gravatar
kpeter (Peter Kovacs)
kpeter@inf.elte.hu
Implement the scaling Price Refinement heuristic in CostScaling (#417) instead of Early Termination. These two heuristics are similar, but the newer one is faster and not only makes it possible to skip some epsilon phases, but it can improve the performance of the other phases, as well.
0 1 0
default
1 file changed with 232 insertions and 27 deletions:
↑ Collapse diff ↑
Show white space 4 line context
... ...
@@ -981,28 +981,227 @@
981 981
    }
982 982

	
983
    // Early termination heuristic
984
    bool earlyTermination() {
985
      const double EARLY_TERM_FACTOR = 3.0;
983
    // Price (potential) refinement heuristic
984
    bool priceRefinement() {
986 985

	
987
      // Build a static residual graph
988
      _arc_vec.clear();
989
      _cost_vec.clear();
990
      for (int j = 0; j != _res_arc_num; ++j) {
991
        if (_res_cap[j] > 0) {
992
          _arc_vec.push_back(IntPair(_source[j], _target[j]));
993
          _cost_vec.push_back(_cost[j] + 1);
986
      // Stack for stroing the topological order
987
      IntVector stack(_res_node_num);
988
      int stack_top;
989

	
990
      // Perform phases
991
      while (topologicalSort(stack, stack_top)) {
992

	
993
        // Compute node ranks in the acyclic admissible network and
994
        // store the nodes in buckets
995
        for (int i = 0; i != _res_node_num; ++i) {
996
          _rank[i] = 0;
994 997
        }
998
        const int bucket_end = _root + 1;
999
        for (int r = 0; r != _max_rank; ++r) {
1000
          _buckets[r] = bucket_end;
995 1001
      }
996
      _sgr.build(_res_node_num, _arc_vec.begin(), _arc_vec.end());
1002
        int top_rank = 0;
1003
        for ( ; stack_top >= 0; --stack_top) {
1004
          int u = stack[stack_top], v;
1005
          int rank_u = _rank[u];
997 1006

	
998
      // Run Bellman-Ford algorithm to check if the current flow is optimal
999
      BellmanFord<StaticDigraph, LargeCostArcMap> bf(_sgr, _cost_map);
1000
      bf.init(0);
1001
      bool done = false;
1002
      int K = int(EARLY_TERM_FACTOR * std::sqrt(double(_res_node_num)));
1003
      for (int i = 0; i < K && !done; ++i) {
1004
        done = bf.processNextWeakRound();
1007
          LargeCost rc, pi_u = _pi[u];
1008
          int last_out = _first_out[u+1];
1009
          for (int a = _first_out[u]; a != last_out; ++a) {
1010
            if (_res_cap[a] > 0) {
1011
              v = _target[a];
1012
              rc = _cost[a] + pi_u - _pi[v];
1013
              if (rc < 0) {
1014
                LargeCost nrc = static_cast<LargeCost>((-rc - 0.5) / _epsilon);
1015
                if (nrc < LargeCost(_max_rank)) {
1016
                  int new_rank_v = rank_u + static_cast<int>(nrc);
1017
                  if (new_rank_v > _rank[v]) {
1018
                    _rank[v] = new_rank_v;
1005 1019
      }
1006
      return done;
1020
                }
1021
              }
1022
            }
1023
          }
1024

	
1025
          if (rank_u > 0) {
1026
            top_rank = std::max(top_rank, rank_u);
1027
            int bfirst = _buckets[rank_u];
1028
            _bucket_next[u] = bfirst;
1029
            _bucket_prev[bfirst] = u;
1030
            _buckets[rank_u] = u;
1031
          }
1032
        }
1033

	
1034
        // Check if the current flow is epsilon-optimal
1035
        if (top_rank == 0) {
1036
          return true;
1037
        }
1038

	
1039
        // Process buckets in top-down order
1040
        for (int rank = top_rank; rank > 0; --rank) {
1041
          while (_buckets[rank] != bucket_end) {
1042
            // Remove the first node from the current bucket
1043
            int u = _buckets[rank];
1044
            _buckets[rank] = _bucket_next[u];
1045

	
1046
            // Search the outgoing arcs of u
1047
            LargeCost rc, pi_u = _pi[u];
1048
            int last_out = _first_out[u+1];
1049
            int v, old_rank_v, new_rank_v;
1050
            for (int a = _first_out[u]; a != last_out; ++a) {
1051
              if (_res_cap[a] > 0) {
1052
                v = _target[a];
1053
                old_rank_v = _rank[v];
1054

	
1055
                if (old_rank_v < rank) {
1056

	
1057
                  // Compute the new rank of node v
1058
                  rc = _cost[a] + pi_u - _pi[v];
1059
                  if (rc < 0) {
1060
                    new_rank_v = rank;
1061
                  } else {
1062
                    LargeCost nrc = rc / _epsilon;
1063
                    new_rank_v = 0;
1064
                    if (nrc < LargeCost(_max_rank)) {
1065
                      new_rank_v = rank - 1 - static_cast<int>(nrc);
1066
                    }
1067
                  }
1068

	
1069
                  // Change the rank of node v
1070
                  if (new_rank_v > old_rank_v) {
1071
                    _rank[v] = new_rank_v;
1072

	
1073
                    // Remove v from its old bucket
1074
                    if (old_rank_v > 0) {
1075
                      if (_buckets[old_rank_v] == v) {
1076
                        _buckets[old_rank_v] = _bucket_next[v];
1077
                      } else {
1078
                        int pv = _bucket_prev[v], nv = _bucket_next[v];
1079
                        _bucket_next[pv] = nv;
1080
                        _bucket_prev[nv] = pv;
1081
                      }
1082
                    }
1083

	
1084
                    // Insert v into its new bucket
1085
                    int nv = _buckets[new_rank_v];
1086
                    _bucket_next[v] = nv;
1087
                    _bucket_prev[nv] = v;
1088
                    _buckets[new_rank_v] = v;
1089
                  }
1090
                }
1091
              }
1092
            }
1093

	
1094
            // Refine potential of node u
1095
            _pi[u] -= rank * _epsilon;
1096
          }
1097
        }
1098

	
1099
      }
1100

	
1101
      return false;
1102
    }
1103

	
1104
    // Find and cancel cycles in the admissible network and
1105
    // determine topological order using DFS
1106
    bool topologicalSort(IntVector &stack, int &stack_top) {
1107
      const int MAX_CYCLE_CANCEL = 1;
1108

	
1109
      BoolVector reached(_res_node_num, false);
1110
      BoolVector processed(_res_node_num, false);
1111
      IntVector pred(_res_node_num);
1112
      for (int i = 0; i != _res_node_num; ++i) {
1113
        _next_out[i] = _first_out[i];
1114
      }
1115
      stack_top = -1;
1116

	
1117
      int cycle_cnt = 0;
1118
      for (int start = 0; start != _res_node_num; ++start) {
1119
        if (reached[start]) continue;
1120

	
1121
        // Start DFS search from this start node
1122
        pred[start] = -1;
1123
        int tip = start, v;
1124
        while (true) {
1125
          // Check the outgoing arcs of the current tip node
1126
          reached[tip] = true;
1127
          LargeCost pi_tip = _pi[tip];
1128
          int a, last_out = _first_out[tip+1];
1129
          for (a = _next_out[tip]; a != last_out; ++a) {
1130
            if (_res_cap[a] > 0) {
1131
              v = _target[a];
1132
              if (_cost[a] + pi_tip - _pi[v] < 0) {
1133
                if (!reached[v]) {
1134
                  // A new node is reached
1135
                  reached[v] = true;
1136
                  pred[v] = tip;
1137
                  _next_out[tip] = a;
1138
                  tip = v;
1139
                  a = _next_out[tip];
1140
                  last_out = _first_out[tip+1];
1141
                  break;
1142
                }
1143
                else if (!processed[v]) {
1144
                  // A cycle is found
1145
                  ++cycle_cnt;
1146
                  _next_out[tip] = a;
1147

	
1148
                  // Find the minimum residual capacity along the cycle
1149
                  Value d, delta = _res_cap[a];
1150
                  int u, delta_node = tip;
1151
                  for (u = tip; u != v; ) {
1152
                    u = pred[u];
1153
                    d = _res_cap[_next_out[u]];
1154
                    if (d <= delta) {
1155
                      delta = d;
1156
                      delta_node = u;
1157
                    }
1158
                  }
1159

	
1160
                  // Augment along the cycle
1161
                  _res_cap[a] -= delta;
1162
                  _res_cap[_reverse[a]] += delta;
1163
                  for (u = tip; u != v; ) {
1164
                    u = pred[u];
1165
                    int ca = _next_out[u];
1166
                    _res_cap[ca] -= delta;
1167
                    _res_cap[_reverse[ca]] += delta;
1168
                  }
1169

	
1170
                  // Check the maximum number of cycle canceling
1171
                  if (cycle_cnt >= MAX_CYCLE_CANCEL) {
1172
                    return false;
1173
                  }
1174

	
1175
                  // Roll back search to delta_node
1176
                  if (delta_node != tip) {
1177
                    for (u = tip; u != delta_node; u = pred[u]) {
1178
                      reached[u] = false;
1179
                    }
1180
                    tip = delta_node;
1181
                    a = _next_out[tip] + 1;
1182
                    last_out = _first_out[tip+1];
1183
                    break;
1184
                  }
1185
                }
1186
              }
1187
            }
1188
          }
1189

	
1190
          // Step back to the previous node
1191
          if (a == last_out) {
1192
            processed[tip] = true;
1193
            stack[++stack_top] = tip;
1194
            tip = pred[tip];
1195
            if (tip < 0) {
1196
              // Finish DFS from the current start node
1197
              break;
1198
            }
1199
            ++_next_out[tip];
1200
          }
1201
        }
1202

	
1203
      }
1204

	
1205
      return (cycle_cnt == 0);
1007 1206
    }
1008 1207

	
... ...
@@ -1103,5 +1302,5 @@
1103 1302
    void startAugment(int max_length) {
1104 1303
      // Paramters for heuristics
1105
      const int EARLY_TERM_EPSILON_LIMIT = 1000;
1304
      const int PRICE_REFINEMENT_LIMIT = 2;
1106 1305
      const double GLOBAL_UPDATE_FACTOR = 1.0;
1107 1306
      const int global_update_skip = static_cast<int>(GLOBAL_UPDATE_FACTOR *
... ...
@@ -1113,10 +1312,13 @@
1113 1312
      BoolVector path_arc(_res_arc_num, false);
1114 1313
      int relabel_cnt = 0;
1314
      int eps_phase_cnt = 0;
1115 1315
      for ( ; _epsilon >= 1; _epsilon = _epsilon < _alpha && _epsilon > 1 ?
1116 1316
                                        1 : _epsilon / _alpha )
1117 1317
      {
1118
        // Early termination heuristic
1119
        if (_epsilon <= EARLY_TERM_EPSILON_LIMIT) {
1120
          if (earlyTermination()) break;
1318
        ++eps_phase_cnt;
1319

	
1320
        // Price refinement heuristic
1321
        if (eps_phase_cnt >= PRICE_REFINEMENT_LIMIT) {
1322
          if (priceRefinement()) continue;
1121 1323
        }
1122 1324

	
... ...
@@ -1225,5 +1427,5 @@
1225 1427
    void startPush() {
1226 1428
      // Paramters for heuristics
1227
      const int EARLY_TERM_EPSILON_LIMIT = 1000;
1429
      const int PRICE_REFINEMENT_LIMIT = 2;
1228 1430
      const double GLOBAL_UPDATE_FACTOR = 2.0;
1229 1431

	
... ...
@@ -1236,10 +1438,13 @@
1236 1438
      LargeCostVector hyper_cost(_res_node_num);
1237 1439
      int relabel_cnt = 0;
1440
      int eps_phase_cnt = 0;
1238 1441
      for ( ; _epsilon >= 1; _epsilon = _epsilon < _alpha && _epsilon > 1 ?
1239 1442
                                        1 : _epsilon / _alpha )
1240 1443
      {
1241
        // Early termination heuristic
1242
        if (_epsilon <= EARLY_TERM_EPSILON_LIMIT) {
1243
          if (earlyTermination()) break;
1444
        ++eps_phase_cnt;
1445

	
1446
        // Price refinement heuristic
1447
        if (eps_phase_cnt >= PRICE_REFINEMENT_LIMIT) {
1448
          if (priceRefinement()) continue;
1244 1449
        }
1245 1450

	
0 comments (0 inline)