Multiobjective genetic algorithm partitioning for hierarchical learning of high-dimensional pattern spaces: a learning-follows-decomposition strategy

In this paper, we present a novel approach to partitioning pattern spaces using a multiobjective genetic algorithm for identifying (near-)optimal subspaces for hierarchical learning. Our approach of "learning-follows-decomposition" is a generic solution to complex high-dimensional problems where the input space is partitioned prior to the hierarchical neural domain instead of by competitive learning. In this technique, clusters are generated on the basis of fitness of purpose--that is, they are explicitly optimized for their subsequent mapping onto the hierarchical classifier. Results of partitioning pattern spaces are presented. This strategy of preprocessing the data and explicitly optimizing the partitions for subsequent mapping onto a hierarchical classifier is found both to reduce the learning complexity and the classification time with no degradation in overall classification error rate. The classification performance of various algorithms is compared and it is suggested that the neural modules are superior for learning the localized decision surfaces of such partitions and offer better generalization.