1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/lemon/grosso_locatelli_pullan_mc.h Sun Sep 12 08:32:46 2010 +0200
1.3 @@ -0,0 +1,680 @@
1.4 +/* -*- mode: C++; indent-tabs-mode: nil; -*-
1.5 + *
1.6 + * This file is a part of LEMON, a generic C++ optimization library.
1.7 + *
1.8 + * Copyright (C) 2003-2010
1.9 + * Egervary Jeno Kombinatorikus Optimalizalasi Kutatocsoport
1.10 + * (Egervary Research Group on Combinatorial Optimization, EGRES).
1.11 + *
1.12 + * Permission to use, modify and distribute this software is granted
1.13 + * provided that this copyright notice appears in all copies. For
1.14 + * precise terms see the accompanying LICENSE file.
1.15 + *
1.16 + * This software is provided "AS IS" with no warranty of any kind,
1.17 + * express or implied, and with no claim as to its suitability for any
1.18 + * purpose.
1.19 + *
1.20 + */
1.21 +
1.22 +#ifndef LEMON_GROSSO_LOCATELLI_PULLAN_MC_H
1.23 +#define LEMON_GROSSO_LOCATELLI_PULLAN_MC_H
1.24 +
1.25 +/// \ingroup approx_algs
1.26 +///
1.27 +/// \file
1.28 +/// \brief The iterated local search algorithm of Grosso, Locatelli, and Pullan
1.29 +/// for the maximum clique problem
1.30 +
1.31 +#include <vector>
1.32 +#include <limits>
1.33 +#include <lemon/core.h>
1.34 +#include <lemon/random.h>
1.35 +
1.36 +namespace lemon {
1.37 +
1.38 + /// \addtogroup approx_algs
1.39 + /// @{
1.40 +
1.41 + /// \brief Implementation of the iterated local search algorithm of Grosso,
1.42 + /// Locatelli, and Pullan for the maximum clique problem
1.43 + ///
1.44 + /// \ref GrossoLocatelliPullanMc implements the iterated local search
1.45 + /// algorithm of Grosso, Locatelli, and Pullan for solving the \e maximum
1.46 + /// \e clique \e problem \ref grosso08maxclique.
1.47 + /// It is to find the largest complete subgraph (\e clique) in an
1.48 + /// undirected graph, i.e., the largest set of nodes where each
1.49 + /// pair of nodes is connected.
1.50 + ///
1.51 + /// This class provides a simple but highly efficient and robust heuristic
1.52 + /// method that quickly finds a large clique, but not necessarily the
1.53 + /// largest one.
1.54 + ///
1.55 + /// \tparam GR The undirected graph type the algorithm runs on.
1.56 + ///
1.57 + /// \note %GrossoLocatelliPullanMc provides three different node selection
1.58 + /// rules, from which the most powerful one is used by default.
1.59 + /// For more information, see \ref SelectionRule.
1.60 + template <typename GR>
1.61 + class GrossoLocatelliPullanMc
1.62 + {
1.63 + public:
1.64 +
1.65 + /// \brief Constants for specifying the node selection rule.
1.66 + ///
1.67 + /// Enum type containing constants for specifying the node selection rule
1.68 + /// for the \ref run() function.
1.69 + ///
1.70 + /// During the algorithm, nodes are selected for addition to the current
1.71 + /// clique according to the applied rule.
1.72 + /// In general, the PENALTY_BASED rule turned out to be the most powerful
1.73 + /// and the most robust, thus it is the default option.
1.74 + /// However, another selection rule can be specified using the \ref run()
1.75 + /// function with the proper parameter.
1.76 + enum SelectionRule {
1.77 +
1.78 + /// A node is selected randomly without any evaluation at each step.
1.79 + RANDOM,
1.80 +
1.81 + /// A node of maximum degree is selected randomly at each step.
1.82 + DEGREE_BASED,
1.83 +
1.84 + /// A node of minimum penalty is selected randomly at each step.
1.85 + /// The node penalties are updated adaptively after each stage of the
1.86 + /// search process.
1.87 + PENALTY_BASED
1.88 + };
1.89 +
1.90 + private:
1.91 +
1.92 + TEMPLATE_GRAPH_TYPEDEFS(GR);
1.93 +
1.94 + typedef std::vector<int> IntVector;
1.95 + typedef std::vector<char> BoolVector;
1.96 + typedef std::vector<BoolVector> BoolMatrix;
1.97 + // Note: vector<char> is used instead of vector<bool> for efficiency reasons
1.98 +
1.99 + const GR &_graph;
1.100 + IntNodeMap _id;
1.101 +
1.102 + // Internal matrix representation of the graph
1.103 + BoolMatrix _gr;
1.104 + int _n;
1.105 +
1.106 + // The current clique
1.107 + BoolVector _clique;
1.108 + int _size;
1.109 +
1.110 + // The best clique found so far
1.111 + BoolVector _best_clique;
1.112 + int _best_size;
1.113 +
1.114 + // The "distances" of the nodes from the current clique.
1.115 + // _delta[u] is the number of nodes in the clique that are
1.116 + // not connected with u.
1.117 + IntVector _delta;
1.118 +
1.119 + // The current tabu set
1.120 + BoolVector _tabu;
1.121 +
1.122 + // Random number generator
1.123 + Random _rnd;
1.124 +
1.125 + private:
1.126 +
1.127 + // Implementation of the RANDOM node selection rule.
1.128 + class RandomSelectionRule
1.129 + {
1.130 + private:
1.131 +
1.132 + // References to the algorithm instance
1.133 + const BoolVector &_clique;
1.134 + const IntVector &_delta;
1.135 + const BoolVector &_tabu;
1.136 + Random &_rnd;
1.137 +
1.138 + // Pivot rule data
1.139 + int _n;
1.140 +
1.141 + public:
1.142 +
1.143 + // Constructor
1.144 + RandomSelectionRule(GrossoLocatelliPullanMc &mc) :
1.145 + _clique(mc._clique), _delta(mc._delta), _tabu(mc._tabu),
1.146 + _rnd(mc._rnd), _n(mc._n)
1.147 + {}
1.148 +
1.149 + // Return a node index for a feasible add move or -1 if no one exists
1.150 + int nextFeasibleAddNode() const {
1.151 + int start_node = _rnd[_n];
1.152 + for (int i = start_node; i != _n; i++) {
1.153 + if (_delta[i] == 0 && !_tabu[i]) return i;
1.154 + }
1.155 + for (int i = 0; i != start_node; i++) {
1.156 + if (_delta[i] == 0 && !_tabu[i]) return i;
1.157 + }
1.158 + return -1;
1.159 + }
1.160 +
1.161 + // Return a node index for a feasible swap move or -1 if no one exists
1.162 + int nextFeasibleSwapNode() const {
1.163 + int start_node = _rnd[_n];
1.164 + for (int i = start_node; i != _n; i++) {
1.165 + if (!_clique[i] && _delta[i] == 1 && !_tabu[i]) return i;
1.166 + }
1.167 + for (int i = 0; i != start_node; i++) {
1.168 + if (!_clique[i] && _delta[i] == 1 && !_tabu[i]) return i;
1.169 + }
1.170 + return -1;
1.171 + }
1.172 +
1.173 + // Return a node index for an add move or -1 if no one exists
1.174 + int nextAddNode() const {
1.175 + int start_node = _rnd[_n];
1.176 + for (int i = start_node; i != _n; i++) {
1.177 + if (_delta[i] == 0) return i;
1.178 + }
1.179 + for (int i = 0; i != start_node; i++) {
1.180 + if (_delta[i] == 0) return i;
1.181 + }
1.182 + return -1;
1.183 + }
1.184 +
1.185 + // Update internal data structures between stages (if necessary)
1.186 + void update() {}
1.187 +
1.188 + }; //class RandomSelectionRule
1.189 +
1.190 +
1.191 + // Implementation of the DEGREE_BASED node selection rule.
1.192 + class DegreeBasedSelectionRule
1.193 + {
1.194 + private:
1.195 +
1.196 + // References to the algorithm instance
1.197 + const BoolVector &_clique;
1.198 + const IntVector &_delta;
1.199 + const BoolVector &_tabu;
1.200 + Random &_rnd;
1.201 +
1.202 + // Pivot rule data
1.203 + int _n;
1.204 + IntVector _deg;
1.205 +
1.206 + public:
1.207 +
1.208 + // Constructor
1.209 + DegreeBasedSelectionRule(GrossoLocatelliPullanMc &mc) :
1.210 + _clique(mc._clique), _delta(mc._delta), _tabu(mc._tabu),
1.211 + _rnd(mc._rnd), _n(mc._n), _deg(_n)
1.212 + {
1.213 + for (int i = 0; i != _n; i++) {
1.214 + int d = 0;
1.215 + BoolVector &row = mc._gr[i];
1.216 + for (int j = 0; j != _n; j++) {
1.217 + if (row[j]) d++;
1.218 + }
1.219 + _deg[i] = d;
1.220 + }
1.221 + }
1.222 +
1.223 + // Return a node index for a feasible add move or -1 if no one exists
1.224 + int nextFeasibleAddNode() const {
1.225 + int start_node = _rnd[_n];
1.226 + int node = -1, max_deg = -1;
1.227 + for (int i = start_node; i != _n; i++) {
1.228 + if (_delta[i] == 0 && !_tabu[i] && _deg[i] > max_deg) {
1.229 + node = i;
1.230 + max_deg = _deg[i];
1.231 + }
1.232 + }
1.233 + for (int i = 0; i != start_node; i++) {
1.234 + if (_delta[i] == 0 && !_tabu[i] && _deg[i] > max_deg) {
1.235 + node = i;
1.236 + max_deg = _deg[i];
1.237 + }
1.238 + }
1.239 + return node;
1.240 + }
1.241 +
1.242 + // Return a node index for a feasible swap move or -1 if no one exists
1.243 + int nextFeasibleSwapNode() const {
1.244 + int start_node = _rnd[_n];
1.245 + int node = -1, max_deg = -1;
1.246 + for (int i = start_node; i != _n; i++) {
1.247 + if (!_clique[i] && _delta[i] == 1 && !_tabu[i] &&
1.248 + _deg[i] > max_deg) {
1.249 + node = i;
1.250 + max_deg = _deg[i];
1.251 + }
1.252 + }
1.253 + for (int i = 0; i != start_node; i++) {
1.254 + if (!_clique[i] && _delta[i] == 1 && !_tabu[i] &&
1.255 + _deg[i] > max_deg) {
1.256 + node = i;
1.257 + max_deg = _deg[i];
1.258 + }
1.259 + }
1.260 + return node;
1.261 + }
1.262 +
1.263 + // Return a node index for an add move or -1 if no one exists
1.264 + int nextAddNode() const {
1.265 + int start_node = _rnd[_n];
1.266 + int node = -1, max_deg = -1;
1.267 + for (int i = start_node; i != _n; i++) {
1.268 + if (_delta[i] == 0 && _deg[i] > max_deg) {
1.269 + node = i;
1.270 + max_deg = _deg[i];
1.271 + }
1.272 + }
1.273 + for (int i = 0; i != start_node; i++) {
1.274 + if (_delta[i] == 0 && _deg[i] > max_deg) {
1.275 + node = i;
1.276 + max_deg = _deg[i];
1.277 + }
1.278 + }
1.279 + return node;
1.280 + }
1.281 +
1.282 + // Update internal data structures between stages (if necessary)
1.283 + void update() {}
1.284 +
1.285 + }; //class DegreeBasedSelectionRule
1.286 +
1.287 +
1.288 + // Implementation of the PENALTY_BASED node selection rule.
1.289 + class PenaltyBasedSelectionRule
1.290 + {
1.291 + private:
1.292 +
1.293 + // References to the algorithm instance
1.294 + const BoolVector &_clique;
1.295 + const IntVector &_delta;
1.296 + const BoolVector &_tabu;
1.297 + Random &_rnd;
1.298 +
1.299 + // Pivot rule data
1.300 + int _n;
1.301 + IntVector _penalty;
1.302 +
1.303 + public:
1.304 +
1.305 + // Constructor
1.306 + PenaltyBasedSelectionRule(GrossoLocatelliPullanMc &mc) :
1.307 + _clique(mc._clique), _delta(mc._delta), _tabu(mc._tabu),
1.308 + _rnd(mc._rnd), _n(mc._n), _penalty(_n, 0)
1.309 + {}
1.310 +
1.311 + // Return a node index for a feasible add move or -1 if no one exists
1.312 + int nextFeasibleAddNode() const {
1.313 + int start_node = _rnd[_n];
1.314 + int node = -1, min_p = std::numeric_limits<int>::max();
1.315 + for (int i = start_node; i != _n; i++) {
1.316 + if (_delta[i] == 0 && !_tabu[i] && _penalty[i] < min_p) {
1.317 + node = i;
1.318 + min_p = _penalty[i];
1.319 + }
1.320 + }
1.321 + for (int i = 0; i != start_node; i++) {
1.322 + if (_delta[i] == 0 && !_tabu[i] && _penalty[i] < min_p) {
1.323 + node = i;
1.324 + min_p = _penalty[i];
1.325 + }
1.326 + }
1.327 + return node;
1.328 + }
1.329 +
1.330 + // Return a node index for a feasible swap move or -1 if no one exists
1.331 + int nextFeasibleSwapNode() const {
1.332 + int start_node = _rnd[_n];
1.333 + int node = -1, min_p = std::numeric_limits<int>::max();
1.334 + for (int i = start_node; i != _n; i++) {
1.335 + if (!_clique[i] && _delta[i] == 1 && !_tabu[i] &&
1.336 + _penalty[i] < min_p) {
1.337 + node = i;
1.338 + min_p = _penalty[i];
1.339 + }
1.340 + }
1.341 + for (int i = 0; i != start_node; i++) {
1.342 + if (!_clique[i] && _delta[i] == 1 && !_tabu[i] &&
1.343 + _penalty[i] < min_p) {
1.344 + node = i;
1.345 + min_p = _penalty[i];
1.346 + }
1.347 + }
1.348 + return node;
1.349 + }
1.350 +
1.351 + // Return a node index for an add move or -1 if no one exists
1.352 + int nextAddNode() const {
1.353 + int start_node = _rnd[_n];
1.354 + int node = -1, min_p = std::numeric_limits<int>::max();
1.355 + for (int i = start_node; i != _n; i++) {
1.356 + if (_delta[i] == 0 && _penalty[i] < min_p) {
1.357 + node = i;
1.358 + min_p = _penalty[i];
1.359 + }
1.360 + }
1.361 + for (int i = 0; i != start_node; i++) {
1.362 + if (_delta[i] == 0 && _penalty[i] < min_p) {
1.363 + node = i;
1.364 + min_p = _penalty[i];
1.365 + }
1.366 + }
1.367 + return node;
1.368 + }
1.369 +
1.370 + // Update internal data structures between stages (if necessary)
1.371 + void update() {}
1.372 +
1.373 + }; //class PenaltyBasedSelectionRule
1.374 +
1.375 + public:
1.376 +
1.377 + /// \brief Constructor.
1.378 + ///
1.379 + /// Constructor.
1.380 + /// The global \ref rnd "random number generator instance" is used
1.381 + /// during the algorithm.
1.382 + ///
1.383 + /// \param graph The undirected graph the algorithm runs on.
1.384 + GrossoLocatelliPullanMc(const GR& graph) :
1.385 + _graph(graph), _id(_graph), _rnd(rnd)
1.386 + {}
1.387 +
1.388 + /// \brief Constructor with random seed.
1.389 + ///
1.390 + /// Constructor with random seed.
1.391 + ///
1.392 + /// \param graph The undirected graph the algorithm runs on.
1.393 + /// \param seed Seed value for the internal random number generator
1.394 + /// that is used during the algorithm.
1.395 + GrossoLocatelliPullanMc(const GR& graph, int seed) :
1.396 + _graph(graph), _id(_graph), _rnd(seed)
1.397 + {}
1.398 +
1.399 + /// \brief Constructor with random number generator.
1.400 + ///
1.401 + /// Constructor with random number generator.
1.402 + ///
1.403 + /// \param graph The undirected graph the algorithm runs on.
1.404 + /// \param random A random number generator that is used during the
1.405 + /// algorithm.
1.406 + GrossoLocatelliPullanMc(const GR& graph, const Random& random) :
1.407 + _graph(graph), _id(_graph), _rnd(random)
1.408 + {}
1.409 +
1.410 + /// \name Execution Control
1.411 + /// @{
1.412 +
1.413 + /// \brief Runs the algorithm.
1.414 + ///
1.415 + /// This function runs the algorithm.
1.416 + ///
1.417 + /// \param step_num The maximum number of node selections (steps)
1.418 + /// during the search process.
1.419 + /// This parameter controls the running time and the success of the
1.420 + /// algorithm. For larger values, the algorithm runs slower but it more
1.421 + /// likely finds larger cliques. For smaller values, the algorithm is
1.422 + /// faster but probably gives worse results.
1.423 + /// \param rule The node selection rule. For more information, see
1.424 + /// \ref SelectionRule.
1.425 + ///
1.426 + /// \return The size of the found clique.
1.427 + int run(int step_num = 100000,
1.428 + SelectionRule rule = PENALTY_BASED)
1.429 + {
1.430 + init();
1.431 + switch (rule) {
1.432 + case RANDOM:
1.433 + return start<RandomSelectionRule>(step_num);
1.434 + case DEGREE_BASED:
1.435 + return start<DegreeBasedSelectionRule>(step_num);
1.436 + case PENALTY_BASED:
1.437 + return start<PenaltyBasedSelectionRule>(step_num);
1.438 + }
1.439 + return 0; // avoid warning
1.440 + }
1.441 +
1.442 + /// @}
1.443 +
1.444 + /// \name Query Functions
1.445 + /// @{
1.446 +
1.447 + /// \brief The size of the found clique
1.448 + ///
1.449 + /// This function returns the size of the found clique.
1.450 + ///
1.451 + /// \pre run() must be called before using this function.
1.452 + int cliqueSize() const {
1.453 + return _best_size;
1.454 + }
1.455 +
1.456 + /// \brief Gives back the found clique in a \c bool node map
1.457 + ///
1.458 + /// This function gives back the characteristic vector of the found
1.459 + /// clique in the given node map.
1.460 + /// It must be a \ref concepts::WriteMap "writable" node map with
1.461 + /// \c bool (or convertible) value type.
1.462 + ///
1.463 + /// \pre run() must be called before using this function.
1.464 + template <typename CliqueMap>
1.465 + void cliqueMap(CliqueMap &map) const {
1.466 + for (NodeIt n(_graph); n != INVALID; ++n) {
1.467 + map[n] = static_cast<bool>(_best_clique[_id[n]]);
1.468 + }
1.469 + }
1.470 +
1.471 + /// \brief Iterator to list the nodes of the found clique
1.472 + ///
1.473 + /// This iterator class lists the nodes of the found clique.
1.474 + /// Before using it, you must allocate a GrossoLocatelliPullanMc instance
1.475 + /// and call its \ref GrossoLocatelliPullanMc::run() "run()" method.
1.476 + ///
1.477 + /// The following example prints out the IDs of the nodes in the found
1.478 + /// clique.
1.479 + /// \code
1.480 + /// GrossoLocatelliPullanMc<Graph> mc(g);
1.481 + /// mc.run();
1.482 + /// for (GrossoLocatelliPullanMc<Graph>::CliqueNodeIt n(mc);
1.483 + /// n != INVALID; ++n)
1.484 + /// {
1.485 + /// std::cout << g.id(n) << std::endl;
1.486 + /// }
1.487 + /// \endcode
1.488 + class CliqueNodeIt
1.489 + {
1.490 + private:
1.491 + NodeIt _it;
1.492 + BoolNodeMap _map;
1.493 +
1.494 + public:
1.495 +
1.496 + /// Constructor
1.497 +
1.498 + /// Constructor.
1.499 + /// \param mc The algorithm instance.
1.500 + CliqueNodeIt(const GrossoLocatelliPullanMc &mc)
1.501 + : _map(mc._graph)
1.502 + {
1.503 + mc.cliqueMap(_map);
1.504 + for (_it = NodeIt(mc._graph); _it != INVALID && !_map[_it]; ++_it) ;
1.505 + }
1.506 +
1.507 + /// Conversion to \c Node
1.508 + operator Node() const { return _it; }
1.509 +
1.510 + bool operator==(Invalid) const { return _it == INVALID; }
1.511 + bool operator!=(Invalid) const { return _it != INVALID; }
1.512 +
1.513 + /// Next node
1.514 + CliqueNodeIt &operator++() {
1.515 + for (++_it; _it != INVALID && !_map[_it]; ++_it) ;
1.516 + return *this;
1.517 + }
1.518 +
1.519 + /// Postfix incrementation
1.520 +
1.521 + /// Postfix incrementation.
1.522 + ///
1.523 + /// \warning This incrementation returns a \c Node, not a
1.524 + /// \c CliqueNodeIt as one may expect.
1.525 + typename GR::Node operator++(int) {
1.526 + Node n=*this;
1.527 + ++(*this);
1.528 + return n;
1.529 + }
1.530 +
1.531 + };
1.532 +
1.533 + /// @}
1.534 +
1.535 + private:
1.536 +
1.537 + // Adds a node to the current clique
1.538 + void addCliqueNode(int u) {
1.539 + if (_clique[u]) return;
1.540 + _clique[u] = true;
1.541 + _size++;
1.542 + BoolVector &row = _gr[u];
1.543 + for (int i = 0; i != _n; i++) {
1.544 + if (!row[i]) _delta[i]++;
1.545 + }
1.546 + }
1.547 +
1.548 + // Removes a node from the current clique
1.549 + void delCliqueNode(int u) {
1.550 + if (!_clique[u]) return;
1.551 + _clique[u] = false;
1.552 + _size--;
1.553 + BoolVector &row = _gr[u];
1.554 + for (int i = 0; i != _n; i++) {
1.555 + if (!row[i]) _delta[i]--;
1.556 + }
1.557 + }
1.558 +
1.559 + // Initialize data structures
1.560 + void init() {
1.561 + _n = countNodes(_graph);
1.562 + int ui = 0;
1.563 + for (NodeIt u(_graph); u != INVALID; ++u) {
1.564 + _id[u] = ui++;
1.565 + }
1.566 + _gr.clear();
1.567 + _gr.resize(_n, BoolVector(_n, false));
1.568 + ui = 0;
1.569 + for (NodeIt u(_graph); u != INVALID; ++u) {
1.570 + for (IncEdgeIt e(_graph, u); e != INVALID; ++e) {
1.571 + int vi = _id[_graph.runningNode(e)];
1.572 + _gr[ui][vi] = true;
1.573 + _gr[vi][ui] = true;
1.574 + }
1.575 + ++ui;
1.576 + }
1.577 +
1.578 + _clique.clear();
1.579 + _clique.resize(_n, false);
1.580 + _size = 0;
1.581 + _best_clique.clear();
1.582 + _best_clique.resize(_n, false);
1.583 + _best_size = 0;
1.584 + _delta.clear();
1.585 + _delta.resize(_n, 0);
1.586 + _tabu.clear();
1.587 + _tabu.resize(_n, false);
1.588 + }
1.589 +
1.590 + // Executes the algorithm
1.591 + template <typename SelectionRuleImpl>
1.592 + int start(int max_select) {
1.593 + // Options for the restart rule
1.594 + const bool delta_based_restart = true;
1.595 + const int restart_delta_limit = 4;
1.596 +
1.597 + if (_n == 0) return 0;
1.598 + if (_n == 1) {
1.599 + _best_clique[0] = true;
1.600 + _best_size = 1;
1.601 + return _best_size;
1.602 + }
1.603 +
1.604 + // Iterated local search
1.605 + SelectionRuleImpl sel_method(*this);
1.606 + int select = 0;
1.607 + IntVector restart_nodes;
1.608 +
1.609 + while (select < max_select) {
1.610 +
1.611 + // Perturbation/restart
1.612 + if (delta_based_restart) {
1.613 + restart_nodes.clear();
1.614 + for (int i = 0; i != _n; i++) {
1.615 + if (_delta[i] >= restart_delta_limit)
1.616 + restart_nodes.push_back(i);
1.617 + }
1.618 + }
1.619 + int rs_node = -1;
1.620 + if (restart_nodes.size() > 0) {
1.621 + rs_node = restart_nodes[_rnd[restart_nodes.size()]];
1.622 + } else {
1.623 + rs_node = _rnd[_n];
1.624 + }
1.625 + BoolVector &row = _gr[rs_node];
1.626 + for (int i = 0; i != _n; i++) {
1.627 + if (_clique[i] && !row[i]) delCliqueNode(i);
1.628 + }
1.629 + addCliqueNode(rs_node);
1.630 +
1.631 + // Local search
1.632 + _tabu.clear();
1.633 + _tabu.resize(_n, false);
1.634 + bool tabu_empty = true;
1.635 + int max_swap = _size;
1.636 + while (select < max_select) {
1.637 + select++;
1.638 + int u;
1.639 + if ((u = sel_method.nextFeasibleAddNode()) != -1) {
1.640 + // Feasible add move
1.641 + addCliqueNode(u);
1.642 + if (tabu_empty) max_swap = _size;
1.643 + }
1.644 + else if ((u = sel_method.nextFeasibleSwapNode()) != -1) {
1.645 + // Feasible swap move
1.646 + int v = -1;
1.647 + BoolVector &row = _gr[u];
1.648 + for (int i = 0; i != _n; i++) {
1.649 + if (_clique[i] && !row[i]) {
1.650 + v = i;
1.651 + break;
1.652 + }
1.653 + }
1.654 + addCliqueNode(u);
1.655 + delCliqueNode(v);
1.656 + _tabu[v] = true;
1.657 + tabu_empty = false;
1.658 + if (--max_swap <= 0) break;
1.659 + }
1.660 + else if ((u = sel_method.nextAddNode()) != -1) {
1.661 + // Non-feasible add move
1.662 + addCliqueNode(u);
1.663 + }
1.664 + else break;
1.665 + }
1.666 + if (_size > _best_size) {
1.667 + _best_clique = _clique;
1.668 + _best_size = _size;
1.669 + if (_best_size == _n) return _best_size;
1.670 + }
1.671 + sel_method.update();
1.672 + }
1.673 +
1.674 + return _best_size;
1.675 + }
1.676 +
1.677 + }; //class GrossoLocatelliPullanMc
1.678 +
1.679 + ///@}
1.680 +
1.681 +} //namespace lemon
1.682 +
1.683 +#endif //LEMON_GROSSO_LOCATELLI_PULLAN_MC_H