Module ojalgo
Package org.ojalgo.optimisation.convex

Class ConvexSolver.Configuration

  • Enclosing class:
    ConvexSolver
    public static final class ConvexSolver.Configuration
extends java.lang.Object

    • Field Detail

      • myCombinedScaleFactor

        private boolean myCombinedScaleFactor

      • myExtendedPrecision

        private boolean myExtendedPrecision

      • myIterativePreconditioner

        private java.util.function.Supplier<Preconditioner> myIterativePreconditioner

      • myIterativeSolver

        private java.util.function.Supplier<IterativeSolverTask> myIterativeSolver

      • mySmallDiagonal

        private double mySmallDiagonal

    • Constructor Detail

      • Configuration

        public Configuration()

    • Method Detail

      • combinedScaleFactor

        public ConvexSolver.Configuration combinedScaleFactor​(boolean combinedScaleFactor)
        Only relevant with extended precision. With the extended precision solver the primal and dual variables are scaled (shift and zoom) to iteratively generate subproblems. In theory there are different scaling factors for the primal and dual variables, but forcing them to be the same enables simplifications resulting in significant performance gains.

        The default is to use the same scaling factor for both primal and dual variables. By setting this to false, you switch to a slower more complex, but theoretically more accurate and flexible algorithm.

        See Also:
        extendedPrecision(boolean)

      • extendedPrecision

        public ConvexSolver.Configuration extendedPrecision​(boolean extendedPrecision)
        With extended precision the usual solver is wrapped by a master algorithm, implemented in Quadruple precision, that iteratively refines (zoom and shift) the problem to be solved by the delegate solver. This enables to handle constraints with very high accuracy.

        The iterative refinement solver cannot handle general inequality constraints, only simple variable bounds (modelled as inequality constraints).

        This is an experimental feature!

        Setting this to true you should most likely also set the Optimisation.Options.solution to something matching that allows for higher precision.

      • isCombinedScaleFactor

        public boolean isCombinedScaleFactor()

      • isExtendedPrecision

        public boolean isExtendedPrecision()

      • newIterativeSolver

        public IterativeSolverTask newIterativeSolver​(int maxIterations)
        Returns a new iterative solver instance configured with the current accuracy, maximum iterations, and preconditioner settings.

      • smallDiagonal

        public double smallDiagonal()

      • smallDiagonal

        public ConvexSolver.Configuration smallDiagonal​(double factor)
        The [Q] matrix (of quadratic terms) is "inverted" using a matrix decomposition returned by newSolverSPD(Structure2D). If, after decomposition, MatrixDecomposition.Solver.isSolvable() returns false a small constant is added to the diagonal.

        The small constant will be the largest absolute element times this small diagonal factor.

        This is only meant to handle minor, unexpected, deficiencies.

      • solverGeneral

        public ConvexSolver.Configuration solverGeneral​(java.util.function.Function<Structure2D,​MatrixDecomposition.Solver<java.lang.Double>> factory)
        This matrix decomposition should be able to "invert" the full KKT systsem body matrix (which is symmetric) and/or its Schur complement with regards to the [Q] matrix (of quadratic terms).

      • solverSPD

        public ConvexSolver.Configuration solverSPD​(java.util.function.Function<Structure2D,​MatrixDecomposition.Solver<java.lang.Double>> factory)
        The [Q] matrix (of quadratic terms) is supposed to be symmetric positive definite (or at least semidefinite), but in reality there are usually many deficiencies. This matrix decomposition should handle "inverting" the [Q] matrix.