Main Content

Explore patternsearch Algorithms

Starting with R2022b, patternsearch has four algorithms:

  • "classic"

  • "nups" (Nonuniform Pattern Search)

  • "nups-gps"

  • "nups-mads"

This example shows how the choice of algorithm affects a bounded problem with a nonsmooth objective function.

Objective Function

The objective function for this example is based on the 2-D ps_example function that is available when you run the example.

type ps_example
function f = ps_example(x)
%PS_EXAMPLE objective function for patternsearch.

%   Copyright 2003-2021 The MathWorks, Inc.

f = zeros(1,size(x,1));
for i = 1:size(x,1)
    if  x(i,1) < -5
        f(i) = (x(i,1)+5)^2 + abs(x(i,2));
    elseif x(i,1) < -3
        f(i) = -2*sin(x(i,1)) + abs(x(i,2));
    elseif x(i,1) < 0
        f(i) = 0.5*x(i,1) + 2 + abs(x(i,2));
    elseif x(i,1) >= 0
        f(i) = .3*sqrt(x(i,1)) + 5/2 +abs(x(i,2));
    end
end

Extend the ps_example function to any even number of dimensions by creating pseudorandom offsets from the origin for every two dimensions, and adding the resulting objectives. See the code for the testps helper function at the end of this example.

Specify N = 10 dimensions, and set bounds of –20 to 20 for each component. Start patternsearch from the point x0 = [0 0 ... 0].

N = 10;
lb = -20*ones(1,N);
ub = -lb;
x0 = zeros(size(lb));

Run Classic Algorithm

Set options to use the "classic" algorithm and to display the objective function values.

optscl = optimoptions("patternsearch",Algorithm="classic",PlotFcn="psplotbestf");

Run the optimization.

[xclassic,fvalclassic,eflagclassic,outputclassic] = ...
    patternsearch(@testps,x0,[],[],[],[],lb,ub,[],optscl)
patternsearch stopped because the mesh size was less than options.MeshTolerance.

xclassic = 1×10

   -7.9575    5.9058   -8.5047   -5.5481   -8.9402   -2.8633   -7.5058    0.8919   -5.6967    2.9322

fvalclassic = -10.0000
eflagclassic = 1
outputclassic = struct with fields:
         function: @testps
      problemtype: 'boundconstraints'
       pollmethod: 'gpspositivebasis2n'
    maxconstraint: 0
     searchmethod: []
       iterations: 218
        funccount: 3487
         meshsize: 9.5367e-07
         rngstate: [1x1 struct]
          message: 'patternsearch stopped because the mesh size was less than options.MeshTolerance.'

Run NUPS Algorithm

Set options to use the "nups" algorithm. To obtain reproducible results, set the random seed.

optsnups = optimoptions(optscl,Algorithm="nups");
rng default % For reproducibility
[xnups,fvalnups,eflagnups,outputnups] = ...
    patternsearch(@testps,x0,[],[],[],[],lb,ub,[],optsnups)
patternsearch stopped because the mesh size was less than options.MeshTolerance.

xnups = 1×10

   -7.9575    5.9058   -8.5047   -5.5480   -8.9401   -2.8633   -7.5058    0.8919   -5.6967    2.9322

fvalnups = -9.9999
eflagnups = 1
outputnups = struct with fields:
         function: @testps
      problemtype: 'boundconstraints'
       pollmethod: 'nups'
    maxconstraint: 0
     searchmethod: []
       iterations: 183
        funccount: 1827
         meshsize: 8.5928e-07
         rngstate: [1x1 struct]
          message: 'patternsearch stopped because the mesh size was less than options.MeshTolerance.'

In this case, the "nups" algorithm reaches essentially the same solution as the "classic" algorithm while using fewer function evaluations.

Further Explorations

You can also try the "nups-gps" and "nups-mads" algorithms to see how they perform on this problem. Although you cannot always predict which algorithm works best on a problem, "nups" can be a good starting algorithm because it incorporates many adaptive features, making it likely to be the best algorithm.

Helper Function

This code creates the testps helper function.

function y = testps(x) 
    N = numel(x); % Number of variables
    if mod(N,2) == 1
        disp("Number of variables must be even.")
        return
    end
    strm = RandStream("twister",Seed=1); % Set Seed for consistency
    % Use RandStream to avoid affecting the global stream
    dsp = 5*randn(strm,size(x));
    z = x - dsp; % Include random offsets
    y = 0;
    for i = 1:N/2
        y = y + ps_example(z([2*i-1,2*i]));
    end
end

See Also

Related Topics