Hybrid GBAR/Nonlinear Time-Series Method for Generation of Synthetic VBR Video Traffic

Linear analysis of a data set examines any evident structure in the data through linear correlations and implicitly assumes that any dynamics of the system are modeled linearly, where small perturbations of initial conditions lead to small changes in the unfolding dynamics. Linear analysis attributes all irregular behavior of the system to stochastic external excitation of the system. However, stochastic excitation of linear equations is not the only source of irregularity in the output of a system. For nonlinear chaotic systems, irregular outputs can be produced from deterministic equations of motion in an autonomous manner; that is with time-independent inputs. We use methods from nonlinear time series to generate synthetic VBR video data. We propose a hybrid algorithm which combines methods from nonlinear time-series analysis to produce downsampled VBR data at the relevant time scale and interpolates between this downsampled data by using the GBAR model for synthetic VBR video data. We contrast this hybrid method with the purely stochastic GBAR method and with a purely deterministic nonlinear time-series method, using both H.263 and MPEG4 data. We compare pdfs and autocorrelation functions of the synthetic and true data. For MPEG4 data the characterisitc autocorrelation structure, present due to GOP coding, is reproduced well by the nonlinear deterministic algorithm. Such synthetic VBR video data are useful in many different aspects of performance evaluation and resource allocation for networks.