How to optimize the variables of a regression model

Hello,

I have created a regression model using Statistics and Machine Learning Toolbox that is dependent on 6 variables. After exporting the model to the workspace, I now want to determine the global minimum of the function using "surrogateopt".

With my code, the solver varies the attributes in the specified interval, but the result of the output "fval" does not change.

Can anyone help me with this problem?

objconstr = @(x)trainedModel.predictFcn(data);

lb = [-10,-10,-10,-10,-10,-10];

ub = [10,10,10,10,10,10];

[x,fval] = surrogateopt(objconstr,lb,ub);

trainedModel.predictFcn is the name of the function

data is a table with the 6 attributes

Best

Simon

Answer 1

[Alan Weiss]

I think that you need to connect your data argument to your x argument. Something like this:

function f = objconstr(x,trainedModel)
% Convert x to a table of the correct type.
% For example,
mytable = array2table(x,'VariableNames',...
   ["name1" "name2" "name3" "name4" "name5" "name6"); % Put in the appropriate names
y = trainedModel.predictFcn(mytable);
end

Then call surrogateopt on the objective function @(x)objconstr(x,trainedModel).

Alan Weiss

MATLAB mathematical toolbox documentation

[Simon Koch]

Thank you very much for your answer.

I have written the function in a separate script (objconstr.m).

function f = objconstr(x,trainedModel)
data = array2table(x,'VariableNames',["L1","L2","L3","R1","R2","R3"]);
f = trainedModel.predictFcn(data);
end

Now when I call surrogatopt in my original script (optimisation.m), an error message appears.

lb = [-1,-1,-1,-1,-1];
ub = [1,1,1,1,1];
[x,fval] = surrogateopt(@(x)objconstr(x,trainedModel),lb,ub);

Unable to find function @(x)trainedModel.predictFcn(TestData_new) within C:\Users\PCUser\Documents\MATLAB\optimisation.m.

Simon Koch

[Alan Weiss]

Please perform the following experiment. Get trainedModel into your workspace and make sure that objconstr is on your MATLAB path. Then define fun = @(x)objconstr(x,trainedModel). Then take

x = zeros(1,6); % or any other feasible point
val = fun(x)

See whether this throws the same error. If so, then you have to figure out why before trying to optimize fun. If not, then I think that you are good to go, using fun inside surrogateopt.

Alan Weiss

MATLAB mathematical toolbox documentation