This paper is an attempt to answer the above question. A novel
algorithm, known as Multiple Evolutionary Algorithm (MultiEA),
is proposed. A portfolio is first formed by selecting several state
of the art EAs. A novel predictive measure is reported to predict
the performance of individual algorithms if they were extrapolated
to the same number of evaluations in the nearest future. The
algorithm with the best predicted performance is chosen to run for
one generation. New search information is received and the
history is updated. The predicted performance of the algorithms is
updated and the algorithm with the best predicted performance is
re-selected, which may or may not be the same algorithm in the
last generation. Experimental results show that the measure is
stable and is a reasonably effective predictor. The idea is simple
and natural. It is parameter-less; it does not introduce any new
control parameter to the algorithm, thus avoiding the challenging
parameter tuning and control problem.