ConicIP.jl

Michael P. Friedlander and Gabriel Goh
August 2016

ConicIP.jl: A Pure Julia Conic QP Solver

ConicIP (Conic Interior Point) is an interior-point solver inspired by cvxopt for optimizing quadratic objectives with linear equality constraints, and polyhedral, second-order cone constraints. (Semidefinite cone constraints are available, but only supported as an experimental feature.) Because ConicIP is written in Julia, it allows abstract input and allows callbacks for its most computationaly intensive internal routines.

Basic Usage

ConicIP has the interface

sol = conicIP( Q , c , A , b , 𝐾 , G , d )

For the problem

minimize    ½yᵀQy - cᵀy
s.t         Ay ≧𝐾 b,  𝐾 = 𝐾₁  × ⋯ × 𝐾ⱼ
            Gy  = d

𝐾 is a list of tuples of the form (Cone Type ∈ {"R", "Q"}, Cone Dimension) specifying the cone 𝐾ᵢ. For example, the cone 𝐾 = 𝑅² × 𝑄³ × 𝑅² has the following specification:

𝐾 = [ ("R",2) , ("Q",3),  ("R",2) ]

ConicIP returns sol, a structure containing error information (sol.status), the primal variables (sol.y), dual variables (sol.v, sol.w), and convergence information.

To solve the problem

minimize    ½yᵀQy - cᵀy
such that   y ≧ 0

for example, use ConicIP as follows

using ConicIP

n = 1000

Q = sparse(randn(n,n))
Q = Q'*Q
c = ones(n,1)
A = speye(n)
b = zeros(n,1)
𝐾 = [("R",n)]

sol = conicIP(Q, c, A, b, 𝐾, verbose=true);

For a more detailed example involving callback functions, refer to this notebook.

Usage with modelling libraries

ConicIP is integrated with MathProgBase and can be used as a solver in JuMP and Convex.

JuMP.jl

using JuMP
using ConicIP

m = Model(solver = ConicIPSolver())
@variable(m, x[1:10] >= 0)
@constraint(m, sum(x) == 1.0)
@objective(m, Min, sum(x))
status = solve(m)
getvalue(x) # should be [0.1 0.1 ⋯ 0.1]

Note: JuMP does not currently allow mixing quadratic objectives with conic constraints.

Convex.jl

using Convex
using ConicIP

set_default_solver(ConicIPSolver())
x = Variable(10)
p = minimize( sum(x), [x >= 0, sum(x) == 1])
solve!(p)
x # should be [0.1 0.1 ⋯ 0.1]