Reently it has been pointed out that not only high diversity of solutions
inthe objetivespae but alsohighdiversity ofsolutions intheeient set
an be ofinterestfor the deisionmaker[68 ,102 ℄. For instane, ifaspei
pointon thePareto front isseletedbythedeisionmaker,itmight alsobe
interesting to onsider dierent possible realizations to this solution in the
deision spae. Hene, if there aretwo dierent pre-images of the seleted
point on the Pareto front in the eient set, both of them are of potential
interest for thedeision maker. Thissituation is illustrated inFigure 5.3.
More preisely, the dierene between the lassial seletion priniple to
our proposed approah an be formalized as follows. Let
A
denote an ap- proximation set on whih we would like to apply ranking, and let~xA
and~xB
be two solutions inA
. In the lassial seletion method, as employed bythe NSGA-IIor SMS-EMOA algorithms, a solution~xA
is preferred to a solution~xB
if~xA
hasa better dominane rank than~xB
inA
,with respet tonon-dominated sorting. Giventhat~xA
and~xB
sharethesamedominane rank inA
, then~xA
is preferred to~xB
, if and only if~xA
ontributes moreDecision Space
Objective Space
Figure5.3: Diversityfordeisionmaking: Illustrativeexampleforasenario
where two adjaent pointson the Pareto front are mapped onto two points
in two ompletely dierent regions in the deision spae. Unitsand sales
arearbitrary.
to the diversityofthe approximation setintheobjetive spaethan
~xB
. In the proposed seletion priniple,~xA
remains preferable to~xB
, if~xA
has a betterdominanerank than~xB
inA
. However, given that~xA
and~xB
share the same dominane rank inA
, then~xA
is preferred to~xB
, if and only if it ontributes moreto the diversity intheaggregated spae (i.e.,inbothobjetiveanddeisionspaes). Thisprinipleanbeinstantiated indierent
ways,dependingon thediversitymeasuredened on theaggregated spae.
Multi-objetive optimization methods aim at maintaining diversity, by
their denition, and indeed, one of the popular mehanisms for diversity
maintenane is therowding onept [67 ℄, whih is also applied, yet dier-
ently, as a single-objetive nihing tehnique. Thus, the important ompo-
nent ofdiversity is the linkingelement between theelds ofmulti-objetive
andmulti-modaloptimization. However,inmulti-objetiveoptimizationthe
diversitymaintenane istypiallysought intheobjetivespae, for thesake
ofobtainingafairoverageoftheParetofront,whilenottakenintoaount
for the Paretooptimal setinthedeision spae.
5.2.1 Related Work
Severaldierent studies treated relatedtopisto thework presentedinthis
hapter. We reviewthem hereshortly.
NihingforMOEA: TheNPGA Nihingtehniqueshavebeenalready
multi-objetiveoptimization,knownastheNihed-ParetoGA(NPGA).The
algorithm wasa variant of thetness sharing nihing method, whereas the
nihing distane metri wasset toonsiderthe objetive spae only. These-
letionwasbasedontheso-alledParetodominationtournaments or onthe
minimalniheount,otherwise. TheNPGAwasalassialexampleofusing
an existingsingle-objetive nihing tehnique, ina straightforward manner,
for multi-objetive optimization - only by redening the nihing distane
measure and the seletion mehanism. However, its kernel was the simple
GA, whih typially suers from limited performane in high-dimensional
ontinuous landsapes,and itlaked anyself-adaptation mehanism.
TheOmni-Optimizer Deb'sso-alledOmni-Optimizer[68 ℄isonsidered
to be one of the rst and only attempts of introduing a generi optimiza-
tion routine whih aims at overing the four ategories of funtion opti-
mization: Single-objetive uni-global, single-objetive multi-global, multi-
objetive uni-global, and multi-objetive multi-global problems. Also, it is
one of the rst attempts to take diversityinthe deisionspae into onsid-
eration.
In priniple, this algorithm extends the NSGA-II by onsidering addi-
tionallythediversityinthedeisionspae. Thisisimplementedbymeansof
therowdingdistanealulationinthedeisionspaeforalltheindividuals.
The assignedrowding distane isdened asfollows:
if rowd_dist_obj
(i)
>
avg_rowd_dist_obj or rowd_dist_de(i)
>
avg_rowd_dist_dethen rowd_dist
(i) =
max(
rowd_dist_obj(i),
rowd_dist_de(i))
else rowd_dist(i) =
min(
rowd_dist_obj(i),
rowd_dist_de(i))
i.e., ifthe individual hasabove-the-average rowding distane, eitherin thedeision or objetive spae, the larger of them is assigned to it, otherwise
thesmallerofthetwodistanes isassigned. Thisriterionis rathergeneral,
andstrongly reliesonuniform distributionofpeaks aswell asontheir equal
tness values. Also,thesalabilityofthetwodierent spaes isnottreated.
We would like to speulate that it is expeted to experiene diulties on
non-uniform multi-modal landsapes, for instane. From the pratial per-
spetive,thealgorithm wasreportedin[68 ℄tobetestedonlyonasingletest
funtion, onstruted by Deb for this purpose, with uniformly-distributed
equi-tness minima landsape. Weshall revisitthis test-funtion inour ex-
perimental proedure.
Deision-Spae Diversity as an Independent Objetive Toolo and
Benini [104 ℄ also promoted the issue of geneti diversity in multi-objetive
diversity of trial solutions in the deision spae, quantied by means of a
overage funtion,asanindependentobjetive,subjet to maximization, in
the ongoingmulti-objetive searh. This GA-based approah wasshown to
outperform the NSGA on a set of
30D
bi-riteria minimization problems introdued byZitzler etal.[105 ℄.Self-AdaptationinMulti-ObjetiveOptimization Self-adaptationof
strategy parameters [106 ℄ hasbeome afundamental omponent intheevo-
lutionary optimization routine. Moreover, the self-adaptation of the muta-
tionstrategy parameters hasbeenshown tobeneessaryforeient single-
objetiveoptimization within ES[106℄.
Self-adaptation is expeted to fail inthe lassial multi-objetive optimiza-
tion routine. This is due to the fat that given oniting objetives, a
suessfulmutation toward one objetive isnot neessarilya suessfulmu-
tation toward theothers andhene shouldnot beseleted.
Bühe, Müller and Koumoutsakos [107 ℄ onduted a pioneering study of
self-adaptation in multi-objetive optimization. They onsidered three dif-
ferent lasses of multi-objetive algorithms - independent sampling, ooper-
ative populationsearh withdominane riterion and ooperative population
searh without dominane riterion. Three representatives - CMEA, SPEA
andSDM-mathingthelassesrespetively,weretestedonamulti-objetive
generalizationofthespheremodel,andomparedwithrespettoeahother.
Self-adaptation had been plugged-in into theevolutionaryoremehanisms
of the algorithms, in a limited way (rotation angles, for instane, were not
always adapted). The onlusion was thatself-adaptation didnot work for
ooperativepopulationsearheswhihusethedominaneriterion inthet-
ness assignment (SPEA), andthis resultwasreassuredbytestingmorerep-
resentativesfromthatlassofalgorithms, suhastheNSGA-IIandSPEA2.
However, self-adaptationouldworkfor theCMEAand SDM,whihdo not
usedominane, butrather onsiderasingleobjetive foroptimization while
the otherobjetivesaretreated asonstraints. Theonludingmessagewas
lear self-adaptation doesnot work inits lassial denitionupononsid-
ering multiple objetives ashadbeen speulated.
Reently, the self-adaptation obstale was treated suessfully by using
the so-alled hyper-volume indiator (also known as S-metri) [99 ℄ as a se-
letionriterion,similarto[100 ℄,intheMulti-Objetive CMA-ES[33 ℄,tobe
disussed next. Asimilar approah,yetemploying asimpler ESkernel,was
also reportedreently in[108 ℄.
CMA-ES forMulti-Objetive Optimization Analgorithm for multi-
objetiveoptimizationwithaCMAkernelwasintrodued reently [33 ℄,em-
riterion, followedbythe maximizationof thePareto front hyper-volume as
a seondaryriterion. Crowding distanewasalsoonsidered asan alterna-
tive seondary seletion riterion. In many ways, this algorithm resembles
our nihing framework. However, its diversity preservation stems from the
outome ofseletion withrespetto multiple riteria,rather than from the
spatialenforement ofspeiationbymeansofanihe denition. Itisimpor-
tant to noteinthis ontext,that thehyper-volume indiator is well-dened
asameasureofdiversityand solution-setqualityintheobjetivespae, but
annotbe applied asanindiator ofdiversity inthesearh spae.