gravatar
kpeter (Peter Kovacs)
kpeter@inf.elte.hu
More options for run() in scaling MCF algorithms (#180) - Three methods can be selected and the scaling factor can be given for CostScaling. - The scaling factor can be given for CapacityScaling.
0 2 0
default
2 files changed with 71 insertions and 42 deletions:
↑ Collapse diff ↑
Ignore white space 24 line context
... ...
@@ -164,25 +164,25 @@
164 164
    ValueVector _lower;
165 165
    ValueVector _upper;
166 166
    CostVector _cost;
167 167
    ValueVector _supply;
168 168

	
169 169
    ValueVector _res_cap;
170 170
    CostVector _pi;
171 171
    ValueVector _excess;
172 172
    IntVector _excess_nodes;
173 173
    IntVector _deficit_nodes;
174 174

	
175 175
    Value _delta;
176
    int _phase_num;
176
    int _factor;
177 177
    IntVector _pred;
178 178

	
179 179
  public:
180 180
  
181 181
    /// \brief Constant for infinite upper bounds (capacities).
182 182
    ///
183 183
    /// Constant for infinite upper bounds (capacities).
184 184
    /// It is \c std::numeric_limits<Value>::infinity() if available,
185 185
    /// \c std::numeric_limits<Value>::max() otherwise.
186 186
    const Value INF;
187 187

	
188 188
  private:
... ...
@@ -504,58 +504,58 @@
504 504
    /// For example,
505 505
    /// \code
506 506
    ///   CapacityScaling<ListDigraph> cs(graph);
507 507
    ///   cs.lowerMap(lower).upperMap(upper).costMap(cost)
508 508
    ///     .supplyMap(sup).run();
509 509
    /// \endcode
510 510
    ///
511 511
    /// This function can be called more than once. All the parameters
512 512
    /// that have been given are kept for the next call, unless
513 513
    /// \ref reset() is called, thus only the modified parameters
514 514
    /// have to be set again. See \ref reset() for examples.
515 515
    /// However the underlying digraph must not be modified after this
516
    /// class have been constructed, since it copies the digraph.
516
    /// class have been constructed, since it copies and extends the graph.
517 517
    ///
518
    /// \param scaling Enable or disable capacity scaling.
519
    /// If the maximum upper bound and/or the amount of total supply
520
    /// is rather small, the algorithm could be slightly faster without
521
    /// scaling.
518
    /// \param factor The capacity scaling factor. It must be larger than
519
    /// one to use scaling. If it is less or equal to one, then scaling
520
    /// will be disabled.
522 521
    ///
523 522
    /// \return \c INFEASIBLE if no feasible flow exists,
524 523
    /// \n \c OPTIMAL if the problem has optimal solution
525 524
    /// (i.e. it is feasible and bounded), and the algorithm has found
526 525
    /// optimal flow and node potentials (primal and dual solutions),
527 526
    /// \n \c UNBOUNDED if the digraph contains an arc of negative cost
528 527
    /// and infinite upper bound. It means that the objective function
529 528
    /// is unbounded on that arc, however note that it could actually be
530 529
    /// bounded over the feasible flows, but this algroithm cannot handle
531 530
    /// these cases.
532 531
    ///
533 532
    /// \see ProblemType
534
    ProblemType run(bool scaling = true) {
535
      ProblemType pt = init(scaling);
533
    ProblemType run(int factor = 4) {
534
      _factor = factor;
535
      ProblemType pt = init();
536 536
      if (pt != OPTIMAL) return pt;
537 537
      return start();
538 538
    }
539 539

	
540 540
    /// \brief Reset all the parameters that have been given before.
541 541
    ///
542 542
    /// This function resets all the paramaters that have been given
543 543
    /// before using functions \ref lowerMap(), \ref upperMap(),
544 544
    /// \ref costMap(), \ref supplyMap(), \ref stSupply().
545 545
    ///
546 546
    /// It is useful for multiple run() calls. If this function is not
547 547
    /// used, all the parameters given before are kept for the next
548 548
    /// \ref run() call.
549
    /// However the underlying digraph must not be modified after this
549
    /// However, the underlying digraph must not be modified after this
550 550
    /// class have been constructed, since it copies and extends the graph.
551 551
    ///
552 552
    /// For example,
553 553
    /// \code
554 554
    ///   CapacityScaling<ListDigraph> cs(graph);
555 555
    ///
556 556
    ///   // First run
557 557
    ///   cs.lowerMap(lower).upperMap(upper).costMap(cost)
558 558
    ///     .supplyMap(sup).run();
559 559
    ///
560 560
    ///   // Run again with modified cost map (reset() is not called,
561 561
    ///   // so only the cost map have to be set again)
... ...
@@ -668,25 +668,25 @@
668 668
    template <typename PotentialMap>
669 669
    void potentialMap(PotentialMap &map) const {
670 670
      for (NodeIt n(_graph); n != INVALID; ++n) {
671 671
        map.set(n, _pi[_node_id[n]]);
672 672
      }
673 673
    }
674 674

	
675 675
    /// @}
676 676

	
677 677
  private:
678 678

	
679 679
    // Initialize the algorithm
680
    ProblemType init(bool scaling) {
680
    ProblemType init() {
681 681
      if (_node_num == 0) return INFEASIBLE;
682 682

	
683 683
      // Check the sum of supply values
684 684
      _sum_supply = 0;
685 685
      for (int i = 0; i != _root; ++i) {
686 686
        _sum_supply += _supply[i];
687 687
      }
688 688
      if (_sum_supply > 0) return INFEASIBLE;
689 689
      
690 690
      // Initialize maps
691 691
      for (int i = 0; i != _root; ++i) {
692 692
        _pi[i] = 0;
... ...
@@ -749,39 +749,37 @@
749 749
      } else {
750 750
        _pi[_root] = 0;
751 751
        _excess[_root] = 0;
752 752
        for (int a = _first_out[_root]; a != _res_arc_num; ++a) {
753 753
          _res_cap[a] = 1;
754 754
          _res_cap[_reverse[a]] = 0;
755 755
          _cost[a] = 0;
756 756
          _cost[_reverse[a]] = 0;
757 757
        }
758 758
      }
759 759

	
760 760
      // Initialize delta value
761
      if (scaling) {
761
      if (_factor > 1) {
762 762
        // With scaling
763 763
        Value max_sup = 0, max_dem = 0;
764 764
        for (int i = 0; i != _node_num; ++i) {
765 765
          if ( _excess[i] > max_sup) max_sup =  _excess[i];
766 766
          if (-_excess[i] > max_dem) max_dem = -_excess[i];
767 767
        }
768 768
        Value max_cap = 0;
769 769
        for (int j = 0; j != _res_arc_num; ++j) {
770 770
          if (_res_cap[j] > max_cap) max_cap = _res_cap[j];
771 771
        }
772 772
        max_sup = std::min(std::min(max_sup, max_dem), max_cap);
773
        _phase_num = 0;
774
        for (_delta = 1; 2 * _delta <= max_sup; _delta *= 2)
775
          ++_phase_num;
773
        for (_delta = 1; 2 * _delta <= max_sup; _delta *= 2) ;
776 774
      } else {
777 775
        // Without scaling
778 776
        _delta = 1;
779 777
      }
780 778

	
781 779
      return OPTIMAL;
782 780
    }
783 781

	
784 782
    ProblemType start() {
785 783
      // Execute the algorithm
786 784
      ProblemType pt;
787 785
      if (_delta > 1)
... ...
@@ -802,26 +800,24 @@
802 800
        for (int i = 0; i != _node_num; ++i) {
803 801
          _pi[i] -= pr;
804 802
        }        
805 803
      }
806 804
      
807 805
      return pt;
808 806
    }
809 807

	
810 808
    // Execute the capacity scaling algorithm
811 809
    ProblemType startWithScaling() {
812 810
      // Perform capacity scaling phases
813 811
      int s, t;
814
      int phase_cnt = 0;
815
      int factor = 4;
816 812
      ResidualDijkstra _dijkstra(*this);
817 813
      while (true) {
818 814
        // Saturate all arcs not satisfying the optimality condition
819 815
        for (int u = 0; u != _node_num; ++u) {
820 816
          for (int a = _first_out[u]; a != _first_out[u+1]; ++a) {
821 817
            int v = _target[a];
822 818
            Cost c = _cost[a] + _pi[u] - _pi[v];
823 819
            Value rc = _res_cap[a];
824 820
            if (c < 0 && rc >= _delta) {
825 821
              _excess[u] -= rc;
826 822
              _excess[v] += rc;
827 823
              _res_cap[a] = 0;
... ...
@@ -878,26 +874,25 @@
878 874
          while ((a = _pred[u]) != -1) {
879 875
            _res_cap[a] -= d;
880 876
            _res_cap[_reverse[a]] += d;
881 877
            u = _source[a];
882 878
          }
883 879
          _excess[s] -= d;
884 880
          _excess[t] += d;
885 881

	
886 882
          if (_excess[s] < _delta) ++next_node;
887 883
        }
888 884

	
889 885
        if (_delta == 1) break;
890
        if (++phase_cnt == _phase_num / 4) factor = 2;
891
        _delta = _delta <= factor ? 1 : _delta / factor;
886
        _delta = _delta <= _factor ? 1 : _delta / _factor;
892 887
      }
893 888

	
894 889
      return OPTIMAL;
895 890
    }
896 891

	
897 892
    // Execute the successive shortest path algorithm
898 893
    ProblemType startWithoutScaling() {
899 894
      // Find excess nodes
900 895
      _excess_nodes.clear();
901 896
      for (int i = 0; i != _node_num; ++i) {
902 897
        if (_excess[i] > 0) _excess_nodes.push_back(i);
903 898
      }
Ignore white space 6 line context
... ...
@@ -101,24 +101,28 @@
101 101
  /// specified, then default values will be used.
102 102
  ///
103 103
  /// \tparam GR The digraph type the algorithm runs on.
104 104
  /// \tparam V The value type used for flow amounts, capacity bounds
105 105
  /// and supply values in the algorithm. By default it is \c int.
106 106
  /// \tparam C The value type used for costs and potentials in the
107 107
  /// algorithm. By default it is the same as \c V.
108 108
  ///
109 109
  /// \warning Both value types must be signed and all input data must
110 110
  /// be integer.
111 111
  /// \warning This algorithm does not support negative costs for such
112 112
  /// arcs that have infinite upper bound.
113
  ///
114
  /// \note %CostScaling provides three different internal methods,
115
  /// from which the most efficient one is used by default.
116
  /// For more information, see \ref Method.
113 117
#ifdef DOXYGEN
114 118
  template <typename GR, typename V, typename C, typename TR>
115 119
#else
116 120
  template < typename GR, typename V = int, typename C = V,
117 121
             typename TR = CostScalingDefaultTraits<GR, V, C> >
118 122
#endif
119 123
  class CostScaling
120 124
  {
121 125
  public:
122 126

	
123 127
    /// The type of the digraph
124 128
    typedef typename TR::Digraph Digraph;
... ...
@@ -150,24 +154,51 @@
150 154
      /// The problem has optimal solution (i.e. it is feasible and
151 155
      /// bounded), and the algorithm has found optimal flow and node
152 156
      /// potentials (primal and dual solutions).
153 157
      OPTIMAL,
154 158
      /// The digraph contains an arc of negative cost and infinite
155 159
      /// upper bound. It means that the objective function is unbounded
156 160
      /// on that arc, however note that it could actually be bounded
157 161
      /// over the feasible flows, but this algroithm cannot handle
158 162
      /// these cases.
159 163
      UNBOUNDED
160 164
    };
161 165

	
166
    /// \brief Constants for selecting the internal method.
167
    ///
168
    /// Enum type containing constants for selecting the internal method
169
    /// for the \ref run() function.
170
    ///
171
    /// \ref CostScaling provides three internal methods that differ mainly
172
    /// in their base operations, which are used in conjunction with the
173
    /// relabel operation.
174
    /// By default, the so called \ref PARTIAL_AUGMENT
175
    /// "Partial Augment-Relabel" method is used, which proved to be
176
    /// the most efficient and the most robust on various test inputs.
177
    /// However, the other methods can be selected using the \ref run()
178
    /// function with the proper parameter.
179
    enum Method {
180
      /// Local push operations are used, i.e. flow is moved only on one
181
      /// admissible arc at once.
182
      PUSH,
183
      /// Augment operations are used, i.e. flow is moved on admissible
184
      /// paths from a node with excess to a node with deficit.
185
      AUGMENT,
186
      /// Partial augment operations are used, i.e. flow is moved on 
187
      /// admissible paths started from a node with excess, but the
188
      /// lengths of these paths are limited. This method can be viewed
189
      /// as a combined version of the previous two operations.
190
      PARTIAL_AUGMENT
191
    };
192

	
162 193
  private:
163 194

	
164 195
    TEMPLATE_DIGRAPH_TYPEDEFS(GR);
165 196

	
166 197
    typedef std::vector<int> IntVector;
167 198
    typedef std::vector<char> BoolVector;
168 199
    typedef std::vector<Value> ValueVector;
169 200
    typedef std::vector<Cost> CostVector;
170 201
    typedef std::vector<LargeCost> LargeCostVector;
171 202

	
172 203
  private:
173 204
  
... ...
@@ -496,47 +527,47 @@
496 527
    /// \ref upperMap(), \ref costMap(), \ref supplyMap(), \ref stSupply().
497 528
    /// For example,
498 529
    /// \code
499 530
    ///   CostScaling<ListDigraph> cs(graph);
500 531
    ///   cs.lowerMap(lower).upperMap(upper).costMap(cost)
501 532
    ///     .supplyMap(sup).run();
502 533
    /// \endcode
503 534
    ///
504 535
    /// This function can be called more than once. All the parameters
505 536
    /// that have been given are kept for the next call, unless
506 537
    /// \ref reset() is called, thus only the modified parameters
507 538
    /// have to be set again. See \ref reset() for examples.
508
    /// However the underlying digraph must not be modified after this
509
    /// class have been constructed, since it copies the digraph.
539
    /// However, the underlying digraph must not be modified after this
540
    /// class have been constructed, since it copies and extends the graph.
510 541
    ///
511
    /// \param partial_augment By default the algorithm performs
512
    /// partial augment and relabel operations in the cost scaling
513
    /// phases. Set this parameter to \c false for using local push and
514
    /// relabel operations instead.
542
    /// \param method The internal method that will be used in the
543
    /// algorithm. For more information, see \ref Method.
544
    /// \param factor The cost scaling factor. It must be larger than one.
515 545
    ///
516 546
    /// \return \c INFEASIBLE if no feasible flow exists,
517 547
    /// \n \c OPTIMAL if the problem has optimal solution
518 548
    /// (i.e. it is feasible and bounded), and the algorithm has found
519 549
    /// optimal flow and node potentials (primal and dual solutions),
520 550
    /// \n \c UNBOUNDED if the digraph contains an arc of negative cost
521 551
    /// and infinite upper bound. It means that the objective function
522 552
    /// is unbounded on that arc, however note that it could actually be
523 553
    /// bounded over the feasible flows, but this algroithm cannot handle
524 554
    /// these cases.
525 555
    ///
526
    /// \see ProblemType
527
    ProblemType run(bool partial_augment = true) {
556
    /// \see ProblemType, Method
557
    ProblemType run(Method method = PARTIAL_AUGMENT, int factor = 8) {
558
      _alpha = factor;
528 559
      ProblemType pt = init();
529 560
      if (pt != OPTIMAL) return pt;
530
      start(partial_augment);
561
      start(method);
531 562
      return OPTIMAL;
532 563
    }
533 564

	
534 565
    /// \brief Reset all the parameters that have been given before.
535 566
    ///
536 567
    /// This function resets all the paramaters that have been given
537 568
    /// before using functions \ref lowerMap(), \ref upperMap(),
538 569
    /// \ref costMap(), \ref supplyMap(), \ref stSupply().
539 570
    ///
540 571
    /// It is useful for multiple run() calls. If this function is not
541 572
    /// used, all the parameters given before are kept for the next
542 573
    /// \ref run() call.
... ...
@@ -672,27 +703,24 @@
672 703
        map.set(n, static_cast<Cost>(_pi[_node_id[n]]));
673 704
      }
674 705
    }
675 706

	
676 707
    /// @}
677 708

	
678 709
  private:
679 710

	
680 711
    // Initialize the algorithm
681 712
    ProblemType init() {
682 713
      if (_res_node_num == 0) return INFEASIBLE;
683 714

	
684
      // Scaling factor
685
      _alpha = 8;
686

	
687 715
      // Check the sum of supply values
688 716
      _sum_supply = 0;
689 717
      for (int i = 0; i != _root; ++i) {
690 718
        _sum_supply += _supply[i];
691 719
      }
692 720
      if (_sum_supply > 0) return INFEASIBLE;
693 721
      
694 722

	
695 723
      // Initialize vectors
696 724
      for (int i = 0; i != _res_node_num; ++i) {
697 725
        _pi[i] = 0;
698 726
        _excess[i] = _supply[i];
... ...
@@ -808,30 +836,39 @@
808 836
          int ra = _reverse[a];
809 837
          _res_cap[a] = 1;
810 838
          _res_cap[ra] = 0;
811 839
          _cost[a] = 0;
812 840
          _cost[ra] = 0;
813 841
        }
814 842
      }
815 843
      
816 844
      return OPTIMAL;
817 845
    }
818 846

	
819 847
    // Execute the algorithm and transform the results
820
    void start(bool partial_augment) {
848
    void start(Method method) {
849
      // Maximum path length for partial augment
850
      const int MAX_PATH_LENGTH = 4;
851
      
821 852
      // Execute the algorithm
822
      if (partial_augment) {
823
        startPartialAugment();
824
      } else {
825
        startPushRelabel();
853
      switch (method) {
854
        case PUSH:
855
          startPush();
856
          break;
857
        case AUGMENT:
858
          startAugment();
859
          break;
860
        case PARTIAL_AUGMENT:
861
          startAugment(MAX_PATH_LENGTH);
862
          break;
826 863
      }
827 864

	
828 865
      // Compute node potentials for the original costs
829 866
      _arc_vec.clear();
830 867
      _cost_vec.clear();
831 868
      for (int j = 0; j != _res_arc_num; ++j) {
832 869
        if (_res_cap[j] > 0) {
833 870
          _arc_vec.push_back(IntPair(_source[j], _target[j]));
834 871
          _cost_vec.push_back(_scost[j]);
835 872
        }
836 873
      }
837 874
      _sgr.build(_res_node_num, _arc_vec.begin(), _arc_vec.end());
... ...
@@ -842,32 +879,29 @@
842 879
      bf.init(0);
843 880
      bf.start();
844 881

	
845 882
      // Handle non-zero lower bounds
846 883
      if (_have_lower) {
847 884
        int limit = _first_out[_root];
848 885
        for (int j = 0; j != limit; ++j) {
849 886
          if (!_forward[j]) _res_cap[j] += _lower[j];
850 887
        }
851 888
      }
852 889
    }
853 890

	
854
    /// Execute the algorithm performing partial augmentation and
855
    /// relabel operations
856
    void startPartialAugment() {
891
    /// Execute the algorithm performing augment and relabel operations
892
    void startAugment(int max_length = std::numeric_limits<int>::max()) {
857 893
      // Paramters for heuristics
858 894
      const int BF_HEURISTIC_EPSILON_BOUND = 1000;
859 895
      const int BF_HEURISTIC_BOUND_FACTOR  = 3;
860
      // Maximum augment path length
861
      const int MAX_PATH_LENGTH = 4;
862 896

	
863 897
      // Perform cost scaling phases
864 898
      IntVector pred_arc(_res_node_num);
865 899
      std::vector<int> path_nodes;
866 900
      for ( ; _epsilon >= 1; _epsilon = _epsilon < _alpha && _epsilon > 1 ?
867 901
                                        1 : _epsilon / _alpha )
868 902
      {
869 903
        // "Early Termination" heuristic: use Bellman-Ford algorithm
870 904
        // to check if the current flow is optimal
871 905
        if (_epsilon <= BF_HEURISTIC_EPSILON_BOUND) {
872 906
          _arc_vec.clear();
873 907
          _cost_vec.clear();
... ...
@@ -916,25 +950,25 @@
916 950
          while (_active_nodes.size() > 0 &&
917 951
                 _excess[_active_nodes.front()] <= 0) {
918 952
            _active_nodes.pop_front();
919 953
          }
920 954
          if (_active_nodes.size() == 0) break;
921 955
          int start = _active_nodes.front();
922 956
          path_nodes.clear();
923 957
          path_nodes.push_back(start);
924 958

	
925 959
          // Find an augmenting path from the start node
926 960
          int tip = start;
927 961
          while (_excess[tip] >= 0 &&
928
                 int(path_nodes.size()) <= MAX_PATH_LENGTH) {
962
                 int(path_nodes.size()) <= max_length) {
929 963
            int u;
930 964
            LargeCost min_red_cost, rc;
931 965
            int last_out = _sum_supply < 0 ?
932 966
              _first_out[tip+1] : _first_out[tip+1] - 1;
933 967
            for (int a = _next_out[tip]; a != last_out; ++a) {
934 968
              if (_res_cap[a] > 0 &&
935 969
                  _cost[a] + _pi[_source[a]] - _pi[_target[a]] < 0) {
936 970
                u = _target[a];
937 971
                pred_arc[u] = a;
938 972
                _next_out[tip] = a;
939 973
                tip = u;
940 974
                path_nodes.push_back(tip);
... ...
@@ -975,25 +1009,25 @@
975 1009
            _res_cap[pa] -= delta;
976 1010
            _res_cap[_reverse[pa]] += delta;
977 1011
            _excess[u] -= delta;
978 1012
            _excess[v] += delta;
979 1013
            if (_excess[v] > 0 && _excess[v] <= delta)
980 1014
              _active_nodes.push_back(v);
981 1015
          }
982 1016
        }
983 1017
      }
984 1018
    }
985 1019

	
986 1020
    /// Execute the algorithm performing push and relabel operations
987
    void startPushRelabel() {
1021
    void startPush() {
988 1022
      // Paramters for heuristics
989 1023
      const int BF_HEURISTIC_EPSILON_BOUND = 1000;
990 1024
      const int BF_HEURISTIC_BOUND_FACTOR  = 3;
991 1025

	
992 1026
      // Perform cost scaling phases
993 1027
      BoolVector hyper(_res_node_num, false);
994 1028
      for ( ; _epsilon >= 1; _epsilon = _epsilon < _alpha && _epsilon > 1 ?
995 1029
                                        1 : _epsilon / _alpha )
996 1030
      {
997 1031
        // "Early Termination" heuristic: use Bellman-Ford algorithm
998 1032
        // to check if the current flow is optimal
999 1033
        if (_epsilon <= BF_HEURISTIC_EPSILON_BOUND) {
0 comments (0 inline)