Logo Logo
Hilfe
Kontakt
Switch language to English
Statistical Learning in High Energy and Astrophysics
Statistical Learning in High Energy and Astrophysics
This thesis studies the performance of statistical learning methods in high energy and astrophysics where they have become a standard tool in physics analysis. They are used to perform complex classification or regression by intelligent pattern recognition. This kind of artificial intelligence is achieved by the principle ``learning from examples'': The examples describe the relationship between detector events and their classification. The application of statistical learning methods is either motivated by the lack of knowledge about this relationship or by tight time restrictions. In the first case learning from examples is the only possibility since no theory is available which would allow to build an algorithm in the classical way. In the second case a classical algorithm exists but is too slow to cope with the time restrictions. It is therefore replaced by a pattern recognition machine which implements a fast statistical learning method. But even in applications where some kind of classical algorithm had done a good job, statistical learning methods convinced by their remarkable performance. This thesis gives an introduction to statistical learning methods and how they are applied correctly in physics analysis. Their flexibility and high performance will be discussed by showing intriguing results from high energy and astrophysics. These include the development of highly efficient triggers, powerful purification of event samples and exact reconstruction of hidden event parameters. The presented studies also show typical problems in the application of statistical learning methods. They should be only second choice in all cases where an algorithm based on prior knowledge exists. Some examples in physics analyses are found where these methods are not used in the right way leading either to wrong predictions or bad performance. Physicists also often hesitate to profit from these methods because they fear that statistical learning methods cannot be controlled in a physically correct way. Besides there are many different statistical learning methods to choose from and all the different methods have their advantages and disadvantages -- compared to each other and to classical algorithms. By discussing several examples from high energy and astrophysics experiments the principles, advantages and weaknesses of all popular statistical learning methods will be analysed. A focus will be put on neural networks as they form some kind of standard among different learning methods in physics analysis.
statistics neural networks support vector machines random forest H1 MAGIC XEUS
Zimmermann, Jens
2005
Englisch
Universitätsbibliothek der Ludwig-Maximilians-Universität München
Zimmermann, Jens (2005): Statistical Learning in High Energy and Astrophysics. Dissertation, LMU München: Fakultät für Physik
[thumbnail of Zimmermann_Jens.pdf]
Vorschau
PDF
Zimmermann_Jens.pdf

9MB

Abstract

This thesis studies the performance of statistical learning methods in high energy and astrophysics where they have become a standard tool in physics analysis. They are used to perform complex classification or regression by intelligent pattern recognition. This kind of artificial intelligence is achieved by the principle ``learning from examples'': The examples describe the relationship between detector events and their classification. The application of statistical learning methods is either motivated by the lack of knowledge about this relationship or by tight time restrictions. In the first case learning from examples is the only possibility since no theory is available which would allow to build an algorithm in the classical way. In the second case a classical algorithm exists but is too slow to cope with the time restrictions. It is therefore replaced by a pattern recognition machine which implements a fast statistical learning method. But even in applications where some kind of classical algorithm had done a good job, statistical learning methods convinced by their remarkable performance. This thesis gives an introduction to statistical learning methods and how they are applied correctly in physics analysis. Their flexibility and high performance will be discussed by showing intriguing results from high energy and astrophysics. These include the development of highly efficient triggers, powerful purification of event samples and exact reconstruction of hidden event parameters. The presented studies also show typical problems in the application of statistical learning methods. They should be only second choice in all cases where an algorithm based on prior knowledge exists. Some examples in physics analyses are found where these methods are not used in the right way leading either to wrong predictions or bad performance. Physicists also often hesitate to profit from these methods because they fear that statistical learning methods cannot be controlled in a physically correct way. Besides there are many different statistical learning methods to choose from and all the different methods have their advantages and disadvantages -- compared to each other and to classical algorithms. By discussing several examples from high energy and astrophysics experiments the principles, advantages and weaknesses of all popular statistical learning methods will be analysed. A focus will be put on neural networks as they form some kind of standard among different learning methods in physics analysis.