Automated test modeling for VBR video

TES (Transform-Expand-Sample) is a versatile class of stationary stochastic processes which can model arbitrary marginals, a wide variety of autocorrelation functions (e.g., monotone, oscillatory and others), and broad range of sample path behaviors (e.g., directional and reversible). TES model specification consists of two sets. of parameters. The first set, which is algorithmically determined, guarantees an exact match to the empirical distribution (histogram). The remaining set largely determines the autocorTelation structure. In order to approximate the empirical autocorrelation function, the M modeling methodology to-date employs a heuristic search approach on a large parametric apace. An interactive visual modeling environment, called TEStool, was designed and implemented to support heuristic searches for TES models under human control. This approach has several drawbacks. First, effective TES modeling requires qualitative understanding of TES processes as well as experience; second, the search scope speed is fundamentally limited by the speed of the human response; and third, modeling precision is constrained by screen resolution as perceived by the human eye. This talk describes a TES modeling algorithm, which largely automates the modeling process, shifting the modeling burden from the human to the computer. The algorithm is cast in nonlinear programming setting with the objective of minimizing a weighted square distance between the empirical autocorrelation function and its candidate TES model counterpart. Steepest-Descent search is feasible due to the availability of analytical formulas for the computation of TES autocorrelation functions and their partial derivatives. Furthermore, the nice local behavior of the objective function brings about good overall performance of the algorithm. Finally, the efficacy of this approach is illustrate via two examples from the domain of VBR (variable bit rate) compressed video.