On averaging the best samples in evolutionary computation