Evolutionary Algorithm for Particle Trajectories Reconstruction

Significant progress has been made in the derivative-free optimization algorithms over past two decades, causing rise in popularity of those methods. The most prominent area of their application is parameters estimation of complex models during machine learning process. Complex model may take forms of non-linear, non-smooth, non-convex functions but also discontinuities might be present. Therefore, group of traditional derivative based methods are failing in this respect. One of the most promising stochastic methods is the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). This stochastic and derivative-free method, is suitable for optimization problems, where very little assumptions on the nature of the underlying objective function can be made. Fewer assumptions will allow to solve larger set of problems, particularly desirable when dealing with difficult objective functions. Another benefit is better separation between model and optimization algorithm, providing means for flexible, multi-model algorithms. Within this work, feasibility of application and performance of the CMA-ES method for particle trajectory reconstruction were studied. The method was used in order to train a model for each particle trajectory by continuous optimization of its parameters. The result of the studies is a novel method of particle trajectory reconstruction, targeting small and medium experiments as well as detector prototype testing, where high reconstruction efficiency and short development time is of greater importance than cutting edge execution time. Additionally, the same optimization technique has been used for training a multiclass classifier, for the purpose of nuclei identification produced in heavy ion reactions at intermediate energies. Resulting algorithm is the first fully automated software tool, which can be used for identification in other nuclear physics experiments as well. In summary, the objective of this work was development of two algorithms based on stochastic optimization methods. The first one has been used to find and reconstruct particle trajectories within gaseous detectors. Whereas the second algorithm has been applied to identify nuclear reaction products, registered by telescope detectors.