public void CheckBpWithSampleNetwork()
{
float GetWeightDifference(NetworkConnection[][] first, NetworkConnection[][] second)
{
var difference = 0f;
for (var i = 0; i < first.Length; i++)
{
for (var j = 0; j < first[i].Length; j++)
{
difference += Math.Abs(first[i][j].Weight - second[i][j].Weight);
}
}
return(difference);
}
float[][] GenerateSampleDataset(RecurrentNetwork sample, int size)
{
var result = new float[size][];
for (var i = 0; i < size; i++)
{
result[i] = new float[3];
var input1 = Neat.Utils.RandomSource.Range(-1f, 1f);
var input2 = Neat.Utils.RandomSource.Range(-1f, 1f);
sample.Sensors[0] = input1;
sample.Sensors[1] = input2;
sample.Activate();
result[i][0] = input1;
result[i][1] = input2;
result[i][2] = sample.Effectors[0];
}
return(result);
}
// Arrange
var sampleConnections =
new NetworkConnection[BiasCount + 4][]; // 0 : bias, 1, 2: inputs, 3: output, 4: hidden
sampleConnections[0] = new NetworkConnection[1];
sampleConnections[0][0] = new NetworkConnection(3, 1f); // bias to output
sampleConnections[1] = new NetworkConnection[1];
sampleConnections[1][0] = new NetworkConnection(4, 0.5f); // input to hidden
sampleConnections[2] = new NetworkConnection[1];
sampleConnections[2][0] = new NetworkConnection(4, -0.5f); // input to hidden
sampleConnections[3] = new NetworkConnection[1];
sampleConnections[3][0] = new NetworkConnection(4, 0.2f); // output to hidden
sampleConnections[4] = new NetworkConnection[1];
sampleConnections[4][0] = new NetworkConnection(3, -1f); // hidden to output
var sampleNetwork = new RecurrentNetwork(sampleConnections, 3, 1, new DummyNeatChromosomeEncoder());
var targetConnections =
new NetworkConnection[BiasCount + 4][]; // 0 : bias, 1, 2: inputs, 3: output, 4: hidden
targetConnections[0] = new NetworkConnection[1];
targetConnections[0][0] = new NetworkConnection(3, 0.8f); // bias to output
targetConnections[1] = new NetworkConnection[1];
targetConnections[1][0] = new NetworkConnection(4, 0.7f); // input to hidden
targetConnections[2] = new NetworkConnection[1];
targetConnections[2][0] = new NetworkConnection(4, -0.3f); // input to hidden
targetConnections[3] = new NetworkConnection[1];
targetConnections[3][0] = new NetworkConnection(4, 0.1f); // output to hidden
targetConnections[4] = new NetworkConnection[1];
targetConnections[4][0] = new NetworkConnection(3, -0.8f); // hidden to output
var targetNetwork = new RecurrentNetwork(targetConnections, 3, 1, new DummyNeatChromosomeEncoder());
// Act
var beforeTrainDifference = GetWeightDifference(sampleConnections, targetConnections);
for (var i = 0; i < 5000; i++)
{
targetNetwork.Train(GenerateSampleDataset(sampleNetwork, 5), 0.5f);
}
var afterTrainDifference = GetWeightDifference(sampleConnections, targetConnections);
// Assert
Assert.True(afterTrainDifference < beforeTrainDifference, "Training failed!");
}
public void CheckRecurrentBp()
{
float GetSampleSetError(float[][] floats, RecurrentNetwork network1)
{
var error = 0f;
foreach (var sample in floats)
{
network1.Sensors[0] = sample[0];
network1.Activate();
error += Math.Abs(sample[1] - network1.Effectors[0]);
}
return(error);
}
void FlushNetworkState(RecurrentNetwork recurrentNetwork)
{
for (var i = 0; i < 3; i++)
{
recurrentNetwork.Sensors[0] = 0f;
recurrentNetwork.Activate();
}
}
float[][] GenerateSampleDataset(int size)
{
var result = new float[size][];
var previous = 0f;
var current = 0f;
for (var i = 0; i < size; i++)
{
result[i] = new float[2];
result[i][1] = current + previous; // previous inputs sum for output
previous = current;
current = Neat.Utils.RandomSource.Range(-0.5f, 0.5f);
result[i][0] = current;
}
return(result);
}
// Arrange
var connections = new NetworkConnection[BiasCount + 3][]; // 0 : bias, 1: input, 2: output, 3: hidden
connections[0] = new NetworkConnection[1];
connections[0][0] = new NetworkConnection(2, 1f); // bias to output
connections[1] = new NetworkConnection[2];
connections[1][0] = new NetworkConnection(2, 2f); // input to output
connections[1][1] = new NetworkConnection(3, -1f); // input to hidden
connections[2] = new NetworkConnection[0]; // no output self-links
connections[3] = new NetworkConnection[1];
connections[3][0] = new NetworkConnection(2, 0f); // hidden to output
var network = new RecurrentNetwork(connections, 2, 1, new DummyNeatChromosomeEncoder());
// Act
FlushNetworkState(network);
var sampleDataSet = GenerateSampleDataset(10);
var beforeTrainError = GetSampleSetError(sampleDataSet, network);
for (var i = 0; i < 20; i++)
{
// flush network state
FlushNetworkState(network);
network.Train(sampleDataSet, 0.1f);
}
FlushNetworkState(network);
var afterTrainError = GetSampleSetError(sampleDataSet, network);
for (var i = 0; i < 20; i++)
{
// flush network state
FlushNetworkState(network);
network.Train(sampleDataSet, 0.1f);
}
FlushNetworkState(network);
var afterSecondTrainError = GetSampleSetError(sampleDataSet, network);
// Assert
Assert.True(afterTrainError < beforeTrainError, "Training failed!");
Assert.True(afterSecondTrainError < afterTrainError, "Second training failed!");
}