/// <summary>
/// Creates the next generation of programs. This method also keeps tabs on memory
/// as it creates the new Population. If we run out of physical memory then the
/// new Population size is changed to wherever we ran out of memory.
/// </summary>
/// <param name="nGeneration">Which generation is currently being processed</param>
/// <param name="Population">Reference to the Population just computed for fitness</param>
/// <param name="Fitness">Fitness object with the fitness computation results</param>
/// <param name="AutoReproduce">List of programs to automatically reproduce into the next generation</param>
/// <returns>Reference to the new Population to use</returns>
public GPPopulation ComputeNext(int nGeneration, GPPopulation Population, GPFitness Fitness, List <GPProgram> AutoReproduce)
{
m_PopCurrent = Population;
m_Fitness = Fitness;
//
// Setup the program selection delegate
switch (m_ModelerConfig.Profile.Reproduction)
{
case GPEnums.Reproduction.Tournament:
ExecProgramSelection = new DELProgramSelection(Fitness.FitnessSelection.SelectProgramTournament);
break;
case GPEnums.Reproduction.OverSelection:
ExecProgramSelection = new DELProgramSelection(Fitness.FitnessSelection.SelectProgramOverSelection);
break;
}
//
// Create a Tree Factory object - It helps in support of the mutation and
// crossover operations
m_TreeFactory = new GPProgramTreeFactory(m_ModelerConfig, m_InputDimension);
return(ComputeNext(Population, Fitness, AutoReproduce));
}
/// <summary>
/// Use the SPEA2 algorithm to compute fitness, with RawFitness and Program
/// Complexity the two objectives to minimize.
/// </summary>
/// <param name="RawFitness"></param>
/// <param name="Population"></param>
public override void PrepareFitness(double[] RawFitness, GPPopulation Population)
{
m_RawFitness = RawFitness;
//
// --- Run The SPEA2 Algorithm ---
m_Archive = UpdateArchive(Population);
//
// Create a new RawFitness based upon the Archive SPEA2 fitness values
double[] SPEA2Fitness = new double[m_Archive.Count];
for (int Program = 0; Program < m_Archive.Count; Program++)
{
SPEA2Fitness[Program] = m_Archive[Program].SPEA2Fitness;
}
//
// In case the archive size changes, let the base class know so it can resize
// its storage to match the new size.
base.PopulationSize = m_Archive.Count;
//
// Use the base class to get the final program selection stats computed.
base.PrepareFitness(SPEA2Fitness, Population);
}
/// <summary>
/// Create the next generation of programs.
/// </summary>
public void ComputeNextGeneration(int Generation)
{
if (m_PopulatonFactory == null)
{
m_PopulatonFactory = new GPPopulationFactory(this, this.Training.Columns);
}
//
// Create the next generation of programs
GPPopulation popNew = m_PopulatonFactory.ComputeNext(Generation, m_Population, m_Fitness, m_SeedPrograms);
//
// If the Population size changed, update the configuration
if (popNew.Count != m_Population.Count)
{
this.Profile.PopulationSize = popNew.Count;
Console.WriteLine("GPServer: Memory is low, Population size reduced to ({0})", popNew.Count);
//
// Reset the Fitness object
m_Fitness.TerminateProcessingThreads();
m_Fitness = null;
}
m_Population = popNew;
//
// Clear out any seed programs we just used
ResetSeedPrograms();
}
/// <summary>
/// Select two parents, based upon fitness, for a crossover operation.
/// Place the two children into the new Population.
/// </summary>
/// <param name="PopNew">Population to add the newly created program to</param>
private void Crossover(GPPopulation PopNew, GPFitnessObjectiveBase FitnessSelection)
{
GPProgram Child1 = (GPProgram)FitnessSelection.Programs(ExecProgramSelection()).Clone();
GPProgram Child2 = (GPProgram)FitnessSelection.Programs(ExecProgramSelection()).Clone();
//
// Convert to trees
Child1.ConvertToTree(m_ModelerConfig.FunctionSet, false);
Child2.ConvertToTree(m_ModelerConfig.FunctionSet, false);
//
// Perform the crossover
m_TreeFactory.Attach(Child1);
Child2 = m_TreeFactory.Crossover(Child2);
//
// Convert to arrays
Child1.ConvertToArray(m_ModelerConfig.FunctionSet);
Child2.ConvertToArray(m_ModelerConfig.FunctionSet);
//
// Add them to the new Population
PopNew.Programs.Add(Child1);
if (PopNew.Count < m_PopCurrent.Count)
{
PopNew.Programs.Add(Child2);
}
}
/// <summary>
/// This method instructs the server object to clean up as much memory
/// as possible.
/// </summary>
public void ForceCleanup()
{
m_Population = null;
m_Fitness.TerminateProcessingThreads();
m_Fitness = null;
m_Training = null;
System.GC.Collect();
}
/// <summary>
/// Select a program based upon fitness to reproduce into the next generation.
/// </summary>
/// <param name="PopNew">Population to add the reproduced program into</param>
private void Reproduce(GPPopulation PopNew, GPFitnessObjectiveBase FitnessSelection)
{
GPProgram Copy = (GPProgram)FitnessSelection.Programs(ExecProgramSelection()).Clone();
//
// Add this copied child into the new Population
PopNew.Programs.Add(Copy);
}
/// <summary>
/// Assign the Population complexity values for use in the MOO domination calculation
/// </summary>
/// <param name="Population"></param>
private int[] PrepareComplexity(GPPopulation Population)
{
int[] Complexity = new int[Population.Count];
for (int Program = 0; Program < Population.Count; Program++)
{
Complexity[Program] = Population.Programs[Program].CountNodes;
}
return(Complexity);
}
/// <summary>
/// Creates the initial Population on this server. The Population size
/// is sent as a parameter, rather than relying upon the profile, because
/// this might be one of many remote objects and it is only working with
/// some subset of the Population.
/// </summary>
/// <param name="PopulationSize">How many individuals to create initially</param>
public void InitializePopulation(int PopulationSize)
{
//
// Go ahead and replace, there is no need to have any knowledge of the
// original Population size;
m_Profile.PopulationSize = PopulationSize;
m_Population = new GPPopulation(this);
m_Population.Build(this.Profile.PopulationBuild, (short)Training.Columns);
//
// Need to get this initialized somewhere, so this seems like a good time
ResetSeedPrograms();
}
/// <summary>
/// Works through the Population to compute the fitness for each program, keeping
/// track of a few statistics, such as best program, worst fitness and
/// average fitness. Makes a call to the abstract method "ComputeFitness" that
/// must be implemented by all GPFitness derived classes.
/// </summary>
/// <param name="Generation">Current modeling generation</param>
/// <param name="Population">Current generation Population of programs</param>
/// <returns></returns>
private int EvaluatePopulation(int Generation, GPPopulation Population)
{
//
// Check to see if the Population size changed, if so, reallocate the array
// that holds the fitness values
if (Population.Count != m_PrevPopulationSize)
{
if (m_Config.Profile.SPEA2MultiObjective)
{
m_FitnessSelection = new GPFitnessSPEA2(Population.Count);
}
else
{
m_FitnessSelection = new GPFitnessSingle(Population.Count);
}
}
FitnessMaximum = 0.0;
FitnessAverage = 0.0;
//
// Go through each program in the Population. First, compute the raw
// fitness result of the program. Once that is done, a call to the
// abstract ComputeFitness function is made where the User Defined Fitness
// function is called.
m_BestProgramGeneration = 0;
m_WorstProgramGeneration = 0;
m_CurrentProgram = 0;
m_CompletedPrograms = 0;
m_Population = Population;
//
// Create the event to be signed when the last program is done processing
wh_LastProgramDone = new EventWaitHandle(false, EventResetMode.ManualReset);
//
// Signal the suspend event
wh_Suspend.Set();
//
// Wait for the last program to be computed
wh_LastProgramDone.WaitOne();
//
// Finish up the average Population fitness
FitnessAverage /= Population.Count;
return(m_BestProgramGeneration);
}
/// <summary>
/// Performs the fitness computation over the Population
/// </summary>
/// <param name="Generation"></param>
/// <param name="Population"></param>
/// <returns></returns>
public GPProgram Compute(int Generation, GPPopulation Population)
{
m_Abort = false;
//
// Go through the Population and compute the fitness of each
// program, returning the best program index.
BestProgram = EvaluatePopulation(Generation, Population);
m_BestProgramRef = Population.Programs[BestProgram];
if (!m_Abort)
{
m_FitnessSelection.PrepareFitness(this.FitnessMeasure, Population);
}
//
// return the best program
return(m_BestProgramRef);
}
/// <summary>
/// Select a program, based on fitness, for mutation. Place
// the mutated program into the new Population.
/// </summary>
/// <param name="PopNew">Population to add the newly created program to</param>
private void Mutate(GPPopulation PopNew, GPFitnessObjectiveBase FitnessSelection)
{
GPProgram Copy = (GPProgram)FitnessSelection.Programs(ExecProgramSelection()).Clone();
//
// Convert it to a tree
Copy.ConvertToTree(m_ModelerConfig.FunctionSet, false);
//
// Perform the mutation
m_TreeFactory.Attach(Copy);
m_TreeFactory.Mutate();
//
// Return it back to an array
Copy.ConvertToArray(m_ModelerConfig.FunctionSet);
//
// Add this mutated individual to the new Population
PopNew.Programs.Add(Copy);
}
/// <summary>
/// Copys all nondominated programs from ArchivePopulation and PopCurrent into
/// the next archive...which is returned from the method
/// </summary>
/// <param name="ArchivePopulation"></param>
/// <param name="Population"></param>
private List <SPEA2ProgramStats> UpdateArchive(GPPopulation Population)
{
SPEA2ProgramStats[] SPEA2Population = new SPEA2ProgramStats[Population.Count];
//
// Grab the complexity of each program. We do this here, rather than getting it
// from the Population object because it saves some computation time due to the
// complexity values being reused several times.
int[] Complexity = PrepareComplexity(Population);
//
// Init the min/max objective values - arbitrarily choose the first program's complexity value
m_MinObjectiveFitness = m_MaxObjectiveFitness = m_RawFitness[0];
m_MinObjectiveComplexity = m_MaxObjectiveComplexity = Complexity[0];
//
// Compute how many solutions each program in the current Population is dominated by
for (int ProgramA = 0; ProgramA < Population.Count; ProgramA++)
{
SPEA2Population[ProgramA] = new SPEA2ProgramStats();
SPEA2Population[ProgramA].Program = Population.Programs[ProgramA];
SPEA2Population[ProgramA].RawFitness = m_RawFitness[ProgramA];
SPEA2Population[ProgramA].Complexity = Complexity[ProgramA];
SPEA2Population[ProgramA].Strength = 0;
//
// Test ProgramA against all programs in the current Population
for (int ProgramB = 0; ProgramB < Population.Count; ProgramB++)
{
if ((m_RawFitness[ProgramB] > SPEA2Population[ProgramA].RawFitness) &&
(Complexity[ProgramB] > SPEA2Population[ProgramA].Complexity))
{
SPEA2Population[ProgramA].Strength++;
}
}
//
// Test ProgramA against all programs in the current archive
for (int ProgramB = 0; ProgramB < Archive.Count; ProgramB++)
{
if (Archive[ProgramB].RawFitness > SPEA2Population[ProgramA].RawFitness &&
Archive[ProgramB].Complexity > SPEA2Population[ProgramA].Complexity)
{
SPEA2Population[ProgramA].Strength++;
}
}
//
// Maintain the min/max objective values from the current Population
m_MinObjectiveFitness = Math.Min(SPEA2Population[ProgramA].RawFitness, m_MinObjectiveFitness);
m_MinObjectiveComplexity = Math.Min(SPEA2Population[ProgramA].Complexity, m_MinObjectiveComplexity);
m_MaxObjectiveFitness = Math.Max(SPEA2Population[ProgramA].RawFitness, m_MaxObjectiveFitness);
m_MaxObjectiveComplexity = Math.Max(SPEA2Population[ProgramA].Complexity, m_MaxObjectiveComplexity);
}
//
// Count how many solutions each program in the archive is dominated by
for (int ProgramA = 0; ProgramA < Archive.Count; ProgramA++)
{
SPEA2ProgramStats ProgramAStats = Archive[ProgramA];
ProgramAStats.Strength = 0;
//
// Test ProgramA against all programs in the current Population
for (int ProgramB = 0; ProgramB < Population.Count; ProgramB++)
{
if (m_RawFitness[ProgramB] > ProgramAStats.RawFitness &&
Complexity[ProgramB] > ProgramAStats.Complexity)
{
ProgramAStats.Strength++;
}
}
//
// Test ProgramA against all programs in the current archive
for (int ProgramB = 0; ProgramB < Archive.Count; ProgramB++)
{
if (Archive[ProgramB].RawFitness > ProgramAStats.RawFitness &&
Archive[ProgramB].Complexity > ProgramAStats.Complexity)
{
ProgramAStats.Strength++;
}
}
Archive[ProgramA] = ProgramAStats;
//
// Maintain the min/max objective values from the current archive
m_MinObjectiveFitness = Math.Min(ProgramAStats.RawFitness, m_MinObjectiveFitness);
m_MinObjectiveComplexity = Math.Min(ProgramAStats.Complexity, m_MinObjectiveComplexity);
m_MaxObjectiveFitness = Math.Max(ProgramAStats.RawFitness, m_MaxObjectiveFitness);
m_MaxObjectiveComplexity = Math.Max(ProgramAStats.Complexity, m_MaxObjectiveComplexity);
}
if (m_MinObjectiveComplexity == m_MaxObjectiveComplexity)
{
m_MaxObjectiveComplexity += 1.0;
}
//
// Now, we can compute the SPEA2 fitness for each individual by summing
// the strenth values for each program it is dominated by.
for (int ProgramA = 0; ProgramA < Population.Count; ProgramA++)
{
SPEA2Population[ProgramA].SPEA2Fitness = 0.0;
//
// Test ProgramA against all programs in the current Population
for (int ProgramB = 0; ProgramB < Population.Count; ProgramB++)
{
if (m_RawFitness[ProgramB] < SPEA2Population[ProgramA].RawFitness &&
Complexity[ProgramB] < SPEA2Population[ProgramA].Complexity)
{
SPEA2Population[ProgramA].SPEA2Fitness += SPEA2Population[ProgramB].Strength;
}
}
//
// Test ProgramA against all programs in the current archive
for (int ProgramB = 0; ProgramB < Archive.Count; ProgramB++)
{
if (Archive[ProgramB].RawFitness < SPEA2Population[ProgramA].RawFitness &&
Archive[ProgramB].Complexity < SPEA2Population[ProgramA].Complexity)
{
SPEA2Population[ProgramA].SPEA2Fitness += Archive[ProgramB].Strength;
}
}
}
//
// Now, do it for the current archive
for (int ProgramA = 0; ProgramA < Archive.Count; ProgramA++)
{
SPEA2ProgramStats ProgramAStats = Archive[ProgramA];
ProgramAStats.SPEA2Fitness = 0.0;
//
// Test ProgramA against all programs in the current Population
for (int ProgramB = 0; ProgramB < Population.Count; ProgramB++)
{
if (SPEA2Population[ProgramB].RawFitness < ProgramAStats.RawFitness &&
SPEA2Population[ProgramB].Complexity < ProgramAStats.Complexity)
{
ProgramAStats.SPEA2Fitness += SPEA2Population[ProgramB].Strength;
}
}
//
// Test ProgramA against all programs in the current archive
for (int ProgramB = 0; ProgramB < Archive.Count; ProgramB++)
{
if (Archive[ProgramB].RawFitness < ProgramAStats.RawFitness &&
Archive[ProgramB].Complexity < ProgramAStats.Complexity)
{
ProgramAStats.SPEA2Fitness += Archive[ProgramB].Strength;
}
}
Archive[ProgramA] = ProgramAStats;
}
//
// Compute the Density estimate for each program in the current Population
for (int ProgramA = 0; ProgramA < Population.Count; ProgramA++)
{
SPEA2Population[ProgramA].Density = ComputeDensity(ref SPEA2Population[ProgramA], SPEA2Population, Archive);
}
//
// Compute the Density estimate for each program in the current archive
for (int ProgramA = 0; ProgramA < Archive.Count; ProgramA++)
{
SPEA2ProgramStats ProgramAStats = Archive[ProgramA];
ProgramAStats.Density = ComputeDensity(ref ProgramAStats, SPEA2Population, Archive);
Archive[ProgramA] = ProgramAStats;
}
//
// Final SPEA2 fitness calculation: F(i)=R(i)+D(i)
for (int ProgramA = 0; ProgramA < Population.Count; ProgramA++)
{
SPEA2Population[ProgramA].SPEA2Fitness = SPEA2Population[ProgramA].SPEA2Fitness + SPEA2Population[ProgramA].Density;
}
for (int ProgramA = 0; ProgramA < Archive.Count; ProgramA++)
{
SPEA2ProgramStats ProgramAStats = Archive[ProgramA];
ProgramAStats.SPEA2Fitness = ProgramAStats.SPEA2Fitness + ProgramAStats.Density;
Archive[ProgramA] = ProgramAStats;
}
//
// Add all nondominated programs from the current Population into the new archive
// The max complexity program allowed in the archive is 10,000. There are two reasons for this...
// *The List container has an "short" sized counter, so programs bigger than 16 bits can't be stored
// in the compressed array representation. Setting a value of 10,000 keeps programs from
// getting too out of hand.
// *A program size of 10,000 is ridiculous in size, so don't allow it.
List <SPEA2ProgramStats> New_Archive = new List <SPEA2ProgramStats>();
SortedDictionary <int, int> New_ArchivePrograms = new SortedDictionary <int, int>();
for (int Program = 0; Program < Population.Programs.Count; Program++)
{
if (SPEA2Population[Program].SPEA2Fitness < 1.0 && SPEA2Population[Program].Complexity > 2 &&
SPEA2Population[Program].Complexity < 10000)
{
New_Archive.Add(SPEA2Population[Program]);
New_ArchivePrograms.Add(Program, Program);
}
}
//
// Add all nondominated programs for the current archive into the new archive
for (int Program = 0; Program < Archive.Count; Program++)
{
if (Archive[Program].SPEA2Fitness < 1.0 && Archive[Program].Complexity > 2)
{
New_Archive.Add(Archive[Program]);
}
}
#if GPLOG
GPLog.ReportLine("Archive Size: " + New_Archive.Count, true);
#endif
//
// Get the archive to the right size
int ArchiveSize = Population.Count / ARCHIVESCALE;
if (New_Archive.Count < ArchiveSize)
{
ExpandArchive(New_Archive, ArchiveSize, SPEA2Population, New_ArchivePrograms);
}
else if (New_Archive.Count > ArchiveSize)
{
ShrinkArchive(New_Archive, ArchiveSize);
}
return(New_Archive);
}
public virtual void PrepareFitness(double[] RawFitness, GPPopulation Population)
{
m_Population = Population;
this.RawFitness = RawFitness;
}
public GPPopulation ComputeNext(GPPopulation PopCurrent, GPFitness Fitness, List <GPProgram> AutoReproduce)
{
GPPopulation PopNew = new GPPopulation(m_ModelerConfig);
//
// Add the new programs into the next generation automatically
foreach (GPProgram Seed in AutoReproduce)
{
PopNew.Programs.Add(Seed);
}
//
// Automatically reproduce the best program into the next generation
PopNew.Programs.Add((GPProgram)Fitness.BestProgramRef.Clone());
//
// Now, go ahead and fill out the rest of the Population
while (PopNew.Count < m_ModelerConfig.Profile.PopulationSize)
{
//
// Make a probabilistic selection between...
// Reproduction
// Mutation
// Crossover
double Choice = GPUtilities.rngNextDouble();
double Cumulative = m_ModelerConfig.Profile.ProbabilityReproductionD;
bool bSelected = false;
if (Choice <= Cumulative)
{
Reproduce(PopNew, Fitness.FitnessSelection);
bSelected = true;
}
Cumulative += m_ModelerConfig.Profile.ProbabilityMutationD;
if (!bSelected && (Choice <= Cumulative))
{
Mutate(PopNew, Fitness.FitnessSelection);
bSelected = true;
}
Cumulative += m_ModelerConfig.Profile.ProbabilityCrossoverD;
if (!bSelected && (Choice <= Cumulative))
{
Crossover(PopNew, Fitness.FitnessSelection);
bSelected = true;
}
//
// Only check every 100 times to save a little time on doing this
// check.
if (PopNew.Count % 100 == 0)
{
if (!CheckMemory())
{
//
// Do a garbage collect just to be sure and check again
System.GC.Collect();
if (!CheckMemory())
{
break;
}
}
}
}
return(PopNew);
}