Non-greedy adaptive vector quantizers

Abstract

Kohonen’s Learning Vector Quantization (LVQ) technique easily gets trapped in local minima of the distortion surface, resulting suboptimal vector quantizers. The reason is that the behavior of competitive learning on which the LVQ bases is greedy, that is, it only accepts new solutions which maximally reduce the distortion. In this paper, a new and non-greedy adaptive vector quantization scheme is developed which applied a simulated annealing - a randomized search algorithm to the learning procedure and has the capabilities of hill-climbing and approaching global optima. Therefore, this scheme has the advantage of global optimization over the Kohonen’s LVQ scheme. The adaptation (learning) equations are derived and the design schedule procedure is presented.