% generate some data (50 data points defined in two dimensions;

% class assignment is 0 or 1 for each data point)

x1 = randn(50,1);
x2 = randn(50,1);
y = (2*x1 + x2 + randn(size(x1))) > 1;
% visualize the data
figure; hold on;
h1 = scatter(x1(y==0),x2(y==0),50,'k','filled'); % black dots for 0
h2 = scatter(x1(y==1),x2(y==1),50,'w','filled'); % white dots for 1
set([h1 h2],'MarkerEdgeColor',[.5 .5 .5]); % outline dots in gray
legend([h1 h2],{'y==0' 'y==1'},'Location','NorthEastOutside');
xlabel('x_1');
ylabel('x_2');

define optimization options

options = optimset('Display','iter','FunValCheck','on', ...

'MaxFunEvals',Inf,'MaxIter',Inf, ...
'TolFun',1e-6,'TolX',1e-6);
% define logistic function (this takes a matrix of values and
% applies the logistic function to each value)
logisticfun = @(x) 1./(1+exp(-x));
% construct regressor matrix
X = [x1 x2];
X(:,end+1) = 1; % we need to add a constant regressor

1
% define initial seed
params0 = zeros(1,size(X,2));
% define bounds
paramslb = []; % [] means no bounds
paramsub = [];
% define cost function (negative-log-likelihood)
% here, the eps is a hack to ensure that the log returns a finite
number.
costfun = @(pp) -sum((y .* log(logisticfun(X*pp')+eps)) + ...
((1-y) .* log(1-logisticfun(X*pp')+eps)));
% fit the logistic regression model by finding
% parameter values that minimize the cost function
[params,resnorm,residual,exitflag,output] = ...
lsqnonlin(costfun,params0,paramslb,paramsub,options);

Warning: Trust-region-reflective algorithm requires at least as many

equations
as variables; using Levenberg-Marquardt algorithm instead.

First-Order
Norm of
Iteration Func-count Residual optimality Lambda
step
0 4 1201.13 416 0.01
1 8 277.244 56 0.001
1.78489
2 12 148.623 27.9 0.0001
3.8386
3 22 132.374 12.2 100
0.320084
4 26 128.393 5.39 10
0.162361
5 31 127.375 3.18 100
0.0812274
6 35 127.017 2.88 10
0.046166
7 39 126.662 7.23 1
0.328918
8 45 125.814 3.97 100
0.0750915
9 49 125.552 2.49 10
0.0399111
10 53 125.487 4.43 1
0.271921
11 59 124.925 2.61 100
0.0594392
12 63 124.708 2.19 10
0.0360605
13 68 124.598 1.87 100
0.0249691
14 72 124.526 1.64 10
0.0197256
15 76 124.173 2.27 1
0.171948

2
 16 82 124.075 1.55 100
0.0238948
17 86 124.017 1.14 10
0.0178019
18 90 123.814 1.9 1
0.148335
19 96 123.727 1.55 100
0.022521
20 100 123.678 1.3 10
0.0165903
21 104 123.576 2.11 1
0.133349
22 110 123.487 1.43 100
0.0229012
23 114 123.442 1.03 10
0.0160575
24 118 123.404 1.78 1
0.123001
25 124 123.319 1.39 100
0.0223498
26 128 123.276 1.12 10
0.0157401
27 133 123.251 0.934 100
0.0117664
28 137 123.235 0.808 10
0.00950573
29 141 123.157 0.986 1
0.0828874
30 147 123.133 0.726 100
0.0115103
31 151 123.119 0.561 10
0.00889812
32 155 123.069 1.01 1
0.074771
33 161 123.047 0.813 100
0.011244
34 165 123.034 0.68 10
0.00848042
35 169 123.005 0.961 1
0.0689362
36 175 122.983 0.691 100
0.0113644
37 179 122.971 0.518 10
0.00826277
38 183 122.957 0.951 1
0.0646201
39 189 122.935 0.74 100
0.0113329
40 193 122.923 0.598 10
0.00813652
41 197 122.922 0.96 1
0.0618885
42 203 122.899 0.677 100
0.0116673

3
 43 207 122.887 0.494 10
0.0081493
44 212 122.881 0.375 100
0.00600049
45 216 122.877 0.297 10
0.00475153
46 220 122.86 0.542 1
0.0406697
47 226 122.854 0.44 100
0.00587151
48 230 122.85 0.371 10
0.00452326
49 234 122.84 0.483 1
0.037564
50 240 122.834 0.357 100
0.00586685
51 244 122.83 0.274 10
0.00437758
52 248 122.824 0.508 1
0.0351516
53 254 122.818 0.4 100
0.00585419
54 258 122.815 0.327 10
0.00428393
55 262 122.812 0.482 1
0.0334166
56 268 122.806 0.349 100
0.00595175
57 272 122.803 0.26 10
0.00425006
58 276 122.803 0.488 1
0.0321513

Local minimum possible.

lsqnonlin stopped because the final change in the sum of squares

relative to
its initial value is less than the selected value of the function
tolerance.

compute the output (class assignment) of the

model for each data point
modelfit = logisticfun(X*params') >= 0.5;
% calculate percent correct (percentage of data points
% that are correctly classified by the model)
pctcorrect = sum(modelfit==y) / length(y) * 100;
% visualize the model
% prepare a grid of points to evaluate the model at

4
ax = axis;
xvals = linspace(ax(1),ax(2),100);
yvals = linspace(ax(3),ax(4),100);
[xx,yy] = meshgrid(xvals,yvals);
% construct regressor matrix
X = [xx(:) yy(:)];
X(:,end+1) = 1;
% evaluate model at the points (but don't perform the final
thresholding)
outputimage = reshape(logisticfun(X*params'),[length(yvals)
length(xvals)]);
% visualize the image (the default coordinate system for images
% is 1:N where N is the number of pixels along each dimension.
% we have to move the image to the proper position; we
% accomplish this by setting XData and YData.)
h3 = imagesc(outputimage,[0 1]); % the range of the logistic function
is 0 to 1
set(h3,'XData',xvals,'YData',yvals);
colormap(hot);
colorbar;
% visualize the decision boundary associated with the model
% by computing the 0.5-contour of the image
[c4,h4] = contour(xvals,yvals,outputimage,[.5 .5]);
set(h4,'LineWidth',3,'LineColor',[0 0 1]); % make the line thick and
blue
% send the image to the bottom so that we can see the data points
uistack(h3,'bottom');
% send the contour to the top
uistack(h4,'top');
% restore the original axis range
axis(ax);
% report the accuracy of the model in the title
title(sprintf('Classification accuracy is %.1f%%',pctcorrect));

5
Published with MATLAB® R2018b