This paper presents a new evolutionary method for the cluster validation index (CVI), namely eCVI. The proposed method learns CVI from the generated training data set using the genetic programming (GP), and then outputs the optimal number of clusters after taking parameters of a test data set into the learned CVI. Each chromosome encodes a possible CVI as a function of the number of clusters, density measure of clusters, and some random factors. Fitness function evaluating each candidate is defined by the difference between the actual number of clusters from training data set and the number of clusters computed by the current CVI. Because of the adaptive nature of GP, the proposed eCVI is reliable and robust in various types of data sets. Experimental results provide grounds for the dominance of eCVI over several widely-known CVIs.