Application of machine learning (ML) algorithms in high-energy physics is evolving rapidly. In particular, they could be used for the optimization of track selection criteria in the analysis of experimental data on hadronic collisions. Using Monte Carlo simulations, one can train ML classifiers to separate correctly reconstructed primary tracks from secondary and fake tracks based on their features such as a number of clusters in TPCs, distance of closest approach to an interaction vertex etc. In this paper, we present the procedure of track selection optimization based on ML techniques and applied to EPOS1.99 simulations of proton-proton interactions obtained via Shine Offline Framework. With this approach, an increase of a fraction of the selected primary tracks and reduced contamination by the secondary tracks is obtained. In case of a complex geometry of an experimental facility like NA61/SHINE, improvement of track selection leads also to a widening of the kinematical acceptance.