMaxSprintPass = -1; |
1803 | } |
1804 | if (iPass > 20) |
1805 | sumArtificials = 0.0; |
1806 | if ((small.objectiveValue()*optimizationDirection_ > lastObjective - 1.0e-7 && iPass > 5 && sumArtificials < 1.0e-8) || |
1807 | (!small.numberIterations() && iPass) || |
1808 | iPass == maxSprintPass - 1 || small.status() == 3) { |
1809 | |
1810 | break; // finished |
1811 | } else { |
1812 | lastObjective = small.objectiveValue() * optimizationDirection_; |
1813 | double tolerance; |
1814 | double averageNegDj = sumNegative / static_cast<double> (numberNegative + 1); |
1815 | if (numberNegative + numberSort > smallNumberColumns) |
1816 | tolerance = -dualTolerance_; |
1817 | else |
1818 | tolerance = 10.0 * averageNegDj; |
1819 | int saveN = numberSort; |
1820 | for (iColumn = 0; iColumn < numberColumns; iColumn++) { |
1821 | double dj = djs[iColumn] * optimizationDirection_; |
1822 | double value = fullSolution[iColumn]; |
1823 | if (model2->getColumnStatus(iColumn) != ClpSimplex::basic) { |
1824 | if (dj < -dualTolerance_ && value < columnUpper[iColumn]) |
1825 | dj = dj; |
1826 | else if (dj > dualTolerance_ && value > columnLower[iColumn]) |
1827 | dj = -dj; |
1828 | else if (columnUpper[iColumn] > columnLower[iColumn]) |
1829 | dj = fabs(dj); |
1830 | else |
1831 | dj = 1.0e50; |
1832 | if (dj < tolerance) { |
1833 | weight[numberSort] = dj; |
1834 | sort[numberSort++] = iColumn; |
1835 | } |
1836 | } |
1837 | } |
1838 | // sort |
1839 | CoinSort_2(weight + saveN, weight + numberSort, sort + saveN); |
1840 | numberSort = CoinMin(smallNumberColumns, numberSort); |
1841 | } |
1842 | } |
1843 | if (interrupt) |
1844 | currentModel = model2; |
1845 | for (i = 0; i < numberArtificials; i++) |
1846 | sort[i] = i + originalNumberColumns; |
1847 | model2->deleteColumns(numberArtificials, sort); |
1848 | if (network) { |
1849 | int iRow = numberRows - 1; |
1850 | model2->deleteRows(1, &iRow); |
1851 | } |
1852 | delete [] weight; |
1853 | delete [] sort; |
1854 | delete [] whichRows; |
1855 | if (saveLower) { |
1856 | // unperturb and clean |
1857 | for (iRow = 0; iRow < numberRows; iRow++) { |
1858 | double diffLower = saveLower[iRow] - model2->rowLower_[iRow]; |
1859 | double diffUpper = saveUpper[iRow] - model2->rowUpper_[iRow]; |
1860 | model2->rowLower_[iRow] = saveLower[iRow]; |
1861 | model2->rowUpper_[iRow] = saveUpper[iRow]; |
1862 | if (diffLower) |
1863 | assert (!diffUpper || fabs(diffLower - diffUpper) < 1.0e-5); |
1864 | else |
1865 | diffLower = diffUpper; |
1866 | model2->rowActivity_[iRow] += diffLower; |
1867 | } |
1868 | delete [] saveLower; |
1869 | delete [] saveUpper; |
1870 | } |
1871 | model2->primal(1); |
1872 | model2->setPerturbation(savePerturbation); |
1873 | if (model2 != originalModel2) { |
1874 | originalModel2->moveInfo(*model2); |
1875 | delete model2; |
1876 | model2 = originalModel2; |
1877 | } |
1878 | time2 = CoinCpuTime(); |
1879 | timeCore = time2 - timeX; |
1880 | handler_->message(CLP_INTERVAL_TIMING, messages_) |
1881 | << "Sprint" << timeCore << time2 - time1 |
1882 | << CoinMessageEol; |
1883 | timeX = time2; |
1884 | model2->setNumberIterations(model2->numberIterations() + totalIterations); |
1885 | } else if (method == ClpSolve::useBarrier || method == ClpSolve::useBarrierNoCross) { |
1886 | #ifndef SLIM_CLP |
1887 | //printf("***** experimental pretty crude barrier\n"); |
1888 | //#define SAVEIT 2 |
1889 | #ifndef SAVEIT |
1890 | #define BORROW |
1891 | #endif |
1892 | #ifdef BORROW |
1893 | ClpInterior barrier; |
1894 | barrier.borrowModel(*model2); |
1895 | #else |
1896 | ClpInterior barrier(*model2); |
1897 | #endif |
1898 | if (interrupt) |
1899 | currentModel2 = &barrier; |
1900 | int barrierOptions = options.getSpecialOption(4); |
1901 | int aggressiveGamma = 0; |
1902 | bool presolveInCrossover = false; |
1903 | bool scale = false; |
1904 | bool doKKT = false; |
1905 | bool forceFixing = false; |
1906 | int speed = 0; |
1907 | if (barrierOptions & 16) { |
1908 | barrierOptions &= ~16; |
1909 | doKKT = true; |
1910 | } |
1911 | if (barrierOptions&(32 + 64 + 128)) { |
1912 | aggressiveGamma = (barrierOptions & (32 + 64 + 128)) >> 5; |
1913 | barrierOptions &= ~(32 + 64 + 128); |
1914 | } |
1915 | if (barrierOptions & 256) { |
1916 | barrierOptions &= ~256; |
1917 | presolveInCrossover = true; |
1918 | } |
1919 | if (barrierOptions & 512) { |
1920 | barrierOptions &= ~512; |
1921 | forceFixing = true; |
1922 | } |
1923 | if (barrierOptions & 1024) { |
1924 | barrierOptions &= ~1024; |
1925 | barrier.setProjectionTolerance(1.0e-9); |
1926 | } |
1927 | if (barrierOptions&(2048 | 4096)) { |
1928 | speed = (barrierOptions & (2048 | 4096)) >> 11; |
1929 | barrierOptions &= ~(2048 | 4096); |
1930 | } |
1931 | if (barrierOptions & 8) { |
1932 | barrierOptions &= ~8; |
1933 | scale = true; |
1934 | } |
1935 | // If quadratic force KKT |
1936 | if (quadraticObj) { |
1937 | doKKT = true; |
1938 | } |
1939 | switch (barrierOptions) { |
1940 | case 0: |
1941 | default: |
1942 | if (!doKKT) { |
1943 | ClpCholeskyBase * cholesky = new ClpCholeskyBase(); |
1944 | cholesky->setIntegerParameter(0, speed); |
1945 | barrier.setCholesky(cholesky); |
1946 | } else { |
1947 | ClpCholeskyBase * cholesky = new ClpCholeskyBase(); |
1948 | cholesky->setKKT(true); |
1949 | barrier.setCholesky(cholesky); |
1950 | } |
1951 | break; |
1952 | case 1: |
1953 | if (!doKKT) { |
1954 | ClpCholeskyDense * cholesky = new ClpCholeskyDense(); |
1955 | barrier.setCholesky(cholesky); |
1956 | } else { |
1957 | ClpCholeskyDense * cholesky = new ClpCholeskyDense(); |
1958 | cholesky->setKKT(true); |
1959 | barrier.setCholesky(cholesky); |
1960 | } |
1961 | break; |
1962 | #ifdef COIN_HAS_WSMP |
1963 | case 2: { |
1964 | ClpCholeskyWssmp * cholesky = new ClpCholeskyWssmp(CoinMax(100, model2->numberRows() / 10)); |
1965 | barrier.setCholesky(cholesky); |
1966 | assert (!doKKT); |
1967 | } |
1968 | break; |
1969 | case 3: |
1970 | if (!doKKT) { |
1971 | ClpCholeskyWssmp * cholesky = new ClpCholeskyWssmp(); |
1972 | barrier.setCholesky(cholesky); |
1973 | } else { |
1974 | ClpCholeskyWssmpKKT * cholesky = new ClpCholeskyWssmpKKT(CoinMax(100, model2->numberRows() / 10)); |
1975 | barrier.setCholesky(cholesky); |
1976 | } |
1977 | break; |
1978 | #endif |
1979 | #ifdef UFL_BARRIER |
1980 | case 4: |
1981 | if (!doKKT) { |
1982 | ClpCholeskyUfl * cholesky = new ClpCholeskyUfl(); |
1983 | barrier.setCholesky(cholesky); |
1984 | } else { |
1985 | ClpCholeskyUfl * cholesky = new ClpCholeskyUfl(); |
1986 | cholesky->setKKT(true); |
1987 | barrier.setCholesky(cholesky); |
1988 | } |
1989 | break; |
1990 | #endif |
1991 | #ifdef TAUCS_BARRIER |
1992 | case 5: { |
1993 | ClpCholeskyTaucs * cholesky = new ClpCholeskyTaucs(); |
1994 | barrier.setCholesky(cholesky); |
1995 | assert (!doKKT); |
1996 | } |
1997 | break; |
1998 | #endif |
1999 | #ifdef COIN_HAS_MUMPS |
2000 | case 6: { |
2001 | ClpCholeskyMumps * cholesky = new ClpCholeskyMumps(); |
2002 | barrier.setCholesky(cholesky); |
2003 | assert (!doKKT); |
2004 | } |
2005 | break; |
2006 | #endif |
2007 | } |
2008 | int numberRows = model2->numberRows(); |
2009 | int numberColumns = model2->numberColumns(); |
2010 | int saveMaxIts = model2->maximumIterations(); |
2011 | if (saveMaxIts < 1000) { |
2012 | barrier.setMaximumBarrierIterations(saveMaxIts); |
2013 | model2->setMaximumIterations(10000000); |
2014 | } |
2015 | #ifndef SAVEIT |
2016 | //barrier.setDiagonalPerturbation(1.0e-25); |
2017 | if (aggressiveGamma) { |
2018 | switch (aggressiveGamma) { |
2019 | case 1: |
2020 | barrier.setGamma(1.0e-5); |
2021 | barrier.setDelta(1.0e-5); |
2022 | break; |
2023 | case 2: |
2024 | barrier.setGamma(1.0e-7); |
2025 | break; |
2026 | case 3: |
2027 | barrier.setDelta(1.0e-5); |
2028 | break; |
2029 | case 4: |
2030 | barrier.setGamma(1.0e-3); |
2031 | barrier.setDelta(1.0e-3); |
2032 | break; |
2033 | case 5: |
2034 | barrier.setGamma(1.0e-3); |
2035 | break; |
2036 | case 6: |
2037 | barrier.setDelta(1.0e-3); |
2038 | break; |
2039 | } |
2040 | } |
2041 | if (scale) |
2042 | barrier.scaling(1); |
2043 | else |
2044 | barrier.scaling(0); |
2045 | barrier.primalDual(); |
2046 | #elif SAVEIT==1 |
2047 | barrier.primalDual(); |
2048 | #else |
2049 | model2->restoreModel("xx.save" ); |
2050 | // move solutions |
2051 | CoinMemcpyN(model2->primalRowSolution(), |
2052 | numberRows, barrier.primalRowSolution()); |
2053 | CoinMemcpyN(model2->dualRowSolution(), |
2054 | numberRows, barrier.dualRowSolution()); |
2055 | CoinMemcpyN(model2->primalColumnSolution(), |
2056 | numberColumns, barrier.primalColumnSolution()); |
2057 | CoinMemcpyN(model2->dualColumnSolution(), |
2058 | numberColumns, barrier.dualColumnSolution()); |
2059 | #endif |
2060 | time2 = CoinCpuTime(); |
2061 | timeCore = time2 - timeX; |
2062 | handler_->message(CLP_INTERVAL_TIMING, messages_) |
2063 | << "Barrier" << timeCore << time2 - time1 |
2064 | << CoinMessageEol; |
2065 | timeX = time2; |
2066 | int maxIts = barrier.maximumBarrierIterations(); |
2067 | int barrierStatus = barrier.status(); |
2068 | double gap = barrier.complementarityGap(); |
2069 | // get which variables are fixed |
2070 | double * saveLower = NULL; |
2071 | double * saveUpper = NULL; |
2072 | ClpPresolve pinfo2; |
2073 | ClpSimplex * saveModel2 = NULL; |
2074 | bool = false; |
2075 | int numberFixed = barrier.numberFixed(); |
2076 | if (numberFixed) { |
2077 | int numberRows = barrier.numberRows(); |
2078 | int numberColumns = barrier.numberColumns(); |
2079 | int numberTotal = numberRows + numberColumns; |
2080 | saveLower = new double [numberTotal]; |
2081 | saveUpper = new double [numberTotal]; |
2082 | CoinMemcpyN(barrier.columnLower(), numberColumns, saveLower); |
2083 | CoinMemcpyN(barrier.rowLower(), numberRows, saveLower + numberColumns); |
2084 | CoinMemcpyN(barrier.columnUpper(), numberColumns, saveUpper); |
2085 | CoinMemcpyN(barrier.rowUpper(), numberRows, saveUpper + numberColumns); |
2086 | } |
2087 | if (((numberFixed * 20 > barrier.numberRows() && numberFixed > 5000) || forceFixing) && |
2088 | presolveInCrossover) { |
2089 | // may as well do presolve |
2090 | if (!forceFixing) { |
2091 | barrier.fixFixed(); |
2092 | } else { |
2093 | // Fix |
2094 | int n = barrier.numberColumns(); |
2095 | double * lower = barrier.columnLower(); |
2096 | double * upper = barrier.columnUpper(); |
2097 | double * solution = barrier.primalColumnSolution(); |
2098 | int nFix = 0; |
2099 | for (int i = 0; i < n; i++) { |
2100 | if (barrier.fixedOrFree(i) && lower[i] < upper[i]) { |
2101 | double value = solution[i]; |
2102 | if (value < lower[i] + 1.0e-6 && value - lower[i] < upper[i] - value) { |
2103 | solution[i] = lower[i]; |
2104 | upper[i] = lower[i]; |
2105 | nFix++; |
2106 | } else if (value > upper[i] - 1.0e-6 && value - lower[i] > upper[i] - value) { |
2107 | solution[i] = upper[i]; |
2108 | lower[i] = upper[i]; |
2109 | nFix++; |
2110 | } |
2111 | } |
2112 | } |
2113 | #ifdef CLP_INVESTIGATE |
2114 | printf("%d columns fixed\n" , nFix); |
2115 | #endif |
2116 | int nr = barrier.numberRows(); |
2117 | lower = barrier.rowLower(); |
2118 | upper = barrier.rowUpper(); |
2119 | solution = barrier.primalRowSolution(); |
2120 | nFix = 0; |
2121 | for (int i = 0; i < nr; i++) { |
2122 | if (barrier.fixedOrFree(i + n) && lower[i] < upper[i]) { |
2123 | double value = solution[i]; |
2124 | if (value < lower[i] + 1.0e-6 && value - lower[i] < upper[i] - value) { |
2125 | solution[i] = lower[i]; |
2126 | upper[i] = lower[i]; |
2127 | nFix++; |
2128 | } else if (value > upper[i] - 1.0e-6 && value - lower[i] > upper[i] - value) { |
2129 | solution[i] = upper[i]; |
2130 | lower[i] = upper[i]; |
2131 | nFix++; |
2132 | } |
2133 | } |
2134 | } |
2135 | #ifdef CLP_INVESTIGATE |
2136 | printf("%d row slacks fixed\n" , nFix); |
2137 | #endif |
2138 | } |
2139 | saveModel2 = model2; |
2140 | extraPresolve = true; |
2141 | } else if (numberFixed) { |
2142 | // Set fixed to bounds (may have restored earlier solution) |
2143 | if (!forceFixing) { |
2144 | barrier.fixFixed(false); |
2145 | } else { |
2146 | // Fix |
2147 | int n = barrier.numberColumns(); |
2148 | double * lower = barrier.columnLower(); |
2149 | double * upper = barrier.columnUpper(); |
2150 | double * solution = barrier.primalColumnSolution(); |
2151 | int nFix = 0; |
2152 | for (int i = 0; i < n; i++) { |
2153 | if (barrier.fixedOrFree(i) && lower[i] < upper[i]) { |
2154 | double value = solution[i]; |
2155 | if (value < lower[i] + 1.0e-8 && value - lower[i] < upper[i] - value) { |
2156 | solution[i] = lower[i]; |
2157 | upper[i] = lower[i]; |
2158 | nFix++; |
2159 | } else if (value > upper[i] - 1.0e-8 && value - lower[i] > upper[i] - value) { |
2160 | solution[i] = upper[i]; |
2161 | lower[i] = upper[i]; |
2162 | nFix++; |
2163 | } else { |
2164 | //printf("fixcol %d %g <= %g <= %g\n", |
2165 | // i,lower[i],solution[i],upper[i]); |
2166 | } |
2167 | } |
2168 | } |
2169 | #ifdef CLP_INVESTIGATE |
2170 | printf("%d columns fixed\n" , nFix); |
2171 | #endif |
2172 | int nr = barrier.numberRows(); |
2173 | lower = barrier.rowLower(); |
2174 | upper = barrier.rowUpper(); |
2175 | solution = barrier.primalRowSolution(); |
2176 | nFix = 0; |
2177 | for (int i = 0; i < nr; i++) { |
2178 | if (barrier.fixedOrFree(i + n) && lower[i] < upper[i]) { |
2179 | double value = solution[i]; |
2180 | if (value < lower[i] + 1.0e-5 && value - lower[i] < upper[i] - value) { |
2181 | solution[i] = lower[i]; |
2182 | upper[i] = lower[i]; |
2183 | nFix++; |
2184 | } else if (value > upper[i] - 1.0e-5 && value - lower[i] > upper[i] - value) { |
2185 | solution[i] = upper[i]; |
2186 | lower[i] = upper[i]; |
2187 | nFix++; |
2188 | } else { |
2189 | //printf("fixrow %d %g <= %g <= %g\n", |
2190 | // i,lower[i],solution[i],upper[i]); |
2191 | } |
2192 | } |
2193 | } |
2194 | #ifdef CLP_INVESTIGATE |
2195 | printf("%d row slacks fixed\n" , nFix); |
2196 | #endif |
2197 | } |
2198 | } |
2199 | #ifdef BORROW |
2200 | barrier.returnModel(*model2); |
2201 | double * rowPrimal = new double [numberRows]; |
2202 | double * columnPrimal = new double [numberColumns]; |
2203 | double * rowDual = new double [numberRows]; |
2204 | double * columnDual = new double [numberColumns]; |
2205 | // move solutions other way |
2206 | CoinMemcpyN(model2->primalRowSolution(), |
2207 | numberRows, rowPrimal); |
2208 | CoinMemcpyN(model2->dualRowSolution(), |
2209 | numberRows, rowDual); |
2210 | CoinMemcpyN(model2->primalColumnSolution(), |
2211 | numberColumns, columnPrimal); |
2212 | CoinMemcpyN(model2->dualColumnSolution(), |
2213 | numberColumns, columnDual); |
2214 | #else |
2215 | double * rowPrimal = barrier.primalRowSolution(); |
2216 | double * columnPrimal = barrier.primalColumnSolution(); |
2217 | double * rowDual = barrier.dualRowSolution(); |
2218 | double * columnDual = barrier.dualColumnSolution(); |
2219 | // move solutions |
2220 | CoinMemcpyN(rowPrimal, |
2221 | numberRows, model2->primalRowSolution()); |
2222 | CoinMemcpyN(rowDual, |
2223 | numberRows, model2->dualRowSolution()); |
2224 | CoinMemcpyN(columnPrimal, |
2225 | numberColumns, model2->primalColumnSolution()); |
2226 | CoinMemcpyN(columnDual, |
2227 | numberColumns, model2->dualColumnSolution()); |
2228 | #endif |
2229 | if (saveModel2) { |
2230 | // do presolve |
2231 | model2 = pinfo2.presolvedModel(*model2, dblParam_[ClpPresolveTolerance], |
2232 | false, 5, true); |
2233 | if (!model2) { |
2234 | model2 = saveModel2; |
2235 | saveModel2 = NULL; |
2236 | int numberRows = model2->numberRows(); |
2237 | int numberColumns = model2->numberColumns(); |
2238 | CoinMemcpyN(saveLower, numberColumns, model2->columnLower()); |
2239 | CoinMemcpyN(saveLower + numberColumns, numberRows, model2->rowLower()); |
2240 | delete [] saveLower; |
2241 | CoinMemcpyN(saveUpper, numberColumns, model2->columnUpper()); |
2242 | CoinMemcpyN(saveUpper + numberColumns, numberRows, model2->rowUpper()); |
2243 | delete [] saveUpper; |
2244 | saveLower = NULL; |
2245 | saveUpper = NULL; |
2246 | } |
2247 | } |
2248 | if (method == ClpSolve::useBarrier || barrierStatus < 0) { |
2249 | if (maxIts && barrierStatus < 4 && !quadraticObj) { |
2250 | //printf("***** crossover - needs more thought on difficult models\n"); |
2251 | #if SAVEIT==1 |
2252 | model2->ClpSimplex::saveModel("xx.save" ); |
2253 | #endif |
2254 | // make sure no status left |
2255 | model2->createStatus(); |
2256 | // solve |
2257 | if (!forceFixing) |
2258 | model2->setPerturbation(100); |
2259 | if (model2->factorizationFrequency() == 200) { |
2260 | // User did not touch preset |
2261 | model2->defaultFactorizationFrequency(); |
2262 | } |
2263 | #if 1 |
2264 | // throw some into basis |
2265 | if(!forceFixing) { |
2266 | int numberRows = model2->numberRows(); |
2267 | int numberColumns = model2->numberColumns(); |
2268 | double * dsort = new double[numberColumns]; |
2269 | int * sort = new int[numberColumns]; |
2270 | int n = 0; |
2271 | const double * columnLower = model2->columnLower(); |
2272 | const double * columnUpper = model2->columnUpper(); |
2273 | double * primalSolution = model2->primalColumnSolution(); |
2274 | const double * dualSolution = model2->dualColumnSolution(); |
2275 | double tolerance = 10.0 * primalTolerance_; |
2276 | int i; |
2277 | for ( i = 0; i < numberRows; i++) |
2278 | model2->setRowStatus(i, superBasic); |
2279 | for ( i = 0; i < numberColumns; i++) { |
2280 | double distance = CoinMin(columnUpper[i] - primalSolution[i], |
2281 | primalSolution[i] - columnLower[i]); |
2282 | if (distance > tolerance) { |
2283 | if (fabs(dualSolution[i]) < 1.0e-5) |
2284 | distance *= 100.0; |
2285 | dsort[n] = -distance; |
2286 | sort[n++] = i; |
2287 | model2->setStatus(i, superBasic); |
2288 | } else if (distance > primalTolerance_) { |
2289 | model2->setStatus(i, superBasic); |
2290 | } else if (primalSolution[i] <= columnLower[i] + primalTolerance_) { |
2291 | model2->setStatus(i, atLowerBound); |
2292 | primalSolution[i] = columnLower[i]; |
2293 | } else { |
2294 | model2->setStatus(i, atUpperBound); |
2295 | primalSolution[i] = columnUpper[i]; |
2296 | } |
2297 | } |
2298 | CoinSort_2(dsort, dsort + n, sort); |
2299 | n = CoinMin(numberRows, n); |
2300 | for ( i = 0; i < n; i++) { |
2301 | int iColumn = sort[i]; |
2302 | model2->setStatus(iColumn, basic); |
2303 | } |
2304 | delete [] sort; |
2305 | delete [] dsort; |
2306 | // model2->allSlackBasis(); |
2307 | if (gap < 1.0e-3 * static_cast<double> (numberRows + numberColumns)) { |
2308 | if (saveUpper) { |
2309 | int numberRows = model2->numberRows(); |
2310 | int numberColumns = model2->numberColumns(); |
2311 | CoinMemcpyN(saveLower, numberColumns, model2->columnLower()); |
2312 | CoinMemcpyN(saveLower + numberColumns, numberRows, model2->rowLower()); |
2313 | CoinMemcpyN(saveUpper, numberColumns, model2->columnUpper()); |
2314 | CoinMemcpyN(saveUpper + numberColumns, numberRows, model2->rowUpper()); |
2315 | delete [] saveLower; |
2316 | delete [] saveUpper; |
2317 | saveLower = NULL; |
2318 | saveUpper = NULL; |
2319 | } |
2320 | int numberRows = model2->numberRows(); |
2321 | int numberColumns = model2->numberColumns(); |
2322 | // just primal values pass |
2323 | double saveScale = model2->objectiveScale(); |
2324 | model2->setObjectiveScale(1.0e-3); |
2325 | model2->primal(2); |
2326 | model2->setObjectiveScale(saveScale); |
2327 | // save primal solution and copy back dual |
2328 | CoinMemcpyN(model2->primalRowSolution(), |
2329 | numberRows, rowPrimal); |
2330 | CoinMemcpyN(rowDual, |
2331 | numberRows, model2->dualRowSolution()); |
2332 | CoinMemcpyN(model2->primalColumnSolution(), |
2333 | numberColumns, columnPrimal); |
2334 | CoinMemcpyN(columnDual, |
2335 | numberColumns, model2->dualColumnSolution()); |
2336 | //model2->primal(1); |
2337 | // clean up reduced costs and flag variables |
2338 | { |
2339 | double * dj = model2->dualColumnSolution(); |
2340 | double * cost = model2->objective(); |
2341 | double * saveCost = new double[numberColumns]; |
2342 | CoinMemcpyN(cost, numberColumns, saveCost); |
2343 | double * saveLower = new double[numberColumns]; |
2344 | double * lower = model2->columnLower(); |
2345 | CoinMemcpyN(lower, numberColumns, saveLower); |
2346 | double * saveUpper = new double[numberColumns]; |
2347 | double * upper = model2->columnUpper(); |
2348 | CoinMemcpyN(upper, numberColumns, saveUpper); |
2349 | int i; |
2350 | double tolerance = 10.0 * dualTolerance_; |
2351 | for ( i = 0; i < numberColumns; i++) { |
2352 | if (model2->getStatus(i) == basic) { |
2353 | dj[i] = 0.0; |
2354 | } else if (model2->getStatus(i) == atLowerBound) { |
2355 | if (optimizationDirection_ * dj[i] < tolerance) { |
2356 | if (optimizationDirection_ * dj[i] < 0.0) { |
2357 | //if (dj[i]<-1.0e-3) |
2358 | //printf("bad dj at lb %d %g\n",i,dj[i]); |
2359 | cost[i] -= dj[i]; |
2360 | dj[i] = 0.0; |
2361 | } |
2362 | } else { |
2363 | upper[i] = lower[i]; |
2364 | } |
2365 | } else if (model2->getStatus(i) == atUpperBound) { |
2366 | if (optimizationDirection_ * dj[i] > tolerance) { |
2367 | if (optimizationDirection_ * dj[i] > 0.0) { |
2368 | //if (dj[i]>1.0e-3) |
2369 | //printf("bad dj at ub %d %g\n",i,dj[i]); |
2370 | cost[i] -= dj[i]; |
2371 | dj[i] = 0.0; |
2372 | } |
2373 | } else { |
2374 | lower[i] = upper[i]; |
2375 | } |
2376 | } |
2377 | } |
2378 | // just dual values pass |
2379 | //model2->setLogLevel(63); |
2380 | //model2->setFactorizationFrequency(1); |
2381 | model2->dual(2); |
2382 | CoinMemcpyN(saveCost, numberColumns, cost); |
2383 | delete [] saveCost; |
2384 | CoinMemcpyN(saveLower, numberColumns, lower); |
2385 | delete [] saveLower; |
2386 | CoinMemcpyN(saveUpper, numberColumns, upper); |
2387 | delete [] saveUpper; |
2388 | } |
2389 | } |
2390 | // and finish |
2391 | // move solutions |
2392 | CoinMemcpyN(rowPrimal, |
2393 | numberRows, model2->primalRowSolution()); |
2394 | CoinMemcpyN(columnPrimal, |
2395 | numberColumns, model2->primalColumnSolution()); |
2396 | } |
2397 | double saveScale = model2->objectiveScale(); |
2398 | model2->setObjectiveScale(1.0e-3); |
2399 | model2->primal(2); |
2400 | model2->setObjectiveScale(saveScale); |
2401 | model2->primal(1); |
2402 | #else |
2403 | // just primal |
2404 | model2->primal(1); |
2405 | #endif |
2406 | } else if (barrierStatus == 4) { |
2407 | // memory problems |
2408 | model2->setPerturbation(savePerturbation); |
2409 | model2->createStatus(); |
2410 | model2->dual(); |
2411 | } else if (maxIts && quadraticObj) { |
2412 | // make sure no status left |
2413 | model2->createStatus(); |
2414 | // solve |
2415 | model2->setPerturbation(100); |
2416 | model2->reducedGradient(1); |
2417 | } |
2418 | } |
2419 | |
2420 | //model2->setMaximumIterations(saveMaxIts); |
2421 | #ifdef BORROW |
2422 | delete [] rowPrimal; |
2423 | delete [] columnPrimal; |
2424 | delete [] rowDual; |
2425 | delete [] columnDual; |
2426 | #endif |
2427 | if (extraPresolve) { |
2428 | pinfo2.postsolve(true); |
2429 | delete model2; |
2430 | model2 = saveModel2; |
2431 | } |
2432 | if (saveUpper) { |
2433 | if (!forceFixing) { |
2434 | int numberRows = model2->numberRows(); |
2435 | int numberColumns = model2->numberColumns(); |
2436 | CoinMemcpyN(saveLower, numberColumns, model2->columnLower()); |
2437 | CoinMemcpyN(saveLower + numberColumns, numberRows, model2->rowLower()); |
2438 | CoinMemcpyN(saveUpper, numberColumns, model2->columnUpper()); |
2439 | CoinMemcpyN(saveUpper + numberColumns, numberRows, model2->rowUpper()); |
2440 | } |
2441 | delete [] saveLower; |
2442 | delete [] saveUpper; |
2443 | saveLower = NULL; |
2444 | saveUpper = NULL; |
2445 | if (method != ClpSolve::useBarrierNoCross) |
2446 | model2->primal(1); |
2447 | } |
2448 | model2->setPerturbation(savePerturbation); |
2449 | time2 = CoinCpuTime(); |
2450 | timeCore = time2 - timeX; |
2451 | handler_->message(CLP_INTERVAL_TIMING, messages_) |
2452 | << "Crossover" << timeCore << time2 - time1 |
2453 | << CoinMessageEol; |
2454 | timeX = time2; |
2455 | #else |
2456 | abort(); |
2457 | #endif |
2458 | } else { |
2459 | assert (method != ClpSolve::automatic); // later |
2460 | time2 = 0.0; |
2461 | } |
2462 | if (saveMatrix) { |
2463 | if (model2 == this) { |
2464 | // delete and replace |
2465 | delete model2->clpMatrix(); |
2466 | model2->replaceMatrix(saveMatrix); |
2467 | } else { |
2468 | delete saveMatrix; |
2469 | } |
2470 | } |
2471 | numberIterations = model2->numberIterations(); |
2472 | finalStatus = model2->status(); |
2473 | int finalSecondaryStatus = model2->secondaryStatus(); |
2474 | if (presolve == ClpSolve::presolveOn) { |
2475 | int saveLevel = logLevel(); |
2476 | if ((specialOptions_ & 1024) == 0) |
2477 | setLogLevel(CoinMin(1, saveLevel)); |
2478 | else |
2479 | setLogLevel(CoinMin(0, saveLevel)); |
2480 | pinfo->postsolve(true); |
2481 | numberIterations_ = 0; |
2482 | delete pinfo; |
2483 | pinfo = NULL; |
2484 | factorization_->areaFactor(model2->factorization()->adjustedAreaFactor()); |
2485 | time2 = CoinCpuTime(); |
2486 | timePresolve += time2 - timeX; |
2487 | handler_->message(CLP_INTERVAL_TIMING, messages_) |
2488 | << "Postsolve" << time2 - timeX << time2 - time1 |
2489 | << CoinMessageEol; |
2490 | timeX = time2; |
2491 | if (!presolveToFile) |
2492 | delete model2; |
2493 | if (interrupt) |
2494 | currentModel = this; |
2495 | // checkSolution(); already done by postSolve |
2496 | setLogLevel(saveLevel); |
2497 | int oldStatus=problemStatus_; |
2498 | setProblemStatus(finalStatus); |
2499 | setSecondaryStatus(finalSecondaryStatus); |
2500 | int rcode=eventHandler()->event(ClpEventHandler::presolveAfterFirstSolve); |
2501 | if (finalStatus != 3 && rcode < 0 && (finalStatus || oldStatus == -1)) { |
2502 | int savePerturbation = perturbation(); |
2503 | if (!finalStatus || (moreSpecialOptions_ & 2) == 0) { |
2504 | if (finalStatus == 2) { |
2505 | // unbounded - get feasible first |
2506 | double save = optimizationDirection_; |
2507 | optimizationDirection_ = 0.0; |
2508 | primal(1); |
2509 | optimizationDirection_ = save; |
2510 | primal(1); |
2511 | } else if (finalStatus == 1) { |
2512 | dual(); |
2513 | } else { |
2514 | setPerturbation(100); |
2515 | primal(1); |
2516 | } |
2517 | } else { |
2518 | // just set status |
2519 | problemStatus_ = finalStatus; |
2520 | } |
2521 | setPerturbation(savePerturbation); |
2522 | numberIterations += numberIterations_; |
2523 | numberIterations_ = numberIterations; |
2524 | finalStatus = status(); |
2525 | time2 = CoinCpuTime(); |
2526 | handler_->message(CLP_INTERVAL_TIMING, messages_) |
2527 | << "Cleanup" << time2 - timeX << time2 - time1 |
2528 | << CoinMessageEol; |
2529 | timeX = time2; |
2530 | } else if (rcode >= 0) { |
2531 | primal(1); |
2532 | } else { |
2533 | secondaryStatus_ = finalSecondaryStatus; |
2534 | } |
2535 | } else if (model2 != this) { |
2536 | // not presolved - but different model used (sprint probably) |
2537 | CoinMemcpyN(model2->primalRowSolution(), |
2538 | numberRows_, this->primalRowSolution()); |
2539 | CoinMemcpyN(model2->dualRowSolution(), |
2540 | numberRows_, this->dualRowSolution()); |
2541 | CoinMemcpyN(model2->primalColumnSolution(), |
2542 | numberColumns_, this->primalColumnSolution()); |
2543 | CoinMemcpyN(model2->dualColumnSolution(), |
2544 | numberColumns_, this->dualColumnSolution()); |
2545 | CoinMemcpyN(model2->statusArray(), |
2546 | numberColumns_ + numberRows_, this->statusArray()); |
2547 | objectiveValue_ = model2->objectiveValue_; |
2548 | numberIterations_ = model2->numberIterations_; |
2549 | problemStatus_ = model2->problemStatus_; |
2550 | secondaryStatus_ = model2->secondaryStatus_; |
2551 | delete model2; |
2552 | } |
2553 | if (method != ClpSolve::useBarrierNoCross && |
2554 | method != ClpSolve::useBarrier) |
2555 | setMaximumIterations(saveMaxIterations); |
2556 | std::string statusMessage[] = {"Unknown" , "Optimal" , "PrimalInfeasible" , "DualInfeasible" , "Stopped" , |
2557 | "Errors" , "User stopped" |
2558 | }; |
2559 | assert (finalStatus >= -1 && finalStatus <= 5); |
2560 | numberIterations_ = numberIterations; |
2561 | handler_->message(CLP_TIMING, messages_) |
2562 | << statusMessage[finalStatus+1] << objectiveValue() << numberIterations << time2 - time1; |
2563 | handler_->printing(presolve == ClpSolve::presolveOn) |
2564 | << timePresolve; |
2565 | handler_->printing(timeIdiot != 0.0) |
2566 | << timeIdiot; |
2567 | handler_->message() << CoinMessageEol; |
2568 | if (interrupt) |
2569 | signal(SIGINT, saveSignal); |
2570 | perturbation_ = savePerturbation; |
2571 | scalingFlag_ = saveScaling; |
2572 | // If faking objective - put back correct one |
2573 | if (savedObjective) { |
2574 | delete objective_; |
2575 | objective_ = savedObjective; |
2576 | } |
2577 | if (options.getSpecialOption(1) == 2 && |
2578 | options.getExtraInfo(1) > 1000000) { |
2579 | ClpObjective * savedObjective = objective_; |
2580 | // make up zero objective |
2581 | double * obj = new double[numberColumns_]; |
2582 | for (int i = 0; i < numberColumns_; i++) |
2583 | obj[i] = 0.0; |
2584 | objective_ = new ClpLinearObjective(obj, numberColumns_); |
2585 | delete [] obj; |
2586 | primal(1); |
2587 | delete objective_; |
2588 | objective_ = savedObjective; |
2589 | finalStatus = status(); |
2590 | } |
2591 | eventHandler()->event(ClpEventHandler::presolveEnd); |
2592 | delete pinfo; |
2593 | return finalStatus; |
2594 | } |
2595 | // General solve |
2596 | int |
2597 | ClpSimplex::initialSolve() |
2598 | { |
2599 | // Default so use dual |
2600 | ClpSolve options; |
2601 | return initialSolve(options); |
2602 | } |
2603 | // General dual solve |
2604 | int |
2605 | ClpSimplex::initialDualSolve() |
2606 | { |
2607 | ClpSolve options; |
2608 | // Use dual |
2609 | options.setSolveType(ClpSolve::useDual); |
2610 | return initialSolve(options); |
2611 | } |
2612 | // General primal solve |
2613 | int |
2614 | ClpSimplex::initialPrimalSolve() |
2615 | { |
2616 | ClpSolve options; |
2617 | // Use primal |
2618 | options.setSolveType(ClpSolve::usePrimal); |
2619 | return initialSolve(options); |
2620 | } |
2621 | // barrier solve, not to be followed by crossover |
2622 | int |
2623 | ClpSimplex::initialBarrierNoCrossSolve() |
2624 | { |
2625 | ClpSolve options; |
2626 | // Use primal |
2627 | options.setSolveType(ClpSolve::useBarrierNoCross); |
2628 | return initialSolve(options); |
2629 | } |
2630 | |
2631 | // General barrier solve |
2632 | int |
2633 | ClpSimplex::initialBarrierSolve() |
2634 | { |
2635 | ClpSolve options; |
2636 | // Use primal |
2637 | options.setSolveType(ClpSolve::useBarrier); |
2638 | return initialSolve(options); |
2639 | } |
2640 | |
2641 | // Default constructor |
2642 | ClpSolve::ClpSolve ( ) |
2643 | { |
2644 | method_ = automatic; |
2645 | presolveType_ = presolveOn; |
2646 | numberPasses_ = 5; |
2647 | int i; |
2648 | for (i = 0; i < 7; i++) |
2649 | options_[i] = 0; |
2650 | // say no +-1 matrix |
2651 | options_[3] = 1; |
2652 | for (i = 0; i < 7; i++) |
2653 | extraInfo_[i] = -1; |
2654 | independentOptions_[0] = 0; |
2655 | // But switch off slacks |
2656 | independentOptions_[1] = 512; |
2657 | // Substitute up to 3 |
2658 | independentOptions_[2] = 3; |
2659 | |
2660 | } |
2661 | // Constructor when you really know what you are doing |
2662 | ClpSolve::ClpSolve ( SolveType method, PresolveType presolveType, |
2663 | int numberPasses, int options[6], |
2664 | int [6], int independentOptions[3]) |
2665 | { |
2666 | method_ = method; |
2667 | presolveType_ = presolveType; |
2668 | numberPasses_ = numberPasses; |
2669 | int i; |
2670 | for (i = 0; i < 6; i++) |
2671 | options_[i] = options[i]; |
2672 | options_[6] = 0; |
2673 | for (i = 0; i < 6; i++) |
2674 | extraInfo_[i] = extraInfo[i]; |
2675 | extraInfo_[6] = 0; |
2676 | for (i = 0; i < 3; i++) |
2677 | independentOptions_[i] = independentOptions[i]; |
2678 | } |
2679 | |
2680 | // Copy constructor. |
2681 | ClpSolve::ClpSolve(const ClpSolve & rhs) |
2682 | { |
2683 | method_ = rhs.method_; |
2684 | presolveType_ = rhs.presolveType_; |
2685 | numberPasses_ = rhs.numberPasses_; |
2686 | int i; |
2687 | for ( i = 0; i < 7; i++) |
2688 | options_[i] = rhs.options_[i]; |
2689 | for ( i = 0; i < 7; i++) |
2690 | extraInfo_[i] = rhs.extraInfo_[i]; |
2691 | for (i = 0; i < 3; i++) |
2692 | independentOptions_[i] = rhs.independentOptions_[i]; |
2693 | } |
2694 | // Assignment operator. This copies the data |
2695 | ClpSolve & |
2696 | ClpSolve::operator=(const ClpSolve & rhs) |
2697 | { |
2698 | if (this != &rhs) { |
2699 | method_ = rhs.method_; |
2700 | presolveType_ = rhs.presolveType_; |
2701 | numberPasses_ = rhs.numberPasses_; |
2702 | int i; |
2703 | for (i = 0; i < 7; i++) |
2704 | options_[i] = rhs.options_[i]; |
2705 | for (i = 0; i < 7; i++) |
2706 | extraInfo_[i] = rhs.extraInfo_[i]; |
2707 | for (i = 0; i < 3; i++) |
2708 | independentOptions_[i] = rhs.independentOptions_[i]; |
2709 | } |
2710 | return *this; |
2711 | |
2712 | } |
2713 | // Destructor |
2714 | ClpSolve::~ClpSolve ( ) |
2715 | { |
2716 | } |
2717 | // See header file for details |
2718 | void |
2719 | ClpSolve::setSpecialOption(int which, int value, int ) |
2720 | { |
2721 | options_[which] = value; |
2722 | extraInfo_[which] = extraInfo; |
2723 | } |
2724 | int |
2725 | ClpSolve::getSpecialOption(int which) const |
2726 | { |
2727 | return options_[which]; |
2728 | } |
2729 | |
2730 | // Solve types |
2731 | void |
2732 | ClpSolve::setSolveType(SolveType method, int /*extraInfo*/) |
2733 | { |
2734 | method_ = method; |
2735 | } |
2736 | |
2737 | ClpSolve::SolveType |
2738 | ClpSolve::getSolveType() |
2739 | { |
2740 | return method_; |
2741 | } |
2742 | |
2743 | // Presolve types |
2744 | void |
2745 | ClpSolve::setPresolveType(PresolveType amount, int ) |
2746 | { |
2747 | presolveType_ = amount; |
2748 | numberPasses_ = extraInfo; |
2749 | } |
2750 | ClpSolve::PresolveType |
2751 | ClpSolve::getPresolveType() |
2752 | { |
2753 | return presolveType_; |
2754 | } |
2755 | // Extra info for idiot (or sprint) |
2756 | int |
2757 | ClpSolve::(int which) const |
2758 | { |
2759 | return extraInfo_[which]; |
2760 | } |
2761 | int |
2762 | ClpSolve::getPresolvePasses() const |
2763 | { |
2764 | return numberPasses_; |
2765 | } |
2766 | /* Say to return at once if infeasible, |
2767 | default is to solve */ |
2768 | void |
2769 | ClpSolve::setInfeasibleReturn(bool trueFalse) |
2770 | { |
2771 | independentOptions_[0] = trueFalse ? 1 : 0; |
2772 | } |
2773 | #include <string> |
2774 | // Generates code for above constructor |
2775 | void |
2776 | ClpSolve::generateCpp(FILE * fp) |
2777 | { |
2778 | std::string solveType[] = { |
2779 | "ClpSolve::useDual" , |
2780 | "ClpSolve::usePrimal" , |
2781 | "ClpSolve::usePrimalorSprint" , |
2782 | "ClpSolve::useBarrier" , |
2783 | "ClpSolve::useBarrierNoCross" , |
2784 | "ClpSolve::automatic" , |
2785 | "ClpSolve::notImplemented" |
2786 | }; |
2787 | std::string presolveType[] = { |
2788 | "ClpSolve::presolveOn" , |
2789 | "ClpSolve::presolveOff" , |
2790 | "ClpSolve::presolveNumber" , |
2791 | "ClpSolve::presolveNumberCost" |
2792 | }; |
2793 | fprintf(fp, "3 ClpSolve::SolveType method = %s;\n" , solveType[method_].c_str()); |
2794 | fprintf(fp, "3 ClpSolve::PresolveType presolveType = %s;\n" , |
2795 | presolveType[presolveType_].c_str()); |
2796 | fprintf(fp, "3 int numberPasses = %d;\n" , numberPasses_); |
2797 | fprintf(fp, "3 int options[] = {%d,%d,%d,%d,%d,%d};\n" , |
2798 | options_[0], options_[1], options_[2], |
2799 | options_[3], options_[4], options_[5]); |
2800 | fprintf(fp, "3 int extraInfo[] = {%d,%d,%d,%d,%d,%d};\n" , |
2801 | extraInfo_[0], extraInfo_[1], extraInfo_[2], |
2802 | extraInfo_[3], extraInfo_[4], extraInfo_[5]); |
2803 | fprintf(fp, "3 int independentOptions[] = {%d,%d,%d};\n" , |
2804 | independentOptions_[0], independentOptions_[1], independentOptions_[2]); |
2805 | fprintf(fp, "3 ClpSolve clpSolve(method,presolveType,numberPasses,\n" ); |
2806 | fprintf(fp, "3 options,extraInfo,independentOptions);\n" ); |
2807 | } |
2808 | //############################################################################# |
2809 | #include "ClpNonLinearCost.hpp" |
2810 | |
2811 | ClpSimplexProgress::ClpSimplexProgress () |
2812 | { |
2813 | int i; |
2814 | for (i = 0; i < CLP_PROGRESS; i++) { |
2815 | objective_[i] = COIN_DBL_MAX; |
2816 | infeasibility_[i] = -1.0; // set to an impossible value |
2817 | realInfeasibility_[i] = COIN_DBL_MAX; |
2818 | numberInfeasibilities_[i] = -1; |
2819 | iterationNumber_[i] = -1; |
2820 | } |
2821 | #ifdef CLP_PROGRESS_WEIGHT |
2822 | for (i = 0; i < CLP_PROGRESS_WEIGHT; i++) { |
2823 | objectiveWeight_[i] = COIN_DBL_MAX; |
2824 | infeasibilityWeight_[i] = -1.0; // set to an impossible value |
2825 | realInfeasibilityWeight_[i] = COIN_DBL_MAX; |
2826 | numberInfeasibilitiesWeight_[i] = -1; |
2827 | iterationNumberWeight_[i] = -1; |
2828 | } |
2829 | drop_ = 0.0; |
2830 | best_ = 0.0; |
2831 | #endif |
2832 | initialWeight_ = 0.0; |
2833 | for (i = 0; i < CLP_CYCLE; i++) { |
2834 | //obj_[i]=COIN_DBL_MAX; |
2835 | in_[i] = -1; |
2836 | out_[i] = -1; |
2837 | way_[i] = 0; |
2838 | } |
2839 | numberTimes_ = 0; |
2840 | numberBadTimes_ = 0; |
2841 | numberReallyBadTimes_ = 0; |
2842 | numberTimesFlagged_ = 0; |
2843 | model_ = NULL; |
2844 | oddState_ = 0; |
2845 | } |
2846 | |
2847 | |
2848 | //----------------------------------------------------------------------------- |
2849 | |
2850 | ClpSimplexProgress::~ClpSimplexProgress () |
2851 | { |
2852 | } |
2853 | // Copy constructor. |
2854 | ClpSimplexProgress::ClpSimplexProgress(const ClpSimplexProgress &rhs) |
2855 | { |
2856 | int i; |
2857 | for (i = 0; i < CLP_PROGRESS; i++) { |
2858 | objective_[i] = rhs.objective_[i]; |
2859 | infeasibility_[i] = rhs.infeasibility_[i]; |
2860 | realInfeasibility_[i] = rhs.realInfeasibility_[i]; |
2861 | numberInfeasibilities_[i] = rhs.numberInfeasibilities_[i]; |
2862 | iterationNumber_[i] = rhs.iterationNumber_[i]; |
2863 | } |
2864 | #ifdef CLP_PROGRESS_WEIGHT |
2865 | for (i = 0; i < CLP_PROGRESS_WEIGHT; i++) { |
2866 | objectiveWeight_[i] = rhs.objectiveWeight_[i]; |
2867 | infeasibilityWeight_[i] = rhs.infeasibilityWeight_[i]; |
2868 | realInfeasibilityWeight_[i] = rhs.realInfeasibilityWeight_[i]; |
2869 | numberInfeasibilitiesWeight_[i] = rhs.numberInfeasibilitiesWeight_[i]; |
2870 | iterationNumberWeight_[i] = rhs.iterationNumberWeight_[i]; |
2871 | } |
2872 | drop_ = rhs.drop_; |
2873 | best_ = rhs.best_; |
2874 | #endif |
2875 | initialWeight_ = rhs.initialWeight_; |
2876 | for (i = 0; i < CLP_CYCLE; i++) { |
2877 | //obj_[i]=rhs.obj_[i]; |
2878 | in_[i] = rhs.in_[i]; |
2879 | out_[i] = rhs.out_[i]; |
2880 | way_[i] = rhs.way_[i]; |
2881 | } |
2882 | numberTimes_ = rhs.numberTimes_; |
2883 | numberBadTimes_ = rhs.numberBadTimes_; |
2884 | numberReallyBadTimes_ = rhs.numberReallyBadTimes_; |
2885 | numberTimesFlagged_ = rhs.numberTimesFlagged_; |
2886 | model_ = rhs.model_; |
2887 | oddState_ = rhs.oddState_; |
2888 | } |
2889 | // Copy constructor.from model |
2890 | ClpSimplexProgress::ClpSimplexProgress(ClpSimplex * model) |
2891 | { |
2892 | model_ = model; |
2893 | reset(); |
2894 | initialWeight_ = 0.0; |
2895 | } |
2896 | // Fill from model |
2897 | void |
2898 | ClpSimplexProgress::fillFromModel ( ClpSimplex * model ) |
2899 | { |
2900 | model_ = model; |
2901 | reset(); |
2902 | initialWeight_ = 0.0; |
2903 | } |
2904 | // Assignment operator. This copies the data |
2905 | ClpSimplexProgress & |
2906 | ClpSimplexProgress::operator=(const ClpSimplexProgress & rhs) |
2907 | { |
2908 | if (this != &rhs) { |
2909 | int i; |
2910 | for (i = 0; i < CLP_PROGRESS; i++) { |
2911 | objective_[i] = rhs.objective_[i]; |
2912 | infeasibility_[i] = rhs.infeasibility_[i]; |
2913 | realInfeasibility_[i] = rhs.realInfeasibility_[i]; |
2914 | numberInfeasibilities_[i] = rhs.numberInfeasibilities_[i]; |
2915 | iterationNumber_[i] = rhs.iterationNumber_[i]; |
2916 | } |
2917 | #ifdef CLP_PROGRESS_WEIGHT |
2918 | for (i = 0; i < CLP_PROGRESS_WEIGHT; i++) { |
2919 | objectiveWeight_[i] = rhs.objectiveWeight_[i]; |
2920 | infeasibilityWeight_[i] = rhs.infeasibilityWeight_[i]; |
2921 | realInfeasibilityWeight_[i] = rhs.realInfeasibilityWeight_[i]; |
2922 | numberInfeasibilitiesWeight_[i] = rhs.numberInfeasibilitiesWeight_[i]; |
2923 | iterationNumberWeight_[i] = rhs.iterationNumberWeight_[i]; |
2924 | } |
2925 | drop_ = rhs.drop_; |
2926 | best_ = rhs.best_; |
2927 | #endif |
2928 | initialWeight_ = rhs.initialWeight_; |
2929 | for (i = 0; i < CLP_CYCLE; i++) { |
2930 | //obj_[i]=rhs.obj_[i]; |
2931 | in_[i] = rhs.in_[i]; |
2932 | out_[i] = rhs.out_[i]; |
2933 | way_[i] = rhs.way_[i]; |
2934 | } |
2935 | numberTimes_ = rhs.numberTimes_; |
2936 | numberBadTimes_ = rhs.numberBadTimes_; |
2937 | numberReallyBadTimes_ = rhs.numberReallyBadTimes_; |
2938 | numberTimesFlagged_ = rhs.numberTimesFlagged_; |
2939 | model_ = rhs.model_; |
2940 | oddState_ = rhs.oddState_; |
2941 | } |
2942 | return *this; |
2943 | } |
2944 | // Seems to be something odd about exact comparison of doubles on linux |
2945 | static bool equalDouble(double value1, double value2) |
2946 | { |
2947 | |
2948 | union { |
2949 | double d; |
2950 | int i[2]; |
2951 | } v1, v2; |
2952 | v1.d = value1; |
2953 | v2.d = value2; |
2954 | if (sizeof(int) * 2 == sizeof(double)) |
2955 | return (v1.i[0] == v2.i[0] && v1.i[1] == v2.i[1]); |
2956 | else |
2957 | return (v1.i[0] == v2.i[0]); |
2958 | } |
2959 | int |
2960 | ClpSimplexProgress::looping() |
2961 | { |
2962 | if (!model_) |
2963 | return -1; |
2964 | double objective = model_->rawObjectiveValue(); |
2965 | if (model_->algorithm() < 0) |
2966 | objective -= model_->bestPossibleImprovement(); |
2967 | double infeasibility; |
2968 | double realInfeasibility = 0.0; |
2969 | int numberInfeasibilities; |
2970 | int iterationNumber = model_->numberIterations(); |
2971 | numberTimesFlagged_ = 0; |
2972 | if (model_->algorithm() < 0) { |
2973 | // dual |
2974 | infeasibility = model_->sumPrimalInfeasibilities(); |
2975 | numberInfeasibilities = model_->numberPrimalInfeasibilities(); |
2976 | } else { |
2977 | //primal |
2978 | infeasibility = model_->sumDualInfeasibilities(); |
2979 | realInfeasibility = model_->nonLinearCost()->sumInfeasibilities(); |
2980 | numberInfeasibilities = model_->numberDualInfeasibilities(); |
2981 | } |
2982 | int i; |
2983 | int numberMatched = 0; |
2984 | int matched = 0; |
2985 | int nsame = 0; |
2986 | for (i = 0; i < CLP_PROGRESS; i++) { |
2987 | bool matchedOnObjective = equalDouble(objective, objective_[i]); |
2988 | bool matchedOnInfeasibility = equalDouble(infeasibility, infeasibility_[i]); |
2989 | bool matchedOnInfeasibilities = |
2990 | (numberInfeasibilities == numberInfeasibilities_[i]); |
2991 | |
2992 | if (matchedOnObjective && matchedOnInfeasibility && matchedOnInfeasibilities) { |
2993 | matched |= (1 << i); |
2994 | // Check not same iteration |
2995 | if (iterationNumber != iterationNumber_[i]) { |
2996 | numberMatched++; |
2997 | // here mainly to get over compiler bug? |
2998 | if (model_->messageHandler()->logLevel() > 10) |
2999 | printf("%d %d %d %d %d loop check\n" , i, numberMatched, |
3000 | matchedOnObjective, matchedOnInfeasibility, |
3001 | matchedOnInfeasibilities); |
3002 | } else { |
3003 | // stuck but code should notice |
3004 | nsame++; |
3005 | } |
3006 | } |
3007 | if (i) { |
3008 | objective_[i-1] = objective_[i]; |
3009 | infeasibility_[i-1] = infeasibility_[i]; |
3010 | realInfeasibility_[i-1] = realInfeasibility_[i]; |
3011 | numberInfeasibilities_[i-1] = numberInfeasibilities_[i]; |
3012 | iterationNumber_[i-1] = iterationNumber_[i]; |
3013 | } |
3014 | } |
3015 | objective_[CLP_PROGRESS-1] = objective; |
3016 | infeasibility_[CLP_PROGRESS-1] = infeasibility; |
3017 | realInfeasibility_[CLP_PROGRESS-1] = realInfeasibility; |
3018 | numberInfeasibilities_[CLP_PROGRESS-1] = numberInfeasibilities; |
3019 | iterationNumber_[CLP_PROGRESS-1] = iterationNumber; |
3020 | if (nsame == CLP_PROGRESS) |
3021 | numberMatched = CLP_PROGRESS; // really stuck |
3022 | if (model_->progressFlag()) |
3023 | numberMatched = 0; |
3024 | numberTimes_++; |
3025 | if (numberTimes_ < 10) |
3026 | numberMatched = 0; |
3027 | // skip if just last time as may be checking something |
3028 | if (matched == (1 << (CLP_PROGRESS - 1))) |
3029 | numberMatched = 0; |
3030 | if (numberMatched && model_->clpMatrix()->type() < 15) { |
3031 | model_->messageHandler()->message(CLP_POSSIBLELOOP, model_->messages()) |
3032 | << numberMatched |
3033 | << matched |
3034 | << numberTimes_ |
3035 | << CoinMessageEol; |
3036 | numberBadTimes_++; |
3037 | if (numberBadTimes_ < 10) { |
3038 | // make factorize every iteration |
3039 | model_->forceFactorization(1); |
3040 | if (numberBadTimes_ < 2) { |
3041 | startCheck(); // clear other loop check |
3042 | if (model_->algorithm() < 0) { |
3043 | // dual - change tolerance |
3044 | model_->setCurrentDualTolerance(model_->currentDualTolerance() * 1.05); |
3045 | // if infeasible increase dual bound |
3046 | if (model_->dualBound() < 1.0e17) { |
3047 | model_->setDualBound(model_->dualBound() * 1.1); |
3048 | static_cast<ClpSimplexDual *>(model_)->resetFakeBounds(0); |
3049 | } |
3050 | } else { |
3051 | // primal - change tolerance |
3052 | if (numberBadTimes_ > 3) |
3053 | model_->setCurrentPrimalTolerance(model_->currentPrimalTolerance() * 1.05); |
3054 | // if infeasible increase infeasibility cost |
3055 | if (model_->nonLinearCost()->numberInfeasibilities() && |
3056 | model_->infeasibilityCost() < 1.0e17) { |
3057 | model_->setInfeasibilityCost(model_->infeasibilityCost() * 1.1); |
3058 | } |
3059 | } |
3060 | } else { |
3061 | // flag |
3062 | int iSequence; |
3063 | if (model_->algorithm() < 0) { |
3064 | // dual |
3065 | if (model_->dualBound() > 1.0e14) |
3066 | model_->setDualBound(1.0e14); |
3067 | iSequence = in_[CLP_CYCLE-1]; |
3068 | } else { |
3069 | // primal |
3070 | if (model_->infeasibilityCost() > 1.0e14) |
3071 | model_->setInfeasibilityCost(1.0e14); |
3072 | iSequence = out_[CLP_CYCLE-1]; |
3073 | } |
3074 | if (iSequence >= 0) { |
3075 | char x = model_->isColumn(iSequence) ? 'C' : 'R'; |
3076 | if (model_->messageHandler()->logLevel() >= 63) |
3077 | model_->messageHandler()->message(CLP_SIMPLEX_FLAG, model_->messages()) |
3078 | << x << model_->sequenceWithin(iSequence) |
3079 | << CoinMessageEol; |
3080 | // if Gub then needs to be sequenceIn_ |
3081 | int save = model_->sequenceIn(); |
3082 | model_->setSequenceIn(iSequence); |
3083 | model_->setFlagged(iSequence); |
3084 | model_->setSequenceIn(save); |
3085 | //printf("flagging %d from loop\n",iSequence); |
3086 | startCheck(); |
3087 | } else { |
3088 | // Give up |
3089 | if (model_->messageHandler()->logLevel() >= 63) |
3090 | printf("***** All flagged?\n" ); |
3091 | return 4; |
3092 | } |
3093 | // reset |
3094 | numberBadTimes_ = 2; |
3095 | } |
3096 | return -2; |
3097 | } else { |
3098 | // look at solution and maybe declare victory |
3099 | if (infeasibility < 1.0e-4) { |
3100 | return 0; |
3101 | } else { |
3102 | model_->messageHandler()->message(CLP_LOOP, model_->messages()) |
3103 | << CoinMessageEol; |
3104 | #ifndef NDEBUG |
3105 | printf("debug loop ClpSimplex A\n" ); |
3106 | abort(); |
3107 | #endif |
3108 | return 3; |
3109 | } |
3110 | } |
3111 | } |
3112 | return -1; |
3113 | } |
3114 | // Resets as much as possible |
3115 | void |
3116 | ClpSimplexProgress::reset() |
3117 | { |
3118 | int i; |
3119 | for (i = 0; i < CLP_PROGRESS; i++) { |
3120 | if (model_->algorithm() >= 0) |
3121 | objective_[i] = COIN_DBL_MAX; |
3122 | else |
3123 | objective_[i] = -COIN_DBL_MAX; |
3124 | infeasibility_[i] = -1.0; // set to an impossible value |
3125 | realInfeasibility_[i] = COIN_DBL_MAX; |
3126 | numberInfeasibilities_[i] = -1; |
3127 | iterationNumber_[i] = -1; |
3128 | } |
3129 | #ifdef CLP_PROGRESS_WEIGHT |
3130 | for (i = 0; i < CLP_PROGRESS_WEIGHT; i++) { |
3131 | objectiveWeight_[i] = COIN_DBL_MAX; |
3132 | infeasibilityWeight_[i] = -1.0; // set to an impossible value |
3133 | realInfeasibilityWeight_[i] = COIN_DBL_MAX; |
3134 | numberInfeasibilitiesWeight_[i] = -1; |
3135 | iterationNumberWeight_[i] = -1; |
3136 | } |
3137 | drop_ = 0.0; |
3138 | best_ = 0.0; |
3139 | #endif |
3140 | for (i = 0; i < CLP_CYCLE; i++) { |
3141 | //obj_[i]=COIN_DBL_MAX; |
3142 | in_[i] = -1; |
3143 | out_[i] = -1; |
3144 | way_[i] = 0; |
3145 | } |
3146 | numberTimes_ = 0; |
3147 | numberBadTimes_ = 0; |
3148 | numberReallyBadTimes_ = 0; |
3149 | numberTimesFlagged_ = 0; |
3150 | oddState_ = 0; |
3151 | } |
3152 | // Returns previous objective (if -1) - current if (0) |
3153 | double |
3154 | ClpSimplexProgress::lastObjective(int back) const |
3155 | { |
3156 | return objective_[CLP_PROGRESS-1-back]; |
3157 | } |
3158 | // Returns previous infeasibility (if -1) - current if (0) |
3159 | double |
3160 | ClpSimplexProgress::lastInfeasibility(int back) const |
3161 | { |
3162 | return realInfeasibility_[CLP_PROGRESS-1-back]; |
3163 | } |
3164 | // Sets real primal infeasibility |
3165 | void |
3166 | ClpSimplexProgress::setInfeasibility(double value) |
3167 | { |
3168 | for (int i = 1; i < CLP_PROGRESS; i++) |
3169 | realInfeasibility_[i-1] = realInfeasibility_[i]; |
3170 | realInfeasibility_[CLP_PROGRESS-1] = value; |
3171 | } |
3172 | // Modify objective e.g. if dual infeasible in dual |
3173 | void |
3174 | ClpSimplexProgress::modifyObjective(double value) |
3175 | { |
3176 | objective_[CLP_PROGRESS-1] = value; |
3177 | } |
3178 | // Returns previous iteration number (if -1) - current if (0) |
3179 | int |
3180 | ClpSimplexProgress::lastIterationNumber(int back) const |
3181 | { |
3182 | return iterationNumber_[CLP_PROGRESS-1-back]; |
3183 | } |
3184 | // clears iteration numbers (to switch off panic) |
3185 | void |
3186 | ClpSimplexProgress::clearIterationNumbers() |
3187 | { |
3188 | for (int i = 0; i < CLP_PROGRESS; i++) |
3189 | iterationNumber_[i] = -1; |
3190 | } |
3191 | // Start check at beginning of whileIterating |
3192 | void |
3193 | ClpSimplexProgress::startCheck() |
3194 | { |
3195 | int i; |
3196 | for (i = 0; i < CLP_CYCLE; i++) { |
3197 | //obj_[i]=COIN_DBL_MAX; |
3198 | in_[i] = -1; |
3199 | out_[i] = -1; |
3200 | way_[i] = 0; |
3201 | } |
3202 | } |
3203 | // Returns cycle length in whileIterating |
3204 | int |
3205 | ClpSimplexProgress::cycle(int in, int out, int wayIn, int wayOut) |
3206 | { |
3207 | int i; |
3208 | #if 0 |
3209 | if (model_->numberIterations() > 206571) { |
3210 | printf("in %d out %d\n" , in, out); |
3211 | for (i = 0; i < CLP_CYCLE; i++) |
3212 | printf("cy %d in %d out %d\n" , i, in_[i], out_[i]); |
3213 | } |
3214 | #endif |
3215 | int matched = 0; |
3216 | // first see if in matches any out |
3217 | for (i = 1; i < CLP_CYCLE; i++) { |
3218 | if (in == out_[i]) { |
3219 | // even if flip then suspicious |
3220 | matched = -1; |
3221 | break; |
3222 | } |
3223 | } |
3224 | #if 0 |
3225 | if (!matched || in_[0] < 0) { |
3226 | // can't be cycle |
3227 | for (i = 0; i < CLP_CYCLE - 1; i++) { |
3228 | //obj_[i]=obj_[i+1]; |
3229 | in_[i] = in_[i+1]; |
3230 | out_[i] = out_[i+1]; |
3231 | way_[i] = way_[i+1]; |
3232 | } |
3233 | } else { |
3234 | // possible cycle |
3235 | matched = 0; |
3236 | for (i = 0; i < CLP_CYCLE - 1; i++) { |
3237 | int k; |
3238 | char wayThis = way_[i]; |
3239 | int inThis = in_[i]; |
3240 | int outThis = out_[i]; |
3241 | //double objThis = obj_[i]; |
3242 | for(k = i + 1; k < CLP_CYCLE; k++) { |
3243 | if (inThis == in_[k] && outThis == out_[k] && wayThis == way_[k]) { |
3244 | int distance = k - i; |
3245 | if (k + distance < CLP_CYCLE) { |
3246 | // See if repeats |
3247 | int j = k + distance; |
3248 | if (inThis == in_[j] && outThis == out_[j] && wayThis == way_[j]) { |
3249 | matched = distance; |
3250 | break; |
3251 | } |
3252 | } else { |
3253 | matched = distance; |
3254 | break; |
3255 | } |
3256 | } |
3257 | } |
3258 | //obj_[i]=obj_[i+1]; |
3259 | in_[i] = in_[i+1]; |
3260 | out_[i] = out_[i+1]; |
3261 | way_[i] = way_[i+1]; |
3262 | } |
3263 | } |
3264 | #else |
3265 | if (matched && in_[0] >= 0) { |
3266 | // possible cycle - only check [0] against all |
3267 | matched = 0; |
3268 | int nMatched = 0; |
3269 | char way0 = way_[0]; |
3270 | int in0 = in_[0]; |
3271 | int out0 = out_[0]; |
3272 | //double obj0 = obj_[i]; |
3273 | for(int k = 1; k < CLP_CYCLE - 4; k++) { |
3274 | if (in0 == in_[k] && out0 == out_[k] && way0 == way_[k]) { |
3275 | nMatched++; |
3276 | // See if repeats |
3277 | int end = CLP_CYCLE - k; |
3278 | int j; |
3279 | for ( j = 1; j < end; j++) { |
3280 | if (in_[j+k] != in_[j] || out_[j+k] != out_[j] || way_[j+k] != way_[j]) |
3281 | break; |
3282 | } |
3283 | if (j == end) { |
3284 | matched = k; |
3285 | break; |
3286 | } |
3287 | } |
3288 | } |
3289 | // If three times then that is too much even if not regular |
3290 | if (matched <= 0 && nMatched > 1) |
3291 | matched = 100; |
3292 | } |
3293 | for (i = 0; i < CLP_CYCLE - 1; i++) { |
3294 | //obj_[i]=obj_[i+1]; |
3295 | in_[i] = in_[i+1]; |
3296 | out_[i] = out_[i+1]; |
3297 | way_[i] = way_[i+1]; |
3298 | } |
3299 | #endif |
3300 | int way = 1 - wayIn + 4 * (1 - wayOut); |
3301 | //obj_[i]=model_->objectiveValue(); |
3302 | in_[CLP_CYCLE-1] = in; |
3303 | out_[CLP_CYCLE-1] = out; |
3304 | way_[CLP_CYCLE-1] = static_cast<char>(way); |
3305 | return matched; |
3306 | } |
3307 | #include "CoinStructuredModel.hpp" |
3308 | // Solve using structure of model and maybe in parallel |
3309 | int |
3310 | ClpSimplex::solve(CoinStructuredModel * model) |
3311 | { |
3312 | // analyze structure |
3313 | int numberRowBlocks = model->numberRowBlocks(); |
3314 | int numberColumnBlocks = model->numberColumnBlocks(); |
3315 | int numberElementBlocks = model->numberElementBlocks(); |
3316 | if (numberElementBlocks == 1) { |
3317 | loadProblem(*model, false); |
3318 | return dual(); |
3319 | } |
3320 | // For now just get top level structure |
3321 | CoinModelBlockInfo * blockInfo = new CoinModelBlockInfo [numberElementBlocks]; |
3322 | for (int i = 0; i < numberElementBlocks; i++) { |
3323 | CoinStructuredModel * subModel = |
3324 | dynamic_cast<CoinStructuredModel *>(model->block(i)); |
3325 | CoinModel * thisBlock; |
3326 | if (subModel) { |
3327 | thisBlock = subModel->coinModelBlock(blockInfo[i]); |
3328 | model->setCoinModel(thisBlock, i); |
3329 | } else { |
3330 | thisBlock = dynamic_cast<CoinModel *>(model->block(i)); |
3331 | assert (thisBlock); |
3332 | // just fill in info |
3333 | CoinModelBlockInfo info = CoinModelBlockInfo(); |
3334 | int whatsSet = thisBlock->whatIsSet(); |
3335 | info.matrix = static_cast<char>(((whatsSet & 1) != 0) ? 1 : 0); |
3336 | info.rhs = static_cast<char>(((whatsSet & 2) != 0) ? 1 : 0); |
3337 | info.rowName = static_cast<char>(((whatsSet & 4) != 0) ? 1 : 0); |
3338 | info.integer = static_cast<char>(((whatsSet & 32) != 0) ? 1 : 0); |
3339 | info.bounds = static_cast<char>(((whatsSet & 8) != 0) ? 1 : 0); |
3340 | info.columnName = static_cast<char>(((whatsSet & 16) != 0) ? 1 : 0); |
3341 | // Which block |
3342 | int iRowBlock = model->rowBlock(thisBlock->getRowBlock()); |
3343 | info.rowBlock = iRowBlock; |
3344 | int iColumnBlock = model->columnBlock(thisBlock->getColumnBlock()); |
3345 | info.columnBlock = iColumnBlock; |
3346 | blockInfo[i] = info; |
3347 | } |
3348 | } |
3349 | int * rowCounts = new int [numberRowBlocks]; |
3350 | CoinZeroN(rowCounts, numberRowBlocks); |
3351 | int * columnCounts = new int [numberColumnBlocks+1]; |
3352 | CoinZeroN(columnCounts, numberColumnBlocks); |
3353 | int decomposeType = 0; |
3354 | for (int i = 0; i < numberElementBlocks; i++) { |
3355 | int iRowBlock = blockInfo[i].rowBlock; |
3356 | int iColumnBlock = blockInfo[i].columnBlock; |
3357 | rowCounts[iRowBlock]++; |
3358 | columnCounts[iColumnBlock]++; |
3359 | } |
3360 | if (numberRowBlocks == numberColumnBlocks || |
3361 | numberRowBlocks == numberColumnBlocks + 1) { |
3362 | // could be Dantzig-Wolfe |
3363 | int numberG1 = 0; |
3364 | for (int i = 0; i < numberRowBlocks; i++) { |
3365 | if (rowCounts[i] > 1) |
3366 | numberG1++; |
3367 | } |
3368 | bool masterColumns = (numberColumnBlocks == numberRowBlocks); |
3369 | if ((masterColumns && numberElementBlocks == 2 * numberRowBlocks - 1) |
3370 | || (!masterColumns && numberElementBlocks == 2 * numberRowBlocks)) { |
3371 | if (numberG1 < 2) |
3372 | decomposeType = 1; |
3373 | } |
3374 | } |
3375 | if (!decomposeType && (numberRowBlocks == numberColumnBlocks || |
3376 | numberRowBlocks == numberColumnBlocks - 1)) { |
3377 | // could be Benders |
3378 | int numberG1 = 0; |
3379 | for (int i = 0; i < numberColumnBlocks; i++) { |
3380 | if (columnCounts[i] > 1) |
3381 | numberG1++; |
3382 | } |
3383 | bool masterRows = (numberColumnBlocks == numberRowBlocks); |
3384 | if ((masterRows && numberElementBlocks == 2 * numberColumnBlocks - 1) |
3385 | || (!masterRows && numberElementBlocks == 2 * numberColumnBlocks)) { |
3386 | if (numberG1 < 2) |
3387 | decomposeType = 2; |
3388 | } |
3389 | } |
3390 | delete [] rowCounts; |
3391 | delete [] columnCounts; |
3392 | delete [] blockInfo; |
3393 | // decide what to do |
3394 | switch (decomposeType) { |
3395 | // No good |
3396 | case 0: |
3397 | loadProblem(*model, false); |
3398 | return dual(); |
3399 | // DW |
3400 | case 1: |
3401 | return solveDW(model); |
3402 | // Benders |
3403 | case 2: |
3404 | return solveBenders(model); |
3405 | } |
3406 | return 0; // to stop compiler warning |
3407 | } |
3408 | /* This loads a model from a CoinStructuredModel object - returns number of errors. |
3409 | If originalOrder then keep to order stored in blocks, |
3410 | otherwise first column/rows correspond to first block - etc. |
3411 | If keepSolution true and size is same as current then |
3412 | keeps current status and solution |
3413 | */ |
3414 | int |
3415 | ClpSimplex::loadProblem ( CoinStructuredModel & coinModel, |
3416 | bool originalOrder, |
3417 | bool keepSolution) |
3418 | { |
3419 | unsigned char * status = NULL; |
3420 | double * psol = NULL; |
3421 | double * dsol = NULL; |
3422 | int numberRows = coinModel.numberRows(); |
3423 | int numberColumns = coinModel.numberColumns(); |
3424 | int numberRowBlocks = coinModel.numberRowBlocks(); |
3425 | int numberColumnBlocks = coinModel.numberColumnBlocks(); |
3426 | int numberElementBlocks = coinModel.numberElementBlocks(); |
3427 | if (status_ && numberRows_ && numberRows_ == numberRows && |
3428 | numberColumns_ == numberColumns && keepSolution) { |
3429 | status = new unsigned char [numberRows_+numberColumns_]; |
3430 | CoinMemcpyN(status_, numberRows_ + numberColumns_, status); |
3431 | psol = new double [numberRows_+numberColumns_]; |
3432 | CoinMemcpyN(columnActivity_, numberColumns_, psol); |
3433 | CoinMemcpyN(rowActivity_, numberRows_, psol + numberColumns_); |
3434 | dsol = new double [numberRows_+numberColumns_]; |
3435 | CoinMemcpyN(reducedCost_, numberColumns_, dsol); |
3436 | CoinMemcpyN(dual_, numberRows_, dsol + numberColumns_); |
3437 | } |
3438 | int returnCode = 0; |
3439 | double * rowLower = new double [numberRows]; |
3440 | double * rowUpper = new double [numberRows]; |
3441 | double * columnLower = new double [numberColumns]; |
3442 | double * columnUpper = new double [numberColumns]; |
3443 | double * objective = new double [numberColumns]; |
3444 | int * integerType = new int [numberColumns]; |
3445 | CoinBigIndex numberElements = 0; |
3446 | // Bases for blocks |
3447 | int * rowBase = new int[numberRowBlocks]; |
3448 | CoinFillN(rowBase, numberRowBlocks, -1); |
3449 | // And row to put it |
3450 | int * whichRow = new int [numberRows]; |
3451 | int * columnBase = new int[numberColumnBlocks]; |
3452 | CoinFillN(columnBase, numberColumnBlocks, -1); |
3453 | // And column to put it |
3454 | int * whichColumn = new int [numberColumns]; |
3455 | for (int iBlock = 0; iBlock < numberElementBlocks; iBlock++) { |
3456 | CoinModel * block = coinModel.coinBlock(iBlock); |
3457 | numberElements += block->numberElements(); |
3458 | //and set up elements etc |
3459 | double * associated = block->associatedArray(); |
3460 | // If strings then do copies |
3461 | if (block->stringsExist()) |
3462 | returnCode += block->createArrays(rowLower, rowUpper, columnLower, columnUpper, |
3463 | objective, integerType, associated); |
3464 | const CoinModelBlockInfo & info = coinModel.blockType(iBlock); |
3465 | int iRowBlock = info.rowBlock; |
3466 | int iColumnBlock = info.columnBlock; |
3467 | if (rowBase[iRowBlock] < 0) { |
3468 | rowBase[iRowBlock] = block->numberRows(); |
3469 | // Save block number |
3470 | whichRow[numberRows-numberRowBlocks+iRowBlock] = iBlock; |
3471 | } else { |
3472 | assert(rowBase[iRowBlock] == block->numberRows()); |
3473 | } |
3474 | if (columnBase[iColumnBlock] < 0) { |
3475 | columnBase[iColumnBlock] = block->numberColumns(); |
3476 | // Save block number |
3477 | whichColumn[numberColumns-numberColumnBlocks+iColumnBlock] = iBlock; |
3478 | } else { |
3479 | assert(columnBase[iColumnBlock] == block->numberColumns()); |
3480 | } |
3481 | } |
3482 | // Fill arrays with defaults |
3483 | CoinFillN(rowLower, numberRows, -COIN_DBL_MAX); |
3484 | CoinFillN(rowUpper, numberRows, COIN_DBL_MAX); |
3485 | CoinFillN(columnLower, numberColumns, 0.0); |
3486 | CoinFillN(columnUpper, numberColumns, COIN_DBL_MAX); |
3487 | CoinFillN(objective, numberColumns, 0.0); |
3488 | CoinFillN(integerType, numberColumns, 0); |
3489 | int n = 0; |
3490 | for (int iBlock = 0; iBlock < numberRowBlocks; iBlock++) { |
3491 | int k = rowBase[iBlock]; |
3492 | rowBase[iBlock] = n; |
3493 | assert (k >= 0); |
3494 | // block number |
3495 | int jBlock = whichRow[numberRows-numberRowBlocks+iBlock]; |
3496 | if (originalOrder) { |
3497 | memcpy(whichRow + n, coinModel.coinBlock(jBlock)->originalRows(), k * sizeof(int)); |
3498 | } else { |
3499 | CoinIotaN(whichRow + n, k, n); |
3500 | } |
3501 | n += k; |
3502 | } |
3503 | assert (n == numberRows); |
3504 | n = 0; |
3505 | for (int iBlock = 0; iBlock < numberColumnBlocks; iBlock++) { |
3506 | int k = columnBase[iBlock]; |
3507 | columnBase[iBlock] = n; |
3508 | assert (k >= 0); |
3509 | if (k) { |
3510 | // block number |
3511 | int jBlock = whichColumn[numberColumns-numberColumnBlocks+iBlock]; |
3512 | if (originalOrder) { |
3513 | memcpy(whichColumn + n, coinModel.coinBlock(jBlock)->originalColumns(), |
3514 | k * sizeof(int)); |
3515 | } else { |
3516 | CoinIotaN(whichColumn + n, k, n); |
3517 | } |
3518 | n += k; |
3519 | } |
3520 | } |
3521 | assert (n == numberColumns); |
3522 | bool gotIntegers = false; |
3523 | for (int iBlock = 0; iBlock < numberElementBlocks; iBlock++) { |
3524 | CoinModel * block = coinModel.coinBlock(iBlock); |
3525 | const CoinModelBlockInfo & info = coinModel.blockType(iBlock); |
3526 | int iRowBlock = info.rowBlock; |
3527 | int iRowBase = rowBase[iRowBlock]; |
3528 | int iColumnBlock = info.columnBlock; |
3529 | int iColumnBase = columnBase[iColumnBlock]; |
3530 | if (info.rhs) { |
3531 | int nRows = block->numberRows(); |
3532 | const double * lower = block->rowLowerArray(); |
3533 | const double * upper = block->rowUpperArray(); |
3534 | for (int i = 0; i < nRows; i++) { |
3535 | int put = whichRow[i+iRowBase]; |
3536 | rowLower[put] = lower[i]; |
3537 | rowUpper[put] = upper[i]; |
3538 | } |
3539 | } |
3540 | if (info.bounds) { |
3541 | int nColumns = block->numberColumns(); |
3542 | const double * lower = block->columnLowerArray(); |
3543 | const double * upper = block->columnUpperArray(); |
3544 | const double * obj = block->objectiveArray(); |
3545 | for (int i = 0; i < nColumns; i++) { |
3546 | int put = whichColumn[i+iColumnBase]; |
3547 | columnLower[put] = lower[i]; |
3548 | columnUpper[put] = upper[i]; |
3549 | objective[put] = obj[i]; |
3550 | } |
3551 | } |
3552 | if (info.integer) { |
3553 | gotIntegers = true; |
3554 | int nColumns = block->numberColumns(); |
3555 | const int * type = block->integerTypeArray(); |
3556 | for (int i = 0; i < nColumns; i++) { |
3557 | int put = whichColumn[i+iColumnBase]; |
3558 | integerType[put] = type[i]; |
3559 | } |
3560 | } |
3561 | } |
3562 | gutsOfLoadModel(numberRows, numberColumns, |
3563 | columnLower, columnUpper, objective, rowLower, rowUpper, NULL); |
3564 | delete [] rowLower; |
3565 | delete [] rowUpper; |
3566 | delete [] columnLower; |
3567 | delete [] columnUpper; |
3568 | delete [] objective; |
3569 | // Do integers if wanted |
3570 | if (gotIntegers) { |
3571 | for (int iColumn = 0; iColumn < numberColumns; iColumn++) { |
3572 | if (integerType[iColumn]) |
3573 | setInteger(iColumn); |
3574 | } |
3575 | } |
3576 | delete [] integerType; |
3577 | setObjectiveOffset(coinModel.objectiveOffset()); |
3578 | // Space for elements |
3579 | int * row = new int[numberElements]; |
3580 | int * column = new int[numberElements]; |
3581 | double * element = new double[numberElements]; |
3582 | numberElements = 0; |
3583 | for (int iBlock = 0; iBlock < numberElementBlocks; iBlock++) { |
3584 | CoinModel * block = coinModel.coinBlock(iBlock); |
3585 | const CoinModelBlockInfo & info = coinModel.blockType(iBlock); |
3586 | int iRowBlock = info.rowBlock; |
3587 | int iRowBase = rowBase[iRowBlock]; |
3588 | int iColumnBlock = info.columnBlock; |
3589 | int iColumnBase = columnBase[iColumnBlock]; |
3590 | if (info.rowName) { |
3591 | int numberItems = block->rowNames()->numberItems(); |
3592 | assert( block->numberRows() >= numberItems); |
3593 | if (numberItems) { |
3594 | const char *const * rowNames = block->rowNames()->names(); |
3595 | for (int i = 0; i < numberItems; i++) { |
3596 | int put = whichRow[i+iRowBase]; |
3597 | std::string name = rowNames[i]; |
3598 | setRowName(put, name); |
3599 | } |
3600 | } |
3601 | } |
3602 | if (info.columnName) { |
3603 | int numberItems = block->columnNames()->numberItems(); |
3604 | assert( block->numberColumns() >= numberItems); |
3605 | if (numberItems) { |
3606 | const char *const * columnNames = block->columnNames()->names(); |
3607 | for (int i = 0; i < numberItems; i++) { |
3608 | int put = whichColumn[i+iColumnBase]; |
3609 | std::string name = columnNames[i]; |
3610 | setColumnName(put, name); |
3611 | } |
3612 | } |
3613 | } |
3614 | if (info.matrix) { |
3615 | CoinPackedMatrix matrix2; |
3616 | const CoinPackedMatrix * matrix = block->packedMatrix(); |
3617 | if (!matrix) { |
3618 | double * associated = block->associatedArray(); |
3619 | block->createPackedMatrix(matrix2, associated); |
3620 | matrix = &matrix2; |
3621 | } |
3622 | // get matrix data pointers |
3623 | const int * row2 = matrix->getIndices(); |
3624 | const CoinBigIndex * columnStart = matrix->getVectorStarts(); |
3625 | const double * elementByColumn = matrix->getElements(); |
3626 | const int * columnLength = matrix->getVectorLengths(); |
3627 | int n = matrix->getNumCols(); |
3628 | assert (matrix->isColOrdered()); |
3629 | for (int iColumn = 0; iColumn < n; iColumn++) { |
3630 | CoinBigIndex j; |
3631 | int jColumn = whichColumn[iColumn+iColumnBase]; |
3632 | for (j = columnStart[iColumn]; |
3633 | j < columnStart[iColumn] + columnLength[iColumn]; j++) { |
3634 | row[numberElements] = whichRow[row2[j] + iRowBase]; |
3635 | column[numberElements] = jColumn; |
3636 | element[numberElements++] = elementByColumn[j]; |
3637 | } |
3638 | } |
3639 | } |
3640 | } |
3641 | delete [] whichRow; |
3642 | delete [] whichColumn; |
3643 | delete [] rowBase; |
3644 | delete [] columnBase; |
3645 | CoinPackedMatrix * matrix = |
3646 | new CoinPackedMatrix (true, row, column, element, numberElements); |
3647 | matrix_ = new ClpPackedMatrix(matrix); |
3648 | matrix_->setDimensions(numberRows, numberColumns); |
3649 | delete [] row; |
3650 | delete [] column; |
3651 | delete [] element; |
3652 | createStatus(); |
3653 | if (status) { |
3654 | // copy back |
3655 | CoinMemcpyN(status, numberRows_ + numberColumns_, status_); |
3656 | CoinMemcpyN(psol, numberColumns_, columnActivity_); |
3657 | CoinMemcpyN(psol + numberColumns_, numberRows_, rowActivity_); |
3658 | CoinMemcpyN(dsol, numberColumns_, reducedCost_); |
3659 | CoinMemcpyN(dsol + numberColumns_, numberRows_, dual_); |
3660 | delete [] status; |
3661 | delete [] psol; |
3662 | delete [] dsol; |
3663 | } |
3664 | optimizationDirection_ = coinModel.optimizationDirection(); |
3665 | return returnCode; |
3666 | } |
3667 | /* If input negative scales objective so maximum <= -value |
3668 | and returns scale factor used. If positive unscales and also |
3669 | redoes dual stuff |
3670 | */ |
3671 | double |
3672 | ClpSimplex::scaleObjective(double value) |
3673 | { |
3674 | double * obj = objective(); |
3675 | double largest = 0.0; |
3676 | if (value < 0.0) { |
3677 | value = - value; |
3678 | for (int i = 0; i < numberColumns_; i++) { |
3679 | largest = CoinMax(largest, fabs(obj[i])); |
3680 | } |
3681 | if (largest > value) { |
3682 | double scaleFactor = value / largest; |
3683 | for (int i = 0; i < numberColumns_; i++) { |
3684 | obj[i] *= scaleFactor; |
3685 | reducedCost_[i] *= scaleFactor; |
3686 | } |
3687 | for (int i = 0; i < numberRows_; i++) { |
3688 | dual_[i] *= scaleFactor; |
3689 | } |
3690 | largest /= value; |
3691 | } else { |
3692 | // no need |
3693 | largest = 1.0; |
3694 | } |
3695 | } else { |
3696 | // at end |
3697 | if (value != 1.0) { |
3698 | for (int i = 0; i < numberColumns_; i++) { |
3699 | obj[i] *= value; |
3700 | reducedCost_[i] *= value; |
3701 | } |
3702 | for (int i = 0; i < numberRows_; i++) { |
3703 | dual_[i] *= value; |
3704 | } |
3705 | computeObjectiveValue(); |
3706 | } |
3707 | } |
3708 | return largest; |
3709 | } |
3710 | // Solve using Dantzig-Wolfe decomposition and maybe in parallel |
3711 | int |
3712 | ClpSimplex::solveDW(CoinStructuredModel * model) |
3713 | { |
3714 | double time1 = CoinCpuTime(); |
3715 | int numberColumns = model->numberColumns(); |
3716 | int numberRowBlocks = model->numberRowBlocks(); |
3717 | int numberColumnBlocks = model->numberColumnBlocks(); |
3718 | int numberElementBlocks = model->numberElementBlocks(); |
3719 | // We already have top level structure |
3720 | CoinModelBlockInfo * blockInfo = new CoinModelBlockInfo [numberElementBlocks]; |
3721 | for (int i = 0; i < numberElementBlocks; i++) { |
3722 | CoinModel * thisBlock = model->coinBlock(i); |
3723 | assert (thisBlock); |
3724 | // just fill in info |
3725 | CoinModelBlockInfo info = CoinModelBlockInfo(); |
3726 | int whatsSet = thisBlock->whatIsSet(); |
3727 | info.matrix = static_cast<char>(((whatsSet & 1) != 0) ? 1 : 0); |
3728 | info.rhs = static_cast<char>(((whatsSet & 2) != 0) ? 1 : 0); |
3729 | info.rowName = static_cast<char>(((whatsSet & 4) != 0) ? 1 : 0); |
3730 | info.integer = static_cast<char>(((whatsSet & 32) != 0) ? 1 : 0); |
3731 | info.bounds = static_cast<char>(((whatsSet & 8) != 0) ? 1 : 0); |
3732 | info.columnName = static_cast<char>(((whatsSet & 16) != 0) ? 1 : 0); |
3733 | // Which block |
3734 | int iRowBlock = model->rowBlock(thisBlock->getRowBlock()); |
3735 | info.rowBlock = iRowBlock; |
3736 | int iColumnBlock = model->columnBlock(thisBlock->getColumnBlock()); |
3737 | info.columnBlock = iColumnBlock; |
3738 | blockInfo[i] = info; |
3739 | } |
3740 | // make up problems |
3741 | int numberBlocks = numberRowBlocks - 1; |
3742 | // Find master rows and columns |
3743 | int * rowCounts = new int [numberRowBlocks]; |
3744 | CoinZeroN(rowCounts, numberRowBlocks); |
3745 | int * columnCounts = new int [numberColumnBlocks+1]; |
3746 | CoinZeroN(columnCounts, numberColumnBlocks); |
3747 | int iBlock; |
3748 | for (iBlock = 0; iBlock < numberElementBlocks; iBlock++) { |
3749 | int iRowBlock = blockInfo[iBlock].rowBlock; |
3750 | rowCounts[iRowBlock]++; |
3751 | int iColumnBlock = blockInfo[iBlock].columnBlock; |
3752 | columnCounts[iColumnBlock]++; |
3753 | } |
3754 | int * whichBlock = new int [numberElementBlocks]; |
3755 | int masterRowBlock = -1; |
3756 | for (iBlock = 0; iBlock < numberRowBlocks; iBlock++) { |
3757 | if (rowCounts[iBlock] > 1) { |
3758 | if (masterRowBlock == -1) { |
3759 | masterRowBlock = iBlock; |
3760 | } else { |
3761 | // Can't decode |
3762 | masterRowBlock = -2; |
3763 | break; |
3764 | } |
3765 | } |
3766 | } |
3767 | int masterColumnBlock = -1; |
3768 | int kBlock = 0; |
3769 | for (iBlock = 0; iBlock < numberColumnBlocks; iBlock++) { |
3770 | int count = columnCounts[iBlock]; |
3771 | columnCounts[iBlock] = kBlock; |
3772 | kBlock += count; |
3773 | } |
3774 | for (iBlock = 0; iBlock < numberElementBlocks; iBlock++) { |
3775 | int iColumnBlock = blockInfo[iBlock].columnBlock; |
3776 | whichBlock[columnCounts[iColumnBlock]] = iBlock; |
3777 | columnCounts[iColumnBlock]++; |
3778 | } |
3779 | for (iBlock = numberColumnBlocks - 1; iBlock >= 0; iBlock--) |
3780 | columnCounts[iBlock+1] = columnCounts[iBlock]; |
3781 | columnCounts[0] = 0; |
3782 | for (iBlock = 0; iBlock < numberColumnBlocks; iBlock++) { |
3783 | int count = columnCounts[iBlock+1] - columnCounts[iBlock]; |
3784 | if (count == 1) { |
3785 | int kBlock = whichBlock[columnCounts[iBlock]]; |
3786 | int iRowBlock = blockInfo[kBlock].rowBlock; |
3787 | if (iRowBlock == masterRowBlock) { |
3788 | if (masterColumnBlock == -1) { |
3789 | masterColumnBlock = iBlock; |
3790 | } else { |
3791 | // Can't decode |
3792 | masterColumnBlock = -2; |
3793 | break; |
3794 | } |
3795 | } |
3796 | } |
3797 | } |
3798 | if (masterRowBlock < 0 || masterColumnBlock == -2) { |
3799 | // What now |
3800 | abort(); |
3801 | } |
3802 | delete [] rowCounts; |
3803 | // create all data |
3804 | const CoinPackedMatrix ** top = new const CoinPackedMatrix * [numberColumnBlocks]; |
3805 | ClpSimplex * sub = new ClpSimplex [numberBlocks]; |
3806 | ClpSimplex master; |
3807 | // Set offset |
3808 | master.setObjectiveOffset(model->objectiveOffset()); |
3809 | kBlock = 0; |
3810 | int masterBlock = -1; |
3811 | for (iBlock = 0; iBlock < numberColumnBlocks; iBlock++) { |
3812 | top[kBlock] = NULL; |
3813 | int start = columnCounts[iBlock]; |
3814 | int end = columnCounts[iBlock+1]; |
3815 | assert (end - start <= 2); |
3816 | for (int j = start; j < end; j++) { |
3817 | int jBlock = whichBlock[j]; |
3818 | int iRowBlock = blockInfo[jBlock].rowBlock; |
3819 | int iColumnBlock = blockInfo[jBlock].columnBlock; |
3820 | assert (iColumnBlock == iBlock); |
3821 | if (iColumnBlock != masterColumnBlock && iRowBlock == masterRowBlock) { |
3822 | // top matrix |
3823 | top[kBlock] = model->coinBlock(jBlock)->packedMatrix(); |
3824 | } else { |
3825 | const CoinPackedMatrix * matrix |
3826 | = model->coinBlock(jBlock)->packedMatrix(); |
3827 | // Get pointers to arrays |
3828 | const double * rowLower; |
3829 | const double * rowUpper; |
3830 | const double * columnLower; |
3831 | const double * columnUpper; |
3832 | const double * objective; |
3833 | model->block(iRowBlock, iColumnBlock, rowLower, rowUpper, |
3834 | columnLower, columnUpper, objective); |
3835 | if (iColumnBlock != masterColumnBlock) { |
3836 | // diagonal block |
3837 | sub[kBlock].loadProblem(*matrix, columnLower, columnUpper, |
3838 | objective, rowLower, rowUpper); |
3839 | if (true) { |
3840 | double * lower = sub[kBlock].columnLower(); |
3841 | double * upper = sub[kBlock].columnUpper(); |
3842 | int n = sub[kBlock].numberColumns(); |
3843 | for (int i = 0; i < n; i++) { |
3844 | lower[i] = CoinMax(-1.0e8, lower[i]); |
3845 | upper[i] = CoinMin(1.0e8, upper[i]); |
3846 | } |
3847 | } |
3848 | if (optimizationDirection_ < 0.0) { |
3849 | double * obj = sub[kBlock].objective(); |
3850 | int n = sub[kBlock].numberColumns(); |
3851 | for (int i = 0; i < n; i++) |
3852 | obj[i] = - obj[i]; |
3853 | } |
3854 | if (this->factorizationFrequency() == 200) { |
3855 | // User did not touch preset |
3856 | sub[kBlock].defaultFactorizationFrequency(); |
3857 | } else { |
3858 | // make sure model has correct value |
3859 | sub[kBlock].setFactorizationFrequency(this->factorizationFrequency()); |
3860 | } |
3861 | sub[kBlock].setPerturbation(50); |
3862 | // Set columnCounts to be diagonal block index for cleanup |
3863 | columnCounts[kBlock] = jBlock; |
3864 | } else { |
3865 | // master |
3866 | masterBlock = jBlock; |
3867 | master.loadProblem(*matrix, columnLower, columnUpper, |
3868 | objective, rowLower, rowUpper); |
3869 | if (optimizationDirection_ < 0.0) { |
3870 | double * obj = master.objective(); |
3871 | int n = master.numberColumns(); |
3872 | for (int i = 0; i < n; i++) |
3873 | obj[i] = - obj[i]; |
3874 | } |
3875 | } |
3876 | } |
3877 | } |
3878 | if (iBlock != masterColumnBlock) |
3879 | kBlock++; |
3880 | } |
3881 | delete [] whichBlock; |
3882 | delete [] blockInfo; |
3883 | // For now master must have been defined (does not have to have columns) |
3884 | assert (master.numberRows()); |
3885 | assert (masterBlock >= 0); |
3886 | int numberMasterRows = master.numberRows(); |
3887 | // Overkill in terms of space |
3888 | int spaceNeeded = CoinMax(numberBlocks * (numberMasterRows + 1), |
3889 | 2 * numberMasterRows); |
3890 | int * rowAdd = new int[spaceNeeded]; |
3891 | double * elementAdd = new double[spaceNeeded]; |
3892 | spaceNeeded = numberBlocks; |
3893 | int * columnAdd = new int[spaceNeeded+1]; |
3894 | double * objective = new double[spaceNeeded]; |
3895 | // Add in costed slacks |
3896 | int firstArtificial = master.numberColumns(); |
3897 | int lastArtificial = firstArtificial; |
3898 | if (true) { |
3899 | const double * lower = master.rowLower(); |
3900 | const double * upper = master.rowUpper(); |
3901 | int kCol = 0; |
3902 | for (int iRow = 0; iRow < numberMasterRows; iRow++) { |
3903 | if (lower[iRow] > -1.0e10) { |
3904 | rowAdd[kCol] = iRow; |
3905 | elementAdd[kCol++] = 1.0; |
3906 | } |
3907 | if (upper[iRow] < 1.0e10) { |
3908 | rowAdd[kCol] = iRow; |
3909 | elementAdd[kCol++] = -1.0; |
3910 | } |
3911 | } |
3912 | if (kCol > spaceNeeded) { |
3913 | spaceNeeded = kCol; |
3914 | delete [] columnAdd; |
3915 | delete [] objective; |
3916 | columnAdd = new int[spaceNeeded+1]; |
3917 | objective = new double[spaceNeeded]; |
3918 | } |
3919 | for (int i = 0; i < kCol; i++) { |
3920 | columnAdd[i] = i; |
3921 | objective[i] = 1.0e13; |
3922 | } |
3923 | columnAdd[kCol] = kCol; |
3924 | master.addColumns(kCol, NULL, NULL, objective, |
3925 | columnAdd, rowAdd, elementAdd); |
3926 | lastArtificial = master.numberColumns(); |
3927 | } |
3928 | int maxPass = 500; |
3929 | int iPass; |
3930 | double lastObjective = 1.0e31; |
3931 | // Create convexity rows for proposals |
3932 | int numberMasterColumns = master.numberColumns(); |
3933 | master.resize(numberMasterRows + numberBlocks, numberMasterColumns); |
3934 | if (this->factorizationFrequency() == 200) { |
3935 | // User did not touch preset |
3936 | master.defaultFactorizationFrequency(); |
3937 | } else { |
3938 | // make sure model has correct value |
3939 | master.setFactorizationFrequency(this->factorizationFrequency()); |
3940 | } |
3941 | master.setPerturbation(50); |
3942 | // Arrays to say which block and when created |
3943 | int maximumColumns = 2 * numberMasterRows + 10 * numberBlocks; |
3944 | whichBlock = new int[maximumColumns]; |
3945 | int * when = new int[maximumColumns]; |
3946 | int numberColumnsGenerated = numberBlocks; |
3947 | // fill in rhs and add in artificials |
3948 | { |
3949 | double * rowLower = master.rowLower(); |
3950 | double * rowUpper = master.rowUpper(); |
3951 | int iBlock; |
3952 | columnAdd[0] = 0; |
3953 | for (iBlock = 0; iBlock < numberBlocks; iBlock++) { |
3954 | int iRow = iBlock + numberMasterRows; |
3955 | rowLower[iRow] = 1.0; |
3956 | rowUpper[iRow] = 1.0; |
3957 | rowAdd[iBlock] = iRow; |
3958 | elementAdd[iBlock] = 1.0; |
3959 | objective[iBlock] = 1.0e13; |
3960 | columnAdd[iBlock+1] = iBlock + 1; |
3961 | when[iBlock] = -1; |
3962 | whichBlock[iBlock] = iBlock; |
3963 | } |
3964 | master.addColumns(numberBlocks, NULL, NULL, objective, |
3965 | columnAdd, rowAdd, elementAdd); |
3966 | } |
3967 | // and resize matrix to double check clp will be happy |
3968 | //master.matrix()->setDimensions(numberMasterRows+numberBlocks, |
3969 | // numberMasterColumns+numberBlocks); |
3970 | std::cout << "Time to decompose " << CoinCpuTime() - time1 << " seconds" << std::endl; |
3971 | for (iPass = 0; iPass < maxPass; iPass++) { |
3972 | printf("Start of pass %d\n" , iPass); |
3973 | // Solve master - may be infeasible |
3974 | //master.scaling(0); |
3975 | if (0) { |
3976 | master.writeMps("yy.mps" ); |
3977 | } |
3978 | // Correct artificials |
3979 | double sumArtificials = 0.0; |
3980 | if (iPass) { |
3981 | double * upper = master.columnUpper(); |
3982 | double * solution = master.primalColumnSolution(); |
3983 | double * obj = master.objective(); |
3984 | sumArtificials = 0.0; |
3985 | for (int i = firstArtificial; i < lastArtificial; i++) { |
3986 | sumArtificials += solution[i]; |
3987 | //assert (solution[i]>-1.0e-2); |
3988 | if (solution[i] < 1.0e-6) { |
3989 | #if 0 |
3990 | // Could take out |
3991 | obj[i] = 0.0; |
3992 | upper[i] = 0.0; |
3993 | #else |
3994 | obj[i] = 1.0e7; |
3995 | upper[i] = 1.0e-1; |
3996 | #endif |
3997 | solution[i] = 0.0; |
3998 | master.setColumnStatus(i, isFixed); |
3999 | } else { |
4000 | upper[i] = solution[i] + 1.0e-5 * (1.0 + solution[i]); |
4001 | } |
4002 | } |
4003 | printf("Sum of artificials before solve is %g\n" , sumArtificials); |
4004 | } |
4005 | // scale objective to be reasonable |
4006 | double scaleFactor = master.scaleObjective(-1.0e9); |
4007 | { |
4008 | double * dual = master.dualRowSolution(); |
4009 | int n = master.numberRows(); |
4010 | memset(dual, 0, n * sizeof(double)); |
4011 | double * solution = master.primalColumnSolution(); |
4012 | master.clpMatrix()->times(1.0, solution, dual); |
4013 | double sum = 0.0; |
4014 | double * lower = master.rowLower(); |
4015 | double * upper = master.rowUpper(); |
4016 | for (int iRow = 0; iRow < n; iRow++) { |
4017 | double value = dual[iRow]; |
4018 | if (value > upper[iRow]) |
4019 | sum += value - upper[iRow]; |
4020 | else if (value < lower[iRow]) |
4021 | sum -= value - lower[iRow]; |
4022 | } |
4023 | printf("suminf %g\n" , sum); |
4024 | lower = master.columnLower(); |
4025 | upper = master.columnUpper(); |
4026 | n = master.numberColumns(); |
4027 | for (int iColumn = 0; iColumn < n; iColumn++) { |
4028 | double value = solution[iColumn]; |
4029 | if (value > upper[iColumn] + 1.0e-5) |
4030 | sum += value - upper[iColumn]; |
4031 | else if (value < lower[iColumn] - 1.0e-5) |
4032 | sum -= value - lower[iColumn]; |
4033 | } |
4034 | printf("suminf %g\n" , sum); |
4035 | } |
4036 | master.primal(1); |
4037 | // Correct artificials |
4038 | sumArtificials = 0.0; |
4039 | { |
4040 | double * solution = master.primalColumnSolution(); |
4041 | for (int i = firstArtificial; i < lastArtificial; i++) { |
4042 | sumArtificials += solution[i]; |
4043 | } |
4044 | printf("Sum of artificials after solve is %g\n" , sumArtificials); |
4045 | } |
4046 | master.scaleObjective(scaleFactor); |
4047 | int problemStatus = master.status(); // do here as can change (delcols) |
4048 | if (master.numberIterations() == 0 && iPass) |
4049 | break; // finished |
4050 | if (master.objectiveValue() > lastObjective - 1.0e-7 && iPass > 555) |
4051 | break; // finished |
4052 | lastObjective = master.objectiveValue(); |
4053 | // mark basic ones and delete if necessary |
4054 | int iColumn; |
4055 | numberColumnsGenerated = master.numberColumns() - numberMasterColumns; |
4056 | for (iColumn = 0; iColumn < numberColumnsGenerated; iColumn++) { |
4057 | if (master.getStatus(iColumn + numberMasterColumns) == ClpSimplex::basic) |
4058 | when[iColumn] = iPass; |
4059 | } |
4060 | if (numberColumnsGenerated + numberBlocks > maximumColumns) { |
4061 | // delete |
4062 | int numberKeep = 0; |
4063 | int numberDelete = 0; |
4064 | int * whichDelete = new int[numberColumnsGenerated]; |
4065 | for (iColumn = 0; iColumn < numberColumnsGenerated; iColumn++) { |
4066 | if (when[iColumn] > iPass - 7) { |
4067 | // keep |
4068 | when[numberKeep] = when[iColumn]; |
4069 | whichBlock[numberKeep++] = whichBlock[iColumn]; |
4070 | } else { |
4071 | // delete |
4072 | whichDelete[numberDelete++] = iColumn + numberMasterColumns; |
4073 | } |
4074 | } |
4075 | numberColumnsGenerated -= numberDelete; |
4076 | master.deleteColumns(numberDelete, whichDelete); |
4077 | delete [] whichDelete; |
4078 | } |
4079 | const double * dual = NULL; |
4080 | bool deleteDual = false; |
4081 | if (problemStatus == 0) { |
4082 | dual = master.dualRowSolution(); |
4083 | } else if (problemStatus == 1) { |
4084 | // could do composite objective |
4085 | dual = master.infeasibilityRay(); |
4086 | deleteDual = true; |
4087 | printf("The sum of infeasibilities is %g\n" , |
4088 | master.sumPrimalInfeasibilities()); |
4089 | } else if (!master.numberColumns()) { |
4090 | assert(!iPass); |
4091 | dual = master.dualRowSolution(); |
4092 | memset(master.dualRowSolution(), |
4093 | 0, (numberMasterRows + numberBlocks)*sizeof(double)); |
4094 | } else { |
4095 | abort(); |
4096 | } |
4097 | // Scale back on first time |
4098 | if (!iPass) { |
4099 | double * dual2 = master.dualRowSolution(); |
4100 | for (int i = 0; i < numberMasterRows + numberBlocks; i++) { |
4101 | dual2[i] *= 1.0e-7; |
4102 | } |
4103 | dual = master.dualRowSolution(); |
4104 | } |
4105 | // Create objective for sub problems and solve |
4106 | columnAdd[0] = 0; |
4107 | int numberProposals = 0; |
4108 | for (iBlock = 0; iBlock < numberBlocks; iBlock++) { |
4109 | int numberColumns2 = sub[iBlock].numberColumns(); |
4110 | double * saveObj = new double [numberColumns2]; |
4111 | double * objective2 = sub[iBlock].objective(); |
4112 | memcpy(saveObj, objective2, numberColumns2 * sizeof(double)); |
4113 | // new objective |
4114 | top[iBlock]->transposeTimes(dual, objective2); |
4115 | int i; |
4116 | if (problemStatus == 0) { |
4117 | for (i = 0; i < numberColumns2; i++) |
4118 | objective2[i] = saveObj[i] - objective2[i]; |
4119 | } else { |
4120 | for (i = 0; i < numberColumns2; i++) |
4121 | objective2[i] = -objective2[i]; |
4122 | } |
4123 | // scale objective to be reasonable |
4124 | double scaleFactor = |
4125 | sub[iBlock].scaleObjective((sumArtificials > 1.0e-5) ? -1.0e-4 : -1.0e9); |
4126 | if (iPass) { |
4127 | sub[iBlock].primal(); |
4128 | } else { |
4129 | sub[iBlock].dual(); |
4130 | } |
4131 | sub[iBlock].scaleObjective(scaleFactor); |
4132 | if (!sub[iBlock].isProvenOptimal() && |
4133 | !sub[iBlock].isProvenDualInfeasible()) { |
4134 | memset(objective2, 0, numberColumns2 * sizeof(double)); |
4135 | sub[iBlock].primal(); |
4136 | if (problemStatus == 0) { |
4137 | for (i = 0; i < numberColumns2; i++) |
4138 | objective2[i] = saveObj[i] - objective2[i]; |
4139 | } else { |
4140 | for (i = 0; i < numberColumns2; i++) |
4141 | objective2[i] = -objective2[i]; |
4142 | } |
4143 | double scaleFactor = sub[iBlock].scaleObjective(-1.0e9); |
4144 | sub[iBlock].primal(1); |
4145 | sub[iBlock].scaleObjective(scaleFactor); |
4146 | } |
4147 | memcpy(objective2, saveObj, numberColumns2 * sizeof(double)); |
4148 | // get proposal |
4149 | if (sub[iBlock].numberIterations() || !iPass) { |
4150 | double objValue = 0.0; |
4151 | int start = columnAdd[numberProposals]; |
4152 | // proposal |
4153 | if (sub[iBlock].isProvenOptimal()) { |
4154 | const double * solution = sub[iBlock].primalColumnSolution(); |
4155 | top[iBlock]->times(solution, elementAdd + start); |
4156 | for (i = 0; i < numberColumns2; i++) |
4157 | objValue += solution[i] * saveObj[i]; |
4158 | // See if good dj and pack down |
4159 | int number = start; |
4160 | double dj = objValue; |
4161 | if (problemStatus) |
4162 | dj = 0.0; |
4163 | double smallest = 1.0e100; |
4164 | double largest = 0.0; |
4165 | for (i = 0; i < numberMasterRows; i++) { |
4166 | double value = elementAdd[start+i]; |
4167 | if (fabs(value) > 1.0e-15) { |
4168 | dj -= dual[i] * value; |
4169 | smallest = CoinMin(smallest, fabs(value)); |
4170 | largest = CoinMax(largest, fabs(value)); |
4171 | rowAdd[number] = i; |
4172 | elementAdd[number++] = value; |
4173 | } |
4174 | } |
4175 | // and convexity |
4176 | dj -= dual[numberMasterRows+iBlock]; |
4177 | rowAdd[number] = numberMasterRows + iBlock; |
4178 | elementAdd[number++] = 1.0; |
4179 | // if elements large then scale? |
4180 | //if (largest>1.0e8||smallest<1.0e-8) |
4181 | printf("For subproblem %d smallest - %g, largest %g - dj %g\n" , |
4182 | iBlock, smallest, largest, dj); |
4183 | if (dj < -1.0e-6 || !iPass) { |
4184 | // take |
4185 | objective[numberProposals] = objValue; |
4186 | columnAdd[++numberProposals] = number; |
4187 | when[numberColumnsGenerated] = iPass; |
4188 | whichBlock[numberColumnsGenerated++] = iBlock; |
4189 | } |
4190 | } else if (sub[iBlock].isProvenDualInfeasible()) { |
4191 | // use ray |
4192 | const double * solution = sub[iBlock].unboundedRay(); |
4193 | top[iBlock]->times(solution, elementAdd + start); |
4194 | for (i = 0; i < numberColumns2; i++) |
4195 | objValue += solution[i] * saveObj[i]; |
4196 | // See if good dj and pack down |
4197 | int number = start; |
4198 | double dj = objValue; |
4199 | double smallest = 1.0e100; |
4200 | double largest = 0.0; |
4201 | for (i = 0; i < numberMasterRows; i++) { |
4202 | double value = elementAdd[start+i]; |
4203 | if (fabs(value) > 1.0e-15) { |
4204 | dj -= dual[i] * value; |
4205 | smallest = CoinMin(smallest, fabs(value)); |
4206 | largest = CoinMax(largest, fabs(value)); |
4207 | rowAdd[number] = i; |
4208 | elementAdd[number++] = value; |
4209 | } |
4210 | } |
4211 | // if elements large or small then scale? |
4212 | //if (largest>1.0e8||smallest<1.0e-8) |
4213 | printf("For subproblem ray %d smallest - %g, largest %g - dj %g\n" , |
4214 | iBlock, smallest, largest, dj); |
4215 | if (dj < -1.0e-6) { |
4216 | // take |
4217 | objective[numberProposals] = objValue; |
4218 | columnAdd[++numberProposals] = number; |
4219 | when[numberColumnsGenerated] = iPass; |
4220 | whichBlock[numberColumnsGenerated++] = iBlock; |
4221 | } |
4222 | } else { |
4223 | abort(); |
4224 | } |
4225 | } |
4226 | delete [] saveObj; |
4227 | } |
4228 | if (deleteDual) |
4229 | delete [] dual; |
4230 | if (numberProposals) |
4231 | master.addColumns(numberProposals, NULL, NULL, objective, |
4232 | columnAdd, rowAdd, elementAdd); |
4233 | } |
4234 | std::cout << "Time at end of D-W " << CoinCpuTime() - time1 << " seconds" << std::endl; |
4235 | //master.scaling(0); |
4236 | //master.primal(1); |
4237 | loadProblem(*model); |
4238 | // now put back a good solution |
4239 | double * lower = new double[numberMasterRows+numberBlocks]; |
4240 | double * upper = new double[numberMasterRows+numberBlocks]; |
4241 | numberColumnsGenerated += numberMasterColumns; |
4242 | double * sol = new double[numberColumnsGenerated]; |
4243 | const double * solution = master.primalColumnSolution(); |
4244 | const double * masterLower = master.rowLower(); |
4245 | const double * masterUpper = master.rowUpper(); |
4246 | double * fullSolution = primalColumnSolution(); |
4247 | const double * fullLower = columnLower(); |
4248 | const double * fullUpper = columnUpper(); |
4249 | const double * rowSolution = master.primalRowSolution(); |
4250 | double * fullRowSolution = primalRowSolution(); |
4251 | const int * rowBack = model->coinBlock(masterBlock)->originalRows(); |
4252 | int numberRows2 = model->coinBlock(masterBlock)->numberRows(); |
4253 | const int * columnBack = model->coinBlock(masterBlock)->originalColumns(); |
4254 | int numberColumns2 = model->coinBlock(masterBlock)->numberColumns(); |
4255 | for (int iRow = 0; iRow < numberRows2; iRow++) { |
4256 | int kRow = rowBack[iRow]; |
4257 | setRowStatus(kRow, master.getRowStatus(iRow)); |
4258 | fullRowSolution[kRow] = rowSolution[iRow]; |
4259 | } |
4260 | for (int iColumn = 0; iColumn < numberColumns2; iColumn++) { |
4261 | int kColumn = columnBack[iColumn]; |
4262 | setStatus(kColumn, master.getStatus(iColumn)); |
4263 | fullSolution[kColumn] = solution[iColumn]; |
4264 | } |
4265 | for (iBlock = 0; iBlock < numberBlocks; iBlock++) { |
4266 | // move basis |
4267 | int kBlock = columnCounts[iBlock]; |
4268 | const int * rowBack = model->coinBlock(kBlock)->originalRows(); |
4269 | int numberRows2 = model->coinBlock(kBlock)->numberRows(); |
4270 | const int * columnBack = model->coinBlock(kBlock)->originalColumns(); |
4271 | int numberColumns2 = model->coinBlock(kBlock)->numberColumns(); |
4272 | for (int iRow = 0; iRow < numberRows2; iRow++) { |
4273 | int kRow = rowBack[iRow]; |
4274 | setRowStatus(kRow, sub[iBlock].getRowStatus(iRow)); |
4275 | } |
4276 | for (int iColumn = 0; iColumn < numberColumns2; iColumn++) { |
4277 | int kColumn = columnBack[iColumn]; |
4278 | setStatus(kColumn, sub[iBlock].getStatus(iColumn)); |
4279 | } |
4280 | // convert top bit to by rows |
4281 | CoinPackedMatrix topMatrix = *top[iBlock]; |
4282 | topMatrix.reverseOrdering(); |
4283 | // zero solution |
4284 | memset(sol, 0, numberColumnsGenerated * sizeof(double)); |
4285 | |
4286 | for (int i = numberMasterColumns; i < numberColumnsGenerated; i++) { |
4287 | if (whichBlock[i-numberMasterColumns] == iBlock) |
4288 | sol[i] = solution[i]; |
4289 | } |
4290 | memset(lower, 0, (numberMasterRows + numberBlocks)*sizeof(double)); |
4291 | master.clpMatrix()->times(1.0, sol, lower); |
4292 | for (int iRow = 0; iRow < numberMasterRows; iRow++) { |
4293 | double value = lower[iRow]; |
4294 | if (masterUpper[iRow] < 1.0e20) |
4295 | upper[iRow] = value; |
4296 | else |
4297 | upper[iRow] = COIN_DBL_MAX; |
4298 | if (masterLower[iRow] > -1.0e20) |
4299 | lower[iRow] = value; |
4300 | else |
4301 | lower[iRow] = -COIN_DBL_MAX; |
4302 | } |
4303 | sub[iBlock].addRows(numberMasterRows, lower, upper, |
4304 | topMatrix.getVectorStarts(), |
4305 | topMatrix.getVectorLengths(), |
4306 | topMatrix.getIndices(), |
4307 | topMatrix.getElements()); |
4308 | sub[iBlock].primal(1); |
4309 | const double * subSolution = sub[iBlock].primalColumnSolution(); |
4310 | const double * subRowSolution = sub[iBlock].primalRowSolution(); |
4311 | // move solution |
4312 | for (int iRow = 0; iRow < numberRows2; iRow++) { |
4313 | int kRow = rowBack[iRow]; |
4314 | fullRowSolution[kRow] = subRowSolution[iRow]; |
4315 | } |
4316 | for (int iColumn = 0; iColumn < numberColumns2; iColumn++) { |
4317 | int kColumn = columnBack[iColumn]; |
4318 | fullSolution[kColumn] = subSolution[iColumn]; |
4319 | } |
4320 | } |
4321 | for (int iColumn = 0; iColumn < numberColumns; iColumn++) { |
4322 | if (fullSolution[iColumn] < fullUpper[iColumn] - 1.0e-8 && |
4323 | fullSolution[iColumn] > fullLower[iColumn] + 1.0e-8) { |
4324 | if (getStatus(iColumn) != ClpSimplex::basic) { |
4325 | if (columnLower_[iColumn] > -1.0e30 || |
4326 | columnUpper_[iColumn] < 1.0e30) |
4327 | setStatus(iColumn, ClpSimplex::superBasic); |
4328 | else |
4329 | setStatus(iColumn, ClpSimplex::isFree); |
4330 | } |
4331 | } else if (fullSolution[iColumn] >= fullUpper[iColumn] - 1.0e-8) { |
4332 | // may help to make rest non basic |
4333 | if (getStatus(iColumn) != ClpSimplex::basic) |
4334 | setStatus(iColumn, ClpSimplex::atUpperBound); |
4335 | } else if (fullSolution[iColumn] <= fullLower[iColumn] + 1.0e-8) { |
4336 | // may help to make rest non basic |
4337 | if (getStatus(iColumn) != ClpSimplex::basic) |
4338 | setStatus(iColumn, ClpSimplex::atLowerBound); |
4339 | } |
4340 | } |
4341 | //int numberRows=model->numberRows(); |
4342 | //for (int iRow=0;iRow<numberRows;iRow++) |
4343 | //setRowStatus(iRow,ClpSimplex::superBasic); |
4344 | std::cout << "Time before cleanup of full model " << CoinCpuTime() - time1 << " seconds" << std::endl; |
4345 | primal(1); |
4346 | std::cout << "Total time " << CoinCpuTime() - time1 << " seconds" << std::endl; |
4347 | delete [] columnCounts; |
4348 | delete [] sol; |
4349 | delete [] lower; |
4350 | delete [] upper; |
4351 | delete [] whichBlock; |
4352 | delete [] when; |
4353 | delete [] columnAdd; |
4354 | delete [] rowAdd; |
4355 | delete [] elementAdd; |
4356 | delete [] objective; |
4357 | delete [] top; |
4358 | delete [] sub; |
4359 | return 0; |
4360 | } |
4361 | // Solve using Benders decomposition and maybe in parallel |
4362 | int |
4363 | ClpSimplex::solveBenders(CoinStructuredModel * model) |
4364 | { |
4365 | double time1 = CoinCpuTime(); |
4366 | //int numberColumns = model->numberColumns(); |
4367 | int numberRowBlocks = model->numberRowBlocks(); |
4368 | int numberColumnBlocks = model->numberColumnBlocks(); |
4369 | int numberElementBlocks = model->numberElementBlocks(); |
4370 | // We already have top level structure |
4371 | CoinModelBlockInfo * blockInfo = new CoinModelBlockInfo [numberElementBlocks]; |
4372 | for (int i = 0; i < numberElementBlocks; i++) { |
4373 | CoinModel * thisBlock = model->coinBlock(i); |
4374 | assert (thisBlock); |
4375 | // just fill in info |
4376 | CoinModelBlockInfo info = CoinModelBlockInfo(); |
4377 | int whatsSet = thisBlock->whatIsSet(); |
4378 | info.matrix = static_cast<char>(((whatsSet & 1) != 0) ? 1 : 0); |
4379 | info.rhs = static_cast<char>(((whatsSet & 2) != 0) ? 1 : 0); |
4380 | info.rowName = static_cast<char>(((whatsSet & 4) != 0) ? 1 : 0); |
4381 | info.integer = static_cast<char>(((whatsSet & 32) != 0) ? 1 : 0); |
4382 | info.bounds = static_cast<char>(((whatsSet & 8) != 0) ? 1 : 0); |
4383 | info.columnName = static_cast<char>(((whatsSet & 16) != 0) ? 1 : 0); |
4384 | // Which block |
4385 | int iRowBlock = model->rowBlock(thisBlock->getRowBlock()); |
4386 | info.rowBlock = iRowBlock; |
4387 | int iColumnBlock = model->columnBlock(thisBlock->getColumnBlock()); |
4388 | info.columnBlock = iColumnBlock; |
4389 | blockInfo[i] = info; |
4390 | } |
4391 | // make up problems |
4392 | int numberBlocks = numberColumnBlocks - 1; |
4393 | // Find master columns and rows |
4394 | int * columnCounts = new int [numberColumnBlocks]; |
4395 | CoinZeroN(columnCounts, numberColumnBlocks); |
4396 | int * rowCounts = new int [numberRowBlocks+1]; |
4397 | CoinZeroN(rowCounts, numberRowBlocks); |
4398 | int iBlock; |
4399 | for (iBlock = 0; iBlock < numberElementBlocks; iBlock++) { |
4400 | int iColumnBlock = blockInfo[iBlock].columnBlock; |
4401 | columnCounts[iColumnBlock]++; |
4402 | int iRowBlock = blockInfo[iBlock].rowBlock; |
4403 | rowCounts[iRowBlock]++; |
4404 | } |
4405 | int * whichBlock = new int [numberElementBlocks]; |
4406 | int masterColumnBlock = -1; |
4407 | for (iBlock = 0; iBlock < numberColumnBlocks; iBlock++) { |
4408 | if (columnCounts[iBlock] > 1) { |
4409 | if (masterColumnBlock == -1) { |
4410 | masterColumnBlock = iBlock; |
4411 | } else { |
4412 | // Can't decode |
4413 | masterColumnBlock = -2; |
4414 | break; |
4415 | } |
4416 | } |
4417 | } |
4418 | int masterRowBlock = -1; |
4419 | int kBlock = 0; |
4420 | for (iBlock = 0; iBlock < numberRowBlocks; iBlock++) { |
4421 | int count = rowCounts[iBlock]; |
4422 | rowCounts[iBlock] = kBlock; |
4423 | kBlock += count; |
4424 | } |
4425 | for (iBlock = 0; iBlock < numberElementBlocks; iBlock++) { |
4426 | int iRowBlock = blockInfo[iBlock].rowBlock; |
4427 | whichBlock[rowCounts[iRowBlock]] = iBlock; |
4428 | rowCounts[iRowBlock]++; |
4429 | } |
4430 | for (iBlock = numberRowBlocks - 1; iBlock >= 0; iBlock--) |
4431 | rowCounts[iBlock+1] = rowCounts[iBlock]; |
4432 | rowCounts[0] = 0; |
4433 | for (iBlock = 0; iBlock < numberRowBlocks; iBlock++) { |
4434 | int count = rowCounts[iBlock+1] - rowCounts[iBlock]; |
4435 | if (count == 1) { |
4436 | int kBlock = whichBlock[rowCounts[iBlock]]; |
4437 | int iColumnBlock = blockInfo[kBlock].columnBlock; |
4438 | if (iColumnBlock == masterColumnBlock) { |
4439 | if (masterRowBlock == -1) { |
4440 | masterRowBlock = iBlock; |
4441 | } else { |
4442 | // Can't decode |
4443 | masterRowBlock = -2; |
4444 | break; |
4445 | } |
4446 | } |
4447 | } |
4448 | } |
4449 | if (masterColumnBlock < 0 || masterRowBlock == -2) { |
4450 | // What now |
4451 | abort(); |
4452 | } |
4453 | delete [] columnCounts; |
4454 | // create all data |
4455 | const CoinPackedMatrix ** first = new const CoinPackedMatrix * [numberRowBlocks]; |
4456 | ClpSimplex * sub = new ClpSimplex [numberBlocks]; |
4457 | ClpSimplex master; |
4458 | // Set offset |
4459 | master.setObjectiveOffset(model->objectiveOffset()); |
4460 | kBlock = 0; |
4461 | int masterBlock = -1; |
4462 | for (iBlock = 0; iBlock < numberRowBlocks; iBlock++) { |
4463 | first[kBlock] = NULL; |
4464 | int start = rowCounts[iBlock]; |
4465 | int end = rowCounts[iBlock+1]; |
4466 | assert (end - start <= 2); |
4467 | for (int j = start; j < end; j++) { |
4468 | int jBlock = whichBlock[j]; |
4469 | int iColumnBlock = blockInfo[jBlock].columnBlock; |
4470 | int iRowBlock = blockInfo[jBlock].rowBlock; |
4471 | assert (iRowBlock == iBlock); |
4472 | if (iRowBlock != masterRowBlock && iColumnBlock == masterColumnBlock) { |
4473 | // first matrix |
4474 | first[kBlock] = model->coinBlock(jBlock)->packedMatrix(); |
4475 | } else { |
4476 | const CoinPackedMatrix * matrix |
4477 | = model->coinBlock(jBlock)->packedMatrix(); |
4478 | // Get pointers to arrays |
4479 | const double * columnLower; |
4480 | const double * columnUpper; |
4481 | const double * rowLower; |
4482 | const double * rowUpper; |
4483 | const double * objective; |
4484 | model->block(iRowBlock, iColumnBlock, rowLower, rowUpper, |
4485 | columnLower, columnUpper, objective); |
4486 | if (iRowBlock != masterRowBlock) { |
4487 | // diagonal block |
4488 | sub[kBlock].loadProblem(*matrix, columnLower, columnUpper, |
4489 | objective, rowLower, rowUpper); |
4490 | if (optimizationDirection_ < 0.0) { |
4491 | double * obj = sub[kBlock].objective(); |
4492 | int n = sub[kBlock].numberColumns(); |
4493 | for (int i = 0; i < n; i++) |
4494 | obj[i] = - obj[i]; |
4495 | } |
4496 | if (this->factorizationFrequency() == 200) { |
4497 | // User did not touch preset |
4498 | sub[kBlock].defaultFactorizationFrequency(); |
4499 | } else { |
4500 | // make sure model has correct value |
4501 | sub[kBlock].setFactorizationFrequency(this->factorizationFrequency()); |
4502 | } |
4503 | sub[kBlock].setPerturbation(50); |
4504 | // Set rowCounts to be diagonal block index for cleanup |
4505 | rowCounts[kBlock] = jBlock; |
4506 | } else { |
4507 | // master |
4508 | masterBlock = jBlock; |
4509 | master.loadProblem(*matrix, columnLower, columnUpper, |
4510 | objective, rowLower, rowUpper); |
4511 | if (optimizationDirection_ < 0.0) { |
4512 | double * obj = master.objective(); |
4513 | int n = master.numberColumns(); |
4514 | for (int i = 0; i < n; i++) |
4515 | obj[i] = - obj[i]; |
4516 | } |
4517 | } |
4518 | } |
4519 | } |
4520 | if (iBlock != masterRowBlock) |
4521 | kBlock++; |
4522 | } |
4523 | delete [] whichBlock; |
4524 | delete [] blockInfo; |
4525 | // For now master must have been defined (does not have to have rows) |
4526 | assert (master.numberColumns()); |
4527 | assert (masterBlock >= 0); |
4528 | int numberMasterColumns = master.numberColumns(); |
4529 | // Overkill in terms of space |
4530 | int spaceNeeded = CoinMax(numberBlocks * (numberMasterColumns + 1), |
4531 | 2 * numberMasterColumns); |
4532 | int * columnAdd = new int[spaceNeeded]; |
4533 | double * elementAdd = new double[spaceNeeded]; |
4534 | spaceNeeded = numberBlocks; |
4535 | int * rowAdd = new int[spaceNeeded+1]; |
4536 | double * objective = new double[spaceNeeded]; |
4537 | int maxPass = 500; |
4538 | int iPass; |
4539 | double lastObjective = -1.0e31; |
4540 | // Create columns for proposals |
4541 | int numberMasterRows = master.numberRows(); |
4542 | master.resize(numberMasterColumns + numberBlocks, numberMasterRows); |
4543 | if (this->factorizationFrequency() == 200) { |
4544 | // User did not touch preset |
4545 | master.defaultFactorizationFrequency(); |
4546 | } else { |
4547 | // make sure model has correct value |
4548 | master.setFactorizationFrequency(this->factorizationFrequency()); |
4549 | } |
4550 | master.setPerturbation(50); |
4551 | // Arrays to say which block and when created |
4552 | int maximumRows = 2 * numberMasterColumns + 10 * numberBlocks; |
4553 | whichBlock = new int[maximumRows]; |
4554 | int * when = new int[maximumRows]; |
4555 | int numberRowsGenerated = numberBlocks; |
4556 | // Add extra variables |
4557 | { |
4558 | int iBlock; |
4559 | columnAdd[0] = 0; |
4560 | for (iBlock = 0; iBlock < numberBlocks; iBlock++) { |
4561 | objective[iBlock] = 1.0; |
4562 | columnAdd[iBlock+1] = 0; |
4563 | when[iBlock] = -1; |
4564 | whichBlock[iBlock] = iBlock; |
4565 | } |
4566 | master.addColumns(numberBlocks, NULL, NULL, objective, |
4567 | columnAdd, rowAdd, elementAdd); |
4568 | } |
4569 | std::cout << "Time to decompose " << CoinCpuTime() - time1 << " seconds" << std::endl; |
4570 | for (iPass = 0; iPass < maxPass; iPass++) { |
4571 | printf("Start of pass %d\n" , iPass); |
4572 | // Solve master - may be unbounded |
4573 | //master.scaling(0); |
4574 | if (1) { |
4575 | master.writeMps("yy.mps" ); |
4576 | } |
4577 | master.dual(); |
4578 | int problemStatus = master.status(); // do here as can change (delcols) |
4579 | if (master.numberIterations() == 0 && iPass) |
4580 | break; // finished |
4581 | if (master.objectiveValue() < lastObjective + 1.0e-7 && iPass > 555) |
4582 | break; // finished |
4583 | lastObjective = master.objectiveValue(); |
4584 | // mark non-basic rows and delete if necessary |
4585 | int iRow; |
4586 | numberRowsGenerated = master.numberRows() - numberMasterRows; |
4587 | for (iRow = 0; iRow < numberRowsGenerated; iRow++) { |
4588 | if (master.getStatus(iRow + numberMasterRows) != ClpSimplex::basic) |
4589 | when[iRow] = iPass; |
4590 | } |
4591 | if (numberRowsGenerated > maximumRows) { |
4592 | // delete |
4593 | int numberKeep = 0; |
4594 | int numberDelete = 0; |
4595 | int * whichDelete = new int[numberRowsGenerated]; |
4596 | for (iRow = 0; iRow < numberRowsGenerated; iRow++) { |
4597 | if (when[iRow] > iPass - 7) { |
4598 | // keep |
4599 | when[numberKeep] = when[iRow]; |
4600 | whichBlock[numberKeep++] = whichBlock[iRow]; |
4601 | } else { |
4602 | // delete |
4603 | whichDelete[numberDelete++] = iRow + numberMasterRows; |
4604 | } |
4605 | } |
4606 | numberRowsGenerated -= numberDelete; |
4607 | master.deleteRows(numberDelete, whichDelete); |
4608 | delete [] whichDelete; |
4609 | } |
4610 | const double * primal = NULL; |
4611 | bool deletePrimal = false; |
4612 | if (problemStatus == 0) { |
4613 | primal = master.primalColumnSolution(); |
4614 | } else if (problemStatus == 2) { |
4615 | // could do composite objective |
4616 | primal = master.unboundedRay(); |
4617 | deletePrimal = true; |
4618 | printf("The sum of infeasibilities is %g\n" , |
4619 | master.sumPrimalInfeasibilities()); |
4620 | } else if (!master.numberRows()) { |
4621 | assert(!iPass); |
4622 | primal = master.primalColumnSolution(); |
4623 | memset(master.primalColumnSolution(), |
4624 | 0, numberMasterColumns * sizeof(double)); |
4625 | } else { |
4626 | abort(); |
4627 | } |
4628 | // Create rhs for sub problems and solve |
4629 | rowAdd[0] = 0; |
4630 | int numberProposals = 0; |
4631 | for (iBlock = 0; iBlock < numberBlocks; iBlock++) { |
4632 | int numberRows2 = sub[iBlock].numberRows(); |
4633 | double * saveLower = new double [numberRows2]; |
4634 | double * lower2 = sub[iBlock].rowLower(); |
4635 | double * saveUpper = new double [numberRows2]; |
4636 | double * upper2 = sub[iBlock].rowUpper(); |
4637 | // new rhs |
4638 | CoinZeroN(saveUpper, numberRows2); |
4639 | first[iBlock]->times(primal, saveUpper); |
4640 | for (int i = 0; i < numberRows2; i++) { |
4641 | double value = saveUpper[i]; |
4642 | saveLower[i] = lower2[i]; |
4643 | saveUpper[i] = upper2[i]; |
4644 | if (saveLower[i] > -1.0e30) |
4645 | lower2[i] -= value; |
4646 | if (saveUpper[i] < 1.0e30) |
4647 | upper2[i] -= value; |
4648 | } |
4649 | sub[iBlock].dual(); |
4650 | memcpy(lower2, saveLower, numberRows2 * sizeof(double)); |
4651 | memcpy(upper2, saveUpper, numberRows2 * sizeof(double)); |
4652 | // get proposal |
4653 | if (sub[iBlock].numberIterations() || !iPass) { |
4654 | double objValue = 0.0; |
4655 | int start = rowAdd[numberProposals]; |
4656 | // proposal |
4657 | if (sub[iBlock].isProvenOptimal()) { |
4658 | const double * solution = sub[iBlock].dualRowSolution(); |
4659 | first[iBlock]->transposeTimes(solution, elementAdd + start); |
4660 | for (int i = 0; i < numberRows2; i++) { |
4661 | if (solution[i] < -dualTolerance_) { |
4662 | // at upper |
4663 | assert (saveUpper[i] < 1.0e30); |
4664 | objValue += solution[i] * saveUpper[i]; |
4665 | } else if (solution[i] > dualTolerance_) { |
4666 | // at lower |
4667 | assert (saveLower[i] > -1.0e30); |
4668 | objValue += solution[i] * saveLower[i]; |
4669 | } |
4670 | } |
4671 | |
4672 | // See if cuts off and pack down |
4673 | int number = start; |
4674 | double infeas = objValue; |
4675 | double smallest = 1.0e100; |
4676 | double largest = 0.0; |
4677 | for (int i = 0; i < numberMasterColumns; i++) { |
4678 | double value = elementAdd[start+i]; |
4679 | if (fabs(value) > 1.0e-15) { |
4680 | infeas -= primal[i] * value; |
4681 | smallest = CoinMin(smallest, fabs(value)); |
4682 | largest = CoinMax(largest, fabs(value)); |
4683 | columnAdd[number] = i; |
4684 | elementAdd[number++] = -value; |
4685 | } |
4686 | } |
4687 | columnAdd[number] = numberMasterColumns + iBlock; |
4688 | elementAdd[number++] = -1.0; |
4689 | // if elements large then scale? |
4690 | //if (largest>1.0e8||smallest<1.0e-8) |
4691 | printf("For subproblem %d smallest - %g, largest %g - infeas %g\n" , |
4692 | iBlock, smallest, largest, infeas); |
4693 | if (infeas < -1.0e-6 || !iPass) { |
4694 | // take |
4695 | objective[numberProposals] = objValue; |
4696 | rowAdd[++numberProposals] = number; |
4697 | when[numberRowsGenerated] = iPass; |
4698 | whichBlock[numberRowsGenerated++] = iBlock; |
4699 | } |
4700 | } else if (sub[iBlock].isProvenPrimalInfeasible()) { |
4701 | // use ray |
4702 | const double * solution = sub[iBlock].infeasibilityRay(); |
4703 | first[iBlock]->transposeTimes(solution, elementAdd + start); |
4704 | for (int i = 0; i < numberRows2; i++) { |
4705 | if (solution[i] < -dualTolerance_) { |
4706 | // at upper |
4707 | assert (saveUpper[i] < 1.0e30); |
4708 | objValue += solution[i] * saveUpper[i]; |
4709 | } else if (solution[i] > dualTolerance_) { |
4710 | // at lower |
4711 | assert (saveLower[i] > -1.0e30); |
4712 | objValue += solution[i] * saveLower[i]; |
4713 | } |
4714 | } |
4715 | // See if good infeas and pack down |
4716 | int number = start; |
4717 | double infeas = objValue; |
4718 | double smallest = 1.0e100; |
4719 | double largest = 0.0; |
4720 | for (int i = 0; i < numberMasterColumns; i++) { |
4721 | double value = elementAdd[start+i]; |
4722 | if (fabs(value) > 1.0e-15) { |
4723 | infeas -= primal[i] * value; |
4724 | smallest = CoinMin(smallest, fabs(value)); |
4725 | largest = CoinMax(largest, fabs(value)); |
4726 | columnAdd[number] = i; |
4727 | elementAdd[number++] = -value; |
4728 | } |
4729 | } |
4730 | // if elements large or small then scale? |
4731 | //if (largest>1.0e8||smallest<1.0e-8) |
4732 | printf("For subproblem ray %d smallest - %g, largest %g - infeas %g\n" , |
4733 | iBlock, smallest, largest, infeas); |
4734 | if (infeas < -1.0e-6) { |
4735 | // take |
4736 | objective[numberProposals] = objValue; |
4737 | rowAdd[++numberProposals] = number; |
4738 | when[numberRowsGenerated] = iPass; |
4739 | whichBlock[numberRowsGenerated++] = iBlock; |
4740 | } |
4741 | } else { |
4742 | abort(); |
4743 | } |
4744 | } |
4745 | delete [] saveLower; |
4746 | delete [] saveUpper; |
4747 | } |
4748 | if (deletePrimal) |
4749 | delete [] primal; |
4750 | if (numberProposals) { |
4751 | master.addRows(numberProposals, NULL, objective, |
4752 | rowAdd, columnAdd, elementAdd); |
4753 | } |
4754 | } |
4755 | std::cout << "Time at end of Benders " << CoinCpuTime() - time1 << " seconds" << std::endl; |
4756 | //master.scaling(0); |
4757 | //master.primal(1); |
4758 | loadProblem(*model); |
4759 | // now put back a good solution |
4760 | const double * columnSolution = master.primalColumnSolution(); |
4761 | double * fullColumnSolution = primalColumnSolution(); |
4762 | const int * columnBack = model->coinBlock(masterBlock)->originalColumns(); |
4763 | int numberColumns2 = model->coinBlock(masterBlock)->numberColumns(); |
4764 | const int * rowBack = model->coinBlock(masterBlock)->originalRows(); |
4765 | int numberRows2 = model->coinBlock(masterBlock)->numberRows(); |
4766 | for (int iColumn = 0; iColumn < numberColumns2; iColumn++) { |
4767 | int kColumn = columnBack[iColumn]; |
4768 | setColumnStatus(kColumn, master.getColumnStatus(iColumn)); |
4769 | fullColumnSolution[kColumn] = columnSolution[iColumn]; |
4770 | } |
4771 | for (int iRow = 0; iRow < numberRows2; iRow++) { |
4772 | int kRow = rowBack[iRow]; |
4773 | setStatus(kRow, master.getStatus(iRow)); |
4774 | //fullSolution[kRow]=solution[iRow]; |
4775 | } |
4776 | for (iBlock = 0; iBlock < numberBlocks; iBlock++) { |
4777 | // move basis |
4778 | int kBlock = rowCounts[iBlock]; |
4779 | const int * columnBack = model->coinBlock(kBlock)->originalColumns(); |
4780 | int numberColumns2 = model->coinBlock(kBlock)->numberColumns(); |
4781 | const int * rowBack = model->coinBlock(kBlock)->originalRows(); |
4782 | int numberRows2 = model->coinBlock(kBlock)->numberRows(); |
4783 | const double * subColumnSolution = sub[iBlock].primalColumnSolution(); |
4784 | for (int iColumn = 0; iColumn < numberColumns2; iColumn++) { |
4785 | int kColumn = columnBack[iColumn]; |
4786 | setColumnStatus(kColumn, sub[iBlock].getColumnStatus(iColumn)); |
4787 | fullColumnSolution[kColumn] = subColumnSolution[iColumn]; |
4788 | } |
4789 | for (int iRow = 0; iRow < numberRows2; iRow++) { |
4790 | int kRow = rowBack[iRow]; |
4791 | setStatus(kRow, sub[iBlock].getStatus(iRow)); |
4792 | setStatus(kRow, atLowerBound); |
4793 | } |
4794 | } |
4795 | double * fullSolution = primalRowSolution(); |
4796 | CoinZeroN(fullSolution, numberRows_); |
4797 | times(1.0, fullColumnSolution, fullSolution); |
4798 | //int numberColumns=model->numberColumns(); |
4799 | //for (int iColumn=0;iColumn<numberColumns;iColumn++) |
4800 | //setColumnStatus(iColumn,ClpSimplex::superBasic); |
4801 | std::cout << "Time before cleanup of full model " << CoinCpuTime() - time1 << " seconds" << std::endl; |
4802 | this->primal(1); |
4803 | std::cout << "Total time " << CoinCpuTime() - time1 << " seconds" << std::endl; |
4804 | delete [] rowCounts; |
4805 | //delete [] sol; |
4806 | //delete [] lower; |
4807 | //delete [] upper; |
4808 | delete [] whichBlock; |
4809 | delete [] when; |
4810 | delete [] rowAdd; |
4811 | delete [] columnAdd; |
4812 | delete [] elementAdd; |
4813 | delete [] objective; |
4814 | delete [] first; |
4815 | delete [] sub; |
4816 | return 0; |
4817 | } |
4818 | |