public static IEnumerable <SolutionGenome> ReadSolutions(string solutionFile, List <ShapeTemplate> shapes, Stock stock, EAConfig config)
{
using (var solFile = new StreamReader(solutionFile, false))
{
List <Gene> genes = null;
int shapeCounter = 0;
string line;
while ((line = solFile.ReadLine()) != null)
{
line = line.Trim();
if (line.Contains("[Solution]"))
{
if (genes != null && genes.Any())
{
yield return(SolutionGenome.ConstructConfig(genes, stock.Length, config));
}
genes = new List <Gene>();
shapeCounter = 0;
}
else if (line != "" && genes != null)
{
var lineData = line.Split(new [] { ',' }).Select(num => Convert.ToInt32(num)).ToList();
var point = new Point(lineData[0], lineData[1]);
var rotation = (ClockwiseRotation)(lineData[2]);
genes.Add(new Gene(shapes[shapeCounter], point, rotation));
shapeCounter += 1;
}
}
if (genes != null && genes.Any())
{
yield return(SolutionGenome.ConstructConfig(genes, stock.Length, config));
}
}
}
public static SolutionGenome GenerateRandomSolution(IEnumerable <ShapeTemplate> shapes, Stock stock, EAConfig config)
{
var solution = SolutionGenome.ConstructRandom(shapes, stock.Length, config);
solution.Repair(stock);
return(solution);
}
public static int SolutionDistance(SolutionGenome sln1, SolutionGenome sln2)
{
var shapes = sln1.Genes.Select(g => g.Template);
var sln1ShapeGenes = sln1.Genes.ToDictionary(g => g.Template, g => g);
var sln2ShapeGenes = sln2.Genes.ToDictionary(g => g.Template, g => g);
return(shapes
.Select(s => Tuple.Create(sln1ShapeGenes[s], sln2ShapeGenes[s]))
.Select(gp => (int)Math.Pow(gp.Item1.Origin.ManhattanDistance(gp.Item2.Origin), 2))
.Sum());
}
public void Run(EAConfig config, Stock stock, IEnumerable <ShapeTemplate> shapes)
{
var initialSolutions = new List <SolutionGenome>();
if (config.SolutionInit != "" && File.Exists(config.SolutionInit))
{
initialSolutions.AddRange(ReadSolutions(config.SolutionInit, shapes.ToList(), stock, config));
}
else
{
initialSolutions.AddRange(Enumerable
.Range(initialSolutions.Count(), config.NumParents)
.Select(i => GenerateRandomSolution(shapes, stock, config))
);
}
int evalCounter = 0;
Func <SolutionGenome, List <int> > evaluateSolution = (individual) =>
{
var evals = new List <int>();
if (config.Fitness.Length)
{
evals.Add(stock.Length - individual.SolutionLength);
}
if (config.Fitness.Width)
{
evals.Add(stock.Width - individual.SolutionWidth);
}
if (config.Fitness.Cut)
{
var placements = new Dictionary <Point, Gene>();
foreach (var gene in individual.Genes)
{
foreach (var point in gene.Phenotype().Points)
{
placements[point] = gene;
}
}
// Maximizing adjacent points is minimizing cuts
evals.Add(
individual.Genes.Sum(g => {
return(g.Phenotype().AdjacentPoints
.Where(p => placements.ContainsKey(p))
.Count());
})
);
}
if (config.Fitness.Adjacents)
{
var placements = new Dictionary <Point, Gene>();
foreach (var gene in individual.Genes)
{
foreach (var point in gene.Phenotype().Points)
{
placements[point] = gene;
}
}
// Minimize number of shapes that are adjacent to each other, also, keep it positive
evals.Add(
(individual.Genes.Count() * individual.Genes.Count()) - individual.Genes.Sum(g => {
return(new HashSet <Gene>(
g.Phenotype().AdjacentPoints
.Where(p => placements.ContainsKey(p))
.Select(p => placements[p])
).Count());
})
);
}
return(evals);
};
Func <IEnumerable <SolutionGenome>, List <EvalNode <SolutionGenome> > > evaluate = (population) =>
{
evalCounter += population.Count();
var popFits = new Dictionary <SolutionGenome, List <int> >();
foreach (var individual in population)
{
popFits[individual] = evaluateSolution(individual);
}
var fittedPopulation = new List <EvalNode <SolutionGenome> >();
var levels = new Dictionary <SolutionGenome, int>();
var dominatedBy = new Dictionary <SolutionGenome, List <SolutionGenome> >();
if (popFits[population.First()].Count() > 1)
{
foreach (var individual in population)
{
var fitnessTypes = popFits[individual];
dominatedBy[individual] = new List <SolutionGenome>();
foreach (var dominator in population)
{
var fitPairs = fitnessTypes
.Zip(popFits[dominator], (i, d) => Tuple.Create(i, d));
var ltFits = fitPairs.Where(el => el.Item1 > el.Item2);
var rtFits = fitPairs.Where(el => el.Item1 < el.Item2);
bool isDominated = ltFits.Count() == 0 && rtFits.Any();
if (isDominated)
{
dominatedBy[individual].Add(dominator);
}
}
}
var unFitted = new List <SolutionGenome>(population);
int levelCounter = 1;
while (unFitted.Count() > 0)
{
foreach (var individual in unFitted.ToList())
{
bool fit = true;
foreach (var dominator in dominatedBy[individual])
{
if (!levels.ContainsKey(dominator))
{
fit = false;
break;
}
}
if (fit)
{
levels[individual] = -levelCounter;
unFitted.Remove(individual);
}
}
levelCounter += 1;
}
}
foreach (var individual in population)
{
var fitnessTypes = popFits[individual];
int fitness = 0;
if (fitnessTypes.Count() == 1)
{
fitness = fitnessTypes.First();
}
else if (fitnessTypes.Count() > 1)
{
fitness = levels[individual];
}
var newFit = new EvalNode <SolutionGenome>(individual, fitness);
fittedPopulation.Add(newFit);
}
if (config.Sharing)
{
foreach (var eFit in fittedPopulation)
{
foreach (var nicheFit in fittedPopulation)
{
eFit.Fitness += (int)Math.Floor(nicheFit.Fitness / Math.Pow(1 + SolutionDistance(eFit.Individual, nicheFit.Individual), 2));
}
}
}
return(fittedPopulation);
};
BestPopulation = new List <SolutionGenome>();
int unchangedBest = 0;
int generationCounter = 0;
Func <IEnumerable <SolutionGenome>, bool> terminate = (population) =>
{
generationCounter += 1;
var _evalPopulation = evaluate((new [] { population, BestPopulation }).SelectMany(p => p))
.ToList();
int maxFitness = _evalPopulation.Max(i => i.Fitness);
// We need this unique population handling to prevent best pop bloat in moea
var uniquePops = new Dictionary <string, EvalNode <SolutionGenome> >();
foreach (var e in _evalPopulation.Where(s => s.Fitness == maxFitness))
{
uniquePops[String.Join(",", evaluateSolution(e.Individual))] = e;
}
var evalPopulation = uniquePops.Values.ToList();
bool isNewBest = true;
var bestUniquePops = BestPopulation.ToDictionary(i => String.Join(",", evaluateSolution(i)), i => i);
foreach (var newBest in evalPopulation)
{
if (!bestUniquePops.ContainsKey(String.Join(",", evaluateSolution(newBest.Individual))))
{
isNewBest = false;
unchangedBest = 0;
break;
}
}
BestPopulation = evalPopulation.Select(e => e.Individual).ToList();
if (isNewBest)
{
unchangedBest += 1;
}
Console.WriteLine("Evals {0}\tLevels: {1}\tAvg Fitness: {2}\tBest Fitnesss {3}\tMutations: {4:0.000} {5:0.000} {6:0.000}\tCrossover: {7:0.000}",
evalCounter,
(new HashSet <int>(_evalPopulation.Select(e => e.Fitness))).Count(),
String.Join(",", population
.Select(s => evaluateSolution(s))
.Aggregate(new List <int> {
0, 0, 0, 0
}, (lhs, rhs) => lhs.Zip(rhs, (l, r) => l + r).ToList())
.Select(total => String.Format("{0:0.000}", total / (double)population.Count()))
),
String.Join(",", population
.Select(s => evaluateSolution(s))
.Aggregate(new List <int> {
0, 0, 0, 0
}, (lhs, rhs) => lhs.Zip(rhs, (l, r) => Math.Max(l, r)).ToList())
),
population.Sum(p => p.RateCreepRandom) / population.Count(),
population.Sum(p => p.RateRotateRandom) / population.Count(),
population.Sum(p => p.RateSlideRandom) / population.Count(),
population.Sum(p => p.RateAdjacencyCrossover) / population.Count()
);
bool evalLimitReached = config.Termination.EvalLimit != 0 && config.Termination.EvalLimit <= evalCounter;
bool generationLimitReached =
config.Termination.GenerationLimit != 0 && config.Termination.GenerationLimit < generationCounter;
bool unchangedBestReached = config.Termination.UnchangedBestLimit != 0 && config.Termination.UnchangedBestLimit < unchangedBest;
return(evalLimitReached || generationLimitReached || unchangedBestReached);
};
Func <IEnumerable <SolutionGenome>,
IEnumerable <SolutionGenome>,
IEnumerable <SolutionGenome> > survive;
switch (config.SurvivalSelection.SelectionWeight)
{
case SelectionWeight.Truncate:
survive = (parents, offspring) =>
{
return(evaluate(
new[] { parents.SkipWhile(s => config.SurvivalSelection.DropParents), offspring }
.SelectMany(p => p)
)
.OrderByDescending(o => o.Fitness)
.Select(o => o.Individual)
.Take(parents.Count()));
};
break;
case SelectionWeight.Random:
survive = EASurvivalSelection <SolutionGenome> .CreateTournamentSelector(
(kChoices) => kChoices.ToList().ChooseSingle(),
evaluate,
config.SurvivalSelection.SelectPool,
config.SurvivalSelection.Replacement,
config.SurvivalSelection.DropParents,
config.NumParents
);
break;
case SelectionWeight.Best:
survive = EASurvivalSelection <SolutionGenome> .CreateTournamentSelector(
(kChoices) => kChoices.MaxByValue((k) => k.Fitness),
evaluate,
config.SurvivalSelection.SelectPool,
config.SurvivalSelection.Replacement,
config.SurvivalSelection.DropParents,
config.NumParents
);
break;
case SelectionWeight.Fitness:
survive = EASurvivalSelection <SolutionGenome> .CreateTournamentSelector(
(kChoices) =>
{
var totalFitness = kChoices.Sum(k => k.Fitness);
var fitPick = CmnRandom.Random.Next(0, totalFitness - 1);
return(kChoices.First(k =>
{
fitPick -= k.Fitness;
return fitPick <= 0;
}));
},
evaluate,
config.SurvivalSelection.SelectPool,
config.SurvivalSelection.Replacement,
config.SurvivalSelection.DropParents,
config.NumParents
);
break;
case SelectionWeight.Rank:
survive = EASurvivalSelection <SolutionGenome> .CreateTournamentSelector(
(kChoices) => kChoices
.OrderBy((k) => - k.Fitness)
.Select((k, index) => Tuple.Create(k, config.SurvivalSelection.RateP *Math.Pow(1 - config.SurvivalSelection.RateP, index)))
.Select(ki => Tuple.Create(ki.Item1, CmnRandom.Random.NextDouble() < ki.Item2))
.OrderBy(kp => !kp.Item2)
.First().Item1,
evaluate,
config.SurvivalSelection.SelectPool,
config.SurvivalSelection.Replacement,
config.SurvivalSelection.DropParents,
config.NumParents
);
break;
case SelectionWeight.Crowding:
survive = EASurvivalSelection <SolutionGenome> .CreateCrowdingSelector(
(kChoices) => { return(kChoices
.OrderBy((k) => - k.Fitness)
.Select((k, index) => Tuple.Create(k, config.SurvivalSelection.RateP *Math.Pow(1 - config.SurvivalSelection.RateP, index)))
.Select(ki => Tuple.Create(ki.Item1, CmnRandom.Random.NextDouble() < ki.Item2))
.OrderBy(kp => !kp.Item2)
.First().Item1); },
evaluate,
SolutionDistance,
config.SurvivalSelection.SelectPool,
config.NumParents
);
break;
default:
throw new NotImplementedException("Selection weight for survival not found");
}
Func <IEnumerable <SolutionGenome>,
IEnumerable <SolutionGenome> > breed;
switch (config.ParentSelection.SelectionWeight)
{
case SelectionWeight.None:
breed = (population) =>
{
return(Enumerable
.Range(0, config.NumOffspring)
.Select(i => GenerateRandomSolution(shapes, stock, config)));
};
break;
case SelectionWeight.Random:
breed = EAParentSelection <SolutionGenome> .CreateTournamentSelector(
SolutionGenome.GetParentBreeder(stock),
(kChoices) => kChoices.ToList().ChooseSingle(),
evaluate,
config.ParentSelection.SelectPool,
config.ParentSelection.Replacement,
config.NumOffspring
);
break;
case SelectionWeight.Best:
breed = EAParentSelection <SolutionGenome> .CreateTournamentSelector(
SolutionGenome.GetParentBreeder(stock),
(kChoices) => kChoices.MaxByValue((k) => k.Fitness),
evaluate,
config.ParentSelection.SelectPool,
config.ParentSelection.Replacement,
config.NumOffspring
);
break;
case SelectionWeight.Fitness:
breed = EAParentSelection <SolutionGenome> .CreateTournamentSelector(
SolutionGenome.GetParentBreeder(stock),
(kChoices) =>
{
var totalFitness = kChoices.Sum(k => k.Fitness);
var fitPick = CmnRandom.Random.Next(0, totalFitness - 1);
return(kChoices.First(k =>
{
fitPick -= k.Fitness;
return fitPick <= 0;
}));
},
evaluate,
config.ParentSelection.SelectPool,
config.ParentSelection.Replacement,
config.NumOffspring
);
break;
case SelectionWeight.Rank:
breed = EAParentSelection <SolutionGenome> .CreateTournamentSelector(
SolutionGenome.GetParentBreeder(stock),
(kChoices) => kChoices
.OrderBy((k) => - k.Fitness)
.Select((k, index) => Tuple.Create(k, config.ParentSelection.RateP *Math.Pow(1 - config.ParentSelection.RateP, index)))
.Select(ki => Tuple.Create(ki.Item1, CmnRandom.Random.NextDouble() < ki.Item2))
.OrderBy(kp => !kp.Item2)
.First().Item1,
evaluate,
config.ParentSelection.SelectPool,
config.ParentSelection.Replacement,
config.NumOffspring
);
break;
default:
throw new NotImplementedException("Selection weight for parents not handled");
}
Action <IEnumerable <SolutionGenome> > mutator = (population) =>
{
foreach (var individual in population)
{
foreach (var gene in individual.Genes)
{
if (CmnRandom.Random.NextDouble() < individual.RateCreepRandom)
{
gene.CreepRandomize(stock.Length);
}
if (CmnRandom.Random.NextDouble() < individual.RateRotateRandom)
{
gene.RotateRandomize();
}
if (CmnRandom.Random.NextDouble() < individual.RateSlideRandom)
{
gene.SlideRandomize(stock.Length);
}
}
individual.Repair(stock);
}
};
var ea = new EvolveSolution <SolutionGenome>(
initialSolutions,
breed,
mutator,
survive,
terminate
);
var newLogData = new List <Tuple <int, List <double> > >();
foreach (var population in ea.Solve())
{
var avg = population
.Select(s => evaluateSolution(s))
.Aggregate(new List <int> {
0, 0, 0, 0
}, (lhs, rhs) => lhs.Zip(rhs, (l, r) => l + r).ToList())
.Select(total => total / (double)population.Count());
var best = population
.Select(s => evaluateSolution(s))
.Aggregate(new List <int> {
0, 0, 0, 0
}, (lhs, rhs) => lhs.Zip(rhs, (l, r) => Math.Max(l, r)).ToList());
var avgBest = avg
.Zip(best, (a, b) => new List <double>(new double[] { a, b }))
.SelectMany(i => i);
newLogData.Add(Tuple.Create(evalCounter, avgBest.ToList()));
}
logFileData.Add(newLogData);
}