Re: [TMVA-users] beginner question: regression, parameter estimation

[Margar S. <Mar...@ce...>]

Hi Helge,

after thinking a bit more I probably would need redfinition of error
function and 2 step back-propagation. It looks beyond the current
functionality, but something like these steps:

o) take K events which have the same target value and apply online learning
o) calculate average of output from the previous step
o) take the same K events, apply online learning again with redefined error
function such that the average output from the previous step is closest to
the target

Don't know if this is clear and probably will require huge number of
changes in TMVA MLP.

Best regards,
      Margar

On Wed, Nov 28, 2012 at 3:07 PM, Helge Voss <Hel...@ce...> wrote:

> Hi Margar,
>
>
>   // BatchSize       <int>        Batch size: number of events/batch,
> only set if in Batch Mode,
>    //                                          -1 for
> BatchSize=number_of_events
>
> isn't that exactly what you want ?
>
> cheers,
>
> Helge
>
> y
> 2012/11/27 Margar Simonyan <Mar...@ce...>:
> > Hi Helge,
> >
> > thank you for quick reply. So BPMode=batch is the bulk learning
> described in
> > the user guide. BatchSize is fixed, it would be nice to be able to chain
> > TTrees as an input where BatchSize is determined from each Tree.
> >
> > Best regards,
> >           Margar
> >
> >
> > On Tue, Nov 27, 2012 at 10:33 PM, Helge Voss <Hel...@ce...> wrote:
> >>
> >> Hi Margar,
> >>
> >> well. BPMode=batch  is for batch-learning and "sequential" for online
> >> learning
> >> and if you choose "batch" then you can define the "size" of the sample
> at
> >> each
> >> weight update step... given by BatchSize ..
> >>
> >> cheers,
> >>
> >> Helge
> >>
> >>
> >> 2012/11/27 Margar Simonyan <Mar...@ce...>:
> >> > Dear TMVA Experts,
> >> >
> >> > I have no prior experience with TMVA and before starting I would like
> to
> >> > ask
> >> > some basic questions.
> >> >
> >> > The idea is to use regression to reduce the dimensionality of the task
> >> > and
> >> > do parameter estimation. For simplicity let's assume N-dimensional
> >> > Gaussian
> >> > model and one is interested in mean value of Gaussian's in 2
> dimensions
> >> > (<x1>, <x2>). Suppose I have samples from the model using different
> <x1>
> >> > and
> >> > <x2> values for training of a single neural network. I read in the
> user
> >> > guide that online learning is implemented, but bulk learning is
> >> > explained as
> >> > well. If one event is used (online) it is nearly impossible to predict
> >> > <x1>
> >> > and <x2>, while with say 100 events it is easier. So it seams the bulk
> >> > learning is required for this particular problem.
> >> >
> >> > Is there any way to solve this problem in current TMVA? From the guide
> >> > it is
> >> > not clear to me the purpose of BatchSize and BPMode=batch options.
> >> >
> >> > Another approach would be to build a network with 100N input nodes
> >> > instead
> >> > of N, but I suppose this requires huge amount of training samples.
> >> >
> >> > Thanks,
> >> >          Margar
> >> >
> >> > P.S. my RooFit based approach is not working very well either with my
> >> > current PDF.
> >> >
> >> >
> >> >
> -------------------------------------------------------------------------
> >> >  Dr Margar Simonyan,  post-doctoral researcher
> >> >  Niels Bohr Institute, Copenhagen University
> >> >
> >> >
> -------------------------------------------------------------------------
> >> >
> >> >
> >> >
> ------------------------------------------------------------------------------
> >> > Monitor your physical, virtual and cloud infrastructure from a single
> >> > web console. Get in-depth insight into apps, servers, databases,
> vmware,
> >> > SAP, cloud infrastructure, etc. Download 30-day Free Trial.
> >> > Pricing starts from $795 for 25 servers or applications!
> >> > http://p.sf.net/sfu/zoho_dev2dev_nov
> >> > _______________________________________________
> >> > TMVA-users mailing list
> >> > TMV...@li...
> >> > https://lists.sourceforge.net/lists/listinfo/tmva-users
> >> >
> >
> >
>