Automatic finding of good classifiers following a biologically inspired metaphor

Abstract

The design of nearest neighbour classifiers can be seen as the partitioning of the whole domain in different regions that can be directly mapped to a class. The definition of the limits of these regions is the goal of any nearest neighbour based algorithm. These limits can be described by the location and class of a reduced set of prototypes and the nearest neighbour rule. The nearest neighbour rule can be defined by any distance metric, while the set of prototypes is the matter of design. To compute this set of prototypes, most of the algorithms in the literature require some crucial parameters as the number of prototypes to use, and a smoothing parameter. In this work, an evolutionary approach based on Nearest Neighbour Classifiers (ENNC) is introduced where no parameters are involved, thus overcoming all the problems derived from the use of the above mentioned parameters. The algorithm follows a biological metaphor where each prototype is identified with an animal, and the regions of the prototypes with the territory of the animals. These animals evolve in a competitive environment with a limited set of resources, emerging a population of animals able to survive in the environment, i.e. emerging a right set of prototypes for the above classification objectives. The approach has been tested using different domains, showing successful results, both in the classification accuracy and the distribution and number of the prototypes achieved.