30.07.2013 Views

Progressively Interactive Evolutionary Multi-Objective Optimization ...

Progressively Interactive Evolutionary Multi-Objective Optimization ...

Progressively Interactive Evolutionary Multi-Objective Optimization ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

conditions (Reklaitis et al., 1983) with a pre-specified error limit. We fix this error<br />

limit to 10 −2 in all problems of this study here. The solutions which meet<br />

this KKT satisfaction criterion are assigned an ‘optimal tag’ for further processing.<br />

For handling non-differentiable problems, a non-gradient, adaptive step-size<br />

basedhill-climbing procedure(Nolle,2006)canbe used.<br />

Step4: Updatingthearchive: The optimally tagged members, if feasible with respect<br />

to the upper level constraints (G), are then compared with the current archive<br />

members. If these members are non-dominated when compared to the members<br />

of the archive, they become eligible to be added into the archive. The dominated<br />

members in the archive are also eliminated, thus the archive always keeps nondominated<br />

solutions. We limit the size of archive to 10Nu. If and when more<br />

members are to be entered in the archive, the archive size is maintained to the<br />

abovelimitbyeliminatingextramembersusingthecrowdingdistance(CDa)measure.<br />

Step5: Formationofthecombinedpopulation: Steps 1 to 4 are repeated until the<br />

population QT is filled with newly created solutions. Each member of QT is now<br />

evaluated with F and G. Populations PT and QT are combined together to form<br />

RT. The combined population RT is then ranked according to constrained nondomination<br />

(Deb et al., 2002) based on upper level objectives (F) and upper level<br />

constraints (G). Solutions are thus, assigned a non-dominated rank (NDu) and<br />

members within an identical non-dominated rank are assigned a crowding distance(CDu)computed<br />

inthe F-space.<br />

Step7: Choosing halfthepopulation: Fromthecombinedpopulation RT ofsize 2Nu,<br />

half of its members are retained in this step. First, the members of rank NDu = 1<br />

are considered. From them, solutions having NDl = 1 are noted one by one in<br />

the order of reducing crowding distance CDu. For each such solution, the entire<br />

Nl subpopulation from its source population (either PT or QT) are copied in an<br />

intermediate population ST. If a subpopulation is already copied in ST and a<br />

future solution from the same subpopulation is found to have NDu = NDl =<br />

1, the subpopulation is not copied again. When all members of NDu = 1 are<br />

considered, a similar consideration is continued with NDu = 2 and so on till ST<br />

has Nu populationmembers.<br />

Step6: Upgradingoldlowerlevelsubpopulations: Each subpopulation of ST which<br />

arenotcreatedinthecurrentgenerationaremodifiedusingthelowerlevelNSGA-<br />

II procedure (Step 2) applied with f and g. This step helps progress each lower<br />

level populationtowardstheir individual Pareto-optimalfrontiers.<br />

The final population is renamed as PT +1. This marks the end of one generation of the<br />

overallH-BLEMOalgorithm.<br />

5.4 AlgorithmicComplexity<br />

With self-adaptive operations to update population sizes and number of generations,<br />

it becomes difficult to compute an exact number of function evaluations (FE) needed<br />

inthe proposedH-BLEMOalgorithm. However,using the maximumallowablevalues<br />

of these parameters, we estimate that the worst case function evaluations is Nu(2Tu +<br />

1)(tmax l + 1) + FELS. Here Tu is the number of upper level generations and FELS is<br />

the total function evaluations used by the local search (LS) algorithm. Lower level<br />

95

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!