Abstract. In this paper we consider a sorting problem from railway optimization called train classification, which is NP-hard in general. We introduce two new variants of an earlier developed 2-approximation as well as a new heuristic for finding feasible classification schedules, i.e. solutions for the train classification problem. We evaluate the four algorithms experimentally using various synthetic and real-world traffic instances and further compare them to an exact IP approach. It turns out that the heuristic matches up to the basic approximation, but both are clearly outperformed by our heuristically improved 2-approximations. Finally, with an average objective value of only 5.4 % above optimal, the best algorithm gets close to the real-world schedules of the IP approach, so we obtain very satisfactory practical schedules extremely quickly.