public RLPolicy(int _Input, int _Output, int _HL, String Letter, String LearnType, int ERun)
{
NetWorkInit = new NetworkParse[3];
ReportTrainingCSV = new Reporter(_HL, Letter, LearnType, ERun);
// int HiddenLayerSize = _Input / 3 * 2 + _Output;
int HiddenLayerSize = _HL;
NetWorkInit[0] = new NetworkParse(_Input, HiddenLayerSize, "INPUT", true, "None");
NetWorkInit[1] = new NetworkParse(HiddenLayerSize, _Output, "HIDDEN", true, "Sigmoid");
NetWorkInit[2] = new NetworkParse(_Output, _Output, "OUTPUT", false, "Sigmoid");
double [] temp = new double[_Output];
if (LearnType == "E")
{
DoEnhancedLearning = true;
}
inputCount = _Input;
OutputCount = _Output;
Net = new ANN(NetWorkInit, NewLearnRate);
double [] inputCode = new double[inputCount];
double [] result = new double[OutputCount];
EnhancedLearning = new ELCache(20, ERun, NewLearnRate, InvestigationRate);
}
public ANN CacheLearn(PastData NewEntry, ANN Net)
{
double LowError = double.MaxValue;
HighError = double.MinValue;
LowPos = -1;
HighPos = -1;
LastEntry = NewEntry;
PreviousData.Add(NewEntry);
for (int a = 0; a < PreviousData.Count(); a++)
{
double ErrorLevel = 0;
double[] NewInput = PreviousData[a].GetInput();
double SingleTarget = PreviousData[a].GetTarget(0);
double[] result = Net.feedForward(NewInput);
double[] NewTarget = (double[])result.Clone();
NewTarget[0] = SingleTarget;
if (EnhancedLearning > 0)
{
Net.Learn(NewTarget, result, NewLearnRate);
//EnhancedLearn(iAction, Net, NewInput, SingleTarget);
}
ErrorLevel = Net.LearnError(NewTarget, result);
if (ErrorLevel < LowError)
{
LowError = ErrorLevel;
LowPos = a;
}
if (ErrorLevel > HighError)
{
HighError = ErrorLevel;
HighPos = a;
}
}
if (EnhancedLearning > 0)
{
EnhancedLearning--;
}
return(Net);
}
public void EnhancedLearn(int iAction, ANN Net, double[] NewInput, double SingleTarget)
{
double SetLevel = HighError * 0.9;
double ErrorLevel;
int CheckCount = 2;
do
{
double[] result = Net.feedForward(NewInput);
double[] NewTarget = (double[])result.Clone();
NewTarget[iAction] = SingleTarget;
Net.Learn(NewTarget, result, NewLearnRate);
CheckCount--;
ErrorLevel = Net.LearnError(NewTarget, result, iAction);
} while (ErrorLevel > SetLevel & CheckCount > 0);
}
public void Learn(double _reward, double[] ID, int LastAction)
{
double[] Target = new double[1];
double TotalError = 0;
Target[0] = _reward;
// TypeName = "Standard";
TotalError = Net.LearnError(Target, Result);
if (Learning)
{
// new Learn System
if (DoEnhancedLearning)
{
Net = EnhancedLearning.CacheLearn(new PastData(Target, Result, ID, RND, TotalError), Net);
Net = EnhancedLearning.UpdateCache(Net);
}
else
{
// normal Learn
Net.Learn(Target, Result, NewLearnRate);
}
}
if (DoEnhancedLearning)
{
NewLearnRate = EnhancedLearning.GetLearnRate();
InvestigationRate = EnhancedLearning.GetInvestigationRate();
Learning = EnhancedLearning.UpdateTraining(Learning, TotalError);
}
ReportTrainingCSV.WriteErrorRate(TotalError, LastAction, NewLearnRate, InvestigationRate, _reward, EnhancedLearning.CacheCount(), EnhancedLearning.GetPreviousAverage(), EnhancedLearning.EnhancedLearning, EnhancedLearning.HighPos, EnhancedLearning.LowPos, Net.LearnCount);
}
public ANN UpdateCache(ANN Net)
{
int count = PreviousData.Count;
if (count > 2)
{
int Pos = HighPos;
double[] NewInput = PreviousData[Pos].GetInput();
double SingleTarget = PreviousData[Pos].GetTarget(0);
double SetLevel = HighError * 0.9;
double ErrorLevel;
int CheckCount = 2;
do
{
double[] result = Net.feedForward(NewInput);
double[] NewTarget = (double[])result.Clone();
NewTarget[0] = SingleTarget;
Net.Learn(NewTarget, result, NewLearnRate);
CheckCount--;
ErrorLevel = Net.LearnError(NewTarget, result);
} while (ErrorLevel > SetLevel & CheckCount > 0);
if (count > CacheSize)
{
PreviousData.RemoveAt(LowPos);
}
}
return(Net);
}
