Different results depending the MIP solver used

[xgandibleux]

Here a tiny GAP with 3 objectives.
It is solved with the 3 algorithms (KirlikSayin, TambyVanderpooten, DominguezRios) x the 4 MIP solvers (HiGHS, GLPK, CPLEX, Gurobi).
In checking only result_count(GAP) I collected 42 points except:

• 0 point for KirlikSayin with CPLEX and Gurobi
• 43 points for TambyVanderpooten with HiGHS

All packages used are the latest versions.

3-GAP model

m=3; n=10
p1 = [5 1 10 8 3 5 3 3 7 2; 10 6 1 6 8 3 2 10 6 1; 2 3 1 6 9 7 1 5 4 8]
p2 = [4 6 4 3 1 6 8 2 9 7; 8 8 8 2 4 8 8 1 10 1; 8 7 8 5 9 2 2 7 10 10]
p3 = [4 3 6 4 7 5 9 5 8 4; 8 6 2 2 6 8 5 2 2 3; 2 8 10 3 5 7 5 9 5 5]
w = [5 9 3 5 10 5 7 10 7 8; 4 8 8 6 10 8 10 7 5 1; 10 7 5 8 8 2 8 1 10 3]
b = [34, 33, 31]

using JuMP
using GLPK
import MultiObjectiveAlgorithms as MOA

GAP = Model()
@variable(GAP, x[1:m, 1:n], Bin)
@expression(GAP, objective1, sum(p1[i,j]*x[i,j] for i = 1:m, j = 1:n) )
@expression(GAP, objective2, sum(p2[i,j]*x[i,j] for i = 1:m, j = 1:n) )
@expression(GAP, objective3, sum(p3[i,j]*x[i,j] for i = 1:m, j = 1:n) )
@objective(GAP, Max, [objective1,objective2,objective3])
@constraint(GAP, [i=1:m], sum(w[i,j]*x[i,j] for j = 1:n) <= b[i])
@constraint(GAP, [j=1:n], sum(x[i,j] for i = 1:m) == 1)

set_optimizer(GAP, () -> MOA.Optimizer(GLPK.Optimizer))
#set_attribute(GAP, MOA.Algorithm(), MOA.KirlikSayin())
#set_attribute(GAP, MOA.Algorithm(), MOA.DominguezRios())
set_attribute(GAP, MOA.Algorithm(), MOA.TambyVanderpooten())
set_silent(GAP)
optimize!(GAP)

YN = []
for i in 1:result_count(GAP)
    y = round.(Int, objective_value(GAP; result = i))
    push!(YN,y)
    println(i, ": z = ", y)
end

results

using CPLEX. or using Gurobi
set_attribute(GAP, MOA.Algorithm(), MOA.TambyVanderpooten())
result_count(GAP) = 42
set_attribute(GAP, MOA.Algorithm(), MOA.DominguezRios())
result_count(GAP) = 42
set_attribute(GAP, MOA.Algorithm(), MOA.KirlikSayin())
result_count(GAP) = 0

using HiGHS
set_attribute(GAP, MOA.Algorithm(), MOA.TambyVanderpooten())
result_count(GAP) = 43
set_attribute(GAP, MOA.Algorithm(), MOA.DominguezRios())
result_count(GAP) = 42
set_attribute(GAP, MOA.Algorithm(), MOA.KirlikSayin())
result_count(GAP) = 42

using GLPK.
set_attribute(GAP, MOA.Algorithm(), MOA.TambyVanderpooten())
result_count(GAP) = 42
set_attribute(GAP, MOA.Algorithm(), MOA.DominguezRios())
result_count(GAP) = 42
set_attribute(GAP, MOA.Algorithm(), MOA.KirlikSayin())
result_count(GAP) = 42

config

[a076750e] CPLEX v1.1.1
[60bf3e95] GLPK v1.2.1
[2e9cd046] Gurobi v1.7.4
[87dc4568] HiGHS v1.18.1
[4076af6c] JuMP v1.26.0
[0327d340] MultiObjectiveAlgorithms v1.4.3

Gurobi Optimizer version 12.0.0 build v12.0.0rc1 (mac64[arm] - Darwin 23.6.0 23G93)
CPLEX Version identifier: 22.1.1.0

[kofgokhan]

For the second stage, if I change:

MOI.delete.(model, ε_constraints)
MOI.delete(model, zₖ_constraint)
if !_is_scalar_status_optimal(model)
    _remove_rectangle(L, _Rectangle(_project(yI, k), uᵢ))
    continue
end

to

if !_is_scalar_status_optimal(model)
    _remove_rectangle(L, _Rectangle(_project(yI, k), uᵢ))
    MOI.delete.(model, ε_constraints)
    MOI.delete(model, zₖ_constraint)
    continue
end
...
MOI.delete.(model, ε_constraints)
MOI.delete(model, zₖ_constraint)

and add delete to the end of the iteration, MOA.KirlikSayin will return 42 points using CPLEX.

I think _is_scalar_status_optimal(model) returns different results after delete when we use HiGHS vs CPLEX.

[odow]

Ah, yes, if you modify the model the solver is allowed to throw away stored solutions. I'll make a fix