Abstract:
Maintaining proper balance between exploitative and explorative operations of an
evolutionary algorithm is essential for preventing premature convergence to local optima
and for sustaining sufficient convergence speed throughout the evolution. This thesis
introduces Recurring Multistage Evolutionary Algorithm (RMEA), a completely new
framework to balance the exploitative and explorative features of the conventional
evolutionary algorithm. The basis of RMEA is repeatedly alternating three different
stages of evolution, each with its oWn explorative or exploitative objective and genetic
operators. As the stages of RMEA repeat, the conflicting goals of exploitation and
exploration are distributed gracefully across the generations of the different stages. The
key concept of RMEA is to combine dissimilar information across the population for
search space exploration and to combine similar information within population
neighborhood for local exploitation. Performance of RMEA has been evaluated on a
number of benchmark numerical optimization problems and results are compared with
several existing algorithms. Experimental results show that RMEA performs better
optimization with a higher rate of convergence for most of the problems. Also, an
in-depth experimental study is carried out about the roles of the different stages and
operators of RMEA, as well as the sensitivity of its parameters. An adaptive variant of
RMEA is also proposed, which adjusts its parameters in an adaptive manner during the
evolution, and does not require any problem specific knowledge from the user.
