/// <summary>
/// Constructs a new neural network.
/// </summary>
/// <returns>
/// A constructed neural network.
/// </returns>
private static ISupervisedLearnerNetwork ConstructNetwork()
{
// Create a backpropagation network for supervised learning.
var network = new BackpropagationNetwork(new RmsCalculator());
// Use two activation functions.
// Linear will be used for the input nodes.
// Sigmoid will be used for all other nodes.
var linearFunction = new LinearActivationFunction();
var sigmoidFunction = new HyperbolicTangentFunction();
// INPUT LAYER: Create an input layer with two neurons.
var inputNeuron1 = new BackpropagationNeuron(GetNextRandomWeight(), linearFunction);
var inputNeuron2 = new BackpropagationNeuron(GetNextRandomWeight(), linearFunction);
network.AddInputNode(inputNeuron1);
network.AddInputNode(inputNeuron2);
// OUTPUT LAYER: Create an ouput layer with one neuron.
var outputNeuron = new BackpropagationNeuron(GetNextRandomWeight(), sigmoidFunction);
network.AddOutputNode(outputNeuron);
// HIDDEN LAYER: Create a hidden layer with three neurons.
// (Note that the hidden layer isn't defined explicitly like the output and input layers.)
var hiddenNeuron1 = new BackpropagationNeuron(GetNextRandomWeight(), sigmoidFunction);
var hiddenNeuron2 = new BackpropagationNeuron(GetNextRandomWeight(), sigmoidFunction);
var hiddenNeuron3 = new BackpropagationNeuron(GetNextRandomWeight(), sigmoidFunction);
// So far, the network isn't usable, because no nodes are ever connected by default.
// Manually wire the input layer to the hidden layer. Notice that we'll connect every
// input neuron to every output neuron in order to create a proper feedforward network.
// If we wanted a different structure, we would use a different wiring pattern.
new BackpropagationConnection(GetNextRandomWeight(), inputNeuron1, hiddenNeuron1);
new BackpropagationConnection(GetNextRandomWeight(), inputNeuron1, hiddenNeuron2);
new BackpropagationConnection(GetNextRandomWeight(), inputNeuron1, hiddenNeuron3);
new BackpropagationConnection(GetNextRandomWeight(), inputNeuron2, hiddenNeuron1);
new BackpropagationConnection(GetNextRandomWeight(), inputNeuron2, hiddenNeuron2);
new BackpropagationConnection(GetNextRandomWeight(), inputNeuron2, hiddenNeuron3);
// Manually wire the hidden layer to the output layer.
new BackpropagationConnection(GetNextRandomWeight(), hiddenNeuron1, outputNeuron);
new BackpropagationConnection(GetNextRandomWeight(), hiddenNeuron2, outputNeuron);
new BackpropagationConnection(GetNextRandomWeight(), hiddenNeuron3, outputNeuron);
return network;
}
/// <summary>
/// Constructs a new neural network with nested sub-networks.
/// </summary>
/// <returns>
/// A constructed neural network.
/// </returns>
private static ISupervisedLearnerNetwork ConstructNestedNetwork()
{
// The activation function for the hidden and output layers.
var sigmoidFunction = new HyperbolicTangentFunction();
// Create a backpropagation network for supervised learning.
var nestedNetwork = new BackpropagationNetwork(new RmsCalculator());
// INPUT LAYER: 2 Sub-networks
// (Each of these sub-networks are used to "associate" the two pairs of inputs which
// represent the OR operations that feed into the XOR operation.)
var subNetwork1 = ConstructSubnetwork();
var subNetwork2 = ConstructSubnetwork();
nestedNetwork.AddInputNode(subNetwork1);
nestedNetwork.AddInputNode(subNetwork2);
// OUTPUT LAYER: Create an ouput layer with one neuron.
var outputNeuron = new BackpropagationNeuron(GetNextRandomWeight(), sigmoidFunction);
nestedNetwork.AddOutputNode(outputNeuron);
// HIDDEN LAYER: Create a hidden layer with three neurons.
// (Note that the hidden layer isn't defined explicitly like the output and input layers.)
var hiddenNeuron1 = new BackpropagationNeuron(GetNextRandomWeight(), sigmoidFunction);
var hiddenNeuron2 = new BackpropagationNeuron(GetNextRandomWeight(), sigmoidFunction);
var hiddenNeuron3 = new BackpropagationNeuron(GetNextRandomWeight(), sigmoidFunction);
// Manually wire the network to be a feedforward network just as in Demo #1.
// Notice that the sub-networks receive connections the same as regular neurons do.
//
// Input Layer to Hidden Layer
new BackpropagationConnection(GetNextRandomWeight(), subNetwork1, hiddenNeuron1);
new BackpropagationConnection(GetNextRandomWeight(), subNetwork1, hiddenNeuron2);
new BackpropagationConnection(GetNextRandomWeight(), subNetwork1, hiddenNeuron3);
new BackpropagationConnection(GetNextRandomWeight(), subNetwork2, hiddenNeuron1);
new BackpropagationConnection(GetNextRandomWeight(), subNetwork2, hiddenNeuron2);
new BackpropagationConnection(GetNextRandomWeight(), subNetwork2, hiddenNeuron3);
// Hidden Layer to Output Layer
new BackpropagationConnection(GetNextRandomWeight(), hiddenNeuron1, outputNeuron);
new BackpropagationConnection(GetNextRandomWeight(), hiddenNeuron2, outputNeuron);
new BackpropagationConnection(GetNextRandomWeight(), hiddenNeuron3, outputNeuron);
return nestedNetwork;
}