Contents

Constrained optimization with FMINSEARCH

function testoptimize

Introduction

Usage

first, define a test function:

    clc, rosen = @(x) (1-x(1))^2 + 105*(x(2)-x(1)^2)^2;

this is the classical Rosenbr\"uck function, which has a global minimum at f(x) = f([1, 1]) = 0. The function is relatively hard to optimize, because that minimum is located in a long narrow ``valley'':

    k = 0; range = -5:0.1:5;
    z = zeros(101);
    for i = range
        m = 0; k = k + 1;
        for j = range
            m = m + 1;
            z(k, m) = rosen([i, j]);
        end
    end
    [y, x] = meshgrid(range, range);
    surf(x, y, z, 'linestyle', 'none'), view(-150, 30), axis tight

Optimizing the fully unconstrained problem with OPTIMIZE indeed finds the global minimum:

    solution = optimize(rosen, [3 3])

solution =

    1.0000
    1.0000

Imposing a lower bound on the variables gives

    [solution, fval] = optimize(rosen, [3 3], [2 2], [])

solution =

    2.0000
    4.0000


fval =

    1.0000

in the figure, this looks like

    zz = z;   zz(x > 2 & y > 2) = inf;
    ZZ = z;   ZZ(x < 2) = inf;  ZZ(y < 2) = inf;
    figure, hold on
    surf(x, y, zz, 'linestyle', 'none', 'FaceAlpha', 0.2)
    surf(x, y, ZZ, 'linestyle', 'none')
    view(-150, 30), grid on, axis tight
    plot3(solution(1), solution(2), fval+1e3, 'g.', 'MarkerSize', 20)

Similarly, imposing an upper bound yields

    solution = optimize(rosen, [3 3], [], [0.5 0.5])

    zz = z;   zz(x < 0.5 & y < 0.5) = inf;
    ZZ = z;   ZZ(x > 0.5) = inf;  ZZ(y > 0.5) = inf;
    figure, hold on
    surf(x, y, zz, 'linestyle', 'none', 'FaceAlpha', 0.2)
    surf(x, y, ZZ, 'linestyle', 'none')
    view(150, 30), grid on, axis tight
    plot3(solution(1), solution(2), fval+1e3, 'g.', 'MarkerSize', 20)

solution =

    0.5000
    0.2500

Optimize with x(2) fixed at 3. In this case, OPTIMIZE simply removes the variable before FMINSEARCH sees it, essentially reducing the dimensionality of the problem. This is particularly useful when the number of dimensions N becomes large.

    optimize(rosen, [3 3], [-inf 3], [inf 3])

ans =

    1.7314
    3.0000

Also general nonlinear constraints can be used. A simple example:

nonlinear inequality:

nonlinear equality :

    options = optimset('TolFun', 1e-8, 'TolX', 1e-8);
    [sol, fval, exitflag, output] = optimize(rosen, [3 -3], [], [], [], ...
        [], [], [], @nonlcon, [], options);

Note that nonlcon is a subfunction, listed below. In a figure, this looks like

    zz = z;   zz(sqrt(x.^2 + y.^2) <= 1) = inf;
    ZZ = z;   ZZ(sqrt(x.^2 + y.^2) >= 1.2) = inf;
    zZ = z;   zZ(x.^2 + y.^3 >= 0.5 + 0.1) = inf;
              zZ(x.^2 + y.^3 <= 0.5 - 0.1) = inf;
    figure, hold on
    surf(x, y, zz, 'linestyle', 'none', 'FaceAlpha', 0.2)
    surf(x, y, ZZ, 'linestyle', 'none')
    xX = x(isfinite(zZ)); xX = xX(:);
    yY = y(isfinite(zZ)); xX = xX(:);
    zZ = zZ(isfinite(zZ)); zZ = zZ(:);
    [xX, inds] = sort(xX); yY = yY(inds); zZ = zZ(inds);
    xyz = [xX, yY, zZ];
    for i = 1:length(xX)-1 % line-command is *somewhat* inconvenient...
        l = line( [xyz(i, 1); xyz(i+1, 1)],[xyz(i, 2); xyz(i+1, 2)], [xyz(i, 3); xyz(i+1, 3)]);
        set(l, 'color', 'r', 'linewidth', 2)
    end
    view(150, 50), grid on, axis tight
    plot3(sol(1), sol(2), fval+1e3, 'g.', 'MarkerSize', 20)

Note that the output structure contains a field ``constrviolation'':

    output

output = 

         iterations: 123
          funcCount: 230
          algorithm: 'Nelder-Mead simplex direct search'
            message: [1x196 char]
    constrviolation: [1x1 struct]

The contents of which shows that all constraints have been satisfied:

    output.constrviolation

ans = 

      nonl_eq: {[0]  [0]}
    nonl_ineq: {[0]  [0]}

end

function [c, ceq] = nonlcon(x)
    c = norm(x) - 1;
    ceq = x(1)^2 + x(2)^3 - 0.5;
end

Published with MATLAB® 7.7