added a new test of infeas. prob. with 2 constraints#159
added a new test of infeas. prob. with 2 constraints#159mlubin merged 5 commits intojump-dev:masterfrom
Conversation
test/contlinear.jl
Outdated
| cd1 = MOI.get(m, MOI.ConstraintDual(), c1) | ||
| cd2 = MOI.get(m, MOI.ConstraintDual(), c2) | ||
| bndyd = MOI.get(m, MOI.ConstraintDual(), bndy) | ||
| @test (cd2-bndyd)/cd1 ≈ 3 atol=atol rtol=rtol |
There was a problem hiding this comment.
Could you add tests for the signs of the dual values ?
I would also rather write - 3 cd1 + cd2 ≈ bndyd and 2 cd1 ≈ bndxd to make it clearer why the test should be true by being closer to the dual constraint.
There was a problem hiding this comment.
We could also add a test for the objective -7a + 2b > 0
test/contlinear.jl
Outdated
| @test cd1 < atol | ||
| @test cd2 < atol | ||
| @test - 3 * cd1 + cd2 ≈ bndyd atol=atol rtol=rtol | ||
| @test 2 * cd1 ≈ bndxd atol=atol rtol=rtol |
| @test MOI.canget(m, MOI.ConstraintDual(), c1) | ||
| cd1 = MOI.get(m, MOI.ConstraintDual(), c1) | ||
| cd2 = MOI.get(m, MOI.ConstraintDual(), c2) | ||
| bndyd = MOI.get(m, MOI.ConstraintDual(), bndy) |
|
|
||
| function contlineartest(solver::MOI.AbstractSolver; atol=Base.rtoldefault(Float64), rtol=Base.rtoldefault(Float64)) | ||
| linear1test(solver, atol=atol, rtol=rtol) | ||
| linear2test(solver, atol=atol, rtol=rtol) |
There was a problem hiding this comment.
You should add linear12test in this function
blegat
left a comment
blegat
left a comment
There was a problem hiding this comment.
Test passing with SDPA and CSDP
test/contlinear.jl
Outdated
| @test cd2 < atol | ||
| @test - 3 * cd1 + cd2 ≈ -bndyd atol=atol rtol=rtol | ||
| @test 2 * cd1 ≈ -bndxd atol=atol rtol=rtol | ||
| @test -7 * cd1 + 2 * cd2 > -atol |
There was a problem hiding this comment.
Shouldn't this be > atol? The dual objective needs to be strictly positive.
There was a problem hiding this comment.
I m not sure. For me, having a tolerance towards a solver meaning that we accept to have an error but only within that tolerance. Namely, abs(expectedValue - actualValue) <= absoluteTolerance must be true
There was a problem hiding this comment.
Here we have no actual value to test, we're checking for strict positivity. If -7 * cd1 + 2 * cd2 == 0 then we do not have a valid infeasibility proof.
There was a problem hiding this comment.
Here is my reasoning:
The test would be correct with 0 tolerance if there exist epsilon > 0 such that -7 * cd1 + 2 * cd2 = epsilon.
Therefore the test would be correct with absoluteTolerance if there exist epsilon > 0 such that -7 * cd1 + 2 * cd2 = actualvalue and abs(epsilon - actualValue) <= absoluteTolerance
There was a problem hiding this comment.
We can't let epsilon get arbitrarily small (smaller than absoluteTolerance). The current tests accepts cd1, cd2, bndxd, bndyd = 0.0, 0.0 , 0.0 , 0.0 which is incorrect, and I'd say an important case to catch.
There was a problem hiding this comment.
I understand your concern but I don't think using > atol is the correct solution because it is not consistant with the general definition of tolerance ("accepting incorrect answers as long as they are not too far from the correct answer"). It is better to define precisely what atol means for us. Solvers accept many tolerances such as feasibility/optimality tolerance... Since we use only atol everywhere I assume it s the largest one. We can enforce that infeasibility/unbounded ray are correct as well with atol tolerance, when scaled such as the biggest component in the vector has absolute value 1. In this case we need to first scale the ray using infinite norm, and then test the exact values using atol.
There was a problem hiding this comment.
Ok, since you both agree on it.
There was a problem hiding this comment.
I agree that this is a nominally improper use of atol. Rescaling the ray and then comparing with a value would also be okay. We are going to have to revisit tolerances in the future anyway: #144.
(this is IMO not the right solution to avoid the 0 vector being an accepted ray, but mlubin and blegat agrees on it)
3 participants
replaces #156