Optimization of tau identification in ATLAS experiment using multivariate tools

In elementary particle physics the efficient analysis of huge amount of collected data require the use of sophisticated selection and analysis algorithms. We have implemented a Support Vector Machine (SVM) integrated with the CERN TMVA/ROOT package. SVM approach to signal and background separation is based on building a separating hyperplane defined by the support vectors. The margin between them and the hyperplane is maximized. The extensions to a non-linear separation is performed by mapping the input vectors into a high dimensional space, in which data can be linearly separated. The use of kernel functions allows to perform computations in a high dimension feature space without explicitly knowing a mapping function. Our SVM implementation is based on Platt's Sequential Minimal Optimization (SMO) algorithm and includes various kernel functions like a linear function, polynomial and Gaussian. The identification of hadronic decays of tau leptons in the ATLAS experiment using a tau1P3P package is performed using, beside the baseline cut analysis, also multivariate analysis tools: neural network, PDE_RS and our implementation of the SVM algorithm. The use and the comparison of the three algorithms is presented.