Improved evolution strategies for high-dimensional optimization
Wang Xiang-zhong, YU Shou-yi · Control theory & applications · 2006
For high-dimensional continuous function optimization,manners of mutation and reproduction of classical evolution strategies(CES) are investigated.Concepts of all-gene mutation and single-gene mutation are proposed and it is proofed through theoretical analysis and simulation that single-gene mutation significantly outperforms all-gene mutation in both local searching capability and computation costs.Since parameters of CES cannot properly track the process of evolution because of their strong randomness,a strategy parameter is introduced which descends in the process of evolution.Finally,a new ES,called(μ+λ+k)-ES,is established which is characterized with single-gene Gaussian plus uniform mutation,elitist-reproduction,descending strategy parameter and small population size.Simulation results on a set of 100-dimensional typical test functions are presented.