// Constructor //
public NeuralNetwork(string Name, NodeLinkMaster Links, NodeSet Nodes, ResponseNodeSet Responses, NeuralRule Rule, Matrix Data)
{
// Set values //
this._Links = Links;
this._Nodes = Nodes;
this._Responses = Responses;
this._Rule = Rule;
this._Data = Data;
this._name = Name;
// Initialize //
this.Initialize();
}
public RProp(ResponseNodeSet Master)
: this(Master, 0.5, 1.2)
{
}
public RProp(ResponseNodeSet Master, double Down, double Up)
: base(Master)
{
this._DownAxon = Down;
this._UpAxon = Up;
}
public BProp(ResponseNodeSet Master)
: this(Master, 0.25, 0.025)
{
}
public BProp(ResponseNodeSet Master, double LearningRate, double Momentum)
: base(Master)
{
this._LearningRate = LearningRate;
this._Momentum = Momentum;
}
public HProp(ResponseNodeSet Master)
: base(Master)
{
}
public HProp(ResponseNodeSet Master, double Down, double Up)
: base(Master, Down, Up)
{
}
public QProp(ResponseNodeSet Master, double Slope, double Acceleration, double Scale)
: base(Master)
{
this._Slope = Slope;
this._Acceleration = Acceleration;
this._Scale = Scale;
}
public MProp(ResponseNodeSet Master, double Gradient, double Momentum)
: base(Master)
{
this._GradientWeight = Gradient;
this._MomentumWeight = Momentum;
}
public iRPropMinus(ResponseNodeSet Master)
: base(Master)
{
}
public iRPropMinus(ResponseNodeSet Master, double Down, double Up)
: base(Master, Down, Up)
{
}
public NeuralRule(ResponseNodeSet Master)
{
this._master = Master;
}
public RPropPlus(ResponseNodeSet Master)
: base(Master)
{
}
public static NeuralRule Construct(string Token, ResponseNodeSet Master)
{
string[] toks = Token.ToLower().Split(',',';');
for (int i = 0; i < toks.Length; i++)
{
toks[i] = toks[i].Trim();
}
switch (toks[0])
{
case "bprop":
double bp_lr = (toks.Length < 2 ? 0.35 : double.Parse(toks[1]));
double bp_m = (toks.Length < 3 ? 0.035 : double.Parse(toks[2]));
return new BProp(Master, bp_lr, bp_m);
case "rprop":
double rp0_down = (toks.Length < 2 ? 0.5 : double.Parse(toks[1]));
double rp0_up = (toks.Length < 3 ? 1.2 : double.Parse(toks[2]));
return new RProp(Master, rp0_down, rp0_up);
case "rprop+":
double rp1_down = (toks.Length < 2 ? 0.5 : double.Parse(toks[1]));
double rp1_up = (toks.Length < 3 ? 1.2 : double.Parse(toks[2]));
return new RPropPlus(Master, rp1_down, rp1_up);
case "rprop-":
double rp2_down = (toks.Length < 2 ? 0.5 : double.Parse(toks[1]));
double rp2_up = (toks.Length < 3 ? 1.2 : double.Parse(toks[2]));
return new RPropMinus(Master, rp2_down, rp2_up);
case "irprop+":
double rp3_down = (toks.Length < 2 ? 0.5 : double.Parse(toks[1]));
double rp3_up = (toks.Length < 3 ? 1.2 : double.Parse(toks[2]));
return new iRPropPlus(Master, rp3_down, rp3_up);
case "irprop-":
double rp4_down = (toks.Length < 2 ? 0.5 : double.Parse(toks[1]));
double rp4_up = (toks.Length < 3 ? 1.2 : double.Parse(toks[2]));
return new iRPropMinus(Master, rp4_down, rp4_up);
case "mprop":
double mp_lr = (toks.Length < 2 ? 0.01 : double.Parse(toks[1]));
double mp_m = (toks.Length < 3 ? 0.001 : double.Parse(toks[2]));
return new MProp(Master, mp_lr, mp_m);
case "qprop":
return new QProp(Master);
case "hprop":
double hp_down = (toks.Length < 2 ? 0.5 : double.Parse(toks[1]));
double hp_up = (toks.Length < 3 ? 1.2 : double.Parse(toks[2]));
return new HProp(Master, hp_down, hp_up);
}
throw new ArgumentException(string.Format("Rule '{0}' is invalid", Token));
}
// Render the network //
public NeuralNetwork Construct()
{
// Create the scrubber //
NeuralDataFactory scrubber = new NeuralDataFactory(this._SourceData, this._XValues, this._YValues, this._Where);
// Create the data layer //
this._DataLayer = new NN_Layer(this._DataBias, scrubber.InputKey, scrubber.Columns);
// Create the prediction layer //
this._PredictionLayer = new NN_Layer(this._PredictionReducer, this._PredictionActivation, scrubber.OutputKey);
// Link predictions to last hidden //
if (this._HiddenLayers.Count == 0)
{
NN_Layer.LinkChildren(this._Links, this._PredictionLayer, this._DataLayer);
}
else
{
// Data -> First.Hidden
NN_Layer.LinkChildren(this._Links, this._HiddenLayers.First(), this._DataLayer);
// Last.Hidden -> Predictions //
NN_Layer.LinkChildren(this._Links, this._PredictionLayer, this._HiddenLayers.Last());
// Hidden -> Hidden //
for (int i = 0; i < this._HiddenLayers.Count - 1; i++)
NN_Layer.LinkChildren(this._Links, this._HiddenLayers[i + 1], this._HiddenLayers[i]);
}
// Build a nodeset //
NodeSet nodes = new NodeSet(this._Links);
// Build a response set //
ResponseNodeSet responses = new ResponseNodeSet(this._Links);
// Neural rule //
NeuralRule rule = RuleFactory.Construct(this.DefaultRule, responses);
// Construct //
return new NeuralNetwork(this._name, this._Links, nodes, responses, rule, scrubber.ScrubbedData);
}
