Entropy constrained clustering algorithm guided by differential evolution

Entropy constrained vector quantization (ECVQ) is a clustering technique [1] that has been successfully used to describe effliciently large amounts of data collected by the NASA Earth Observing System. The manipulation of this algorithm requires the user to set two parameters: the entropy Lagrange multiplier, and the initial guess for the number of clusters. In this work, we describe an integrated solution that uses a differential evolution algorithm to determine these two parameters. By optimizing two objective functions, entropy and distortion, we find that the solution that best describes the data is located at the inflection point in the Pareto front, i.e. at the point where the tradeoff between the two competing objectives does not favor either one(12).