Implementation of smoothing stabilization

[rrsadykov]

No description provided.

[codecov]

Codecov Report

Merging #348 into master will decrease coverage by 2.23%.
The diff coverage is 50.68%.

@@            Coverage Diff             @@
##           master     #348      +/-   ##
==========================================
- Coverage   76.03%   73.79%   -2.24%     
==========================================
  Files          51       52       +1     
  Lines        3764     3931     +167     
==========================================
+ Hits         2862     2901      +39     
- Misses        902     1030     +128     

Impacted Files	Coverage Δ
src/Algorithm/Algorithm.jl	100.00% <ø> (ø)
src/Algorithm/branching/branchingalgo.jl	88.54% <0.00%> (ø)
src/Algorithm/conquer.jl	71.01% <ø> (ø)
src/Algorithm/storage.jl	78.15% <ø> (-0.37%)	⬇️
src/Algorithm/colgenstabilization.jl	8.33% <8.33%> (ø)
src/Algorithm/colgen.jl	81.01% <80.35%> (-2.10%)	⬇️
src/Algorithm/basic/solveipform.jl	87.50% <100.00%> (+0.54%)	⬆️
src/Algorithm/formstorages.jl	37.22% <100.00%> (+0.92%)	⬆️
src/ColunaBase/solsandbounds.jl	53.52% <0.00%> (-4.23%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 22a3693...3e74b6f. Read the comment docs.

[rrsadykov]

@guimarqu What is the best way to do a linear combination of two DualSolution? For example, if we have solutions (constr_1 = 2, constr_2 = 2) and (constr_1 = 1, constr_3 = 1), then linear combination with coefficient 0.5 should be (constr_1 = 1.5, constr_2 = 1, constr_3 = 0.5). In addition, the coefficient should depend on the constraint type (coefficient should be different for convexity constraints).

[guimarqu]

@guimarqu What is the best way to do a linear combination of two DualSolution? For example, if we have solutions (constr_1 = 2, constr_2 = 2) and (constr_1 = 1, constr_3 = 1), then linear combination with coefficient 0.5 should be (constr_1 = 1.5, constr_2 = 1, constr_3 = 0.5). In addition, the coefficient should depend on the constraint type (coefficient should be different for convexity constraints).

If the coefficient depends on the duty, it’s better to use a loop.

[rrsadykov]

If the coefficient is independent, is it possible to perform the linear combination in O(n)?
Because if one uses a loop, the complexity is O(n\log n), no?
I think that the best is to do the linear combination with the same coefficient in O(n), and then make another loop in O(n) to correct the coefficients of convexity constraints.

[guimarqu]

Loop is O(n), lookup is O(log n).
If you do something like that, you are in O(n) :

constrids = [constrid for (constrid, _) in getconstr(form)] # sorted ? If not, we should consider keeping the constraint dict sorted
constrvals = zeros(Float64, length(constrids)

pos = 1
for (constrid, constrval) in dualsolution # sorted and O(n)
    while constrids[pos] < constrid
         pos += 1
    end
    if ! (constrids[pos] > constrid)
        constrvals[pos] += coefficient * constrval
    end
end

[guimarqu]

It's not a good idea to keep the constraints dict sorted. If we assume that no constraints are added during column generation, we can sort the constrids once at the beginning of colgen.

[rrsadykov]

Ok, I will first do a simple version in O(n\log n). We will see latter if we need to improve that

[rrsadykov]

@guimarqu In Solution.sol, ConstrId are sorted or not?

[guimarqu]

@guimarqu In Solution.sol, ConstrId are sorted or not?

Yes, they are.

[rrsadykov]

@guimarqu Guillaume, can you review and merge? All tests without stabilization work.
I prefer to do an "intermediate" merge now as there are many changes in column generation algorithm and it will simplify merging with other branches in the future.

For the moment, stabilization is not quite ready yet, but the column generation algorithm is fully prepared for it. I think I need a couple days more to finish the implementation and add tests.

Can you leave the branch after merge so I can continue

[guimarqu]

• colgen code looks much better thanks to SubprobInfo
• EmptyStorage becomes FormulationStorage ( Names for Data/Storage/StorageState #330 )
• new algorithm for stabilisation.

Only style issues :

• there is a lot of comments to delete
• return is mandatory at the end of a method (style guide https://github.com/invenia/BlueStyle)

On github, 1 branch = 1 PR, you can create a new branch from this one or after the merge of this one into master.

[guimarqu]

Do you want to finish stabilization before release of 0.3 (we can wait until mid-June max) ?

[rrsadykov]

I think I can finish this week