Personal tools
Document Actions

TMVA

by kaivoss last modified 2006-06-27 11:11

C++

Major Toolset

Categorization

TMVA is a ROOT based toolkit for parallel multivariate data analysis. In factory mode TMVA performs the training, testing and performance evaluation of various implemented discrimination algorithms. Homepage: http://tmva.sourceforge.net

Andreas Höcker (CERN), Helge Voss (MPI-KP Heidelberg), Kai Voss (U. of Victoria), Xavier Prudent (LAPP-Annecy)

The Toolkit for Multivariate Analysis (TMVA) provides a ROOT-integrated environment for the parallel processing and evaluation of MVA techniques to discriminate signal from background samples. It presently includes (ranked by complexity):
  • Rectangular cut optimisation
  • Correlated likelihood estimator (PDE approach)
  • Multi-dimensional likelihood estimator (PDE - range-search approach)
  • Fisher (and Mahalanobis) discriminant
  • H-Matrix (chi-squared) estimator
  • Artificial Neural Network (two different implementations)
  • Boosted Decision Trees
The TMVA package includes an implementation for each of these discrimination techniques, their training and testing (performance evaluation). In addition all these methods can be tested in parallel, and hence their performance on a particular data set may easily be compared. For the comparison of the efficiency and background rejection of all the different methods, the analysis job gives some tabulated efficiency / background values in the log file, as well as the efficiencye vs. background rejection curves along with detailed other evaluation information (like ranking, correlation matrixes and alike) in a ROOT file. These results can then be easily displayed using the ROOT macros provided with TMVA.

The C++ code runs alternatively as a standalone executable, or as a root script, where libTMVA.so is linked as a shared library. Each method that is trained in this application writes the training results in individual "weight" files, which are either text or ROOT files. Two single compact C++ classes (reader/TMV_Reader and reader/TMVA_ReaderTools) are provided, which interpret the training files, and which can be included in any C++ or ROOT-based analysis job.

Special emphasis has been put on the comparative performance assessment of the various MVA methods. The training, testing and evaluation phases are performed in parallel for the various methods (through a factory). The evaluation accomodates several numerical performance estimators as well as various ROOT plots. ROOT scripts to conveniently access these plots are provided.


kai.voss@cern.ch
tmva-v1-0-tar.gz by Site Administrator — last modified 2006-05-30 15:55
 

Powered by Plone, the Open Source Content Management System