tests the performance was usually worse, hence solely recombination
was nally used, too. The EA terminates when a
given time or iteration limit is reached.
We embedded dierent variants of local search utilizing
the neighborhoods described in Section 3. SPRr was omitted
as it yielded poorer results in preliminary tests, presumably
because of the similarity to PDG recombination. Each
of the variants uses a single neighborhood structure and applies
a random neighbor step function; i.e. a random move
is performed and the new solution is accepted if it is better
than the original one. Improved trees are re-encoded in
the chromosome in a Lamarckian manner. A local search
phase terminates after a certain number of consecutive nonimproving
moves.
We further developed a simple progressive search (SPS)
roughly following the idea in [11]. In this original work, the
algorithm starts with the large SPR neighborhood and progressively
reduces it by limiting the distance between the
removal and insertion positions of a subtree until the neighborhood
converges to NNI having a distance limit of one.
In contrast, we use several dierent neighborhoods without
this smooth transition though maintaining the idea to reduce
the size of the applied neighborhood: In the rst third
of the MA|either w.r.t. an iteration or time limit| we apply
the Step local search, followed by Swap in the second
third, and nally Rotate in the last third, whereas the local
search variants are the ones described before.
Although a single application of one of the local search
variants is relatively fast, trying to improve every ospring
would dramatically increase the overall run-time without a
tests the performance was usually worse, hence solely recombinationwas nally used, too. The EA terminates when agiven time or iteration limit is reached.We embedded di erent variants of local search utilizingthe neighborhoods described in Section 3. SPRr was omittedas it yielded poorer results in preliminary tests, presumablybecause of the similarity to PDG recombination. Eachof the variants uses a single neighborhood structure and appliesa random neighbor step function; i.e. a random moveis performed and the new solution is accepted if it is betterthan the original one. Improved trees are re-encoded inthe chromosome in a Lamarckian manner. A local searchphase terminates after a certain number of consecutive nonimprovingmoves.We further developed a simple progressive search (SPS)roughly following the idea in [11]. In this original work, thealgorithm starts with the large SPR neighborhood and progressivelyreduces it by limiting the distance between theremoval and insertion positions of a subtree until the neighborhoodconverges to NNI having a distance limit of one.In contrast, we use several di erent neighborhoods withoutthis smooth transition though maintaining the idea to reducethe size of the applied neighborhood: In the rst thirdof the MA|either w.r.t. an iteration or time limit| we applythe Step local search, followed by Swap in the secondthird, and nally Rotate in the last third, whereas the localsearch variants are the ones described before.Although a single application of one of the local searchvariants is relatively fast, trying to improve every o springwould dramatically increase the overall run-time without a
การแปล กรุณารอสักครู่..
