Application Of Based On The Improved BP Algorithm In The Identification Of High-energy Particles

High-energy physics is an important branch of physics discipline. Its main goal is to study the structures and properties of the substances in the microscopic world which is smaller than nuclei, to study the phenomenon of the transformation of these substances into each other at very high energies, and to explore the reason and the physics laws behind these phenomena. High-energy physics is based on experiments which generally need to record and process large amounts of data used to describe the system characteristics and operation status. These descriptions contain a large number of indicator variables of the system features with a huge number of samples and a certain degree of randomness, where the sample data forms a huge complex data ocean. The complex data processing became a great difficulty in high energy physics experiment. So, in recent years, the neural network method is used in the field of high energy physics to deal with the problem.There are two aspects of neural network applications in high energy physics. One is the on-site processing and the other is the off-site processing. For on-site processing, neural network can be integrated into the processing chips in the detector, which can greatly improve the detection rate benefiting from a large number of the small parallel processing units. These chips have been used in many high-energy physics experiments. Neural network is also widely used in off-site data processing due to its unique ability of parallel computing.Artificial neural network is a kind of intelligent system, which is capable of distributed information storage and parallel processing by simulating network structure, function and interconnection features of the human brain, It has unique benefits of robustness, self-learning, self-organizing and adaptability. It has very strong classification ability and the ability of approximating arbitrary complex nonlinear function at any precision. It has important significance in making up for the inadequacy of existing science and technology, and in dealing with the high complexity of nonlinear problems. Artificial neural network has been applied in many engineering fields, such as image recognition, information processing, economic forecasting, and system modeling. Since1988, Artificial neural network methods have been introduced in high energy physics experiments and have been widely applied to the identification of the quark-gluon injection, electronic hadron, top quark, and the Higgs particle searching, and has produced satisfying results.At present, the more mature neural network in the practical applications model is Back Propagation (BP) algorithm. BP neural network has good learning ability and generalization ability, which has been applied to the fields of nonlinear modeling, function approximation and pattern classification, etc. However, step size of the traditional BP algorithm in the process of learning is generally determined artificially, and is usually fixed. This is a big shortcoming, the reason is that the total error E is a complex and high nonlinear function, and it is difficult to obtain optimal step size using traditional optimization techniques and analytic methods. If learning step size is too big, the learning optimization process may produce oscillation or even divergence; if learning step size is too small, the learning process will need more iteration, and may fall into local minimum values. So how to determine the optimal step size in the search direction is one of the important problems for a long time. The traditional BP algorithm uses a negative gradient descend algorithm, which has many shortcoming:learning often takes long training time to converge, and it may fall into local minima. This can seriously affect the application of BP neural network in high energy physics.In view of the limitations of traditional BP algorithm, many scholars inland or overseas put forward some improved algorithms, which can be classified into three kinds:one is based on the improvement of numerical calculation, such as Newton’s method and least square method. The second is based on the improvement of gradient method, such as variable learning rate back propagation algorithm and the adaptive learning rate adjustment algorithm. The third is hybrid algorithm, such as dynamic algorithm, fuzzy algorithm, BP algorithm and genetic algorithm, or other hybrid algorithms. However, because most of these algorithms use fixed step size, and may often fail to obtain ideal results.This paper introduces a modified conjugate gradient optimization algorithm. The algorithm uses the conjugate direction as the search direction, and uses quadratic interpolation method to determine the optimal step size to improve the traditional BP algorithm. The algorithm is applied to the identification of high-energy particles. In the application, our algorithm can obtain optimal step size in the search direction for minimizing the objective function, and can overcome the local vibration problem. So, the fast convergence of the objective function is obtained and the stability of the algorithm is improved. Experiments show that our new BP neural network algorithm can effectively improve the identification of particles in high energy physics.In addition, our method is also applied to classification of the Iris data set and Seeds data set to further illustrate the effectiveness and applicability of the conjugate gradient optimization algorithm. The two data sets are two standard test data sets in UCI database, and have been widely used in the study of the classification and pattern recognition. Experiments on these data sets also show satisfactory results.