CHEN et al.:EAWITHDOUBLE-LEVELARCHIVESFORMULTIOBJECTIVEOPTIMIZATION 1853
The archive participates in the whole evolutionary process and
the updating strategy for the archive significantly influences
the performance of algorithm.
Considering the issues that: 1) a selection process prefer-
ring nondominated solutions would be adopted to preserve
individuals closer to the PS and 2) a technique is required
to achieve diversified distribution, the PAES and the NSGA-II
are proposed. Addressing issue 1), the PAES uses only one
parent and one offspring in each generation and the NSGA-II
proposed the nondominated sorting. Addressing issue 2), in
PAES positions of historically best solutions in an archive
are referred to, and in NSGA-II the crowding distance is
introduced for the selection of individuals residing in a less
crowded region.
There are other research efforts concerning solution distri-
bution strategies, which work on distributing the population
with diversity with certain mechanism. These efforts include
clustering the candidate individuals [18], [19]toachieve
highly spreading population, using multiple populations [15],
quantizing the solution space for selection [46], using dynamic
multiple populations [16], [17], devising new strategies to
estimate the density of objective space, territory definition
around each individual [35], and estimating the density of
solution space [30].
More recently, one class of MOEAs have been developed
based on problem decomposition and have gained much atten-
tion [20]–[23], [36], [37]. These algorithms achieve the two
MOEA goals in a different way from optimizing the multi-
objectives as a whole collectively. By decomposing the multi-
objective problem into different single-objective problems, the
task of finding solutions close to the true PF for the multiob-
jective problem is first replaced by that of optimizing a set
of single-objective sub-problems. In this case, the distribution
of solutions on PF merely depends on the methods of decom-
position at the beginning and of recombination at the end.
Among the MOEAs based on decomposition, the MOEA/D
proposed by Zhang and Li [24]isarepresentative.This
algorithm, after decomposing an MOEA problem into mul-
tiple single-objective scalar optimization problems, optimizes
all the scalar objectives simultaneously using only single-
objective evolutionary operators for simplicity and speed.
Since the framework of MOEA/D is compatible with existing
single-objective reproduction operators, an enhanced version
of MOEA/D [25]whichadoptsthereproductionoperatorof
DE [26]hasbeenproposed.Algorithmsinthisclassarecom-
petitive for a high convergence speed, high compatibility with
single-objective evolutionary operators, and good coverage to
the PF, as validated by a set of benchmark tests. Based on
problem decomposition, another decomposition method has
been proposed by Liu et al. [31]. The algorithm decom-
poses the multiobjective problem into a number of simple
multiobjective problems and assign each sub-problem one sub-
population to conquer the lack of sub-population diversity
resulted from the elitism in MOEA/D.
To enhance the local exploitation ability of MOEAs,
MOEAs are hybridized with local search strategies. In [38],
asynchronousparticlelocalsearch(SPLS)isadopted
in multiobjective particle swarm optimization (MOPSO).
Ke et al. [39]proposedamemeticalgorithmbasedon
decomposition (MOMAD), which hybridize the Pareto local
search with problem decomposition.
For many-objective optimization, research efforts have been
paid on strategies for working with large number of objectives
efficiently. Wang et al. [40]proposedapreference-inspired
coevolutionary algorithm (PICEA) for many-objective opti-
mization, which coevolved a population of solutions together
with a set of decision-maker preferences. Deb and Jain [41]
designed a many-objective particle swarm optimization algo-
rithm based on reference point. For set quality measurement
in many-objective MOEAs, Bader and Zitzler [42]designeda
hypervolumn estimation algorithm HypE. In [43], a shift-based
density estimation (SDE) strategy is proposed in order to make
Pareto-based MOEAs suitable for many-objective EAs.
B. Definition of Sub-Problems
The MOEA-DLA adopts both the problem-level archive and
sub-problem-level sub-archives, with each sub-archive serves
one corresponding sub-problem. As the proposed method is
closely related to the definition of sub-problems, this sec-
tion introduces the problem decomposition approach and the
definition of sub-problems.
1) Problem Decomposition: Several decomposition meth-
ods have been proposed to decompose a multiobjective prob-
lem into a series of single-objective sub-problems. In this
paper, two most commonly used decomposition approaches
are briefly introduced as follows. For more information on
decomposition methods, one can refer to [44]and[45].
a) Weighted sum approach: The weighted sum approach
is an intuitive approach to problem reduction. The method
considers a convex combination of all the objectives. To
reduce a multiobjective problem to a single-objective one,
define a weight vector w = (w
1
,...,w
m
)
T
,wherew
i
≥
0foralli = 1, 2, . . . , m and
!
m
i=1
w
i
= 1. One
weight vector w yields one sub-problem of the multiobjective
problem
minimize g
w
(x) =
m
"
i=1
w
i
f
i
(x) (1)
where f
i
(x) is the objective function of the ith objective
and whereby altering w can control the positions of optimal
solutions on the PF explicitly.
b) Chebyshev approach: In the Chebyshev (also known
as Tchebycheff) approach [1], the weight vector w is similar
to the weighted sum approach, but the sub-problems are
defined as
minimize g
w
(x) = max
1≤i≤M
{
w
i
·|f
i
(x) − z
i
|
}
(2)
where z
i
= (z
1
, z
2
,...,z
M
) is the reference point defined as
z = inf( f
j
(x)) for each j = 1, 2,...,M.
The Chebyshev approach to optimal solutions for a sub-
problem is controlled by the weight vector, which is similar to
the weighted sum approach. Here the reference point is asso-
ciated with the offset of PF. Problems with shifted PF in the
objective space can be handled by defining suitable reference