A new selection scheme for steady-state evolution strategies
Jürgen Wakunda, Andreas Zell · 2000
In this paper, a new selection scheme for a (m + 1) steady-state evolution strategy is described: the so-called median selection. In contrast to generational algorithms, only one individual is generated and evaluated in one step of the algorithm and is immediately integrated into the population. Previous steady-state algorithms are similar to the (m + l) selection scheme in evolution strategies, which has a disadvantage in the fast self-adaptation of mutation step-length. This is compensated by the presented median selection, which is oriented at the (m, l) selection. The median selection is compared with other steady-state selection schemes and with (m, l) selection. As a result, median selection achieves better or equally good results as the other selection schemes for a large number of benchmark functions. Additionally, it is shown that the use of a sequential steady-state evolution strategy is advantageous even on oneprocessor computers.