/// <summary>
///
/// </summary>
/// <param name="design"></param>
/// <param name="response"></param>
/// <param name="weights"></param>
/// <param name="distribution"></param>
/// <param name="maxIterations"></param>
/// <param name="absoluteTolerance"></param>
/// <param name="relativeTolerance"></param>
/// <returns></returns>
public static (double[] Coefficients, double[] WeightedResponse) RegressIrls([NotNull][ItemNotNull] this double[][] design, [NotNull] double[] response, [NotNull] double[] weights, [NotNull] IDistribution distribution, int maxIterations = 100, double absoluteTolerance = 1e-8, double relativeTolerance = default)
{
if (response is null)
{
throw new ArgumentNullException(nameof(response));
}
if (design is null)
{
throw new ArgumentNullException(nameof(design));
}
if (weights is null)
{
throw new ArgumentNullException(nameof(weights));
}
double[] oldResiduals = new double[design.Length];
double[] wlsCoefficients = new double[design[0].Length];
double[] wlsResponse = new double[weights.Length];
double[] wlsWeights = new double[weights.Length];
Array.Copy(weights, wlsWeights, wlsWeights.Length);
double[] meanResponse = distribution.InitialMean(response);
double[] linearResponse = distribution.Predict(meanResponse);
for (int i = 0; i < maxIterations; i++)
{
wlsWeights =
distribution.Weight(meanResponse)
.Multiply(weights);
wlsResponse =
distribution.LinkFunction
.FirstDerivative(meanResponse)
.Multiply(response.Subtract(meanResponse))
.Add(linearResponse);
wlsCoefficients =
design.RegressWls(wlsResponse, wlsWeights);
linearResponse =
design.MatrixProduct(wlsCoefficients);
meanResponse =
distribution.Fit(linearResponse);
double[] residuals =
response.Subtract(linearResponse);
if (HasConverged(residuals, oldResiduals, absoluteTolerance, relativeTolerance))
{
break;
}
Array.Copy(residuals, oldResiduals, oldResiduals.Length);
}
return(wlsCoefficients, wlsResponse);
}