Using a scouting predator-prey optimizer to train support vector machines with non psd kernels

Abstract

In this paper, we investigate the use of an heterogeneous particle swarm optimizer, the scouting predator-prey optimizer, to train support vector machines with non positive definite kernels, including distance substitution based kernels. These kernels can arise in practical applications, resulting in multi-modal optimization problems where tra- ditional algorithms can struggle to find the global optimum. We compare the scouting predator-prey algorithm with the previous best evolutionary approach to this problem and a standard quadratic programming based algorithm, on a large set of of benchmark problems, using various non positive definite kernels. The use of cooperating scout particles allows the proposed algorithm to be more efficient than the other evolutionary approach, which is based on an evolution strategy. Both are shown to perform better than the standard algorithm in several dataset/kernel in- stances, a result that underlines the usefulness of evolutionary training algorithms for support vector machines.