joptimise

joptimise.jl (Fr: j'optimise) is a Julia wrapper to Ipopt and SNOPT for solving nonlinear programming problems with gradient methods. Parts borrowed from Snow.jl and SNOPT7.jl.

Dependencies

Ipopt, FiniteDiff, ForwardDiff, ReverseDiff, SparseArrays

Environment setup

For using SNOPT, users must also have an active license.

Windows (Windows Subsystem for Linux)

Usage on native Windows is discouraged; setting up a Julia environment on WSL seems to work well with SNOPT. Set as environment variables in ~/.bashrc (working on WSL):

export SNOPT_LICENSE="$HOME/path-to/snopt7.lic"
export SNOPT_SO="$HOME/path-to/libsnopt7/libsnopt7.so"

Then, internally, the SNOPT library is called via

const snoptlib = ENV["SNOPT_SO"]

Installation

(@v1.6) pkg> dev https://github.com/Yuricst/joptimise.git

Usage

Import module, define objective function which mutates the constraint value g and returns the objective value:

using joptimise

function rosenbrock!(g, x)
    # compute objective
    f = (1 - x[1])^2 + 100*(x[2] - x[1]^2)^2
    # constraint
    g[1] = x[1]^2 + x[2]^2 - 1.0
    return f
end

# initial guess
x0 = [4.0; 4.0]
# bounds on variables
lx = [-5.0; -5.0]
ux = [5.0; 5.0]
# bounds on constraints
lg = [0.0]
ug = [0.0]
# number of constraints
ng = 1

then call SNOPT

sn_options = Dict(
    "Major feasibility tolerance" => 1.e-6,
    "Major optimality tolerance"  => 1.e-6,
    "Minor feasibility tolerance" => 1.e-6,
    "Major iterations limit" => 1000,
    "Major print level" => 1,
)

xopt, fopt, info = minimize(rosenbrock!, x0, ng; lx=lx, ux=ux, lg=lg, ug=ug, solver="snopt", options=sn_options)

or IPOPT

ip_options = Dict(
    "max_iter" => 2500,
    "tol" => 1e-6,
    "print_level" => 5,
)

xopt, fopt, info = minimize(rosenbrock!, x0, ng; lx=lx, ux=ux, lg=lg, ug=ug, solver="ipopt", options=ip_options)

Specifying derivative method

By default, minimize() will compute derivatives of the objective and constraints using forward finite-difference from ForwardDiff. This may be altered to central-difference, forward-mode or reverse-mode AD from ForwardDiff or ReverseDiff. This is passed as the kwargs derivatives; the possible options are ForwardFD(), CentralFD(), ForwardAD(), or ReverseAD(). For example, if using forward-mode AD,

xopt, fopt, info = minimize(rosenbrock!, x0, ng; lx=lx, ux=ux, lg=lg, ug=ug, solver="snopt", options=sn_options, derivatives=ForwardAD())

Todo

• user-specified derivative
• optional output file using SNOPT