An Evolution Strategy for On-line Optimisation of Dynamic Objective Functions

摘要

We review recent research in Evolution Strategy (ES) operators with particular emphasis on improving convergence within small populations. We then report on the results of applying some of these operators in a problem domain where performance is critically dependent on population size and where the objective function is dynamic, i.e. changing shape as optimization proceeds. The ES must operate quickly and efficiently, acting as the exploration component of a larger on-line learning architecture. We found that the use of a derandomised mutation operator and. intermediate recombination resulted in a considerable performance improvement.