TongCHEN779/MultiMomOpt

Multi-order moment problem based on Lasserre's hierarchy

MultiMomOpt

Sublevel hierarchy is an intermediate hierarchy between the d-th order and (d+1)-th order Lasserre's hierarchy for sparse and non-sparse polynomial optimization problems. This repository is the code for paper A Sublevel Moment-SOS Hierarchy for Polynomial Optimization. For more information, contact me: toch@di.ku.dk.

Dependencies

JuMP, AMD, LinearAlgebra, MAT, SparseArrays, LightGraphs, GraphPlot, DynamicPolynomials, MosekTools, Printf, Dualization, Random, SumOfSquares, ProgressBars, StatsBase.

Usage

Maximize x1*x2, subject to x1^2+x2^2<=1.

using MultiMomOpt
typ = "max"; obj = [1 1 1];
options = Dict();
options["silent"] = true; options["quad"] = false;
MomConst = Array{Any}(undef, 1);
MomConst[1] = Dict();
MomConst[1]["basis"] = [1 0; 0 1];
MomConst[1]["ord"] = 2;
LocConst = Array{Any}(undef, 1);
LocConst[1] = Dict();
LocConst[1]["basis"] = [1 0; 0 1];
LocConst[1]["pol"] = [1 0 0; -1 2 0; -1 0 2];
LocConst[1]["typ"] = ">=";
LocConst[1]["ord"] = 1;
OptVal, running_time, status = solve_moment_manual(typ, obj, MomConst, LocConst, options);

Examples

Max-Cut

using MultiMomOpt
vars = matread("maxcut.mat");
A = vars["A"]; W = ones(size(A, 1), size(A, 1));
options = Dict();
options["level"] = 3; options["clique"] = "off"; options["ord"] = 2; options["silent"] = true; options["quad"] = true;
OptVal, running_time, status = solve_moment_maxcut(A, W, options)

Max-Cliq

using MultiMomOpt
vars = matread("mac_5.mat");
A = vars["A"];
options = Dict();
options["level"] = 3; options["clique"] = "off"; options["ord"] = 2; options["silent"] = true; options["quad"] = true;
OptVal, running_time, status = solve_moment_mac(A, options)

MIQCP

using MultiMomOpt
vars = matread("mip.mat");
A = vars["L"];
options = Dict();
options["level"] = 15; options["clique"] = "off"; options["ord"] = 2; options["silent"] = true; options["quad"] = true;
OptVal, running_time, status = solve_moment_mip(A, options);

QCQP

using MultiMomOpt
vars = matread("qcqp_5_1.mat");
A = vars["A"]; b = vars["b"];
options = Dict();
options["level"] = 3; options["clique"] = "off"; options["ord"] = 2; options["silent"] = true; options["quad"] = true;
OptVal, running_time, status = solve_moment_qcqp(A, b, options)

Lipschitz Constant Estimation problem (one hidden layer)

using MultiMomOpt
vars = matread("lip_test.mat");
A = vars["A"]; b = vars["b"]; c = vars["c"]; x00 = vars["x00"]; eps = 0.1;
options = Dict();
options["range"] = "global"; options["level"] = 6; options["clique"] = "off"; options["ord"] = 2; options["silent"] = true; options["quad"] = true;
OptVal, running_time, status = solve_moment_lip(A, b, c, x00, eps, options);

Robustness Certification problem

using MultiMomOpt
vars = matread("lip_test.mat");
A = vars["A"]; b = vars["b"]; c = vars["c"]; x00 = vars["x00"]; eps = 0.1;
options = Dict();
options["range"] = "global"; options["level"] = 0; options["clique"] = "off"; options["ord"] = 2; options["silent"] = true; options["quad"] = true;
OptVal, running_time, status = solve_moment_cert(A, b, c, x00, eps, options);

References

The Lipschitz Constant Estimation problem is referred to Semialgebraic Optimization for Lipschitz Constants of ReLU Networks. The Robustness Certification problem is referred to Semidefinite relaxations for certifying robustness to adversarial examples. For more information, contact me: toch@di.ku.dk.

Ellipsoid Problem

Finding the minimum volume ellipsoid of the image under ReLU networks

Dependencies

Plots, SumOfSquares, DynamicPolynomials, MosekTools, LinearAlgebra, Printf, MathOptInterface, StatsBase

Usage

n = 10; Q10 = Matrix(-I(n)); b10 = zeros(n,1); c10 = 1;
L10 = 2; P10 = Array{Any}(undef, L10); ξ10 = Array{Any}(undef, L10);
for i = 1:L10-1
    P10[i] = randn(n,n)*0.5; ξ10[i] = randn(n,1);
    for j = 1:n
        P10[i][j,j] = 1
    end
end
P10[L10] = rand(2,n)*0.5; ξ10[L10] = rand(2,1); P10[L10][1,1] = 1; P10[L10][2,2] = 1;

p11, m = OuterApproximationPlotSampling(Q10, b10, c10, 1, P10, ξ10, "Morari", L10);
p21 = OuterApproximationPlotSampling(Q10, b10, c10, 2, P10, ξ10, "sublevel", L10, lv=2, morari=m, meth="cycle_v");
p31 = OuterApproximationPlotSampling(Q10, b10, c10, 2, P10, ξ10, "sublevel", L10, lv=2, morari=m, meth="order_v");
p12 = OuterApproximationPlotSampling(Q10, b10, c10, 2, P10, ξ10, "sublevel", L10, lv=3, morari=m, meth="order_v");
p22 = OuterApproximationPlotSampling(Q10, b10, c10, 2, P10, ξ10, "sublevel", L10, lv=4, morari=m, meth="order_v");
p32 = OuterApproximationPlotSampling(Q10, b10, c10, 2, P10, ξ10, "sublevel", L10, lv=5, morari=m, meth="order_v");
p13 = OuterApproximationPlotSampling(Q10, b10, c10, 2, P10, ξ10, "sublevel", L10, lv=6, morari=m, meth="order_v");
p23 = OuterApproximationPlotSampling(Q10, b10, c10, 2, P10, ξ10, "sublevel", L10, lv=7, morari=m, meth="order_v");
p33 = OuterApproximationPlotSampling(Q10, b10, c10, 2, P10, ξ10, "sublevel", L10, lv=8, morari=m, meth="order_v");

plot(p11, p21, p31, p12, p22, p32, p13, p23, p33, layout = Plots.grid(3,3), fmt = :png)

References

Reach-SDP: Reachability Analysis of Closed-Loop Systems with Neural Network Controllers via Semidefinite Programming. For more information, contact me: toch@di.ku.dk.