An approach to constructing discrete models of packet losses suitable for a wide variety of communication network applications is studied. It is based on estimating parameters of probabilistic automata described via so-called pseudo-Markov chains. The new technique is applied both to approximating a discrete time analog process at the output of known channel models and to the experimental data stream. Comparison of models is performed by computing probabilities of more than m losses out of n transmitted packets (P (≥ m, n)). It is shown that for the Rician fading channel with exponential correlation and correlation determined by a Bessel filter, the obtained rank-two and rank-three discrete modes, respectively, provide high accuracy coincidence of P (≥ m, n) performances. The rank-three discrete model computed on the experimental data stream obtained from the LTE network provides significantly better approximation of P (≥ m, n) performance than that obtained by the Baum-Welch algorithm.