GA-based approaches for finding the minimum reference set for nearest neighbor classification

In this paper, we examine the ability of genetic algorithms to find a compact reference set for nearest neighbor classification. The task of genetic algorithms is to select a small number of reference patterns from a large number of given training patterns. Our pattern selection problem has two objectives: to maximize the classification performance of the reference set and to minimize the size of the reference set. In our genetic algorithm, they are combined into a single scalar fitness function using a constant weight for each objective. Thus our pattern selection problem is handled as a single-objective combinatorial optimization problem with 0-1 variables where "1" means the inclusion of the corresponding pattern in the reference set and "0" means the exclusion. In this paper, we first briefly explain our genetic algorithm for the pattern selection problem for nearest neighbor classification. Next we examine the ability of the genetic algorithm to find a compact reference set by computer simulations on commonly used real-world pattern classification problems. Finally, we suggest some extensions of our genetic algorithm.