An EM-Algorithm for MAP Fitting from Real Traffic Data

Abstract

For model based analysis of computer and telecommunication systems an appropriate representation of arrival and service processes is very important. Especially representations that can be used in analytical or numerical solution approaches like phase type (PH) distributions or Markovian arrival processes (MAPs) are useful. This paper presents an algorithm to fit the parameters of a MAP according to measured data. The proposed algorithm is of the expectation-maximization (EM-) type and extends known approaches for the parameter fitting of PH-distributions and hidden Markov chains. It is shown that the algorithm generates MAPs which approximate traces very well and especially capture the autocorrelation in the trace. Furthermore the approach can be combined with other more efficient but less accurate fitting techniques by computing initial MAPs with those techniques and improving these MAPs with the approach presented in this paper.