why use negative values for lambda,...

[caprone]

HI why use negative values for lambda, gamma and alpha ????

[caprone]

a ok, i cheked source; autoxgbparset = makeParamSet( makeNumericParam("eta", lower = 0.01, upper = 0.2), makeNumericParam("gamma", lower = -7, upper = 6, trafo = function(x) 2^x), makeIntegerParam("max_depth", lower = 3, upper = 20), makeNumericParam("colsample_bytree", lower = 0.5, upper = 1), makeNumericParam("colsample_bylevel", lower = 0.5, upper = 1), makeNumericParam("lambda", lower = -10, upper = 10, trafo = function(x) 2^x), makeNumericParam("alpha", lower = -10, upper = 10, trafo = function(x) 2^x), makeNumericParam("subsample", lower = 0.5, upper = 1) ) but, when i run the cntrl, seems that 'trafo' doesn't work, then it returned me negative values...for gamma, lambda e alpha..

EDIT:

ISSUE IS HERE:

res$optim.result$x

it return correct 'recommended parameters' but for lambda, gamma and alpha; for these return 'pre' trafo values, then return also negative values

[ja-thomas]

Hi,

these values are tuned on log-scale to make the tuning more efficient. OptPath entries always use the values as seen by the optimizer, hence the untransformed values. The printer shows the correct transformed values and they are also saved in the final.learner object if you want to train the learner on new data (If you want to train it on the same data you can just use the build.final.model argument).

If you directly need the transformed values you can easily extract them from the final.learner with getHyperPars(res$final.learner).

I don't plan to change the internal representation of the optim.result, since it is done created by mlrMBO and not directly by autoxgboost.

Are the untransformed values returned anywhere else by autoxgboost, except for accessig them by res$optim.result$x?

[caprone]

a OK perfect, thanks;

@ja-thomas "Are the untransformed values returned anywhere else by autoxgboost, except for accessig them by res$optim.result$x?"

No, seems not

ps. Great work, very useful

[liuyanguu]

Very helpful. I spent quite some time to find the recommended parameters... Thanks for the discussion.

[ja-thomas]

cool :)