public MLLMFileInterpolant(string fileName, alglib.idwinterpolant myInterpolant, List <double> myPPMs)
{
MyFile = fileName;
MyInterpolant = myInterpolant;
MyPPMs = myPPMs;
double average = MyPPMs.Average();
double std = PatternTools.pTools.Stdev(MyPPMs, false);
lowerBound = average - 3 * std;
upperBound = average + 3 * std;
}
private void Learn(List <SQTScan> scans, Parameters myParams)
{
//First lets group the Scans by filename
Dictionary <string, List <SQTScan> > byFileName = (from scan in scans.AsParallel()
group scan by scan.FileName into theGroup
select new { theGroup.Key, theGroup }).ToDictionary(a => a.Key, a => a.theGroup.ToList());
interpolants = new ConcurrentBag <MLLMFileInterpolant>();
Parallel.ForEach(byFileName, kvp =>
{
Console.WriteLine("Learning lockmass function from " + kvp.Key);
List <SQTScan> learningScans = kvp.Value;
//Lets eliminate all labeled and unlabeled decoys
learningScans.RemoveAll(a => a.CountNumberForwardNames(myParams.LabeledDecoyTag) == 0);
learningScans.RemoveAll(a => a.CountNumberForwardNames(myParams.UnlabeledDecoyTag) == 0);
//Prepare a training dataset
List <double> x_MZ = new List <double>(learningScans.Count);
List <double> y_ScanNumber = new List <double>(learningScans.Count);
List <double> z_ppm = new List <double>(learningScans.Count);
//First, lets find ppm bounds for outliers
for (int i = 0; i < 4; i++)
{
List <double> ppms = (from scan in learningScans
select scan.PPM_Orbitrap).ToList();
double ppmAverage = ppms.Average();
double ppmSTD = PatternTools.pTools.Stdev(ppms, false);
learningScans.RemoveAll(a => a.PPM_Orbitrap > ppmAverage + (6 - i) * ppmSTD || a.PPM_Orbitrap < ppmAverage - (6 - i) * ppmSTD);
}
List <List <double> > trainMatrixTMP = new List <List <double> >(learningScans.Count);
List <SQTScan> testScans = new List <SQTScan>(learningScans.Count);
List <List <double> > fullMatrix = new List <List <double> >(learningScans.Count);
int theCounter = 0;
foreach (SQTScan scan in learningScans)
{
if (Math.Abs(scan.MeasuredMH - scan.TheoreticalMH) < 0.8)
{
z_ppm.Add(scan.PPM_Orbitrap);
y_ScanNumber.Add(scan.ScanNumber);
double mz = GetMZ(scan);
x_MZ.Add(mz);
theCounter++;
List <double> v = new List <double> {
{ mz }, { scan.ScanNumber }, { scan.PPM_Orbitrap }
};
fullMatrix.Add(v);
if (theCounter < 2)
{
trainMatrixTMP.Add(v);
}
else
{
theCounter = 0;
testScans.Add(scan);
}
}
}
double[,] FullMatrix = PatternTools.pTools.ConvertListofListsToDoubleArray(fullMatrix);
double[,] TrainMatrix = PatternTools.pTools.ConvertListofListsToDoubleArray(trainMatrixTMP);
//Tune Parameters
int lowestNW = -1;
int lowestNQ = -1;
double lowestAVGRMS = double.MaxValue;
alglib.idwinterpolant z = new alglib.idwinterpolant();
Console.WriteLine("Optimizing");
for (int nw = 10; nw < 55; nw += 2)
{
for (int nq = 5; nq < 20; nq += 2)
{
alglib.idwbuildnoisy(TrainMatrix, TrainMatrix.GetLength(0), 2, 1, nq, nw, out z);
//Study the error
List <double> ABSErr = new List <double>(learningScans.Count);
foreach (SQTScan scan in testScans)
{
double[] iv = new double[2];
iv[0] = GetMZ(scan);
iv[1] = scan.ScanNumber;
double result = alglib.idwcalc(z, iv);
ABSErr.Add(Math.Abs(scan.PPM_Orbitrap - result));
}
Console.Write(".");
double averageError = ABSErr.Average();
if (averageError < lowestAVGRMS)
{
lowestAVGRMS = averageError;
lowestNW = nw;
lowestNQ = nq;
Console.WriteLine("(ABSErr:" + lowestAVGRMS + " NW:" + nw + " NQ:" + nq + ") ");
}
}
}
//Do the final training
alglib.idwbuildnoisy(FullMatrix, z_ppm.Count, 2, 1, lowestNQ, lowestNW, out z);
Console.WriteLine("\nAverage RMS ppm error after optimization: " + lowestAVGRMS);
interpolants.Add(new MLLMFileInterpolant(kvp.Key, z, z_ppm));
}
);
}
