Re: VarProp[i]=FVerySmart ;

Hi Alexandros,

The "VerySmart" doesn't work. It is only there for backward compatibility, but clearly we should send out a warning message. Will do so in the svn trunk.

What happens if you use it should be exactly the same as using: VarProp [i]=NotEnforced (which is the default). It means that TMVA fits two cuts per input variable (min and max), while for the "FSmart" option a single cut is fitted (lower or greater than). This means that you multiply by two the number of parameters without the FSmart option and hence need much (!) more MC samples (or populations in case you use the GA) to converge properly.

Note that the FSmart option should only be employed when really the distributions can be optimally separated by a single cut - as, for example it would be the case for two Gaussians with same width but different biases. If the tails of the signal or background distributions leak to both sides of the other one, you shouldn't use it!

This is explained in the Users Guide (on: http://tmva.sf.net)

Cheers,
Andreas

On 22 Sep 2009, at 17:42 22 Sep 2009, Alexandros Attikis wrote:

> Hi ROOTers,
>
> I am currently using the TMVA Version 4.0.1 (ROOT 5.24/00) and i am
> not sure if the Conguration option reference for MVA method Cuts:
>
> VarProp[i]=FVerySmart;
>
> is currently working properly or not?
>
> The setup of TMVA is fine as everything works great when I use
> VarProp[i]=FSmart, I get a smooth ROC curve. (only 1 efficiency bin
> not fllled)
>
> However, replacing VarProp[i]=FSmart wtih VarProp[i]=FVerySmart for
> all variables completely ruins the curve
> because 99 efficiency bins are basically empty.
> Any suggestions ?
>
> Cheers,
> Alexandros
>
Received on Tue Sep 22 2009 - 18:40:08 CEST