Estimation of von mises-fisher distribution algorithm, with application to support vector classification
Adetunji David Ajimakin, V. Susheela Devi · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2021
The most common method in the Evolutionary Algorithm community to handle constraints is to use penalties. The simplest being the death penalty, which rejects solutions that violate constraints. However, its inefficiency in search spaces possessing small feasible regions spurred research into adaptive penalties and other competitive methods. A major criticism of these approaches is that they require the user to fine-tune parameters or design problem-dependent operators. We propose to do away with penalty functions for problems over the Euclidean space when the constraint is an equality concerning the Euclidean distance. This paper describes an evolutionary algorithm on the unit hypersphere based on representing the population with the von Mises-Fisher probability distribution from the field of Directional statistics. We demonstrate its utility by solving the support vector classification problem for a few datasets.