public BpNeuron(int Layer, int NeuronNumber, NeuronClass NeuronType, ArtificialNetwork AINetwork)
{
this.Location = new Vector2(Layer, NeuronNumber);
this.NeuronType = NeuronType;
this.Bias = NeuralFunctions.NextDouble(MIN, MAX);
this.AINetwork = AINetwork;
}
public unsafe override async Task <double> Fire(List <double> Inputsf, NetworkState State)
{
if (!((BackPropagationNetwork)this.AINetwork).TrainingMode)
{
if (((BackPropagationNetwork)this.AINetwork).HiddenLayers + 1 == this.GetLocation().X)
{
Program.WriteLine("NeuronType: " + this.ToString());
Program.WriteLine("Editing neuron class from HIDDEN to OUTPUT");
this.NeuronType = NeuronClass.OUTPUT;
}
}
//Console.SetOut(Program.Debug);
//Console.WriteLine("Firing Neuron " + this.ToString());
List <double> Weights = GetWeights(Inputsf.Count, State);
//Console.WriteLine("Received Weights, count = " + Weights.Count);
List <double> Inputs = Inputsf.ToList();
if (Inputs.Count != Weights.Count)
{
throw new BrainFireException("The number of inputs != the number of weights. Aborting");
}
if (State == NetworkState.FORWARD_PROPAGATION)
{
if (NeuronType == NeuronClass.HIDDEN)
{
//simply multiply and sum. Then run through Sigmoid function
double value = 0.0;
for (int i = 0; i < Inputs.Count; i++)
{
value += Inputs[i] * Weights[i];
}
//Console.WriteLine("Total Sum: " + value);
value += this.Bias;
//Console.WriteLine("Total Sum (Bias) : " + value);
value = NeuralFunctions.SimpleSquash(value);
//Console.WriteLine("Sigmoid: " + value);
this.Output = value;
}
else if (NeuronType == NeuronClass.OUTPUT)
{
//multiply and sum, then calculate this neurons error. Then, update the bias term
double value = 0.0;
for (int i = 0; i < Inputs.Count; i++)
{
value += Inputs[i] * Weights[i];
}
value += this.Bias;
//Program.WriteLine("Total Sum: " + value);
Console.SetOut(Program.ConsoleOut);
this.Output = NeuralFunctions.SimpleSquash(value);
Console.SetOut(Program.Debug);
//NeuralFunctions.TanSigmoid(value, &value);
if (((BackPropagationNetwork)this.AINetwork).TrainingMode)
{
//Console.WriteLine("Sigmoid: " + value);
double T = ((BackPropagationNetwork)this.AINetwork).ExpectedOutputs[this.GetLocation().Y];
this.DeltaKOutput = this.Output * (1 - this.Output) * (T - this.Output);
//Console.WriteLine("DeltaK: " + this.DeltaKOutput);
//Update bias
double DeltaBias = this.DeltaKOutput * ((BackPropagationNetwork)this.AINetwork).LearningRate;
//Console.WriteLine("Delta Bias: " + DeltaBias);
this.DeltaBiasOutput = DeltaBias;
this.Bias += this.DeltaBiasOutput;
//Console.WriteLine("New Bias: " + Bias);
}
//this.Output = this.DeltaKOutput;
}
}
else if (State == NetworkState.BACK_PROPAGATION)
{
if (NeuronType == NeuronClass.HIDDEN)
{
//primary role of the neuron in this state is to update weights and biases
//required Inputs[] are DeltaK values from output neuron
double DeltaX = 0.0;
for (int i = 0; i < Inputs.Count; i++)
{
//Console.WriteLine("Operation[+]: " + Inputs[i] + " x " + Weights[i]);
DeltaX += Weights[i] * Inputs[i];
}
//Console.WriteLine("Total Sum: " + DeltaX);
//get previous output
DeltaX = DeltaX * (1 - this.Output) * this.Output;
//Console.WriteLine("DeltaX: " + DeltaX);
double DeltaBias = DeltaX * ((BackPropagationNetwork)this.AINetwork).LearningRate;
//Console.WriteLine("Delta Bias: " + DeltaBias);
//Console.WriteLine("Old Bias: " + this.Bias);
if (this.DeltaW == null)
{
this.DeltaW = DeltaX * ((BackPropagationNetwork)this.AINetwork).LearningRate * this.Output;
for (int i = 0; i < Inputs.Count; i++)
{
((BackPropagationNetwork)this.AINetwork).Weights[this.GetLocation().X][i + (this.Location.Y * Inputs.Count)] += this.DeltaW;
}
}
else
{
double DeltaWCur = DeltaX * ((BackPropagationNetwork)this.AINetwork).LearningRate * this.Output;
this.DeltaW = DeltaWCur + (((BackPropagationNetwork)this.AINetwork).Momentum * this.DeltaW); //Momentum term
for (int i = 0; i < Inputs.Count; i++)
{
((BackPropagationNetwork)this.AINetwork).Weights[this.GetLocation().X][i + (this.Location.Y * Inputs.Count)] += this.DeltaW;
}
}
this.Bias += DeltaBias;
//Console.WriteLine("New Bias: " + this.Bias);
this.Output = DeltaX;
}
else if (NeuronType == NeuronClass.INPUT)
{
for (int i = 0; i < Inputs.Count; i++)
{
double WeightDelta = Inputs[i] * ((BackPropagationNetwork)this.AINetwork).LearningRate * this.InputNeuronInput;
//Console.WriteLine("Weight Before: " + ((BackPropagationNetwork)this.AINetwork).Weights[this.GetLocation().X][i + (this.Location.Y * Inputs.Count)]);
//Console.WriteLine("Weight Delta: " + WeightDelta);
((BackPropagationNetwork)this.AINetwork).Weights[this.GetLocation().X][i + (this.Location.Y * Inputs.Count)] += WeightDelta;
//Console.WriteLine("New Weight: " + ((BackPropagationNetwork)this.AINetwork).Weights[this.GetLocation().X][i + (this.Location.Y * Inputs.Count)]);
}
}
}
Console.SetOut(Program.Debug);
Console.WriteLine("Output: " + this.Output);
return(this.Output);
}
