5. MEMETIC ALGORITHM
The basic idea of a memetic algorithm (MA) is|at least as
it is most commonly interpreted|to extend an evolutionary
algorithm by a local search component in a suitable way to
increase the solution quality or speed up the search; for a
more general introduction to MAs see [17].
For the CTP, we use the EA variant that has been found
to perform best in [4] as a basis. As already mentioned in
Section 2 it always applies prune-delete-graft (PDG) recombination
and no mutation. The EA selects parents via binary
tournaments with replacement and works in a steady-state
fashion, i.e. in each iteration one ospring is derived, and
it always replaces the worst solution in the population with
one exception: To avoid duplicates and enforce a minimum
diversity, new solutions having the same TreeRank score as
solutions in the population are immediately discarded. The
initial population consists of a copy of the input tree collection
T and otherwise randomly created trees on L without
any bias. In addition to [4] we also tried to use recombination
and mutation together, though according to preliminary