1 | // Copyright (C) 2002, International Business Machines |
---|
2 | // Corporation and others. All Rights Reserved. |
---|
3 | |
---|
4 | |
---|
5 | /* Notes on implementation of primal simplex algorithm. |
---|
6 | |
---|
7 | When primal feasible(A): |
---|
8 | |
---|
9 | If dual feasible, we are optimal. Otherwise choose an infeasible |
---|
10 | basic variable to enter basis from a bound (B). We now need to find an |
---|
11 | outgoing variable which will leave problem primal feasible so we get |
---|
12 | the column of the tableau corresponding to the incoming variable |
---|
13 | (with the correct sign depending if variable will go up or down). |
---|
14 | |
---|
15 | We now perform a ratio test to determine which outgoing variable will |
---|
16 | preserve primal feasibility (C). If no variable found then problem |
---|
17 | is unbounded (in primal sense). If there is a variable, we then |
---|
18 | perform pivot and repeat. Trivial? |
---|
19 | |
---|
20 | ------------------------------------------- |
---|
21 | |
---|
22 | A) How do we get primal feasible? All variables have fake costs |
---|
23 | outside their feasible region so it is trivial to declare problem |
---|
24 | feasible. OSL did not have a phase 1/phase 2 approach but |
---|
25 | instead effectively put an extra cost on infeasible basic variables |
---|
26 | I am taking the same approach here, although it is generalized |
---|
27 | to allow for non-linear costs and dual information. |
---|
28 | |
---|
29 | In OSL, this weight was changed heuristically, here at present |
---|
30 | it is only increased if problem looks finished. If problem is |
---|
31 | feasible I check for unboundedness. If not unbounded we |
---|
32 | could play with going into dual. As long as weights increase |
---|
33 | any algorithm would be finite. |
---|
34 | |
---|
35 | B) Which incoming variable to choose is a virtual base class. |
---|
36 | For difficult problems steepest edge is preferred while for |
---|
37 | very easy (large) problems we will need partial scan. |
---|
38 | |
---|
39 | C) Sounds easy, but this is hardest part of algorithm. |
---|
40 | 1) Instead of stopping at first choice, we may be able |
---|
41 | to allow that variable to go through bound and if objective |
---|
42 | still improving choose again. These mini iterations can |
---|
43 | increase speed by orders of magnitude but we may need to |
---|
44 | go to more of a bucket choice of variable rather than looking |
---|
45 | at them one by one (for speed). |
---|
46 | 2) Accuracy. Basic infeasibilities may be less than |
---|
47 | tolerance. Pivoting on these makes objective go backwards. |
---|
48 | OSL modified cost so a zero move was made, Gill et al |
---|
49 | modified so a strictly positive move was made. |
---|
50 | The two problems are that re-factorizations can |
---|
51 | change rinfeasibilities above and below tolerances and that when |
---|
52 | finished we need to reset costs and try again. |
---|
53 | 3) Degeneracy. Gill et al helps but may not be enough. We |
---|
54 | may need more. Also it can improve speed a lot if we perturb |
---|
55 | the costs significantly. |
---|
56 | |
---|
57 | References: |
---|
58 | Forrest and Goldfarb, Steepest-edge simplex algorithms for |
---|
59 | linear programming - Mathematical Programming 1992 |
---|
60 | Forrest and Tomlin, Implementing the simplex method for |
---|
61 | the Optimization Subroutine Library - IBM Systems Journal 1992 |
---|
62 | Gill, Murray, Saunders, Wright A Practical Anti-Cycling |
---|
63 | Procedure for Linear and Nonlinear Programming SOL report 1988 |
---|
64 | |
---|
65 | |
---|
66 | TODO: |
---|
67 | |
---|
68 | a) Better recovery procedures. At present I never check on forward |
---|
69 | progress. There is checkpoint/restart with reducing |
---|
70 | re-factorization frequency, but this is only on singular |
---|
71 | factorizations. |
---|
72 | b) Fast methods for large easy problems (and also the option for |
---|
73 | the code to automatically choose which method). |
---|
74 | c) We need to be able to stop in various ways for OSI - this |
---|
75 | is fairly easy. |
---|
76 | |
---|
77 | */ |
---|
78 | |
---|
79 | #if defined(_MSC_VER) |
---|
80 | // Turn off compiler warning about long names |
---|
81 | # pragma warning(disable:4786) |
---|
82 | #endif |
---|
83 | |
---|
84 | #include <math.h> |
---|
85 | |
---|
86 | #include "CoinHelperFunctions.hpp" |
---|
87 | #include "ClpSimplexPrimal.hpp" |
---|
88 | #include "ClpFactorization.hpp" |
---|
89 | #include "ClpNonLinearCost.hpp" |
---|
90 | #include "OsiPackedMatrix.hpp" |
---|
91 | #include "OsiIndexedVector.hpp" |
---|
92 | #include "OsiWarmStartBasis.hpp" |
---|
93 | #include "ClpPrimalColumnPivot.hpp" |
---|
94 | #include "ClpMessage.hpp" |
---|
95 | #include <cfloat> |
---|
96 | #include <cassert> |
---|
97 | #include <string> |
---|
98 | #include <stdio.h> |
---|
99 | #include <iostream> |
---|
100 | // This returns a non const array filled with input from scalar |
---|
101 | // or actual array |
---|
102 | template <class T> inline T* |
---|
103 | copyOfArray( const T * array, const int size, T value) |
---|
104 | { |
---|
105 | T * arrayNew = new T[size]; |
---|
106 | if (array) |
---|
107 | CoinDisjointCopyN(array,size,arrayNew); |
---|
108 | else |
---|
109 | CoinFillN ( arrayNew, size,value); |
---|
110 | return arrayNew; |
---|
111 | } |
---|
112 | |
---|
113 | // This returns a non const array filled with actual array (or NULL) |
---|
114 | template <class T> inline T* |
---|
115 | copyOfArray( const T * array, const int size) |
---|
116 | { |
---|
117 | if (array) { |
---|
118 | T * arrayNew = new T[size]; |
---|
119 | CoinDisjointCopyN(array,size,arrayNew); |
---|
120 | return arrayNew; |
---|
121 | } else { |
---|
122 | return NULL; |
---|
123 | } |
---|
124 | } |
---|
125 | // primal |
---|
126 | int ClpSimplexPrimal::primal (int ifValuesPass ) |
---|
127 | { |
---|
128 | |
---|
129 | /* |
---|
130 | Method |
---|
131 | |
---|
132 | It tries to be a single phase approach with a weight of 1.0 being |
---|
133 | given to getting optimal and a weight of infeasibilityCost_ being |
---|
134 | given to getting primal feasible. In this version I have tried to |
---|
135 | be clever in a stupid way. The idea of fake bounds in dual |
---|
136 | seems to work so the primal analogue would be that of getting |
---|
137 | bounds on reduced costs (by a presolve approach) and using |
---|
138 | these for being above or below feasible region. I decided to waste |
---|
139 | memory and keep these explicitly. This allows for non-linear |
---|
140 | costs! |
---|
141 | |
---|
142 | The code is designed to take advantage of sparsity so arrays are |
---|
143 | seldom zeroed out from scratch or gone over in their entirety. |
---|
144 | The only exception is a full scan to find incoming variable for |
---|
145 | Dantzig row choice. For steepest edge we keep an updated list |
---|
146 | of dual infeasibilities (actually squares). |
---|
147 | On easy problems we don't need full scan - just |
---|
148 | pick first reasonable. |
---|
149 | |
---|
150 | One problem is how to tackle degeneracy and accuracy. At present |
---|
151 | I am using the modification of costs which I put in OSL and which was |
---|
152 | extended by Gill et al. I am still not sure of the exact details. |
---|
153 | |
---|
154 | The flow of primal is three while loops as follows: |
---|
155 | |
---|
156 | while (not finished) { |
---|
157 | |
---|
158 | while (not clean solution) { |
---|
159 | |
---|
160 | Factorize and/or clean up solution by changing bounds so |
---|
161 | primal feasible. If looks finished check fake primal bounds. |
---|
162 | Repeat until status is iterating (-1) or finished (0,1,2) |
---|
163 | |
---|
164 | } |
---|
165 | |
---|
166 | while (status==-1) { |
---|
167 | |
---|
168 | Iterate until no pivot in or out or time to re-factorize. |
---|
169 | |
---|
170 | Flow is: |
---|
171 | |
---|
172 | choose pivot column (incoming variable). if none then |
---|
173 | we are primal feasible so looks as if done but we need to |
---|
174 | break and check bounds etc. |
---|
175 | |
---|
176 | Get pivot column in tableau |
---|
177 | |
---|
178 | Choose outgoing row. If we don't find one then we look |
---|
179 | primal unbounded so break and check bounds etc. (Also the |
---|
180 | pivot tolerance is larger after any iterations so that may be |
---|
181 | reason) |
---|
182 | |
---|
183 | If we do find outgoing row, we may have to adjust costs to |
---|
184 | keep going forwards (anti-degeneracy). Check pivot will be stable |
---|
185 | and if unstable throw away iteration and break to re-factorize. |
---|
186 | If minor error re-factorize after iteration. |
---|
187 | |
---|
188 | Update everything (this may involve changing bounds on |
---|
189 | variables to stay primal feasible. |
---|
190 | |
---|
191 | } |
---|
192 | |
---|
193 | } |
---|
194 | |
---|
195 | At present we never check we are going forwards. I overdid that in |
---|
196 | OSL so will try and make a last resort. |
---|
197 | |
---|
198 | Needs partial scan pivot in option. |
---|
199 | |
---|
200 | May need other anti-degeneracy measures, especially if we try and use |
---|
201 | loose tolerances as a way to solve in fewer iterations. |
---|
202 | |
---|
203 | I like idea of dynamic scaling. This gives opportunity to decouple |
---|
204 | different implications of scaling for accuracy, iteration count and |
---|
205 | feasibility tolerance. |
---|
206 | |
---|
207 | */ |
---|
208 | |
---|
209 | // sanity check |
---|
210 | assert (numberRows_==matrix_->getNumRows()); |
---|
211 | assert (numberColumns_==matrix_->getNumCols()); |
---|
212 | // for moment all arrays must exist |
---|
213 | assert(columnLower_); |
---|
214 | assert(columnUpper_); |
---|
215 | assert(rowLower_); |
---|
216 | assert(rowUpper_); |
---|
217 | |
---|
218 | |
---|
219 | algorithm_ = +1; |
---|
220 | primalTolerance_=dblParam_[OsiPrimalTolerance]; |
---|
221 | dualTolerance_=dblParam_[OsiDualTolerance]; |
---|
222 | |
---|
223 | // put in standard form (and make row copy) |
---|
224 | // create modifiable copies of model rim and do optional scaling |
---|
225 | createRim(7+8+16,true); |
---|
226 | |
---|
227 | // save infeasibility cost |
---|
228 | double saveInfeasibilityCost = infeasibilityCost_; |
---|
229 | |
---|
230 | int iRow,iColumn; |
---|
231 | // Do initial factorization |
---|
232 | // and set certain stuff |
---|
233 | // We can either set increasing rows so ...IsBasic gives pivot row |
---|
234 | // or we can just increment iBasic one by one |
---|
235 | // for now let ...iBasic give pivot row |
---|
236 | factorization_->increasingRows(2); |
---|
237 | // row activities have negative sign |
---|
238 | factorization_->slackValue(-1.0); |
---|
239 | factorization_->zeroTolerance(1.0e-13); |
---|
240 | |
---|
241 | |
---|
242 | // If user asked for perturbation - do it |
---|
243 | int savePerturbation = perturbation_; |
---|
244 | |
---|
245 | if (perturbation_<100) |
---|
246 | perturb(); |
---|
247 | |
---|
248 | // for primal we will change bounds using infeasibilityCost_ |
---|
249 | if (nonLinearCost_==NULL) { |
---|
250 | // get a valid nonlinear cost function |
---|
251 | delete nonLinearCost_; |
---|
252 | nonLinearCost_= new ClpNonLinearCost(this); |
---|
253 | } |
---|
254 | |
---|
255 | // save if sparse factorization wanted |
---|
256 | int saveSparse = factorization_->sparseThreshold(); |
---|
257 | |
---|
258 | // loop round to clean up solution if values pass |
---|
259 | int numberThrownOut = -1; |
---|
260 | int firstSuperBasic=numberRows_+numberColumns_; |
---|
261 | int totalNumberThrownOut=0; |
---|
262 | while(numberThrownOut) { |
---|
263 | int status = internalFactorize(0+10*ifValuesPass); |
---|
264 | if (status<0) |
---|
265 | return 1; // some error |
---|
266 | else |
---|
267 | totalNumberThrownOut+= status; |
---|
268 | |
---|
269 | // for this we need clean basis so it is after factorize |
---|
270 | numberThrownOut=gutsOfSolution(rowActivityWork_,columnActivityWork_, |
---|
271 | ifValuesPass); |
---|
272 | totalNumberThrownOut+= numberThrownOut; |
---|
273 | |
---|
274 | // find first superbasic - columns, then rows |
---|
275 | if (ifValuesPass) { |
---|
276 | nextSuperBasic(firstSuperBasic); |
---|
277 | if (firstSuperBasic==numberRows_+numberColumns_) |
---|
278 | ifValuesPass=0; // signal no values pass |
---|
279 | } |
---|
280 | } |
---|
281 | |
---|
282 | if (totalNumberThrownOut) |
---|
283 | handler_->message(CLP_SINGULARITIES,messages_) |
---|
284 | <<totalNumberThrownOut |
---|
285 | <<OsiMessageEol; |
---|
286 | |
---|
287 | problemStatus_ = -1; |
---|
288 | numberIterations_=0; |
---|
289 | |
---|
290 | int lastCleaned=0; // last time objective or bounds cleaned up |
---|
291 | |
---|
292 | // number of times we have declared optimality |
---|
293 | numberTimesOptimal_=0; |
---|
294 | |
---|
295 | // Say no pivot has occurred (for steepest edge and updates) |
---|
296 | pivotRow_=-2; |
---|
297 | |
---|
298 | // This says whether to restore things etc |
---|
299 | int factorType=0; |
---|
300 | // Save iteration number |
---|
301 | int saveNumber = -1; |
---|
302 | /* |
---|
303 | Status of problem: |
---|
304 | 0 - optimal |
---|
305 | 1 - infeasible |
---|
306 | 2 - unbounded |
---|
307 | -1 - iterating |
---|
308 | -2 - factorization wanted |
---|
309 | -3 - redo checking without factorization |
---|
310 | -4 - looks infeasible |
---|
311 | -5 - looks unbounded |
---|
312 | */ |
---|
313 | while (problemStatus_<0) { |
---|
314 | // clear |
---|
315 | for (iRow=0;iRow<4;iRow++) { |
---|
316 | rowArray_[iRow]->clear(); |
---|
317 | } |
---|
318 | |
---|
319 | for (iColumn=0;iColumn<2;iColumn++) { |
---|
320 | columnArray_[iColumn]->clear(); |
---|
321 | } |
---|
322 | |
---|
323 | // give matrix (and model costs and bounds a chance to be |
---|
324 | // refreshed (normally null) |
---|
325 | matrix_->refresh(this); |
---|
326 | // If getting nowhere - why not give it a kick |
---|
327 | #if 0 |
---|
328 | // primal perturbation not coded yet |
---|
329 | if (perturbation_<101&&numberIterations_>2*(numberRows_+numberColumns_)) |
---|
330 | perturb(); |
---|
331 | #endif |
---|
332 | // If we have done no iterations - special |
---|
333 | if (saveNumber==numberIterations_) |
---|
334 | factorType=3; |
---|
335 | // may factorize, checks if problem finished |
---|
336 | statusOfProblemInPrimal(lastCleaned,factorType); |
---|
337 | |
---|
338 | // Say good factorization |
---|
339 | factorType=1; |
---|
340 | if (saveSparse) { |
---|
341 | // use default at present |
---|
342 | factorization_->sparseThreshold(0); |
---|
343 | factorization_->goSparse(); |
---|
344 | } |
---|
345 | |
---|
346 | // Say no pivot has occurred (for steepest edge and updates) |
---|
347 | pivotRow_=-2; |
---|
348 | |
---|
349 | // Save iteration number |
---|
350 | saveNumber = numberIterations_; |
---|
351 | |
---|
352 | // Iterate |
---|
353 | whileIterating(firstSuperBasic); |
---|
354 | } |
---|
355 | |
---|
356 | // if infeasible get real values |
---|
357 | if (problemStatus_) { |
---|
358 | infeasibilityCost_=0.0; |
---|
359 | createRim(7); |
---|
360 | nonLinearCost_->checkInfeasibilities(true); |
---|
361 | sumPrimalInfeasibilities_=nonLinearCost_->sumInfeasibilities(); |
---|
362 | numberPrimalInfeasibilities_= nonLinearCost_->numberInfeasibilities(); |
---|
363 | // and get good feasible duals |
---|
364 | computeDuals(); |
---|
365 | } |
---|
366 | // at present we are leaving factorization around |
---|
367 | // maybe we should empty it |
---|
368 | deleteRim(); |
---|
369 | handler_->message(CLP_SIMPLEX_FINISHED+problemStatus_,messages_) |
---|
370 | <<objectiveValue() |
---|
371 | <<OsiMessageEol; |
---|
372 | // Restore any saved stuff |
---|
373 | perturbation_ = savePerturbation; |
---|
374 | factorization_->sparseThreshold(saveSparse); |
---|
375 | infeasibilityCost_ = saveInfeasibilityCost; |
---|
376 | return problemStatus_; |
---|
377 | } |
---|
378 | /* |
---|
379 | Reasons to come out: |
---|
380 | -1 iterations etc |
---|
381 | -2 inaccuracy |
---|
382 | -3 slight inaccuracy (and done iterations) |
---|
383 | -4 end of values pass and done iterations |
---|
384 | +0 looks optimal (might be infeasible - but we will investigate) |
---|
385 | +2 looks unbounded |
---|
386 | +3 max iterations |
---|
387 | */ |
---|
388 | int |
---|
389 | ClpSimplexPrimal::whileIterating(int & firstSuperBasic) |
---|
390 | { |
---|
391 | |
---|
392 | // Say if values pass |
---|
393 | int ifValuesPass=0; |
---|
394 | int returnCode=-1; |
---|
395 | if (firstSuperBasic<numberRows_+numberColumns_) |
---|
396 | ifValuesPass=1; |
---|
397 | int saveNumber = numberIterations_; |
---|
398 | // status stays at -1 while iterating, >=0 finished, -2 to invert |
---|
399 | // status -3 to go to top without an invert |
---|
400 | while (problemStatus_==-1) { |
---|
401 | #ifdef CLP_DEBUG |
---|
402 | { |
---|
403 | int i; |
---|
404 | // not [1] as has information |
---|
405 | for (i=0;i<4;i++) { |
---|
406 | if (i!=1) |
---|
407 | rowArray_[i]->checkClear(); |
---|
408 | } |
---|
409 | for (i=0;i<2;i++) { |
---|
410 | columnArray_[i]->checkClear(); |
---|
411 | } |
---|
412 | } |
---|
413 | #endif |
---|
414 | #if CLP_DEBUG>2 |
---|
415 | // very expensive |
---|
416 | if (numberIterations_>0&&numberIterations_<-2534) { |
---|
417 | handler_->setLogLevel(63); |
---|
418 | double saveValue = objectiveValue_; |
---|
419 | double * saveRow1 = new double[numberRows_]; |
---|
420 | double * saveRow2 = new double[numberRows_]; |
---|
421 | memcpy(saveRow1,rowReducedCost_,numberRows_*sizeof(double)); |
---|
422 | memcpy(saveRow2,rowActivityWork_,numberRows_*sizeof(double)); |
---|
423 | double * saveColumn1 = new double[numberColumns_]; |
---|
424 | double * saveColumn2 = new double[numberColumns_]; |
---|
425 | memcpy(saveColumn1,reducedCostWork_,numberColumns_*sizeof(double)); |
---|
426 | memcpy(saveColumn2,columnActivityWork_,numberColumns_*sizeof(double)); |
---|
427 | createRim(7); |
---|
428 | gutsOfSolution(rowActivityWork_,columnActivityWork_); |
---|
429 | printf("xxx %d old obj %g, recomputed %g, sum primal inf %g\n", |
---|
430 | numberIterations_, |
---|
431 | saveValue,objectiveValue_,sumPrimalInfeasibilities_); |
---|
432 | memcpy(rowReducedCost_,saveRow1,numberRows_*sizeof(double)); |
---|
433 | memcpy(rowActivityWork_,saveRow2,numberRows_*sizeof(double)); |
---|
434 | memcpy(reducedCostWork_,saveColumn1,numberColumns_*sizeof(double)); |
---|
435 | memcpy(columnActivityWork_,saveColumn2,numberColumns_*sizeof(double)); |
---|
436 | delete [] saveRow1; |
---|
437 | delete [] saveRow2; |
---|
438 | delete [] saveColumn1; |
---|
439 | delete [] saveColumn2; |
---|
440 | objectiveValue_=saveValue; |
---|
441 | } |
---|
442 | #endif |
---|
443 | if (!ifValuesPass) { |
---|
444 | // choose column to come in |
---|
445 | // can use pivotRow_ to update weights |
---|
446 | // pass in list of cost changes so can do row updates (rowArray_[1]) |
---|
447 | // NOTE rowArray_[0] is used by computeDuals which is a |
---|
448 | // slow way of getting duals but might be used |
---|
449 | primalColumn(rowArray_[1],rowArray_[2],rowArray_[3], |
---|
450 | columnArray_[0],columnArray_[1]); |
---|
451 | } else { |
---|
452 | // in values pass |
---|
453 | if (ifValuesPass>0) { |
---|
454 | nextSuperBasic(firstSuperBasic); |
---|
455 | if (firstSuperBasic==numberRows_+numberColumns_) |
---|
456 | ifValuesPass=-1; // signal end of values pass after this |
---|
457 | } else { |
---|
458 | // end of values pass - initialize weights etc |
---|
459 | primalColumnPivot_->saveWeights(this,5); |
---|
460 | ifValuesPass=0; |
---|
461 | if(saveNumber != numberIterations_) { |
---|
462 | problemStatus_=-2; // factorize now |
---|
463 | pivotRow_=-1; // say no weights update |
---|
464 | returnCode=-4; |
---|
465 | break; |
---|
466 | } |
---|
467 | |
---|
468 | // and get variable |
---|
469 | primalColumn(rowArray_[1],rowArray_[2],rowArray_[3], |
---|
470 | columnArray_[0],columnArray_[1]); |
---|
471 | } |
---|
472 | } |
---|
473 | pivotRow_=-1; |
---|
474 | sequenceOut_=-1; |
---|
475 | rowArray_[1]->clear(); |
---|
476 | if (sequenceIn_>=0) { |
---|
477 | // we found a pivot column |
---|
478 | #ifdef CLP_DEBUG |
---|
479 | if ((handler_->logLevel()&32)) { |
---|
480 | char x = isColumn(sequenceIn_) ? 'C' :'R'; |
---|
481 | std::cout<<"pivot column "<< |
---|
482 | x<<sequenceWithin(sequenceIn_)<<std::endl; |
---|
483 | } |
---|
484 | #endif |
---|
485 | // update the incoming column |
---|
486 | unpack(rowArray_[1]); |
---|
487 | // save reduced cost |
---|
488 | double saveDj = dualIn_; |
---|
489 | factorization_->updateColumn(rowArray_[2],rowArray_[1],true); |
---|
490 | // do ratio test and re-compute dj |
---|
491 | primalRow(rowArray_[1],rowArray_[3],rowArray_[2],rowArray_[0], |
---|
492 | ifValuesPass); |
---|
493 | if (ifValuesPass) { |
---|
494 | saveDj=dualIn_; |
---|
495 | if (pivotRow_==-1||(pivotRow_>=0&&fabs(alpha_)<1.0e-5)) { |
---|
496 | if(fabs(dualIn_)<1.0e2*dualTolerance_) { |
---|
497 | // try other way |
---|
498 | directionIn_=-directionIn_; |
---|
499 | primalRow(rowArray_[1],rowArray_[3],rowArray_[2],rowArray_[0], |
---|
500 | 0); |
---|
501 | } |
---|
502 | if (pivotRow_==-1||(pivotRow_>=0&&fabs(alpha_)<1.0e-5)) { |
---|
503 | // reject it |
---|
504 | char x = isColumn(sequenceIn_) ? 'C' :'R'; |
---|
505 | handler_->message(CLP_SIMPLEX_FLAG,messages_) |
---|
506 | <<x<<sequenceWithin(sequenceIn_) |
---|
507 | <<OsiMessageEol; |
---|
508 | setFlagged(sequenceIn_); |
---|
509 | lastBadIteration_ = numberIterations_; // say be more cautious |
---|
510 | rowArray_[1]->clear(); |
---|
511 | pivotRow_=-1; |
---|
512 | continue; |
---|
513 | } |
---|
514 | } |
---|
515 | } |
---|
516 | |
---|
517 | #ifdef CLP_DEBUG |
---|
518 | if ((handler_->logLevel()&32)) |
---|
519 | printf("btran dj %g, ftran dj %g\n",saveDj,dualIn_); |
---|
520 | #endif |
---|
521 | if ((saveDj*dualIn_<1.0e-20&&!ifValuesPass)|| |
---|
522 | fabs(saveDj-dualIn_)>1.0e-5*(1.0+fabs(saveDj))) { |
---|
523 | handler_->message(CLP_PRIMAL_DJ,messages_) |
---|
524 | <<saveDj<<dualIn_ |
---|
525 | <<OsiMessageEol; |
---|
526 | if(saveNumber != numberIterations_) { |
---|
527 | problemStatus_=-2; // factorize now |
---|
528 | rowArray_[1]->clear(); |
---|
529 | pivotRow_=-1; // say no weights update |
---|
530 | returnCode=-2; |
---|
531 | break; |
---|
532 | } else { |
---|
533 | // take on more relaxed criterion |
---|
534 | if (saveDj*dualIn_<1.0e-20|| |
---|
535 | fabs(saveDj-dualIn_)>1.0e-4*(1.0+fabs(dualIn_))) { |
---|
536 | // need to reject something |
---|
537 | char x = isColumn(sequenceIn_) ? 'C' :'R'; |
---|
538 | handler_->message(CLP_SIMPLEX_FLAG,messages_) |
---|
539 | <<x<<sequenceWithin(sequenceIn_) |
---|
540 | <<OsiMessageEol; |
---|
541 | setFlagged(sequenceIn_); |
---|
542 | lastBadIteration_ = numberIterations_; // say be more cautious |
---|
543 | rowArray_[1]->clear(); |
---|
544 | pivotRow_=-1; |
---|
545 | continue; |
---|
546 | } |
---|
547 | } |
---|
548 | } |
---|
549 | if (pivotRow_>=0) { |
---|
550 | // if stable replace in basis |
---|
551 | int updateStatus = factorization_->replaceColumn(rowArray_[2], |
---|
552 | pivotRow_, |
---|
553 | alpha_); |
---|
554 | if (updateStatus==1) { |
---|
555 | // slight error |
---|
556 | if (factorization_->pivots()>5) { |
---|
557 | problemStatus_=-2; // factorize now |
---|
558 | returnCode=-3; |
---|
559 | } |
---|
560 | } else if (updateStatus==2) { |
---|
561 | // major error |
---|
562 | // later we may need to unwind more e.g. fake bounds |
---|
563 | if(saveNumber != numberIterations_) { |
---|
564 | problemStatus_=-2; // factorize now |
---|
565 | returnCode=-2; |
---|
566 | break; |
---|
567 | } else { |
---|
568 | // need to reject something |
---|
569 | char x = isColumn(sequenceIn_) ? 'C' :'R'; |
---|
570 | handler_->message(CLP_SIMPLEX_FLAG,messages_) |
---|
571 | <<x<<sequenceWithin(sequenceIn_) |
---|
572 | <<OsiMessageEol; |
---|
573 | setFlagged(sequenceIn_); |
---|
574 | lastBadIteration_ = numberIterations_; // say be more cautious |
---|
575 | rowArray_[1]->clear(); |
---|
576 | pivotRow_=-1; |
---|
577 | continue; |
---|
578 | } |
---|
579 | } else if (updateStatus==3) { |
---|
580 | // out of memory |
---|
581 | // increase space if not many iterations |
---|
582 | if (factorization_->pivots()< |
---|
583 | 0.5*factorization_->maximumPivots()&& |
---|
584 | factorization_->pivots()<200) |
---|
585 | factorization_->areaFactor( |
---|
586 | factorization_->areaFactor() * 1.1); |
---|
587 | problemStatus_=-2; // factorize now |
---|
588 | } |
---|
589 | // here do part of steepest - ready for next iteration |
---|
590 | primalColumnPivot_->updateWeights(rowArray_[1]); |
---|
591 | } else { |
---|
592 | if (pivotRow_==-1) { |
---|
593 | // no outgoing row is valid |
---|
594 | #ifdef CLP_DEBUG |
---|
595 | if (handler_->logLevel()&32) |
---|
596 | printf("** no row pivot\n"); |
---|
597 | #endif |
---|
598 | if (!factorization_->pivots()) { |
---|
599 | problemStatus_=-5; //say looks unbounded |
---|
600 | // do ray |
---|
601 | delete [] ray_; |
---|
602 | ray_ = new double [numberColumns_]; |
---|
603 | CoinFillN(ray_,numberColumns_,0.0); |
---|
604 | int number=rowArray_[1]->getNumElements(); |
---|
605 | int * index = rowArray_[1]->getIndices(); |
---|
606 | double * array = rowArray_[1]->denseVector(); |
---|
607 | double way=-directionIn_; |
---|
608 | int i; |
---|
609 | double zeroTolerance=1.0e-12; |
---|
610 | if (sequenceIn_<numberColumns_) |
---|
611 | ray_[sequenceIn_]=directionIn_; |
---|
612 | for (i=0;i<number;i++) { |
---|
613 | int iRow=index[i]; |
---|
614 | int iPivot=pivotVariable_[iRow]; |
---|
615 | double arrayValue = array[iRow]; |
---|
616 | if (iPivot<numberColumns_&&fabs(arrayValue)>=zeroTolerance) |
---|
617 | ray_[iPivot] = way* array[iRow]; |
---|
618 | } |
---|
619 | } |
---|
620 | rowArray_[0]->clear(); |
---|
621 | returnCode=2; |
---|
622 | break; |
---|
623 | } else { |
---|
624 | // flipping from bound to bound |
---|
625 | } |
---|
626 | } |
---|
627 | |
---|
628 | // update primal solution |
---|
629 | |
---|
630 | double objectiveChange=0.0; |
---|
631 | // Cost on pivot row may change - may need to change dualIn |
---|
632 | double oldCost=0.0; |
---|
633 | if (pivotRow_>=0) |
---|
634 | oldCost = cost(pivotVariable_[pivotRow_]); |
---|
635 | // rowArray_[1] is not empty - used to update djs |
---|
636 | updatePrimalsInPrimal(rowArray_[1],theta_, objectiveChange); |
---|
637 | if (pivotRow_>=0) |
---|
638 | dualIn_ += (oldCost-cost(pivotVariable_[pivotRow_])); |
---|
639 | |
---|
640 | int whatNext=housekeeping(objectiveChange); |
---|
641 | |
---|
642 | if (whatNext==1) { |
---|
643 | problemStatus_ =-2; // refactorize |
---|
644 | } else if (whatNext==2) { |
---|
645 | // maximum iterations or equivalent |
---|
646 | problemStatus_= 3; |
---|
647 | returnCode=3; |
---|
648 | break; |
---|
649 | } |
---|
650 | } else { |
---|
651 | // no pivot column |
---|
652 | #ifdef CLP_DEBUG |
---|
653 | if (handler_->logLevel()&32) |
---|
654 | printf("** no column pivot\n"); |
---|
655 | #endif |
---|
656 | if (nonLinearCost_->numberInfeasibilities()) |
---|
657 | problemStatus_=-4; // might be infeasible |
---|
658 | returnCode=0; |
---|
659 | break; |
---|
660 | } |
---|
661 | } |
---|
662 | return returnCode; |
---|
663 | } |
---|
664 | /* Checks if finished. Updates status */ |
---|
665 | void |
---|
666 | ClpSimplexPrimal::statusOfProblemInPrimal(int & lastCleaned,int type) |
---|
667 | { |
---|
668 | if (type==2) { |
---|
669 | // trouble - restore solution |
---|
670 | memcpy(status_ ,saveStatus_,(numberColumns_+numberRows_)*sizeof(char)); |
---|
671 | memcpy(rowActivityWork_,savedSolution_+numberColumns_ , |
---|
672 | numberRows_*sizeof(double)); |
---|
673 | memcpy(columnActivityWork_,savedSolution_ , |
---|
674 | numberColumns_*sizeof(double)); |
---|
675 | forceFactorization_=1; // a bit drastic but .. |
---|
676 | pivotRow_=-1; // say no weights update |
---|
677 | changeMade_++; // say change made |
---|
678 | } |
---|
679 | int tentativeStatus = problemStatus_; |
---|
680 | if (problemStatus_>-3||problemStatus_==-4) { |
---|
681 | // factorize |
---|
682 | // later on we will need to recover from singularities |
---|
683 | // also we could skip if first time |
---|
684 | // do weights |
---|
685 | // This may save pivotRow_ for use |
---|
686 | primalColumnPivot_->saveWeights(this,1); |
---|
687 | |
---|
688 | if (type) { |
---|
689 | // is factorization okay? |
---|
690 | if (internalFactorize(1)) { |
---|
691 | // no - restore previous basis |
---|
692 | assert (type==1); |
---|
693 | memcpy(status_ ,saveStatus_,(numberColumns_+numberRows_)*sizeof(char)); |
---|
694 | memcpy(rowActivityWork_,savedSolution_+numberColumns_ , |
---|
695 | numberRows_*sizeof(double)); |
---|
696 | memcpy(columnActivityWork_,savedSolution_ , |
---|
697 | numberColumns_*sizeof(double)); |
---|
698 | forceFactorization_=1; // a bit drastic but .. |
---|
699 | type = 2; |
---|
700 | assert (internalFactorize(1)==0); |
---|
701 | changeMade_++; // say change made |
---|
702 | } |
---|
703 | } |
---|
704 | if (problemStatus_!=-4) |
---|
705 | problemStatus_=-3; |
---|
706 | } |
---|
707 | // at this stage status is -3 or -5 if looks unbounded |
---|
708 | // get primal and dual solutions |
---|
709 | // put back original bounds and then check |
---|
710 | createRim(7); |
---|
711 | gutsOfSolution(rowActivityWork_, columnActivityWork_); |
---|
712 | handler_->message(CLP_SIMPLEX_STATUS,messages_) |
---|
713 | <<numberIterations_<<objectiveValue(); |
---|
714 | handler_->printing(sumPrimalInfeasibilities_>0.0) |
---|
715 | <<sumPrimalInfeasibilities_<<numberPrimalInfeasibilities_; |
---|
716 | handler_->printing(sumDualInfeasibilities_>0.0) |
---|
717 | <<sumDualInfeasibilities_<<numberDualInfeasibilities_; |
---|
718 | handler_->printing(numberDualInfeasibilitiesWithoutFree_ |
---|
719 | <numberDualInfeasibilities_) |
---|
720 | <<numberDualInfeasibilities_- |
---|
721 | numberDualInfeasibilitiesWithoutFree_; |
---|
722 | handler_->message()<<OsiMessageEol; |
---|
723 | assert (primalFeasible()); |
---|
724 | // we may wish to say it is optimal even if infeasible |
---|
725 | bool alwaysOptimal = (specialOptions_&1)!=0; |
---|
726 | if (dualFeasible()||problemStatus_==-4||(type==3&&problemStatus_!=-5)) { |
---|
727 | if (nonLinearCost_->numberInfeasibilities()&&!alwaysOptimal) { |
---|
728 | //may need infeasiblity cost changed |
---|
729 | // we can see if we can construct a ray |
---|
730 | // make up a new objective |
---|
731 | double saveWeight = infeasibilityCost_; |
---|
732 | // save nonlinear cost as we are going to switch off costs |
---|
733 | ClpNonLinearCost * nonLinear = nonLinearCost_; |
---|
734 | infeasibilityCost_=1.0e100; |
---|
735 | // put back original bounds |
---|
736 | createRim(7); |
---|
737 | nonLinearCost_->checkInfeasibilities(true); |
---|
738 | nonLinearCost_=NULL; |
---|
739 | // scale |
---|
740 | int i; |
---|
741 | for (i=0;i<numberRows_+numberColumns_;i++) |
---|
742 | cost_[i] *= 1.0e-100; |
---|
743 | gutsOfSolution(rowActivityWork_, columnActivityWork_); |
---|
744 | nonLinearCost_=nonLinear; |
---|
745 | infeasibilityCost_=saveWeight; |
---|
746 | if (infeasibilityCost_>=1.0e20|| |
---|
747 | numberDualInfeasibilities_==0) { |
---|
748 | // we are infeasible - use as ray |
---|
749 | delete [] ray_; |
---|
750 | ray_ = new double [numberRows_]; |
---|
751 | memcpy(ray_,dual_,numberRows_*sizeof(double)); |
---|
752 | // and get feasible duals |
---|
753 | infeasibilityCost_=0.0; |
---|
754 | createRim(7); |
---|
755 | nonLinearCost_->checkInfeasibilities(true); |
---|
756 | gutsOfSolution(rowActivityWork_, columnActivityWork_); |
---|
757 | // so will exit |
---|
758 | infeasibilityCost_=1.0e30; |
---|
759 | } |
---|
760 | |
---|
761 | if (infeasibilityCost_<1.0e20) { |
---|
762 | infeasibilityCost_ *= 5.0; |
---|
763 | changeMade_++; // say change made |
---|
764 | handler_->message(CLP_PRIMAL_WEIGHT,messages_) |
---|
765 | <<infeasibilityCost_ |
---|
766 | <<OsiMessageEol; |
---|
767 | // put back original bounds and then check |
---|
768 | createRim(7); |
---|
769 | nonLinearCost_->checkInfeasibilities(true); |
---|
770 | gutsOfSolution(rowActivityWork_, columnActivityWork_); |
---|
771 | problemStatus_=-1; //continue |
---|
772 | } else { |
---|
773 | // say infeasible |
---|
774 | problemStatus_ = 1; |
---|
775 | } |
---|
776 | } else { |
---|
777 | // may be optimal |
---|
778 | if ( lastCleaned!=numberIterations_) { |
---|
779 | handler_->message(CLP_PRIMAL_OPTIMAL,messages_) |
---|
780 | <<primalTolerance_ |
---|
781 | <<OsiMessageEol; |
---|
782 | if (numberTimesOptimal_<4) { |
---|
783 | numberTimesOptimal_++; |
---|
784 | changeMade_++; // say change made |
---|
785 | if (numberTimesOptimal_==1) { |
---|
786 | // better to have small tolerance even if slower |
---|
787 | factorization_->zeroTolerance(1.0e-15); |
---|
788 | } |
---|
789 | lastCleaned=numberIterations_; |
---|
790 | handler_->message(CLP_PRIMAL_ORIGINAL,messages_) |
---|
791 | <<OsiMessageEol; |
---|
792 | primalTolerance_=dblParam_[OsiPrimalTolerance]; |
---|
793 | |
---|
794 | // put back original bounds and then check |
---|
795 | createRim(7); |
---|
796 | nonLinearCost_->checkInfeasibilities(true); |
---|
797 | gutsOfSolution(rowActivityWork_, columnActivityWork_); |
---|
798 | problemStatus_ = -1; |
---|
799 | } else { |
---|
800 | problemStatus_=0; // optimal |
---|
801 | if (lastCleaned<numberIterations_) { |
---|
802 | handler_->message(CLP_SIMPLEX_GIVINGUP,messages_) |
---|
803 | <<OsiMessageEol; |
---|
804 | } |
---|
805 | } |
---|
806 | } else { |
---|
807 | problemStatus_=0; // optimal |
---|
808 | } |
---|
809 | } |
---|
810 | } else { |
---|
811 | // see if looks unbounded |
---|
812 | if (problemStatus_==-5) { |
---|
813 | if (nonLinearCost_->numberInfeasibilities()) { |
---|
814 | //we need infeasiblity cost changed |
---|
815 | if (infeasibilityCost_<1.0e20) { |
---|
816 | infeasibilityCost_ *= 5.0; |
---|
817 | changeMade_++; // say change made |
---|
818 | handler_->message(CLP_PRIMAL_WEIGHT,messages_) |
---|
819 | <<infeasibilityCost_ |
---|
820 | <<OsiMessageEol; |
---|
821 | // put back original bounds and then check |
---|
822 | createRim(7); |
---|
823 | gutsOfSolution(rowActivityWork_, columnActivityWork_); |
---|
824 | problemStatus_=-1; //continue |
---|
825 | } else { |
---|
826 | // say unbounded |
---|
827 | problemStatus_ = 2; |
---|
828 | } |
---|
829 | } else { |
---|
830 | // say unbounded |
---|
831 | problemStatus_ = 2; |
---|
832 | } |
---|
833 | } else { |
---|
834 | // carry on |
---|
835 | problemStatus_ = -1; |
---|
836 | } |
---|
837 | } |
---|
838 | if (type==0||type==1) { |
---|
839 | if (!type) { |
---|
840 | // create save arrays |
---|
841 | delete [] saveStatus_; |
---|
842 | delete [] savedSolution_; |
---|
843 | saveStatus_ = new unsigned char [numberRows_+numberColumns_]; |
---|
844 | savedSolution_ = new double [numberRows_+numberColumns_]; |
---|
845 | } |
---|
846 | // save arrays |
---|
847 | memcpy(saveStatus_,status_,(numberColumns_+numberRows_)*sizeof(char)); |
---|
848 | memcpy(savedSolution_+numberColumns_ ,rowActivityWork_, |
---|
849 | numberRows_*sizeof(double)); |
---|
850 | memcpy(savedSolution_ ,columnActivityWork_,numberColumns_*sizeof(double)); |
---|
851 | } |
---|
852 | // restore weights (if saved) - also recompute infeasibility list |
---|
853 | if (tentativeStatus>-3) |
---|
854 | primalColumnPivot_->saveWeights(this,(type <2) ? 2 : 4); |
---|
855 | else |
---|
856 | primalColumnPivot_->saveWeights(this,3); |
---|
857 | if (problemStatus_<0&&!changeMade_) { |
---|
858 | problemStatus_=4; // unknown |
---|
859 | } |
---|
860 | } |
---|
861 | /* |
---|
862 | Row array has pivot column |
---|
863 | This chooses pivot row. |
---|
864 | For speed, we may need to go to a bucket approach when many |
---|
865 | variables go through bounds |
---|
866 | On exit rhsArray will have changes in costs of basic variables |
---|
867 | */ |
---|
868 | void |
---|
869 | ClpSimplexPrimal::primalRow(OsiIndexedVector * rowArray, |
---|
870 | OsiIndexedVector * rhsArray, |
---|
871 | OsiIndexedVector * spareArray, |
---|
872 | OsiIndexedVector * spareArray2, |
---|
873 | int valuesPass) |
---|
874 | { |
---|
875 | if (valuesPass) { |
---|
876 | dualIn_ = cost_[sequenceIn_]; |
---|
877 | |
---|
878 | double * work=rowArray->denseVector(); |
---|
879 | int number=rowArray->getNumElements(); |
---|
880 | int * which=rowArray->getIndices(); |
---|
881 | |
---|
882 | int iIndex; |
---|
883 | |
---|
884 | for (iIndex=0;iIndex<number;iIndex++) { |
---|
885 | |
---|
886 | int iRow = which[iIndex]; |
---|
887 | double alpha = work[iRow]; |
---|
888 | int iPivot=pivotVariable_[iRow]; |
---|
889 | dualIn_ -= alpha*cost(iPivot); |
---|
890 | } |
---|
891 | // determine direction here |
---|
892 | if (dualIn_<-dualTolerance_) { |
---|
893 | directionIn_=1; |
---|
894 | } else if (dualIn_>dualTolerance_) { |
---|
895 | directionIn_=-1; |
---|
896 | } else { |
---|
897 | // towards nearest bound |
---|
898 | if (valueIn_-lowerIn_<upperIn_-valueIn_) { |
---|
899 | directionIn_=-1; |
---|
900 | dualIn_=dualTolerance_; |
---|
901 | } else { |
---|
902 | directionIn_=1; |
---|
903 | dualIn_=-dualTolerance_; |
---|
904 | } |
---|
905 | } |
---|
906 | } |
---|
907 | |
---|
908 | // sequence stays as row number until end |
---|
909 | pivotRow_=-1; |
---|
910 | int numberSwapped=0; |
---|
911 | int numberRemaining=0; |
---|
912 | |
---|
913 | int numberThru =0; // number gone thru a barrier |
---|
914 | int lastThru =0; // number gone thru a barrier on last time |
---|
915 | |
---|
916 | double totalThru=0.0; // for when variables flip |
---|
917 | double acceptablePivot=1.0e-7; |
---|
918 | if (factorization_->pivots()) |
---|
919 | acceptablePivot=1.0e-5; // if we have iterated be more strict |
---|
920 | double bestEverPivot=acceptablePivot; |
---|
921 | int lastPivotRow = -1; |
---|
922 | double lastPivot=0.0; |
---|
923 | double lastTheta=1.0e50; |
---|
924 | int lastNumberSwapped=0; |
---|
925 | |
---|
926 | // use spareArrays to put ones looked at in |
---|
927 | // First one is list of candidates |
---|
928 | // We could compress if we really know we won't need any more |
---|
929 | // Second array has current set of pivot candidates |
---|
930 | // with a backup list saved in double * part of indexed vector |
---|
931 | |
---|
932 | // for zeroing out arrays after |
---|
933 | int maximumSwapped=0; |
---|
934 | // pivot elements |
---|
935 | double * spare; |
---|
936 | // indices |
---|
937 | int * index, * indexSwapped; |
---|
938 | int * saveSwapped; |
---|
939 | spareArray->clear(); |
---|
940 | spareArray2->clear(); |
---|
941 | spare = spareArray->denseVector(); |
---|
942 | index = spareArray->getIndices(); |
---|
943 | saveSwapped = (int *) spareArray2->denseVector(); |
---|
944 | indexSwapped = spareArray2->getIndices(); |
---|
945 | |
---|
946 | // we also need somewhere for effective rhs |
---|
947 | double * rhs=rhsArray->denseVector(); |
---|
948 | |
---|
949 | /* |
---|
950 | First we get a list of possible pivots. We can also see if the |
---|
951 | problem looks unbounded. |
---|
952 | |
---|
953 | At first we increase theta and see what happens. We start |
---|
954 | theta at a reasonable guess. If in right area then we do bit by bit. |
---|
955 | We save possible pivot candidates |
---|
956 | |
---|
957 | */ |
---|
958 | |
---|
959 | // do first pass to get possibles |
---|
960 | // We can also see if unbounded |
---|
961 | // We also re-compute reduced cost |
---|
962 | |
---|
963 | dualIn_ = cost_[sequenceIn_]; |
---|
964 | |
---|
965 | double * work=rowArray->denseVector(); |
---|
966 | int number=rowArray->getNumElements(); |
---|
967 | int * which=rowArray->getIndices(); |
---|
968 | |
---|
969 | // we need to swap sign if coming in from ub |
---|
970 | double way = directionIn_; |
---|
971 | double maximumMovement; |
---|
972 | if (way>0.0) |
---|
973 | maximumMovement = min(1.0e30,upperIn_-valueIn_); |
---|
974 | else |
---|
975 | maximumMovement = min(1.0e30,valueIn_-lowerIn_); |
---|
976 | |
---|
977 | double tentativeTheta = maximumMovement; |
---|
978 | double upperTheta = maximumMovement; |
---|
979 | |
---|
980 | int iIndex; |
---|
981 | |
---|
982 | for (iIndex=0;iIndex<number;iIndex++) { |
---|
983 | |
---|
984 | int iRow = which[iIndex]; |
---|
985 | double alpha = work[iRow]; |
---|
986 | int iPivot=pivotVariable_[iRow]; |
---|
987 | dualIn_ -= alpha*cost(iPivot); |
---|
988 | alpha *= way; |
---|
989 | double oldValue = solution(iPivot); |
---|
990 | // get where in bound sequence |
---|
991 | if (alpha>0.0) { |
---|
992 | // basic variable going towards lower bound |
---|
993 | double bound = lower(iPivot); |
---|
994 | oldValue -= bound; |
---|
995 | } else if (alpha<0.0) { |
---|
996 | // basic variable going towards upper bound |
---|
997 | double bound = upper(iPivot); |
---|
998 | oldValue = bound-oldValue; |
---|
999 | } |
---|
1000 | double value = oldValue-tentativeTheta*fabs(alpha); |
---|
1001 | assert (oldValue>=-primalTolerance_*1.0001); |
---|
1002 | if (value<-primalTolerance_) { |
---|
1003 | // add to list |
---|
1004 | spare[numberRemaining]=alpha; |
---|
1005 | rhs[iRow]=oldValue; |
---|
1006 | index[numberRemaining++]=iRow; |
---|
1007 | double value=oldValue-upperTheta*fabs(alpha); |
---|
1008 | if (value<-primalTolerance_&&fabs(alpha)>=acceptablePivot) |
---|
1009 | upperTheta = (oldValue+primalTolerance_)/fabs(alpha); |
---|
1010 | } |
---|
1011 | } |
---|
1012 | |
---|
1013 | // we need to keep where rhs non-zeros are |
---|
1014 | int numberInRhs=numberRemaining; |
---|
1015 | memcpy(rhsArray->getIndices(),index,numberInRhs*sizeof(int)); |
---|
1016 | rhsArray->setNumElements(numberInRhs); |
---|
1017 | |
---|
1018 | theta_=maximumMovement; |
---|
1019 | |
---|
1020 | double dualCheck = fabs(dualIn_); |
---|
1021 | // but make a bit more pessimistic |
---|
1022 | dualCheck=max(dualCheck-100.0*dualTolerance_,0.99*dualCheck); |
---|
1023 | |
---|
1024 | bool goBackOne = false; |
---|
1025 | |
---|
1026 | if (numberRemaining) { |
---|
1027 | |
---|
1028 | // looks like pivoting |
---|
1029 | // now try until reasonable theta |
---|
1030 | tentativeTheta = max(10.0*upperTheta,1.0e-7); |
---|
1031 | tentativeTheta = min(tentativeTheta,maximumMovement); |
---|
1032 | |
---|
1033 | // loops increasing tentative theta until can't go through |
---|
1034 | |
---|
1035 | while (tentativeTheta <= maximumMovement) { |
---|
1036 | double thruThis = 0.0; |
---|
1037 | |
---|
1038 | double bestPivot=acceptablePivot; |
---|
1039 | pivotRow_ = -1; |
---|
1040 | |
---|
1041 | numberSwapped = 0; |
---|
1042 | |
---|
1043 | upperTheta = maximumMovement; |
---|
1044 | |
---|
1045 | for (iIndex=0;iIndex<numberRemaining;iIndex++) { |
---|
1046 | |
---|
1047 | int iRow = index[iIndex]; |
---|
1048 | double alpha = spare[iIndex]; |
---|
1049 | double oldValue = rhs[iRow]; |
---|
1050 | double value = oldValue-tentativeTheta*fabs(alpha); |
---|
1051 | |
---|
1052 | if (value<-primalTolerance_) { |
---|
1053 | // how much would it cost to go thru |
---|
1054 | thruThis += alpha* |
---|
1055 | nonLinearCost_->changeInCost(pivotVariable_[iRow],alpha); |
---|
1056 | // goes on swapped list (also means candidates if too many) |
---|
1057 | indexSwapped[numberSwapped++]=iRow; |
---|
1058 | if (fabs(alpha)>bestPivot) { |
---|
1059 | bestPivot=fabs(alpha); |
---|
1060 | pivotRow_ = iRow; |
---|
1061 | theta_ = oldValue/bestPivot; |
---|
1062 | } |
---|
1063 | } else { |
---|
1064 | value = oldValue-upperTheta*fabs(alpha); |
---|
1065 | if (value<-primalTolerance_ && fabs(alpha)>=acceptablePivot) |
---|
1066 | upperTheta = (oldValue+primalTolerance_)/fabs(alpha); |
---|
1067 | } |
---|
1068 | } |
---|
1069 | |
---|
1070 | maximumSwapped = max(maximumSwapped,numberSwapped); |
---|
1071 | |
---|
1072 | if (totalThru+thruThis>=dualCheck) { |
---|
1073 | // We should be pivoting in this batch |
---|
1074 | // so compress down to this lot |
---|
1075 | |
---|
1076 | int saveNumber = numberRemaining; |
---|
1077 | numberRemaining=0; |
---|
1078 | for (iIndex=0;iIndex<numberSwapped;iIndex++) { |
---|
1079 | int iRow = indexSwapped[iIndex]; |
---|
1080 | spare[numberRemaining]=way*work[iRow]; |
---|
1081 | index[numberRemaining++]=iRow; |
---|
1082 | } |
---|
1083 | memset(spare+numberRemaining,0, |
---|
1084 | (saveNumber-numberRemaining)*sizeof(double)); |
---|
1085 | int iTry; |
---|
1086 | #define MAXTRY 100 |
---|
1087 | // first get ratio with tolerance |
---|
1088 | for (iTry=0;iTry<MAXTRY;iTry++) { |
---|
1089 | |
---|
1090 | upperTheta=maximumMovement; |
---|
1091 | numberSwapped = 0; |
---|
1092 | |
---|
1093 | for (iIndex=0;iIndex<numberRemaining;iIndex++) { |
---|
1094 | |
---|
1095 | int iRow = index[iIndex]; |
---|
1096 | double alpha = fabs(spare[iIndex]); |
---|
1097 | double oldValue = rhs[iRow]; |
---|
1098 | double value = oldValue-upperTheta*alpha; |
---|
1099 | |
---|
1100 | if (value<-primalTolerance_ && alpha>=acceptablePivot) |
---|
1101 | upperTheta = (oldValue+primalTolerance_)/alpha; |
---|
1102 | |
---|
1103 | } |
---|
1104 | |
---|
1105 | // now look at best in this lot |
---|
1106 | bestPivot=acceptablePivot; |
---|
1107 | pivotRow_=-1; |
---|
1108 | for (iIndex=0;iIndex<numberRemaining;iIndex++) { |
---|
1109 | |
---|
1110 | int iRow = index[iIndex]; |
---|
1111 | double alpha = spare[iIndex]; |
---|
1112 | double oldValue = rhs[iRow]; |
---|
1113 | double value = oldValue-upperTheta*fabs(alpha); |
---|
1114 | |
---|
1115 | if (value<=0) { |
---|
1116 | // how much would it cost to go thru |
---|
1117 | totalThru += alpha* |
---|
1118 | nonLinearCost_->changeInCost(pivotVariable_[iRow],alpha); |
---|
1119 | // goes on swapped list (also means candidates if too many) |
---|
1120 | indexSwapped[numberSwapped++]=iRow; |
---|
1121 | if (fabs(alpha)>bestPivot) { |
---|
1122 | bestPivot=fabs(alpha); |
---|
1123 | theta_ = oldValue/bestPivot; |
---|
1124 | pivotRow_=iRow; |
---|
1125 | } |
---|
1126 | } else { |
---|
1127 | value = oldValue-upperTheta*fabs(alpha); |
---|
1128 | if (value<-primalTolerance_ && fabs(alpha)>=acceptablePivot) |
---|
1129 | upperTheta = (oldValue+primalTolerance_)/fabs(alpha); |
---|
1130 | } |
---|
1131 | } |
---|
1132 | |
---|
1133 | maximumSwapped = max(maximumSwapped,numberSwapped); |
---|
1134 | if (bestPivot<0.1*bestEverPivot&& |
---|
1135 | bestEverPivot>1.0e-6&&bestPivot<1.0e-3) { |
---|
1136 | // back to previous one |
---|
1137 | goBackOne = true; |
---|
1138 | break; |
---|
1139 | } else if (pivotRow_==-1&&upperTheta>largeValue_) { |
---|
1140 | if (lastPivot>acceptablePivot) { |
---|
1141 | // back to previous one |
---|
1142 | goBackOne = true; |
---|
1143 | //break; |
---|
1144 | } else { |
---|
1145 | // can only get here if all pivots so far too small |
---|
1146 | } |
---|
1147 | break; |
---|
1148 | } else if (totalThru>=dualCheck) { |
---|
1149 | break; // no point trying another loop |
---|
1150 | } else { |
---|
1151 | // skip this lot |
---|
1152 | nonLinearCost_->goThru(numberSwapped,way,indexSwapped, work,rhs); |
---|
1153 | lastPivotRow=pivotRow_; |
---|
1154 | lastTheta = theta_; |
---|
1155 | lastThru = numberThru; |
---|
1156 | numberThru += numberSwapped; |
---|
1157 | lastNumberSwapped = numberSwapped; |
---|
1158 | memcpy(saveSwapped,indexSwapped,lastNumberSwapped*sizeof(int)); |
---|
1159 | if (bestPivot>bestEverPivot) |
---|
1160 | bestEverPivot=bestPivot; |
---|
1161 | } |
---|
1162 | } |
---|
1163 | break; |
---|
1164 | } else { |
---|
1165 | // skip this lot |
---|
1166 | nonLinearCost_->goThru(numberSwapped,way,indexSwapped, work,rhs); |
---|
1167 | lastPivotRow=pivotRow_; |
---|
1168 | lastTheta = theta_; |
---|
1169 | lastThru = numberThru; |
---|
1170 | numberThru += numberSwapped; |
---|
1171 | lastNumberSwapped = numberSwapped; |
---|
1172 | memcpy(saveSwapped,indexSwapped,lastNumberSwapped*sizeof(int)); |
---|
1173 | if (bestPivot>bestEverPivot) |
---|
1174 | bestEverPivot=bestPivot; |
---|
1175 | totalThru += thruThis; |
---|
1176 | tentativeTheta = 2.0*upperTheta; |
---|
1177 | } |
---|
1178 | } |
---|
1179 | // can get here without pivotRow_ set but with lastPivotRow |
---|
1180 | if (goBackOne||(pivotRow_<0&&lastPivotRow>=0)) { |
---|
1181 | // back to previous one |
---|
1182 | pivotRow_=lastPivotRow; |
---|
1183 | theta_ = lastTheta; |
---|
1184 | // undo this lot |
---|
1185 | nonLinearCost_->goBack(lastNumberSwapped,saveSwapped,rhs); |
---|
1186 | memcpy(indexSwapped,saveSwapped,lastNumberSwapped*sizeof(int)); |
---|
1187 | numberSwapped = lastNumberSwapped; |
---|
1188 | } |
---|
1189 | } |
---|
1190 | |
---|
1191 | if (pivotRow_>=0) { |
---|
1192 | |
---|
1193 | #define MINIMUMTHETA 1.0e-12 |
---|
1194 | // will we need to increase tolerance |
---|
1195 | #ifdef CLP_DEBUG |
---|
1196 | bool found=false; |
---|
1197 | #endif |
---|
1198 | double largestInfeasibility = primalTolerance_; |
---|
1199 | if (theta_<MINIMUMTHETA) { |
---|
1200 | theta_=MINIMUMTHETA; |
---|
1201 | for (iIndex=0;iIndex<numberSwapped;iIndex++) { |
---|
1202 | int iRow = indexSwapped[iIndex]; |
---|
1203 | #ifdef CLP_DEBUG |
---|
1204 | if (iRow==pivotRow_) |
---|
1205 | found=true; |
---|
1206 | #endif |
---|
1207 | largestInfeasibility = max (largestInfeasibility, |
---|
1208 | -(rhs[iRow]-fabs(work[iRow])*theta_)); |
---|
1209 | } |
---|
1210 | #ifdef CLP_DEBUG |
---|
1211 | assert(found); |
---|
1212 | if (largestInfeasibility>primalTolerance_&&(handler_->logLevel()&32)) |
---|
1213 | printf("Primal tolerance increased from %g to %g\n", |
---|
1214 | primalTolerance_,largestInfeasibility); |
---|
1215 | #endif |
---|
1216 | primalTolerance_ = max(primalTolerance_,largestInfeasibility); |
---|
1217 | } |
---|
1218 | alpha_ = work[pivotRow_]; |
---|
1219 | // translate to sequence |
---|
1220 | sequenceOut_ = pivotVariable_[pivotRow_]; |
---|
1221 | valueOut_ = solution(sequenceOut_); |
---|
1222 | if (way<0.0) |
---|
1223 | theta_ = - theta_; |
---|
1224 | double newValue = valueOut_ - theta_*alpha_; |
---|
1225 | if (alpha_*way<0.0) { |
---|
1226 | directionOut_=-1; // to upper bound |
---|
1227 | if (fabs(theta_)>0.1) |
---|
1228 | upperOut_ = nonLinearCost_->nearest(sequenceOut_,newValue); |
---|
1229 | else |
---|
1230 | upperOut_ = newValue; |
---|
1231 | } else { |
---|
1232 | directionOut_=1; // to lower bound |
---|
1233 | if (fabs(theta_)>0.1) |
---|
1234 | lowerOut_ = nonLinearCost_->nearest(sequenceOut_,newValue); |
---|
1235 | else |
---|
1236 | lowerOut_ = newValue; |
---|
1237 | } |
---|
1238 | dualOut_ = reducedCost(sequenceOut_); |
---|
1239 | } else if (maximumMovement<1.0e20) { |
---|
1240 | // flip |
---|
1241 | pivotRow_ = -2; // so we can tell its a flip |
---|
1242 | sequenceOut_ = sequenceIn_; |
---|
1243 | valueOut_ = valueIn_; |
---|
1244 | dualOut_ = dualIn_; |
---|
1245 | lowerOut_ = lowerIn_; |
---|
1246 | upperOut_ = upperIn_; |
---|
1247 | alpha_ = 0.0; |
---|
1248 | if (way<0.0) { |
---|
1249 | directionOut_=1; // to lower bound |
---|
1250 | theta_ = lowerOut_ - valueOut_; |
---|
1251 | } else { |
---|
1252 | directionOut_=-1; // to upper bound |
---|
1253 | theta_ = upperOut_ - valueOut_; |
---|
1254 | } |
---|
1255 | } |
---|
1256 | |
---|
1257 | // clear arrays |
---|
1258 | |
---|
1259 | memset(spare,0,numberRemaining*sizeof(double)); |
---|
1260 | memset(saveSwapped,0,maximumSwapped*sizeof(int)); |
---|
1261 | |
---|
1262 | // put back original bounds etc |
---|
1263 | nonLinearCost_->goBackAll(rhsArray); |
---|
1264 | |
---|
1265 | rhsArray->clear(); |
---|
1266 | |
---|
1267 | } |
---|
1268 | /* |
---|
1269 | Chooses primal pivot column |
---|
1270 | updateArray has cost updates (also use pivotRow_ from last iteration) |
---|
1271 | Would be faster with separate region to scan |
---|
1272 | and will have this (with square of infeasibility) when steepest |
---|
1273 | For easy problems we can just choose one of the first columns we look at |
---|
1274 | */ |
---|
1275 | void |
---|
1276 | ClpSimplexPrimal::primalColumn(OsiIndexedVector * updates, |
---|
1277 | OsiIndexedVector * spareRow1, |
---|
1278 | OsiIndexedVector * spareRow2, |
---|
1279 | OsiIndexedVector * spareColumn1, |
---|
1280 | OsiIndexedVector * spareColumn2) |
---|
1281 | { |
---|
1282 | sequenceIn_ = primalColumnPivot_->pivotColumn(updates,spareRow1, |
---|
1283 | spareRow2,spareColumn1, |
---|
1284 | spareColumn2); |
---|
1285 | if (sequenceIn_>=0) { |
---|
1286 | valueIn_=solution_[sequenceIn_]; |
---|
1287 | lowerIn_=lower_[sequenceIn_]; |
---|
1288 | upperIn_=upper_[sequenceIn_]; |
---|
1289 | dualIn_=dj_[sequenceIn_]; |
---|
1290 | if (dualIn_>0.0) |
---|
1291 | directionIn_ = -1; |
---|
1292 | else |
---|
1293 | directionIn_ = 1; |
---|
1294 | } else { |
---|
1295 | sequenceIn_ = -1; |
---|
1296 | } |
---|
1297 | } |
---|
1298 | /* The primals are updated by the given array. |
---|
1299 | Returns number of infeasibilities. |
---|
1300 | After rowArray will have list of cost changes |
---|
1301 | */ |
---|
1302 | int |
---|
1303 | ClpSimplexPrimal::updatePrimalsInPrimal(OsiIndexedVector * rowArray, |
---|
1304 | double theta, |
---|
1305 | double & objectiveChange) |
---|
1306 | { |
---|
1307 | double * work=rowArray->denseVector(); |
---|
1308 | int number=rowArray->getNumElements(); |
---|
1309 | int * which=rowArray->getIndices(); |
---|
1310 | |
---|
1311 | int newNumber = 0; |
---|
1312 | |
---|
1313 | nonLinearCost_->setChangeInCost(0.0); |
---|
1314 | int iIndex; |
---|
1315 | |
---|
1316 | for (iIndex=0;iIndex<number;iIndex++) { |
---|
1317 | |
---|
1318 | int iRow = which[iIndex]; |
---|
1319 | double alpha = work[iRow]; |
---|
1320 | int iPivot=pivotVariable_[iRow]; |
---|
1321 | double & value = solutionAddress(iPivot); |
---|
1322 | double change = theta*alpha; |
---|
1323 | value -= change; |
---|
1324 | |
---|
1325 | if (change>0.0) { |
---|
1326 | // going down |
---|
1327 | if (value<=lower(iPivot)+primalTolerance_) { |
---|
1328 | double difference = nonLinearCost_->setOne(iPivot,value); |
---|
1329 | work[iRow] = difference; |
---|
1330 | if (difference) { |
---|
1331 | //change reduced cost on this |
---|
1332 | reducedCostAddress(iPivot) = -difference; |
---|
1333 | which[newNumber++]=iRow; |
---|
1334 | } |
---|
1335 | } else { |
---|
1336 | work[iRow]=0.0; |
---|
1337 | } |
---|
1338 | } else { |
---|
1339 | // going up |
---|
1340 | if (value>=upper(iPivot)-primalTolerance_) { |
---|
1341 | double difference = nonLinearCost_->setOne(iPivot,value); |
---|
1342 | work[iRow] = difference; |
---|
1343 | if (difference) { |
---|
1344 | //change reduced cost on this |
---|
1345 | reducedCostAddress(iPivot) = -difference; |
---|
1346 | which[newNumber++]=iRow; |
---|
1347 | } |
---|
1348 | } else { |
---|
1349 | work[iRow]=0.0; |
---|
1350 | } |
---|
1351 | } |
---|
1352 | } |
---|
1353 | objectiveChange += nonLinearCost_->changeInCost(); |
---|
1354 | rowArray->setNumElements(newNumber); |
---|
1355 | return 0; |
---|
1356 | } |
---|
1357 | void |
---|
1358 | ClpSimplexPrimal::nextSuperBasic(int & firstSuperBasic) |
---|
1359 | { |
---|
1360 | int iColumn; |
---|
1361 | if (firstSuperBasic==numberRows_+numberColumns_) { |
---|
1362 | // initialization |
---|
1363 | iColumn=0; |
---|
1364 | } else { |
---|
1365 | // normal |
---|
1366 | sequenceIn_=firstSuperBasic; |
---|
1367 | valueIn_=solution_[sequenceIn_]; |
---|
1368 | lowerIn_=lower_[sequenceIn_]; |
---|
1369 | upperIn_=upper_[sequenceIn_]; |
---|
1370 | dualIn_=dj_[sequenceIn_]; |
---|
1371 | iColumn=firstSuperBasic+1; |
---|
1372 | } |
---|
1373 | for (;iColumn<numberRows_+numberColumns_;iColumn++) { |
---|
1374 | if (getStatus(iColumn)==ClpSimplex::superBasic) { |
---|
1375 | // is it really super basic |
---|
1376 | if (fabs(solution_[iColumn]-lower_[iColumn])<=primalTolerance_) { |
---|
1377 | solution_[iColumn]=lower_[iColumn]; |
---|
1378 | setStatus(iColumn,ClpSimplex::atLowerBound); |
---|
1379 | } else if (fabs(solution_[iColumn]-upper_[iColumn]) |
---|
1380 | <=primalTolerance_) { |
---|
1381 | solution_[iColumn]=upper_[iColumn]; |
---|
1382 | setStatus(iColumn,ClpSimplex::atUpperBound); |
---|
1383 | } else if (lower_[iColumn]<-1.0e20&&upper_[iColumn]>1.0e20) { |
---|
1384 | setStatus(iColumn,ClpSimplex::isFree); |
---|
1385 | } else { |
---|
1386 | break; |
---|
1387 | } |
---|
1388 | } |
---|
1389 | } |
---|
1390 | firstSuperBasic = iColumn; |
---|
1391 | } |
---|
1392 | // Perturbs problem |
---|
1393 | void |
---|
1394 | ClpSimplexPrimal::perturb() |
---|
1395 | { |
---|
1396 | if (perturbation_>100) |
---|
1397 | return; //perturbed already |
---|
1398 | abort(); |
---|
1399 | } |
---|
1400 | // Do not change infeasibility cost and always say optimal |
---|
1401 | void |
---|
1402 | ClpSimplexPrimal::alwaysOptimal(bool onOff) |
---|
1403 | { |
---|
1404 | if (onOff) |
---|
1405 | specialOptions_ |= 1; |
---|
1406 | else |
---|
1407 | specialOptions_ &= ~1; |
---|
1408 | } |
---|
1409 | |
---|