gaussNewtonNlls.hpp

Go to the documentation of this file.

 
#pragma once
 
#include <array>
#include <functional>
#include <type_traits>
#include <utility>
 
#include "NumCpp/Core/Internal/Error.hpp"
#include "NumCpp/Core/Internal/StaticAsserts.hpp"
#include "NumCpp/Core/Internal/TypeTraits.hpp"
#include "NumCpp/Core/Shape.hpp"
#include "NumCpp/Core/Types.hpp"
#include "NumCpp/Functions/rms.hpp"
#include "NumCpp/Linalg/inv.hpp"
#include "NumCpp/NdArray.hpp"
 
namespace nc::linalg
{
    //============================================================================
    // Method Description:
    template<typename dtype,
             typename... Params,
             std::enable_if_t<std::is_arithmetic_v<dtype>, int>  = 0,
             std::enable_if_t<all_arithmetic_v<Params...>, int>  = 0,
             std::enable_if_t<all_same_v<dtype, Params...>, int> = 0>
    std::pair<NdArray<double>, double>
        gaussNewtonNlls(const uint32                                                              numIterations,
                        const NdArray<dtype>&                                                     coordinates,
                        const NdArray<dtype>&                                                     measurements,
                        const std::function<dtype(const NdArray<dtype>&, const NdArray<dtype>&)>& function,
                        const std::array<std::function<dtype(const NdArray<dtype>&, const NdArray<dtype>&)>,
                                         sizeof...(Params)>&                                      derivatives,
                        Params... initialGuess)
    {
        STATIC_ASSERT_ARITHMETIC(dtype);
 
        const auto coordinatesShape = coordinates.shape();
 
        if (coordinatesShape.rows != measurements.size())
        {
            THROW_INVALID_ARGUMENT_ERROR("coordinates number of rows, and measurements size must be the same.");
        }
 
        NdArray<double> beta = NdArray<dtype>({ initialGuess... }).template astype<double>().transpose();
        NdArray<double> residuals(coordinatesShape.rows, 1);
        NdArray<double> jacobian(coordinatesShape.rows, sizeof...(Params));
 
        const auto colSlice = coordinates.cSlice();
        for (uint32 iteration = 1; iteration <= numIterations; ++iteration)
        {
            for (uint32 measIdx = 0; measIdx < coordinatesShape.rows; ++measIdx)
            {
                const auto coordinate = coordinates(measIdx, colSlice);
 
                residuals[measIdx] =
                    static_cast<double>(measurements[measIdx]) - static_cast<double>(function(coordinate, beta));
 
                for (uint32 paramIdx = 0; paramIdx < sizeof...(Params); ++paramIdx)
                {
                    const auto& derivative      = derivatives[paramIdx];
                    jacobian(measIdx, paramIdx) = static_cast<double>(derivative(coordinate, beta));
                }
            }
 
            // perform the gauss-newton linear algebra
            const auto jacobianT             = jacobian.transpose();
            const auto jacobianPsuedoInverse = linalg::inv(jacobianT.dot(jacobian));
            const auto intermediate          = jacobianPsuedoInverse.dot(jacobianT);
            const auto deltaBeta             = intermediate.dot(residuals);
            beta += deltaBeta;
        }
 
        // calculate the final rms of the residuals
        for (uint32 measIdx = 0; measIdx < coordinatesShape.rows; ++measIdx)
        {
            const auto coordinate = coordinates(measIdx, colSlice);
 
            residuals[measIdx] =
                static_cast<double>(measurements[measIdx]) - static_cast<double>(function(coordinate, beta));
        }
 
        return std::make_pair(beta.flatten(), rms(residuals).item());
    }
} // namespace nc::linalg