private void Initialize() { #region Create parameters Parameters.Add(new ValueLookupParameter <IRandom>("Random", "A pseudo random number generator.")); Parameters.Add(new ValueLookupParameter <BoolValue>("Maximization", "True if the problem is a maximization problem, otherwise false.")); Parameters.Add(new ScopeTreeLookupParameter <DoubleValue>("Quality", "The value which represents the quality of a solution.")); Parameters.Add(new ValueLookupParameter <IOperator>("Selector", "The operator used to select solutions for reproduction.")); Parameters.Add(new ValueLookupParameter <IOperator>("Crossover", "The operator used to cross solutions.")); Parameters.Add(new ValueLookupParameter <PercentValue>("MutationProbability", "The probability that the mutation operator is applied on a solution.")); Parameters.Add(new ValueLookupParameter <IOperator>("Mutator", "The operator used to mutate solutions.")); Parameters.Add(new ValueLookupParameter <IOperator>("Evaluator", "The operator used to evaluate solutions. This operator is executed in parallel, if an engine is used which supports parallelization.")); Parameters.Add(new LookupParameter <IntValue>("EvaluatedSolutions", "The number of evaluated solutions.")); Parameters.Add(new ValueLookupParameter <IntValue>("Elites", "The numer of elite solutions which are kept in each generation.")); Parameters.Add(new ValueLookupParameter <BoolValue>("ReevaluateElites", "Flag to determine if elite individuals should be reevaluated (i.e., if stochastic fitness functions are used.)")); Parameters.Add(new LookupParameter <DoubleValue>("ComparisonFactor", "The comparison factor is used to determine whether the offspring should be compared to the better parent, the worse parent or a quality value linearly interpolated between them. It is in the range [0;1].")); Parameters.Add(new LookupParameter <DoubleValue>("CurrentSuccessRatio", "The current success ratio.")); Parameters.Add(new ValueLookupParameter <DoubleValue>("SuccessRatio", "The ratio of successful to total children that should be achieved.")); Parameters.Add(new LookupParameter <DoubleValue>("SelectionPressure", "The actual selection pressure.")); Parameters.Add(new ValueLookupParameter <DoubleValue>("MaximumSelectionPressure", "The maximum selection pressure that terminates the algorithm.")); Parameters.Add(new ValueLookupParameter <BoolValue>("OffspringSelectionBeforeMutation", "True if the offspring selection step should be applied before mutation, false if it should be applied after mutation.")); Parameters.Add(new ValueLookupParameter <BoolValue>("FillPopulationWithParents", "True if the population should be filled with parent individual or false if worse children should be used when the maximum selection pressure is exceeded.")); #endregion #region Create operators Placeholder selector = new Placeholder(); SubScopesProcessor subScopesProcessor1 = new SubScopesProcessor(); ChildrenCreator childrenCreator = new ChildrenCreator(); ConditionalBranch osBeforeMutationBranch = new ConditionalBranch(); UniformSubScopesProcessor uniformSubScopesProcessor1 = new UniformSubScopesProcessor(); Placeholder crossover1 = new Placeholder(); UniformSubScopesProcessor uniformSubScopesProcessor2 = new UniformSubScopesProcessor(); Placeholder evaluator1 = new Placeholder(); SubScopesCounter subScopesCounter1 = new SubScopesCounter(); WeightedParentsQualityComparator qualityComparer1 = new WeightedParentsQualityComparator(); SubScopesRemover subScopesRemover1 = new SubScopesRemover(); UniformSubScopesProcessor uniformSubScopesProcessor3 = new UniformSubScopesProcessor(); StochasticBranch mutationBranch1 = new StochasticBranch(); Placeholder mutator1 = new Placeholder(); VariableCreator variableCreator1 = new VariableCreator(); VariableCreator variableCreator2 = new VariableCreator(); ConditionalSelector conditionalSelector = new ConditionalSelector(); SubScopesProcessor subScopesProcessor2 = new SubScopesProcessor(); UniformSubScopesProcessor uniformSubScopesProcessor4 = new UniformSubScopesProcessor(); Placeholder evaluator2 = new Placeholder(); SubScopesCounter subScopesCounter2 = new SubScopesCounter(); MergingReducer mergingReducer1 = new MergingReducer(); UniformSubScopesProcessor uniformSubScopesProcessor5 = new UniformSubScopesProcessor(); Placeholder crossover2 = new Placeholder(); StochasticBranch mutationBranch2 = new StochasticBranch(); Placeholder mutator2 = new Placeholder(); UniformSubScopesProcessor uniformSubScopesProcessor6 = new UniformSubScopesProcessor(); Placeholder evaluator3 = new Placeholder(); SubScopesCounter subScopesCounter3 = new SubScopesCounter(); WeightedParentsQualityComparator qualityComparer2 = new WeightedParentsQualityComparator(); SubScopesRemover subScopesRemover2 = new SubScopesRemover(); OffspringSelector offspringSelector = new OffspringSelector(); SubScopesProcessor subScopesProcessor3 = new SubScopesProcessor(); BestSelector bestSelector = new BestSelector(); WorstSelector worstSelector = new WorstSelector(); RightReducer rightReducer = new RightReducer(); LeftReducer leftReducer = new LeftReducer(); MergingReducer mergingReducer2 = new MergingReducer(); ConditionalBranch reevaluateElitesBranch = new ConditionalBranch(); UniformSubScopesProcessor uniformSubScopesProcessor7 = new UniformSubScopesProcessor(); Placeholder evaluator4 = new Placeholder(); SubScopesCounter subScopesCounter4 = new SubScopesCounter(); selector.Name = "Selector (placeholder)"; selector.OperatorParameter.ActualName = SelectorParameter.Name; childrenCreator.ParentsPerChild = new IntValue(2); osBeforeMutationBranch.Name = "Apply OS before mutation?"; osBeforeMutationBranch.ConditionParameter.ActualName = OffspringSelectionBeforeMutationParameter.Name; crossover1.Name = "Crossover (placeholder)"; crossover1.OperatorParameter.ActualName = CrossoverParameter.Name; uniformSubScopesProcessor2.Parallel.Value = true; evaluator1.Name = "Evaluator (placeholder)"; evaluator1.OperatorParameter.ActualName = EvaluatorParameter.Name; subScopesCounter1.Name = "Increment EvaluatedSolutions"; subScopesCounter1.ValueParameter.ActualName = EvaluatedSolutionsParameter.Name; qualityComparer1.ComparisonFactorParameter.ActualName = ComparisonFactorParameter.Name; qualityComparer1.LeftSideParameter.ActualName = QualityParameter.Name; qualityComparer1.MaximizationParameter.ActualName = MaximizationParameter.Name; qualityComparer1.RightSideParameter.ActualName = QualityParameter.Name; qualityComparer1.ResultParameter.ActualName = "SuccessfulOffspring"; subScopesRemover1.RemoveAllSubScopes = true; mutationBranch1.ProbabilityParameter.ActualName = MutationProbabilityParameter.Name; mutationBranch1.RandomParameter.ActualName = RandomParameter.Name; mutator1.Name = "Mutator (placeholder)"; mutator1.OperatorParameter.ActualName = MutatorParameter.Name; variableCreator1.Name = "MutatedOffspring = true"; variableCreator1.CollectedValues.Add(new ValueParameter <BoolValue>("MutatedOffspring", null, new BoolValue(true), false)); variableCreator2.Name = "MutatedOffspring = false"; variableCreator2.CollectedValues.Add(new ValueParameter <BoolValue>("MutatedOffspring", null, new BoolValue(false), false)); conditionalSelector.ConditionParameter.ActualName = "MutatedOffspring"; conditionalSelector.ConditionParameter.Depth = 1; conditionalSelector.CopySelected.Value = false; uniformSubScopesProcessor4.Parallel.Value = true; evaluator2.Name = "Evaluator (placeholder)"; evaluator2.OperatorParameter.ActualName = EvaluatorParameter.Name; subScopesCounter2.Name = "Increment EvaluatedSolutions"; subScopesCounter2.ValueParameter.ActualName = EvaluatedSolutionsParameter.Name; crossover2.Name = "Crossover (placeholder)"; crossover2.OperatorParameter.ActualName = CrossoverParameter.Name; mutationBranch2.ProbabilityParameter.ActualName = MutationProbabilityParameter.Name; mutationBranch2.RandomParameter.ActualName = RandomParameter.Name; mutator2.Name = "Mutator (placeholder)"; mutator2.OperatorParameter.ActualName = MutatorParameter.Name; uniformSubScopesProcessor6.Parallel.Value = true; evaluator3.Name = "Evaluator (placeholder)"; evaluator3.OperatorParameter.ActualName = EvaluatorParameter.Name; subScopesCounter3.Name = "Increment EvaluatedSolutions"; subScopesCounter3.ValueParameter.ActualName = EvaluatedSolutionsParameter.Name; qualityComparer2.ComparisonFactorParameter.ActualName = ComparisonFactorParameter.Name; qualityComparer2.LeftSideParameter.ActualName = QualityParameter.Name; qualityComparer2.MaximizationParameter.ActualName = MaximizationParameter.Name; qualityComparer2.RightSideParameter.ActualName = QualityParameter.Name; qualityComparer2.ResultParameter.ActualName = "SuccessfulOffspring"; subScopesRemover2.RemoveAllSubScopes = true; offspringSelector.CurrentSuccessRatioParameter.ActualName = CurrentSuccessRatioParameter.Name; offspringSelector.MaximumSelectionPressureParameter.ActualName = MaximumSelectionPressureParameter.Name; offspringSelector.SelectionPressureParameter.ActualName = SelectionPressureParameter.Name; offspringSelector.SuccessRatioParameter.ActualName = SuccessRatioParameter.Name; offspringSelector.OffspringPopulationParameter.ActualName = "OffspringPopulation"; offspringSelector.OffspringPopulationWinnersParameter.ActualName = "OffspringPopulationWinners"; offspringSelector.SuccessfulOffspringParameter.ActualName = "SuccessfulOffspring"; offspringSelector.FillPopulationWithParentsParameter.ActualName = FillPopulationWithParentsParameter.Name; bestSelector.CopySelected = new BoolValue(false); bestSelector.MaximizationParameter.ActualName = MaximizationParameter.Name; bestSelector.NumberOfSelectedSubScopesParameter.ActualName = ElitesParameter.Name; bestSelector.QualityParameter.ActualName = QualityParameter.Name; worstSelector.CopySelected = new BoolValue(false); worstSelector.MaximizationParameter.ActualName = MaximizationParameter.Name; worstSelector.NumberOfSelectedSubScopesParameter.ActualName = ElitesParameter.Name; worstSelector.QualityParameter.ActualName = QualityParameter.Name; reevaluateElitesBranch.ConditionParameter.ActualName = "ReevaluateElites"; reevaluateElitesBranch.Name = "Reevaluate elites ?"; uniformSubScopesProcessor7.Parallel.Value = true; evaluator4.Name = "Evaluator (placeholder)"; evaluator4.OperatorParameter.ActualName = EvaluatorParameter.Name; subScopesCounter4.Name = "Increment EvaluatedSolutions"; subScopesCounter4.ValueParameter.ActualName = EvaluatedSolutionsParameter.Name; #endregion #region Create operator graph OperatorGraph.InitialOperator = selector; selector.Successor = subScopesProcessor1; subScopesProcessor1.Operators.Add(new EmptyOperator()); subScopesProcessor1.Operators.Add(childrenCreator); subScopesProcessor1.Successor = offspringSelector; childrenCreator.Successor = osBeforeMutationBranch; osBeforeMutationBranch.TrueBranch = uniformSubScopesProcessor1; osBeforeMutationBranch.FalseBranch = uniformSubScopesProcessor5; osBeforeMutationBranch.Successor = null; uniformSubScopesProcessor1.Operator = crossover1; uniformSubScopesProcessor1.Successor = uniformSubScopesProcessor2; crossover1.Successor = null; uniformSubScopesProcessor2.Operator = evaluator1; uniformSubScopesProcessor2.Successor = subScopesCounter1; evaluator1.Successor = qualityComparer1; qualityComparer1.Successor = subScopesRemover1; subScopesRemover1.Successor = null; subScopesCounter1.Successor = uniformSubScopesProcessor3; uniformSubScopesProcessor3.Operator = mutationBranch1; uniformSubScopesProcessor3.Successor = conditionalSelector; mutationBranch1.FirstBranch = mutator1; mutationBranch1.SecondBranch = variableCreator2; mutationBranch1.Successor = null; mutator1.Successor = variableCreator1; variableCreator1.Successor = null; variableCreator2.Successor = null; conditionalSelector.Successor = subScopesProcessor2; subScopesProcessor2.Operators.Add(new EmptyOperator()); subScopesProcessor2.Operators.Add(uniformSubScopesProcessor4); subScopesProcessor2.Successor = mergingReducer1; uniformSubScopesProcessor4.Operator = evaluator2; uniformSubScopesProcessor4.Successor = subScopesCounter2; evaluator2.Successor = null; subScopesCounter2.Successor = null; mergingReducer1.Successor = null; uniformSubScopesProcessor5.Operator = crossover2; uniformSubScopesProcessor5.Successor = uniformSubScopesProcessor6; crossover2.Successor = mutationBranch2; mutationBranch2.FirstBranch = mutator2; mutationBranch2.SecondBranch = null; mutationBranch2.Successor = null; mutator2.Successor = null; uniformSubScopesProcessor6.Operator = evaluator3; uniformSubScopesProcessor6.Successor = subScopesCounter3; evaluator3.Successor = qualityComparer2; qualityComparer2.Successor = subScopesRemover2; subScopesRemover2.Successor = null; subScopesCounter3.Successor = null; offspringSelector.OffspringCreator = selector; offspringSelector.Successor = subScopesProcessor3; subScopesProcessor3.Operators.Add(bestSelector); subScopesProcessor3.Operators.Add(worstSelector); subScopesProcessor3.Successor = mergingReducer2; bestSelector.Successor = rightReducer; rightReducer.Successor = reevaluateElitesBranch; reevaluateElitesBranch.TrueBranch = uniformSubScopesProcessor7; uniformSubScopesProcessor7.Operator = evaluator4; uniformSubScopesProcessor7.Successor = subScopesCounter4; subScopesCounter4.Successor = null; reevaluateElitesBranch.FalseBranch = null; reevaluateElitesBranch.Successor = null; worstSelector.Successor = leftReducer; leftReducer.Successor = null; mergingReducer2.Successor = null; #endregion }
IEnumerator Start() { var logPath = string.Format("logs_{0}", DateTime.Now.Ticks); Debug.LogFormat("Logging to {0}", logPath); if (!Directory.Exists(logPath)) { Directory.CreateDirectory(logPath); } elitesLog = File.CreateText(Path.Combine(logPath, "elites.csv")); populationLog = File.CreateText(Path.Combine(logPath, "populations.csv")); generationLog = File.CreateText(Path.Combine(logPath, "generations.csv")); speciesLog = File.CreateText(Path.Combine(logPath, "species.csv")); var populationSize = 100; var innovations = new InnovationCollection(); var mutations = new MutationCollection(); mutations.Add(0.01f, new AddNeuronMutator(innovations)); // 0.1% mutations.Add(0.05f, new AddSynapseMutator(innovations)); // 1% mutations.Add(0.05f, new ToggleSynapseMutator(0.125f)); mutations.Add(0.20f, new PerturbNeuronMutator(0.5f, 0.25f)); // 98% vvv mutations.Add(0.20f, new PerturbSynapseMutator(0.5f, 0.25f)); mutations.Add(0.20f, new ReplaceNeuronMutator(0.5f)); mutations.Add(0.20f, new ReplaceSynapseMutator(0.5f)); var eliteSelector = new EliteSelector(); var crossover = new MultipointCrossover(); var offspringSelector = new OffspringSelector(crossover); var distanceMetric = new DistanceMetric(2.0f, 2.0f, 1.0f); var speciation = new Speciation(10, 6.0f, 0.3f, distanceMetric); var neuronGenes = new [] { new NeuronGene(innovations.GetInitialNeuronInnovationId(0), NeuronType.UpperNeuron), new NeuronGene(innovations.GetInitialNeuronInnovationId(1), NeuronType.LowerNeuron), new NeuronGene(innovations.GetInitialNeuronInnovationId(2), NeuronType.PositionNeuron), new NeuronGene(innovations.GetInitialNeuronInnovationId(3), NeuronType.SpeedNeuron), }.ToGeneList(); var synapseGenes = new GeneList <SynapseGene>(); var protoGenotype = new Genotype(neuronGenes, synapseGenes); var genotypes = new GenotypeStream(protoGenotype) .Take(populationSize).ToArray(); var species = new Specie[0]; var generation = 0; var elitePhenotypes = new List <Phenotype>(); var offspringPhenotypes = genotypes.Select(gt => new Phenotype(gt)).ToList(); while (true) { yield return(StartCoroutine(EvaluatePopulation(offspringPhenotypes))); var phenotypes = new List <Phenotype>(elitePhenotypes.Count + offspringPhenotypes.Count); phenotypes.AddRange(elitePhenotypes); phenotypes.AddRange(offspringPhenotypes); Assert.AreEqual(phenotypes.Count, populationSize, "Population size must remain constant"); var longest = phenotypes.OrderByDescending(pt => pt.BestDuration).First(); Debug.LogFormat("[{0}] Fitness: {1}, Duration: {2}s ({3}, {4}) (Longest)", generation, longest.MeanFitness, longest.MeanDuration, longest.Genotype.NeuronCount, longest.Genotype.SynapseCount); var best = phenotypes.OrderByDescending(pt => pt.Fitness).First(); Debug.LogFormat("[{0}] Fitness: {1}, Duration: {2}s ({3}, {4}) (Best)", generation, best.MeanFitness, best.MeanDuration, best.Genotype.NeuronCount, best.Genotype.SynapseCount); elitesLog.WriteLine(string.Join(",", new string[] { generation.ToString(), best.Fitness.ToString(), best.BestDuration.ToString(), JSON.Serialize(best.Genotype.ToJSON()), })); var meanComplexity = phenotypes.Aggregate(0.0f, (sum, pt) => sum + (float)pt.Genotype.Complexity, (sum) => sum / (float)phenotypes.Count); var meanNeuronCount = phenotypes.Aggregate(0.0f, (sum, pt) => sum + (float)pt.Genotype.NeuronCount, (sum) => sum / (float)phenotypes.Count); var meanSynapseCount = phenotypes.Aggregate(0.0f, (sum, pt) => sum + (float)pt.Genotype.SynapseCount, (sum) => sum / (float)phenotypes.Count); var meanFitness = phenotypes.Aggregate(0.0f, (sum, pt) => sum + (float)pt.Fitness, (sum) => sum / (float)phenotypes.Count); var stdevFitness = phenotypes.Aggregate(0.0f, (sum, pt) => sum + Mathf.Pow(pt.Fitness - meanFitness, 2.0f), (sum) => Mathf.Sqrt(sum / (float)phenotypes.Count)); var stdevComplexity = phenotypes.Aggregate(0.0f, (sum, pt) => sum + Mathf.Pow((float)pt.Genotype.Complexity - meanComplexity, 2.0f), (sum) => Mathf.Sqrt(sum / (float)phenotypes.Count)); var stdevNeuronCount = phenotypes.Aggregate(0.0f, (sum, pt) => sum + Mathf.Pow((float)pt.Genotype.NeuronCount - meanNeuronCount, 2.0f), (sum) => Mathf.Sqrt(sum / (float)phenotypes.Count)); var stdevSynapseCount = phenotypes.Aggregate(0.0f, (sum, pt) => sum + Mathf.Pow((float)pt.Genotype.SynapseCount - meanSynapseCount, 2.0f), (sum) => Mathf.Sqrt(sum / (float)phenotypes.Count)); species = speciation.Speciate(species, phenotypes.ToArray()); Debug.LogFormat("[{0}] Species Count: {1} (Threshold: {2})", generation, species.Length, speciation.DistanceThreshold); var meanAdjustedFitness = phenotypes.Aggregate(0.0f, (sum, pt) => sum + (float)pt.AdjustedFitness, (sum) => sum / (float)phenotypes.Count); // standard deviation: // take the square root of the average of the squared deviations of the values from their average value var stdevAdjustedFitness = phenotypes.Aggregate(0.0f, (sum, pt) => sum + Mathf.Pow(pt.AdjustedFitness - meanAdjustedFitness, 2.0f), (sum) => Mathf.Sqrt(sum / (float)phenotypes.Count)); generationLog.WriteLine(new [] { generation, best.Fitness, best.MeanDuration, meanAdjustedFitness, stdevAdjustedFitness, meanFitness, stdevFitness, meanComplexity, stdevComplexity, meanNeuronCount, stdevNeuronCount, meanSynapseCount, stdevSynapseCount }.Stringify()); foreach (var sp in species) { speciesLog.WriteLine(new [] { generation, sp.SpeciesId, sp.Count, sp.BestFitness, sp.MeanFitness, sp.MeanAdjustedFitness, sp.MeanComplexity, }.Stringify()); foreach (var pt in sp) { populationLog.WriteLine(new [] { generation, sp.SpeciesId, pt.Fitness, pt.AdjustedFitness, pt.BestDuration, pt.Genotype.Complexity, pt.Genotype.NeuronCount, pt.Genotype.SynapseCount, }.Stringify()); } } var eliteCount = species.Length; elitePhenotypes = eliteSelector.Select(species, eliteCount); var offspringCount = populationSize - elitePhenotypes.Count; var offspringGenotypes = offspringSelector.Select(species, offspringCount); var mutationResults = mutations.Mutate(offspringGenotypes); Debug.LogFormat("[{0}] Mutations: Added Neurons: {1}, Added Synapses: {2}, Perturbed Neurons: {3}, Perturbed Synapses: {4}, Replaced Neurons: {5}, Replaced Synapses: {6}, Toggled Synapses: {7}, Pruned Synapses: {8}, Orphaned Neurons: {9}", generation, mutationResults.addedNeurons, mutationResults.addedSynapses, mutationResults.perturbedNeurons, mutationResults.perturbedSynapses, mutationResults.replacedNeurons, mutationResults.replacedSynapses, mutationResults.toggledSynapses, mutationResults.prunedSynapses, mutationResults.orphanedNeurons); offspringPhenotypes = offspringGenotypes.Select(gt => new Phenotype(gt)).ToList(); generation++; // Flush these so we can preview results while it runs generationLog.Flush(); populationLog.Flush(); speciesLog.Flush(); } }