ohinder / onephase.jl Goto Github PK

This package is the implementation of a one-phase interior point method that finds KKT points of nonconvex optimization problems.

License: Other

Julia 95.81% Shell 4.19%

mathematical-programming optimization optimization-algorithms

onephase.jl's People

Contributors

Stargazers

Watchers

Forkers

lrsantos11 amuwa carolinesnakama fadihamad94

onephase.jl's Issues

warm start dual variables JuMP

Create a test set of for Pkg.test()

get extra linear solver working

get MUMPS/MA57 going

handle maximize and minimization problems

Upgrade code to latest julia

update complementarity penalty

$(s_i y_i - \mu)^3 / (\mu^2 * \beta_2^2)$

print summary info on quality of point obtained.

Direct interface to JumP

Make it so we can interface with JuMP directly.

Fix HSL linear solver

Usage

If I run

using OnePhase, JuMP

m = Model(OnePhase())
@variable(m, x, start=-3)
@objective(m, Min, x)
@NLconstraint(m, x^2 >= 1.0)
@NLconstraint(m, x >= -1.0)

status = solve(m)

I get

ERROR: MethodError: objects of type Module are not callable

as an error. I might have a different version of JuMP. However,
replacing m = Model(OnePhase())
with m = Model(solver=OnePhaseSolver()) works on my laptop.

Moreover, if I run

using OnePhase, CUTEst
nlp = CUTEstModel("CHAIN")
iter = one_phase_solve(nlp);

I get the error

projecting onto bounds ...
ERROR: MethodError: no method matching cons(::CUTEst.CUTEstModel, ::Array{Float64,1})
Closest candidates are:
  cons(::OnePhase.MathProgNLPModel, ::Array{Float64,N} where N) at /home/jomi/.julia/v0.6/OnePhase/src/JuMPinterface.jl:178
Stacktrace:
 [1] eval_a at /home/jomi/.julia/v0.6/OnePhase/src/utils/Class_cutest.jl:138 [inlined]
 [2] eval_init(::OnePhase.Class_CUTEst, ::OnePhase.Class_parameters, ::advanced_timer.class_advanced_timer, ::Array{Float64,1}) at /home/jomi/.julia/v0.6/OnePhase/src/init/other.jl:4
 [3] gertz_init(::OnePhase.Class_CUTEst, ::OnePhase.Class_parameters, ::advanced_timer.class_advanced_timer) at /home/jomi/.julia/v0.6/OnePhase/src/init/gertz_init.jl:12
 [4] one_phase_solve(::CUTEst.CUTEstModel, ::OnePhase.Class_parameters) at /home/jomi/.julia/v0.6/OnePhase/src/IPM/one_phase.jl:25
 [5] one_phase_solve(::CUTEst.CUTEstModel) at /home/jomi/.julia/v0.6/OnePhase/src/IPM/one_phase.jl:14

Solver speed

Go through code and fix speed issues. Note that display progress is taking longer than it should.

See:
https://discourse.julialang.org/t/speeding-up-julia/4794
https://julialang.org/blog/2013/09/fast-numeric
https://docs.julialang.org/en/release-0.4/manual/performance-tips/

Update JuMP interface

change inequalities in base code

Currently we use
a(x) >= 0, but paper uses a(x) <= 0. Should make it consistent, i.e., change code.

address andreas's feedback

due better job of explaining merit function acceptance condition
figure out why IPOPT fails more often than in IPOPT paper (I think it is MUMPs, https://github.com/JuliaSmoothOptimizers/CUTEst.jl may use different default problem sizes, maybe other factors).
make replication easier/more transparent, i.e., add benchmarking code to repo
IPOPT fails on problems with more equalities than inequalities