Session:

EVH - Application of Hybrid Evolutionary Algorithms to Complex Optimization Problems

Title:

On Fitness, Niching Strategies, and Hybrid Niche Size Estimation for Discovering an Unknown Number of Clusters in Noisy Data

Authors:

Olfa Nasraoui
Elizabeth Leon

Abstract:

Clustering is crucial to many applications in pattern recognition, data mining, and machine learning. Evolutionary techniques have been used with success in clustering. However, most existing evolutionary clustering techniques still suffer from several drawbacks. After surveying existing evolutionary clustering techniques, we show that (i) robustness to noise can be achieved with a robust fitness measure, while (ii) scalability of the search space with respect to the number of clusters and to the size of the data can be achieved by encoding a single cluster prototype in the chromosome, (iii) the resulting multimodal optimization problem should be solved using a \emph{niching} strategy, which will also allow the determination of the number of clusters automatically, and finally that (iv) a hybrid Piccard niche size estimation strategy is the key to implement a sound mating restricion, and is the key to successful fitness sharing.