# HG changeset patch
# User Peter Kovacs <kpeter@inf.elte.hu>
# Date 1249582363 -7200
# Node ID 83ce7ce39f21e45540ad13118842fefec217cfe1
# Parent d66ff32624e2cec102cedea55dca5e9f67ce1931
Rework and fix the implementation of MinMeanCycle (#179)
- Fix the handling of the cycle means.
- Many implementation improvements:
- More efficient data storage for the strongly connected
components.
- Better handling of BFS queues.
- Merge consecutive BFS searches (perform two BFS searches
instead of three).
This version is about two times faster on average and an order of
magnitude faster if there are a lot of strongly connected components.
diff --git a/lemon/min_mean_cycle.h b/lemon/min_mean_cycle.h
a
|
b
|
|
74 | 74 | // The length of the arcs |
75 | 75 | const LengthMap &_length; |
76 | 76 | |
77 | | // The total length of the found cycle |
78 | | Value _cycle_length; |
79 | | // The number of arcs on the found cycle |
80 | | int _cycle_size; |
81 | | // The found cycle |
| 77 | // Data for the found cycles |
| 78 | bool _curr_found, _best_found; |
| 79 | Value _curr_length, _best_length; |
| 80 | int _curr_size, _best_size; |
| 81 | Node _curr_node, _best_node; |
| 82 | |
82 | 83 | Path *_cycle_path; |
| 84 | bool _local_path; |
83 | 85 | |
84 | | bool _local_path; |
85 | | bool _cycle_found; |
86 | | Node _cycle_node; |
| 86 | // Internal data used by the algorithm |
| 87 | typename Digraph::template NodeMap<Arc> _policy; |
| 88 | typename Digraph::template NodeMap<bool> _reached; |
| 89 | typename Digraph::template NodeMap<int> _level; |
| 90 | typename Digraph::template NodeMap<double> _dist; |
87 | 91 | |
88 | | typename Digraph::template NodeMap<bool> _reached; |
89 | | typename Digraph::template NodeMap<double> _dist; |
90 | | typename Digraph::template NodeMap<Arc> _policy; |
91 | | |
| 92 | // Data for storing the strongly connected components |
| 93 | int _comp_num; |
92 | 94 | typename Digraph::template NodeMap<int> _comp; |
93 | | int _comp_num; |
94 | | |
95 | | std::vector<Node> _nodes; |
96 | | std::vector<Arc> _arcs; |
| 95 | std::vector<std::vector<Node> > _comp_nodes; |
| 96 | std::vector<Node>* _nodes; |
| 97 | typename Digraph::template NodeMap<std::vector<Arc> > _in_arcs; |
| 98 | |
| 99 | // Queue used for BFS search |
| 100 | std::vector<Node> _queue; |
| 101 | int _qfront, _qback; |
| 102 | |
97 | 103 | Tolerance<double> _tol; |
98 | 104 | |
99 | 105 | public: |
… |
… |
|
106 | 112 | /// \param length The lengths (costs) of the arcs. |
107 | 113 | MinMeanCycle( const Digraph &digraph, |
108 | 114 | const LengthMap &length ) : |
109 | | _gr(digraph), _length(length), _cycle_length(0), _cycle_size(-1), |
110 | | _cycle_path(NULL), _local_path(false), _reached(digraph), |
111 | | _dist(digraph), _policy(digraph), _comp(digraph) |
| 115 | _gr(digraph), _length(length), _cycle_path(NULL), _local_path(false), |
| 116 | _policy(digraph), _reached(digraph), _level(digraph), _dist(digraph), |
| 117 | _comp(digraph), _in_arcs(digraph) |
112 | 118 | {} |
113 | 119 | |
114 | 120 | /// Destructor. |
… |
… |
|
172 | 178 | /// |
173 | 179 | /// \return \c true if a directed cycle exists in the digraph. |
174 | 180 | bool findMinMean() { |
175 | | // Initialize |
176 | | _tol.epsilon(1e-6); |
177 | | if (!_cycle_path) { |
178 | | _local_path = true; |
179 | | _cycle_path = new Path; |
180 | | } |
181 | | _cycle_path->clear(); |
182 | | _cycle_found = false; |
183 | | |
| 181 | // Initialize and find strongly connected components |
| 182 | init(); |
| 183 | findComponents(); |
| 184 | |
184 | 185 | // Find the minimum cycle mean in the components |
185 | | _comp_num = stronglyConnectedComponents(_gr, _comp); |
186 | 186 | for (int comp = 0; comp < _comp_num; ++comp) { |
187 | | if (!initCurrentComponent(comp)) continue; |
| 187 | // Find the minimum mean cycle in the current component |
| 188 | if (!buildPolicyGraph(comp)) continue; |
188 | 189 | while (true) { |
189 | | if (!findPolicyCycles()) break; |
190 | | contractPolicyGraph(comp); |
| 190 | findPolicyCycle(); |
191 | 191 | if (!computeNodeDistances()) break; |
192 | 192 | } |
| 193 | // Update the best cycle (global minimum mean cycle) |
| 194 | if ( !_best_found || (_curr_found && |
| 195 | _curr_length * _best_size < _best_length * _curr_size) ) { |
| 196 | _best_found = true; |
| 197 | _best_length = _curr_length; |
| 198 | _best_size = _curr_size; |
| 199 | _best_node = _curr_node; |
| 200 | } |
193 | 201 | } |
194 | | return _cycle_found; |
| 202 | return _best_found; |
195 | 203 | } |
196 | 204 | |
197 | 205 | /// \brief Find a minimum mean directed cycle. |
… |
… |
|
203 | 211 | /// |
204 | 212 | /// \pre \ref findMinMean() must be called before using this function. |
205 | 213 | bool findCycle() { |
206 | | if (!_cycle_found) return false; |
207 | | _cycle_path->addBack(_policy[_cycle_node]); |
208 | | for ( Node v = _cycle_node; |
209 | | (v = _gr.target(_policy[v])) != _cycle_node; ) { |
| 214 | if (!_best_found) return false; |
| 215 | _cycle_path->addBack(_policy[_best_node]); |
| 216 | for ( Node v = _best_node; |
| 217 | (v = _gr.target(_policy[v])) != _best_node; ) { |
210 | 218 | _cycle_path->addBack(_policy[v]); |
211 | 219 | } |
212 | 220 | return true; |
… |
… |
|
225 | 233 | /// |
226 | 234 | /// This function returns the total length of the found cycle. |
227 | 235 | /// |
228 | | /// \pre \ref run() or \ref findCycle() must be called before |
| 236 | /// \pre \ref run() or \ref findMinMean() must be called before |
229 | 237 | /// using this function. |
230 | 238 | Value cycleLength() const { |
231 | | return _cycle_length; |
| 239 | return _best_length; |
232 | 240 | } |
233 | 241 | |
234 | 242 | /// \brief Return the number of arcs on the found cycle. |
235 | 243 | /// |
236 | 244 | /// This function returns the number of arcs on the found cycle. |
237 | 245 | /// |
238 | | /// \pre \ref run() or \ref findCycle() must be called before |
| 246 | /// \pre \ref run() or \ref findMinMean() must be called before |
239 | 247 | /// using this function. |
240 | 248 | int cycleArcNum() const { |
241 | | return _cycle_size; |
| 249 | return _best_size; |
242 | 250 | } |
243 | 251 | |
244 | 252 | /// \brief Return the mean length of the found cycle. |
245 | 253 | /// |
246 | 254 | /// This function returns the mean length of the found cycle. |
247 | 255 | /// |
248 | | /// \note <tt>mmc.cycleMean()</tt> is just a shortcut of the |
| 256 | /// \note <tt>alg.cycleMean()</tt> is just a shortcut of the |
249 | 257 | /// following code. |
250 | 258 | /// \code |
251 | | /// return double(mmc.cycleLength()) / mmc.cycleArcNum(); |
| 259 | /// return static_cast<double>(alg.cycleLength()) / alg.cycleArcNum(); |
252 | 260 | /// \endcode |
253 | 261 | /// |
254 | 262 | /// \pre \ref run() or \ref findMinMean() must be called before |
255 | 263 | /// using this function. |
256 | 264 | double cycleMean() const { |
257 | | return double(_cycle_length) / _cycle_size; |
| 265 | return static_cast<double>(_best_length) / _best_size; |
258 | 266 | } |
259 | 267 | |
260 | 268 | /// \brief Return the found cycle. |
… |
… |
|
274 | 282 | |
275 | 283 | private: |
276 | 284 | |
277 | | // Initialize the internal data structures for the current strongly |
278 | | // connected component and create the policy graph. |
279 | | // The policy graph can be represented by the _policy map because |
280 | | // the out-degree of every node is 1. |
281 | | bool initCurrentComponent(int comp) { |
282 | | // Find the nodes of the current component |
283 | | _nodes.clear(); |
284 | | for (NodeIt n(_gr); n != INVALID; ++n) { |
285 | | if (_comp[n] == comp) _nodes.push_back(n); |
| 285 | // Initialize |
| 286 | void init() { |
| 287 | _tol.epsilon(1e-6); |
| 288 | if (!_cycle_path) { |
| 289 | _local_path = true; |
| 290 | _cycle_path = new Path; |
286 | 291 | } |
287 | | if (_nodes.size() <= 1) return false; |
288 | | // Find the arcs of the current component |
289 | | _arcs.clear(); |
290 | | for (ArcIt e(_gr); e != INVALID; ++e) { |
291 | | if ( _comp[_gr.source(e)] == comp && |
292 | | _comp[_gr.target(e)] == comp ) |
293 | | _arcs.push_back(e); |
| 292 | _queue.resize(countNodes(_gr)); |
| 293 | _best_found = false; |
| 294 | _best_length = 0; |
| 295 | _best_size = 1; |
| 296 | _cycle_path->clear(); |
| 297 | } |
| 298 | |
| 299 | // Find strongly connected components and initialize _comp_nodes |
| 300 | // and _in_arcs |
| 301 | void findComponents() { |
| 302 | _comp_num = stronglyConnectedComponents(_gr, _comp); |
| 303 | _comp_nodes.resize(_comp_num); |
| 304 | if (_comp_num == 1) { |
| 305 | _comp_nodes[0].clear(); |
| 306 | for (NodeIt n(_gr); n != INVALID; ++n) { |
| 307 | _comp_nodes[0].push_back(n); |
| 308 | _in_arcs[n].clear(); |
| 309 | for (InArcIt a(_gr, n); a != INVALID; ++a) { |
| 310 | _in_arcs[n].push_back(a); |
| 311 | } |
| 312 | } |
| 313 | } else { |
| 314 | for (int i = 0; i < _comp_num; ++i) |
| 315 | _comp_nodes[i].clear(); |
| 316 | for (NodeIt n(_gr); n != INVALID; ++n) { |
| 317 | int k = _comp[n]; |
| 318 | _comp_nodes[k].push_back(n); |
| 319 | _in_arcs[n].clear(); |
| 320 | for (InArcIt a(_gr, n); a != INVALID; ++a) { |
| 321 | if (_comp[_gr.source(a)] == k) _in_arcs[n].push_back(a); |
| 322 | } |
| 323 | } |
294 | 324 | } |
295 | | // Initialize _reached, _dist, _policy maps |
296 | | for (int i = 0; i < int(_nodes.size()); ++i) { |
297 | | _reached[_nodes[i]] = false; |
298 | | _policy[_nodes[i]] = INVALID; |
| 325 | } |
| 326 | |
| 327 | // Build the policy graph in the given strongly connected component |
| 328 | // (the out-degree of every node is 1) |
| 329 | bool buildPolicyGraph(int comp) { |
| 330 | _nodes = &(_comp_nodes[comp]); |
| 331 | if (_nodes->size() < 1 || |
| 332 | (_nodes->size() == 1 && _in_arcs[(*_nodes)[0]].size() == 0)) { |
| 333 | return false; |
299 | 334 | } |
300 | | Node u; Arc e; |
301 | | for (int j = 0; j < int(_arcs.size()); ++j) { |
302 | | e = _arcs[j]; |
303 | | u = _gr.source(e); |
304 | | if (!_reached[u] || _length[e] < _dist[u]) { |
305 | | _dist[u] = _length[e]; |
306 | | _policy[u] = e; |
307 | | _reached[u] = true; |
| 335 | for (int i = 0; i < int(_nodes->size()); ++i) { |
| 336 | _dist[(*_nodes)[i]] = std::numeric_limits<double>::max(); |
| 337 | } |
| 338 | Node u, v; |
| 339 | Arc e; |
| 340 | for (int i = 0; i < int(_nodes->size()); ++i) { |
| 341 | v = (*_nodes)[i]; |
| 342 | for (int j = 0; j < int(_in_arcs[v].size()); ++j) { |
| 343 | e = _in_arcs[v][j]; |
| 344 | u = _gr.source(e); |
| 345 | if (_length[e] < _dist[u]) { |
| 346 | _dist[u] = _length[e]; |
| 347 | _policy[u] = e; |
| 348 | } |
308 | 349 | } |
309 | 350 | } |
310 | 351 | return true; |
311 | 352 | } |
312 | 353 | |
313 | | // Find all cycles in the policy graph. |
314 | | // Set _cycle_found to true if a cycle is found and set |
315 | | // _cycle_length, _cycle_size, _cycle_node to represent the minimum |
316 | | // mean cycle in the policy graph. |
317 | | bool findPolicyCycles() { |
318 | | typename Digraph::template NodeMap<int> level(_gr, -1); |
319 | | bool curr_cycle_found = false; |
| 354 | // Find the minimum mean cycle in the policy graph |
| 355 | void findPolicyCycle() { |
| 356 | for (int i = 0; i < int(_nodes->size()); ++i) { |
| 357 | _level[(*_nodes)[i]] = -1; |
| 358 | } |
320 | 359 | Value clength; |
321 | 360 | int csize; |
322 | | int path_cnt = 0; |
323 | 361 | Node u, v; |
324 | | // Searching for cycles |
325 | | for (int i = 0; i < int(_nodes.size()); ++i) { |
326 | | if (level[_nodes[i]] < 0) { |
327 | | u = _nodes[i]; |
328 | | level[u] = path_cnt; |
329 | | while (level[u = _gr.target(_policy[u])] < 0) |
330 | | level[u] = path_cnt; |
331 | | if (level[u] == path_cnt) { |
332 | | // A cycle is found |
333 | | curr_cycle_found = true; |
334 | | clength = _length[_policy[u]]; |
335 | | csize = 1; |
336 | | for (v = u; (v = _gr.target(_policy[v])) != u; ) { |
337 | | clength += _length[_policy[v]]; |
338 | | ++csize; |
339 | | } |
340 | | if ( !_cycle_found || |
341 | | clength * _cycle_size < _cycle_length * csize ) { |
342 | | _cycle_found = true; |
343 | | _cycle_length = clength; |
344 | | _cycle_size = csize; |
345 | | _cycle_node = u; |
346 | | } |
| 362 | _curr_found = false; |
| 363 | for (int i = 0; i < int(_nodes->size()); ++i) { |
| 364 | u = (*_nodes)[i]; |
| 365 | if (_level[u] >= 0) continue; |
| 366 | for (; _level[u] < 0; u = _gr.target(_policy[u])) { |
| 367 | _level[u] = i; |
| 368 | } |
| 369 | if (_level[u] == i) { |
| 370 | // A cycle is found |
| 371 | clength = _length[_policy[u]]; |
| 372 | csize = 1; |
| 373 | for (v = u; (v = _gr.target(_policy[v])) != u; ) { |
| 374 | clength += _length[_policy[v]]; |
| 375 | ++csize; |
347 | 376 | } |
348 | | ++path_cnt; |
349 | | } |
350 | | } |
351 | | return curr_cycle_found; |
352 | | } |
353 | | |
354 | | // Contract the policy graph to be connected by cutting all cycles |
355 | | // except for the main cycle (i.e. the minimum mean cycle). |
356 | | void contractPolicyGraph(int comp) { |
357 | | // Find the component of the main cycle using reverse BFS search |
358 | | typename Digraph::template NodeMap<int> found(_gr, false); |
359 | | std::deque<Node> queue; |
360 | | queue.push_back(_cycle_node); |
361 | | found[_cycle_node] = true; |
362 | | Node u, v; |
363 | | while (!queue.empty()) { |
364 | | v = queue.front(); queue.pop_front(); |
365 | | for (InArcIt e(_gr, v); e != INVALID; ++e) { |
366 | | u = _gr.source(e); |
367 | | if (_policy[u] == e && !found[u]) { |
368 | | found[u] = true; |
369 | | queue.push_back(u); |
370 | | } |
371 | | } |
372 | | } |
373 | | // Connect all other nodes to this component using reverse BFS search |
374 | | queue.clear(); |
375 | | for (int i = 0; i < int(_nodes.size()); ++i) |
376 | | if (found[_nodes[i]]) queue.push_back(_nodes[i]); |
377 | | int found_cnt = queue.size(); |
378 | | while (found_cnt < int(_nodes.size())) { |
379 | | v = queue.front(); queue.pop_front(); |
380 | | for (InArcIt e(_gr, v); e != INVALID; ++e) { |
381 | | u = _gr.source(e); |
382 | | if (_comp[u] == comp && !found[u]) { |
383 | | found[u] = true; |
384 | | ++found_cnt; |
385 | | _policy[u] = e; |
386 | | queue.push_back(u); |
| 377 | if ( !_curr_found || |
| 378 | (clength * _curr_size < _curr_length * csize) ) { |
| 379 | _curr_found = true; |
| 380 | _curr_length = clength; |
| 381 | _curr_size = csize; |
| 382 | _curr_node = u; |
387 | 383 | } |
388 | 384 | } |
389 | 385 | } |
390 | 386 | } |
391 | 387 | |
392 | | // Compute node distances in the policy graph and update the |
393 | | // policy graph if the node distances can be improved. |
| 388 | // Contract the policy graph and compute node distances |
394 | 389 | bool computeNodeDistances() { |
395 | | // Compute node distances using reverse BFS search |
396 | | double cycle_mean = double(_cycle_length) / _cycle_size; |
397 | | typename Digraph::template NodeMap<int> found(_gr, false); |
398 | | std::deque<Node> queue; |
399 | | queue.push_back(_cycle_node); |
400 | | found[_cycle_node] = true; |
401 | | _dist[_cycle_node] = 0; |
| 390 | // Find the component of the main cycle and compute node distances |
| 391 | // using reverse BFS |
| 392 | for (int i = 0; i < int(_nodes->size()); ++i) { |
| 393 | _reached[(*_nodes)[i]] = false; |
| 394 | } |
| 395 | double curr_mean = double(_curr_length) / _curr_size; |
| 396 | _qfront = _qback = 0; |
| 397 | _queue[0] = _curr_node; |
| 398 | _reached[_curr_node] = true; |
| 399 | _dist[_curr_node] = 0; |
402 | 400 | Node u, v; |
403 | | while (!queue.empty()) { |
404 | | v = queue.front(); queue.pop_front(); |
405 | | for (InArcIt e(_gr, v); e != INVALID; ++e) { |
| 401 | Arc e; |
| 402 | while (_qfront <= _qback) { |
| 403 | v = _queue[_qfront++]; |
| 404 | for (int j = 0; j < int(_in_arcs[v].size()); ++j) { |
| 405 | e = _in_arcs[v][j]; |
406 | 406 | u = _gr.source(e); |
407 | | if (_policy[u] == e && !found[u]) { |
408 | | found[u] = true; |
409 | | _dist[u] = _dist[v] + _length[e] - cycle_mean; |
410 | | queue.push_back(u); |
| 407 | if (_policy[u] == e && !_reached[u]) { |
| 408 | _reached[u] = true; |
| 409 | _dist[u] = _dist[v] + _length[e] - curr_mean; |
| 410 | _queue[++_qback] = u; |
411 | 411 | } |
412 | 412 | } |
413 | 413 | } |
414 | | // Improving node distances |
| 414 | |
| 415 | // Connect all other nodes to this component and compute node |
| 416 | // distances using reverse BFS |
| 417 | _qfront = 0; |
| 418 | while (_qback < int(_nodes->size())-1) { |
| 419 | v = _queue[_qfront++]; |
| 420 | for (int j = 0; j < int(_in_arcs[v].size()); ++j) { |
| 421 | e = _in_arcs[v][j]; |
| 422 | u = _gr.source(e); |
| 423 | if (!_reached[u]) { |
| 424 | _reached[u] = true; |
| 425 | _policy[u] = e; |
| 426 | _dist[u] = _dist[v] + _length[e] - curr_mean; |
| 427 | _queue[++_qback] = u; |
| 428 | } |
| 429 | } |
| 430 | } |
| 431 | |
| 432 | // Improve node distances |
415 | 433 | bool improved = false; |
416 | | for (int j = 0; j < int(_arcs.size()); ++j) { |
417 | | Arc e = _arcs[j]; |
418 | | u = _gr.source(e); v = _gr.target(e); |
419 | | double delta = _dist[v] + _length[e] - cycle_mean; |
420 | | if (_tol.less(delta, _dist[u])) { |
421 | | improved = true; |
422 | | _dist[u] = delta; |
423 | | _policy[u] = e; |
| 434 | for (int i = 0; i < int(_nodes->size()); ++i) { |
| 435 | v = (*_nodes)[i]; |
| 436 | for (int j = 0; j < int(_in_arcs[v].size()); ++j) { |
| 437 | e = _in_arcs[v][j]; |
| 438 | u = _gr.source(e); |
| 439 | double delta = _dist[v] + _length[e] - curr_mean; |
| 440 | if (_tol.less(delta, _dist[u])) { |
| 441 | _dist[u] = delta; |
| 442 | _policy[u] = e; |
| 443 | improved = true; |
| 444 | } |
424 | 445 | } |
425 | 446 | } |
426 | 447 | return improved; |