Академический Документы
Профессиональный Документы
Культура Документы
I. I NTRODUCTION
optimal vector. Denote the Pareto set (PS) as the set of all the
Pareto-optimal solutions. A set that contains all the Paretooptimal objective vectors is called the Pareto front, denoted
by PF such that PF = {F(x) Rm |x PS}. A population
that is denoted as P is a set of solutions.
The solving of MOPs has been of great use in many
diverse areas including engineering, computer science, industry, economics, and physics; therefore, it has attracted
much attention. Over the past decades, many MOP algorithms
have been proposed, which can be divided into the following categories: traditional algorithms, evolutionary algorithms
(EA), memetic algorithms (MA), particle swarm optimization
algorithms (PSO), ant colony algorithms (ACA), simulated
annealing algorithms (SA), artificial immune systems (AIS),
tabu search algorithms (TS), scatter search algorithms (SS),
and so on [1].
In these algorithms, some famous methods are based on
EA called multiobjective evolutionary algorithms (MOEAs).
The vector evaluated genetic algorithm (VEGA) [2], multiobjective genetic algorithm (MOGA) [3], niched Pareto genetic
algorithm (NPGA) [4], and nondominated sorting genetic algorithm (NSGA) [5] were the early representative algorithms. In
1999, Zitzler and Thiele [6] presented the strength Pareto EA
(SPEA), which employs an external archive to store these nondominated solutions previously found. Since the appearance of
the SPEA, most researchers have tried to combine an external
archive and the current population with their MOEAs. At each
generation, a combined population with the external and the
current population is first constructed and the fitness of an individual is evaluated depending on the number of these external
nondominated solutions which dominate it. Furthermore, a
clustering technique is used to keep diversity. Another version
of SPEA (SPEA2) was proposed in [7], which incorporates a
fine-grained fitness assignment strategy, a density estimation
technique, and an enhanced archive truncation method in
contrast to its predecessor. There are another two popular
algorithms NSGA [5] and its improved version NSGA-II [8].
The NSGA-II is noticeably more efficient than its previous
version. It uses a fast nondominated sorting approach to
alleviate the computational complexity in the former method
and a mating pool is constructed by a selection operator to
select the best solutions. NSGA-II tends to spread quickly and
appropriately in most cases, but it has some difficulties to find
these nondominated vectors that lie in some certain regions
of the search space [1]. In 2000, a famous MOEA called
Pareto-archived evolution strategy (PAES) [9] was presented
c 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
1089-778X
51
52
(2)
w+k,i,t = xk,i,t + c u k,i,t vk,i,t
(3)
w
k,i,t = x k,i,t c u k,i,t vk,i,t
(4)
by the value of u k,i,t vk,i,t for giving disturbance
5) else if w
k,i,t x i,t ,
replace x i,t by w
k,i,t ;
6) else
do nothing;
7) end if.
It is obvious that replacing the current solution x i,t with
any neighborhood solution that dominates x i,t would help
the current solution converge to the Pareto-optimal front. If
w+k,i,t x i,t or w
k,i,t x i,t , one neighborhood solution is
also chosen to replace x i,t . The reason is that it can enhance
the diversity of these nondominated solutions found so far.
As a result, a new population Pt is created by performing
the local search schema on the current population Pt . All
solutions in the new population are as good as or better than
their counterparts in the current population.
The experimental results in Section III-F will show that this
simple method can get quite good convergence. The pseudo
code of the proposed local search is given in Algorithm 1.
2) More Discussion About Proposed Local Search Schema:
To discuss the feasibility of the proposed local search schema,
the DE is used for comparison. We first introduce the basic
DE briefly, and then discuss these superiorities of the proposed
local search schema over the general DE.
a) The basic concept of DE
DE maintains a population of N solutions: x i,t , i =
1, . . . , N at each generation of the algorithm, where the
index i denotes the ith solution of the population and
t denotes the generation number of the algorithm. DE
contains three main operators: mutation, crossover, and
selection. The mutation operation is defined as
1
2
3
+ c (xk,t
xk,t
) k {1, ..., n}
z k,i,t = xk,t
(6)
1
1
2
2
1
2
, x2,t
, . . . , xn,t
), x 2t (x1,t
, x2,t
, . . . , xn,t
), and
where x 1t (x1,t
3
3
3
x 3t (x1,t
, x2,t
, . . . , xn,t
) are three different solutions chosen randomly from the current population, respectively.
zi,t (z 1,i,t , z 2,i,t , . . . , z n,i,t )is the perturbed solution. c is a
control parameter, which is set in (0, 2] by Storn and
Price [24]. Once zi,t is generated by (6), a trial solution
wi,t (w1,i,t , w2,i,t , . . . , wn,i,t ) is produced by the crossover
operation which is defined as
z , randj CR j = CJ
wj,i,t = j,i,t
(7)
xj,i,t , other wise
53
54
is obvious that the point of the high density space has a low
chance to be selected so that the spread is not good enough.
The crowded comparison mechanism has been used for
diversity maintenance of the obtained solutions in many
MOEAs. In addition, there are several studies, which
slightly modified the original crowding distance in [74][78].
Kukkonen and Deb [77] proposed an improved pruning of nondominated solutions for bi-objective optimization problems.
This method removes the solution that has the smallest crowding distance value one by one and recalculates the crowding
distance value after each removal until the number of the
remaining solutions is equal to the population size. In [78], a
fast and effective method which is based on crowding distance
using the nearest neighborhood of solutions in Euclidean sense
is proposed for pruning of the nondominated solutions.
In the proposed NSLS, we replace the crowded comparison mechanism in NSGA-II [8] with the farthest-candidate
approach that can solve above difficulty to some extent. The
suggested method is inspired by the best-candidate sampling
algorithm [79] in sampling theory. Suppose that we are going
to select K best points from F points, whenever a new point
is to be selected, the candidate point in the unselected points
which is farthest from the selected points is accepted. Here, the
distance is defined by Euclidean Norm. The boundary points
(solutions with the smallest and largest function values) are
selected first. Algorithm 2 is shown in Pseudo code, where
Paccept stores the selected solutions, D [x] stores the minimum
Euclidean distance between x and the unselected points, and
dis(x, x ) is a function calculating the Euclidean distance
between solution x and x . The complexity of this procedure is
governed by the selecting procedure. Thus, the computational
complexity of the proposed algorithm is O(mF 2 ), where F
is the number of the original points and m is the number
of the objectives. As F 2N, so the worst cast complexity
of the farthest-candidate approach is O(mN 2 ). The proposed
method is good at dealing with above case in Fig. 1 [line (b)
is the selection results of the proposed method], from which
the farthest-candidate method has a better spread than that of
the crowded comparison mechanism. In addition, the overall
crowding distance value in NSGA-II is calculated as the sum
of solution distance values corresponding to each objective and
the space distance is actually the Euclidean distance. Thus,
End For
For each x in F Paccept
D[x] argmin x Paccept dis(x, x )
End For
For i = 1 to K |Paccept |
x 1 = argmax x(F Paccept ) (D[x])
For each x 2 in F Paccept
D[x 2 ] min(D [x 2 ] , dis(x 1 , x 2 ))
Paccept Paccept x 1
End For
End For
xF
55
56
Fig. 2. Plots of the nondominated solutions with the lowest IGD-metric values found by NSGA-II, MOEA/D-DE, NSLS-C, and NSLS in 20 runs in the
objective space on ZDT2, ZDT4, and LZ07 F9 where N = 100 and T = 250.
TABLE I
N UMBER OF P OINTS IN E ACH Z HANG S P F F ILE
57
TABLE II
GDPS-M ETRIC VALUES OF N ONDOMINATED S OLUTIONS F OUND BY E ACH A LGORITHM ON P ROBLEMS ZDT S W HERE N = 100 AND T = 250
d (Ei , ) + X |d (X, ) d|
(12)
= i=1 m
1
d(X, )
d =
|| X
(13)
(14)
58
TABLE III
IGD-M ETRIC VALUES OF N ONDOMINATED S OLUTIONS F OUND BY E ACH A LGORITHM ON P ROBLEMS ZDT S W HERE N = 100 AND T = 250
59
Fig. 3. Plots of the nondominated solutions with the lowest IGD-metric values found by NSGA-II, MOEA/D-DE, NSLS-C, and NSLS in 20 runs in the
objective space on UF7 where N = 100 and T = 250.
60
TABLE IV
IGD-M ETRIC VALUES OF N ONDOMINATED S OLUTIONS F OUND BY E ACH A LGORITHM ON P ROBLEMS LZ07 S W HERE N = 100 AND T = 250
TABLE V
IGD-M ETRIC VALUES OF N ONDOMINATED S OLUTIONS F OUND BY E ACH A LGORITHM ON P ROBLEMS UF S W HERE N = 100 AND T = 250
61
62
TABLE VI
IGD-M ETRIC VALUES OF N ONDOMINATED S OLUTIONS F OUND BY E ACH A LGORITHM ON P ROBLEMS LZ07 S W HERE N = 300 AND T = 500
63
Fig. 5. Plots of the nondominated solutions with the lowest IGD-metric values found by NSGA-II, MOEA/D-DE, NSLS-C, and NSLS in 20 runs in the
objective space on LZ07 F1 - LZ07 F3 where N = 300 and T = 500.
64
Fig. 6. Plots of the nondominated solutions with the lowest IGD-metric values found by NSGA-II, MOEA/D-DE, NSLS-C, and NSLS in 20 runs in the
objective space on LZ07 F4, LZ07 F5, and LZ07 F7 where N = 300 and T = 500.
65
Fig. 7. Plots of the nondominated solutions with the lowest IGD-metric values found by NSGA-II, MOEA/D-DE, NSLS-C, and NSLS in 20 runs in the
objective space on LZ07 F8, LZ07 F9, and UF4 where N = 300 and T = 500.
66
Fig. 8. Plots of the nondominated solutions with the lowest IGD-metric values found by NSGA-II, MOEA/D-DE, NSLS-C, and NSLS in 20 runs in the
objective space on UF5, UF6, and UF7 where N = 300 and T = 500.
67
Fig. 9. Plots of the nondominated solutions with the lowest IGD-metric values found by NSGA-II, MOEA/D-DE, NSLS-C, and NSLS in 20 runs in the
objective space on LZ07 F6, UF9, and UF10 where N = 595 and T = 500, and the three images in the last line are the corresponding Pareto-optimal front
of LZ07 F6, UF9, and UF10.
the p-values are all smaller than = 0.05, which reveals that
NSLS outperforms all the other algorithms on both of the two
experiments with different parameter settings of the population
size and the number of iterations. Likewise, from Table XVI,
which shows the results on the basis of the GD and the spread
metrics, it is obvious that NSLS is better than the other
algorithms too.
I. Sensitivity in NSLS
1) Sensitivity of : is a control parameter adopted in the
proposed local search schema in NSLS. To study the sensitivity of the performance to in NSLS, we have tested different
68
TABLE VIII
IGD-M ETRIC VALUES OF N ONDOMINATED S OLUTIONS F OUND BY E ACH A LGORITHM IN 20 RUNS W HERE N = 100 AND T = 250
TABLE IX
GD- METRIC VALUES OF N ONDOMINATED S OLUTIONS F OUND BY E ACH A LGORITHM ON B I - OBJECTIVE
P ROBLEMS OF LZ07 S W HERE N = 300 AND T = 500
TABLE X
GD-M ETRIC VALUES OF N ONDOMINATED S OLUTIONS F OUND BY E ACH A LGORITHM ON T RI -O BJECTIVE
P ROBLEMS LZ07 F6, UF9, AND UF10 WHERE N = 595 AND T = 500
TABLE XI
-M ETRIC VALUES OF N ONDOMINATED S OLUTIONS F OUND BY E ACH A LGORITHM ON B I - OBJECTIVE
P ROBLEMS OF LZ07 S W HERE N = 300 AND T = 500
TABLE XII
-M ETRIC VALUES OF N ONDOMINATED S OLUTIONS F OUND BY E ACH A LGORITHM ON T RI -O BJECTIVE
P ROBLEMS : LZ07 F6, UF9, AND UF10 W HERE N = 595 AND T = 500
Fig. 10.
69
Fig. 12. GD-metric values versus number of generations for LZ07 F9.
Fig. 13. Mean IGD-metric values versus the value of in NSLS for ZDT1.
Fig. 11. IGD-metric values versus the number of function evaluations for
LZ07 F5.
IV. C ONCLUSION
settings of in the implementation of NSLS on ZDT1. All
parameter settings are similar to that in Section III-D, except
the setting of . Fig. 13 shows the mean IGD-metric values
in 20 runs versus the value of in NSLS on ZDT1. It is
clear from Fig. 13 that NSLS performs well with from 0
to 1.0 on ZDT1. And it works well for all values of with
a small difference. Therefore, it can be concluded that NSLS
is not very sensitive to the setting of , at least for the MOPs
that are more or less similar to these problems adopted in
this paper.
2) Sensitivity of : is another control parameter utilized in the proposed local search schema in NSLS. To
discuss the sensitivity of the performance to in NSLS, we
have tested different settings of in the implementation of
NSLS on ZDT1. All parameter settings are similar to that in
Section III-D, except the setting of . It is clear from Fig. 14
that NSLS performs well with from 0 to 0.2 on ZDT1. And
it works well for all values of with a small difference and
achieves the best value with = 0.1. Therefore, we can claim
that NSLS is not very sensitive to the setting of too.
70
TABLE XIII
R ESULTS OF NSLS AND NSLS-V W HERE N = 300 AND T = 500
TABLE XVI
S TATISTICAL R ESULTS BY W ILCOXON T EST FOR GD AND M ETRICS
C ONSIDERING UF9, UF10, AND LZ07 S W HERE N = 300 AND T = 500
Fig. 14. Mean IGD-metric values versus the value of in NSLS for ZDT1.
TABLE XIV
S TATISTICAL R ESULTS BY W ILCOXON T EST F OR IGD-M ETRIC
C ONSIDERING ZDTS, LZ 07 S , AND UF S W HERE N = 100 AND T = 250
TABLE XV
S TATISTICAL R ESULTS BY W ILCOXON T EST FOR IGD-M ETRIC
C ONSIDERING LZ 07 S AND UF S W HERE N = 300 AND T = 500
71
72
73