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 spa ces using a multiobjective genetic algorithm for identifying (near-)op timal subspaces for hierarchical learning. Our approach of ''learning- follows-decomposition'' is a generic solution to complex high-dimensio nal problems where the input space is partitioned prior to the hierarc hical neural domain instead of by competitive learning, In this techni que, 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 pr esented, This strategy of preprocessing the data and explicitly optimi zing the partitions for subsequent mapping onto a hierarchical classif ier is found both to reduce the learning complexity and the classifica tion time with no degradation in overall classification error rate. Th e classification performance of various algorithms is compared and it is suggested that the neural modules are superior for learning the loc alized decision surfaces of such partitions and offer better generaliz ation.