Double-layer evolutionary algorithm for distributed optimization of particle detection on the Grid

24th IUPAP Conference on Computational Physics

物理学;计算机科学

Padée, Adam^1,2,3 ; Kurek, Krzysztof^4,5 ; Zaremba, Krzysztof^1

Warsaw University of Technology, Institute of Radioelectronics, ul. Nowowiejska 15/19, 00-665 Warsaw, Poland^1

National Centre for Nuclear Research, S´wierk Computing Centre, ul. Andrzeja Soltana 7, 05-400 Otwock-S´wierk, Poland^2

University of Warsaw, Interdisciplinary Centre for Mathematical and Computational Modeling (ICM), ul. Pawin´skiego 5a, 02-106 Warsaw, Poland^3

National Centre for Nuclear Research, Theoretical Physics Department, ul. Andrzeja Soltana 7, 05-400 Otwock-S´wierk, Poland^4

Cardinal Stefan Wyszyn´ski University in Warsaw, Faculty of Mathematics and Natural Sciences, School of Exact Sciences, ul. Wóycickiego 1/3, 01-938 Warsaw, Poland^5

关键词: Automatic approaches;    Binary representations;    Distributed architecture;    Distributed optimization;    Numerical parameters;    Position-Sensitive Detectors;    Reconstruction efficiency;    Track reconstruction;   

Others  :  https://iopscience.iop.org/article/10.1088/1742-6596/454/1/012084/pdf
DOI  :  10.1088/1742-6596/454/1/012084

学科分类：计算机科学（综合）

来源: IOP

【 摘 要 】

Reconstruction of particle tracks from information collected by position-sensitive detectors is an important procedure in HEP experiments. It is usually controlled by a set of numerical parameters which have to be manually optimized. This paper proposes an automatic approach to this task by utilizing evolutionary algorithm (EA) operating on both real-valued and binary representations. Because of computational complexity of the task a special distributed architecture of the algorithm is proposed, designed to be run in grid environment. It is two-level hierarchical hybrid utilizing asynchronous master-slave EA on the level of clusters and island model EA on the level of the grid. The technical aspects of usage of production grid infrastructure are covered, including communication protocols on both levels. The paper deals also with the problem of heterogeneity of the resources, presenting efficiency tests on a benchmark function. These tests confirm that even relatively small islands (clusters) can be beneficial to the optimization process when connected to the larger ones. Finally a real-life usage example is presented, which is an optimization of track reconstruction in Large Angle Spectrometer of NA-58 COMPASS experiment held at CERN, using a sample of Monte Carlo simulated data. The overall reconstruction efficiency gain, achieved by the proposed method, is more than 4%, compared to the manually optimized parameters.