NewtonConjugateGradientOptimizer Class

public sealed class NewtonConjugateGradientOptimizer : MultidimensionalOptimizer

Public NotInheritable Class NewtonConjugateGradientOptimizer
	Inherits MultidimensionalOptimizer

public ref class NewtonConjugateGradientOptimizer sealed : public MultidimensionalOptimizer

[<SealedAttribute>]
type NewtonConjugateGradientOptimizer = 
    class
        inherit MultidimensionalOptimizer
    end

Inheritance

Object  →  ManagedIterativeAlgorithm<Vector<Double>, Double, OptimizationSolutionReport>  →  MultidimensionalOptimizer<Vector<Double>, OptimizationSolutionReport>  →  MultidimensionalOptimizer  →  NewtonConjugateGradientOptimizer

Supply first derivatives and a Hessian function returning a matrix or a multiplication-only linear operator. Optional bounds are enforced throughout the solve. The optimizer uses Hessian-vector products without reconstructing a dense Hessian.

A supplied curvature approximation, such as Gauss-Newton for a sum of squared residuals, may replace exact Hessian products or entries. Use the original objective sign, and keep the objective and gradient callbacks exact. Final first-order checks apply to those callbacks.

Successful termination requires bound feasibility, projected stationarity and complementarity. Custom convergence tests add stopping conditions; a small value or step change alone cannot certify a solution. A stationary point of a nonconvex function is not a global optimum certificate.

Reference