// main function of the neural network to train and store the outputs using backprop & feedforward
public void compute()
{
int epoch = 0; // used for iterating through tests N times
// hidden layer of neurons
Neuron hidden1 = new Neuron();
Neuron hidden2 = new Neuron();
Neuron hidden3 = new Neuron();
Neuron hidden4 = new Neuron();
//output neuron
Neuron output = new Neuron();
output.randomizeWeights();
hidden1.randomizeWeights();
hidden2.randomizeWeights();
hidden3.randomizeWeights();
hidden4.randomizeWeights();
//double errorRate = 1;
//output.error = 1;
while (epoch < 100) // iterate based on the output error OR use epochs > 100 for reliable numbers
{
//errorRate = 0;
epoch++;
for (int i = 0; i < 16; i++)
{
// set each neurons inputs to the training set
hidden1.inputs = new double[] { inputs[i, 0], inputs[i, 1], inputs[i, 2], inputs[i, 3] };
hidden2.inputs = new double[] { inputs[i, 0], inputs[i, 1], inputs[i, 2], inputs[i, 3] };
hidden3.inputs = new double[] { inputs[i, 0], inputs[i, 1], inputs[i, 2], inputs[i, 3] };
hidden4.inputs = new double[] { inputs[i, 0], inputs[i, 1], inputs[i, 2], inputs[i, 3] };
// set the outputnode inputs to the results of the 4 neurons sigmoid transfer
output.inputs = new double[] { hidden1.output(), hidden2.output(), hidden3.output(), hidden4.output() };
// display the training set and the "output of the output" neuron
// we set the output nodes inputs above, and now we're using those inputs * weights + bias formula in the output function
textBox1.AppendText(inputs[i, 0] + " " + inputs[i, 1] + " " + inputs[i, 2] + " " + inputs[i, 3] + " = " + output.output() + "\r\n");
outputResults.Add(output.output());
neuronOutputs.Add(new double[] { hidden1.output(), hidden2.output(), hidden3.output(), hidden4.output() });
// calculate the error
output.error = sigmoid.derivative(output.output()) * (answers[i] - output.output());
output.tweakWeights();
// calculate the neurons error rates based on their individual outputs * the error rate of the output neuron * the weights of the output neuron
// the sigmoid deriv is used for adjusting the weights in order to reduce the error rate of each neuron
hidden1.error = sigmoid.derivative(hidden1.output()) * output.error * output.weights[0];
hidden2.error = sigmoid.derivative(hidden2.output()) * output.error * output.weights[1];
hidden3.error = sigmoid.derivative(hidden3.output()) * output.error * output.weights[2];
hidden4.error = sigmoid.derivative(hidden4.output()) * output.error * output.weights[3];
// the line below is irrelevant since our sigmoid function takes care of the error rate
//double error = answers[i] - output.output();
// tweak the neurons weights using the error rates by adding the error rate * input value to the weights
hidden1.tweakWeights();
hidden2.tweakWeights();
hidden3.tweakWeights();
hidden4.tweakWeights();
//errorRate += Math.Abs(error);
}
}
}
