1 | // Copyright (C) 2002, International Business Machines |
---|
2 | // Corporation and others. All Rights Reserved. |
---|
3 | |
---|
4 | |
---|
5 | /* Notes on implementation of dual simplex algorithm. |
---|
6 | |
---|
7 | When dual feasible: |
---|
8 | |
---|
9 | If primal feasible, we are optimal. Otherwise choose an infeasible |
---|
10 | basic variable to leave basis (normally going to nearest bound) (B). We |
---|
11 | now need to find an incoming variable which will leave problem |
---|
12 | dual feasible so we get the row of the tableau corresponding to |
---|
13 | the basic variable (with the correct sign depending if basic variable |
---|
14 | above or below feasibility region - as that affects whether reduced |
---|
15 | cost on outgoing variable has to be positive or negative). |
---|
16 | |
---|
17 | We now perform a ratio test to determine which incoming variable will |
---|
18 | preserve dual feasibility (C). If no variable found then problem |
---|
19 | is infeasible (in primal sense). If there is a variable, we then |
---|
20 | perform pivot and repeat. Trivial? |
---|
21 | |
---|
22 | ------------------------------------------- |
---|
23 | |
---|
24 | A) How do we get dual feasible? If all variables have bounds then |
---|
25 | it is trivial to get feasible by putting non-basic variables to |
---|
26 | correct bounds. OSL did not have a phase 1/phase 2 approach but |
---|
27 | instead effectively put fake bounds on variables and this is the |
---|
28 | approach here, although I had hoped to make it cleaner. |
---|
29 | |
---|
30 | If there is a weight of X on getting dual feasible: |
---|
31 | Non-basic variables with negative reduced costs are put to |
---|
32 | lesser of their upper bound and their lower bound + X. |
---|
33 | Similarly, mutatis mutandis, for positive reduced costs. |
---|
34 | |
---|
35 | Free variables should normally be in basis, otherwise I have |
---|
36 | coding which may be able to come out (and may not be correct). |
---|
37 | |
---|
38 | In OSL, this weight was changed heuristically, here at present |
---|
39 | it is only increased if problem looks finished. If problem is |
---|
40 | feasible I check for unboundedness. If not unbounded we |
---|
41 | could play with going into primal. As long as weights increase |
---|
42 | any algorithm would be finite. |
---|
43 | |
---|
44 | B) Which outgoing variable to choose is a virtual base class. |
---|
45 | For difficult problems steepest edge is preferred while for |
---|
46 | very easy (large) problems we will need partial scan. |
---|
47 | |
---|
48 | C) Sounds easy, but this is hardest part of algorithm. |
---|
49 | 1) Instead of stopping at first choice, we may be able |
---|
50 | to flip that variable to other bound and if objective |
---|
51 | still improving choose again. These mini iterations can |
---|
52 | increase speed by orders of magnitude but we may need to |
---|
53 | go to more of a bucket choice of variable rather than looking |
---|
54 | at them one by one (for speed). |
---|
55 | 2) Accuracy. Reduced costs may be of wrong sign but less than |
---|
56 | tolerance. Pivoting on these makes objective go backwards. |
---|
57 | OSL modified cost so a zero move was made, Gill et al |
---|
58 | (in primal analogue) modified so a strictly positive move was |
---|
59 | made. It is not quite as neat in dual but that is what we |
---|
60 | try and do. The two problems are that re-factorizations can |
---|
61 | change reduced costs above and below tolerances and that when |
---|
62 | finished we need to reset costs and try again. |
---|
63 | 3) Degeneracy. Gill et al helps but may not be enough. We |
---|
64 | may need more. Also it can improve speed a lot if we perturb |
---|
65 | the costs significantly. |
---|
66 | |
---|
67 | References: |
---|
68 | Forrest and Goldfarb, Steepest-edge simplex algorithms for |
---|
69 | linear programming - Mathematical Programming 1992 |
---|
70 | Forrest and Tomlin, Implementing the simplex method for |
---|
71 | the Optimization Subroutine Library - IBM Systems Journal 1992 |
---|
72 | Gill, Murray, Saunders, Wright A Practical Anti-Cycling |
---|
73 | Procedure for Linear and Nonlinear Programming SOL report 1988 |
---|
74 | |
---|
75 | |
---|
76 | TODO: |
---|
77 | |
---|
78 | a) Better recovery procedures. At present I never check on forward |
---|
79 | progress. There is checkpoint/restart with reducing |
---|
80 | re-factorization frequency, but this is only on singular |
---|
81 | factorizations. |
---|
82 | b) Fast methods for large easy problems (and also the option for |
---|
83 | the code to automatically choose which method). |
---|
84 | c) We need to be able to stop in various ways for OSI - this |
---|
85 | is fairly easy. |
---|
86 | |
---|
87 | */ |
---|
88 | |
---|
89 | #if defined(_MSC_VER) |
---|
90 | // Turn off compiler warning about long names |
---|
91 | # pragma warning(disable:4786) |
---|
92 | #endif |
---|
93 | |
---|
94 | #include <math.h> |
---|
95 | |
---|
96 | #include "CoinHelperFunctions.hpp" |
---|
97 | #include "ClpSimplexDual.hpp" |
---|
98 | #include "ClpFactorization.hpp" |
---|
99 | #include "OsiPackedMatrix.hpp" |
---|
100 | #include "OsiIndexedVector.hpp" |
---|
101 | #include "OsiWarmStartBasis.hpp" |
---|
102 | #include "ClpDualRowDantzig.hpp" |
---|
103 | #include "ClpMessage.hpp" |
---|
104 | #include <cfloat> |
---|
105 | #include <cassert> |
---|
106 | #include <string> |
---|
107 | #include <stdio.h> |
---|
108 | #include <iostream> |
---|
109 | // This returns a non const array filled with input from scalar |
---|
110 | // or actual array |
---|
111 | template <class T> inline T* |
---|
112 | copyOfArray( const T * array, const int size, T value) |
---|
113 | { |
---|
114 | T * arrayNew = new T[size]; |
---|
115 | if (array) |
---|
116 | CoinDisjointCopyN(array,size,arrayNew); |
---|
117 | else |
---|
118 | CoinFillN ( arrayNew, size,value); |
---|
119 | return arrayNew; |
---|
120 | } |
---|
121 | |
---|
122 | // This returns a non const array filled with actual array (or NULL) |
---|
123 | template <class T> inline T* |
---|
124 | copyOfArray( const T * array, const int size) |
---|
125 | { |
---|
126 | if (array) { |
---|
127 | T * arrayNew = new T[size]; |
---|
128 | CoinDisjointCopyN(array,size,arrayNew); |
---|
129 | return arrayNew; |
---|
130 | } else { |
---|
131 | return NULL; |
---|
132 | } |
---|
133 | } |
---|
134 | // dual |
---|
135 | int ClpSimplexDual::dual ( ) |
---|
136 | { |
---|
137 | |
---|
138 | /* *** Method |
---|
139 | This is a vanilla version of dual simplex. |
---|
140 | |
---|
141 | It tries to be a single phase approach with a weight of 1.0 being |
---|
142 | given to getting optimal and a weight of dualBound_ being |
---|
143 | given to getting dual feasible. In this version I have used the |
---|
144 | idea that this weight can be thought of as a fake bound. If the |
---|
145 | distance between the lower and upper bounds on a variable is less |
---|
146 | than the feasibility weight then we are always better off flipping |
---|
147 | to other bound to make dual feasible. If the distance is greater |
---|
148 | then we make up a fake bound dualBound_ away from one bound. |
---|
149 | If we end up optimal or primal infeasible, we check to see if |
---|
150 | bounds okay. If so we have finished, if not we increase dualBound_ |
---|
151 | and continue (after checking if unbounded). I am undecided about |
---|
152 | free variables - there is coding but I am not sure about it. At |
---|
153 | present I put them in basis anyway. |
---|
154 | |
---|
155 | The code is designed to take advantage of sparsity so arrays are |
---|
156 | seldom zeroed out from scratch or gone over in their entirety. |
---|
157 | The only exception is a full scan to find outgoing variable. This |
---|
158 | will be changed to keep an updated list of infeasibilities (or squares |
---|
159 | if steepest edge). Also on easy problems we don't need full scan - just |
---|
160 | pick first reasonable. |
---|
161 | |
---|
162 | One problem is how to tackle degeneracy and accuracy. At present |
---|
163 | I am using the modification of costs which I put in OSL and which was |
---|
164 | extended by Gill et al. I am still not sure of the exact details. |
---|
165 | |
---|
166 | The flow of dual is three while loops as follows: |
---|
167 | |
---|
168 | while (not finished) { |
---|
169 | |
---|
170 | while (not clean solution) { |
---|
171 | |
---|
172 | Factorize and/or clean up solution by flipping variables so |
---|
173 | dual feasible. If looks finished check fake dual bounds. |
---|
174 | Repeat until status is iterating (-1) or finished (0,1,2) |
---|
175 | |
---|
176 | } |
---|
177 | |
---|
178 | while (status==-1) { |
---|
179 | |
---|
180 | Iterate until no pivot in or out or time to re-factorize. |
---|
181 | |
---|
182 | Flow is: |
---|
183 | |
---|
184 | choose pivot row (outgoing variable). if none then |
---|
185 | we are primal feasible so looks as if done but we need to |
---|
186 | break and check bounds etc. |
---|
187 | |
---|
188 | Get pivot row in tableau |
---|
189 | |
---|
190 | Choose incoming column. If we don't find one then we look |
---|
191 | primal infeasible so break and check bounds etc. (Also the |
---|
192 | pivot tolerance is larger after any iterations so that may be |
---|
193 | reason) |
---|
194 | |
---|
195 | If we do find incoming column, we may have to adjust costs to |
---|
196 | keep going forwards (anti-degeneracy). Check pivot will be stable |
---|
197 | and if unstable throw away iteration (we will need to implement |
---|
198 | flagging of basic variables sometime) and break to re-factorize. |
---|
199 | If minor error re-factorize after iteration. |
---|
200 | |
---|
201 | Update everything (this may involve flipping variables to stay |
---|
202 | dual feasible. |
---|
203 | |
---|
204 | } |
---|
205 | |
---|
206 | } |
---|
207 | |
---|
208 | At present we never check we are going forwards. I overdid that in |
---|
209 | OSL so will try and make a last resort. |
---|
210 | |
---|
211 | Needs partial scan pivot out option. |
---|
212 | Needs dantzig, uninitialized and full steepest edge options (can still |
---|
213 | use partial scan) |
---|
214 | |
---|
215 | May need other anti-degeneracy measures, especially if we try and use |
---|
216 | loose tolerances as a way to solve in fewer iterations. |
---|
217 | |
---|
218 | I like idea of dynamic scaling. This gives opportunity to decouple |
---|
219 | different implications of scaling for accuracy, iteration count and |
---|
220 | feasibility tolerance. |
---|
221 | |
---|
222 | */ |
---|
223 | |
---|
224 | |
---|
225 | // sanity check |
---|
226 | assert (numberRows_==matrix_->getNumRows()); |
---|
227 | assert (numberColumns_==matrix_->getNumCols()); |
---|
228 | // for moment all arrays must exist |
---|
229 | assert(columnLower_); |
---|
230 | assert(columnUpper_); |
---|
231 | assert(rowLower_); |
---|
232 | assert(rowUpper_); |
---|
233 | |
---|
234 | #ifdef CLP_DEBUG |
---|
235 | int debugIteration=-1; |
---|
236 | #endif |
---|
237 | |
---|
238 | algorithm_ = -1; |
---|
239 | dualTolerance_=dblParam_[OsiDualTolerance]; |
---|
240 | primalTolerance_=dblParam_[OsiPrimalTolerance]; |
---|
241 | |
---|
242 | // put in standard form (and make row copy) |
---|
243 | // create modifiable copies of model rim and do optional scaling |
---|
244 | createRim(7+8+16,true); |
---|
245 | |
---|
246 | // save dual bound |
---|
247 | double saveDualBound_ = dualBound_; |
---|
248 | |
---|
249 | int iRow,iColumn; |
---|
250 | // Do initial factorization |
---|
251 | // and set certain stuff |
---|
252 | // We can either set increasing rows so ...IsBasic gives pivot row |
---|
253 | // or we can just increment iBasic one by one |
---|
254 | // for now let ...iBasic give pivot row |
---|
255 | factorization_->increasingRows(2); |
---|
256 | // row activities have negative sign |
---|
257 | factorization_->slackValue(-1.0); |
---|
258 | factorization_->zeroTolerance(1.0e-13); |
---|
259 | if (internalFactorize(0)) |
---|
260 | return 1; // some error |
---|
261 | |
---|
262 | // If user asked for perturbation - do it |
---|
263 | int savePerturbation = perturbation_; |
---|
264 | |
---|
265 | if (perturbation_<100) |
---|
266 | perturb(); |
---|
267 | |
---|
268 | double objectiveChange; |
---|
269 | // for dual we will change bounds using dualBound_ |
---|
270 | // for this we need clean basis so it is after factorize |
---|
271 | gutsOfSolution(rowActivityWork_,columnActivityWork_); |
---|
272 | changeBounds(true,NULL,objectiveChange); |
---|
273 | |
---|
274 | problemStatus_ = -1; |
---|
275 | numberIterations_=0; |
---|
276 | |
---|
277 | int lastCleaned=0; // last time objective or bounds cleaned up |
---|
278 | |
---|
279 | // number of times we have declared optimality |
---|
280 | numberTimesOptimal_=0; |
---|
281 | |
---|
282 | // This says whether to restore things etc |
---|
283 | int factorType=0; |
---|
284 | /* |
---|
285 | Status of problem: |
---|
286 | 0 - optimal |
---|
287 | 1 - infeasible |
---|
288 | 2 - unbounded |
---|
289 | -1 - iterating |
---|
290 | -2 - factorization wanted |
---|
291 | -3 - redo checking without factorization |
---|
292 | -4 - looks infeasible |
---|
293 | */ |
---|
294 | while (problemStatus_<0) { |
---|
295 | // clear |
---|
296 | for (iRow=0;iRow<4;iRow++) { |
---|
297 | rowArray_[iRow]->clear(); |
---|
298 | } |
---|
299 | |
---|
300 | for (iColumn=0;iColumn<2;iColumn++) { |
---|
301 | columnArray_[iColumn]->clear(); |
---|
302 | } |
---|
303 | |
---|
304 | // give matrix (and model costs and bounds a chance to be |
---|
305 | // refreshed (normally null) |
---|
306 | matrix_->refresh(this); |
---|
307 | // If getting nowhere - why not give it a kick |
---|
308 | #if 0 |
---|
309 | // does not seem to work too well - do some more work |
---|
310 | if (perturbation_<101&&numberIterations_>2*(numberRows_+numberColumns_)) |
---|
311 | perturb(); |
---|
312 | #endif |
---|
313 | // may factorize, checks if problem finished |
---|
314 | statusOfProblemInDual(lastCleaned,factorType); |
---|
315 | |
---|
316 | // Say good factorization |
---|
317 | factorType=1; |
---|
318 | |
---|
319 | // status stays at -1 while iterating, >=0 finished, -2 to invert |
---|
320 | // status -3 to go to top without an invert |
---|
321 | while (problemStatus_==-1) { |
---|
322 | #ifdef CLP_DEBUG |
---|
323 | { |
---|
324 | int i; |
---|
325 | for (i=0;i<4;i++) { |
---|
326 | rowArray_[i]->checkClear(); |
---|
327 | } |
---|
328 | for (i=0;i<2;i++) { |
---|
329 | columnArray_[i]->checkClear(); |
---|
330 | } |
---|
331 | } |
---|
332 | #endif |
---|
333 | #if CLP_DEBUG>2 |
---|
334 | // very expensive |
---|
335 | if (numberIterations_>0&&numberIterations_<-801) { |
---|
336 | handler_->setLogLevel(63); |
---|
337 | double saveValue = objectiveValue_; |
---|
338 | double * saveRow1 = new double[numberRows_]; |
---|
339 | double * saveRow2 = new double[numberRows_]; |
---|
340 | memcpy(saveRow1,rowReducedCost_,numberRows_*sizeof(double)); |
---|
341 | memcpy(saveRow2,rowActivityWork_,numberRows_*sizeof(double)); |
---|
342 | double * saveColumn1 = new double[numberColumns_]; |
---|
343 | double * saveColumn2 = new double[numberColumns_]; |
---|
344 | memcpy(saveColumn1,reducedCostWork_,numberColumns_*sizeof(double)); |
---|
345 | memcpy(saveColumn2,columnActivityWork_,numberColumns_*sizeof(double)); |
---|
346 | gutsOfSolution(rowActivityWork_,columnActivityWork_); |
---|
347 | printf("xxx %d old obj %g, recomputed %g, sum dual inf %g\n", |
---|
348 | numberIterations_, |
---|
349 | saveValue,objectiveValue_,sumDualInfeasibilities_); |
---|
350 | if (saveValue>objectiveValue_+1.0e-2) |
---|
351 | printf("**bad**\n"); |
---|
352 | memcpy(rowReducedCost_,saveRow1,numberRows_*sizeof(double)); |
---|
353 | memcpy(rowActivityWork_,saveRow2,numberRows_*sizeof(double)); |
---|
354 | memcpy(reducedCostWork_,saveColumn1,numberColumns_*sizeof(double)); |
---|
355 | memcpy(columnActivityWork_,saveColumn2,numberColumns_*sizeof(double)); |
---|
356 | delete [] saveRow1; |
---|
357 | delete [] saveRow2; |
---|
358 | delete [] saveColumn1; |
---|
359 | delete [] saveColumn2; |
---|
360 | objectiveValue_=saveValue; |
---|
361 | } |
---|
362 | #endif |
---|
363 | #ifdef CLP_DEBUG |
---|
364 | { |
---|
365 | int iSequence, number=numberRows_+numberColumns_; |
---|
366 | for (iSequence=0;iSequence<number;iSequence++) { |
---|
367 | double lowerValue=lower_[iSequence]; |
---|
368 | double upperValue=upper_[iSequence]; |
---|
369 | double value=solution_[iSequence]; |
---|
370 | if(getStatus(iSequence)!=ClpSimplex::basic) { |
---|
371 | assert(lowerValue>-1.0e20); |
---|
372 | assert(upperValue<1.0e20); |
---|
373 | } |
---|
374 | switch(getStatus(iSequence)) { |
---|
375 | |
---|
376 | case ClpSimplex::basic: |
---|
377 | break; |
---|
378 | case ClpSimplex::isFree: |
---|
379 | case ClpSimplex::superBasic: |
---|
380 | break; |
---|
381 | case ClpSimplex::atUpperBound: |
---|
382 | assert (fabs(value-upperValue)<=primalTolerance_) ; |
---|
383 | break; |
---|
384 | case ClpSimplex::atLowerBound: |
---|
385 | assert (fabs(value-lowerValue)<=primalTolerance_) ; |
---|
386 | break; |
---|
387 | } |
---|
388 | } |
---|
389 | } |
---|
390 | if(numberIterations_==debugIteration) { |
---|
391 | printf("dodgy iteration coming up\n"); |
---|
392 | } |
---|
393 | #endif |
---|
394 | // choose row to go out |
---|
395 | dualRow(); |
---|
396 | if (pivotRow_>=0) { |
---|
397 | // we found a pivot row |
---|
398 | handler_->message(CLP_SIMPLEX_PIVOTROW,messages_) |
---|
399 | <<pivotRow_ |
---|
400 | <<OsiMessageEol; |
---|
401 | // check accuracy of weights |
---|
402 | dualRowPivot_->checkAccuracy(); |
---|
403 | // get sign for finding row of tableau |
---|
404 | rowArray_[0]->insert(pivotRow_,directionOut_); |
---|
405 | factorization_->updateColumnTranspose(rowArray_[1],rowArray_[0]); |
---|
406 | // put row of tableau in rowArray[0] and columnArray[0] |
---|
407 | matrix_->transposeTimes(this,-1.0, |
---|
408 | rowArray_[0],columnArray_[1],columnArray_[0]); |
---|
409 | // rowArray has pi equivalent |
---|
410 | // do ratio test |
---|
411 | dualColumn(rowArray_[0],columnArray_[0],columnArray_[1], |
---|
412 | rowArray_[3]); |
---|
413 | if (sequenceIn_>=0) { |
---|
414 | // normal iteration |
---|
415 | // update the incoming column |
---|
416 | unpack(rowArray_[1]); |
---|
417 | factorization_->updateColumn(rowArray_[2],rowArray_[1],true); |
---|
418 | // and update dual weights (can do in parallel - with extra array) |
---|
419 | dualRowPivot_->updateWeights(rowArray_[0],rowArray_[2], |
---|
420 | rowArray_[1]); |
---|
421 | // see if update stable |
---|
422 | double btranAlpha = -alpha_*directionOut_; // for check |
---|
423 | alpha_=(*rowArray_[1])[pivotRow_]; |
---|
424 | #ifdef CLP_DEBUG |
---|
425 | if ((handler_->logLevel()&32)) |
---|
426 | printf("btran alpha %g, ftran alpha %g\n",btranAlpha,alpha_); |
---|
427 | #endif |
---|
428 | if (fabs(btranAlpha)<1.0e-12||fabs(alpha_)<1.0e-12|| |
---|
429 | fabs(btranAlpha-alpha_)>1.0e-7*(1.0+fabs(alpha_))) { |
---|
430 | handler_->message(CLP_DUAL_CHECK,messages_) |
---|
431 | <<btranAlpha |
---|
432 | <<alpha_ |
---|
433 | <<OsiMessageEol; |
---|
434 | dualRowPivot_->unrollWeights(); |
---|
435 | if (factorization_->pivots()) { |
---|
436 | problemStatus_=-2; // factorize now |
---|
437 | rowArray_[0]->clear(); |
---|
438 | rowArray_[1]->clear(); |
---|
439 | columnArray_[0]->clear(); |
---|
440 | break; |
---|
441 | } else { |
---|
442 | // take on more relaxed criterion |
---|
443 | if (fabs(btranAlpha)<1.0e-12||fabs(alpha_)<1.0e-12|| |
---|
444 | fabs(btranAlpha-alpha_)>1.0e-4*(1.0+fabs(alpha_))) { |
---|
445 | // need to reject something |
---|
446 | char x = isColumn(sequenceOut_) ? 'C' :'R'; |
---|
447 | handler_->message(CLP_SIMPLEX_FLAG,messages_) |
---|
448 | <<x<<sequenceWithin(sequenceOut_) |
---|
449 | <<OsiMessageEol; |
---|
450 | setFlagged(sequenceOut_); |
---|
451 | rowArray_[0]->clear(); |
---|
452 | rowArray_[1]->clear(); |
---|
453 | columnArray_[0]->clear(); |
---|
454 | continue; |
---|
455 | } |
---|
456 | } |
---|
457 | } |
---|
458 | // update duals BEFORE replaceColumn so can do updateColumn |
---|
459 | objectiveChange=0.0; |
---|
460 | // do duals first as variables may flip bounds |
---|
461 | // rowArray_[0] and columnArray_[0] may have flips |
---|
462 | // so use rowArray_[3] for work array from here on |
---|
463 | int nswapped = |
---|
464 | updateDualsInDual(rowArray_[0],columnArray_[0],rowArray_[2],theta_, |
---|
465 | objectiveChange); |
---|
466 | // which will change basic solution |
---|
467 | if (nswapped) { |
---|
468 | #ifdef CLP_DEBUG |
---|
469 | if ((handler_->logLevel()&16)) |
---|
470 | printf("old dualOut_ %g, v %g, l %g, u %g - new ", |
---|
471 | dualOut_,valueOut_,lowerOut_,upperOut_); |
---|
472 | double oldOut=dualOut_; |
---|
473 | #endif |
---|
474 | factorization_->updateColumn(rowArray_[3],rowArray_[2],false); |
---|
475 | dualRowPivot_->updatePrimalSolution(rowArray_[2], |
---|
476 | 1.0,objectiveChange); |
---|
477 | |
---|
478 | // recompute dualOut_ |
---|
479 | valueOut_ = solution_[sequenceOut_]; |
---|
480 | if (directionOut_<0) { |
---|
481 | dualOut_ = valueOut_ - upperOut_; |
---|
482 | } else { |
---|
483 | dualOut_ = lowerOut_ - valueOut_; |
---|
484 | } |
---|
485 | #ifdef CLP_DEBUG |
---|
486 | if ((handler_->logLevel()&16)) |
---|
487 | printf("%g\n",dualOut_); |
---|
488 | assert(dualOut_<=oldOut); |
---|
489 | #endif |
---|
490 | if(dualOut_<0.0&&factorization_->pivots()) { |
---|
491 | // going backwards - factorize |
---|
492 | dualRowPivot_->unrollWeights(); |
---|
493 | problemStatus_=-2; // factorize now |
---|
494 | break; |
---|
495 | } |
---|
496 | } |
---|
497 | // amount primal will move |
---|
498 | double movement = -dualOut_*directionOut_/alpha_; |
---|
499 | // so objective should increase by fabs(dj)*movement |
---|
500 | // but we already have objective change - so check will be good |
---|
501 | if (objectiveChange+fabs(movement*dualIn_)<-1.0e-5) { |
---|
502 | #ifdef CLP_DEBUG |
---|
503 | if (handler_->logLevel()&32) |
---|
504 | printf("movement %g, swap change %g, rest %g * %g\n", |
---|
505 | objectiveChange+fabs(movement*dualIn_), |
---|
506 | objectiveChange,movement,dualIn_); |
---|
507 | #endif |
---|
508 | if(factorization_->pivots()>5) { |
---|
509 | // going backwards - factorize |
---|
510 | dualRowPivot_->unrollWeights(); |
---|
511 | problemStatus_=-2; // factorize now |
---|
512 | break; |
---|
513 | } |
---|
514 | } |
---|
515 | // if stable replace in basis |
---|
516 | int updateStatus = factorization_->replaceColumn(rowArray_[2], |
---|
517 | pivotRow_, |
---|
518 | alpha_); |
---|
519 | if (updateStatus==1) { |
---|
520 | // slight error |
---|
521 | if (factorization_->pivots()>5) |
---|
522 | problemStatus_=-2; // factorize now |
---|
523 | } else if (updateStatus==2) { |
---|
524 | // major error |
---|
525 | dualRowPivot_->unrollWeights(); |
---|
526 | // later we may need to unwind more e.g. fake bounds |
---|
527 | if (factorization_->pivots()) { |
---|
528 | problemStatus_=-2; // factorize now |
---|
529 | break; |
---|
530 | } else { |
---|
531 | // need to reject something |
---|
532 | char x = isColumn(sequenceOut_) ? 'C' :'R'; |
---|
533 | handler_->message(CLP_SIMPLEX_FLAG,messages_) |
---|
534 | <<x<<sequenceWithin(sequenceOut_) |
---|
535 | <<OsiMessageEol; |
---|
536 | setFlagged(sequenceOut_); |
---|
537 | rowArray_[0]->clear(); |
---|
538 | rowArray_[1]->clear(); |
---|
539 | columnArray_[0]->clear(); |
---|
540 | continue; |
---|
541 | } |
---|
542 | } else if (updateStatus==3) { |
---|
543 | // out of memory |
---|
544 | // increase space if not many iterations |
---|
545 | if (factorization_->pivots()< |
---|
546 | 0.5*factorization_->maximumPivots()&& |
---|
547 | factorization_->pivots()<200) |
---|
548 | factorization_->areaFactor( |
---|
549 | factorization_->areaFactor() * 1.1); |
---|
550 | problemStatus_=-2; // factorize now |
---|
551 | } |
---|
552 | // update primal solution |
---|
553 | if (theta_<0.0) { |
---|
554 | #ifdef CLP_DEBUG |
---|
555 | if (handler_->logLevel()&32) |
---|
556 | printf("negative theta %g\n",theta_); |
---|
557 | #endif |
---|
558 | theta_=0.0; |
---|
559 | } |
---|
560 | // do actual flips |
---|
561 | flipBounds(rowArray_[0],columnArray_[0],theta_); |
---|
562 | dualRowPivot_->updatePrimalSolution(rowArray_[1], |
---|
563 | movement, |
---|
564 | objectiveChange); |
---|
565 | #ifdef CLP_DEBUG |
---|
566 | double oldobj=objectiveValue_; |
---|
567 | #endif |
---|
568 | int whatNext=housekeeping(objectiveChange); |
---|
569 | // and set bounds correctly |
---|
570 | originalBound(sequenceIn_); |
---|
571 | changeBound(sequenceOut_); |
---|
572 | #ifdef CLP_DEBUG |
---|
573 | if (objectiveValue_<oldobj-1.0e-5&&(handler_->logLevel()&16)) |
---|
574 | printf("obj backwards %g %g\n",objectiveValue_,oldobj); |
---|
575 | #endif |
---|
576 | if (whatNext==1) { |
---|
577 | problemStatus_ =-2; // refactorize |
---|
578 | } else if (whatNext==2) { |
---|
579 | // maximum iterations or equivalent |
---|
580 | problemStatus_= 3; |
---|
581 | break; |
---|
582 | } |
---|
583 | } else { |
---|
584 | // no incoming column is valid |
---|
585 | #ifdef CLP_DEBUG |
---|
586 | if (handler_->logLevel()&32) |
---|
587 | printf("** no column pivot\n"); |
---|
588 | #endif |
---|
589 | if (factorization_->pivots()<5) { |
---|
590 | problemStatus_=-4; //say looks infeasible |
---|
591 | // create ray anyway |
---|
592 | delete [] ray_; |
---|
593 | ray_ = new double [ numberRows_]; |
---|
594 | CoinDisjointCopyN(rowArray_[0]->denseVector(),numberRows_,ray_); |
---|
595 | } |
---|
596 | rowArray_[0]->clear(); |
---|
597 | columnArray_[0]->clear(); |
---|
598 | break; |
---|
599 | } |
---|
600 | } else { |
---|
601 | // no pivot row |
---|
602 | #ifdef CLP_DEBUG |
---|
603 | if (handler_->logLevel()&32) |
---|
604 | printf("** no row pivot\n"); |
---|
605 | #endif |
---|
606 | if (!factorization_->pivots()) { |
---|
607 | // may have crept through - so may be optimal |
---|
608 | //problemStatus_=-5; //say looks unbounded |
---|
609 | problemStatus_=0; |
---|
610 | // check any flagged variables |
---|
611 | int iRow; |
---|
612 | for (iRow=0;iRow<numberRows_;iRow++) { |
---|
613 | int iPivot=pivotVariable_[iRow]; |
---|
614 | if (flagged(iPivot)) |
---|
615 | break; |
---|
616 | } |
---|
617 | if (iRow<numberRows_) { |
---|
618 | #ifdef CLP_DEBUG |
---|
619 | std::cerr<<"Flagged variables at end - infeasible?"<<std::endl; |
---|
620 | #endif |
---|
621 | problemStatus_=-4; //say looks infeasible |
---|
622 | // create ray anyway |
---|
623 | delete [] ray_; |
---|
624 | ray_ = new double [ numberRows_]; |
---|
625 | CoinDisjointCopyN(rowArray_[0]->denseVector(),numberRows_,ray_); |
---|
626 | } |
---|
627 | } |
---|
628 | break; |
---|
629 | } |
---|
630 | } |
---|
631 | } |
---|
632 | |
---|
633 | // at present we are leaving factorization around |
---|
634 | // maybe we should empty it |
---|
635 | deleteRim(); |
---|
636 | handler_->message(CLP_SIMPLEX_FINISHED+problemStatus_,messages_) |
---|
637 | <<objectiveValue() |
---|
638 | <<OsiMessageEol; |
---|
639 | // Restore any saved stuff |
---|
640 | perturbation_ = savePerturbation; |
---|
641 | dualBound_ = saveDualBound_; |
---|
642 | return problemStatus_; |
---|
643 | } |
---|
644 | /* The duals are updated by the given arrays. |
---|
645 | Returns number of infeasibilities. |
---|
646 | rowArray and columnarray will have flipped |
---|
647 | The output vector has movement (row length array) */ |
---|
648 | int |
---|
649 | ClpSimplexDual::updateDualsInDual(OsiIndexedVector * rowArray, |
---|
650 | OsiIndexedVector * columnArray, |
---|
651 | OsiIndexedVector * outputArray, |
---|
652 | double theta, |
---|
653 | double & objectiveChange) |
---|
654 | { |
---|
655 | |
---|
656 | outputArray->clear(); |
---|
657 | |
---|
658 | double * work; |
---|
659 | int number; |
---|
660 | int * which; |
---|
661 | |
---|
662 | int numberInfeasibilities=0; |
---|
663 | int numberRowInfeasibilities=0; |
---|
664 | |
---|
665 | // see whether we will be doing full recompute |
---|
666 | bool fullRecompute= (rowArray->getNumElements()==numberRows_&& |
---|
667 | columnArray->getNumElements()==numberColumns_); |
---|
668 | int numberAtFake=0; |
---|
669 | |
---|
670 | // use a tighter tolerance except for all being okay |
---|
671 | double tolerance = dualTolerance_; |
---|
672 | |
---|
673 | double changeObj=0.0; |
---|
674 | |
---|
675 | int iSection; |
---|
676 | |
---|
677 | for (iSection=0;iSection<2;iSection++) { |
---|
678 | int i; |
---|
679 | double * solution = solutionRegion(iSection); |
---|
680 | double * reducedCost = djRegion(iSection); |
---|
681 | double * lower = lowerRegion(iSection); |
---|
682 | double * upper = upperRegion(iSection); |
---|
683 | double * cost = costRegion(iSection); |
---|
684 | int addSequence; |
---|
685 | if (!iSection) { |
---|
686 | addSequence = numberColumns_; |
---|
687 | work = rowArray->denseVector(); |
---|
688 | number = rowArray->getNumElements(); |
---|
689 | which = rowArray->getIndices(); |
---|
690 | } else { |
---|
691 | // set number of infeasibilities in row array |
---|
692 | addSequence=0; |
---|
693 | numberRowInfeasibilities=numberInfeasibilities; |
---|
694 | rowArray->setNumElements(numberInfeasibilities); |
---|
695 | numberInfeasibilities=0; |
---|
696 | work = columnArray->denseVector(); |
---|
697 | number = columnArray->getNumElements(); |
---|
698 | which = columnArray->getIndices(); |
---|
699 | } |
---|
700 | |
---|
701 | for (i=0;i<number;i++) { |
---|
702 | int iSequence = which[i]; |
---|
703 | double alphaI = work[iSequence]; |
---|
704 | double value = reducedCost[iSequence]-theta*alphaI; |
---|
705 | work[iSequence]=0.0; |
---|
706 | reducedCost[iSequence]=value; |
---|
707 | |
---|
708 | if (!fixed(iSequence+addSequence)) { |
---|
709 | double movement=0.0; |
---|
710 | FakeBound bound = getFakeBound(iSequence+addSequence); |
---|
711 | Status status = getStatus(iSequence+addSequence); |
---|
712 | |
---|
713 | switch(status) { |
---|
714 | |
---|
715 | case ClpSimplex::basic: |
---|
716 | case ClpSimplex::superBasic: |
---|
717 | break; |
---|
718 | case ClpSimplex::isFree: |
---|
719 | if (fabs(value)>tolerance) { |
---|
720 | #ifdef CLP_DEBUG |
---|
721 | if (handler_->logLevel()&32) |
---|
722 | printf("%d %d, free has dj of %g, alpha %g\n", |
---|
723 | iSection,iSequence,value,alphaI); |
---|
724 | #endif |
---|
725 | } |
---|
726 | break; |
---|
727 | case ClpSimplex::atUpperBound: |
---|
728 | if (value>tolerance) { |
---|
729 | // to lower bound (if swap) |
---|
730 | // put back alpha |
---|
731 | which[numberInfeasibilities++]=iSequence; |
---|
732 | work[iSequence]=alphaI; |
---|
733 | movement = lower[iSequence]-upper[iSequence]; |
---|
734 | #ifdef CLP_DEBUG |
---|
735 | if ((handler_->logLevel()&32)) |
---|
736 | printf("%d %d, new dj %g, alpha %g, movement %g\n", |
---|
737 | iSection,iSequence,value,alphaI,movement); |
---|
738 | #endif |
---|
739 | changeObj += movement*cost[iSequence]; |
---|
740 | if (bound==ClpSimplexDual::bothFake|| |
---|
741 | bound==ClpSimplexDual::lowerFake) |
---|
742 | numberAtFake++; |
---|
743 | } else if (fullRecompute) { |
---|
744 | // at correct bound |
---|
745 | if (bound==ClpSimplexDual::bothFake|| |
---|
746 | bound==ClpSimplexDual::upperFake) { |
---|
747 | // but flip if dj would allow |
---|
748 | if (bound==ClpSimplexDual::upperFake&& |
---|
749 | value>=-tolerance) { |
---|
750 | movement = lower[iSequence]-upper[iSequence]; |
---|
751 | setStatus(iSequence+addSequence,ClpSimplex::atLowerBound); |
---|
752 | solution[iSequence] = lower[iSequence]; |
---|
753 | changeObj += movement*cost[iSequence]; |
---|
754 | } else { |
---|
755 | numberAtFake++; |
---|
756 | } |
---|
757 | } |
---|
758 | } |
---|
759 | break; |
---|
760 | case ClpSimplex::atLowerBound: |
---|
761 | if (value<-tolerance) { |
---|
762 | // to upper bound |
---|
763 | // put back alpha |
---|
764 | which[numberInfeasibilities++]=iSequence; |
---|
765 | work[iSequence]=alphaI; |
---|
766 | movement = upper[iSequence] - lower[iSequence]; |
---|
767 | #ifdef CLP_DEBUG |
---|
768 | if ((handler_->logLevel()&32)) |
---|
769 | printf("%d %d, new dj %g, alpha %g, movement %g\n", |
---|
770 | iSection,iSequence,value,alphaI,movement); |
---|
771 | #endif |
---|
772 | changeObj += movement*cost[iSequence]; |
---|
773 | if (bound==ClpSimplexDual::bothFake|| |
---|
774 | bound==ClpSimplexDual::upperFake) |
---|
775 | numberAtFake++; |
---|
776 | } else if (fullRecompute) { |
---|
777 | // at correct bound |
---|
778 | if (bound==ClpSimplexDual::bothFake|| |
---|
779 | bound==ClpSimplexDual::lowerFake) { |
---|
780 | // but flip if dj would allow |
---|
781 | if (bound==ClpSimplexDual::lowerFake&& |
---|
782 | value<=tolerance) { |
---|
783 | movement = upper[iSequence] - lower[iSequence]; |
---|
784 | setStatus(iSequence+addSequence,ClpSimplex::atUpperBound); |
---|
785 | solution[iSequence] = upper[iSequence]; |
---|
786 | changeObj += movement*cost[iSequence]; |
---|
787 | } else { |
---|
788 | numberAtFake++; |
---|
789 | } |
---|
790 | } |
---|
791 | } |
---|
792 | break; |
---|
793 | } |
---|
794 | if (!fullRecompute) { |
---|
795 | if (movement) { |
---|
796 | if (!iSection) { |
---|
797 | // row (sign ?) |
---|
798 | outputArray->quickAdd(iSequence,-movement); |
---|
799 | } else { |
---|
800 | matrix_->add(this,outputArray,iSequence,movement); |
---|
801 | } |
---|
802 | } |
---|
803 | } |
---|
804 | } |
---|
805 | } |
---|
806 | } |
---|
807 | #ifdef CLP_DEBUG |
---|
808 | if (fullRecompute&&numberAtFake&&(handler_->logLevel()&16)!=0) |
---|
809 | printf("%d fake after full update\n",numberAtFake); |
---|
810 | #endif |
---|
811 | outputArray->stopQuickAdd(); |
---|
812 | // set number of infeasibilities |
---|
813 | columnArray->setNumElements(numberInfeasibilities); |
---|
814 | numberInfeasibilities += numberRowInfeasibilities; |
---|
815 | if (fullRecompute) { |
---|
816 | // do actual flips |
---|
817 | flipBounds(rowArray,columnArray,theta); |
---|
818 | } |
---|
819 | objectiveChange += changeObj; |
---|
820 | return numberInfeasibilities; |
---|
821 | } |
---|
822 | /* |
---|
823 | Chooses dual pivot row |
---|
824 | Would be faster with separate region to scan |
---|
825 | and will have this (with square of infeasibility) when steepest |
---|
826 | For easy problems we can just choose one of the first rows we look at |
---|
827 | */ |
---|
828 | void |
---|
829 | ClpSimplexDual::dualRow() |
---|
830 | { |
---|
831 | // get pivot row using whichever method it is |
---|
832 | pivotRow_=dualRowPivot_->pivotRow(); |
---|
833 | if (pivotRow_>=0) { |
---|
834 | int iPivot=pivotVariable_[pivotRow_]; |
---|
835 | sequenceOut_ = iPivot; |
---|
836 | if (iPivot>=numberColumns_) { |
---|
837 | // slack |
---|
838 | iPivot-=numberColumns_; |
---|
839 | valueOut_=rowActivityWork_[iPivot]; |
---|
840 | lowerOut_=rowLowerWork_[iPivot]; |
---|
841 | upperOut_=rowUpperWork_[iPivot]; |
---|
842 | } else { |
---|
843 | // column |
---|
844 | valueOut_=columnActivityWork_[iPivot]; |
---|
845 | lowerOut_=columnLowerWork_[iPivot]; |
---|
846 | upperOut_=columnUpperWork_[iPivot]; |
---|
847 | } |
---|
848 | // if we have problems we could try other way and hope we get a |
---|
849 | // zero pivot? |
---|
850 | if (valueOut_>upperOut_) { |
---|
851 | directionOut_ = -1; |
---|
852 | dualOut_ = valueOut_ - upperOut_; |
---|
853 | } else { |
---|
854 | directionOut_ = 1; |
---|
855 | dualOut_ = lowerOut_ - valueOut_; |
---|
856 | } |
---|
857 | #ifdef CLP_DEBUG |
---|
858 | assert(dualOut_>=0.0); |
---|
859 | #endif |
---|
860 | } |
---|
861 | return ; |
---|
862 | } |
---|
863 | // Checks if any fake bounds active - if so returns number and modifies |
---|
864 | // dualBound_ and everything. |
---|
865 | // Free variables will be left as free |
---|
866 | // Returns number of bounds changed if >=0 |
---|
867 | // Returns -1 if not initialize and no effect |
---|
868 | // Fills in changeVector which can be used to see if unbounded |
---|
869 | // and cost of change vector |
---|
870 | int |
---|
871 | ClpSimplexDual::changeBounds(bool initialize, |
---|
872 | OsiIndexedVector * outputArray, |
---|
873 | double & changeCost) |
---|
874 | { |
---|
875 | if (!initialize) { |
---|
876 | int numberInfeasibilities; |
---|
877 | double newBound; |
---|
878 | newBound = 5.0*dualBound_; |
---|
879 | numberInfeasibilities=0; |
---|
880 | changeCost=0.0; |
---|
881 | // put back original bounds and then check |
---|
882 | createRim(3); |
---|
883 | int iSequence; |
---|
884 | // bounds will get bigger - just look at ones at bounds |
---|
885 | for (iSequence=0;iSequence<numberRows_+numberColumns_;iSequence++) { |
---|
886 | double lowerValue=lower_[iSequence]; |
---|
887 | double upperValue=upper_[iSequence]; |
---|
888 | double value=solution_[iSequence]; |
---|
889 | setFakeBound(iSequence,ClpSimplexDual::noFake); |
---|
890 | switch(getStatus(iSequence)) { |
---|
891 | |
---|
892 | case ClpSimplex::basic: |
---|
893 | break; |
---|
894 | case ClpSimplex::isFree: |
---|
895 | case ClpSimplex::superBasic: |
---|
896 | break; |
---|
897 | case ClpSimplex::atUpperBound: |
---|
898 | if (fabs(value-upperValue)>primalTolerance_) |
---|
899 | numberInfeasibilities++; |
---|
900 | break; |
---|
901 | case ClpSimplex::atLowerBound: |
---|
902 | if (fabs(value-lowerValue)>primalTolerance_) |
---|
903 | numberInfeasibilities++; |
---|
904 | break; |
---|
905 | } |
---|
906 | } |
---|
907 | if (numberInfeasibilities) { |
---|
908 | int iSequence; |
---|
909 | for (iSequence=0;iSequence<numberRows_+numberColumns_;iSequence++) { |
---|
910 | double lowerValue=lower_[iSequence]; |
---|
911 | double upperValue=upper_[iSequence]; |
---|
912 | double newLowerValue; |
---|
913 | double newUpperValue; |
---|
914 | Status status = getStatus(iSequence); |
---|
915 | if (status==ClpSimplex::atUpperBound|| |
---|
916 | status==ClpSimplex::atLowerBound) { |
---|
917 | double value = solution_[iSequence]; |
---|
918 | if (value-lowerValue<=upperValue-value) { |
---|
919 | newLowerValue = max(lowerValue,value-0.666667*newBound); |
---|
920 | newUpperValue = min(upperValue,newLowerValue+newBound); |
---|
921 | } else { |
---|
922 | newUpperValue = min(upperValue,value+0.666667*newBound); |
---|
923 | newLowerValue = max(lowerValue,newUpperValue-newBound); |
---|
924 | } |
---|
925 | lower_[iSequence]=newLowerValue; |
---|
926 | upper_[iSequence]=newUpperValue; |
---|
927 | if (newLowerValue > lowerValue) { |
---|
928 | if (newUpperValue < upperValue) |
---|
929 | setFakeBound(iSequence,ClpSimplexDual::bothFake); |
---|
930 | else |
---|
931 | setFakeBound(iSequence,ClpSimplexDual::lowerFake); |
---|
932 | } else { |
---|
933 | if (newUpperValue < upperValue) |
---|
934 | setFakeBound(iSequence,ClpSimplexDual::upperFake); |
---|
935 | } |
---|
936 | if (status==ClpSimplex::atUpperBound) |
---|
937 | solution_[iSequence] = newUpperValue; |
---|
938 | else |
---|
939 | solution_[iSequence] = newLowerValue; |
---|
940 | double movement = solution_[iSequence] - value; |
---|
941 | if (movement&&outputArray) { |
---|
942 | if (iSequence>=numberColumns_) { |
---|
943 | outputArray->quickAdd(iSequence,-movement); |
---|
944 | changeCost += movement*cost_[iSequence]; |
---|
945 | } else { |
---|
946 | matrix_->add(this,outputArray,iSequence,movement); |
---|
947 | changeCost += movement*cost_[iSequence]; |
---|
948 | } |
---|
949 | } |
---|
950 | } |
---|
951 | } |
---|
952 | dualBound_ = newBound; |
---|
953 | if (outputArray) |
---|
954 | outputArray->stopQuickAdd(); |
---|
955 | } else { |
---|
956 | numberInfeasibilities=-1; |
---|
957 | } |
---|
958 | return numberInfeasibilities; |
---|
959 | } else { |
---|
960 | int iSequence; |
---|
961 | |
---|
962 | for (iSequence=0;iSequence<numberRows_+numberColumns_;iSequence++) { |
---|
963 | Status status = getStatus(iSequence); |
---|
964 | if (status==ClpSimplex::atUpperBound|| |
---|
965 | status==ClpSimplex::atLowerBound) { |
---|
966 | double lowerValue=lower_[iSequence]; |
---|
967 | double upperValue=upper_[iSequence]; |
---|
968 | double value = solution_[iSequence]; |
---|
969 | if (lowerValue>-largeValue_||upperValue<largeValue_) { |
---|
970 | if (lowerValue-value>-0.5*dualBound_|| |
---|
971 | upperValue-value<0.5*dualBound_) { |
---|
972 | if (fabs(lowerValue-value)<=fabs(upperValue-value)) { |
---|
973 | if (upperValue > lowerValue + dualBound_) { |
---|
974 | upper_[iSequence]=lowerValue+dualBound_; |
---|
975 | setFakeBound(iSequence,ClpSimplexDual::upperFake); |
---|
976 | } |
---|
977 | } else { |
---|
978 | if (lowerValue < upperValue - dualBound_) { |
---|
979 | lower_[iSequence]=upperValue-dualBound_; |
---|
980 | setFakeBound(iSequence,ClpSimplexDual::lowerFake); |
---|
981 | } |
---|
982 | } |
---|
983 | } else { |
---|
984 | lower_[iSequence]=-0.5*dualBound_; |
---|
985 | upper_[iSequence]= 0.5*dualBound_; |
---|
986 | setFakeBound(iSequence,ClpSimplexDual::bothFake); |
---|
987 | } |
---|
988 | } |
---|
989 | } |
---|
990 | } |
---|
991 | return 1; |
---|
992 | } |
---|
993 | } |
---|
994 | /* |
---|
995 | Row array has row part of pivot row (as duals so sign may be switched) |
---|
996 | Column array has column part. |
---|
997 | This chooses pivot column. |
---|
998 | Spare array will be needed when we start getting clever. |
---|
999 | We will check for basic so spare array will never overflow. |
---|
1000 | If necessary will modify costs |
---|
1001 | */ |
---|
1002 | void |
---|
1003 | ClpSimplexDual::dualColumn(OsiIndexedVector * rowArray, |
---|
1004 | OsiIndexedVector * columnArray, |
---|
1005 | OsiIndexedVector * spareArray, |
---|
1006 | OsiIndexedVector * spareArray2) |
---|
1007 | { |
---|
1008 | double * work; |
---|
1009 | int number; |
---|
1010 | int * which; |
---|
1011 | double * reducedCost; |
---|
1012 | |
---|
1013 | int iSection; |
---|
1014 | |
---|
1015 | sequenceIn_=-1; |
---|
1016 | int numberPossiblySwapped=0; |
---|
1017 | int numberRemaining=0; |
---|
1018 | |
---|
1019 | double totalThru=0.0; // for when variables flip |
---|
1020 | double acceptablePivot=1.0e-7; |
---|
1021 | if (factorization_->pivots()) |
---|
1022 | acceptablePivot=1.0e-5; // if we have iterated be more strict |
---|
1023 | double bestEverPivot=acceptablePivot; |
---|
1024 | int lastSequence = -1; |
---|
1025 | double lastPivot=0.0; |
---|
1026 | double upperTheta; |
---|
1027 | double newTolerance = dualTolerance_; |
---|
1028 | // will we need to increase tolerance |
---|
1029 | bool thisIncrease=false; |
---|
1030 | // If we think we need to modify costs (not if something from broad sweep) |
---|
1031 | bool modifyCosts=false; |
---|
1032 | // Increase in objective due to swapping bounds (may be negative) |
---|
1033 | double increaseInObjective=0.0; |
---|
1034 | |
---|
1035 | // use spareArrays to put ones looked at in |
---|
1036 | // we are going to flip flop between |
---|
1037 | int iFlip = 0; |
---|
1038 | // Possible list of pivots |
---|
1039 | int interesting[2]; |
---|
1040 | // where possible swapped ones are |
---|
1041 | int swapped[2]; |
---|
1042 | // for zeroing out arrays after |
---|
1043 | int marker[2][2]; |
---|
1044 | // pivot elements |
---|
1045 | double * array[2], * spare, * spare2; |
---|
1046 | // indices |
---|
1047 | int * indices[2], * index, * index2; |
---|
1048 | spareArray->clear(); |
---|
1049 | spareArray2->clear(); |
---|
1050 | array[0] = spareArray->denseVector(); |
---|
1051 | indices[0] = spareArray->getIndices(); |
---|
1052 | spare = array[0]; |
---|
1053 | index = indices[0]; |
---|
1054 | array[1] = spareArray2->denseVector(); |
---|
1055 | indices[1] = spareArray2->getIndices(); |
---|
1056 | int i; |
---|
1057 | double * lower; |
---|
1058 | double * upper; |
---|
1059 | |
---|
1060 | // initialize lists |
---|
1061 | for (i=0;i<2;i++) { |
---|
1062 | interesting[i]=0; |
---|
1063 | swapped[i]=numberColumns_; |
---|
1064 | marker[i][0]=0; |
---|
1065 | marker[i][1]=numberColumns_; |
---|
1066 | } |
---|
1067 | |
---|
1068 | /* |
---|
1069 | First we get a list of possible pivots. We can also see if the |
---|
1070 | problem looks infeasible or whether we want to pivot in free variable. |
---|
1071 | This may make objective go backwards but can only happen a finite |
---|
1072 | number of times and I do want free variables basic. |
---|
1073 | |
---|
1074 | Then we flip back and forth. At the start of each iteration |
---|
1075 | interesting[iFlip] should have possible candidates and swapped[iFlip] |
---|
1076 | will have pivots if we decide to take a previous pivot. |
---|
1077 | At end of each iteration interesting[1-iFlip] should have |
---|
1078 | candidates if we go through this theta and swapped[1-iFlip] |
---|
1079 | pivots if we don't go through. |
---|
1080 | |
---|
1081 | At first we increase theta and see what happens. We start |
---|
1082 | theta at a reasonable guess. If in right area then we do bit by bit. |
---|
1083 | |
---|
1084 | */ |
---|
1085 | |
---|
1086 | // do first pass to get possibles |
---|
1087 | // We can also see if infeasible or pivoting on free |
---|
1088 | double tentativeTheta = 1.0e22; |
---|
1089 | upperTheta = 1.0e31; |
---|
1090 | double freePivot = acceptablePivot; |
---|
1091 | for (iSection=0;iSection<2;iSection++) { |
---|
1092 | |
---|
1093 | int addSequence; |
---|
1094 | |
---|
1095 | if (!iSection) { |
---|
1096 | lower = rowLowerWork_; |
---|
1097 | upper = rowUpperWork_; |
---|
1098 | work = rowArray->denseVector(); |
---|
1099 | number = rowArray->getNumElements(); |
---|
1100 | which = rowArray->getIndices(); |
---|
1101 | reducedCost = rowReducedCost_; |
---|
1102 | addSequence = numberColumns_; |
---|
1103 | } else { |
---|
1104 | lower = columnLowerWork_; |
---|
1105 | upper = columnUpperWork_; |
---|
1106 | work = columnArray->denseVector(); |
---|
1107 | number = columnArray->getNumElements(); |
---|
1108 | which = columnArray->getIndices(); |
---|
1109 | reducedCost = reducedCostWork_; |
---|
1110 | addSequence = 0; |
---|
1111 | } |
---|
1112 | |
---|
1113 | for (i=0;i<number;i++) { |
---|
1114 | int iSequence = which[i]; |
---|
1115 | double alpha = work[iSequence]; |
---|
1116 | if (fixed(iSequence+addSequence)||!alpha) |
---|
1117 | continue; // skip fixed ones or (zeroed out) |
---|
1118 | double oldValue = reducedCost[iSequence]; |
---|
1119 | double value = oldValue-tentativeTheta*alpha; |
---|
1120 | int keep = 0; |
---|
1121 | |
---|
1122 | switch(getStatus(iSequence+addSequence)) { |
---|
1123 | |
---|
1124 | case ClpSimplex::basic: |
---|
1125 | break; |
---|
1126 | case ClpSimplex::isFree: |
---|
1127 | case ClpSimplex::superBasic: |
---|
1128 | if (oldValue>dualTolerance_) { |
---|
1129 | if (value<-newTolerance) |
---|
1130 | keep = 2; |
---|
1131 | } else if (oldValue<-dualTolerance_) { |
---|
1132 | if (value>newTolerance) |
---|
1133 | keep = 2; |
---|
1134 | } else { |
---|
1135 | if (alpha>=acceptablePivot) |
---|
1136 | keep = 2; |
---|
1137 | else if (-alpha>=acceptablePivot) |
---|
1138 | keep = 2; |
---|
1139 | } |
---|
1140 | break; |
---|
1141 | case ClpSimplex::atUpperBound: |
---|
1142 | assert (oldValue<=dualTolerance_*1.0001); |
---|
1143 | if (value>newTolerance) { |
---|
1144 | keep = 1; |
---|
1145 | value = oldValue-upperTheta*alpha; |
---|
1146 | if (value>newTolerance && -alpha>=acceptablePivot) |
---|
1147 | upperTheta = (oldValue-newTolerance)/alpha; |
---|
1148 | } |
---|
1149 | break; |
---|
1150 | case ClpSimplex::atLowerBound: |
---|
1151 | assert (oldValue>=-dualTolerance_*1.0001); |
---|
1152 | if (value<-newTolerance) { |
---|
1153 | keep = 1; |
---|
1154 | value = oldValue-upperTheta*alpha; |
---|
1155 | if (value<-newTolerance && alpha>=acceptablePivot) |
---|
1156 | upperTheta = (oldValue+newTolerance)/alpha; |
---|
1157 | } |
---|
1158 | break; |
---|
1159 | } |
---|
1160 | if (keep) { |
---|
1161 | if (keep==2) { |
---|
1162 | // free - choose largest |
---|
1163 | if (fabs(alpha)>freePivot) { |
---|
1164 | freePivot=fabs(alpha); |
---|
1165 | sequenceIn_ = iSequence + addSequence; |
---|
1166 | theta_=oldValue/alpha; |
---|
1167 | } |
---|
1168 | } else { |
---|
1169 | // add to list |
---|
1170 | spare[numberRemaining]=alpha; |
---|
1171 | index[numberRemaining++]=iSequence+addSequence; |
---|
1172 | } |
---|
1173 | } |
---|
1174 | } |
---|
1175 | } |
---|
1176 | interesting[0]=numberRemaining; |
---|
1177 | marker[0][0] = numberRemaining; |
---|
1178 | |
---|
1179 | if (!numberRemaining) |
---|
1180 | return; // Looks infeasible |
---|
1181 | |
---|
1182 | if (sequenceIn_>=0) { |
---|
1183 | // free variable - always choose |
---|
1184 | } else { |
---|
1185 | |
---|
1186 | theta_=1.0e50; |
---|
1187 | // now flip flop between spare arrays until reasonable theta |
---|
1188 | tentativeTheta = max(10.0*upperTheta,1.0e-7); |
---|
1189 | |
---|
1190 | // loops increasing tentative theta until can't go through |
---|
1191 | |
---|
1192 | while (tentativeTheta < 1.0e22) { |
---|
1193 | double thruThis = 0.0; |
---|
1194 | |
---|
1195 | double bestPivot=acceptablePivot; |
---|
1196 | int bestSequence=-1; |
---|
1197 | |
---|
1198 | numberPossiblySwapped = numberColumns_; |
---|
1199 | numberRemaining = 0; |
---|
1200 | |
---|
1201 | upperTheta = 1.0e50; |
---|
1202 | |
---|
1203 | spare = array[iFlip]; |
---|
1204 | index = indices[iFlip]; |
---|
1205 | spare2 = array[1-iFlip]; |
---|
1206 | index2 = indices[1-iFlip]; |
---|
1207 | |
---|
1208 | // try 3 different ways |
---|
1209 | // 1 bias increase by ones with slightly wrong djs |
---|
1210 | // 2 bias by all |
---|
1211 | // 3 bias by all - tolerance (doesn't seem very good) |
---|
1212 | #define TRYBIAS 1 |
---|
1213 | |
---|
1214 | |
---|
1215 | double increaseInThis=0.0; //objective increase in this loop |
---|
1216 | |
---|
1217 | for (i=0;i<interesting[iFlip];i++) { |
---|
1218 | int iSequence = index[i]; |
---|
1219 | double alpha = spare[i]; |
---|
1220 | double oldValue = dj_[iSequence]; |
---|
1221 | double value = oldValue-tentativeTheta*alpha; |
---|
1222 | |
---|
1223 | if (alpha < 0.0) { |
---|
1224 | //at upper bound |
---|
1225 | if (value>newTolerance) { |
---|
1226 | double range = upper_[iSequence] - lower_[iSequence]; |
---|
1227 | thruThis -= range*alpha; |
---|
1228 | #if TRYBIAS==1 |
---|
1229 | if (oldValue>0.0) |
---|
1230 | increaseInThis -= oldValue*range; |
---|
1231 | #elif TRYBIAS==2 |
---|
1232 | increaseInThis -= oldValue*range; |
---|
1233 | #else |
---|
1234 | increaseInThis -= (oldValue+dualTolerance_)*range; |
---|
1235 | #endif |
---|
1236 | // goes on swapped list (also means candidates if too many) |
---|
1237 | spare2[--numberPossiblySwapped]=alpha; |
---|
1238 | index2[numberPossiblySwapped]=iSequence; |
---|
1239 | if (fabs(alpha)>bestPivot) { |
---|
1240 | bestPivot=fabs(alpha); |
---|
1241 | bestSequence=numberPossiblySwapped; |
---|
1242 | } |
---|
1243 | } else { |
---|
1244 | value = oldValue-upperTheta*alpha; |
---|
1245 | if (value>newTolerance && -alpha>=acceptablePivot) |
---|
1246 | upperTheta = (oldValue-newTolerance)/alpha; |
---|
1247 | spare2[numberRemaining]=alpha; |
---|
1248 | index2[numberRemaining++]=iSequence; |
---|
1249 | } |
---|
1250 | } else { |
---|
1251 | // at lower bound |
---|
1252 | if (value<-newTolerance) { |
---|
1253 | double range = upper_[iSequence] - lower_[iSequence]; |
---|
1254 | thruThis += range*alpha; |
---|
1255 | //?? is this correct - and should we look at good ones |
---|
1256 | #if TRYBIAS==1 |
---|
1257 | if (oldValue<0.0) |
---|
1258 | increaseInThis += oldValue*range; |
---|
1259 | #elif TRYBIAS==2 |
---|
1260 | increaseInThis += oldValue*range; |
---|
1261 | #else |
---|
1262 | increaseInThis += (oldValue-dualTolerance_)*range; |
---|
1263 | #endif |
---|
1264 | // goes on swapped list (also means candidates if too many) |
---|
1265 | spare2[--numberPossiblySwapped]=alpha; |
---|
1266 | index2[numberPossiblySwapped]=iSequence; |
---|
1267 | if (fabs(alpha)>bestPivot) { |
---|
1268 | bestPivot=fabs(alpha); |
---|
1269 | bestSequence=numberPossiblySwapped; |
---|
1270 | } |
---|
1271 | } else { |
---|
1272 | value = oldValue-upperTheta*alpha; |
---|
1273 | if (value<-newTolerance && alpha>=acceptablePivot) |
---|
1274 | upperTheta = (oldValue+newTolerance)/alpha; |
---|
1275 | spare2[numberRemaining]=alpha; |
---|
1276 | index2[numberRemaining++]=iSequence; |
---|
1277 | } |
---|
1278 | } |
---|
1279 | } |
---|
1280 | swapped[1-iFlip]=numberPossiblySwapped; |
---|
1281 | interesting[1-iFlip]=numberRemaining; |
---|
1282 | marker[1-iFlip][0]= max(marker[1-iFlip][0],numberRemaining); |
---|
1283 | marker[1-iFlip][1]= min(marker[1-iFlip][1],numberPossiblySwapped); |
---|
1284 | |
---|
1285 | if (totalThru+thruThis>=fabs(dualOut_)|| |
---|
1286 | increaseInObjective+increaseInThis<0.0) { |
---|
1287 | // We should be pivoting in this batch |
---|
1288 | // so compress down to this lot |
---|
1289 | numberRemaining=0; |
---|
1290 | for (i=numberColumns_-1;i>=swapped[1-iFlip];i--) { |
---|
1291 | spare[numberRemaining]=spare2[i]; |
---|
1292 | index[numberRemaining++]=index2[i]; |
---|
1293 | } |
---|
1294 | interesting[iFlip]=numberRemaining; |
---|
1295 | int iTry; |
---|
1296 | #define MAXTRY 100 |
---|
1297 | // first get ratio with tolerance |
---|
1298 | for (iTry=0;iTry<MAXTRY;iTry++) { |
---|
1299 | |
---|
1300 | upperTheta=1.0e50; |
---|
1301 | numberPossiblySwapped = numberColumns_; |
---|
1302 | numberRemaining = 0; |
---|
1303 | |
---|
1304 | increaseInThis=0.0; //objective increase in this loop |
---|
1305 | |
---|
1306 | thruThis=0.0; |
---|
1307 | |
---|
1308 | spare = array[iFlip]; |
---|
1309 | index = indices[iFlip]; |
---|
1310 | spare2 = array[1-iFlip]; |
---|
1311 | index2 = indices[1-iFlip]; |
---|
1312 | |
---|
1313 | for (i=0;i<interesting[iFlip];i++) { |
---|
1314 | int iSequence=index[i]; |
---|
1315 | double alpha=spare[i]; |
---|
1316 | double oldValue = dj_[iSequence]; |
---|
1317 | double value = oldValue-upperTheta*alpha; |
---|
1318 | |
---|
1319 | if (alpha < 0.0) { |
---|
1320 | //at upper bound |
---|
1321 | if (value>newTolerance) { |
---|
1322 | if (-alpha>=acceptablePivot) { |
---|
1323 | upperTheta = (oldValue-newTolerance)/alpha; |
---|
1324 | } |
---|
1325 | } |
---|
1326 | } else { |
---|
1327 | // at lower bound |
---|
1328 | if (value<-newTolerance) { |
---|
1329 | if (alpha>=acceptablePivot) { |
---|
1330 | upperTheta = (oldValue+newTolerance)/alpha; |
---|
1331 | } |
---|
1332 | } |
---|
1333 | } |
---|
1334 | } |
---|
1335 | bestPivot=acceptablePivot; |
---|
1336 | sequenceIn_=-1; |
---|
1337 | // now choose largest and sum all ones which will go through |
---|
1338 | #define MINIMUMTHETA 1.0e-12 |
---|
1339 | for (i=0;i<interesting[iFlip];i++) { |
---|
1340 | int iSequence=index[i]; |
---|
1341 | double alpha=spare[i]; |
---|
1342 | double value = dj_[iSequence]-upperTheta*alpha; |
---|
1343 | double badDj=0.0; |
---|
1344 | |
---|
1345 | bool addToSwapped=false; |
---|
1346 | |
---|
1347 | if (alpha < 0.0) { |
---|
1348 | //at upper bound |
---|
1349 | if (value>=0.0) { |
---|
1350 | addToSwapped=true; |
---|
1351 | #if TRYBIAS==1 |
---|
1352 | badDj = -max(dj_[iSequence],0.0); |
---|
1353 | #elif TRYBIAS==2 |
---|
1354 | badDj = -dj_[iSequence]; |
---|
1355 | #else |
---|
1356 | badDj = -dj_[iSequence]-dualTolerance_; |
---|
1357 | #endif |
---|
1358 | } |
---|
1359 | } else { |
---|
1360 | // at lower bound |
---|
1361 | if (value<=0.0) { |
---|
1362 | addToSwapped=true; |
---|
1363 | #if TRYBIAS==1 |
---|
1364 | badDj = min(dj_[iSequence],0.0); |
---|
1365 | #elif TRYBIAS==2 |
---|
1366 | badDj = dj_[iSequence]; |
---|
1367 | #else |
---|
1368 | badDj = dj_[iSequence]-dualTolerance_; |
---|
1369 | #endif |
---|
1370 | } |
---|
1371 | } |
---|
1372 | if (!addToSwapped) { |
---|
1373 | // add to list of remaining |
---|
1374 | spare2[numberRemaining]=alpha; |
---|
1375 | index2[numberRemaining++]=iSequence; |
---|
1376 | } else { |
---|
1377 | // add to list of swapped |
---|
1378 | spare2[--numberPossiblySwapped]=alpha; |
---|
1379 | index2[numberPossiblySwapped]=iSequence; |
---|
1380 | // select if largest pivot |
---|
1381 | if (fabs(alpha)>bestPivot) { |
---|
1382 | sequenceIn_ = numberPossiblySwapped; |
---|
1383 | bestPivot = fabs(alpha); |
---|
1384 | theta_ = dj_[iSequence]/alpha; |
---|
1385 | } |
---|
1386 | double range = upper[iSequence] - lower[iSequence]; |
---|
1387 | thruThis += range*fabs(alpha); |
---|
1388 | increaseInThis += badDj*range; |
---|
1389 | } |
---|
1390 | } |
---|
1391 | swapped[1-iFlip]=numberPossiblySwapped; |
---|
1392 | interesting[1-iFlip]=numberRemaining; |
---|
1393 | marker[1-iFlip][0]= max(marker[1-iFlip][0],numberRemaining); |
---|
1394 | marker[1-iFlip][1]= min(marker[1-iFlip][1],numberPossiblySwapped); |
---|
1395 | // If we stop now this will be increase in objective (I think) |
---|
1396 | double increase = (fabs(dualOut_)-totalThru)*theta_; |
---|
1397 | increase += increaseInObjective; |
---|
1398 | if (theta_<0.0) |
---|
1399 | thruThis += fabs(dualOut_); // force using this one |
---|
1400 | if (increaseInObjective<0.0&&increase<0.0&&lastSequence>=0) { |
---|
1401 | // back |
---|
1402 | // We may need to be more careful - we could do by |
---|
1403 | // switch so we always do fine grained? |
---|
1404 | bestPivot=0.0; |
---|
1405 | } else { |
---|
1406 | // add in |
---|
1407 | totalThru += thruThis; |
---|
1408 | increaseInObjective += increaseInThis; |
---|
1409 | } |
---|
1410 | if (bestPivot<0.1*bestEverPivot&& |
---|
1411 | bestEverPivot>1.0e-6&&bestPivot<1.0e-3) { |
---|
1412 | // back to previous one |
---|
1413 | sequenceIn_=lastSequence; |
---|
1414 | // swap regions |
---|
1415 | iFlip = 1-iFlip; |
---|
1416 | break; |
---|
1417 | } else if (sequenceIn_==-1&&upperTheta>largeValue_) { |
---|
1418 | if (lastPivot>acceptablePivot) { |
---|
1419 | // back to previous one |
---|
1420 | sequenceIn_=lastSequence; |
---|
1421 | // swap regions |
---|
1422 | iFlip = 1-iFlip; |
---|
1423 | } else { |
---|
1424 | // can only get here if all pivots too small |
---|
1425 | } |
---|
1426 | break; |
---|
1427 | } else if (totalThru>=fabs(dualOut_)) { |
---|
1428 | modifyCosts=true; // fine grain - we can modify costs |
---|
1429 | break; // no point trying another loop |
---|
1430 | } else { |
---|
1431 | lastSequence=sequenceIn_; |
---|
1432 | if (bestPivot>bestEverPivot) |
---|
1433 | bestEverPivot=bestPivot; |
---|
1434 | iFlip = 1 -iFlip; |
---|
1435 | modifyCosts=true; // fine grain - we can modify costs |
---|
1436 | } |
---|
1437 | } |
---|
1438 | if (iTry==MAXTRY) |
---|
1439 | iFlip = 1-iFlip; // flip back |
---|
1440 | break; |
---|
1441 | } else { |
---|
1442 | // skip this lot |
---|
1443 | if (bestPivot>1.0e-3||bestPivot>bestEverPivot) { |
---|
1444 | bestEverPivot=bestPivot; |
---|
1445 | lastSequence=bestSequence; |
---|
1446 | } else { |
---|
1447 | // keep old swapped |
---|
1448 | memcpy(array[1-iFlip]+swapped[iFlip], |
---|
1449 | array[iFlip]+swapped[iFlip], |
---|
1450 | (numberColumns_-swapped[iFlip])*sizeof(double)); |
---|
1451 | memcpy(indices[1-iFlip]+swapped[iFlip], |
---|
1452 | indices[iFlip]+swapped[iFlip], |
---|
1453 | (numberColumns_-swapped[iFlip])*sizeof(int)); |
---|
1454 | marker[1-iFlip][1] = min(marker[1-iFlip][1],swapped[iFlip]); |
---|
1455 | swapped[1-iFlip]=swapped[iFlip]; |
---|
1456 | } |
---|
1457 | increaseInObjective += increaseInThis; |
---|
1458 | iFlip = 1 - iFlip; // swap regions |
---|
1459 | tentativeTheta = 2.0*upperTheta; |
---|
1460 | totalThru += thruThis; |
---|
1461 | } |
---|
1462 | } |
---|
1463 | |
---|
1464 | // can get here without sequenceIn_ set but with lastSequence |
---|
1465 | if (sequenceIn_<0&&lastSequence>=0) { |
---|
1466 | // back to previous one |
---|
1467 | sequenceIn_=lastSequence; |
---|
1468 | // swap regions |
---|
1469 | iFlip = 1-iFlip; |
---|
1470 | } |
---|
1471 | |
---|
1472 | if (sequenceIn_>=0) { |
---|
1473 | // at this stage sequenceIn_ is just pointer into index array |
---|
1474 | // flip just so we can use iFlip |
---|
1475 | iFlip = 1 -iFlip; |
---|
1476 | spare = array[iFlip]; |
---|
1477 | index = indices[iFlip]; |
---|
1478 | double oldValue; |
---|
1479 | double alpha = spare[sequenceIn_]; |
---|
1480 | sequenceIn_ = indices[iFlip][sequenceIn_]; |
---|
1481 | oldValue = dj_[sequenceIn_]; |
---|
1482 | theta_ = oldValue/alpha; |
---|
1483 | if (theta_<MINIMUMTHETA) { |
---|
1484 | // can't pivot to zero |
---|
1485 | if (oldValue-MINIMUMTHETA*alpha>=-dualTolerance_) { |
---|
1486 | theta_=MINIMUMTHETA; |
---|
1487 | } else if (oldValue-MINIMUMTHETA*alpha>=-newTolerance) { |
---|
1488 | theta_=MINIMUMTHETA; |
---|
1489 | thisIncrease=true; |
---|
1490 | } else { |
---|
1491 | theta_=(oldValue+newTolerance)/alpha; |
---|
1492 | assert(theta_>=0.0); |
---|
1493 | thisIncrease=true; |
---|
1494 | } |
---|
1495 | } |
---|
1496 | // may need to adjust costs so all dual feasible AND pivoted is exactly 0 |
---|
1497 | if (modifyCosts) { |
---|
1498 | int i; |
---|
1499 | double * workRow = rowArray->denseVector(); |
---|
1500 | double * workColumn = columnArray->denseVector(); |
---|
1501 | for (i=numberColumns_-1;i>=swapped[iFlip];i--) { |
---|
1502 | int iSequence=index[i]; |
---|
1503 | double alpha; |
---|
1504 | if (iSequence>=numberColumns_) |
---|
1505 | alpha=workRow[iSequence-numberColumns_]; |
---|
1506 | else |
---|
1507 | alpha=workColumn[iSequence]; |
---|
1508 | double value = dj_[iSequence]-theta_*alpha; |
---|
1509 | |
---|
1510 | // can't be free here |
---|
1511 | |
---|
1512 | if (alpha < 0.0) { |
---|
1513 | //at upper bound |
---|
1514 | if (value>dualTolerance_) { |
---|
1515 | thisIncrease=true; |
---|
1516 | #define MODIFYCOST 2 |
---|
1517 | #if MODIFYCOST |
---|
1518 | // modify cost to hit new tolerance |
---|
1519 | double modification = alpha*theta_-dj_[iSequence] |
---|
1520 | +newTolerance; |
---|
1521 | //modification = min(modification,dualTolerance_); |
---|
1522 | //assert (fabs(modification)<1.0e-7); |
---|
1523 | dj_[iSequence] += modification; |
---|
1524 | cost_[iSequence] += modification; |
---|
1525 | #endif |
---|
1526 | } |
---|
1527 | } else { |
---|
1528 | // at lower bound |
---|
1529 | if (-value>dualTolerance_) { |
---|
1530 | thisIncrease=true; |
---|
1531 | #if MODIFYCOST |
---|
1532 | // modify cost to hit new tolerance |
---|
1533 | double modification = alpha*theta_-dj_[iSequence] |
---|
1534 | -newTolerance; |
---|
1535 | //modification = max(modification,-dualTolerance_); |
---|
1536 | //assert (fabs(modification)<1.0e-7); |
---|
1537 | dj_[iSequence] += modification; |
---|
1538 | cost_[iSequence] += modification; |
---|
1539 | #endif |
---|
1540 | } |
---|
1541 | } |
---|
1542 | } |
---|
1543 | } |
---|
1544 | } |
---|
1545 | } |
---|
1546 | |
---|
1547 | if (sequenceIn_>=0) { |
---|
1548 | if (sequenceIn_>=numberColumns_) { |
---|
1549 | //slack |
---|
1550 | alpha_ = rowArray->denseVector()[sequenceIn_-numberColumns_]; |
---|
1551 | } else { |
---|
1552 | // column |
---|
1553 | alpha_ = columnArray->denseVector()[sequenceIn_]; |
---|
1554 | } |
---|
1555 | lowerIn_ = lower_[sequenceIn_]; |
---|
1556 | upperIn_ = upper_[sequenceIn_]; |
---|
1557 | valueIn_ = solution_[sequenceIn_]; |
---|
1558 | dualIn_ = dj_[sequenceIn_]; |
---|
1559 | |
---|
1560 | if (numberTimesOptimal_) { |
---|
1561 | // can we adjust cost back closer to original |
---|
1562 | //*** add coding |
---|
1563 | } |
---|
1564 | #if MODIFYCOST>1 |
---|
1565 | // modify cost to hit zero exactly |
---|
1566 | // so (dualIn_+modification)==theta_*alpha_ |
---|
1567 | double modification = theta_*alpha_-dualIn_; |
---|
1568 | dualIn_ += modification; |
---|
1569 | dj_[sequenceIn_]=dualIn_; |
---|
1570 | cost_[sequenceIn_] += modification; |
---|
1571 | //assert (fabs(modification)<1.0e-6); |
---|
1572 | #ifdef CLP_DEBUG |
---|
1573 | if ((handler_->logLevel()&32)&&fabs(modification)>1.0e-15) |
---|
1574 | printf("exact %d new cost %g, change %g\n",sequenceIn_, |
---|
1575 | cost_[sequenceIn_],modification); |
---|
1576 | #endif |
---|
1577 | #endif |
---|
1578 | |
---|
1579 | if (alpha_<0.0) { |
---|
1580 | // as if from upper bound |
---|
1581 | directionIn_=-1; |
---|
1582 | upperIn_=valueIn_; |
---|
1583 | } else { |
---|
1584 | // as if from lower bound |
---|
1585 | directionIn_=1; |
---|
1586 | lowerIn_=valueIn_; |
---|
1587 | } |
---|
1588 | } |
---|
1589 | if (thisIncrease) { |
---|
1590 | newTolerance = dualTolerance_+1.0e-4*dblParam_[OsiDualTolerance]; |
---|
1591 | } |
---|
1592 | |
---|
1593 | // clear arrays |
---|
1594 | |
---|
1595 | for (i=0;i<2;i++) { |
---|
1596 | memset(array[i],0,marker[i][0]*sizeof(double)); |
---|
1597 | memset(array[i]+marker[i][1],0, |
---|
1598 | (numberColumns_-marker[i][1])*sizeof(double)); |
---|
1599 | } |
---|
1600 | } |
---|
1601 | /* Checks if tentative optimal actually means unbounded |
---|
1602 | Returns -3 if not, 2 if is unbounded */ |
---|
1603 | int |
---|
1604 | ClpSimplexDual::checkUnbounded(OsiIndexedVector * ray, |
---|
1605 | OsiIndexedVector * spare, |
---|
1606 | double changeCost) |
---|
1607 | { |
---|
1608 | int status=2; // say unbounded |
---|
1609 | factorization_->updateColumn(spare,ray); |
---|
1610 | // get reduced cost |
---|
1611 | int i; |
---|
1612 | int number=ray->getNumElements(); |
---|
1613 | int * index = ray->getIndices(); |
---|
1614 | double * array = ray->denseVector(); |
---|
1615 | for (i=0;i<number;i++) { |
---|
1616 | int iRow=index[i]; |
---|
1617 | int iPivot=pivotVariable_[iRow]; |
---|
1618 | changeCost -= cost(iPivot)*array[iRow]; |
---|
1619 | } |
---|
1620 | double way; |
---|
1621 | if (changeCost>0.0) { |
---|
1622 | //try going down |
---|
1623 | way=1.0; |
---|
1624 | } else if (changeCost<0.0) { |
---|
1625 | //try going up |
---|
1626 | way=-1.0; |
---|
1627 | } else { |
---|
1628 | #ifdef CLP_DEBUG |
---|
1629 | printf("can't decide on up or down\n"); |
---|
1630 | #endif |
---|
1631 | way=0.0; |
---|
1632 | status=-3; |
---|
1633 | } |
---|
1634 | double movement=1.0e10*way; // some largish number |
---|
1635 | double zeroTolerance = 1.0e-14*dualBound_; |
---|
1636 | for (i=0;i<number;i++) { |
---|
1637 | int iRow=index[i]; |
---|
1638 | int iPivot=pivotVariable_[iRow]; |
---|
1639 | double arrayValue = array[iRow]; |
---|
1640 | if (fabs(arrayValue)<zeroTolerance) |
---|
1641 | arrayValue=0.0; |
---|
1642 | double newValue=solution(iPivot)+movement*arrayValue; |
---|
1643 | if (newValue>upper(iPivot)+primalTolerance_|| |
---|
1644 | newValue<lower(iPivot)-primalTolerance_) |
---|
1645 | status=-3; // not unbounded |
---|
1646 | } |
---|
1647 | if (status==2) { |
---|
1648 | // create ray |
---|
1649 | delete [] ray_; |
---|
1650 | ray_ = new double [numberColumns_]; |
---|
1651 | CoinFillN(ray_,numberColumns_,0.0); |
---|
1652 | for (i=0;i<number;i++) { |
---|
1653 | int iRow=index[i]; |
---|
1654 | int iPivot=pivotVariable_[iRow]; |
---|
1655 | double arrayValue = array[iRow]; |
---|
1656 | if (iPivot<numberColumns_&&fabs(arrayValue)>=zeroTolerance) |
---|
1657 | ray_[iPivot] = way* array[iRow]; |
---|
1658 | } |
---|
1659 | } |
---|
1660 | ray->clear(); |
---|
1661 | return status; |
---|
1662 | } |
---|
1663 | /* Checks if finished. Updates status */ |
---|
1664 | void |
---|
1665 | ClpSimplexDual::statusOfProblemInDual(int & lastCleaned,int type) |
---|
1666 | { |
---|
1667 | if (type==2) { |
---|
1668 | // trouble - restore solution |
---|
1669 | memcpy(status_ ,saveStatus_,(numberColumns_+numberRows_)*sizeof(char)); |
---|
1670 | memcpy(rowActivityWork_,savedSolution_+numberColumns_ , |
---|
1671 | numberRows_*sizeof(double)); |
---|
1672 | memcpy(columnActivityWork_,savedSolution_ , |
---|
1673 | numberColumns_*sizeof(double)); |
---|
1674 | forceFactorization_=1; // a bit drastic but .. |
---|
1675 | changeMade_++; // say something changed |
---|
1676 | } |
---|
1677 | int tentativeStatus = problemStatus_; |
---|
1678 | double changeCost; |
---|
1679 | |
---|
1680 | if (problemStatus_>-3) { |
---|
1681 | // factorize |
---|
1682 | // later on we will need to recover from singularities |
---|
1683 | // also we could skip if first time |
---|
1684 | // save dual weights |
---|
1685 | dualRowPivot_->saveWeights(this,1); |
---|
1686 | // is factorization okay? |
---|
1687 | if (internalFactorize(1)) { |
---|
1688 | // no - restore previous basis |
---|
1689 | assert (type==1); |
---|
1690 | changeMade_++; // say something changed |
---|
1691 | memcpy(status_ ,saveStatus_,(numberColumns_+numberRows_)*sizeof(char)); |
---|
1692 | memcpy(rowActivityWork_,savedSolution_+numberColumns_ , |
---|
1693 | numberRows_*sizeof(double)); |
---|
1694 | memcpy(columnActivityWork_,savedSolution_ , |
---|
1695 | numberColumns_*sizeof(double)); |
---|
1696 | // get correct bounds on all variables |
---|
1697 | double dummyChangeCost=0.0; |
---|
1698 | changeBounds(true,rowArray_[2],dummyChangeCost); |
---|
1699 | // throw away change |
---|
1700 | rowArray_[2]->clear(); |
---|
1701 | forceFactorization_=1; // a bit drastic but .. |
---|
1702 | type = 2; |
---|
1703 | assert (internalFactorize(1)==0); |
---|
1704 | } |
---|
1705 | problemStatus_=-3; |
---|
1706 | } |
---|
1707 | // at this stage status is -3 or -4 if looks infeasible |
---|
1708 | // get primal and dual solutions |
---|
1709 | gutsOfSolution(rowActivityWork_,columnActivityWork_); |
---|
1710 | #ifdef CLP_DEBUG |
---|
1711 | if (!rowScale_&&(handler_->logLevel()&32)) { |
---|
1712 | double * objectiveSimplex |
---|
1713 | = copyOfArray(objective_,numberColumns_,0.0); |
---|
1714 | double * rowObjectiveSimplex |
---|
1715 | = copyOfArray(rowObjective_,numberRows_,0.0); |
---|
1716 | int i; |
---|
1717 | double largest; |
---|
1718 | largest=0.0; |
---|
1719 | for (i=0;i<numberRows_;i++) { |
---|
1720 | rowObjectiveSimplex[i] *= optimizationDirection_; |
---|
1721 | double difference = fabs(rowObjectiveWork_[i]-rowObjectiveSimplex[i]); |
---|
1722 | if (difference>largest) |
---|
1723 | largest=difference; |
---|
1724 | } |
---|
1725 | for (i=0;i<numberColumns_;i++) { |
---|
1726 | objectiveSimplex[i] *= optimizationDirection_; |
---|
1727 | double difference = fabs(objectiveWork_[i]-objectiveSimplex[i]); |
---|
1728 | if (difference>largest) |
---|
1729 | largest=difference; |
---|
1730 | } |
---|
1731 | if ((handler_->logLevel()&16)) |
---|
1732 | printf("difference in obj %g\n",largest); |
---|
1733 | delete [] objectiveSimplex; |
---|
1734 | delete [] rowObjectiveSimplex; |
---|
1735 | } |
---|
1736 | #endif |
---|
1737 | handler_->message(CLP_SIMPLEX_STATUS,messages_) |
---|
1738 | <<numberIterations_<<objectiveValue(); |
---|
1739 | handler_->printing(sumPrimalInfeasibilities_>0.0) |
---|
1740 | <<sumPrimalInfeasibilities_<<numberPrimalInfeasibilities_; |
---|
1741 | handler_->printing(sumDualInfeasibilities_>0.0) |
---|
1742 | <<sumDualInfeasibilities_<<numberDualInfeasibilities_; |
---|
1743 | handler_->printing(numberDualInfeasibilitiesWithoutFree_ |
---|
1744 | <numberDualInfeasibilities_) |
---|
1745 | <<numberDualInfeasibilities_- |
---|
1746 | numberDualInfeasibilitiesWithoutFree_; |
---|
1747 | handler_->message()<<OsiMessageEol; |
---|
1748 | while (problemStatus_<=-3) { |
---|
1749 | bool cleanDuals=false; |
---|
1750 | int numberChangedBounds=0; |
---|
1751 | int doOriginalTolerance=0; |
---|
1752 | if ( lastCleaned==numberIterations_) |
---|
1753 | doOriginalTolerance=1; |
---|
1754 | // check optimal |
---|
1755 | if (dualFeasible()||problemStatus_==-4) { |
---|
1756 | if (primalFeasible()) { |
---|
1757 | // may be optimal - or may be bounds are wrong |
---|
1758 | handler_->message(CLP_DUAL_BOUNDS,messages_) |
---|
1759 | <<dualBound_ |
---|
1760 | <<OsiMessageEol; |
---|
1761 | // save solution in case unbounded |
---|
1762 | CoinDisjointCopyN(columnActivityWork_,numberColumns_, |
---|
1763 | columnArray_[0]->denseVector()); |
---|
1764 | CoinDisjointCopyN(rowActivityWork_,numberRows_, |
---|
1765 | rowArray_[2]->denseVector()); |
---|
1766 | numberChangedBounds=changeBounds(false,rowArray_[0],changeCost); |
---|
1767 | if (numberChangedBounds<=0) { |
---|
1768 | //looks optimal - do we need to reset tolerance |
---|
1769 | if (lastCleaned<numberIterations_&&numberTimesOptimal_<4) { |
---|
1770 | doOriginalTolerance=2; |
---|
1771 | numberTimesOptimal_++; |
---|
1772 | changeMade_++; // say something changed |
---|
1773 | if (numberTimesOptimal_==1) { |
---|
1774 | dualTolerance_ = min(dualTolerance_,1.0e-8); |
---|
1775 | // better to have small tolerance even if slower |
---|
1776 | factorization_->zeroTolerance(1.0e-15); |
---|
1777 | } |
---|
1778 | } else { |
---|
1779 | problemStatus_=0; // optimal |
---|
1780 | if (lastCleaned<numberIterations_) { |
---|
1781 | handler_->message(CLP_SIMPLEX_GIVINGUP,messages_) |
---|
1782 | <<OsiMessageEol; |
---|
1783 | } |
---|
1784 | } |
---|
1785 | } else { |
---|
1786 | cleanDuals=true; |
---|
1787 | if (doOriginalTolerance==1) { |
---|
1788 | // check unbounded |
---|
1789 | problemStatus_ = checkUnbounded(rowArray_[0],rowArray_[1], |
---|
1790 | changeCost); |
---|
1791 | if (problemStatus_==2) { |
---|
1792 | // it is unbounded - restore solution |
---|
1793 | // but first add in changes to non-basic |
---|
1794 | int iColumn; |
---|
1795 | double * original = columnArray_[0]->denseVector(); |
---|
1796 | for (iColumn=0;iColumn<numberColumns_;iColumn++) { |
---|
1797 | if(getColumnStatus(iColumn)!= ClpSimplex::basic) |
---|
1798 | ray_[iColumn] += |
---|
1799 | columnActivityWork_[iColumn]-original[iColumn]; |
---|
1800 | columnActivityWork_[iColumn] = original[iColumn]; |
---|
1801 | } |
---|
1802 | CoinDisjointCopyN(rowArray_[2]->denseVector(),numberRows_, |
---|
1803 | rowActivityWork_); |
---|
1804 | } |
---|
1805 | } else { |
---|
1806 | doOriginalTolerance=2; |
---|
1807 | rowArray_[0]->clear(); |
---|
1808 | } |
---|
1809 | } |
---|
1810 | CoinFillN(columnArray_[0]->denseVector(),numberColumns_,0.0); |
---|
1811 | CoinFillN(rowArray_[2]->denseVector(),numberRows_,0.0); |
---|
1812 | } |
---|
1813 | if (problemStatus_==-4) { |
---|
1814 | // may be infeasible - or may be bounds are wrong |
---|
1815 | handler_->message(CLP_DUAL_CHECKB,messages_) |
---|
1816 | <<dualBound_ |
---|
1817 | <<OsiMessageEol; |
---|
1818 | numberChangedBounds=changeBounds(false,NULL,changeCost); |
---|
1819 | if (numberChangedBounds<=0||dualBound_>1.0e20|| |
---|
1820 | (largestPrimalError_>1.0&&dualBound_>1.0e17)) { |
---|
1821 | problemStatus_=1; // infeasible |
---|
1822 | } else { |
---|
1823 | problemStatus_=-1; //iterate |
---|
1824 | cleanDuals=true; |
---|
1825 | doOriginalTolerance=2; |
---|
1826 | // and delete ray which has been created |
---|
1827 | delete [] ray_; |
---|
1828 | ray_ = NULL; |
---|
1829 | } |
---|
1830 | } |
---|
1831 | } else { |
---|
1832 | cleanDuals=true; |
---|
1833 | } |
---|
1834 | if (problemStatus_<0) { |
---|
1835 | if (doOriginalTolerance==2) { |
---|
1836 | // put back original tolerance |
---|
1837 | lastCleaned=numberIterations_; |
---|
1838 | handler_->message(CLP_DUAL_ORIGINAL,messages_) |
---|
1839 | <<OsiMessageEol; |
---|
1840 | |
---|
1841 | perturbation_=102; // stop any perturbations |
---|
1842 | createRim(4); |
---|
1843 | // make sure duals are current |
---|
1844 | computeDuals(); |
---|
1845 | // put back bounds as they were if was optimal |
---|
1846 | if (doOriginalTolerance==2) { |
---|
1847 | changeMade_++; // say something changed |
---|
1848 | changeBounds(true,NULL,changeCost); |
---|
1849 | cleanDuals=true; |
---|
1850 | } |
---|
1851 | } |
---|
1852 | if (cleanDuals) { |
---|
1853 | // make sure dual feasible |
---|
1854 | // look at all rows and columns |
---|
1855 | rowArray_[0]->clear(); |
---|
1856 | CoinIotaN(rowArray_[0]->getIndices(),numberRows_,0); |
---|
1857 | rowArray_[0]->setNumElements(numberRows_); |
---|
1858 | columnArray_[0]->clear(); |
---|
1859 | CoinIotaN(columnArray_[0]->getIndices(),numberColumns_,0); |
---|
1860 | columnArray_[0]->setNumElements(numberColumns_); |
---|
1861 | double objectiveChange=0.0; |
---|
1862 | updateDualsInDual(rowArray_[0],columnArray_[0],rowArray_[1], |
---|
1863 | 0.0,objectiveChange); |
---|
1864 | // for now - recompute all |
---|
1865 | gutsOfSolution(rowActivityWork_,columnActivityWork_); |
---|
1866 | assert(numberDualInfeasibilitiesWithoutFree_==0); |
---|
1867 | if (numberDualInfeasibilities_) { |
---|
1868 | // bad free variables |
---|
1869 | if (primalFeasible()) { |
---|
1870 | std::cerr<<"Free variable problem?"<<std::endl; |
---|
1871 | abort(); // what now |
---|
1872 | } |
---|
1873 | problemStatus_=-1; // carry on as normal |
---|
1874 | } |
---|
1875 | } else { |
---|
1876 | // iterate |
---|
1877 | problemStatus_=-1; |
---|
1878 | } |
---|
1879 | } |
---|
1880 | } |
---|
1881 | if (type==0||type==1) { |
---|
1882 | if (!type) { |
---|
1883 | // create save arrays |
---|
1884 | delete [] saveStatus_; |
---|
1885 | delete [] savedSolution_; |
---|
1886 | saveStatus_ = new unsigned char [numberRows_+numberColumns_]; |
---|
1887 | savedSolution_ = new double [numberRows_+numberColumns_]; |
---|
1888 | } |
---|
1889 | // save arrays |
---|
1890 | memcpy(saveStatus_,status_,(numberColumns_+numberRows_)*sizeof(char)); |
---|
1891 | memcpy(savedSolution_+numberColumns_ ,rowActivityWork_, |
---|
1892 | numberRows_*sizeof(double)); |
---|
1893 | memcpy(savedSolution_ ,columnActivityWork_,numberColumns_*sizeof(double)); |
---|
1894 | } |
---|
1895 | |
---|
1896 | // restore weights (if saved) - also recompute infeasibility list |
---|
1897 | if (tentativeStatus>-3) |
---|
1898 | dualRowPivot_->saveWeights(this,(type <2) ? 2 : 4); |
---|
1899 | else |
---|
1900 | dualRowPivot_->saveWeights(this,3); |
---|
1901 | // unflag all variables (we may want to wait a bit?) |
---|
1902 | int iRow; |
---|
1903 | for (iRow=0;iRow<numberRows_;iRow++) { |
---|
1904 | int iPivot=pivotVariable_[iRow]; |
---|
1905 | clearFlagged(iPivot); |
---|
1906 | } |
---|
1907 | if (problemStatus_<0&&!changeMade_) { |
---|
1908 | problemStatus_=4; // unknown |
---|
1909 | } |
---|
1910 | |
---|
1911 | } |
---|
1912 | /* While updateDualsInDual sees what effect is of flip |
---|
1913 | this does actuall flipping. |
---|
1914 | If change >0.0 then value in array >0.0 => from lower to upper |
---|
1915 | */ |
---|
1916 | void |
---|
1917 | ClpSimplexDual::flipBounds(OsiIndexedVector * rowArray, |
---|
1918 | OsiIndexedVector * columnArray, |
---|
1919 | double change) |
---|
1920 | { |
---|
1921 | double * work; |
---|
1922 | int number; |
---|
1923 | int * which; |
---|
1924 | |
---|
1925 | int iSection; |
---|
1926 | |
---|
1927 | for (iSection=0;iSection<2;iSection++) { |
---|
1928 | int i; |
---|
1929 | double * solution = solutionRegion(iSection); |
---|
1930 | double * lower = lowerRegion(iSection); |
---|
1931 | double * upper = upperRegion(iSection); |
---|
1932 | int addSequence; |
---|
1933 | if (!iSection) { |
---|
1934 | work = rowArray->denseVector(); |
---|
1935 | number = rowArray->getNumElements(); |
---|
1936 | which = rowArray->getIndices(); |
---|
1937 | addSequence = numberColumns_; |
---|
1938 | } else { |
---|
1939 | work = columnArray->denseVector(); |
---|
1940 | number = columnArray->getNumElements(); |
---|
1941 | which = columnArray->getIndices(); |
---|
1942 | addSequence = 0; |
---|
1943 | } |
---|
1944 | |
---|
1945 | for (i=0;i<number;i++) { |
---|
1946 | int iSequence = which[i]; |
---|
1947 | #ifndef NDEBUG |
---|
1948 | double value = work[iSequence]*change; |
---|
1949 | #endif |
---|
1950 | work[iSequence]=0.0; |
---|
1951 | Status status = getStatus(iSequence+addSequence); |
---|
1952 | |
---|
1953 | switch(status) { |
---|
1954 | |
---|
1955 | case ClpSimplex::basic: |
---|
1956 | case ClpSimplex::isFree: |
---|
1957 | case ClpSimplex::superBasic: |
---|
1958 | break; |
---|
1959 | case ClpSimplex::atUpperBound: |
---|
1960 | assert (value<=0.0); |
---|
1961 | // to lower bound |
---|
1962 | setStatus(iSequence+addSequence,ClpSimplex::atLowerBound); |
---|
1963 | solution[iSequence] = lower[iSequence]; |
---|
1964 | break; |
---|
1965 | case ClpSimplex::atLowerBound: |
---|
1966 | assert (value>=0.0); |
---|
1967 | // to upper bound |
---|
1968 | setStatus(iSequence+addSequence,ClpSimplex::atUpperBound); |
---|
1969 | solution[iSequence] = upper[iSequence]; |
---|
1970 | break; |
---|
1971 | } |
---|
1972 | } |
---|
1973 | } |
---|
1974 | rowArray->setNumElements(0); |
---|
1975 | columnArray->setNumElements(0); |
---|
1976 | } |
---|
1977 | // Restores bound to original bound |
---|
1978 | void |
---|
1979 | ClpSimplexDual::originalBound( int iSequence) |
---|
1980 | { |
---|
1981 | if (iSequence>=numberColumns_) { |
---|
1982 | // rows |
---|
1983 | int iRow = iSequence-numberColumns_; |
---|
1984 | rowLowerWork_[iRow]=rowLower_[iRow]; |
---|
1985 | rowUpperWork_[iRow]=rowUpper_[iRow]; |
---|
1986 | if (rowScale_) { |
---|
1987 | if (rowLowerWork_[iRow]>-1.0e50) |
---|
1988 | rowLowerWork_[iRow] *= rowScale_[iRow]; |
---|
1989 | if (rowUpperWork_[iRow]<1.0e50) |
---|
1990 | rowUpperWork_[iRow] *= rowScale_[iRow]; |
---|
1991 | } |
---|
1992 | } else { |
---|
1993 | // columns |
---|
1994 | columnLowerWork_[iSequence]=columnLower_[iSequence]; |
---|
1995 | columnUpperWork_[iSequence]=columnUpper_[iSequence]; |
---|
1996 | if (rowScale_) { |
---|
1997 | double multiplier = 1.0/columnScale_[iSequence]; |
---|
1998 | if (columnLowerWork_[iSequence]>-1.0e50) |
---|
1999 | columnLowerWork_[iSequence] *= multiplier; |
---|
2000 | if (columnUpperWork_[iSequence]<1.0e50) |
---|
2001 | columnUpperWork_[iSequence] *= multiplier; |
---|
2002 | } |
---|
2003 | } |
---|
2004 | setFakeBound(iSequence,ClpSimplex::noFake); |
---|
2005 | } |
---|
2006 | /* As changeBounds but just changes new bounds for a single variable. |
---|
2007 | Returns true if change */ |
---|
2008 | bool |
---|
2009 | ClpSimplexDual::changeBound( int iSequence) |
---|
2010 | { |
---|
2011 | // old values |
---|
2012 | double oldLower=lower_[iSequence]; |
---|
2013 | double oldUpper=upper_[iSequence]; |
---|
2014 | double value=solution_[iSequence]; |
---|
2015 | bool modified=false; |
---|
2016 | originalBound(iSequence); |
---|
2017 | // original values |
---|
2018 | double lowerValue=lower_[iSequence]; |
---|
2019 | double upperValue=upper_[iSequence]; |
---|
2020 | // back to altered values |
---|
2021 | lower_[iSequence] = oldLower; |
---|
2022 | upper_[iSequence] = oldUpper; |
---|
2023 | if (value==oldLower) { |
---|
2024 | if (upperValue > oldLower + dualBound_) { |
---|
2025 | upper_[iSequence]=oldLower+dualBound_; |
---|
2026 | setFakeBound(iSequence,ClpSimplex::upperFake); |
---|
2027 | modified=true; |
---|
2028 | } |
---|
2029 | } else if (value==oldUpper) { |
---|
2030 | if (lowerValue < oldUpper - dualBound_) { |
---|
2031 | lower_[iSequence]=oldUpper-dualBound_; |
---|
2032 | setFakeBound(iSequence,ClpSimplex::lowerFake); |
---|
2033 | modified=true; |
---|
2034 | } |
---|
2035 | } else { |
---|
2036 | assert(value==oldLower||value==oldUpper); |
---|
2037 | } |
---|
2038 | return modified; |
---|
2039 | } |
---|
2040 | // Perturbs problem |
---|
2041 | void |
---|
2042 | ClpSimplexDual::perturb() |
---|
2043 | { |
---|
2044 | if (perturbation_>100) |
---|
2045 | return; //perturbed already |
---|
2046 | int iRow,iColumn; |
---|
2047 | // dual perturbation |
---|
2048 | double perturbation=1.0e-20; |
---|
2049 | // maximum fraction of cost to perturb |
---|
2050 | double maximumFraction = 1.0e-4; |
---|
2051 | if (perturbation_==100) { |
---|
2052 | perturbation = 1.0e-4; |
---|
2053 | for (iRow=0;iRow<numberRows_;iRow++) { |
---|
2054 | double value = fabs(rowActivityWork_[iRow]*rowObjectiveWork_[iRow]); |
---|
2055 | perturbation = max(perturbation,value); |
---|
2056 | } |
---|
2057 | for (iColumn=0;iColumn<numberColumns_;iColumn++) { |
---|
2058 | double value = |
---|
2059 | fabs(columnActivityWork_[iColumn]*objectiveWork_[iColumn]); |
---|
2060 | perturbation = max(perturbation,value); |
---|
2061 | } |
---|
2062 | perturbation *= 1.0e-8; |
---|
2063 | } else if (perturbation_<100) { |
---|
2064 | perturbation = pow(10.0,perturbation_); |
---|
2065 | // user is in charge |
---|
2066 | maximumFraction = 1.0e100; |
---|
2067 | } |
---|
2068 | // modify costs |
---|
2069 | handler_->message(CLP_SIMPLEX_PERTURB,messages_) |
---|
2070 | <<perturbation |
---|
2071 | <<OsiMessageEol; |
---|
2072 | for (iRow=0;iRow<numberRows_;iRow++) { |
---|
2073 | double value = perturbation; |
---|
2074 | double currentValue = rowObjectiveWork_[iRow]; |
---|
2075 | value = min(value,maximumFraction*fabs(currentValue)+1.0e-6); |
---|
2076 | if (rowLowerWork_[iRow]>-largeValue_) { |
---|
2077 | if (fabs(rowLowerWork_[iRow])<fabs(rowUpperWork_[iRow])) |
---|
2078 | value *= drand48(); |
---|
2079 | else |
---|
2080 | value *= -drand48(); |
---|
2081 | } else if (rowUpperWork_[iRow]<largeValue_) { |
---|
2082 | value *= -drand48(); |
---|
2083 | } else { |
---|
2084 | value=0.0; |
---|
2085 | } |
---|
2086 | rowObjectiveWork_[iRow] += value; |
---|
2087 | } |
---|
2088 | for (iColumn=0;iColumn<numberColumns_;iColumn++) { |
---|
2089 | double value = perturbation; |
---|
2090 | double currentValue = objectiveWork_[iColumn]; |
---|
2091 | value = min(value,maximumFraction*fabs(currentValue)+1.0e-6); |
---|
2092 | if (columnLowerWork_[iColumn]>-largeValue_) { |
---|
2093 | if (fabs(columnLowerWork_[iColumn])< |
---|
2094 | fabs(columnUpperWork_[iColumn])) |
---|
2095 | value *= drand48(); |
---|
2096 | else |
---|
2097 | value *= -drand48(); |
---|
2098 | } else if (columnUpperWork_[iColumn]<largeValue_) { |
---|
2099 | value *= -drand48(); |
---|
2100 | } else { |
---|
2101 | value=0.0; |
---|
2102 | } |
---|
2103 | objectiveWork_[iColumn] += value; |
---|
2104 | } |
---|
2105 | // say perturbed |
---|
2106 | perturbation_=102; |
---|
2107 | |
---|
2108 | } |
---|
2109 | |
---|