lemon/network_simplex.h
changeset 1099 ad40f7d32846
parent 877 141f9c0db4a3
parent 888 5205145fabf6
child 895 dca9eed2c375
     1.1 --- a/lemon/network_simplex.h	Fri Aug 09 11:07:27 2013 +0200
     1.2 +++ b/lemon/network_simplex.h	Sun Aug 11 15:28:12 2013 +0200
     1.3 @@ -2,7 +2,7 @@
     1.4   *
     1.5   * This file is a part of LEMON, a generic C++ optimization library.
     1.6   *
     1.7 - * Copyright (C) 2003-2009
     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 @@ -40,15 +40,17 @@
    1.13    /// for finding a \ref min_cost_flow "minimum cost flow".
    1.14    ///
    1.15    /// \ref NetworkSimplex implements the primal Network Simplex algorithm
    1.16 -  /// for finding a \ref min_cost_flow "minimum cost flow".
    1.17 -  /// This algorithm is a specialized version of the linear programming
    1.18 -  /// simplex method directly for the minimum cost flow problem.
    1.19 -  /// It is one of the most efficient solution methods.
    1.20 +  /// for finding a \ref min_cost_flow "minimum cost flow"
    1.21 +  /// \ref amo93networkflows, \ref dantzig63linearprog,
    1.22 +  /// \ref kellyoneill91netsimplex.
    1.23 +  /// This algorithm is a highly efficient specialized version of the
    1.24 +  /// linear programming simplex method directly for the minimum cost
    1.25 +  /// flow problem.
    1.26    ///
    1.27 -  /// In general this class is the fastest implementation available
    1.28 -  /// in LEMON for the minimum cost flow problem.
    1.29 -  /// Moreover it supports both directions of the supply/demand inequality
    1.30 -  /// constraints. For more information see \ref SupplyType.
    1.31 +  /// In general, %NetworkSimplex is the fastest implementation available
    1.32 +  /// in LEMON for this problem.
    1.33 +  /// Moreover, it supports both directions of the supply/demand inequality
    1.34 +  /// constraints. For more information, see \ref SupplyType.
    1.35    ///
    1.36    /// Most of the parameters of the problem (except for the digraph)
    1.37    /// can be given using separate functions, and the algorithm can be
    1.38 @@ -56,17 +58,17 @@
    1.39    /// specified, then default values will be used.
    1.40    ///
    1.41    /// \tparam GR The digraph type the algorithm runs on.
    1.42 -  /// \tparam V The value type used for flow amounts, capacity bounds
    1.43 -  /// and supply values in the algorithm. By default it is \c int.
    1.44 -  /// \tparam C The value type used for costs and potentials in the
    1.45 -  /// algorithm. By default it is the same as \c V.
    1.46 +  /// \tparam V The number type used for flow amounts, capacity bounds
    1.47 +  /// and supply values in the algorithm. By default, it is \c int.
    1.48 +  /// \tparam C The number type used for costs and potentials in the
    1.49 +  /// algorithm. By default, it is the same as \c V.
    1.50    ///
    1.51 -  /// \warning Both value types must be signed and all input data must
    1.52 +  /// \warning Both number types must be signed and all input data must
    1.53    /// be integer.
    1.54    ///
    1.55    /// \note %NetworkSimplex provides five different pivot rule
    1.56    /// implementations, from which the most efficient one is used
    1.57 -  /// by default. For more information see \ref PivotRule.
    1.58 +  /// by default. For more information, see \ref PivotRule.
    1.59    template <typename GR, typename V = int, typename C = V>
    1.60    class NetworkSimplex
    1.61    {
    1.62 @@ -95,7 +97,7 @@
    1.63        /// infinite upper bound.
    1.64        UNBOUNDED
    1.65      };
    1.66 -    
    1.67 +
    1.68      /// \brief Constants for selecting the type of the supply constraints.
    1.69      ///
    1.70      /// Enum type containing constants for selecting the supply type,
    1.71 @@ -113,7 +115,7 @@
    1.72        /// supply/demand constraints in the definition of the problem.
    1.73        LEQ
    1.74      };
    1.75 -    
    1.76 +
    1.77      /// \brief Constants for selecting the pivot rule.
    1.78      ///
    1.79      /// Enum type containing constants for selecting the pivot rule for
    1.80 @@ -122,59 +124,62 @@
    1.81      /// \ref NetworkSimplex provides five different pivot rule
    1.82      /// implementations that significantly affect the running time
    1.83      /// of the algorithm.
    1.84 -    /// By default \ref BLOCK_SEARCH "Block Search" is used, which
    1.85 +    /// By default, \ref BLOCK_SEARCH "Block Search" is used, which
    1.86      /// proved to be the most efficient and the most robust on various
    1.87 -    /// test inputs according to our benchmark tests.
    1.88 -    /// However another pivot rule can be selected using the \ref run()
    1.89 +    /// test inputs.
    1.90 +    /// However, another pivot rule can be selected using the \ref run()
    1.91      /// function with the proper parameter.
    1.92      enum PivotRule {
    1.93  
    1.94 -      /// The First Eligible pivot rule.
    1.95 +      /// The \e First \e Eligible pivot rule.
    1.96        /// The next eligible arc is selected in a wraparound fashion
    1.97        /// in every iteration.
    1.98        FIRST_ELIGIBLE,
    1.99  
   1.100 -      /// The Best Eligible pivot rule.
   1.101 +      /// The \e Best \e Eligible pivot rule.
   1.102        /// The best eligible arc is selected in every iteration.
   1.103        BEST_ELIGIBLE,
   1.104  
   1.105 -      /// The Block Search pivot rule.
   1.106 +      /// The \e Block \e Search pivot rule.
   1.107        /// A specified number of arcs are examined in every iteration
   1.108        /// in a wraparound fashion and the best eligible arc is selected
   1.109        /// from this block.
   1.110        BLOCK_SEARCH,
   1.111  
   1.112 -      /// The Candidate List pivot rule.
   1.113 +      /// The \e Candidate \e List pivot rule.
   1.114        /// In a major iteration a candidate list is built from eligible arcs
   1.115        /// in a wraparound fashion and in the following minor iterations
   1.116        /// the best eligible arc is selected from this list.
   1.117        CANDIDATE_LIST,
   1.118  
   1.119 -      /// The Altering Candidate List pivot rule.
   1.120 +      /// The \e Altering \e Candidate \e List pivot rule.
   1.121        /// It is a modified version of the Candidate List method.
   1.122        /// It keeps only the several best eligible arcs from the former
   1.123        /// candidate list and extends this list in every iteration.
   1.124        ALTERING_LIST
   1.125      };
   1.126 -    
   1.127 +
   1.128    private:
   1.129  
   1.130      TEMPLATE_DIGRAPH_TYPEDEFS(GR);
   1.131  
   1.132 -    typedef std::vector<Arc> ArcVector;
   1.133 -    typedef std::vector<Node> NodeVector;
   1.134      typedef std::vector<int> IntVector;
   1.135 -    typedef std::vector<bool> BoolVector;
   1.136      typedef std::vector<Value> ValueVector;
   1.137      typedef std::vector<Cost> CostVector;
   1.138 +    typedef std::vector<char> BoolVector;
   1.139 +    // Note: vector<char> is used instead of vector<bool> for efficiency reasons
   1.140  
   1.141      // State constants for arcs
   1.142 -    enum ArcStateEnum {
   1.143 +    enum ArcState {
   1.144        STATE_UPPER = -1,
   1.145        STATE_TREE  =  0,
   1.146        STATE_LOWER =  1
   1.147      };
   1.148  
   1.149 +    typedef std::vector<signed char> StateVector;
   1.150 +    // Note: vector<signed char> is used instead of vector<ArcState> for
   1.151 +    // efficiency reasons
   1.152 +
   1.153    private:
   1.154  
   1.155      // Data related to the underlying digraph
   1.156 @@ -194,6 +199,7 @@
   1.157      IntArcMap _arc_id;
   1.158      IntVector _source;
   1.159      IntVector _target;
   1.160 +    bool _arc_mixing;
   1.161  
   1.162      // Node and arc data
   1.163      ValueVector _lower;
   1.164 @@ -213,7 +219,7 @@
   1.165      IntVector _last_succ;
   1.166      IntVector _dirty_revs;
   1.167      BoolVector _forward;
   1.168 -    IntVector _state;
   1.169 +    StateVector _state;
   1.170      int _root;
   1.171  
   1.172      // Temporary data used in the current pivot iteration
   1.173 @@ -222,8 +228,10 @@
   1.174      int stem, par_stem, new_stem;
   1.175      Value delta;
   1.176  
   1.177 +    const Value MAX;
   1.178 +
   1.179    public:
   1.180 -  
   1.181 +
   1.182      /// \brief Constant for infinite upper bounds (capacities).
   1.183      ///
   1.184      /// Constant for infinite upper bounds (capacities).
   1.185 @@ -242,7 +250,7 @@
   1.186        const IntVector  &_source;
   1.187        const IntVector  &_target;
   1.188        const CostVector &_cost;
   1.189 -      const IntVector  &_state;
   1.190 +      const StateVector &_state;
   1.191        const CostVector &_pi;
   1.192        int &_in_arc;
   1.193        int _search_arc_num;
   1.194 @@ -263,7 +271,7 @@
   1.195        // Find next entering arc
   1.196        bool findEnteringArc() {
   1.197          Cost c;
   1.198 -        for (int e = _next_arc; e < _search_arc_num; ++e) {
   1.199 +        for (int e = _next_arc; e != _search_arc_num; ++e) {
   1.200            c = _state[e] * (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.201            if (c < 0) {
   1.202              _in_arc = e;
   1.203 @@ -271,7 +279,7 @@
   1.204              return true;
   1.205            }
   1.206          }
   1.207 -        for (int e = 0; e < _next_arc; ++e) {
   1.208 +        for (int e = 0; e != _next_arc; ++e) {
   1.209            c = _state[e] * (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.210            if (c < 0) {
   1.211              _in_arc = e;
   1.212 @@ -294,7 +302,7 @@
   1.213        const IntVector  &_source;
   1.214        const IntVector  &_target;
   1.215        const CostVector &_cost;
   1.216 -      const IntVector  &_state;
   1.217 +      const StateVector &_state;
   1.218        const CostVector &_pi;
   1.219        int &_in_arc;
   1.220        int _search_arc_num;
   1.221 @@ -311,7 +319,7 @@
   1.222        // Find next entering arc
   1.223        bool findEnteringArc() {
   1.224          Cost c, min = 0;
   1.225 -        for (int e = 0; e < _search_arc_num; ++e) {
   1.226 +        for (int e = 0; e != _search_arc_num; ++e) {
   1.227            c = _state[e] * (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.228            if (c < min) {
   1.229              min = c;
   1.230 @@ -333,7 +341,7 @@
   1.231        const IntVector  &_source;
   1.232        const IntVector  &_target;
   1.233        const CostVector &_cost;
   1.234 -      const IntVector  &_state;
   1.235 +      const StateVector &_state;
   1.236        const CostVector &_pi;
   1.237        int &_in_arc;
   1.238        int _search_arc_num;
   1.239 @@ -352,7 +360,7 @@
   1.240          _next_arc(0)
   1.241        {
   1.242          // The main parameters of the pivot rule
   1.243 -        const double BLOCK_SIZE_FACTOR = 0.5;
   1.244 +        const double BLOCK_SIZE_FACTOR = 1.0;
   1.245          const int MIN_BLOCK_SIZE = 10;
   1.246  
   1.247          _block_size = std::max( int(BLOCK_SIZE_FACTOR *
   1.248 @@ -364,33 +372,32 @@
   1.249        bool findEnteringArc() {
   1.250          Cost c, min = 0;
   1.251          int cnt = _block_size;
   1.252 -        int e, min_arc = _next_arc;
   1.253 -        for (e = _next_arc; e < _search_arc_num; ++e) {
   1.254 +        int e;
   1.255 +        for (e = _next_arc; e != _search_arc_num; ++e) {
   1.256            c = _state[e] * (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.257            if (c < min) {
   1.258              min = c;
   1.259 -            min_arc = e;
   1.260 +            _in_arc = e;
   1.261            }
   1.262            if (--cnt == 0) {
   1.263 -            if (min < 0) break;
   1.264 +            if (min < 0) goto search_end;
   1.265              cnt = _block_size;
   1.266            }
   1.267          }
   1.268 -        if (min == 0 || cnt > 0) {
   1.269 -          for (e = 0; e < _next_arc; ++e) {
   1.270 -            c = _state[e] * (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.271 -            if (c < min) {
   1.272 -              min = c;
   1.273 -              min_arc = e;
   1.274 -            }
   1.275 -            if (--cnt == 0) {
   1.276 -              if (min < 0) break;
   1.277 -              cnt = _block_size;
   1.278 -            }
   1.279 +        for (e = 0; e != _next_arc; ++e) {
   1.280 +          c = _state[e] * (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.281 +          if (c < min) {
   1.282 +            min = c;
   1.283 +            _in_arc = e;
   1.284 +          }
   1.285 +          if (--cnt == 0) {
   1.286 +            if (min < 0) goto search_end;
   1.287 +            cnt = _block_size;
   1.288            }
   1.289          }
   1.290          if (min >= 0) return false;
   1.291 -        _in_arc = min_arc;
   1.292 +
   1.293 +      search_end:
   1.294          _next_arc = e;
   1.295          return true;
   1.296        }
   1.297 @@ -407,7 +414,7 @@
   1.298        const IntVector  &_source;
   1.299        const IntVector  &_target;
   1.300        const CostVector &_cost;
   1.301 -      const IntVector  &_state;
   1.302 +      const StateVector &_state;
   1.303        const CostVector &_pi;
   1.304        int &_in_arc;
   1.305        int _search_arc_num;
   1.306 @@ -428,7 +435,7 @@
   1.307          _next_arc(0)
   1.308        {
   1.309          // The main parameters of the pivot rule
   1.310 -        const double LIST_LENGTH_FACTOR = 1.0;
   1.311 +        const double LIST_LENGTH_FACTOR = 0.25;
   1.312          const int MIN_LIST_LENGTH = 10;
   1.313          const double MINOR_LIMIT_FACTOR = 0.1;
   1.314          const int MIN_MINOR_LIMIT = 3;
   1.315 @@ -445,7 +452,7 @@
   1.316        /// Find next entering arc
   1.317        bool findEnteringArc() {
   1.318          Cost min, c;
   1.319 -        int e, min_arc = _next_arc;
   1.320 +        int e;
   1.321          if (_curr_length > 0 && _minor_count < _minor_limit) {
   1.322            // Minor iteration: select the best eligible arc from the
   1.323            // current candidate list
   1.324 @@ -456,48 +463,44 @@
   1.325              c = _state[e] * (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.326              if (c < min) {
   1.327                min = c;
   1.328 -              min_arc = e;
   1.329 +              _in_arc = e;
   1.330              }
   1.331 -            if (c >= 0) {
   1.332 +            else if (c >= 0) {
   1.333                _candidates[i--] = _candidates[--_curr_length];
   1.334              }
   1.335            }
   1.336 -          if (min < 0) {
   1.337 -            _in_arc = min_arc;
   1.338 -            return true;
   1.339 -          }
   1.340 +          if (min < 0) return true;
   1.341          }
   1.342  
   1.343          // Major iteration: build a new candidate list
   1.344          min = 0;
   1.345          _curr_length = 0;
   1.346 -        for (e = _next_arc; e < _search_arc_num; ++e) {
   1.347 +        for (e = _next_arc; e != _search_arc_num; ++e) {
   1.348            c = _state[e] * (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.349            if (c < 0) {
   1.350              _candidates[_curr_length++] = e;
   1.351              if (c < min) {
   1.352                min = c;
   1.353 -              min_arc = e;
   1.354 +              _in_arc = e;
   1.355              }
   1.356 -            if (_curr_length == _list_length) break;
   1.357 +            if (_curr_length == _list_length) goto search_end;
   1.358            }
   1.359          }
   1.360 -        if (_curr_length < _list_length) {
   1.361 -          for (e = 0; e < _next_arc; ++e) {
   1.362 -            c = _state[e] * (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.363 -            if (c < 0) {
   1.364 -              _candidates[_curr_length++] = e;
   1.365 -              if (c < min) {
   1.366 -                min = c;
   1.367 -                min_arc = e;
   1.368 -              }
   1.369 -              if (_curr_length == _list_length) break;
   1.370 +        for (e = 0; e != _next_arc; ++e) {
   1.371 +          c = _state[e] * (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.372 +          if (c < 0) {
   1.373 +            _candidates[_curr_length++] = e;
   1.374 +            if (c < min) {
   1.375 +              min = c;
   1.376 +              _in_arc = e;
   1.377              }
   1.378 +            if (_curr_length == _list_length) goto search_end;
   1.379            }
   1.380          }
   1.381          if (_curr_length == 0) return false;
   1.382 +
   1.383 +      search_end:
   1.384          _minor_count = 1;
   1.385 -        _in_arc = min_arc;
   1.386          _next_arc = e;
   1.387          return true;
   1.388        }
   1.389 @@ -514,7 +517,7 @@
   1.390        const IntVector  &_source;
   1.391        const IntVector  &_target;
   1.392        const CostVector &_cost;
   1.393 -      const IntVector  &_state;
   1.394 +      const StateVector &_state;
   1.395        const CostVector &_pi;
   1.396        int &_in_arc;
   1.397        int _search_arc_num;
   1.398 @@ -549,7 +552,7 @@
   1.399          _next_arc(0), _cand_cost(ns._search_arc_num), _sort_func(_cand_cost)
   1.400        {
   1.401          // The main parameters of the pivot rule
   1.402 -        const double BLOCK_SIZE_FACTOR = 1.5;
   1.403 +        const double BLOCK_SIZE_FACTOR = 1.0;
   1.404          const int MIN_BLOCK_SIZE = 10;
   1.405          const double HEAD_LENGTH_FACTOR = 0.1;
   1.406          const int MIN_HEAD_LENGTH = 3;
   1.407 @@ -567,7 +570,7 @@
   1.408        bool findEnteringArc() {
   1.409          // Check the current candidate list
   1.410          int e;
   1.411 -        for (int i = 0; i < _curr_length; ++i) {
   1.412 +        for (int i = 0; i != _curr_length; ++i) {
   1.413            e = _candidates[i];
   1.414            _cand_cost[e] = _state[e] *
   1.415              (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.416 @@ -578,39 +581,35 @@
   1.417  
   1.418          // Extend the list
   1.419          int cnt = _block_size;
   1.420 -        int last_arc = 0;
   1.421          int limit = _head_length;
   1.422  
   1.423 -        for (int e = _next_arc; e < _search_arc_num; ++e) {
   1.424 +        for (e = _next_arc; e != _search_arc_num; ++e) {
   1.425            _cand_cost[e] = _state[e] *
   1.426              (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.427            if (_cand_cost[e] < 0) {
   1.428              _candidates[_curr_length++] = e;
   1.429 -            last_arc = e;
   1.430            }
   1.431            if (--cnt == 0) {
   1.432 -            if (_curr_length > limit) break;
   1.433 +            if (_curr_length > limit) goto search_end;
   1.434              limit = 0;
   1.435              cnt = _block_size;
   1.436            }
   1.437          }
   1.438 -        if (_curr_length <= limit) {
   1.439 -          for (int e = 0; e < _next_arc; ++e) {
   1.440 -            _cand_cost[e] = _state[e] *
   1.441 -              (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.442 -            if (_cand_cost[e] < 0) {
   1.443 -              _candidates[_curr_length++] = e;
   1.444 -              last_arc = e;
   1.445 -            }
   1.446 -            if (--cnt == 0) {
   1.447 -              if (_curr_length > limit) break;
   1.448 -              limit = 0;
   1.449 -              cnt = _block_size;
   1.450 -            }
   1.451 +        for (e = 0; e != _next_arc; ++e) {
   1.452 +          _cand_cost[e] = _state[e] *
   1.453 +            (_cost[e] + _pi[_source[e]] - _pi[_target[e]]);
   1.454 +          if (_cand_cost[e] < 0) {
   1.455 +            _candidates[_curr_length++] = e;
   1.456 +          }
   1.457 +          if (--cnt == 0) {
   1.458 +            if (_curr_length > limit) goto search_end;
   1.459 +            limit = 0;
   1.460 +            cnt = _block_size;
   1.461            }
   1.462          }
   1.463          if (_curr_length == 0) return false;
   1.464 -        _next_arc = last_arc + 1;
   1.465 +
   1.466 +      search_end:
   1.467  
   1.468          // Make heap of the candidate list (approximating a partial sort)
   1.469          make_heap( _candidates.begin(), _candidates.begin() + _curr_length,
   1.470 @@ -618,6 +617,7 @@
   1.471  
   1.472          // Pop the first element of the heap
   1.473          _in_arc = _candidates[0];
   1.474 +        _next_arc = e;
   1.475          pop_heap( _candidates.begin(), _candidates.begin() + _curr_length,
   1.476                    _sort_func );
   1.477          _curr_length = std::min(_head_length, _curr_length - 1);
   1.478 @@ -633,69 +633,25 @@
   1.479      /// The constructor of the class.
   1.480      ///
   1.481      /// \param graph The digraph the algorithm runs on.
   1.482 -    NetworkSimplex(const GR& graph) :
   1.483 +    /// \param arc_mixing Indicate if the arcs have to be stored in a
   1.484 +    /// mixed order in the internal data structure.
   1.485 +    /// In special cases, it could lead to better overall performance,
   1.486 +    /// but it is usually slower. Therefore it is disabled by default.
   1.487 +    NetworkSimplex(const GR& graph, bool arc_mixing = false) :
   1.488        _graph(graph), _node_id(graph), _arc_id(graph),
   1.489 +      _arc_mixing(arc_mixing),
   1.490 +      MAX(std::numeric_limits<Value>::max()),
   1.491        INF(std::numeric_limits<Value>::has_infinity ?
   1.492 -          std::numeric_limits<Value>::infinity() :
   1.493 -          std::numeric_limits<Value>::max())
   1.494 +          std::numeric_limits<Value>::infinity() : MAX)
   1.495      {
   1.496 -      // Check the value types
   1.497 +      // Check the number types
   1.498        LEMON_ASSERT(std::numeric_limits<Value>::is_signed,
   1.499          "The flow type of NetworkSimplex must be signed");
   1.500        LEMON_ASSERT(std::numeric_limits<Cost>::is_signed,
   1.501          "The cost type of NetworkSimplex must be signed");
   1.502 -        
   1.503 -      // Resize vectors
   1.504 -      _node_num = countNodes(_graph);
   1.505 -      _arc_num = countArcs(_graph);
   1.506 -      int all_node_num = _node_num + 1;
   1.507 -      int max_arc_num = _arc_num + 2 * _node_num;
   1.508  
   1.509 -      _source.resize(max_arc_num);
   1.510 -      _target.resize(max_arc_num);
   1.511 -
   1.512 -      _lower.resize(_arc_num);
   1.513 -      _upper.resize(_arc_num);
   1.514 -      _cap.resize(max_arc_num);
   1.515 -      _cost.resize(max_arc_num);
   1.516 -      _supply.resize(all_node_num);
   1.517 -      _flow.resize(max_arc_num);
   1.518 -      _pi.resize(all_node_num);
   1.519 -
   1.520 -      _parent.resize(all_node_num);
   1.521 -      _pred.resize(all_node_num);
   1.522 -      _forward.resize(all_node_num);
   1.523 -      _thread.resize(all_node_num);
   1.524 -      _rev_thread.resize(all_node_num);
   1.525 -      _succ_num.resize(all_node_num);
   1.526 -      _last_succ.resize(all_node_num);
   1.527 -      _state.resize(max_arc_num);
   1.528 -
   1.529 -      // Copy the graph (store the arcs in a mixed order)
   1.530 -      int i = 0;
   1.531 -      for (NodeIt n(_graph); n != INVALID; ++n, ++i) {
   1.532 -        _node_id[n] = i;
   1.533 -      }
   1.534 -      int k = std::max(int(std::sqrt(double(_arc_num))), 10);
   1.535 -      i = 0;
   1.536 -      for (ArcIt a(_graph); a != INVALID; ++a) {
   1.537 -        _arc_id[a] = i;
   1.538 -        _source[i] = _node_id[_graph.source(a)];
   1.539 -        _target[i] = _node_id[_graph.target(a)];
   1.540 -        if ((i += k) >= _arc_num) i = (i % k) + 1;
   1.541 -      }
   1.542 -      
   1.543 -      // Initialize maps
   1.544 -      for (int i = 0; i != _node_num; ++i) {
   1.545 -        _supply[i] = 0;
   1.546 -      }
   1.547 -      for (int i = 0; i != _arc_num; ++i) {
   1.548 -        _lower[i] = 0;
   1.549 -        _upper[i] = INF;
   1.550 -        _cost[i] = 1;
   1.551 -      }
   1.552 -      _have_lower = false;
   1.553 -      _stype = GEQ;
   1.554 +      // Reset data structures
   1.555 +      reset();
   1.556      }
   1.557  
   1.558      /// \name Parameters
   1.559 @@ -729,7 +685,7 @@
   1.560      /// This function sets the upper bounds (capacities) on the arcs.
   1.561      /// If it is not used before calling \ref run(), the upper bounds
   1.562      /// will be set to \ref INF on all arcs (i.e. the flow value will be
   1.563 -    /// unbounded from above on each arc).
   1.564 +    /// unbounded from above).
   1.565      ///
   1.566      /// \param map An arc map storing the upper bounds.
   1.567      /// Its \c Value type must be convertible to the \c Value type
   1.568 @@ -768,7 +724,6 @@
   1.569      /// This function sets the supply values of the nodes.
   1.570      /// If neither this function nor \ref stSupply() is used before
   1.571      /// calling \ref run(), the supply of each node will be set to zero.
   1.572 -    /// (It makes sense only if non-zero lower bounds are given.)
   1.573      ///
   1.574      /// \param map A node map storing the supply values.
   1.575      /// Its \c Value type must be convertible to the \c Value type
   1.576 @@ -789,7 +744,6 @@
   1.577      /// and the required flow value.
   1.578      /// If neither this function nor \ref supplyMap() is used before
   1.579      /// calling \ref run(), the supply of each node will be set to zero.
   1.580 -    /// (It makes sense only if non-zero lower bounds are given.)
   1.581      ///
   1.582      /// Using this function has the same effect as using \ref supplyMap()
   1.583      /// with such a map in which \c k is assigned to \c s, \c -k is
   1.584 @@ -809,14 +763,14 @@
   1.585        _supply[_node_id[t]] = -k;
   1.586        return *this;
   1.587      }
   1.588 -    
   1.589 +
   1.590      /// \brief Set the type of the supply constraints.
   1.591      ///
   1.592      /// This function sets the type of the supply/demand constraints.
   1.593      /// If it is not used before calling \ref run(), the \ref GEQ supply
   1.594      /// type will be used.
   1.595      ///
   1.596 -    /// For more information see \ref SupplyType.
   1.597 +    /// For more information, see \ref SupplyType.
   1.598      ///
   1.599      /// \return <tt>(*this)</tt>
   1.600      NetworkSimplex& supplyType(SupplyType supply_type) {
   1.601 @@ -835,7 +789,7 @@
   1.602      ///
   1.603      /// This function runs the algorithm.
   1.604      /// The paramters can be specified using functions \ref lowerMap(),
   1.605 -    /// \ref upperMap(), \ref costMap(), \ref supplyMap(), \ref stSupply(), 
   1.606 +    /// \ref upperMap(), \ref costMap(), \ref supplyMap(), \ref stSupply(),
   1.607      /// \ref supplyType().
   1.608      /// For example,
   1.609      /// \code
   1.610 @@ -844,15 +798,15 @@
   1.611      ///     .supplyMap(sup).run();
   1.612      /// \endcode
   1.613      ///
   1.614 -    /// This function can be called more than once. All the parameters
   1.615 -    /// that have been given are kept for the next call, unless
   1.616 -    /// \ref reset() is called, thus only the modified parameters
   1.617 -    /// have to be set again. See \ref reset() for examples.
   1.618 -    /// However the underlying digraph must not be modified after this
   1.619 -    /// class have been constructed, since it copies and extends the graph.
   1.620 +    /// This function can be called more than once. All the given parameters
   1.621 +    /// are kept for the next call, unless \ref resetParams() or \ref reset()
   1.622 +    /// is used, thus only the modified parameters have to be set again.
   1.623 +    /// If the underlying digraph was also modified after the construction
   1.624 +    /// of the class (or the last \ref reset() call), then the \ref reset()
   1.625 +    /// function must be called.
   1.626      ///
   1.627      /// \param pivot_rule The pivot rule that will be used during the
   1.628 -    /// algorithm. For more information see \ref PivotRule.
   1.629 +    /// algorithm. For more information, see \ref PivotRule.
   1.630      ///
   1.631      /// \return \c INFEASIBLE if no feasible flow exists,
   1.632      /// \n \c OPTIMAL if the problem has optimal solution
   1.633 @@ -863,6 +817,7 @@
   1.634      /// cost and infinite upper bound.
   1.635      ///
   1.636      /// \see ProblemType, PivotRule
   1.637 +    /// \see resetParams(), reset()
   1.638      ProblemType run(PivotRule pivot_rule = BLOCK_SEARCH) {
   1.639        if (!init()) return INFEASIBLE;
   1.640        return start(pivot_rule);
   1.641 @@ -874,11 +829,12 @@
   1.642      /// before using functions \ref lowerMap(), \ref upperMap(),
   1.643      /// \ref costMap(), \ref supplyMap(), \ref stSupply(), \ref supplyType().
   1.644      ///
   1.645 -    /// It is useful for multiple run() calls. If this function is not
   1.646 -    /// used, all the parameters given before are kept for the next
   1.647 -    /// \ref run() call.
   1.648 -    /// However the underlying digraph must not be modified after this
   1.649 -    /// class have been constructed, since it copies and extends the graph.
   1.650 +    /// It is useful for multiple \ref run() calls. Basically, all the given
   1.651 +    /// parameters are kept for the next \ref run() call, unless
   1.652 +    /// \ref resetParams() or \ref reset() is used.
   1.653 +    /// If the underlying digraph was also modified after the construction
   1.654 +    /// of the class or the last \ref reset() call, then the \ref reset()
   1.655 +    /// function must be used, otherwise \ref resetParams() is sufficient.
   1.656      ///
   1.657      /// For example,
   1.658      /// \code
   1.659 @@ -888,20 +844,22 @@
   1.660      ///   ns.lowerMap(lower).upperMap(upper).costMap(cost)
   1.661      ///     .supplyMap(sup).run();
   1.662      ///
   1.663 -    ///   // Run again with modified cost map (reset() is not called,
   1.664 +    ///   // Run again with modified cost map (resetParams() is not called,
   1.665      ///   // so only the cost map have to be set again)
   1.666      ///   cost[e] += 100;
   1.667      ///   ns.costMap(cost).run();
   1.668      ///
   1.669 -    ///   // Run again from scratch using reset()
   1.670 +    ///   // Run again from scratch using resetParams()
   1.671      ///   // (the lower bounds will be set to zero on all arcs)
   1.672 -    ///   ns.reset();
   1.673 +    ///   ns.resetParams();
   1.674      ///   ns.upperMap(capacity).costMap(cost)
   1.675      ///     .supplyMap(sup).run();
   1.676      /// \endcode
   1.677      ///
   1.678      /// \return <tt>(*this)</tt>
   1.679 -    NetworkSimplex& reset() {
   1.680 +    ///
   1.681 +    /// \see reset(), run()
   1.682 +    NetworkSimplex& resetParams() {
   1.683        for (int i = 0; i != _node_num; ++i) {
   1.684          _supply[i] = 0;
   1.685        }
   1.686 @@ -915,6 +873,83 @@
   1.687        return *this;
   1.688      }
   1.689  
   1.690 +    /// \brief Reset the internal data structures and all the parameters
   1.691 +    /// that have been given before.
   1.692 +    ///
   1.693 +    /// This function resets the internal data structures and all the
   1.694 +    /// paramaters that have been given before using functions \ref lowerMap(),
   1.695 +    /// \ref upperMap(), \ref costMap(), \ref supplyMap(), \ref stSupply(),
   1.696 +    /// \ref supplyType().
   1.697 +    ///
   1.698 +    /// It is useful for multiple \ref run() calls. Basically, all the given
   1.699 +    /// parameters are kept for the next \ref run() call, unless
   1.700 +    /// \ref resetParams() or \ref reset() is used.
   1.701 +    /// If the underlying digraph was also modified after the construction
   1.702 +    /// of the class or the last \ref reset() call, then the \ref reset()
   1.703 +    /// function must be used, otherwise \ref resetParams() is sufficient.
   1.704 +    ///
   1.705 +    /// See \ref resetParams() for examples.
   1.706 +    ///
   1.707 +    /// \return <tt>(*this)</tt>
   1.708 +    ///
   1.709 +    /// \see resetParams(), run()
   1.710 +    NetworkSimplex& reset() {
   1.711 +      // Resize vectors
   1.712 +      _node_num = countNodes(_graph);
   1.713 +      _arc_num = countArcs(_graph);
   1.714 +      int all_node_num = _node_num + 1;
   1.715 +      int max_arc_num = _arc_num + 2 * _node_num;
   1.716 +
   1.717 +      _source.resize(max_arc_num);
   1.718 +      _target.resize(max_arc_num);
   1.719 +
   1.720 +      _lower.resize(_arc_num);
   1.721 +      _upper.resize(_arc_num);
   1.722 +      _cap.resize(max_arc_num);
   1.723 +      _cost.resize(max_arc_num);
   1.724 +      _supply.resize(all_node_num);
   1.725 +      _flow.resize(max_arc_num);
   1.726 +      _pi.resize(all_node_num);
   1.727 +
   1.728 +      _parent.resize(all_node_num);
   1.729 +      _pred.resize(all_node_num);
   1.730 +      _forward.resize(all_node_num);
   1.731 +      _thread.resize(all_node_num);
   1.732 +      _rev_thread.resize(all_node_num);
   1.733 +      _succ_num.resize(all_node_num);
   1.734 +      _last_succ.resize(all_node_num);
   1.735 +      _state.resize(max_arc_num);
   1.736 +
   1.737 +      // Copy the graph
   1.738 +      int i = 0;
   1.739 +      for (NodeIt n(_graph); n != INVALID; ++n, ++i) {
   1.740 +        _node_id[n] = i;
   1.741 +      }
   1.742 +      if (_arc_mixing) {
   1.743 +        // Store the arcs in a mixed order
   1.744 +        int k = std::max(int(std::sqrt(double(_arc_num))), 10);
   1.745 +        int i = 0, j = 0;
   1.746 +        for (ArcIt a(_graph); a != INVALID; ++a) {
   1.747 +          _arc_id[a] = i;
   1.748 +          _source[i] = _node_id[_graph.source(a)];
   1.749 +          _target[i] = _node_id[_graph.target(a)];
   1.750 +          if ((i += k) >= _arc_num) i = ++j;
   1.751 +        }
   1.752 +      } else {
   1.753 +        // Store the arcs in the original order
   1.754 +        int i = 0;
   1.755 +        for (ArcIt a(_graph); a != INVALID; ++a, ++i) {
   1.756 +          _arc_id[a] = i;
   1.757 +          _source[i] = _node_id[_graph.source(a)];
   1.758 +          _target[i] = _node_id[_graph.target(a)];
   1.759 +        }
   1.760 +      }
   1.761 +
   1.762 +      // Reset parameters
   1.763 +      resetParams();
   1.764 +      return *this;
   1.765 +    }
   1.766 +
   1.767      /// @}
   1.768  
   1.769      /// \name Query Functions
   1.770 @@ -1024,9 +1059,9 @@
   1.771          for (int i = 0; i != _arc_num; ++i) {
   1.772            Value c = _lower[i];
   1.773            if (c >= 0) {
   1.774 -            _cap[i] = _upper[i] < INF ? _upper[i] - c : INF;
   1.775 +            _cap[i] = _upper[i] < MAX ? _upper[i] - c : INF;
   1.776            } else {
   1.777 -            _cap[i] = _upper[i] < INF + c ? _upper[i] - c : INF;
   1.778 +            _cap[i] = _upper[i] < MAX + c ? _upper[i] - c : INF;
   1.779            }
   1.780            _supply[_source[i]] -= c;
   1.781            _supply[_target[i]] += c;
   1.782 @@ -1054,7 +1089,7 @@
   1.783          _flow[i] = 0;
   1.784          _state[i] = STATE_LOWER;
   1.785        }
   1.786 -      
   1.787 +
   1.788        // Set data for the artificial root node
   1.789        _root = _node_num;
   1.790        _parent[_root] = -1;
   1.791 @@ -1218,7 +1253,7 @@
   1.792        for (int u = first; u != join; u = _parent[u]) {
   1.793          e = _pred[u];
   1.794          d = _forward[u] ?
   1.795 -          _flow[e] : (_cap[e] == INF ? INF : _cap[e] - _flow[e]);
   1.796 +          _flow[e] : (_cap[e] >= MAX ? INF : _cap[e] - _flow[e]);
   1.797          if (d < delta) {
   1.798            delta = d;
   1.799            u_out = u;
   1.800 @@ -1228,8 +1263,8 @@
   1.801        // Search the cycle along the path form the second node to the root
   1.802        for (int u = second; u != join; u = _parent[u]) {
   1.803          e = _pred[u];
   1.804 -        d = _forward[u] ? 
   1.805 -          (_cap[e] == INF ? INF : _cap[e] - _flow[e]) : _flow[e];
   1.806 +        d = _forward[u] ?
   1.807 +          (_cap[e] >= MAX ? INF : _cap[e] - _flow[e]) : _flow[e];
   1.808          if (d <= delta) {
   1.809            delta = d;
   1.810            u_out = u;
   1.811 @@ -1330,7 +1365,7 @@
   1.812        }
   1.813  
   1.814        // Update _rev_thread using the new _thread values
   1.815 -      for (int i = 0; i < int(_dirty_revs.size()); ++i) {
   1.816 +      for (int i = 0; i != int(_dirty_revs.size()); ++i) {
   1.817          u = _dirty_revs[i];
   1.818          _rev_thread[_thread[u]] = u;
   1.819        }
   1.820 @@ -1402,6 +1437,100 @@
   1.821        }
   1.822      }
   1.823  
   1.824 +    // Heuristic initial pivots
   1.825 +    bool initialPivots() {
   1.826 +      Value curr, total = 0;
   1.827 +      std::vector<Node> supply_nodes, demand_nodes;
   1.828 +      for (NodeIt u(_graph); u != INVALID; ++u) {
   1.829 +        curr = _supply[_node_id[u]];
   1.830 +        if (curr > 0) {
   1.831 +          total += curr;
   1.832 +          supply_nodes.push_back(u);
   1.833 +        }
   1.834 +        else if (curr < 0) {
   1.835 +          demand_nodes.push_back(u);
   1.836 +        }
   1.837 +      }
   1.838 +      if (_sum_supply > 0) total -= _sum_supply;
   1.839 +      if (total <= 0) return true;
   1.840 +
   1.841 +      IntVector arc_vector;
   1.842 +      if (_sum_supply >= 0) {
   1.843 +        if (supply_nodes.size() == 1 && demand_nodes.size() == 1) {
   1.844 +          // Perform a reverse graph search from the sink to the source
   1.845 +          typename GR::template NodeMap<bool> reached(_graph, false);
   1.846 +          Node s = supply_nodes[0], t = demand_nodes[0];
   1.847 +          std::vector<Node> stack;
   1.848 +          reached[t] = true;
   1.849 +          stack.push_back(t);
   1.850 +          while (!stack.empty()) {
   1.851 +            Node u, v = stack.back();
   1.852 +            stack.pop_back();
   1.853 +            if (v == s) break;
   1.854 +            for (InArcIt a(_graph, v); a != INVALID; ++a) {
   1.855 +              if (reached[u = _graph.source(a)]) continue;
   1.856 +              int j = _arc_id[a];
   1.857 +              if (_cap[j] >= total) {
   1.858 +                arc_vector.push_back(j);
   1.859 +                reached[u] = true;
   1.860 +                stack.push_back(u);
   1.861 +              }
   1.862 +            }
   1.863 +          }
   1.864 +        } else {
   1.865 +          // Find the min. cost incomming arc for each demand node
   1.866 +          for (int i = 0; i != int(demand_nodes.size()); ++i) {
   1.867 +            Node v = demand_nodes[i];
   1.868 +            Cost c, min_cost = std::numeric_limits<Cost>::max();
   1.869 +            Arc min_arc = INVALID;
   1.870 +            for (InArcIt a(_graph, v); a != INVALID; ++a) {
   1.871 +              c = _cost[_arc_id[a]];
   1.872 +              if (c < min_cost) {
   1.873 +                min_cost = c;
   1.874 +                min_arc = a;
   1.875 +              }
   1.876 +            }
   1.877 +            if (min_arc != INVALID) {
   1.878 +              arc_vector.push_back(_arc_id[min_arc]);
   1.879 +            }
   1.880 +          }
   1.881 +        }
   1.882 +      } else {
   1.883 +        // Find the min. cost outgoing arc for each supply node
   1.884 +        for (int i = 0; i != int(supply_nodes.size()); ++i) {
   1.885 +          Node u = supply_nodes[i];
   1.886 +          Cost c, min_cost = std::numeric_limits<Cost>::max();
   1.887 +          Arc min_arc = INVALID;
   1.888 +          for (OutArcIt a(_graph, u); a != INVALID; ++a) {
   1.889 +            c = _cost[_arc_id[a]];
   1.890 +            if (c < min_cost) {
   1.891 +              min_cost = c;
   1.892 +              min_arc = a;
   1.893 +            }
   1.894 +          }
   1.895 +          if (min_arc != INVALID) {
   1.896 +            arc_vector.push_back(_arc_id[min_arc]);
   1.897 +          }
   1.898 +        }
   1.899 +      }
   1.900 +
   1.901 +      // Perform heuristic initial pivots
   1.902 +      for (int i = 0; i != int(arc_vector.size()); ++i) {
   1.903 +        in_arc = arc_vector[i];
   1.904 +        if (_state[in_arc] * (_cost[in_arc] + _pi[_source[in_arc]] -
   1.905 +            _pi[_target[in_arc]]) >= 0) continue;
   1.906 +        findJoinNode();
   1.907 +        bool change = findLeavingArc();
   1.908 +        if (delta >= MAX) return false;
   1.909 +        changeFlow(change);
   1.910 +        if (change) {
   1.911 +          updateTreeStructure();
   1.912 +          updatePotential();
   1.913 +        }
   1.914 +      }
   1.915 +      return true;
   1.916 +    }
   1.917 +
   1.918      // Execute the algorithm
   1.919      ProblemType start(PivotRule pivot_rule) {
   1.920        // Select the pivot rule implementation
   1.921 @@ -1424,18 +1553,21 @@
   1.922      ProblemType start() {
   1.923        PivotRuleImpl pivot(*this);
   1.924  
   1.925 +      // Perform heuristic initial pivots
   1.926 +      if (!initialPivots()) return UNBOUNDED;
   1.927 +
   1.928        // Execute the Network Simplex algorithm
   1.929        while (pivot.findEnteringArc()) {
   1.930          findJoinNode();
   1.931          bool change = findLeavingArc();
   1.932 -        if (delta >= INF) return UNBOUNDED;
   1.933 +        if (delta >= MAX) return UNBOUNDED;
   1.934          changeFlow(change);
   1.935          if (change) {
   1.936            updateTreeStructure();
   1.937            updatePotential();
   1.938          }
   1.939        }
   1.940 -      
   1.941 +
   1.942        // Check feasibility
   1.943        for (int e = _search_arc_num; e != _all_arc_num; ++e) {
   1.944          if (_flow[e] != 0) return INFEASIBLE;
   1.945 @@ -1452,7 +1584,7 @@
   1.946            }
   1.947          }
   1.948        }
   1.949 -      
   1.950 +
   1.951        // Shift potentials to meet the requirements of the GEQ/LEQ type
   1.952        // optimality conditions
   1.953        if (_sum_supply == 0) {