Support for compute_conflict!(model) #134
Comments
|
If you're working with linear programs (which is the only class supported by Tulip), then a conflict can be identified from a dual proof of infeasibility (which you can retrieve by getting the dual solution of an infeasible problem). I don't know how the internals of
Because Tulip is an interior-point solver, you get the following limitations:
|
|
I am trying to get access to the dual with the JuMP interface, but having problems still. My problem has linear objective and the following constraints: |
|
The interval constraints are an issue in this case? |
|
I checked |
|
Can you give me a reproducible example? |
|
Here is a MWE where the infeasibility comes from two sources:
Can we somehow highlight these two facts from the dual problem? function mwe()
model = Model(Tulip.Optimizer)
### PRODUCERS
pset = zip(["P1","P2"], ["A","B"])
@variable(model, Vp[pset])
@constraint(model, [0.,0.] .≤ Vp .≤ [100.,100.])
### STATIONS
sset = [("S1","A","IN"), ("S1","A","OUT"), ("S1","A-S","OUT"),
("S2","B","IN"), ("S2","B","OUT")]
@variable(model, Vs[sset])
for s in ["S1","S2"]
ssetin = filter(x -> first(x) == s && last(x) == "IN", sset)
ssetout = filter(x -> first(x) == s && last(x) == "OUT", sset)
@constraint(model, 0.0 ≤ sum(Vs[ssetin]) ≤ Inf)
@constraint(model, 0.0 ≤ sum(Vs[ssetout]) ≤ Inf)
end
### BENEFICIATION
bset = zip(["S1","S1","S2"], ["A","A","B"], ["A","A-S","B"])
@variable(model, Vb[bset])
@constraint(model, Vb .≥ 0)
### CONSUMERS
cset = zip(["C1","C2"], ["A","B"])
@variable(model, Vc[cset])
@constraint(model, [200.,100.] .≤ Vc .≤ [Inf,Inf])
### ROUTES
rset = [("R1","A","P1","S1"),("R2","A","S1","C1"),
("R3","B","P2","S2"),("R4","B","S2","C2")]
@variable(model, Vr[rset])
@constraint(model, 0. ≤ sum(Vr[[("R1","A","P1","S1")]]) ≤ Inf)
@constraint(model, 0. ≤ sum(Vr[[("R2","A","S1","C1")]]) ≤ Inf)
@constraint(model, 0. ≤ sum(Vr[[("R3","B","P2","S2")]]) ≤ 50.)
@constraint(model, 0. ≤ sum(Vr[[("R4","B","S2","C2")]]) ≤ Inf)
### MASS BALANCE
for (orig, prod) in pset
rcon = filter(x -> x[3] == orig && x[2] == prod, rset)
@constraint(model, Vp[(orig, prod)] == sum(Vr[rcon]))
end
for (station, prod, side) in sset
if side == "IN"
rcon = filter(x -> x[2] == prod && x[4] == station, rset)
@constraint(model, Vs[(station, prod, side)] == sum(Vr[rcon]))
elseif side == "OUT"
rcon = filter(x -> x[3] == station && x[2] == prod, rset)
@constraint(model, Vs[(station, prod, side)] == sum(Vr[rcon]))
end
end
for (dest, prod) in cset
rcon = filter(x -> x[2] == prod && x[4] == dest, rset)
@constraint(model, Vc[(dest, prod)] == sum(Vr[rcon]))
end
@constraint(model,
Vs[("S1","A","IN")] == sum(Vb[[("S1","A","A"),("S1","A","A-S")]]))
@constraint(model,
Vs[("S1","A","IN")] == sum(Vs[[("S1","A","OUT"),("S1","A-S","OUT")]]))
@constraint(model,
Vs[("S2","B","IN")] == sum(Vb[[("S2","B","B")]]))
@constraint(model,
Vs[("S2","B","IN")] == sum(Vs[[("S2","B","OUT")]]))
### OBJECTIVE
rcosts = sum(Vr[r] for r in rset)
bcosts = sum(Vb[b] for b in bset)
cvalue = sum(Vc[c] for c in cset)
@objective(model, Min, rcosts + bcosts - cvalue)
model
end |
|
Here is how I (manually) print a conflict from an infeasibility proof using JuMP
using Tulip
using Printf
# Add MWE code here
function mwe()
...
end
function conflict()
model = mwe()
optimize!(model)
st = termination_status(model)
dst = dual_status(model)
if st == INFEASIBLE && dst == INFEASIBILITY_CERTIFICATE
ctypes = list_of_constraint_types(model)
for (F, S) in ctypes
cons = all_constraints(model, F, S)
for i in eachindex(cons)
isassigned(cons, i) || continue
con = cons[i]
λ = dual(con) # get dual of constraint
# if dual is non-zero, the constraint is part of the conflict
abs(λ) > 1e-4 && println("[λ = $λ] ", con)
end
end
else
println("Termination status is $(st) (dual status $dst); no conflict to print")
end
return nothing
endThis code outputs the following: [λ = 1.970914335410348] Vp[("P1", "A")] - Vr[("R1", "A", "P1", "S1")] = 0.0
[λ = -1.8082052821554653] Vs[("S1", "A", "IN")] - Vr[("R1", "A", "P1", "S1")] = 0.0
[λ = 1.310308006900506] Vs[("S1", "A", "OUT")] - Vr[("R2", "A", "S1", "C1")] = 0.0
[λ = 1.3103080069936996] Vs[("S1", "A-S", "OUT")] = 0.0
[λ = -1.3707587381583282] Vs[("S2", "B", "IN")] - Vr[("R3", "B", "P2", "S2")] = 0.0
[λ = 0.8477120338037647] Vs[("S2", "B", "OUT")] - Vr[("R4", "B", "S2", "C2")] = 0.0
[λ = -1.227954732915119] Vc[("C1", "A")] - Vr[("R2", "A", "S1", "C1")] = 0.0
[λ = -0.6537346401063114] Vc[("C2", "B")] - Vr[("R4", "B", "S2", "C2")] = 0.0
[λ = 0.27551734649770443] Vs[("S1", "A", "IN")] - Vb[("S1", "A", "A")] - Vb[("S1", "A", "A-S")] = 0.0
[λ = 1.4055739642095522] Vs[("S1", "A", "IN")] - Vs[("S1", "A", "OUT")] - Vs[("S1", "A-S", "OUT")] = 0.0
[λ = -1.0] Vs[("S2", "B", "IN")] - Vb[("S2", "B", "B")] = 0.0
[λ = 1.076829350357877] Vs[("S2", "B", "IN")] - Vs[("S2", "B", "OUT")] = 0.0
[λ = 0.2755173464754485] Vb[("S1", "A", "A")] ≥ 0.0
[λ = 0.2755173464754485] Vb[("S1", "A", "A-S")] ≥ 0.0
[λ = -1.9709143354282603] Vp[("P1", "A")] ∈ [0.0, 100.0]
[λ = 0.127113971372492] Vs[("S1", "A", "IN")] ∈ [0.0, Inf]
[λ = 0.09526595721585247] Vs[("S1", "A", "OUT")] + Vs[("S1", "A-S", "OUT")] ∈ [0.0, Inf]
[λ = 0.29392938776936095] Vs[("S2", "B", "IN")] ∈ [0.0, Inf]
[λ = 0.22911731651387665] Vs[("S2", "B", "OUT")] ∈ [0.0, Inf]
[λ = 1.2279547328428626] Vc[("C1", "A")] ∈ [200.0, Inf]
[λ = 0.6537346400692975] Vc[("C2", "B")] ∈ [100.0, Inf]
[λ = 0.16270905321115217] Vr[("R1", "A", "P1", "S1")] ∈ [0.0, Inf]
[λ = 0.08235327389692947] Vr[("R2", "A", "S1", "C1")] ∈ [0.0, Inf]
[λ = -1.370758738172468] Vr[("R3", "B", "P2", "S2")] ∈ [0.0, 50.0]
[λ = 0.1939773936510756] Vr[("R4", "B", "S2", "C2")] ∈ [0.0, Inf]All these constraints have non-zero dual and therefore form part of the conflict. I just printed the dual proof of infeasibility, so it's not guaranteed to be irreducible (i.e., it's possible that removing one of these constraints doesn't render the model feasible). |
|
Is this result similar to what we get with primal_feasibility_report in
JuMP.jl?
Em qui., 9 de mar. de 2023 17:00, mtanneau ***@***.***>
escreveu:
… Here is how I (manually) print a conflict from an infeasibility proof
using JuMPusing Tulipusing Printf
# Add MWE code herefunction mwe()
...end
function conflict()
model = mwe()
optimize!(model)
st = termination_status(model)
dst = dual_status(model)
if st == INFEASIBLE && dst == INFEASIBILITY_CERTIFICATE
ctypes = list_of_constraint_types(model)
for (F, S) in ctypes
cons = all_constraints(model, F, S)
for i in eachindex(cons)
isassigned(cons, i) || continue
con = cons[i]
λ = dual(con) # get dual of constraint
# if dual is non-zero, the constraint is part of the conflict
abs(λ) > 1e-4 && println("[λ = $λ] ", con)
end
end
else
println("Termination status is $(st) (dual status $dst); no conflict to print")
end
return nothingend
This code outputs the following:
[λ = 1.970914335410348] Vp[("P1", "A")] - Vr[("R1", "A", "P1", "S1")] = 0.0
[λ = -1.8082052821554653] Vs[("S1", "A", "IN")] - Vr[("R1", "A", "P1", "S1")] = 0.0
[λ = 1.310308006900506] Vs[("S1", "A", "OUT")] - Vr[("R2", "A", "S1", "C1")] = 0.0
[λ = 1.3103080069936996] Vs[("S1", "A-S", "OUT")] = 0.0
[λ = -1.3707587381583282] Vs[("S2", "B", "IN")] - Vr[("R3", "B", "P2", "S2")] = 0.0
[λ = 0.8477120338037647] Vs[("S2", "B", "OUT")] - Vr[("R4", "B", "S2", "C2")] = 0.0
[λ = -1.227954732915119] Vc[("C1", "A")] - Vr[("R2", "A", "S1", "C1")] = 0.0
[λ = -0.6537346401063114] Vc[("C2", "B")] - Vr[("R4", "B", "S2", "C2")] = 0.0
[λ = 0.27551734649770443] Vs[("S1", "A", "IN")] - Vb[("S1", "A", "A")] - Vb[("S1", "A", "A-S")] = 0.0
[λ = 1.4055739642095522] Vs[("S1", "A", "IN")] - Vs[("S1", "A", "OUT")] - Vs[("S1", "A-S", "OUT")] = 0.0
[λ = -1.0] Vs[("S2", "B", "IN")] - Vb[("S2", "B", "B")] = 0.0
[λ = 1.076829350357877] Vs[("S2", "B", "IN")] - Vs[("S2", "B", "OUT")] = 0.0
[λ = 0.2755173464754485] Vb[("S1", "A", "A")] ≥ 0.0
[λ = 0.2755173464754485] Vb[("S1", "A", "A-S")] ≥ 0.0
[λ = -1.9709143354282603] Vp[("P1", "A")] ∈ [0.0, 100.0]
[λ = 0.127113971372492] Vs[("S1", "A", "IN")] ∈ [0.0, Inf]
[λ = 0.09526595721585247] Vs[("S1", "A", "OUT")] + Vs[("S1", "A-S", "OUT")] ∈ [0.0, Inf]
[λ = 0.29392938776936095] Vs[("S2", "B", "IN")] ∈ [0.0, Inf]
[λ = 0.22911731651387665] Vs[("S2", "B", "OUT")] ∈ [0.0, Inf]
[λ = 1.2279547328428626] Vc[("C1", "A")] ∈ [200.0, Inf]
[λ = 0.6537346400692975] Vc[("C2", "B")] ∈ [100.0, Inf]
[λ = 0.16270905321115217] Vr[("R1", "A", "P1", "S1")] ∈ [0.0, Inf]
[λ = 0.08235327389692947] Vr[("R2", "A", "S1", "C1")] ∈ [0.0, Inf]
[λ = -1.370758738172468] Vr[("R3", "B", "P2", "S2")] ∈ [0.0, 50.0]
[λ = 0.1939773936510756] Vr[("R4", "B", "S2", "C2")] ∈ [0.0, Inf]
All these constraints have non-zero dual and therefore form part of the
conflict. I just printed the dual proof of infeasibility, so it's not
guaranteed to be irreducible (i.e., it's possible that removing one of
these constraints doesn't render the model feasible).
For interval constraints, if the dual is positive (resp. negative), the
lower (resp. upper) bound is the one that's active.
—
Reply to this email directly, view it on GitHub
<#134 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAZQW3KATTVZFISOX2TGIN3W3IZHDANCNFSM6AAAAAAVUCCQV4>
.
You are receiving this because you authored the thread.Message ID:
***@***.***>
|
No. JuMP's The code I shared identifies a set of constraints that, together, render the model infeasible. There is no primal solution involved in that. |
|
Tank you @mtanneau for crafting the |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Is it easy to add support for compute_conflict!(model) as described in JuMP's docs?
The text was updated successfully, but these errors were encountered: