In this semester I modeled computer network’s traffic with phase type distributions. I fitted these distributions to data traces from real measurements. The aim was to minimize the differences between the statistics from the trace and the fittings.
My first task was studying the phase type distributions. Then I studied the existing phase type fitting methods. Each one has its weaknesses and strengths, for example, one is better in fitting the head part of the trace, the other is better at the tail part, or better at momentum matching, or has higher likelihood parameter. These are introduced shortly in my thesis. Finally my task was to introduce a new method, which combine the strengths of former fittings, and produce better statistics. In the new method I combine a moment matching algorithm with the EM algorithm, which maximizes the likelihood parameter. In addition to get better initial values I used a clustering algorithm, the k-means algorithm. With this new fitting method I get better results in almost every statistical parameter.