Why almost the same optimization function gives different results?

Hello,

I am trying to optimize ECOC classifier as follows:

%data

clear all

load fisheriris

X = meas;Y = species;

rng default

t_gaussian=templateSVM('KernelFunction','gaussian','standardize',true)

Mdl_gaussian = fitcecoc(X,Y,'Coding','onevsall','Learners',t_gaussian,'OptimizeHyperparameters','auto',...

'HyperparameterOptimizationOptions',struct('CVPartition',CVO,'Optimizer','bayesopt','AcquisitionFunctionName',...

'expected-improvement-plus'))

I am wondering why I did not find the same results if I remplace 'OptimizeHyperparameters','auto' with 'OptimizeHyperparameters',{'BoxConstraint','KernelScale'}

rng default

Mdl_g = fitcecoc(X,Y,'Coding','onevsall','Learners',t_gaussian,'OptimizeHyperparameters',{'BoxConstraint','KernelScale'},...

'HyperparameterOptimizationOptions',struct('CVPartition',CVO,'Optimizer','bayesopt','AcquisitionFunctionName',...

'expected-improvement-plus'))

Best regards

Answer 1

[Alan Weiss]

I am not 100% sure, but my reading of the fitcecoc documentation shows that 'auto' has this description:

'auto' — Use {'Coding'} along with the default parameters for the specified Learners:

• Learners = 'svm' (default) — {'BoxConstraint','KernelScale'}

So I think that 'auto' is equivalent to {'Coding','BoxConstraint','KernelScale'}.

Alan Weiss

MATLAB mathematical toolbox documentation

[Nadou]

Hello Alan,

Thank you for your response

This is what I thought also while reading fitcecoc documentation. However, I found different results

Best regards