Adaptive vector quantization .II. Classification and comparison of algorithms

Summary form only given. For pt.I see ibid., p.436, 1997. We review prominent examples of adaptive vector quantization (AVQ) algorithms from prior literature and develop a classification of these algorithms. Well known theorems from rate-distortion theory suggest two approaches to the nonadaptive vector quantization (VQ) of a stationary, ergodic random process. These two nonadaptive VQ approaches have, in turn, inspired two general types of AVQ algorithms for the coding of nonstationary sources. In constrained-distortion AVQ algorithms, the algorithm limits the distortion to some maximum value and then attempts to minimize the rate subject to this distortion constraint. Constrained-rate AVQ algorithms do the opposite, limiting the rate to be less than or equal to some maximum value and attempting to produce a coding with the smallest distortion. A third category of AVQ algorithms, rate-distortion-based algorithms minimize the rate-distortion cost functions. We discuss each of the three categories of AVQ algorithms in detail and mention notable algorithms found in each category. Afterwards, we summarize the discussion with an algorithm taxonomy. Finally, we present experimental results for several prominent AVQ algorithms on an artificial nonstationary random process. Our results suggest that, one, the class of rate-distortion-based algorithms is capable of coding performance superior than that of other algorithms, particularly for low-rate coding, and, two, that complex, batch coding algorithms are not as competitive as simpler, online algorithms.