plot_sparsity (generic function with 1 method)Difference-of-Convex Algorithm with Frank-Wolfe
This example shows the optimization of a difference-of-convex problem of the form:
\[min\limits_{x \in \mathcal{X}} \phi(x) = f(x) - g(x)\]
with $f$, $g$ convex functions with access to subgradients and $f$ smooth.
The DCA-FW algorithm constructs local convex models of $\phi$ by linearizing $g$ and approximately optimizes them with FW. It is a local method that converges to a stationary point
using FrankWolfe
using LinearAlgebra
using Random
using SparseArrays
using StableRNGs
using RandomThe convex functions $f$, $g$ will be generated as random convex quadratics:
\[\begin{align} f(x) & = \frac12 x^\top A x + a^\top x + c \\ g(x) & = \frac12 x^\top B x + b^\top x + d \end{align}\]
Setting up the problem functions and data
const n = 500 # Reduced dimension
# Generate random positive definite matrices to ensure convexity
function generate_problem_data(rng)
A_raw = randn(rng, n, n)
A_raw ./= opnorm(A_raw)
A = A_raw' * A_raw + 0.1 * I
B_raw = randn(rng, n, n)
B_raw ./= opnorm(B_raw)
B = B_raw' * B_raw + 0.1 * I
a = randn(rng, n)
b = randn(rng, n)
c = randn(rng)
d = -359434.0
return A, B, a, b, c, d
end
const rng = StableRNGs.StableRNG(1)
const A, B, a, b, c, d = generate_problem_data(rng)([0.3664716461545843 0.007434201971305423 … 0.004451785375884827 0.015514411527022787; 0.007434201971305423 0.34822813947414344 … 0.013654308840956007 -0.025910486878046375; … ; 0.004451785375884827 0.013654308840956007 … 0.3806261404722967 0.0118969109481474; 0.015514411527022787 -0.025910486878046375 … 0.0118969109481474 0.34299337924542433], [0.36858235928505056 0.0036477368657639084 … -0.014747118492049596 -0.009390758842324287; 0.0036477368657639084 0.36832277324228413 … -0.004708423922308739 0.019319726914150886; … ; -0.014747118492049596 -0.004708423922308739 … 0.3718244447067395 -0.0070589804508730414; -0.009390758842324287 0.019319726914150886 … -0.0070589804508730414 0.36976970853592206], [-0.5056985849854947, 1.8254695540945816, 1.1686948924955953, 0.025865043318803068, 0.7618207158940428, -1.9599078996573032, -0.28701522778989075, -1.210555966368952, -0.3476555815275666, 0.36622748810084865 … -1.2418975260665193, -2.042717239142079, -0.977247840514734, -0.9480533072182336, 0.2354971732415276, 1.1489207372892907, 0.26713737924876685, -0.11111819389928967, 1.2106291210589706, -0.36135542009200855], [-1.7352231531269595, -1.0889348569606767, -1.9696069675033783, -0.6944538146281424, 0.12594479576421444, 2.43890963984164, -0.3332102207771672, -1.598587279152474, 0.8970898814075347, 0.7441204116527911 … -0.07885951683533911, 0.921491279700674, 0.8755353895807455, 0.4142785039140312, 1.0607172005644177, -0.9845905380360871, 0.7603797972500105, 1.0314248598279463, -1.0727076372535254, -0.1591576537893938], -0.6622940864130684, -359434.0)We can now define the two functions
function f(x)
return 0.5 * FrankWolfe.fast_dot(x, A, x) + dot(a, x) + c
end
function grad_f!(storage, x)
mul!(storage, A, x)
storage .+= a
return nothing
end
function g(x)
return 0.5 * FrankWolfe.fast_dot(x, B, x) + dot(b, x) + d
end
function grad_g!(storage, x)
mul!(storage, B, x)
storage .+= b
return nothing
end
# True objective function for verification
# It is not needed by the solver
function phi(x)
return f(x) - g(x)
end
lmo = FrankWolfe.KSparseLMO(5, 1000.0)
x0 = FrankWolfe.compute_extreme_point(lmo, randn(n))
res_dca = FrankWolfe.dca_fw(
f,
grad_f!,
g,
grad_g!,
lmo,
copy(x0),
max_iteration=200, # Outer iterations
max_inner_iteration=10000, # Inner iterations
epsilon=1e-5, # Tolerance for DCA gap
line_search=FrankWolfe.Secant(),
verbose=true,
trajectory=true,
verbose_inner=false,
print_iter=10,
use_corrective_fw=true,
use_dca_early_stopping=true,
grad_f_workspace=collect(x0),
grad_g_workspace=collect(x0),
)
res_boost = FrankWolfe.dca_fw(
f,
grad_f!,
g,
grad_g!,
lmo,
copy(x0),
boosted=true,
max_iteration=200, # Outer iterations
max_inner_iteration=10000, # Inner iterations
epsilon=1e-5, # Tolerance for DCA gap
line_search=FrankWolfe.Secant(),
verbose=true,
trajectory=true,
verbose_inner=false,
print_iter=10,
use_corrective_fw=true,
use_dca_early_stopping=true,
grad_f_workspace=collect(x0),
grad_g_workspace=collect(x0),
)(x = [85 ] = -521.751
[284] = 779.515
[313] = -1000.0
[348] = 943.152
[369] = -755.582
[486] = -1000.0, primal = 10284.822655046242, dca_gap = 0.7536499968576128, iterations = 200, status = FrankWolfe.STATUS_MAXITER, traj_data = Any[(1, 262240.8388926722, 206673.48945768716, 55567.349434985044, 0.000266657), (2, 232480.17113008053, 205323.0813233029, 27157.08980677762, 0.001445949), (3, 214392.43470281863, 196651.49094151697, 17740.94376130167, 0.002114938), (4, 198902.63665069378, 184360.61699834905, 14542.019652344712, 0.002892118), (5, 181669.09573550435, 166430.01393149866, 15239.081804005684, 0.003672325), (6, 163592.85582701804, 145528.58149260594, 18064.27433441209, 0.004378293), (7, 147711.04373584787, 129881.7321535621, 17829.311582285765, 0.005096103), (8, 142613.38159836666, 137864.40163595832, 4748.979962408335, 0.005856983), (9, 140118.5638997993, 138112.54266489664, 2006.0212349026642, 0.006789243), (10, 138541.8827375109, 137448.78572191967, 1093.0970155912091, 0.00758055) … (192, 10297.150958738988, 10296.237293165508, 0.9136655734801025, 0.07228356), (193, 10295.4531352357, 10294.586727716991, 0.8664075187100816, 0.072578791), (194, 10293.802819822915, 10292.858873713882, 0.943946109033277, 0.072863673), (195, 10292.198888173676, 10291.342270252791, 0.8566179208854602, 0.073111646), (196, 10290.639355245396, 10289.663764068495, 0.97559117690097, 0.07337626), (197, 10289.12336484855, 10288.313572655383, 0.8097921931678589, 0.073629623), (198, 10287.649383946555, 10286.85517398215, 0.7942099644044812, 0.073897273), (199, 10286.216214430868, 10285.442568169989, 0.7736462608793687, 0.074174831), (200, 10284.822655046242, 10284.069005049385, 0.7536499968576128, 0.074432081), (200, 10284.822655046242, 10284.069005049385, 0.7536499968576128, 0.074908981)])Plotting the resulting trajectory
We modify the y axis to highlight that we are plotting the DCA gap, not the FW gap
data = [res_dca.traj_data, res_boost.traj_data]
label = ["DCA-FW", "DCA-FW-B"]
p_res = plot_trajectories(data, label, marker_shapes=[:o, :x])
ylabel!(p_res.subplots[3], "DCA gap")display(p_res)This page was generated using Literate.jl.