UncNLPrograms

A set of high-dimensional, nonlinear, and unconstrained optimization problems implemented in native Julia to test solvers using Automatic/Algorithmic Differentiation.

The UncNLPrograms is a subset of CUTEst, the latest evolution in the Constrained and Unconstrained Testing Enviroment. Each program defines an initial dimension, iterate, an objective function and its corresponding gradient.

The UncNLPrograms interface was inspired by the NLPModels package from the JuliaSmoothOptimizers orginization.

Interface

Function	Description
Programs	Lists each selectable program and it's current dimension
SelectProgram	Returns a reference to the specified program
adjdim!	Adjust the specified programs dimensions to the given order
obj	Evaluate the programs objective function at a point x
grad	Evaluate the programs gradient function at a point x
objgrad	Economical evaluation of the programs gradient and objective at a point x
hessAD	Returns the full Hessian matrix of an objective function at a point x (by Forward AD)

Installation

Using Pkg.jl:

julia> using Pkg
julia> Pkg.add(url="https://github.com/danphenderson/UncNLPrograms.jl")

Or from the REPL:

julia> ]
pkg> add "https://github.com/danphenderson/UncNLPrograms.jl"

Simple UncNLPrograms Use Case

# Lists the name and dimension in the programs enviroment
Programs() 

# Store a reference to a program, as specified by the name given in Programs()
# programs are named as their origional .SIF file names
nlp = SelectProgram("SROSENBR")

# Returns the default itterate corresponding to the CUTEst problem in 20 dimensions
adjdim!(nlp, 20)

# Returns $f(x))$ at $x$ = nlp.$x0$, 
# where nlp.x0 is a reference to the default itterate
obj(nlp, nlp.x0) 

# Returns $\nabla x f(x))$ at $x$ = nlp.$x0$ 
grad(nlp, nlp.x0)

# Returns the tuple $(\nabla f(x),  \nabla f(x))$ at $x$ = nlp.$x0$ 
objgrad(nlp, nlp.x0)

# The jacobian of $\nabla f$ at $x$ = nlp.$x0$ using ForwardDiff.jl
hessAD(nlp, nlp.x0)

Development Items

1. Determine a suitable home for UncNLPrograms in the Julia ecosystem and/or utilize an existing structure (i.e. build upon NLPModels, JuMP).
2. Optimize programs' implementation.
3. Check off each item in the Precompile warning list.
4. Overload Base.show(::UncProgram)
5. Confirm julia defined f and $\nabla^2 f(x)$ math the formula in Buckley's report.