/// <summary>
/// Run the genetic algorithm to train the neural network
/// </summary>
private void ExecuteTrainCommand()
{
this.ExecuteStopCommand();
// The background worker allows the GA to be run in a different thread
BackgroundWorker backgroundWorker = new BackgroundWorker();
// Setup of the genetic algorithm
LanderIndividualSettings landerIndividualSettings = new LanderIndividualSettings();
LanderIndividualFactory landerFactory = new LanderIndividualFactory();
GeneticAlgorithm.GeneticAlgorithm ga = new GeneticAlgorithm.GeneticAlgorithm(landerFactory);
LanderIndividual best = null;
int currentIteration = 0;
// Set up the ga
this.environment.Gravity = 2.0;
this.environment.WindSpeed = 0.1;
landerIndividualSettings.StartingFuel = 100;
landerIndividualSettings.StartingHeight = 100;
landerIndividualSettings.StartingHorizontal = 0;
landerIndividualSettings.LanderEnvironment = this.environment;
landerIndividualSettings.CrossoverAlgorithm = LanderIndividualSettings.CrossoverType.OnePoint;
ga.SelectionType = GeneticAlgorithm.GeneticAlgorithm.SelectionTypes.Tournament;
ga.TournamentSize = 4;
ga.CrossoverProbability = 0;
ga.MutationProbability = 1;
ga.CalculationLimit = 60000;
ga.ElitistCount = 4;
ga.PopulationSize = 100;
landerFactory.IndividualSettings = landerIndividualSettings;
// This lambda function handles the iteration events from the ga.
EventHandler<IterationEventArgs> handler = (sender, args) =>
{
backgroundWorker.ReportProgress(0, args);
};
// Set the work lambda function.
backgroundWorker.DoWork += (sender, e) =>
{
// Add the iteration event handler to report progress back to the main thread
ga.IterationEvent += handler;
// Run the ga
best = (LanderIndividual)ga.Run(landerIndividualSettings);
ga.IterationEvent -= handler;
};
// Update the graph to show progress
backgroundWorker.ProgressChanged += (sender, e) =>
{
this.MinFitnessValues.Add(new Tuple<int, double>(currentIteration, ((IterationEventArgs)e.UserState).MinFitness));
this.MaxFitnessValues.Add(new Tuple<int, double>(currentIteration, ((IterationEventArgs)e.UserState).MaxFitness));
this.AvgFitnessValues.Add(new Tuple<int, double>(currentIteration, ((IterationEventArgs)e.UserState).AverageFitness));
currentIteration++;
if (this.MinFitnessValues.Count > 600)
{
this.MinFitnessValues.Clear();
this.MaxFitnessValues.Clear();
this.AvgFitnessValues.Clear();
}
};
backgroundWorker.RunWorkerCompleted += (sender, e) =>
{
this.IsTraining = false;
// Save the resulting neural network
this.neuralNetwork = best.CurrentNeuralNetwork;
Debug.WriteLine(this.neuralNetwork.ToString());
};
// Clear out previous chart data and run the background worker
this.MinFitnessValues.Clear();
this.MaxFitnessValues.Clear();
this.AvgFitnessValues.Clear();
backgroundWorker.WorkerReportsProgress = true;
this.IsTraining = true;
backgroundWorker.RunWorkerAsync();
}
private static double Train(bool showOutput)
{
LanderIndividualSettings landerIndividualSettings = new LanderIndividualSettings();
LanderIndividualFactory landerFactory = new LanderIndividualFactory();
GeneticAlgorithm ga = new GeneticAlgorithm(landerFactory);
LanderIndividual best = null;
int currentIteration = 0;
double bestfitness = 0;
// Set up the ga
LanderSimulator.Model.Environment environment = new LanderSimulator.Model.Environment();
environment.Gravity = 2.0;
environment.WindSpeed = 0.1;
landerIndividualSettings.StartingFuel = 100;
landerIndividualSettings.StartingHeight = 100;
landerIndividualSettings.StartingHorizontal = 0;
landerIndividualSettings.LanderEnvironment = environment;
landerIndividualSettings.CrossoverAlgorithm = LanderIndividualSettings.CrossoverType.Uniform;
ga.SelectionType = GeneticAlgorithm.SelectionTypes.Tournament;
ga.TournamentSize = 5;
ga.CrossoverProbability = 0.98;
ga.MutationProbability = 1;
ga.CalculationLimit = 60000;
ga.ElitistCount = 10;
ga.PopulationSize = 500;
landerFactory.IndividualSettings = landerIndividualSettings;
// This lambda function handles the iteration events from the ga.
EventHandler<IterationEventArgs> handler = (sender, args) =>
{
if (showOutput)
{
Console.CursorLeft = 0;
Console.Write("{0,4:g}{1,12:f3}{2,12:f3}{3,12:f3}{4,12:f3}{5,12:f3}{6,12:f3}",
currentIteration,
args.MinFitness, args.AverageFitness, args.MaxFitness,
args.MinDifference, args.AverageDifference, args.MaxDifference);
}
currentIteration++;
bestfitness = args.MinFitness;
};
ga.IterationEvent += handler;
if (showOutput)
{
Console.WriteLine("{0,4}{1,12}{2,12}{3,12}{4,12}{5,12}{6,12}",
"", "Min Fit", "Avg Fit", "Max Fit",
"Min Dif", "Avg Dif", "Max Dif");
}
// Run the ga
best = (LanderIndividual)ga.Run(landerIndividualSettings);
Console.WriteLine();
neuralNetwork.SetAllWeights(best.CurrentNeuralNetwork.GetAllWeights());
ga.IterationEvent -= handler;
return bestfitness;
}