back to the main page

Algorithm description can be found in this paper (except the last network growth improvement - split neurons):
Title: Application of the neural networks in events classification in the measurement of spin structure of the deuteron
R. Sulej ¹⁾, K. Zaremba ¹⁾, K. Kurek ²⁾ and E. Rondio ^2)
1) Institute of Radioelectronics, Warsaw University of Technology, Warsaw, Poland
²⁾ Institute for Nuclear Studies, Warsaw, Poland

Abstract:
In this paper, we present the application of a neural network for events classification in a high-energy physics experiment. As a network model we use a multi-layer perceptron with a dynamic topology adjustment algorithm. Our solution covers both adding new hidden neuron units and removing unnecessary units. Neural network results are compared to the standard kinematical cuts technique and to the well-known k-nearest neighbor (knn) classifier.

Full-text is published in Meas. Sci. Technol., Vol. 18 (2007), pp. 2486-2490 (abstract). Full text (but not in the MST paper format) is also available here.

Application of the algorithm in the neutrino interactions classification:
Title: Polarization effects in tau production by neutrinos
J. Lagoda, D. Kielczewska, M. Posiadala, R. Sulej, K. Zaremba, T. Kozlowski, K. Kurek, P. Mijakowski, P. Przewlocki, E. Rondio, J. Stepaniak, M. Szeptycka

Abstract:
A direct proof of the existence of n_m → n_t oscillations is important. This proof can be obtained by an observation of the production of taons in charge current reactions n_t + N → t + X. The influence of t polarization on the characteristics of the CC events and on the efficiency of their selection is discussed. The neural network method is used to select t leptons produced in n_t interactions.

Full-text is published in Acta Physica Polonica B, Vol. 38, No. 6 (2007), pp. 2083-2103 (free access to online paper here).

Results of DG/G analysis presented at SPIN2006, Kyoto (neural network trained with dynamic structure algorithm was used in a_LL parametrization):
Title: Measurements of Delta G/G
G. K. Mallot

Abstract:
Our present information on the gluon polarisation DG/G is reviewed. The data from fixed-target lepton-nucleon experiments are in context with the recent data from the RHIC polarised pp collider. The main tools to study DG/G in lepton-nucleon scattering are scaling violations of the g₁ structure functions and longitudinal spin asymmetries in hadron production. Results from high-p_T hadron pairs, inclusive hadrons as well as open-charm production are discussed. At RHIC the most precise data presently came from inclusive p⁰ and jet production. All data indicate that the gluon polarisation is small compared to earlier expectations, but still can make a major contribution to the nucleon spin.

Full-text: Proceedings for SPIN2006, Kyoto or arXiv:hep-ex/0612055v1.

Algorithm has been changed significantly since this paper was published. Anyway, these were our first attempts with dynamic network structure...
Title: Dynamic topology adjustment algorithm for MLP networks
R. Sulej ¹⁾, K. Zaremba ¹⁾, and K. Kurek ^2)
1) Institute of Radioelectronics, Warsaw University of Technology, Warsaw, Poland
²⁾ Institute for Nuclear Studies, Warsaw, Poland

Abstract:
In this paper we present new algorithm for network topology adjustment during the training process. As a network model we use multi-layer perceptron (MLP) trained with various back-propagation techniques. Our solution covers both adding new hidden neuron units and removing unnecessary units. We present the test results on a basic tasks to show some characteristics of our algorithm and compare it with other well known model - Cascade-Correlation network. Also we give a brief view of applications of our algorithm in high energy physics classification and approximation tasks.

Full-text is published in ICAISC proceedings: "Artificial Intelligence and Soft Computing", Academic Publishing House EXIT, Polish Neural Network Society, Academy of Humanities and Economics in Łódź, IEEE Computational Intelligence Society - Poland Chapter, 2006.

Conference poster in pdf format (1.84MB)

Application of the fixed-structure network:
Title: Spin asymmetries for events with high p_T hadrons in DIS and an evaluation of the gluon polarization
The SMC Collaboration

Abstract:
We present a measurement of the longitudinal spin cross section asymmetry for deep-inelastic muon-nucleon interactions with two high transverse momentum hadrons in the final state. Two methods of event classification are used to increase the contribution of the photon-gluon fusion process to above 30%. The most effective one, based on a neural network approach, provides the asymmetries A_p^lN→lhhX = 0.030 ± 0.057(stat) ± 0.010(syst) and A_d^lN→lhhX = 0.070 ± 0.076(stat) ± 0.010(syst). From these values we derive an averaged gluon polarization DG/G = –0.20 ± 0.28(stat) ± 0.10(syst) at an average fraction of nucleon momentum carried by gluons <h> = 0.07.

Full-text is published in Physical Review D, 70:012002 (2004).

And one more old paper:
Title: Selection of Photon Gluon Fusion Events in DIS
K. Kowalik, E. Rondio, R. Sulej, K. Zaremba

Abstract:
A selection of the Photon Gluon Fusion (PGF) process with light quarks for deep inelastic scattering events is presented. This process is directly sensitive to gluon polarization and our goal is to find out the most effective selection on a sample of events simulated for the SMC experiment. We compare two general multi-class classification methods - Bayes method and neural network with a conventional selection procedure. The neural network algorithm presented here is a modification of method belonging to the family of directional minimization algorithms. This method is convenient and effective for photon gluon fusion selection and determination of gluon polarization. Finally we present the estimation for precision of gluon polarization for neural network method.

Full-text is published in Acta Physica Polonica B, Vol. 32, No. 10 (2001), page 2929.

back to the main page