private async Task <StringBuilder> Predict(List <List <string> > trainData,
List <List <string> > rowNames, List <List <string> > testData)
{
StringBuilder sb = null;
try
{
//ml algo rule: iterations can also set rowcount and -1 for mlinstructs removed
int iRowCount = (Shared.GetRowCount(_iterations, trainData.Count) - 1);
//columns of data used and returned in DataResults
_actualColNames = Shared.GetActualColNames(_colNames, _depColNames).ToArray();
//ml instructions associated with actual colNames
List <string> normTypes = Shared.GetNormTypes(trainData[0], _colNames, _depColNames);
//instructions in both row names and datasets
List <string> actualMLInstructs = Shared.GetAlgoInstructs(rowNames);
actualMLInstructs.AddRange(normTypes);
// error allowance
double dbPlusorMinus
= CalculatorHelpers.ConvertStringToDouble(actualMLInstructs[0]);
//converts rows to columns with normalized data
List <List <double> > trainDB = Shared.GetNormalizedDData(trainData,
this.IndicatorQT, _colNames, _depColNames, normTypes, "F2");
List <List <double> > testDB = Shared.GetNormalizedDData(testData,
this.IndicatorQT, _colNames, _depColNames, normTypes, "F2");
//make a new list with same matrix, to be replaced with results
int iColCount = testDB.Count;
if (_subalgorithm == MATHML_SUBTYPES.subalgorithm_03.ToString().ToString())
{
//subalgo02 needs qtm and percent probability of accuracy, qtm, low ci, high ci
iColCount = testDB.Count + 5;
//normtypes need full columns before insertion
normTypes = Shared.FixNormTypes(normTypes, iColCount);
}
//row count comes from original testdata to account for the instructions row
DataResults = CalculatorHelpers.GetList(testData.Count, iColCount);
DataResults[0] = normTypes;
//dep var output count
int numOutput = 1;
//less col[0]
int numInput = trainDB.Count - 1;
int numHidden = 12;
//can truncate the data to iRowCount
double[][] trainInputs = Shared.MakeInputDData(trainDB, iRowCount, this.IndicatorQT,
numInput);
//build a neural network
NeuralNetwork2 nn2 = new NeuralNetwork2(numInput, numHidden, numOutput);
int maxEpochs = iRowCount;
double learnRate = 0.001;
//train nn2
double[] wts = nn2.Train(trainInputs, maxEpochs, learnRate, sb);
//mean squared error
double trainErr = nn2.Error(trainInputs);
//final model accuracy
double trainAcc = nn2.Accuracy(trainInputs, dbPlusorMinus);
//add classified test data to DataResults
bool bHasNewClassifs = await AddNewClassifications(nn2, testDB,
trainAcc, trainErr, iRowCount, dbPlusorMinus, _ciLevel);
}
catch (Exception ex)
{
IndicatorQT.ErrorMessage = ex.Message;
}
return(sb);
}
private async Task <bool> AddNewClassifications(NeuralNetwork2 nn2, List <List <double> > testData,
double trainAccuracy, double mse, int rowCount, double hiLow, int ciPercent)
{
bool bHasNewClassifs = false;
if (ciPercent > 5)
{
hiLow = CalculatorHelpers.GetConfidenceIntervalFromMSE(
ciPercent, rowCount, mse);
}
int iRowLength = 0;
string sResult = string.Empty;
List <double> predictors = new List <double>();
double dbInput = 0;
double dbTarget = 0;
//test data stores cols in testdata[0] first index
for (int r = 0; r < DataResults.Count - 1; r++)
{
if (r == 0)
{
iRowLength = DataResults[r + 1].Count;
}
predictors = new List <double>();
for (int c = 0; c < testData.Count; c++)
{
//prepare mathresults
dbInput = testData[c][r];
DataResults[r + 1][c] = dbInput.ToString("F4");
if (c > 0)
{
predictors.Add(dbInput);
}
}
//predict
double[] forecast = nn2.ComputeOutputs(predictors.ToArray());
double dbMostLikelyQTM = forecast[0];
if (r == 0)
{
DataResults[r][iRowLength - 5] = mse.ToString("F4");
DataResults[r][iRowLength - 4] = trainAccuracy.ToString("F4");
}
dbTarget = testData[0][r];
double dbError = Math.Abs(dbTarget - dbMostLikelyQTM);
this.DataResults[r + 1][iRowLength - 5] = dbError.ToString("F4");
if (dbError < hiLow)
{
DataResults[r + 1][iRowLength - 4] = "true";
}
else
{
DataResults[r + 1][iRowLength - 4] = "false";
}
DataResults[r + 1][iRowLength - 3] = dbMostLikelyQTM.ToString("F4");
if (r == DataResults.Count - 2)
{
this.IndicatorQT.QTM = dbMostLikelyQTM;
this.IndicatorQT.QTL = dbMostLikelyQTM - hiLow;
this.IndicatorQT.QTU = dbMostLikelyQTM + hiLow;
}
//low estimation
DataResults[r + 1][iRowLength - 2] = (dbMostLikelyQTM - hiLow).ToString("F4");
//high estimation
DataResults[r + 1][iRowLength - 1] = (dbMostLikelyQTM + hiLow).ToString("F4");
}
return(bHasNewClassifs);
}
