Multi-objective genetic algorithm partitioning for hierarchical learning of high dimensional spaces

Complex pattern recognition problems of high dimensionality are best addressed through a ´divide-and-conquer´ approach rather than monolithically. We introduce a novel approach to partitioning the pattern space into hyperspheres using a multiobjective genetic algorithm for subsequent mapping onto a hierarchical neural network for subspace learning. In our technique clusters are generated on the basis of ´fitness for purpose´-they are explicitly optimised for their subsequent mapping onto the hierarchical classifier-rather than emerging as some implicit property of the clustering algorithm. Multi-objective genetic algorithms perform optimisation on a vector space of objectives and are able to explore the NP-complete search space for a set of equally viable partitions of the pattern space. The rationale behind this strategy is set-out and the objectives used for (near-) optimum partitioning of feature spaces for hierarchical learning are identified. Implementation details are described in brief and results presented for both high dimensional synthetic and real data