1 | // Copyright (C) 2016 The Qt Company Ltd. |
2 | // SPDX-License-Identifier: LicenseRef-Qt-Commercial OR LGPL-3.0-only OR GPL-2.0-only OR GPL-3.0-only |
3 | |
4 | #include "qsimplex_p.h" |
5 | |
6 | #include <QtCore/qset.h> |
7 | #include <QtCore/qdebug.h> |
8 | |
9 | #include <stdlib.h> |
10 | |
11 | QT_BEGIN_NAMESPACE |
12 | |
13 | using namespace Qt::StringLiterals; |
14 | |
15 | /*! |
16 | \internal |
17 | \class QSimplex |
18 | |
19 | The QSimplex class is a Linear Programming problem solver based on the two-phase |
20 | simplex method. |
21 | |
22 | It takes a set of QSimplexConstraints as its restrictive constraints and an |
23 | additional QSimplexConstraint as its objective function. Then methods to maximize |
24 | and minimize the problem solution are provided. |
25 | |
26 | The two-phase simplex method is based on the following steps: |
27 | First phase: |
28 | 1.a) Modify the original, complex, and possibly not feasible problem, into a new, |
29 | easy to solve problem. |
30 | 1.b) Set as the objective of the new problem, a feasible solution for the original |
31 | complex problem. |
32 | 1.c) Run simplex to optimize the modified problem and check whether a solution for |
33 | the original problem exists. |
34 | |
35 | Second phase: |
36 | 2.a) Go back to the original problem with the feasibl (but not optimal) solution |
37 | found in the first phase. |
38 | 2.b) Set the original objective. |
39 | 3.c) Run simplex to optimize the original problem towards its optimal solution. |
40 | */ |
41 | |
42 | /*! |
43 | \internal |
44 | */ |
45 | QSimplex::QSimplex() : objective(nullptr), rows(0), columns(0), firstArtificial(0), matrix(nullptr) |
46 | { |
47 | } |
48 | |
49 | /*! |
50 | \internal |
51 | */ |
52 | QSimplex::~QSimplex() |
53 | { |
54 | clearDataStructures(); |
55 | } |
56 | |
57 | /*! |
58 | \internal |
59 | */ |
60 | void QSimplex::clearDataStructures() |
61 | { |
62 | if (matrix == nullptr) |
63 | return; |
64 | |
65 | // Matrix |
66 | rows = 0; |
67 | columns = 0; |
68 | firstArtificial = 0; |
69 | free(ptr: matrix); |
70 | matrix = nullptr; |
71 | |
72 | // Constraints |
73 | for (int i = 0; i < constraints.size(); ++i) { |
74 | delete constraints[i]->helper.first; |
75 | delete constraints[i]->artificial; |
76 | delete constraints[i]; |
77 | } |
78 | constraints.clear(); |
79 | |
80 | // Other |
81 | variables.clear(); |
82 | objective = nullptr; |
83 | } |
84 | |
85 | /*! |
86 | \internal |
87 | Sets the new constraints in the simplex solver and returns whether the problem |
88 | is feasible. |
89 | |
90 | This method sets the new constraints, normalizes them, creates the simplex matrix |
91 | and runs the first simplex phase. |
92 | */ |
93 | bool QSimplex::setConstraints(const QList<QSimplexConstraint *> &newConstraints) |
94 | { |
95 | //////////////////////////// |
96 | // Reset to initial state // |
97 | //////////////////////////// |
98 | clearDataStructures(); |
99 | |
100 | if (newConstraints.isEmpty()) |
101 | return true; // we are ok with no constraints |
102 | |
103 | // Make deep copy of constraints. We need this copy because we may change |
104 | // them in the simplification method. |
105 | for (int i = 0; i < newConstraints.size(); ++i) { |
106 | QSimplexConstraint *c = new QSimplexConstraint; |
107 | c->constant = newConstraints[i]->constant; |
108 | c->ratio = newConstraints[i]->ratio; |
109 | c->variables = newConstraints[i]->variables; |
110 | constraints << c; |
111 | } |
112 | |
113 | // Remove constraints of type Var == K and replace them for their value. |
114 | if (!simplifyConstraints(constraints: &constraints)) { |
115 | qWarning(msg: "QSimplex: No feasible solution!" ); |
116 | clearDataStructures(); |
117 | return false; |
118 | } |
119 | |
120 | /////////////////////////////////////// |
121 | // Prepare variables and constraints // |
122 | /////////////////////////////////////// |
123 | |
124 | // Set Variables direct mapping. |
125 | // "variables" is a list that provides a stable, indexed list of all variables |
126 | // used in this problem. |
127 | QSet<QSimplexVariable *> variablesSet; |
128 | for (int i = 0; i < constraints.size(); ++i) { |
129 | const auto &v = constraints.at(i)->variables; |
130 | for (auto it = v.cbegin(), end = v.cend(); it != end; ++it) |
131 | variablesSet.insert(value: it.key()); |
132 | } |
133 | variables = variablesSet.values(); |
134 | |
135 | // Set Variables reverse mapping |
136 | // We also need to be able to find the index for a given variable, to do that |
137 | // we store in each variable its index. |
138 | for (int i = 0; i < variables.size(); ++i) { |
139 | // The variable "0" goes at the column "1", etc... |
140 | variables[i]->index = i + 1; |
141 | } |
142 | |
143 | // Normalize Constraints |
144 | // In this step, we prepare the constraints in two ways: |
145 | // Firstly, we modify all constraints of type "LessOrEqual" or "MoreOrEqual" |
146 | // by the adding slack or surplus variables and making them "Equal" constraints. |
147 | // Secondly, we need every single constraint to have a direct, easy feasible |
148 | // solution. Constraints that have slack variables are already easy to solve, |
149 | // to all the others we add artificial variables. |
150 | // |
151 | // At the end we modify the constraints as follows: |
152 | // - LessOrEqual: SLACK variable is added. |
153 | // - Equal: ARTIFICIAL variable is added. |
154 | // - More or Equal: ARTIFICIAL and SURPLUS variables are added. |
155 | int variableIndex = variables.size(); |
156 | QList <QSimplexVariable *> artificialList; |
157 | |
158 | for (int i = 0; i < constraints.size(); ++i) { |
159 | QSimplexVariable *slack; |
160 | QSimplexVariable *surplus; |
161 | QSimplexVariable *artificial; |
162 | |
163 | Q_ASSERT(constraints[i]->helper.first == 0); |
164 | Q_ASSERT(constraints[i]->artificial == nullptr); |
165 | |
166 | switch(constraints[i]->ratio) { |
167 | case QSimplexConstraint::LessOrEqual: |
168 | slack = new QSimplexVariable; |
169 | slack->index = ++variableIndex; |
170 | constraints[i]->helper.first = slack; |
171 | constraints[i]->helper.second = 1.0; |
172 | break; |
173 | case QSimplexConstraint::MoreOrEqual: |
174 | surplus = new QSimplexVariable; |
175 | surplus->index = ++variableIndex; |
176 | constraints[i]->helper.first = surplus; |
177 | constraints[i]->helper.second = -1.0; |
178 | Q_FALLTHROUGH(); |
179 | case QSimplexConstraint::Equal: |
180 | artificial = new QSimplexVariable; |
181 | constraints[i]->artificial = artificial; |
182 | artificialList += constraints[i]->artificial; |
183 | break; |
184 | } |
185 | } |
186 | |
187 | // All original, slack and surplus have already had its index set |
188 | // at this point. We now set the index of the artificial variables |
189 | // as to ensure they are at the end of the variable list and therefore |
190 | // can be easily removed at the end of this method. |
191 | firstArtificial = variableIndex + 1; |
192 | for (int i = 0; i < artificialList.size(); ++i) |
193 | artificialList[i]->index = ++variableIndex; |
194 | artificialList.clear(); |
195 | |
196 | ///////////////////////////// |
197 | // Fill the Simplex matrix // |
198 | ///////////////////////////// |
199 | |
200 | // One for each variable plus the Basic and BFS columns (first and last) |
201 | columns = variableIndex + 2; |
202 | // One for each constraint plus the objective function |
203 | rows = constraints.size() + 1; |
204 | |
205 | matrix = (qreal *)malloc(size: sizeof(qreal) * columns * rows); |
206 | if (!matrix) { |
207 | qWarning(msg: "QSimplex: Unable to allocate memory!" ); |
208 | return false; |
209 | } |
210 | for (int i = columns * rows - 1; i >= 0; --i) |
211 | matrix[i] = 0.0; |
212 | |
213 | // Fill Matrix |
214 | for (int i = 1; i <= constraints.size(); ++i) { |
215 | QSimplexConstraint *c = constraints[i - 1]; |
216 | |
217 | if (c->artificial) { |
218 | // Will use artificial basic variable |
219 | setValueAt(rowIndex: i, columnIndex: 0, value: c->artificial->index); |
220 | setValueAt(rowIndex: i, columnIndex: c->artificial->index, value: 1.0); |
221 | |
222 | if (c->helper.second != 0.0) { |
223 | // Surplus variable |
224 | setValueAt(rowIndex: i, columnIndex: c->helper.first->index, value: c->helper.second); |
225 | } |
226 | } else { |
227 | // Slack is used as the basic variable |
228 | Q_ASSERT(c->helper.second == 1.0); |
229 | setValueAt(rowIndex: i, columnIndex: 0, value: c->helper.first->index); |
230 | setValueAt(rowIndex: i, columnIndex: c->helper.first->index, value: 1.0); |
231 | } |
232 | |
233 | QHash<QSimplexVariable *, qreal>::const_iterator iter; |
234 | for (iter = c->variables.constBegin(); |
235 | iter != c->variables.constEnd(); |
236 | ++iter) { |
237 | setValueAt(rowIndex: i, columnIndex: iter.key()->index, value: iter.value()); |
238 | } |
239 | |
240 | setValueAt(rowIndex: i, columnIndex: columns - 1, value: c->constant); |
241 | } |
242 | |
243 | // Set objective for the first-phase Simplex. |
244 | // Z = -1 * sum_of_artificial_vars |
245 | for (int j = firstArtificial; j < columns - 1; ++j) |
246 | setValueAt(rowIndex: 0, columnIndex: j, value: 1.0); |
247 | |
248 | // Maximize our objective (artificial vars go to zero) |
249 | solveMaxHelper(); |
250 | |
251 | // If there is a solution where the sum of all artificial |
252 | // variables is zero, then all of them can be removed and yet |
253 | // we will have a feasible (but not optimal) solution for the |
254 | // original problem. |
255 | // Otherwise, we clean up our structures and report there is |
256 | // no feasible solution. |
257 | if ((valueAt(rowIndex: 0, columnIndex: columns - 1) != 0.0) && (qAbs(t: valueAt(rowIndex: 0, columnIndex: columns - 1)) > 0.00001)) { |
258 | qWarning(msg: "QSimplex: No feasible solution!" ); |
259 | clearDataStructures(); |
260 | return false; |
261 | } |
262 | |
263 | // Remove artificial variables. We already have a feasible |
264 | // solution for the first problem, thus we don't need them |
265 | // anymore. |
266 | clearColumns(first: firstArtificial, last: columns - 2); |
267 | |
268 | return true; |
269 | } |
270 | |
271 | /*! |
272 | \internal |
273 | |
274 | Run simplex on the current matrix with the current objective. |
275 | |
276 | This is the iterative method. The matrix lines are combined |
277 | as to modify the variable values towards the best solution possible. |
278 | The method returns when the matrix is in the optimal state. |
279 | */ |
280 | void QSimplex::solveMaxHelper() |
281 | { |
282 | reducedRowEchelon(); |
283 | while (iterate()) ; |
284 | } |
285 | |
286 | /*! |
287 | \internal |
288 | */ |
289 | void QSimplex::setObjective(QSimplexConstraint *newObjective) |
290 | { |
291 | objective = newObjective; |
292 | } |
293 | |
294 | /*! |
295 | \internal |
296 | */ |
297 | void QSimplex::clearRow(int rowIndex) |
298 | { |
299 | qreal *item = matrix + rowIndex * columns; |
300 | for (int i = 0; i < columns; ++i) |
301 | item[i] = 0.0; |
302 | } |
303 | |
304 | /*! |
305 | \internal |
306 | */ |
307 | void QSimplex::clearColumns(int first, int last) |
308 | { |
309 | for (int i = 0; i < rows; ++i) { |
310 | qreal *row = matrix + i * columns; |
311 | for (int j = first; j <= last; ++j) |
312 | row[j] = 0.0; |
313 | } |
314 | } |
315 | |
316 | /*! |
317 | \internal |
318 | */ |
319 | void QSimplex::dumpMatrix() |
320 | { |
321 | qDebug(msg: "---- Simplex Matrix ----\n" ); |
322 | |
323 | QString str(" "_L1 ); |
324 | for (int j = 0; j < columns; ++j) |
325 | str += QString::fromLatin1(ba: " <%1 >" ).arg(a: j, fieldWidth: 2); |
326 | qDebug(msg: "%s" , qPrintable(str)); |
327 | for (int i = 0; i < rows; ++i) { |
328 | str = QString::fromLatin1(ba: "Row %1:" ).arg(a: i, fieldWidth: 2); |
329 | |
330 | qreal *row = matrix + i * columns; |
331 | for (int j = 0; j < columns; ++j) |
332 | str += QString::fromLatin1(ba: "%1" ).arg(a: row[j], fieldWidth: 7, format: 'f', precision: 2); |
333 | qDebug(msg: "%s" , qPrintable(str)); |
334 | } |
335 | qDebug(msg: "------------------------\n" ); |
336 | } |
337 | |
338 | /*! |
339 | \internal |
340 | */ |
341 | void QSimplex::combineRows(int toIndex, int fromIndex, qreal factor) |
342 | { |
343 | if (!factor) |
344 | return; |
345 | |
346 | qreal *from = matrix + fromIndex * columns; |
347 | qreal *to = matrix + toIndex * columns; |
348 | |
349 | for (int j = 1; j < columns; ++j) { |
350 | qreal value = from[j]; |
351 | |
352 | // skip to[j] = to[j] + factor*0.0 |
353 | if (value == 0.0) |
354 | continue; |
355 | |
356 | to[j] += factor * value; |
357 | |
358 | // ### Avoid Numerical errors |
359 | if (qAbs(t: to[j]) < 0.0000000001) |
360 | to[j] = 0.0; |
361 | } |
362 | } |
363 | |
364 | /*! |
365 | \internal |
366 | */ |
367 | int QSimplex::findPivotColumn() |
368 | { |
369 | qreal min = 0; |
370 | int minIndex = -1; |
371 | |
372 | for (int j = 0; j < columns-1; ++j) { |
373 | if (valueAt(rowIndex: 0, columnIndex: j) < min) { |
374 | min = valueAt(rowIndex: 0, columnIndex: j); |
375 | minIndex = j; |
376 | } |
377 | } |
378 | |
379 | return minIndex; |
380 | } |
381 | |
382 | /*! |
383 | \internal |
384 | |
385 | For a given pivot column, find the pivot row. That is, the row with the |
386 | minimum associated "quotient" where: |
387 | |
388 | - quotient is the division of the value in the last column by the value |
389 | in the pivot column. |
390 | - rows with value less or equal to zero are ignored |
391 | - if two rows have the same quotient, lines are chosen based on the |
392 | highest variable index (value in the first column) |
393 | |
394 | The last condition avoids a bug where artificial variables would be |
395 | left behind for the second-phase simplex, and with 'good' |
396 | constraints would be removed before it, what would lead to incorrect |
397 | results. |
398 | */ |
399 | int QSimplex::pivotRowForColumn(int column) |
400 | { |
401 | qreal min = qreal(999999999999.0); // ### |
402 | int minIndex = -1; |
403 | |
404 | for (int i = 1; i < rows; ++i) { |
405 | qreal divisor = valueAt(rowIndex: i, columnIndex: column); |
406 | if (divisor <= 0) |
407 | continue; |
408 | |
409 | qreal quotient = valueAt(rowIndex: i, columnIndex: columns - 1) / divisor; |
410 | if (quotient < min) { |
411 | min = quotient; |
412 | minIndex = i; |
413 | } else if ((quotient == min) && (valueAt(rowIndex: i, columnIndex: 0) > valueAt(rowIndex: minIndex, columnIndex: 0))) { |
414 | minIndex = i; |
415 | } |
416 | } |
417 | |
418 | return minIndex; |
419 | } |
420 | |
421 | /*! |
422 | \internal |
423 | */ |
424 | void QSimplex::reducedRowEchelon() |
425 | { |
426 | for (int i = 1; i < rows; ++i) { |
427 | int factorInObjectiveRow = valueAt(rowIndex: i, columnIndex: 0); |
428 | combineRows(toIndex: 0, fromIndex: i, factor: -1 * valueAt(rowIndex: 0, columnIndex: factorInObjectiveRow)); |
429 | } |
430 | } |
431 | |
432 | /*! |
433 | \internal |
434 | |
435 | Does one iteration towards a better solution for the problem. |
436 | See 'solveMaxHelper'. |
437 | */ |
438 | bool QSimplex::iterate() |
439 | { |
440 | // Find Pivot column |
441 | int pivotColumn = findPivotColumn(); |
442 | if (pivotColumn == -1) |
443 | return false; |
444 | |
445 | // Find Pivot row for column |
446 | int pivotRow = pivotRowForColumn(column: pivotColumn); |
447 | if (pivotRow == -1) { |
448 | qWarning(msg: "QSimplex: Unbounded problem!" ); |
449 | return false; |
450 | } |
451 | |
452 | // Normalize Pivot Row |
453 | qreal pivot = valueAt(rowIndex: pivotRow, columnIndex: pivotColumn); |
454 | if (pivot != 1.0) |
455 | combineRows(toIndex: pivotRow, fromIndex: pivotRow, factor: (1.0 - pivot) / pivot); |
456 | |
457 | // Update other rows |
458 | for (int row=0; row < rows; ++row) { |
459 | if (row == pivotRow) |
460 | continue; |
461 | |
462 | combineRows(toIndex: row, fromIndex: pivotRow, factor: -1 * valueAt(rowIndex: row, columnIndex: pivotColumn)); |
463 | } |
464 | |
465 | // Update first column |
466 | setValueAt(rowIndex: pivotRow, columnIndex: 0, value: pivotColumn); |
467 | |
468 | // dumpMatrix(); |
469 | // qDebug("------------ end of iteration --------------\n"); |
470 | return true; |
471 | } |
472 | |
473 | /*! |
474 | \internal |
475 | |
476 | Both solveMin and solveMax are interfaces to this method. |
477 | |
478 | The enum SolverFactor admits 2 values: Minimum (-1) and Maximum (+1). |
479 | |
480 | This method sets the original objective and runs the second phase |
481 | Simplex to obtain the optimal solution for the problem. As the internal |
482 | simplex solver is only able to _maximize_ objectives, we handle the |
483 | minimization case by inverting the original objective and then |
484 | maximizing it. |
485 | */ |
486 | qreal QSimplex::solver(SolverFactor factor) |
487 | { |
488 | // Remove old objective |
489 | clearRow(rowIndex: 0); |
490 | |
491 | // Set new objective in the first row of the simplex matrix |
492 | qreal resultOffset = 0; |
493 | QHash<QSimplexVariable *, qreal>::const_iterator iter; |
494 | for (iter = objective->variables.constBegin(); |
495 | iter != objective->variables.constEnd(); |
496 | ++iter) { |
497 | |
498 | // Check if the variable was removed in the simplification process. |
499 | // If so, we save its offset to the objective function and skip adding |
500 | // it to the matrix. |
501 | if (iter.key()->index == -1) { |
502 | resultOffset += iter.value() * iter.key()->result; |
503 | continue; |
504 | } |
505 | |
506 | setValueAt(rowIndex: 0, columnIndex: iter.key()->index, value: -1 * factor * iter.value()); |
507 | } |
508 | |
509 | solveMaxHelper(); |
510 | collectResults(); |
511 | |
512 | #ifdef QT_DEBUG |
513 | for (int i = 0; i < constraints.size(); ++i) { |
514 | Q_ASSERT(constraints[i]->isSatisfied()); |
515 | } |
516 | #endif |
517 | |
518 | // Return the value calculated by the simplex plus the value of the |
519 | // fixed variables. |
520 | return (qToUnderlying(e: factor) * valueAt(rowIndex: 0, columnIndex: columns - 1)) + resultOffset; |
521 | } |
522 | |
523 | /*! |
524 | \internal |
525 | Minimize the original objective. |
526 | */ |
527 | qreal QSimplex::solveMin() |
528 | { |
529 | return solver(factor: Minimum); |
530 | } |
531 | |
532 | /*! |
533 | \internal |
534 | Maximize the original objective. |
535 | */ |
536 | qreal QSimplex::solveMax() |
537 | { |
538 | return solver(factor: Maximum); |
539 | } |
540 | |
541 | /*! |
542 | \internal |
543 | |
544 | Reads results from the simplified matrix and saves them in the |
545 | "result" member of each QSimplexVariable. |
546 | */ |
547 | void QSimplex::collectResults() |
548 | { |
549 | // All variables are zero unless overridden below. |
550 | |
551 | // ### Is this really needed? Is there any chance that an |
552 | // important variable remains as non-basic at the end of simplex? |
553 | for (int i = 0; i < variables.size(); ++i) |
554 | variables[i]->result = 0; |
555 | |
556 | // Basic variables |
557 | // Update the variable indicated in the first column with the value |
558 | // in the last column. |
559 | for (int i = 1; i < rows; ++i) { |
560 | int index = valueAt(rowIndex: i, columnIndex: 0) - 1; |
561 | if (index < variables.size()) |
562 | variables[index]->result = valueAt(rowIndex: i, columnIndex: columns - 1); |
563 | } |
564 | } |
565 | |
566 | /*! |
567 | \internal |
568 | |
569 | Looks for single-valued variables and remove them from the constraints list. |
570 | */ |
571 | bool QSimplex::simplifyConstraints(QList<QSimplexConstraint *> *constraints) |
572 | { |
573 | QHash<QSimplexVariable *, qreal> results; // List of single-valued variables |
574 | bool modified = true; // Any chance more optimization exists? |
575 | |
576 | while (modified) { |
577 | modified = false; |
578 | |
579 | // For all constraints |
580 | QList<QSimplexConstraint *>::iterator iter = constraints->begin(); |
581 | while (iter != constraints->end()) { |
582 | QSimplexConstraint *c = *iter; |
583 | if ((c->ratio == QSimplexConstraint::Equal) && (c->variables.size() == 1)) { |
584 | // Check whether this is a constraint of type Var == K |
585 | // If so, save its value to "results". |
586 | QSimplexVariable *variable = c->variables.constBegin().key(); |
587 | qreal result = c->constant / c->variables.value(key: variable); |
588 | |
589 | results.insert(key: variable, value: result); |
590 | variable->result = result; |
591 | variable->index = -1; |
592 | modified = true; |
593 | |
594 | } |
595 | |
596 | // Replace known values among their variables |
597 | QHash<QSimplexVariable *, qreal>::const_iterator r; |
598 | for (r = results.constBegin(); r != results.constEnd(); ++r) { |
599 | if (c->variables.contains(key: r.key())) { |
600 | c->constant -= r.value() * c->variables.take(key: r.key()); |
601 | modified = true; |
602 | } |
603 | } |
604 | |
605 | // Keep it normalized |
606 | if (c->constant < 0) |
607 | c->invert(); |
608 | |
609 | if (c->variables.isEmpty()) { |
610 | // If constraint became empty due to substitution, delete it. |
611 | if (c->isSatisfied() == false) |
612 | // We must ensure that the constraint soon to be deleted would not |
613 | // make the problem unfeasible if left behind. If that's the case, |
614 | // we return false so the simplex solver can properly report that. |
615 | return false; |
616 | |
617 | delete c; |
618 | iter = constraints->erase(pos: iter); |
619 | } else { |
620 | ++iter; |
621 | } |
622 | } |
623 | } |
624 | |
625 | return true; |
626 | } |
627 | |
628 | void QSimplexConstraint::invert() |
629 | { |
630 | constant = -constant; |
631 | ratio = Ratio(2 - ratio); |
632 | |
633 | QHash<QSimplexVariable *, qreal>::iterator iter; |
634 | for (iter = variables.begin(); iter != variables.end(); ++iter) { |
635 | iter.value() = -iter.value(); |
636 | } |
637 | } |
638 | |
639 | QT_END_NAMESPACE |
640 | |