Boundary for target in regression not working

TMVA
1

Dear experts,
I’m using a MLP multi-target regression.
I’ve set a target variable with min and max values with:

dataloader->AddTarget(myvar,"","",xmin,xmax)

However when plotting TestTree->Draw(MLP.myvar) there are values beyond the boundaries.
Best, Marianna.

2

I guess @moneta can help you.

3

Hi,
The min and max values are not used as a contraint to the target. I think you can use as target a transformed variable like arcsin to the target that will avoid that the prediction will go out of the boundary.

Lorenzo

4

HI, I’ve tried to use as target cond := 80 + 50*sin(f) to bound f between 30 and 130. In the TestTree cond is correctly bounded, but not the MLP.cond which shows (attached) cond.pdf (14.2 KB) MLPcond.pdf (13.8 KB) the distribution of f.
Best, Marianna.

5

Hi,

It is not clear to me what the plots are exactly. cond.pdf is the input target transformed as
80+50*sin(f) ?
And MLPcond.pdf is the predicted target by MLP ?

Lorenzo

6

Hi Lorenzo,
yes.
Marianna.

7

OK, but then I don’t understand why you have such distribution for the target cond peaked at 30 and 130.
And what is the problem with the predicted variable ? That is too far off ?

8

Hi, the input variable is expected to peak at 30 and 130, because of the sin function peaking at -+1. My question is why the predicted variable is not bounded between 30 and 130, as the input variable.
Best, Marianna.

9

Sorry for the late reply. There is no constraints in the algorithm that make the prediction to be bound in a given region, so it is expected that the prediction can be outside the range.