Exploitation of Sub-populations in Evolution Strategies for Improved Numerical Optimization

Tenth Midwest Artificial intelligence and Cognitive Science Conference

Volume

Tenth Midwest Artificial intelligence and Cognitive Science Conference

Contents

This paper describes the use of a modified Differential Evolution strategy that identifies multiple solutions to the numerical optimization of multidimensional objective functions. Traditional approaches to this class of problems, such as Newton’s Method, are restricted for use on continuous, differentiable functions; in addition, the solution identified by these approaches is often dependent upon the initial guess. The ability to find the multiple solutions is therefore restricted by one’s ability to choose appropriate initial conditions. The Differential Evolution strategy described in this paper is not restricted by continuity and differentiability requirements, and can therefore robustly exploit the concept of subpopulations to converge to multiple solutions in a multi-dimensional problem space.