|
1 /**************************************************************************** |
|
2 ** |
|
3 ** Copyright (C) 2009 Nokia Corporation and/or its subsidiary(-ies). |
|
4 ** All rights reserved. |
|
5 ** Contact: Nokia Corporation (qt-info@nokia.com) |
|
6 ** |
|
7 ** This file is part of the QtGui module of the Qt Toolkit. |
|
8 ** |
|
9 ** $QT_BEGIN_LICENSE:LGPL$ |
|
10 ** No Commercial Usage |
|
11 ** This file contains pre-release code and may not be distributed. |
|
12 ** You may use this file in accordance with the terms and conditions |
|
13 ** contained in the Technology Preview License Agreement accompanying |
|
14 ** this package. |
|
15 ** |
|
16 ** GNU Lesser General Public License Usage |
|
17 ** Alternatively, this file may be used under the terms of the GNU Lesser |
|
18 ** General Public License version 2.1 as published by the Free Software |
|
19 ** Foundation and appearing in the file LICENSE.LGPL included in the |
|
20 ** packaging of this file. Please review the following information to |
|
21 ** ensure the GNU Lesser General Public License version 2.1 requirements |
|
22 ** will be met: http://www.gnu.org/licenses/old-licenses/lgpl-2.1.html. |
|
23 ** |
|
24 ** In addition, as a special exception, Nokia gives you certain additional |
|
25 ** rights. These rights are described in the Nokia Qt LGPL Exception |
|
26 ** version 1.1, included in the file LGPL_EXCEPTION.txt in this package. |
|
27 ** |
|
28 ** If you have questions regarding the use of this file, please contact |
|
29 ** Nokia at qt-info@nokia.com. |
|
30 ** |
|
31 ** |
|
32 ** |
|
33 ** |
|
34 ** |
|
35 ** |
|
36 ** |
|
37 ** |
|
38 ** $QT_END_LICENSE$ |
|
39 ** |
|
40 ****************************************************************************/ |
|
41 |
|
42 #include "qsimplex_p.h" |
|
43 |
|
44 #include <QtCore/qset.h> |
|
45 #include <QtCore/qdebug.h> |
|
46 |
|
47 #include <stdlib.h> |
|
48 |
|
49 QT_BEGIN_NAMESPACE |
|
50 |
|
51 /*! |
|
52 \internal |
|
53 \class QSimplex |
|
54 |
|
55 The QSimplex class is a Linear Programming problem solver based on the two-phase |
|
56 simplex method. |
|
57 |
|
58 It takes a set of QSimplexConstraints as its restrictive constraints and an |
|
59 additional QSimplexConstraint as its objective function. Then methods to maximize |
|
60 and minimize the problem solution are provided. |
|
61 |
|
62 The two-phase simplex method is based on the following steps: |
|
63 First phase: |
|
64 1.a) Modify the original, complex, and possibly not feasible problem, into a new, |
|
65 easy to solve problem. |
|
66 1.b) Set as the objective of the new problem, a feasible solution for the original |
|
67 complex problem. |
|
68 1.c) Run simplex to optimize the modified problem and check whether a solution for |
|
69 the original problem exists. |
|
70 |
|
71 Second phase: |
|
72 2.a) Go back to the original problem with the feasibl (but not optimal) solution |
|
73 found in the first phase. |
|
74 2.b) Set the original objective. |
|
75 3.c) Run simplex to optimize the original problem towards its optimal solution. |
|
76 */ |
|
77 |
|
78 /*! |
|
79 \internal |
|
80 */ |
|
81 QSimplex::QSimplex() : objective(0), rows(0), columns(0), firstArtificial(0), matrix(0) |
|
82 { |
|
83 } |
|
84 |
|
85 /*! |
|
86 \internal |
|
87 */ |
|
88 QSimplex::~QSimplex() |
|
89 { |
|
90 clearDataStructures(); |
|
91 } |
|
92 |
|
93 /*! |
|
94 \internal |
|
95 */ |
|
96 void QSimplex::clearDataStructures() |
|
97 { |
|
98 if (matrix == 0) |
|
99 return; |
|
100 |
|
101 // Matrix |
|
102 rows = 0; |
|
103 columns = 0; |
|
104 firstArtificial = 0; |
|
105 free(matrix); |
|
106 matrix = 0; |
|
107 |
|
108 // Constraints |
|
109 for (int i = 0; i < constraints.size(); ++i) { |
|
110 delete constraints[i]->helper.first; |
|
111 constraints[i]->helper.first = 0; |
|
112 constraints[i]->helper.second = 0.0; |
|
113 delete constraints[i]->artificial; |
|
114 constraints[i]->artificial = 0; |
|
115 } |
|
116 constraints.clear(); |
|
117 |
|
118 // Other |
|
119 variables.clear(); |
|
120 objective = 0; |
|
121 } |
|
122 |
|
123 /*! |
|
124 \internal |
|
125 Sets the new constraints in the simplex solver and returns whether the problem |
|
126 is feasible. |
|
127 |
|
128 This method sets the new constraints, normalizes them, creates the simplex matrix |
|
129 and runs the first simplex phase. |
|
130 */ |
|
131 bool QSimplex::setConstraints(const QList<QSimplexConstraint *> newConstraints) |
|
132 { |
|
133 //////////////////////////// |
|
134 // Reset to initial state // |
|
135 //////////////////////////// |
|
136 clearDataStructures(); |
|
137 |
|
138 if (newConstraints.isEmpty()) |
|
139 return true; // we are ok with no constraints |
|
140 constraints = newConstraints; |
|
141 |
|
142 /////////////////////////////////////// |
|
143 // Prepare variables and constraints // |
|
144 /////////////////////////////////////// |
|
145 |
|
146 // Set Variables direct mapping. |
|
147 // "variables" is a list that provides a stable, indexed list of all variables |
|
148 // used in this problem. |
|
149 QSet<QSimplexVariable *> variablesSet; |
|
150 for (int i = 0; i < constraints.size(); ++i) |
|
151 variablesSet += \ |
|
152 QSet<QSimplexVariable *>::fromList(constraints[i]->variables.keys()); |
|
153 variables = variablesSet.toList(); |
|
154 |
|
155 // Set Variables reverse mapping |
|
156 // We also need to be able to find the index for a given variable, to do that |
|
157 // we store in each variable its index. |
|
158 for (int i = 0; i < variables.size(); ++i) { |
|
159 // The variable "0" goes at the column "1", etc... |
|
160 variables[i]->index = i + 1; |
|
161 } |
|
162 |
|
163 // Normalize Constraints |
|
164 // In this step, we prepare the constraints in two ways: |
|
165 // Firstly, we modify all constraints of type "LessOrEqual" or "MoreOrEqual" |
|
166 // by the adding slack or surplus variables and making them "Equal" constraints. |
|
167 // Secondly, we need every single constraint to have a direct, easy feasible |
|
168 // solution. Constraints that have slack variables are already easy to solve, |
|
169 // to all the others we add artificial variables. |
|
170 // |
|
171 // At the end we modify the constraints as follows: |
|
172 // - LessOrEqual: SLACK variable is added. |
|
173 // - Equal: ARTIFICIAL variable is added. |
|
174 // - More or Equal: ARTIFICIAL and SURPLUS variables are added. |
|
175 int variableIndex = variables.size(); |
|
176 QList <QSimplexVariable *> artificialList; |
|
177 |
|
178 for (int i = 0; i < constraints.size(); ++i) { |
|
179 QSimplexVariable *slack; |
|
180 QSimplexVariable *surplus; |
|
181 QSimplexVariable *artificial; |
|
182 |
|
183 Q_ASSERT(constraints[i]->helper.first == 0); |
|
184 Q_ASSERT(constraints[i]->artificial == 0); |
|
185 |
|
186 switch(constraints[i]->ratio) { |
|
187 case QSimplexConstraint::LessOrEqual: |
|
188 slack = new QSimplexVariable; |
|
189 slack->index = ++variableIndex; |
|
190 constraints[i]->helper.first = slack; |
|
191 constraints[i]->helper.second = 1.0; |
|
192 break; |
|
193 case QSimplexConstraint::MoreOrEqual: |
|
194 surplus = new QSimplexVariable; |
|
195 surplus->index = ++variableIndex; |
|
196 constraints[i]->helper.first = surplus; |
|
197 constraints[i]->helper.second = -1.0; |
|
198 // fall through |
|
199 case QSimplexConstraint::Equal: |
|
200 artificial = new QSimplexVariable; |
|
201 constraints[i]->artificial = artificial; |
|
202 artificialList += constraints[i]->artificial; |
|
203 break; |
|
204 } |
|
205 } |
|
206 |
|
207 // All original, slack and surplus have already had its index set |
|
208 // at this point. We now set the index of the artificial variables |
|
209 // as to ensure they are at the end of the variable list and therefore |
|
210 // can be easily removed at the end of this method. |
|
211 firstArtificial = variableIndex + 1; |
|
212 for (int i = 0; i < artificialList.size(); ++i) |
|
213 artificialList[i]->index = ++variableIndex; |
|
214 artificialList.clear(); |
|
215 |
|
216 ///////////////////////////// |
|
217 // Fill the Simplex matrix // |
|
218 ///////////////////////////// |
|
219 |
|
220 // One for each variable plus the Basic and BFS columns (first and last) |
|
221 columns = variableIndex + 2; |
|
222 // One for each constraint plus the objective function |
|
223 rows = constraints.size() + 1; |
|
224 |
|
225 matrix = (qreal *)malloc(sizeof(qreal) * columns * rows); |
|
226 if (!matrix) { |
|
227 qWarning() << "QSimplex: Unable to allocate memory!"; |
|
228 return false; |
|
229 } |
|
230 for (int i = columns * rows - 1; i >= 0; --i) |
|
231 matrix[i] = 0.0; |
|
232 |
|
233 // Fill Matrix |
|
234 for (int i = 1; i <= constraints.size(); ++i) { |
|
235 QSimplexConstraint *c = constraints[i - 1]; |
|
236 |
|
237 if (c->artificial) { |
|
238 // Will use artificial basic variable |
|
239 setValueAt(i, 0, c->artificial->index); |
|
240 setValueAt(i, c->artificial->index, 1.0); |
|
241 |
|
242 if (c->helper.second != 0.0) { |
|
243 // Surplus variable |
|
244 setValueAt(i, c->helper.first->index, c->helper.second); |
|
245 } |
|
246 } else { |
|
247 // Slack is used as the basic variable |
|
248 Q_ASSERT(c->helper.second == 1.0); |
|
249 setValueAt(i, 0, c->helper.first->index); |
|
250 setValueAt(i, c->helper.first->index, 1.0); |
|
251 } |
|
252 |
|
253 QHash<QSimplexVariable *, qreal>::const_iterator iter; |
|
254 for (iter = c->variables.constBegin(); |
|
255 iter != c->variables.constEnd(); |
|
256 ++iter) { |
|
257 setValueAt(i, iter.key()->index, iter.value()); |
|
258 } |
|
259 |
|
260 setValueAt(i, columns - 1, c->constant); |
|
261 } |
|
262 |
|
263 // Set objective for the first-phase Simplex. |
|
264 // Z = -1 * sum_of_artificial_vars |
|
265 for (int j = firstArtificial; j < columns - 1; ++j) |
|
266 setValueAt(0, j, 1.0); |
|
267 |
|
268 // Maximize our objective (artificial vars go to zero) |
|
269 solveMaxHelper(); |
|
270 |
|
271 // If there is a solution where the sum of all artificial |
|
272 // variables is zero, then all of them can be removed and yet |
|
273 // we will have a feasible (but not optimal) solution for the |
|
274 // original problem. |
|
275 // Otherwise, we clean up our structures and report there is |
|
276 // no feasible solution. |
|
277 if (valueAt(0, columns - 1) != 0.0) { |
|
278 qWarning() << "QSimplex: No feasible solution!"; |
|
279 clearDataStructures(); |
|
280 return false; |
|
281 } |
|
282 |
|
283 // Remove artificial variables. We already have a feasible |
|
284 // solution for the first problem, thus we don't need them |
|
285 // anymore. |
|
286 clearColumns(firstArtificial, columns - 2); |
|
287 |
|
288 return true; |
|
289 } |
|
290 |
|
291 /*! |
|
292 \internal |
|
293 |
|
294 Run simplex on the current matrix with the current objective. |
|
295 |
|
296 This is the iterative method. The matrix lines are combined |
|
297 as to modify the variable values towards the best solution possible. |
|
298 The method returns when the matrix is in the optimal state. |
|
299 */ |
|
300 void QSimplex::solveMaxHelper() |
|
301 { |
|
302 reducedRowEchelon(); |
|
303 while (iterate()) ; |
|
304 } |
|
305 |
|
306 /*! |
|
307 \internal |
|
308 */ |
|
309 void QSimplex::setObjective(QSimplexConstraint *newObjective) |
|
310 { |
|
311 objective = newObjective; |
|
312 } |
|
313 |
|
314 /*! |
|
315 \internal |
|
316 */ |
|
317 void QSimplex::clearRow(int rowIndex) |
|
318 { |
|
319 qreal *item = matrix + rowIndex * columns; |
|
320 for (int i = 0; i < columns; ++i) |
|
321 item[i] = 0.0; |
|
322 } |
|
323 |
|
324 /*! |
|
325 \internal |
|
326 */ |
|
327 void QSimplex::clearColumns(int first, int last) |
|
328 { |
|
329 for (int i = 0; i < rows; ++i) { |
|
330 qreal *row = matrix + i * columns; |
|
331 for (int j = first; j <= last; ++j) |
|
332 row[j] = 0.0; |
|
333 } |
|
334 } |
|
335 |
|
336 /*! |
|
337 \internal |
|
338 */ |
|
339 void QSimplex::dumpMatrix() |
|
340 { |
|
341 qDebug("---- Simplex Matrix ----\n"); |
|
342 |
|
343 QString str(QLatin1String(" ")); |
|
344 for (int j = 0; j < columns; ++j) |
|
345 str += QString::fromAscii(" <%1 >").arg(j, 2); |
|
346 qDebug("%s", qPrintable(str)); |
|
347 for (int i = 0; i < rows; ++i) { |
|
348 str = QString::fromAscii("Row %1:").arg(i, 2); |
|
349 |
|
350 qreal *row = matrix + i * columns; |
|
351 for (int j = 0; j < columns; ++j) |
|
352 str += QString::fromAscii("%1").arg(row[j], 7, 'f', 2); |
|
353 qDebug("%s", qPrintable(str)); |
|
354 } |
|
355 qDebug("------------------------\n"); |
|
356 } |
|
357 |
|
358 /*! |
|
359 \internal |
|
360 */ |
|
361 void QSimplex::combineRows(int toIndex, int fromIndex, qreal factor) |
|
362 { |
|
363 if (!factor) |
|
364 return; |
|
365 |
|
366 qreal *from = matrix + fromIndex * columns; |
|
367 qreal *to = matrix + toIndex * columns; |
|
368 |
|
369 for (int j = 1; j < columns; ++j) { |
|
370 qreal value = from[j]; |
|
371 |
|
372 // skip to[j] = to[j] + factor*0.0 |
|
373 if (value == 0.0) |
|
374 continue; |
|
375 |
|
376 to[j] += factor * value; |
|
377 |
|
378 // ### Avoid Numerical errors |
|
379 if (qAbs(to[j]) < 0.0000000001) |
|
380 to[j] = 0.0; |
|
381 } |
|
382 } |
|
383 |
|
384 /*! |
|
385 \internal |
|
386 */ |
|
387 int QSimplex::findPivotColumn() |
|
388 { |
|
389 qreal min = 0; |
|
390 int minIndex = -1; |
|
391 |
|
392 for (int j = 0; j < columns-1; ++j) { |
|
393 if (valueAt(0, j) < min) { |
|
394 min = valueAt(0, j); |
|
395 minIndex = j; |
|
396 } |
|
397 } |
|
398 |
|
399 return minIndex; |
|
400 } |
|
401 |
|
402 /*! |
|
403 \internal |
|
404 |
|
405 For a given pivot column, find the pivot row. That is, the row with the |
|
406 minimum associated "quotient" where: |
|
407 |
|
408 - quotient is the division of the value in the last column by the value |
|
409 in the pivot column. |
|
410 - rows with value less or equal to zero are ignored |
|
411 - if two rows have the same quotient, lines are chosen based on the |
|
412 highest variable index (value in the first column) |
|
413 |
|
414 The last condition avoids a bug where artificial variables would be |
|
415 left behind for the second-phase simplex, and with 'good' |
|
416 constraints would be removed before it, what would lead to incorrect |
|
417 results. |
|
418 */ |
|
419 int QSimplex::pivotRowForColumn(int column) |
|
420 { |
|
421 qreal min = qreal(999999999999.0); // ### |
|
422 int minIndex = -1; |
|
423 |
|
424 for (int i = 1; i < rows; ++i) { |
|
425 qreal divisor = valueAt(i, column); |
|
426 if (divisor <= 0) |
|
427 continue; |
|
428 |
|
429 qreal quotient = valueAt(i, columns - 1) / divisor; |
|
430 if (quotient < min) { |
|
431 min = quotient; |
|
432 minIndex = i; |
|
433 } else if ((quotient == min) && (valueAt(i, 0) > valueAt(minIndex, 0))) { |
|
434 minIndex = i; |
|
435 } |
|
436 } |
|
437 |
|
438 return minIndex; |
|
439 } |
|
440 |
|
441 /*! |
|
442 \internal |
|
443 */ |
|
444 void QSimplex::reducedRowEchelon() |
|
445 { |
|
446 for (int i = 1; i < rows; ++i) { |
|
447 int factorInObjectiveRow = valueAt(i, 0); |
|
448 combineRows(0, i, -1 * valueAt(0, factorInObjectiveRow)); |
|
449 } |
|
450 } |
|
451 |
|
452 /*! |
|
453 \internal |
|
454 |
|
455 Does one iteration towards a better solution for the problem. |
|
456 See 'solveMaxHelper'. |
|
457 */ |
|
458 bool QSimplex::iterate() |
|
459 { |
|
460 // Find Pivot column |
|
461 int pivotColumn = findPivotColumn(); |
|
462 if (pivotColumn == -1) |
|
463 return false; |
|
464 |
|
465 // Find Pivot row for column |
|
466 int pivotRow = pivotRowForColumn(pivotColumn); |
|
467 if (pivotRow == -1) { |
|
468 qWarning() << "QSimplex: Unbounded problem!"; |
|
469 return false; |
|
470 } |
|
471 |
|
472 // Normalize Pivot Row |
|
473 qreal pivot = valueAt(pivotRow, pivotColumn); |
|
474 if (pivot != 1.0) |
|
475 combineRows(pivotRow, pivotRow, (1.0 - pivot) / pivot); |
|
476 |
|
477 // Update other rows |
|
478 for (int row=0; row < rows; ++row) { |
|
479 if (row == pivotRow) |
|
480 continue; |
|
481 |
|
482 combineRows(row, pivotRow, -1 * valueAt(row, pivotColumn)); |
|
483 } |
|
484 |
|
485 // Update first column |
|
486 setValueAt(pivotRow, 0, pivotColumn); |
|
487 |
|
488 // dumpMatrix(); |
|
489 // qDebug("------------ end of iteration --------------\n"); |
|
490 return true; |
|
491 } |
|
492 |
|
493 /*! |
|
494 \internal |
|
495 |
|
496 Both solveMin and solveMax are interfaces to this method. |
|
497 |
|
498 The enum solverFactor admits 2 values: Minimum (-1) and Maximum (+1). |
|
499 |
|
500 This method sets the original objective and runs the second phase |
|
501 Simplex to obtain the optimal solution for the problem. As the internal |
|
502 simplex solver is only able to _maximize_ objectives, we handle the |
|
503 minimization case by inverting the original objective and then |
|
504 maximizing it. |
|
505 */ |
|
506 qreal QSimplex::solver(solverFactor factor) |
|
507 { |
|
508 // Remove old objective |
|
509 clearRow(0); |
|
510 |
|
511 // Set new objective |
|
512 QHash<QSimplexVariable *, qreal>::const_iterator iter; |
|
513 for (iter = objective->variables.constBegin(); |
|
514 iter != objective->variables.constEnd(); |
|
515 ++iter) { |
|
516 setValueAt(0, iter.key()->index, -1 * factor * iter.value()); |
|
517 } |
|
518 |
|
519 solveMaxHelper(); |
|
520 collectResults(); |
|
521 |
|
522 #ifdef QT_DEBUG |
|
523 for (int i = 0; i < constraints.size(); ++i) { |
|
524 Q_ASSERT(constraints[i]->isSatisfied()); |
|
525 } |
|
526 #endif |
|
527 |
|
528 return factor * valueAt(0, columns - 1); |
|
529 } |
|
530 |
|
531 /*! |
|
532 \internal |
|
533 Minimize the original objective. |
|
534 */ |
|
535 qreal QSimplex::solveMin() |
|
536 { |
|
537 return solver(Minimum); |
|
538 } |
|
539 |
|
540 /*! |
|
541 \internal |
|
542 Maximize the original objective. |
|
543 */ |
|
544 qreal QSimplex::solveMax() |
|
545 { |
|
546 return solver(Maximum); |
|
547 } |
|
548 |
|
549 /*! |
|
550 \internal |
|
551 |
|
552 Reads results from the simplified matrix and saves them in the |
|
553 "result" member of each QSimplexVariable. |
|
554 */ |
|
555 void QSimplex::collectResults() |
|
556 { |
|
557 // All variables are zero unless overridden below. |
|
558 |
|
559 // ### Is this really needed? Is there any chance that an |
|
560 // important variable remains as non-basic at the end of simplex? |
|
561 for (int i = 0; i < variables.size(); ++i) |
|
562 variables[i]->result = 0; |
|
563 |
|
564 // Basic variables |
|
565 // Update the variable indicated in the first column with the value |
|
566 // in the last column. |
|
567 for (int i = 1; i < rows; ++i) { |
|
568 int index = valueAt(i, 0) - 1; |
|
569 if (index < variables.size()) |
|
570 variables[index]->result = valueAt(i, columns - 1); |
|
571 } |
|
572 } |
|
573 |
|
574 QT_END_NAMESPACE |