Windows long CI runs

[dhendryc]

1. Happens in bcg_directions_error.jl.
2. After in iteration 918, we go into a simplex descent step over the convex hull. However, the step size computed is always zero. Hence, we do this until t + number_of_iterations > 3000 where t=918 is the global FW iteration and number_of_steps counts the number of simplex descent steps. Aka we have do more than 2000 simplex descent steps which takes ages.

One example print out from the simplex descent:

    SD           918   3.153456e+03   3.153456e+03   2.588143e-05   9.904006e+00   9.268976e+01            270            292
number_of_steps = 8
(tolerance, maximum(c), minimum(c), maximum(c) - minimum(c)) = (2.5881426935825402e-5, -171.1251156092066, -171.12514237907425, 2.6769867645271006e-5)
η = 0.021773492508886032
(f(x), f(y)) = (3153.4557328848196, 3153.4557328848796)
(γ, dot_dir / (line_search.L_est * ndir2), dot_dir, ndir2, line_search.L_est) = (0.0, -0.00927877237125842, -4.154008092374794e-13, 5.7778230358668776e-11, 0.7748409796162811)
┌ Warning: Smoothness estimate run away -> hard clipping. Convergence might be not guaranteed.
└ @ FrankWolfe C:\Users\dhend\Documents\Arbeit\ZIB\research_projects\FrankWolfe.jl\src\linesearch.jl:777
(γ, dot_dir / (line_search.L_est * ndir2), dot_dir, ndir2, line_search.L_est) = (0.0, -0.009278772434189943049227665898572937708166942796916697520420849286994679255541762, -4.154008120548570910941155341278458107661182698888117446421716551924196991762983e-13, 5.777823035866878792528764382830880113477336502673130064104500852415977142141153e-11, 0.7748409796162811)
gamma = 0.0

Considering that maximum(c) - minimum(c), so the gap within the active set, is nearly at the tolerance, this could be a numerics issue.

[dhendryc]

• In the callback of the simplex descent, we should probably have t + number_of_steps as iteration, no? Currently, we get of bunch of SD steps with the same iteration counter,

[dhendryc]

Seems like this isn't a Windows issue per se.