Plot gaplotdistance in one plot for multiple runs of genetic algorithm

Dear all,

With the code below I managed to run the genetic algorithm multiple times.

gprMdl2 = fitrgp(X,Y1,'KernelFunction','squaredexponential','OptimizeHyperparameters','auto','HyperparameterOptimizationOptions',struct('AcquisitionFunctionName','expected-improvement-plus'));

for i = 1:3
options = optimoptions('ga','CrossoverFrac',0.9,'PopulationSize',50,'StallGen',50,'Generations',70,'PlotFcn', {'gaplotbestf','gaplotdistance'});
fun = @(X) [abs(((predict(gprMdl2,X)-MFR_exp)/MFR_exp))];
[x_opt, Obj, exitflag,output] = ga(fun,2,[],[],[],[],[0.1 0.1], [0.9 0.9],[],[],options);
end

With PlotFcn I will get the plot of the fitness value vs genration and the average distance vs generation (as shown below). How can I combine the plots of each run into one plot?

[Mario Malic]

Hello,

You can make your own OutputFcn for ga that does what you want. You can do the plotting with plotyy function.

[Tessa Kol]

Dear Mario,

I saw indeed your suggestion also in order post. I tried to save the variables that create the plot by adjusting the gaplotdistance.m file, but I have haven't been able to do that succesfully.

function state = gaplotdistance(options,state,flag)
%GAPLOTDISTANCE Averages several samples of distances between individuals.
%   STATE = GAPLOTDISTANCE(OPTIONS,STATE,FLAG) plots an averaged distance
%   between individuals.
%
%   Example:
%    Create an options structure that uses GAPLOTDISTANCE
%    as the plot function
%     options = optimoptions('ga','PlotFcn',@gaplotdistance);
%
%    (Note: If calling gamultiobj, replace 'ga' with 'gamultiobj') 

%   Copyright 2003-2015 The MathWorks, Inc.

samples = 20;
choices = ceil(sum(options.PopulationSize) * rand(samples,2));
switch flag
    case 'init'
        population = state.Population;
        distance = 0;
        for i = 1:samples
            d = population(choices(i,1),:) - population(choices(i,2),:);
            distance = distance + sqrt( sum ( d.* d));
            save plotdistance.mat distance samples
        end
        plotDist = plot(state.Generation,distance/samples,'.');
        set(gca,'xlimmode','manual','zlimmode','manual', ...
            'alimmode','manual')
        set(gca,'xlim',[1,options.MaxGenerations]);
        set(plotDist,'Tag','gaplotdistance');
        xlabel('Generation','interp','none');
        ylabel('Average Distance');
        title('Average Distance Between Individuals','interp','none')

    case 'iter'
        population = state.Population;
        distance = 0;
        for i = 1:samples
            d = population(choices(i,1),:) - population(choices(i,2),:);
            distance = distance + sqrt( sum ( d.* d));
        end
        plotDist = findobj(get(gca,'Children'),'Tag','gaplotdistance');
        newX = [get(plotDist,'Xdata') state.Generation];
        newY = [get(plotDist,'Ydata') distance/samples];
        set(plotDist,'Xdata',newX,'Ydata',newY);
end

Do you have any suggestions on how make my own OutputFcn to create the respective plots?

[Mario Malic]

First, backup the original gaplotdistance.m file and create your own, under different name and call that one.

In the case of PlotFcn, optimisation solver creates a figure on its own, outside of plot function, so in case of OutputFcn, you should do it prior to calling the solver and set a name or a tag so you can find it in the OutputFcn.

OptimAxes = axes(1)

set(OptimAxes, 'Name', 'CustomOptimPlot')

Regarding the custom output function: Actually plotyy is not recommended anymore, yyaxis should be used. In this example right yyaxis is considered to be gaplotbestf as it has additional case 'done'.

function state = gaplotdistance(options,state,flag)
%GAPLOTDISTANCE Averages several samples of distances between individuals.
%   STATE = GAPLOTDISTANCE(OPTIONS,STATE,FLAG) plots an averaged distance
%   between individuals.
%
%   Example:
%    Create an options structure that uses GAPLOTDISTANCE
%    as the plot function
%     options = optimoptions('ga','PlotFcn',@gaplotdistance);
%
%    (Note: If calling gamultiobj, replace 'ga' with 'gamultiobj') 
%   Copyright 2003-2015 The MathWorks, Inc.

% Getting the handle for axes *** test if you can move this in case 'init'
Axes = findobj('Name', 'CustomOptimPlot')

% options for left - gaplotdistance.m
samples = 20;
choices = ceil(sum(options.PopulationSize) * rand(samples,2)); 

% options for right - gaplotbestf
if size(state.Score,2) > 1
    msg = getString(message('globaloptim:gaplotcommon:PlotFcnUnavailable','gaplotbestf'));
    title(msg,'interp','none');
    return;
end

switch flag
    case 'init'
        yyaxis left
        % Code for first PlotFcn
        
        yyaxis right
        % Code for second PlotFcn
    case 'iter'
        yyaxis left
        % Code for first PlotFcn
        
        yyaxis right
        % Code for second PlotFcn
     case 'done'
        yyaxis right
        % Code for second PlotFcn 
end

There might be some options that interfere with each other, like titles and labels, so consider that as well.

Edit: I will update comment later on the saving the variables

[Tessa Kol]

I tried saving the distance values with the code below:

 case 'iter'
        population = state.Population;
        distance = 0;
        for i = 1:samples
            d = population(choices(i,1),:) - population(choices(i,2),:);
            distance = distance + sqrt( sum ( d.* d));
            test.distance(i) = distance;
            save(horzcat('test',int2str(i),'.mat'),'test')
        end
        

Then the last value of distance devided by number of samples gives the average distance value only for the last generation (which is generation). But I don't know how the save the distance value for all generations.

[Mario Malic]

Unfortunately, I am not familiar with ga, if you're looking for all values of variable d, throughout all optimisations, this would be the way to do it.

function state = customgaplotdistance(options,state,flag)
%GAPLOTDISTANCE Averages several samples of distances between individuals.
%   STATE = GAPLOTDISTANCE(OPTIONS,STATE,FLAG) plots an averaged distance
%   between individuals.
%
%   Example:
%    Create an options structure that uses GAPLOTDISTANCE
%    as the plot function
%     options = optimoptions('ga','PlotFcn',@gaplotdistance);
%
%    (Note: If calling gamultiobj, replace 'ga' with 'gamultiobj') 

%   Copyright 2003-2015 The MathWorks, Inc.

persistent testdist  % change number 1
testdist(1,:) = [0 0]; % initialising the value

samples = 20;
choices = ceil(sum(options.PopulationSize) * rand(samples,2));
switch flag
    case 'init'
        population = state.Population;
        distance = 0;
        for i = 1:samples
            d = population(choices(i,1),:) - population(choices(i,2),:);
            distance = distance + sqrt( sum ( d.* d));
            testdist(end+1,:) = d; % change number 2
        end
        plotDist = plot(state.Generation,distance/samples,'.');
        set(gca,'xlimmode','manual','zlimmode','manual', ...
            'alimmode','manual')
        set(gca,'xlim',[1,options.MaxGenerations]);
        set(plotDist,'Tag','gaplotdistance');
        xlabel('Generation','interp','none');
        ylabel('Average Distance');
        title('Average Distance Between Individuals','interp','none')

    case 'iter'
        population = state.Population;
        distance = 0;
        for i = 1:samples
            d = population(choices(i,1),:) - population(choices(i,2),:);
            distance = distance + sqrt( sum ( d.* d));
            testdist(end+1,:) = d; % change number 3
            assignin('base', 'testdist', testdist) % it might be better to assign it and save it from the main file
%             save('testdist.mat', 'testdist')     % as it will save the file 3000+ times
        end
        plotDist = findobj(get(gca,'Children'),'Tag','gaplotdistance');
        newX = [get(plotDist,'Xdata') state.Generation];
        newY = [get(plotDist,'Ydata') distance/samples];
        set(plotDist,'Xdata',newX,'Ydata',newY);
end

[Tessa Kol]

I changed the following line of code

testdist(end+1,:) = distance

After that I devided every value of the testdist by the number of samples, which is 20. That is how you get the values of the average distance. Than I filterd out the average distance values of every optimization run and put each run in a seperate column. Every 20th value corresponds to the value in the column corresponds to the average distance value in the plot.

A1 = testdis(:,1)/20;
A2 = zeros(1281,3);
A2(1:1281,1)=A1(1:1281,1);
A2(2:1081,2)=A1(1282:2361,1);
A2(2:1221,3)=A1(2362:3581,1);
A3 = A2(41:20:1221,3);
A4 = A2(41:20:1081,2);
A5 = A2(41:20:1281,1);

The resulting plot is shown below

I want to thank you for your help. Can you post your comment as an answer so I can accept it.

[Mario Malic]

Great to hear it works! Actually, I put the distance first, and then I thought, but it doesn't show every single distance out there, so I set it back to d.

Answer 1

[Mario Malic]

Edited according to the change of testdist. More details in comments.

function state = customgaplotdistance(options,state,flag)
%GAPLOTDISTANCE Averages several samples of distances between individuals.
%   STATE = GAPLOTDISTANCE(OPTIONS,STATE,FLAG) plots an averaged distance
%   between individuals.
%
%   Example:
%    Create an options structure that uses GAPLOTDISTANCE
%    as the plot function
%     options = optimoptions('ga','PlotFcn',@gaplotdistance);
%
%    (Note: If calling gamultiobj, replace 'ga' with 'gamultiobj') 
%   Copyright 2003-2015 The MathWorks, Inc.
persistent testdist  % change number 1
testdist(1,:) = [0 0]; % initialising the value
samples = 20;
choices = ceil(sum(options.PopulationSize) * rand(samples,2));
switch flag
    case 'init'
        population = state.Population;
        distance = 0;
        for i = 1:samples
            d = population(choices(i,1),:) - population(choices(i,2),:);
            distance = distance + sqrt( sum ( d.* d));
            testdist(end+1,:) = distance; % change number 2
        end
        plotDist = plot(state.Generation,distance/samples,'.');
        set(gca,'xlimmode','manual','zlimmode','manual', ...
            'alimmode','manual')
        set(gca,'xlim',[1,options.MaxGenerations]);
        set(plotDist,'Tag','gaplotdistance');
        xlabel('Generation','interp','none');
        ylabel('Average Distance');
        title('Average Distance Between Individuals','interp','none')
    case 'iter'
        population = state.Population;
        distance = 0;
        for i = 1:samples
            d = population(choices(i,1),:) - population(choices(i,2),:);
            distance = distance + sqrt( sum ( d.* d));
            testdist(end+1,:) = distance; % change number 3
            assignin('base', 'testdist', testdist) % it might be better to assign it and save it from the main file
%             save('testdist.mat', 'testdist')     % as it will save the file 3000+ times
        end
        plotDist = findobj(get(gca,'Children'),'Tag','gaplotdistance');
        newX = [get(plotDist,'Xdata') state.Generation];
        newY = [get(plotDist,'Ydata') distance/samples];
        set(plotDist,'Xdata',newX,'Ydata',newY);
end

Filtering out each ga run is done by lines below

A1 = testdis(:,1)/20;
A2 = zeros(1281,3);
A2(1:1281,1)=A1(1:1281,1);
A2(2:1081,2)=A1(1282:2361,1);
A2(2:1221,3)=A1(2362:3581,1);
A3 = A2(41:20:1221,3);
A4 = A2(41:20:1081,2);
A5 = A2(41:20:1281,1);