How to optimize hyperparameters for fitcensemble using a customized maximize function?

Question

Yean Lim 2020-11-15

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/648823-how-to-optimize-hyperparameters-for-fitcensemble-using-a-customized-maximize-function

回答： Alan Weiss 2020-11-17

Hi, I have been using fitcensemble method to minimize objective function by default, as I can get the minimum objective result from mdl, which is a ClassificationEnsemble

mdl = fitcensemble(XTrain,yTrain,...
    'OptimizeHyperparameters',auto,... 
    'Learners',template, ...
    'HyperparameterOptimizationOptions',hyperoptsOptions,...
    'ScoreTransform','logit',... 
    'CategoricalPredictors', isCategoricalPredictor,...
    'PredictorNames', predictorNames);
minObj = mdl.HyperparameterOptimizationResults.MinObjective;

Is it possible to get a customized maximize function from fitcensemble function?. If yes, could you provide me an example please?

Thank you

2 个评论
显示无隐藏无

Alan Weiss 2020-11-15

I do not understand what you are trying to do. fitcensemble tries to fit a classification ensemble by minimizing an error. What sense would it make to maximize an error?

Alan Weiss

MATLAB mathematical toolbox documentation

Yean Lim 2020-11-15

Hi I want to maximize a customised accuracy definition from the model given by the fitcensemble

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Alan Weiss 2020-11-17

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/648823-how-to-optimize-hyperparameters-for-fitcensemble-using-a-customized-maximize-function#answer_547118

I am not sure, but maybe you can use the fitcensemble 'Cost' name-value argument. Give the negative of your customized accuracy measure in order that fitcensemble will look for a maximum. Or maybe give a constant minus the customized accuracy where you choose the constant so that the cost will always be positive, but not too large so that the differences still matter.

Alan Weiss

MATLAB mathematical toolbox documentation