public NeuralNetwork(int[] layers, double learning_rate, double momentum) {
_eta = learning_rate;
_alpha = momentum;
_squash = new Sigmoid();
// layers.size is the number of layers
// and each integer in layers is the number of units
// in that layer, layer 0 is input layer layer n is output
// initialize random number generator
_rand = new Random();
// initialize the list of neuron lists
// , the number of neuron lists is the number of layers
_neurons = new List<List<NeuralUnit>>();
// iterate through the layers, newing the number of neurons in each layer
for (int i = 0; i < layers.Length; i++) {
_neurons.Add(new List<NeuralUnit>());
for (int j = 0; j < layers[i]; j++)
_neurons[i].Add(new NeuralUnit(getRandomWeight(), _squash));
}
_last = _neurons.Count - 1;
// initialize the input layer with one input in each input array
foreach (NeuralUnit p in _neurons[0]) {
p.Inputs = new Input[1];
p.Inputs[0] = p.XZero;
}
for (int i = 1; i < _neurons.Count; i++) {
foreach (NeuralUnit p in _neurons[i]) {
p.Inputs = new Input[_neurons[i - 1].Count + 1];
p.Inputs[0] = p.XZero;
}
}
// wire the neurons together, each neuron's output connects
// to one or more neuron inputs, plug output into inputs
for (int i = 0; i < _last; i++) {
for (int j = 0; j < _neurons[i].Count; j++) { //outer layer
NeuralUnit p1 = _neurons[i][j];
// since the output is a shared memory reference between objects
// updating the neuron's output will update all neurons consuming
// the output, also initialize the input weights to random numbers
// inner layer, start adding inputs at one because the first input is x0
foreach (NeuralUnit p2 in _neurons[i + 1]) {
p2.Inputs[j+1] = new Input(ref p1.Oput, getRandomWeight());
}
}
}
//name the neurons
for (int i = 0; i < _neurons.Count; i++) {
for (int j = 0; j < _neurons[i].Count; j++) {
_neurons[i][j].Name = ((char)(i + 105)).ToString() + j.ToString();
_neurons[i][j].Oput.Name = ((char)(i + 105)).ToString() + j.ToString();
}
}
}
private void instantiateAnnFromFile(XmlDocument doc) {
XmlNode node = null;
//get the name of the ANN
//get the squashing finction used
//get the number of layers
//get eta and alpha
//get the units at each layer, their weights,
//and the units they are connected to
foreach (XmlNode n in doc.ChildNodes) {
if (n.Name == "network") {
_name = n.Attributes[0].Value;
string squashFunction = n.Attributes[1].Value;
string layers = n.Attributes[2].Value;
string learningRate = n.Attributes[3].Value;
string learningMomentum = n.Attributes[4].Value;
//init squashing function
_squash = new Sigmoid();
if (squashFunction == "tanh")
_squash = new Tanh();
//init the number of hidden layers
_neurons = new List<List<NeuralUnit>>();
int num = int.Parse(layers);
for (int i = 0; i < num; i++)
_neurons.Add(new List<NeuralUnit>());
_eta = double.Parse(learningRate);
_alpha = double.Parse(learningMomentum);
node = n;
break;
}
}
//parse the IOs
int currentLayer = 0;
foreach (XmlNode n in node.ChildNodes) {
if (n.Name == "layer") {
foreach (XmlNode n2 in n.ChildNodes) {
if (n2.Name == "unit") {
List<double> inputWeights = new List<double>();
string weights = n2.Attributes[1].Value;
while (weights.Contains(",")) {
int i = weights.IndexOf(",");
string weight = weights.Substring(0, i);
if (weight != ",")
inputWeights.Add(double.Parse(weight));
weights = weights.Remove(0, i + 1);
}
NeuralUnit nu = new NeuralUnit(_squash);
nu.Name = n2.Attributes[0].Value;
nu.Oput.Name = nu.Name;
nu.Inputs = new Input[inputWeights.Count];
//this may not be necessary
nu.Inputs[0] = nu.XZero;
nu.Inputs[0].Weight = inputWeights[0];
for (int i = 1; i < inputWeights.Count; i++)
nu.Inputs[i] = new Input(inputWeights[i]);
List<string> connectedUnits = new List<string>();
string units = n2.Attributes[2].Value;
while (units.Contains(",")) {
int i = units.IndexOf(",");
string name = units.Substring(0, i);
if(name != ",")
connectedUnits.Add(name);
units = units.Remove(0, i+1);
}
nu.ConnectedUnits = connectedUnits.ToArray();
_neurons[currentLayer].Add(nu);
}
}
currentLayer++;
}
}
//this is fine for networks without feedback
//for (int i = _neurons.Count - 1; i >= 0; i--) {
// for (int j = 0; j < _neurons[i].Count; j++) {
// NeuralUnit nu = _neurons[i][j];
// for (int k = 0; k < _neurons[i][j].ConnectedUnits.Length; k++) {
// for (int m = 0; m < _neurons[i - 1].Count; m++) {
// if (_neurons[i - 1][m].Name == _neurons[i][j].ConnectedUnits[k]) {
// _neurons[i][j].Inputs[k + 1].Oput = _neurons[i - 1][m].Oput;
// }
// }
// }
// }
//}
_last = _neurons.Count - 1;
// now wire the neurons together
List<NeuralUnit> neurons = new List<NeuralUnit>();
for (int i = 0; i < _neurons.Count; i++) {
for (int j = 0; j < _neurons[i].Count; j++) {
neurons.Add(_neurons[i][j]);
}
}
for (int i = 0; i < neurons.Count; i++) {
for (int j = 0; j < neurons[i].ConnectedUnits.Length; j++) {
for (int k = 0; k < neurons.Count; k++) {
if (neurons[k].Name == neurons[i].ConnectedUnits[j])
neurons[i].Inputs[j+1].Oput = neurons[k].Oput;
}
}
}
}
public NeuralUnit(SquashingFunction sq) {
initialize();
_squash = sq;
}
public NeuralUnit(double weight, SquashingFunction sq) {
initialize();
_xZero.Weight = weight;
_squash = sq;
}
//default constructor
public NeuralUnit() {
initialize();
_squash = new Sigmoid();
}
