Jvanrhijn/lbfgs-playground

Testing repository for LBFGS libraries

LBFGS playground

This repository contains a Fortran 90 implementation of the "online" L-BFGS (o-LBFGS) algorithm outlined in Schraudolf (2007).

Usage

Assume in the following that n denotes the number of variables x to optimize, m denotes the number of curvature pairs to store, f denotes the objective function, and g its gradient. Then, a minimal usage example is as follows.

program main
    use olbfgs, only: initialize_olbfgs, olbfgs_iteration, update_hessian
    implicit none

    ! algorithm parameters
    integer :: n = 100
    integer :: m = 5
    double precision :: step_size = 0.5

    double precision :: value

    double precision, dimension(:) :: x
    double precision, dimension(:) :: grad

    ! initialize parameter vector x
    x = ...

    value = f(x)
    grad = g(x)

    ! First intiialize the algorithm (i.e. allocate storage, set previous gradient/parameter values)
    call initialize_olbfgs(value, grad, n, m)

    do i = 1, 100, 1
        ! perform optimization iteration
        call olbfgs_iteration(x, g, step_size, i)
        ! compute function value, gradient
        value = f(x)
        grad = g(x)
        ! update the Hessian approximation
        call update_hessian(x, g)
        ! output function value
        print *, f(x)
    end do

end program

The call to initialize_olbfgs before calling either update_hessian or olbfgs_iteration is required for allocating memory used internally in the module.