An analysis of evolutionary gradient search

Evolution strategies and gradient strategies are two different approaches to continuous optimization. Salomon's evolutionary gradient search procedure is a hybrid strategy that obtains gradient estimates by borrowing the idea of random variations from evolutionary computation. The present paper applies successful tools and ideas from the theory of evolution strategies to the evolutionary gradient search framework. Performance laws are derived that shed light on the strategy's performance and the influence of its parameters. Comparisons with the (mu/mu, lambda)-ES are presented, and the issue of genetic repair in evolutionary gradient search is discussed. The practically relevant problem of noisy objective function measurements is addressed, and recommendations with regard to the setting of strategy parameters are made.