Fine-grained end-to-end network model via vector quantization and hidden Markov processes

We study and compare modeling an end-to-end network by conventional, bivariate, and exponential observation hidden Markov processes. Furthermore, effects of μ-law, Lindle-Boyde-Gray, and uniform quantization approaches on the modeling granularity is explored. We performed experiments using synthetic representative data from a traffic-modeler autoregressive modular process and the Network Simulator software as well as over-the-Internet experiments with real data to contrast the fidelity produced from each model. Comparing statistical signatures of the model-generated data with those of the training sequence indicates that accompanying Lindle-Boyde-Gray quantization with conventional or bivariate hidden Markov processes significantly improves the modeling fidelity.