public void RunDemo()
{
Console.WriteLine("### BUILDING BLOCK DEMO ###");
//Initialize the pattern matching building block
//for a 2x2 grid and 2 output neurons.
Grid2DControler gpm = new Grid2DControler(2, 2, 2);
gpm.BuildNetwork(true);
//Add 4 patterns to the pattern manager.
//note that the input array is in row by row order,
//and the output numers the (single!) output neuron to train for.
gpm.Patterns.Add("first", 0, new bool[] { true, false, true, false });
gpm.Patterns.Add("second", 1, new bool[] { true, true, false, false });
gpm.Patterns.Add("third", 1, new bool[] { false, true, false, false });
gpm.Patterns.Add("fourth", 0, new bool[] { false, true, true, false });
//Some configuration ...
gpm.TrainingConfiguration.AutoTrainingEpochs.Value = 500;
//Calculate all patterns. The CalculateCurrentNetwork method
//returns the output neuron index that matches best.
//First select a pattern, then calculate it:
Console.WriteLine("# Initial Status");
for (int i = 0; i < gpm.PatternCount; i++)
{
gpm.SelectPattern(i);
Console.Write(gpm.CalculateCurrentNetwork() + ": ");
App.PrintArray(gpm.NeuralNetwork.CollectOutput());
}
Console.WriteLine("Successful Patterns: " + gpm.CountSuccessfulPatterns());
//Train the third pattern one time:
Console.WriteLine("# Third Pattern Training");
gpm.SelectPattern(2);
gpm.TrainCurrentNetwork();
for (int i = 0; i < gpm.PatternCount; i++)
{
gpm.SelectPattern(i);
Console.Write(gpm.CalculateCurrentNetwork() + ": ");
App.PrintArray(gpm.NeuralNetwork.CollectOutput());
}
Console.WriteLine("Successful Patterns: " + gpm.CountSuccessfulPatterns());
//Autotrain all Patterns:
Console.WriteLine("# Auto Training:");
Console.WriteLine(gpm.AutoTrainNetwork() ? "successful" : "failed");
for (int i = 0; i < gpm.PatternCount; i++)
{
gpm.SelectPattern(i);
Console.Write(gpm.CalculateCurrentNetwork() + ": ");
App.PrintArray(gpm.NeuralNetwork.CollectOutput());
}
Console.WriteLine("Successful Patterns: " + gpm.CountSuccessfulPatterns());
Console.WriteLine("=== COMPLETE ===");
Console.WriteLine();
}
public PatternMatching(int x, int y, int output)
{
InitializeComponent();
ProgressBegin();
pm = new Grid2DControler(x, y, output);
BasicConfig bc = pm.BasicConfiguration;
TrainingConfig tc = pm.TrainingConfiguration;
Grid2DConfig gc = new Grid2DConfig(bc.Node);
bc.ActivationType.Value = EActivationType.Symmetric;
bc.BiasNeuronEnable.Value = true;
bc.BiasNeuronOutput.Value = 0.9d;
bc.DeadNeuronDecayEnabled.Value = false;
bc.FlatspotEliminationEnable.Value = true;
bc.FlatspotElimination.Value = 0.05d;
bc.InitialSymmetryBreaking.Value = 0.2d;
bc.LearningRate.Value = 0.3d;
bc.ManhattanTrainingEnable.Value = false;
bc.MomentumTermEnable.Value = false;
bc.SymmetryPreventionEnable.Value = true;
bc.SymmetryPrevention.Value = 0.05d;
bc.WeightDecayEnable.Value = true;
bc.WeightDecay.Value = 0.01d;
bc.LowInput.Value = -3d;
bc.HighInput.Value = 3d;
bc.LowOutput.Value = -0.95d;
bc.HighOutput.Value = 0.95d;
tc.AutoTrainingAttempts.Value = 1;
tc.AutoTrainingEpochs.Value = 400;
tc.AutoTrainingPercentSuccessful.Value = 1.0d;
tc.ShuffleEnable.Value = true; // make it more robust for noise
tc.ShuffleNoiseSigma.Value = 0.2d;
tc.ShuffleSwapProbability.Value = 0.2d; // 0.2 = every 5ht pixel is swapped (20% probability) !!
gc.All2AllEnable.Value = false;
gc.HorizontalLinesEnable.Value = true;
gc.VerticalLinesEnable.Value = true;
gc.RingsEnable.Value = true;
gc.LittleSquaresEnable.Value = true;
pm.BuildNetwork(true);
//pm.NetworkTrainer = new SimpleTrainer();
//pm.NetworkTrainer = new OpenloopTrainer();
//pm.NetworkTrainer = new ConditionalTrainer();
pm.NetworkTrainer = new FeedbackTrainer();
ProgressStatus(50);
pm.Patterns.PatternAdded += pm_OnNewPatternAdded;
pm.PatternSelectionChanged += pm_OnPatternSelectionChanged;
pm.InputChanged += pm_OnInputChanged;
pm.NetworkRebuilt += pm_OnNetworkRebuilt;
outputStat.Init(pm);
inputGrid.Init(pm);
inputGrid.OnGridChanged += inputGrid_OnGridChanged;
ProgressEnd();
}