Evolutionary Gradient Search Revisited


Dirk V. Arnold
Ralf Salomon

Author Addresses: 

Faculty of Computer Science
Dalhousie University
6050 University Ave.
PO Box 15000
Halifax, Nova Scotia, Canada
B3H 4R2


Evolutionary gradient search is an approach to optimization that combines features of gradient strategies with ideas from evolutionary computation. Recently, several modifications to the algorithm have been proposed with the goal of improving its robustness in the presence of noise and its suitability for implementation on parallel computers. In this paper the value of the proposed modifications is studied analytically. A scaling law is derived that describes the performance of the algorithm on the noisy sphere model and allows comparing it with competing strategies. The comparisons yield insights into the interplay of mutation, multirecombination, and selection. Then, the covariance matrix adaptation mechanism originally formulated for evolution strategies is adapted for use with evolutionary gradient search in order to make the algorithm competitive on objective functions with a large condition numbers of their Hessians. The resulting strategy is evaluated experimentally on a number of convex quadratic test functions.

Tech Report Number: 
Report Date: 
January 8, 2005
PDF icon CS-2005-09.pdf342 KB