MINLPLib

STATUS:

MINLPLib.jl is an extensive collection of Mixed-Integer and Continuous Nonlinear Programming Instances in the native JuMP.jl format. Numerous instances in this library are based on the standard MINLPLib.org instances in addition to hundreds of global optimization instances which appear in the litearature. This library can be useful for:

• Benchmarking the Julia-based MINLP solvers, like Alpine.jl and Juniper.jl, by performing rigorous numerical experiments

• Viewing meta information of each instance to assist your analyses and experiments

For more details of the libraries included in this package, please see documentation.

To install, at your Julia command prompt,

Pkg.clone("https://github.com/lanl-ansi/MINLPLib.jl.git")

Fetch an instance by its libname/name just like the following,

using MINLPLib
m = fetch_model("minlp2/blend029")

or

using MINLPLib
m = fetch_model("minlp2", "blend029")

For special built-in model that requires additional inputs,

using MINLPLib
m = fetch_model("special", "multiKND", options=Dict(:K=>3, :N=>3, :D=>1))

The above script initializes an optimization JuMP model multiKND. The formulation is controlled with parameter :K, :N, and :D, which yields the following problem:

Max x[1] * x[2] * x[3] + x[3] * x[4] * x[5] + x[5] * x[6] * x[7]
Subject to
 x[1] + x[2] + x[3] <= 3
 x[3] + x[4] + x[5] <= 3
 x[5] + x[6] + x[7] <= 3
 0.1 <= x[i] <= 10 for all i in {1,2,..,6,7}

To know a instance better, you can do the following to get a dictionary of meta info. Note that as we continue to develop and test. More verified attributes will be added to the meta.

julia> meta = MINLPLib.fetch_meta("minlp2", "blend029")
Dict{String,Any} with 18 entries:
  "NINTVARS"   => 0
  "NLINCONS"   => 202
  "OBJBOUND"   => nothing
  "NAME"       => "blend029"
  "OBJVAL"     => nothing
  "LIBRARY"    => "minlp2"
  "NCONS"      => 214
  "NVARS"      => 103
  "OBJTYPE"    => "linear"
  "NLOPERATOR" => Any["mul"]
  "LOAD"       => 1.7414
  "OBJSENSE"   => "Max"
  "NSDPCONS"   => 0
  "NNLCONS"    => 12
  "NBINVARS"   => 36
  "NQUADCONS"  => 0
  "NSOSCONS"   => 0

Currently, MINLPLib.jl contains over 6000 instances. Experimenting with all of them can be computationally heavy and risky. For your research projects, the scope may be limited to a specific type of problems. Hence, it is necessary to construct you very own library for research experiments. Below, we write a small .jl script that will generate a user instance library (without actually copying the problem) called USERLib that collect all instances with exp function. The directory will be created with name instance/USERLib. Please see our PODLib designed for testing Alpine.jl solver. This customized library contains instances reference as well as full instances. Note that each instance can be drastically different with different formulations and implementations (peak into the library called special and you will see some examples). It is up to you to select and construct your own instance library for more exciting and convenient experiments.

using MINLPLib

# Collect target problem from these libraries
for lib in ["bcp", "global", "ibm", "inf", "minlp", "minlp2"]
    NAMES = MINLPLib.fetch_names(lib)   # Fetch a list of instance names
    for i in NAMES
        Meta = MINLPLib.fetch_meta(lib, i)  # Fetch instance meta
        isempty(Meta) && continue
        !haskey(Meta, "NLOPERATOR") && continue
        if occursin("exp", in Meta["NLOPERATOR"][1])
            MINLPLib.add_to_lib("USERLib", lib, i)  # Collect instance
        else
            continue
        end
    end
end
println(MINLPLib.fetch_names("USERLib"))