int
SolveSubproblem(int CurrentSubproblem, int Subproblems,
GainType * GlobalBestCost)
{
Node *FirstNodeSaved = FirstNode, *N, *Next, *Last = 0;
GainType OptimumSaved = Optimum, Cost, Improvement, GlobalCost;
double LastTime, Time, ExcessSaved = Excess;
int NewDimension = 0, OldDimension = 0, Number, i, InitialTourEdges = 0,
AscentCandidatesSaved = AscentCandidates,
InitialPeriodSaved = InitialPeriod, MaxTrialsSaved = MaxTrials;
BestCost = PLUS_INFINITY;
FirstNode = 0;
N = FirstNodeSaved;
do {
if (N->Subproblem == CurrentSubproblem) {
if (SubproblemsCompressed &&
(((N->SubproblemPred == N->SubBestPred ||
FixedOrCommon(N, N->SubproblemPred) ||
(N->SubBestPred &&
(N->FixedTo1Saved == N->SubBestPred ||
N->FixedTo2Saved == N->SubBestPred))) &&
(N->SubproblemSuc == N->SubBestSuc ||
FixedOrCommon(N, N->SubproblemSuc) ||
(N->SubBestSuc &&
(N->FixedTo1Saved == N->SubBestSuc ||
N->FixedTo2Saved == N->SubBestSuc)))) ||
((N->SubproblemPred == N->SubBestSuc ||
FixedOrCommon(N, N->SubproblemPred) ||
(N->SubBestSuc &&
(N->FixedTo1Saved == N->SubBestSuc ||
N->FixedTo2Saved == N->SubBestSuc))) &&
(N->SubproblemSuc == N->SubBestPred ||
FixedOrCommon(N, N->SubproblemSuc) ||
(N->SubBestPred &&
(N->FixedTo1Saved == N->SubBestPred ||
N->FixedTo2Saved == N->SubBestPred))))))
N->Subproblem = -CurrentSubproblem;
else {
if (!FirstNode)
FirstNode = N;
NewDimension++;
}
N->Head = N->Tail = 0;
if (N->SubBestSuc)
OldDimension++;
}
N->SubBestPred = N->SubBestSuc = 0;
N->FixedTo1Saved = N->FixedTo1;
N->FixedTo2Saved = N->FixedTo2;
} while ((N = N->SubproblemSuc) != FirstNodeSaved);
if ((Number = CurrentSubproblem % Subproblems) == 0)
Number = Subproblems;
if (NewDimension <= 3 || NewDimension == OldDimension) {
if (TraceLevel >= 1 && NewDimension <= 3)
printff
("\nSubproblem %d of %d: Dimension = %d (too small)\n",
Number, Subproblems, NewDimension);
FirstNode = FirstNodeSaved;
return 0;
}
if (AscentCandidates > NewDimension - 1)
AscentCandidates = NewDimension - 1;
if (InitialPeriod < 0) {
InitialPeriod = NewDimension / 2;
if (InitialPeriod < 100)
InitialPeriod = 100;
}
if (Excess < 0)
Excess = 1.0 / NewDimension;
if (MaxTrials == -1)
MaxTrials = NewDimension;
N = FirstNode;
do {
Next = N->SubproblemSuc;
if (N->Subproblem == CurrentSubproblem) {
N->Pred = N->Suc = N;
if (N != FirstNode)
Follow(N, Last);
Last = N;
} else if (Next->Subproblem == CurrentSubproblem
&& !Fixed(Last, Next)) {
if (!Last->FixedTo1
|| Last->FixedTo1->Subproblem != CurrentSubproblem)
Last->FixedTo1 = Next;
else
Last->FixedTo2 = Next;
if (!Next->FixedTo1
|| Next->FixedTo1->Subproblem != CurrentSubproblem)
Next->FixedTo1 = Last;
else
Next->FixedTo2 = Last;
if (C == C_EXPLICIT) {
if (Last->Id > Next->Id)
Last->C[Next->Id] = 0;
else
Next->C[Last->Id] = 0;
}
}
}
while ((N = Next) != FirstNode);
Dimension = NewDimension;
AllocateSegments();
InitializeStatistics();
if (CacheSig)
for (i = 0; i <= CacheMask; i++)
CacheSig[i] = 0;
OptimumSaved = Optimum;
Optimum = 0;
N = FirstNode;
do {
if (N->SubproblemSuc == N->InitialSuc ||
N->SubproblemPred == N->InitialSuc)
InitialTourEdges++;
if (!Fixed(N, N->Suc))
Optimum += Distance(N, N->Suc);
if (N->FixedTo1 && N->Subproblem != N->FixedTo1->Subproblem)
eprintf("Illegal fixed edge (%d,%d)", N->Id, N->FixedTo1->Id);
if (N->FixedTo2 && N->Subproblem != N->FixedTo2->Subproblem)
eprintf("Illegal fixed edge (%d,%d)", N->Id, N->FixedTo2->Id);
}
while ((N = N->Suc) != FirstNode);
if (TraceLevel >= 1)
printff
("\nSubproblem %d of %d: Dimension = %d, Upper bound = "
GainFormat "\n", Number, Subproblems, Dimension, Optimum);
FreeCandidateSets();
CreateCandidateSet();
for (Run = 1; Run <= Runs; Run++) {
LastTime = GetTime();
Cost = Norm != 0 ? FindTour() : Optimum;
/* Merge with subproblem tour */
Last = 0;
N = FirstNode;
do {
if (N->Subproblem == CurrentSubproblem) {
if (Last)
Last->Next = N;
Last = N;
}
}
while ((N = N->SubproblemSuc) != FirstNode);
Last->Next = FirstNode;
Cost = MergeWithTour();
if (MaxPopulationSize > 1) {
/* Genetic algorithm */
for (i = 0; i < PopulationSize; i++)
Cost = MergeTourWithIndividual(i);
if (!HasFitness(Cost)) {
if (PopulationSize < MaxPopulationSize) {
AddToPopulation(Cost);
if (TraceLevel >= 1)
PrintPopulation();
} else if (Cost < Fitness[PopulationSize - 1]) {
ReplaceIndividualWithTour(PopulationSize - 1, Cost);
if (TraceLevel >= 1)
PrintPopulation();
}
}
}
if (Cost < BestCost) {
N = FirstNode;
do {
N->SubBestPred = N->Pred;
N->SubBestSuc = N->Suc;
} while ((N = N->Suc) != FirstNode);
BestCost = Cost;
}
if (Cost < Optimum || (Cost != Optimum && OutputTourFileName)) {
Improvement = Optimum - Cost;
if (Improvement > 0) {
BestCost = GlobalCost = *GlobalBestCost -= Improvement;
Optimum = Cost;
} else
GlobalCost = *GlobalBestCost - Improvement;
N = FirstNode;
do
N->Mark = 0;
while ((N = N->SubproblemSuc) != FirstNode);
do {
N->Mark = N;
if (!N->SubproblemSuc->Mark &&
(N->Subproblem != CurrentSubproblem ||
N->SubproblemSuc->Subproblem != CurrentSubproblem))
N->BestSuc = N->SubproblemSuc;
else if (!N->SubproblemPred->Mark &&
(N->Subproblem != CurrentSubproblem ||
N->SubproblemPred->Subproblem !=
CurrentSubproblem))
N->BestSuc = N->SubproblemPred;
else if (!N->Suc->Mark)
N->BestSuc = N->Suc;
else if (!N->Pred->Mark)
N->BestSuc = N->Pred;
else
N->BestSuc = FirstNode;
}
while ((N = N->BestSuc) != FirstNode);
Dimension = DimensionSaved;
i = 0;
do {
if (ProblemType != ATSP)
BetterTour[++i] = N->Id;
else if (N->Id <= Dimension / 2) {
i++;
if (N->BestSuc->Id != N->Id + Dimension / 2)
BetterTour[i] = N->Id;
else
BetterTour[Dimension / 2 - i + 1] = N->Id;
}
}
while ((N = N->BestSuc) != FirstNode);
BetterTour[0] =
BetterTour[ProblemType !=
ATSP ? Dimension : Dimension / 2];
WriteTour(OutputTourFileName, BetterTour, GlobalCost);
if (Improvement > 0) {
do
if (N->Subproblem != CurrentSubproblem)
break;
while ((N = N->SubproblemPred) != FirstNode);
if (N->SubproblemSuc == N->BestSuc) {
N = FirstNode;
do {
N->BestSuc->SubproblemPred = N;
N = N->SubproblemSuc = N->BestSuc;
}
while (N != FirstNode);
} else {
N = FirstNode;
do
(N->SubproblemPred = N->BestSuc)->SubproblemSuc =
N;
while ((N = N->BestSuc) != FirstNode);
}
RecordBestTour();
WriteTour(TourFileName, BestTour, GlobalCost);
}
Dimension = NewDimension;
if (TraceLevel >= 1) {
printff("*** %d: Cost = " GainFormat, Number, GlobalCost);
if (OptimumSaved != MINUS_INFINITY && OptimumSaved != 0)
printff(", Gap = %04f%%",
100.0 * (GlobalCost -
OptimumSaved) / OptimumSaved);
printff(", Time = %0.2f sec. %s\n",
fabs(GetTime() - LastTime),
GlobalCost < OptimumSaved ? "<" : GlobalCost ==
OptimumSaved ? "=" : "");
}
}
Time = fabs(GetTime() - LastTime);
UpdateStatistics(Cost, Time);
if (TraceLevel >= 1 && Cost != PLUS_INFINITY)
printff("Run %d: Cost = " GainFormat ", Time = %0.2f sec.\n\n",
Run, Cost, Time);
if (PopulationSize >= 2 &&
(PopulationSize == MaxPopulationSize
|| Run >= 2 * MaxPopulationSize) && Run < Runs) {
Node *N;
int Parent1, Parent2;
Parent1 = LinearSelection(PopulationSize, 1.25);
do
Parent2 = LinearSelection(PopulationSize, 1.25);
while (Parent1 == Parent2);
ApplyCrossover(Parent1, Parent2);
N = FirstNode;
do {
int d = C(N, N->Suc);
AddCandidate(N, N->Suc, d, INT_MAX);
AddCandidate(N->Suc, N, d, INT_MAX);
N = N->InitialSuc = N->Suc;
}
while (N != FirstNode);
}
SRandom(++Seed);
if (Norm == 0)
break;
}
if (TraceLevel >= 1)
PrintStatistics();
if (C == C_EXPLICIT) {
N = FirstNode;
do {
for (i = 1; i < N->Id; i++) {
N->C[i] -= N->Pi + NodeSet[i].Pi;
N->C[i] /= Precision;
}
if (N->FixedTo1 && N->FixedTo1 != N->FixedTo1Saved) {
if (N->Id > N->FixedTo1->Id)
N->C[N->FixedTo1->Id] = Distance(N, N->FixedTo1);
else
N->FixedTo1->C[N->Id] = Distance(N, N->FixedTo1);
}
if (N->FixedTo2 && N->FixedTo2 != N->FixedTo2Saved) {
if (N->Id > N->FixedTo2->Id)
N->C[N->FixedTo2->Id] = Distance(N, N->FixedTo2);
else
N->FixedTo2->C[N->Id] = Distance(N, N->FixedTo2);
}
}
while ((N = N->Suc) != FirstNode);
}
FreeSegments();
FreeCandidateSets();
FreePopulation();
if (InitialTourEdges == Dimension) {
do
N->InitialSuc = N->SubproblemSuc;
while ((N = N->SubproblemSuc) != FirstNode);
} else {
do
N->InitialSuc = 0;
while ((N = N->SubproblemSuc) != FirstNode);
}
Dimension = ProblemType != ATSP ? DimensionSaved : 2 * DimensionSaved;
N = FirstNode = FirstNodeSaved;
do {
N->Suc = N->BestSuc = N->SubproblemSuc;
N->Suc->Pred = N;
Next = N->FixedTo1;
N->FixedTo1 = N->FixedTo1Saved;
N->FixedTo1Saved = Next;
Next = N->FixedTo2;
N->FixedTo2 = N->FixedTo2Saved;
N->FixedTo2Saved = Next;
}
while ((N = N->Suc) != FirstNode);
Optimum = OptimumSaved;
Excess = ExcessSaved;
AscentCandidates = AscentCandidatesSaved;
InitialPeriod = InitialPeriodSaved;
MaxTrials = MaxTrialsSaved;
return 1;
}
int tsp_lkh()
{
GainType Cost, OldOptimum;
double Time, LastTime = GetTime();
/* Read the specification of the problem */
ReadParameters();
MaxMatrixDimension = 10000;
init();
AllocateStructures();
CreateCandidateSet();
InitializeStatistics();
BestCost = PLUS_INFINITY;
for (Run = 1; Run <= Runs; Run++) {
LastTime = GetTime();
Cost = FindTour(); /* using the Lin-Kernighan heuristic */
if (MaxPopulationSize > 1) {
/* Genetic algorithm */
int i;
for (i = 0; i < PopulationSize; i++) {
GainType OldCost = Cost;
Cost = MergeTourWithIndividual(i);
if (TraceLevel >= 1 && Cost < OldCost) {
// printff(" Merged with %d: Cost = " GainFormat, i + 1,
/// Cost);
// if (Optimum != MINUS_INFINITY && Optimum != 0)
// printff(", Gap = %0.4f%%",
// 100.0 * (Cost - Optimum) / Optimum);
//printff("\n");
}
}
if (!HasFitness(Cost)) {
if (PopulationSize < MaxPopulationSize) {
AddToPopulation(Cost);
if (TraceLevel >= 1)
PrintPopulation();
}
else if (Cost < Fitness[PopulationSize - 1]) {
i = ReplacementIndividual(Cost);
ReplaceIndividualWithTour(i, Cost);
if (TraceLevel >= 1)
PrintPopulation();
}
}
}
else if (Run > 1)
Cost = MergeTourWithBestTour();
if (Cost < BestCost) {
BestCost = Cost;
RecordBetterTour();
RecordBestTour();
}
OldOptimum = Optimum;
if (Cost < Optimum) {
if (FirstNode->InputSuc) {
Node *N = FirstNode;
while ((N = N->InputSuc = N->Suc) != FirstNode);
}
Optimum = Cost;
//printff("*** New optimum = " GainFormat " ***\n\n", Optimum);
}
Time = fabs(GetTime() - LastTime);
UpdateStatistics(Cost, Time);
/*
if (TraceLevel >= 1 && Cost != PLUS_INFINITY) {
printff("Run %d: Cost = " GainFormat, Run, Cost);
if (Optimum != MINUS_INFINITY && Optimum != 0)
printff(", Gap = %0.4f%%",
100.0 * (Cost - Optimum) / Optimum);
printff(", Time = %0.2f sec. %s\n\n", Time,
Cost < Optimum ? "<" : Cost == Optimum ? "=" : "");
}*/
if (StopAtOptimum && Cost == OldOptimum && MaxPopulationSize >= 1) {
Runs = Run;
break;
}
if (PopulationSize >= 2 &&
(PopulationSize == MaxPopulationSize ||
Run >= 2 * MaxPopulationSize) && Run < Runs) {
Node *N;
int Parent1, Parent2;
Parent1 = LinearSelection(PopulationSize, 1.25);
do
Parent2 = LinearSelection(PopulationSize, 1.25);
while (Parent2 == Parent1);
ApplyCrossover(Parent1, Parent2);
N = FirstNode;
do {
if (ProblemType != HCP && ProblemType != HPP) {
int d = C(N, N->Suc);
AddCandidate(N, N->Suc, d, INT_MAX);
AddCandidate(N->Suc, N, d, INT_MAX);
}
N = N->InitialSuc = N->Suc;
} while (N != FirstNode);
}
SRandom(++Seed);
}
// PrintStatistics();
for (int i = 0; i < TSP_N; i++)
{
TSP_RESULT[i] = BestTour[i] - 1;
}
// printf("%d -¡·",BestTour[i]-1);
return BestCost;
}
int main(int argc, char *argv[])
{
GainType Cost;
double Time, LastTime = GetTime();
/* Read the specification of the problem */
if (argc >= 2)
ParameterFileName = argv[1];
ReadParameters();
MaxMatrixDimension = 10000;
ReadProblem();
if (SubproblemSize > 0) {
if (DelaunayPartitioning)
SolveDelaunaySubproblems();
else if (KarpPartitioning)
SolveKarpSubproblems();
else if (KCenterPartitioning)
SolveKCenterSubproblems();
else if (KMeansPartitioning)
SolveKMeansSubproblems();
else if (RohePartitioning)
SolveRoheSubproblems();
else if (MoorePartitioning || SierpinskiPartitioning)
SolveSFCSubproblems();
else
SolveTourSegmentSubproblems();
return EXIT_SUCCESS;
}
AllocateStructures();
CreateCandidateSet();
InitializeStatistics();
if (Norm != 0)
BestCost = PLUS_INFINITY;
else {
/* The ascent has solved the problem! */
Optimum = BestCost = (GainType) LowerBound;
UpdateStatistics(Optimum, GetTime() - LastTime);
RecordBetterTour();
RecordBestTour();
WriteTour(OutputTourFileName, BestTour, BestCost);
WriteTour(TourFileName, BestTour, BestCost);
Runs = 0;
}
/* Find a specified number (Runs) of local optima */
for (Run = 1; Run <= Runs; Run++) {
LastTime = GetTime();
Cost = FindTour(); /* using the Lin-Kernighan heuristic */
if (MaxPopulationSize > 1) {
/* Genetic algorithm */
int i;
for (i = 0; i < PopulationSize; i++) {
GainType OldCost = Cost;
Cost = MergeTourWithIndividual(i);
if (TraceLevel >= 1 && Cost < OldCost) {
printff(" Merged with %d: Cost = " GainFormat, i + 1,
Cost);
if (Optimum != MINUS_INFINITY && Optimum != 0)
printff(", Gap = %0.4f%%",
100.0 * (Cost - Optimum) / Optimum);
printff("\n");
}
}
if (!HasFitness(Cost)) {
if (PopulationSize < MaxPopulationSize) {
AddToPopulation(Cost);
if (TraceLevel >= 1)
PrintPopulation();
} else if (Cost < Fitness[PopulationSize - 1]) {
i = ReplacementIndividual(Cost);
ReplaceIndividualWithTour(i, Cost);
if (TraceLevel >= 1)
PrintPopulation();
}
}
} else if (Run > 1)
Cost = MergeBetterTourWithBestTour();
if (Cost < BestCost) {
BestCost = Cost;
RecordBetterTour();
RecordBestTour();
WriteTour(OutputTourFileName, BestTour, BestCost);
WriteTour(TourFileName, BestTour, BestCost);
}
if (Cost < Optimum) {
if (FirstNode->InputSuc) {
Node *N = FirstNode;
while ((N = N->InputSuc = N->Suc) != FirstNode);
}
Optimum = Cost;
printff("*** New optimum = " GainFormat " ***\n\n", Optimum);
}
Time = fabs(GetTime() - LastTime);
UpdateStatistics(Cost, Time);
if (TraceLevel >= 1 && Cost != PLUS_INFINITY) {
printff("Run %d: Cost = " GainFormat, Run, Cost);
if (Optimum != MINUS_INFINITY && Optimum != 0)
printff(", Gap = %0.4f%%",
100.0 * (Cost - Optimum) / Optimum);
printff(", Time = %0.2f sec. %s\n\n", Time,
Cost < Optimum ? "<" : Cost == Optimum ? "=" : "");
}
if (PopulationSize >= 2 &&
(PopulationSize == MaxPopulationSize ||
Run >= 2 * MaxPopulationSize) && Run < Runs) {
Node *N;
int Parent1, Parent2;
Parent1 = LinearSelection(PopulationSize, 1.25);
do
Parent2 = LinearSelection(PopulationSize, 1.25);
while (Parent2 == Parent1);
ApplyCrossover(Parent1, Parent2);
N = FirstNode;
do {
int d = C(N, N->Suc);
AddCandidate(N, N->Suc, d, INT_MAX);
AddCandidate(N->Suc, N, d, INT_MAX);
N = N->InitialSuc = N->Suc;
}
while (N != FirstNode);
}
SRandom(++Seed);
}
PrintStatistics();
return EXIT_SUCCESS;
}
ReturnFlag LKH::LKHAlg::run_()
{
GainType Cost, OldOptimum;
double Time, LastTime = GetTime();
if (SubproblemSize > 0) {
if (DelaunayPartitioning)
SolveDelaunaySubproblems();
else if (KarpPartitioning)
SolveKarpSubproblems();
else if (KCenterPartitioning)
SolveKCenterSubproblems();
else if (KMeansPartitioning)
SolveKMeansSubproblems();
else if (RohePartitioning)
SolveRoheSubproblems();
else if (MoorePartitioning || SierpinskiPartitioning)
SolveSFCSubproblems();
else
SolveTourSegmentSubproblems();
}
AllocateStructures();
CreateCandidateSet();
InitializeStatistics();
if (Norm != 0)
BestCost = PLUS_INFINITY;
else {
/* The ascent has solved the problem! */
Optimum = BestCost = (GainType) LowerBound;
UpdateStatistics(Optimum, GetTime() - LastTime);
RecordBetterTour();
RecordBestTour();
WriteTour(OutputTourFileName, BestTour, BestCost);
WriteTour(TourFileName, BestTour, BestCost);
Runs = 0;
}
/* Find a specified number (Runs) of local optima */
for (Run = 1; Run <= Runs; Run++) {
LastTime = GetTime();
Cost = FindTour(); /* using the Lin-Kernighan heuristic */
if (*MaxPopulationSize > 1) {
/* Genetic algorithm */
int i;
for (i = 0; i < *PopulationSize; i++) {
GainType OldCost = Cost;
Cost = MergeTourWithIndividual(i,this);
if (TraceLevel >= 1 && Cost < OldCost) {
printff(" Merged with %d: Cost = " GainFormat, i + 1,
Cost);
if (Optimum != MINUS_INFINITY && Optimum != 0)
printff(", Gap = %0.4f%%",
100.0 * (Cost - Optimum) / Optimum);
printff("\n");
}
}
if (!HasFitness(Cost)) {
if (*PopulationSize < *MaxPopulationSize) {
AddToPopulation(Cost,this);
if (TraceLevel >= 1)
PrintPopulation(this);
} else if (Cost < Fitness.get()[*PopulationSize - 1]) {
i = ReplacementIndividual(Cost,this);
ReplaceIndividualWithTour(i, Cost,this);
if (TraceLevel >= 1)
PrintPopulation(this);
}
}
} else if (Run > 1)
Cost = MergeBetterTourWithBestTour();
if (Cost < BestCost) {
BestCost = Cost;
RecordBetterTour();
RecordBestTour();
WriteTour(OutputTourFileName, BestTour, BestCost);
WriteTour(TourFileName, BestTour, BestCost);
}
OldOptimum = Optimum;
if (Cost < Optimum) {
if (FirstNode->InputSuc) {
Node *N = FirstNode;
while ((N = N->InputSuc = N->Suc) != FirstNode);
}
Optimum = Cost;
printff("*** New optimum = " GainFormat " ***\n\n", Optimum);
}
Time = fabs(GetTime() - LastTime);
UpdateStatistics(Cost, Time);
if (TraceLevel >= 1 && Cost != PLUS_INFINITY) {
printff("Run %d: Cost = " GainFormat, Global::msp_global->m_runId, Cost);
if (Optimum != MINUS_INFINITY && Optimum != 0)
printff(", Gap = %0.4f%%",
100.0 * (Cost - Optimum) / Optimum);
// printff(", Time = %0.2f sec. %s\n\n", Time,
// Cost < Optimum ? "<" : Cost == Optimum ? "=" : "");
printff(", Time = %0.2f sec. \n", Time);
}
if (StopAtOptimum && Cost == OldOptimum && *MaxPopulationSize >= 1) {
Runs = Run;
break;
}
if (*PopulationSize >= 2 &&
(*PopulationSize == *MaxPopulationSize ||
Run >= 2 * *MaxPopulationSize) && Run < Runs) {
Node *N;
int Parent1, Parent2;
Parent1 = LinearSelection(*PopulationSize, 1.25,this);
do
Parent2 = LinearSelection(*PopulationSize, 1.25,this);
while (Parent2 == Parent1);
ApplyCrossover(Parent1, Parent2,this);
N = FirstNode;
do {
int d = (this->*C)(N, N->Suc);
AddCandidate(N, N->Suc, d, INT_MAX);
AddCandidate(N->Suc, N, d, INT_MAX);
N = N->InitialSuc = N->Suc;
}
while (N != FirstNode);
}
mv_cost[Global::msp_global->m_runId]=Cost;
SRandom(++Seed);
}
// PrintStatistics();
freeAll();
FreeStructures();
return Return_Terminate;
}