Runtime Analysis of the $$(\mu + 1)$$-EA on the Dynamic BinVal Function

AbstractWe study evolutionary algorithms in a dynamic setting, where for each generation different fitness function is chosen, and selection performed with respect to the current function. Specifically, we consider Dynamic BinVal, which functions given by linear but order of bits randomly permuted. For \((1 + 1)\)-EA it was known that there an efficiency threshold \(c_0\) mutation parameter, at runtime switches from quasilinear exponential. Previous empirical evidence suggested larger population size \(\mu \), may increase. We prove rigorously this least case \(\varepsilon \)-neighborhood around optimum: \((\mu becomes arbitrarily large if \) chosen enough.However, most surprising result obtained second analysis =2\): increases increasing proximity optimum. In particular, hardest region optimization not (Extended Abstract. A full version available on arxiv [17].)