Codebook Design for Vector Quantisation Using Multiobjective Genetic Algorithms

In this paper, we present a novel approach for codebook design using multiobjective genetic optimization. We partition the pattern space into perceptive classes, which are then clustered into hyperspheres based on multiobjective criteria. Multiobjective genetic algorithms perform optimisation on a vector space of objectives and are able to explore the search space for a set of equally viable and equivalent partitions of the pattern space. One of the objectives searches for compact solutions within a class, which in turn yields the least MSE distortion along with better perceptive quality.