Migration uses selection by torunament

Author: salvacorts

mlp-ea-centralized/native/client/client.go

@@ -1,8 +1,6 @@
 package main
 
 import (
-	"os"
-
 	"github.com/salvacorts/TFG-Parasitic-Metaheuristics/mlp-ea-centralized/common/ga"
 	"google.golang.org/grpc/grpclog"
 
@@ -18,7 +16,7 @@ func main() {
 
 	logrus.SetLevel(logrus.ErrorLevel)
 	client.Log.SetLevel(logrus.InfoLevel)
-	grpclog.SetLoggerV2(grpclog.NewLoggerV2(os.Stdout, os.Stdout, os.Stdout))
+	grpclog.SetLoggerV2(grpclog.NewLoggerV2(nil, nil, ga.Log.Out))
 
 	err := client.Start()
 	if err != nil {

mlp-ea-centralized/native/server/server.go

@@ -22,7 +22,7 @@ func main() {
 	logrus.SetLevel(logrus.ErrorLevel)
 	ga.Log.SetOutput(os.Stdout)
 	ga.Log.SetLevel(logrus.InfoLevel)
-	grpclog.SetLoggerV2(grpclog.NewLoggerV2(ga.Log.Out, ga.Log.Out, ga.Log.Out))
+	grpclog.SetLoggerV2(grpclog.NewLoggerV2(nil, nil, ga.Log.Out))
 
 	_, score, err := ga.TrainMLP(string(fileContent))
 	if err != nil {

mlp-ea-decentralized/common/ga/poolModel.go

@@ -161,8 +161,8 @@ func (pool *PoolModel) Minimize() {
 		for i := range offsprings {
 			for _, op := range pool.ExtraOperators {
 				if pool.Rnd.Float64() <= op.Probability {
-					offsprings[i].Evaluated = false
 					offsprings[i].Genome = op.Operator(offsprings[i].Genome, pool.Rnd)
+					offsprings[i].Evaluated = false
 				}
 			}
 		}
@@ -215,7 +215,6 @@ func (pool *PoolModel) selection(nOffstrings, nCandidates int) []eaopt.Individua
 	return offsprings
 }
 
-// TODO: Get rid of odd arrays here
 func (pool *PoolModel) crossover(in []eaopt.Individual) []eaopt.Individual {
 	offsprings := make([]eaopt.Individual, len(in))
 	for i := 0; i < len(in)-1; i += 2 {
@@ -241,8 +240,8 @@ func (pool *PoolModel) crossover(in []eaopt.Individual) []eaopt.Individual {
 func (pool *PoolModel) mutate(in []eaopt.Individual) []eaopt.Individual {
 	for i := range in {
 		if pool.Rnd.Float64() < pool.MutRate {
-			in[i].Evaluated = false
 			in[i].Genome.Mutate(pool.Rnd)
+			in[i].Evaluated = false
 		}
 	}
 
@@ -513,16 +512,8 @@ func (pool *PoolModel) migrationScheduler() {
 			var conn *grpc.ClientConn
 			dialed := false
 
-			snap := pool.population.Snapshot()
-			indivArr := make([]eaopt.Individual, 0, len(snap))
+			indivArr := pool.selection(pool.NMigrate, 4)
 			migrate := make([]Individual, pool.NMigrate)
-
-			for _, item := range snap {
-				indivArr = append(indivArr, item.Object.(eaopt.Individual).Clone(pool.Rnd))
-			}
-
-			// Sort population and get NMigrate best
-			indivArr = pool.SortFunc(indivArr, pool.SortPrecission)
 			for i := range migrate {
 				migrate[i] = Individual{
 					IndividualID: indivArr[i].ID,

mlp-ea-decentralized/common/mlp/operators.go

@@ -31,7 +31,6 @@ func NewRandMLP(rng *rand.Rand) eaopt.Genome {
 }
 
 // Evaluate a MLP by getting its accuracy
-// TODO: Compare also number of neurons
 func (nn *MultiLayerNetwork) Evaluate() (float64, error) {
 	copy := nn.Clone().(*MultiLayerNetwork)
 

mlp-ea-decentralized/native/server/server.go

@@ -47,7 +47,7 @@ func main() {
 		Classes:     mapped,
 		TrainingSet: train,
 		TrainEpochs: 100,
-		MutateRate:  0.3,
+		MutateRate:  1,
 		FactoryCfg: mlp.MLPFactoryConfig{
 			InputLayers:      len(patterns[0].Features),
 			OutputLayers:     len(mapped),
@@ -80,7 +80,7 @@ func main() {
 		eaopt.ExtraOperator{Operator: mlp.AddNeuron, Probability: 0.3},
 		eaopt.ExtraOperator{Operator: mlp.RemoveNeuron, Probability: 0.15},
 		eaopt.ExtraOperator{Operator: mlp.SubstituteNeuron, Probability: 0.15},
-		eaopt.ExtraOperator{Operator: mlp.Train, Probability: 0.3},
+		eaopt.ExtraOperator{Operator: mlp.Train, Probability: 0.5},
 	}
 	pool.BestCallback = func(pool *ga.PoolModel) {
 		logrus.WithFields(logrus.Fields{

mlp-ea/common/ga.go

@@ -34,6 +34,7 @@ func TrainMLP(csvdata string) (mn.MultiLayerNetwork, float64, error) {
 	ga.NPops = 1
 	ga.PopSize = popSize
 	ga.Model = getGenerationalModelTournament(4)
+	//ga.Model = getSteadyStateModel()
 	ga.Callback = func(ga *eaopt.GA) {
 		Log.WithFields(logrus.Fields{
 			"level":               "info",
@@ -50,8 +51,8 @@ func TrainMLP(csvdata string) (mn.MultiLayerNetwork, float64, error) {
 
 	// Configure MLP
 	Config = MLPConfig{
-		Epochs:      10,
-		Folds:       1,
+		Epochs:      50,
+		Folds:       5,
 		Classes:     &mapped,
 		TrainingSet: &train,
 		FactoryCfg: MLPFactoryConfig{

mlp-ea/common/mlp.go

@@ -38,16 +38,16 @@ func NewRandMLP(rng *rand.Rand) eaopt.Genome {
 func (mlp *MLP) Evaluate() (float64, error) {
 	copy := mlp.Clone().(*MLP)
 
-	train, validation := mv.TrainTestPatternSplit(*Config.TrainingSet, 0.8, 1)
+	scores := mv.MLPKFoldValidation(
+		(*mn.MultiLayerNetwork)(copy), *Config.TrainingSet, Config.Epochs, Config.Folds, 1, *Config.Classes)
 
-	mn.MLPTrain((*mn.MultiLayerNetwork)(copy), train, *Config.Classes,
-		Config.Epochs, true)
-
-	predictions := utils.PredictN((*mn.MultiLayerNetwork)(copy), validation)
-	predictionsR := utils.RoundPredictions(predictions)
-	_, acc := utils.AccuracyN(predictionsR, validation)
+	mean := 0.0
+	for _, s := range scores {
+		mean += s
+	}
+	mean /= float64(len(scores))
 
-	return 100 - acc, nil
+	return 100 - mean, nil
 }
 
 // Mutate modifies the weights of certain neurons, at random, depending on the application rate.