Expert Systems With Applications 158 (2020) 113446

Contents lists available at ScienceDirect

Expert Systems With Applications

journal homepage: www.elsevier.com/locate/eswa

Post Pareto-optimal ranking algorithm for multi-objective optimization

using extended angle dominance
Supoj Choachaicharoenkul, Naruemon Wattanapongsakorn∗
Department of Computer Engineering, Faculty of Engineering, King Mongkut’s University of Technology Thonburi (KMUTT), Bangkok, Thailand

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a solution ranking algorithm to find the outstanding solutions in given set of non-
Received 9 September 2019 dominated solutions of multi-objective optimization problems, which are the results from either Multi-
Revised 21 January 2020
Objective Evolutionary Algorithms (MOEAs) or exact methods. The algorithm enables the decision makers
Accepted 8 April 2020
to identify outstanding solutions without a deep understanding of the problem. The algorithm provides
Available online 6 May 2020
a ranking for all solutions so that they can obtain any top K ranked solutions to implement. This novel
Keywords: parameter-free solution ranking approach is based on two concepts: an extended angle-based dominance
Extended angle-based dominance technique from the algorithm called ADaptive angle-based pruning Algorithm (ADA) for discovering the
Knee solutions knee solutions and the inverse-square law of light for enhancing the diversity of solutions. We evaluate
Implicit preference the performance of the approach on several well-known test problems against well-known knee finding
Multi-objective optimization algorithms as well as on a practical system design and optimization problem to demonstrate the useful-
Solution ranking
ness of the algorithm.
© 2020 Elsevier Ltd. All rights reserved.

1. Introduction and background discover these solutions without the need for the preferences. The
algorithms in this group identify the common preferred solutions:
Multi-objective Optimization Problems (MOPs) can be solved ef- knee solutions, which are the most outstanding or most cost-
ficiently using Multi-objective Optimization Algorithms (MOEAs). effective solutions and they are almost always the most preferred
The results from these algorithms are non-dominated solutions solutions (Deb & Gupta, 2010; Setoguchi, Narukawa, & Ishibuchi,
that represent approximated Pareto-optimal solutions (the best 2015).
trade-off solution set). Unfortunately, presenting the whole Pareto- However, the number of solutions as the results from
optimal solutions to the decision makers can cause trouble for preference-based MOEAs and the needs of decision makers are
them in decision making to select only one or a few solutions mismatched. They need to adjust the parameters of the algorithms,
to implement (Ramirez-Atencia, Mostaghim, & Camacho, 2017; which is a tedious process of trial-and-error in order to get the
Weiszer, Chen, Stewart, & Zhang, 2018). expected number of desired solutions. This paper aims to design
Two mainstream techniques have been developed to overcome a decision support algorithm based on knee solutions concept.
the problem by discarding the less outstanding solutions: explicit The algorithm uses the non-dominated solutions from the existing
and implicit preference-based MOEAs. For explicit preference- MOEAs as input and then rank them all. Consequently, the deci-
based MOEAs, the decision makers need to specify preferences ex- sion makers can choose any top K outstanding solutions to imple-
plicitly, such as a solution of interest or ranking of objectives’ im- ment, where K is a user-defined value. They can utilize the algo-
portance. Then the algorithms will use them as additionally re- rithm without providing any other preference or parameter.
quirement(s) to scope down the results to the region of interests In many fields, optimization, that is, finding extreme values of
(ROIs). However, the preferences are not always available from the mathematical functions, is a vital process. The objective is to ob-
decision maker, so the algorithms are limitedly used (Bechikh, Said, tain the minimum value of quantities such as cost or carbon emis-
& Ghedira, 2011; Branke, Deb, Dierolf, & Osswald, 2004). Con- sion or the maximum value of outcomes such as performance or
versely, some solutions are naturally, repeatedly be chosen by deci- profit. Many real-world functions are too complex to be optimized
sion makers. The implicit preference-based MOEAs are dedicated to using exact methods. Hence, many optimization algorithms apply
approximate solutions to produce acceptable results.
Moreover, many problems require the simultaneous considera-
∗
Corresponding author. tion of multiple aspects. In most of these cases, objectives are in
E-mail address: [email protected] (N. Wattanapongsakorn). conflict and compete with each other. For example, cost or re-

https://doi.org/10.1016/j.eswa.2020.113446
0957-4174/© 2020 Elsevier Ltd. All rights reserved.
2 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446

source consumption are typically for minimizing while the ob-

jectives such as profit or performance are typically for maximiz-
ing. MOPs can be found in many fields, for examples, minimiz-
ing force and energy consumption for robot leg-mechanism (Deb
& Gupta, 2010), maximizing the expected portfolio return while
minimizing the risk for stock portfolio optimization (Shalan &
Ykhlef, 2015), and minimizing the deliver time and cost for a
transportation (Wang, Yuan, Zhao, & Lu, 2020). Optimization prob-
lems of this type are called Multi-objective Optimization Problems
(MOPs). The general form of a MOP can be formulated as follows
(1).
Minimize ( f1 (x ), f2 (x ), . . . , fM (x ) ),
subject to g j (x ) ≥ 0, j = 1, 2, . . . , J, (1)
Fig. 2. Pareto Dominance for a bi-objective minimization problem.
hk ( x ) = 0 , k = 1 , 2 , . . . , K
where D denotes the number of decision variables, M denotes the
number of objectives, x = (x1 ,x2 ,…, xD ) is the feasible set solutions Solving MOPs, that is, finding the solutions on the Pareto front,
in the decision space, (f1 (x), f2 (x),…, fM (x)) denotes a solution that is difficult when using numerical methods (Cui, Geng, Zhu, & Han,
comprises M-objective functions in the objective space, gj (x) and 2017). Heuristic approaches such as Multi-objective Evolutionary
hk (x) are constraints of the problem. Algorithms (MOEAs), in contrast, can fairly easily discover an ap-
Solutions for this type of problem are not the best or extreme proximated Pareto front that is comprised of many non-dominated
solutions for any single objective but are the solutions that rep- solutions. Unfortunately, presenting the whole set of solutions to
resent a reasonable trade-off among all objectives. They are called the decision makers (DMs) not only increases their workload but
non-dominated or Pareto-optimal solutions. The Pareto front is a also provides them with many irrelevant solutions (Li, Deb, & Yao,
graphical presentation of these solutions. 2018). DMs cannot select preferred solutions from such a large
The concept is depicted in Fig. 1. The decision and objective set of possibilities. Hence, for the last decade, many researchers
spaces are the domain and range of the objective functions, re- have focused on finding techniques to help the DM select the pre-
spectively. To construct the objective space, we assign the range ferred solutions from the full set of Pareto-optimal solutions. We
of each function to an axis which, in this case, are f1 and f2 for the can broadly divide these techniques into two categories: explicit
first and the second objectives, respectively. A potentially unlim- preferences and implicit preferences.
ited number of feasible solutions exist in the objective space and In explicit preference methods, a DM must articulate his or her
the Pareto front. preference by providing hints to the preference-based MOEA. This
The non-dominated solutions are based on the Pareto domi- usually involves identifying a region of interest (ROI) in terms of
nance concept. Solution A dominates solution B, represented as a goal point or preferred points in the objective space. The algo-
A≺B, if and only if solution A is superior or as good as solution rithms will then attempt to prune out solutions that are far away
B in all M objectives and there is at least one objective for which from the goal point. Examples of these techniques are presented in
solution A is superior to solution B. For a minimization problem, Hu, Yu, Zheng, and Zou (2017), Liu, Li, Feng, Yu, and Jiao (2018).
the concept can be formalized as in (2). It is worth noting that we In these preference-based techniques, the quality of solutions is
focus only on minimization problems since maximization of x can directly proportional to the level of the DM’s knowledge. The DM
be easily transformed to a minimization of -x. needs to have significant prior knowledge about the preferred so-
lutions, expressed in terms of a region of interest, reference point,
A ≺ B if (∀i ∈ {1...M}, fi (A ) ≤ fi (B )) ∧ (∃i ∈ {1...M}, fi (A ) < fi (B ) )
or trade-off among the objectives. In the real world, many DMs
(2) have limited knowledge about the problems to be solved; they
The concept is illustrated in a bi-objective space in Fig. 2. S2 is cannot provide accurate preferences. Therefore, the utility of these
larger than S1 in both objective values, so S1 ≺S2 . S3 has the same explicit preference-based MOEAs is limited.
value as S1 for the second objective but for the first objective it On the other hand, implicit preference algorithms can prune
is larger, so S1 ≺S3 . S2 cannot dominate S3 , and vice versa, because out most ordinary solutions from Pareto fronts without a prefer-
one objective has a smaller value while the other objective has a ence specified by the DM. The solutions remaining after this prun-
larger value. The dominated area of S1 is visualized by the grey ing are outstanding solutions. For many problems, these tend to be
area. so-called knee solutions, that is, solutions where a minor enhance-
ment in one objective will cause severe degradation in another ob-
jective. The concept of a knee solution can be depicted as in Fig. 3,
showing that the knee solution is at the convex bulge of the graph.
It is an outstanding solution.

2. Existing approaches

In the last two decades, techniques for generating implicit pref-

erences in the form of knee-finding algorithms have received sig-
nificant interest since DMs cannot always specify their prior pref-
erences. Most techniques are posterior; that is, the algorithms take
the non-dominated solutions from MOEAs or exact methods as the
input and present the knee solutions to the DMs. Noteworthy ap-
Fig. 1. The relationship between decision and objective space. proaches are briefly reviewed below.
 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446 3

Vector adaptation based Evolutionary Algorithm (KnRVEA) that hy-

bridized of two MOEAs: KnEA (Zhang et al., 2015) and Reference
Vector guided Evolutionary Algorithm (RVEA) (Cheng, Jin, Olhofer,
& Sendhoff, 2016). For KnEA, KnRVEA borrowed the knee points
searching technique, while for RVEA, KnRVEA borrowed the de-
composition technique into sub problems:
The quality metric focus is to assign a real number score for
each solution on the Pareto front by using scalarization func-
tions. The extreme solution, maximum or minimum value de-
pends on each algorithm, is the knee point solution. Deb and
Gupta (2011) introduced two scalarization functions based on the
bend-angle and trade-off that calculated by using two surrounding
neighbors. The solution with the highest bend-angle or trade-off
Fig. 3. The concept of knee solution.
is the knee solution. Shukla, Braun, and Schmeck (2014) defined
two scalarization functions by summing objective values together
and linear utility function that adapted from Von Neumann and
The angle-based focus is based on the angles between an indi- Morgenstern (1944). The lower, the better the score. Braun, Shukla,
cated non-dominated solution and some others. Branke, Kaußler, and Schmeck (2017) defined a scalarization function by modifying
and Schmeck (2001) extended Pareto dominance to reduce the ASA (Sudeng & Wattanapongsakorn, 2016b). The lower, the bet-
number of dominated solutions, by extending the virtual axes from ter the score. Braun et al. (2017) introduced a Nash bargaining
the typical right angle to an obtuse angle. Branke et al. (2004) pro- solution that based on Nash equilibrium concept in Economics,
posed a knee finding algorithm based on the angles between Chebyshev, proper utility, and the sum of objective values as the
two vectors that are created from two closest neighbors of each metrics for their knee finding ESPEA algorithm. Ye, Liu, Yin, and
direction on the given Pareto front. The widest angle among Jin (2017) used hypervolume (HV) as the metric for their IKNEA
four is assigned as the score of that solution. The solution that algorithm.
has the highest score is considered as the knee point. Deb and Based on our literature review, all the reviewed algorithms can
Gupta (2011) introduced the bending angle technique that cal- identify the same knee regions of Pareto fronts. However, we can
culates the bending angles of two closest adjacent neighbor so- identify some shared weaknesses among existing algorithms.
lutions of an existing Pareto front. The knee point is the solu-
1) Most of them can produce quality solutions only if the DMs can
tion that has the highest bending angle. Leesutthipornchai and
specify suitable parameter values to the algorithms.
Wattanapongsakorn (2013) proposed a solution pruning algorithm
2) Most of them also include several no-knee solutions along with
called ADaptive Angle based pruning algorithm (ADA) which auto-
the knee solutions, which makes it difficult to identify the knee
matically adjusts the extended angles of dominance. Sudeng and
solutions.
Wattanapongsakorn (2015) enhanced the flexibility of the above
3) None of the algorithms above except
concept by introducing a pruning parameter for DMs to control
Bhattacharjee et al. (2017) allow DMs to specify the num-
the number of solutions, creating a method called Angle based in a
ber of solutions they need from the algorithms.
Specific bias parameter pruning Algorithm (ASA). Sudeng and Wat-
4) None of them can rank the selected solutions in terms of qual-
tanapongsakorn (2016b) integrated ASA with Non-dominated Sort-
ity.
ing Genetic Algorithm (NSGA-II) (Deb, Pratap, Agarwal, & Meyari-
van, 2002) to create an algorithm called knee-ASA-NSGA-II (k-ASA- It is worth noting that all knee-based benchmark func-
NSGA-II). Ramirez-Atencia et al. (2017) proposed the Knee-Point tions/problems that the algorithms in the literature review used
based (KPNSGA-II) that used the same concept as above, but in- do not have any mathematical rigor metric. The knee solutions are
stead, let decision makers directly specify the size of extended an- only identified by visualization; however, they locate at the convex
gles. bulge on a Pareto front, which can be identified easily.
The utility-based focus is based on the trade-off metrics that
are calculated by using the trade-off between objective functions 3. Proposed algorithm: RADA
since the knee point is a small gain of an objective but causes
a great loss of some objectives. Branke et al. (2004) proposed a In the proposed algorithm, RADA (Ranked ADA), we introduce
knee finding algorithm called Expected Marginal Utility. It calcu- a novel concept for solution ranking based on knee solutions. The
lates the cost between the second best individual and the best parameter-free algorithm accepts approximated non-dominated so-
individual of each objective by using weighted sum technique, lutions from any MOEA as the input and able to work with discrete
and the solution with the greatest difference is the knee solution. and continuous optimization problems. The main contributions of
Bhattacharjee, Singh, Ryan, and Ray (2017) proposed EMUr that di- RADA are as follows.
vides the Pareto front into regions and uses EMU to identify the • Every solution in the approximated/true Pareto set is ranked by
best solution in each region. The knee solutions are the solutions
quality. So the DMs will have a guideline for choosing a few
that completely better than the EMU values of neighbor regions
solutions to implement.
The maximum distance from the hyperplane focus was intro- • The DMs can choose any top K solutions to implement with low
duced by Das (1999). The concept is to identify the shortest dis-
computational complexity.
tance between each non-dominated solution to a predefined hy- • The solution ranking technique is also based on diversification,
perplane. The solution with the longest distance becomes the knee
so the DMs will have a variety of solutions to implement.
solution. Zhang, Tian, and Jin (2015) introduced Knee Point-Driven • The algorithm is parameter-free; no parameter setting is re-
Evolutionary Algorithm (KnEA) which integrated the above tech-
quested from the DMs. So no tedious process of trial-and-error
nique with NSGA-II from Deb et al. (2002) while Sudeng and
is required for the best performance.
Wattanapongsakorn (2016a) also combined the above technique
with MOEA/D-DE (Li & Zhang, 2009) to create k-MOEA/D-DE. Solution ranking of the algorithm is based on knee solutions;
Dhiman and Kumar (2019) proposed Knee point, and Reference the knee solutions tend to have better ranks while the solutions
4 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446

Algorithm RADA

Input: FN,M (the N unique nondominated solutions for M objectives)

Output: f score1...N (the ranked scores of F)
1: Variables Initialization
2: Normalize F1...N,1...M to Fˆ1...N,1...M
3: Calculate the Threshold Angles
4: for i ← 1...N
5: for j ← 1...N
6: if i = j then
7: for m ← 1...M
8: if solution i angle-dominates solution j
9: calculate and accumulate the angle-dominated scores of solution j
10: end if
11: end for
12: end if
13: end for
14: end for
15: for j ← 1...N
16: calculate Final Score f score j
17: end for
18: return f score1...N

that are too close to the knee solutions are degraded, so the diver-
sity of the top solutions are well-preserved. The RADA algorithm
Table Algorithm is outlined and presented below. In the following
sections, the main concepts of the algorithm are explained step by
step.

3.1. Variable initialization (line 1)

Initialize the following variables to zero:

• Total Angle Penalty Score (TAPS1...N ) is the sum of positive er-

ror scores that are accumulated to a solution if the solution is Fig. 4. Angle Penalty Degree concept.
angle-dominated by other solutions.
• Angle-dominated Times Count (ATC1...N ) is the number of times
that a solution is angle-dominated by other solutions. 3.3. Angle-dominance determination (line 8)

The F1...N,1...M are the non-dominated input solutions. The algo- The extended angle of solution j with respect to solution i at
rithm requires that the objective function values be normalized in objective m (θ i,j, m ) can be calculated as follows (5).
the range of zero to one. This is accomplished using the following ⎡   2
⎤
M
formula (3). a=1,a=m Fˆj,a − Fˆi,a
θi, j,m = tan−1 ⎣ ⎦ (5)
Fˆj,m − Fˆi,m
Fn,m − min(F1...N,m )
Fˆn,m = , where n = 1 . . . N; m = 1 . . . M
max(F1...N,m ) − min(F1...N,m ) Solution i angle-dominates solution j for objective m, denoted
(3) ≺
i j, if the extended angle of solution i to solution j for ob-
m
where n is the solution number, N is a number of solutions, m is
jective m, θ i,j, m , is smaller than the threshold angle δ m for at least
the objective number, M is the number of objectives.
one objective m. This constraint can be formulated as follows (6).
≺
i j if ( θi, j,m < δm (6)
3.2. Threshold angle calculation (line 2) m
The angle-dominance concept can be depicted in Fig. 4. Con-
The threshold angles (δ 1...M ) are the angles that defined the
sidering solutions B and C, they are not dominated by solution
fixed scope of extended angles beyond the typical of Pareto dom-
A by the concept of Pareto Dominance, which was described in
inance. The wider threshold angles, the more solutions can be
(2). However, solution B is angle-dominated by solution A as the
angle-dominated. The threshold angle δ m can be calculated as fol-
extended angle θ A,B,1 < δ 1 . Likewise, θ A,C,2 < δ 2 solution A also
lows (4)(4).
≺ ≺
 IQ S  angle-dominates solution C. Therefore, A
1
B and A
2
C are true.
δm = tan−1 m
, m = 1...M (4)
For the concept of ADA, it performs a round-robin tournament
I Qm
to compare every objective by using the angle-dominance concept.
where m is the objective number, M is a number of objectives. Each angle-dominated solution will be eliminated and the remain-
IQSm is the inter-quartile of sorted values of each objective in ders are knee solutions. However, for RADA, the angle-dominated
Fˆ1..N,m . The IQm is the average distance between two consecutive solutions are not eliminated but accumulated with the penalty
solutions in Fˆ1...N,m . The function tan−1 is an arctangent function. scores instead.
 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446 5

3.4. Penalty score calculation and accumulation (line 9)

The Angle Penalty Score (APS) of solution j for objective m by

solution i, APSi,j,m , can be calculated by multiplying two quantities,
Angle Penalty Degree (APD) and Penalty Intensity Rate (PIR) as fol-
lows (7).
AP Si, j,m = AP Di, j,m × P IRi, j,m (7)
The Angle Penalty Score, APDi,j, m , is the penalty to solution j
when angle-dominated by solution i for objective m. The value can
be calculated as the difference between δ m and θ i,j,m . While the
Penalty Intensity Rate, PIRi,j, m , is a coefficient value between zero Fig. 5. Luminous Intensity (Voudoukis, 2017).
and one that is used for reducing the effect of APDi,j, m . The two
values can be calculated as follows (8), (9).
value of PIR for solution B than for solution C, because of the dis-
≺
δm − θi, j,m if i j, tance from solutions A to B (d1 ) is longer than the distance from
AP Di, j,m = m (8)
solutions A to C (d2 ). The longer the distance between the domi-
0 otherwise
nating and dominated solutions, the smaller the value of PIR and
1 consequently, the smaller the resulting APS. In other words, solu-
P IRi, j,m =  2
(9) tions with longer distances from a particular solution will receive
Fˆj,m −Fˆi,m lower penalties and potentially higher ranks.
max Fˆa,m − min Fˆa,m
a=1...N a=1...N For example, suppose we have only three solutions in a solu-
tion set {So ,S5 ,S30 } and that solution S0 angle-dominates solutions
In the round-robin tournament, the two penalty scores of solu-
S5 and S30 with distances of 5 and 30 units, respectively. Even
tion j are added, which are Total Angle Penalty Score (TAPS j ) and
though both solutions have the same APD values, S30 should be
Angle-dominated Time Count (ATCj ). They are calculated as follows
preferred over the solution S5 because it is more different (more
(10), (11).
diverse) from the solution S0 . Diversity in the obtained solutions is
TAPS j ← TAPS j + AP Si, j,m (10) mentioned here.
The concept is depicted in Fig. 4 demonstrating that solution A
angle-dominates solutions B and C. The APDA,B and APDA,C are cal-
AT C j ← AT C j + 1 (11)
culated according to (8). APDA,B comprises APDA,B, f1 and APDA,B, f2 ,
In ADA, all angle-dominated solutions are eliminated, the re- while APDA,C comprises APDA,C, f1 and APDA,C, f2 . APDA,B, f2 and
mainders are knee solutions. In RADA, we use TAPS and ATC to APDA,C, f1 are zeros as solution A is no angle-dominated by
calculate the final penalty score. The lower score, the better rank. any other solutions on those objectives. Therefore, APDA,B, f1 and
The solutions with ATC = 0 and implicitly TAPS = 0 are the same APDA,C, f2 represent APDA,B and APDA,C , respectively. The APDA,B, f1
as knee solutions in ADA as they are not angle-dominated by any > APDA,C, f2 due to δ − θ 1 is slightly higher than δ − θ 2 . Con-
other solution. We use the term True Outstanding Solutions (TOS) versely, APSA,C tends to be higher than APSA,B as solution B is sig-
for those solutions, and they are at 1st rank. While the inferior so- nificantly far from solution A than solution C, because the APSA,B
lutions will be ranked afterward, based on the TAPS and ATC. is greatly reduced by PIRA,B .
TAPS is based on APS and APS is based on penalty score APD.
However, if we use APD alone, without the distance factor, to cal- 3.5. Final rank score calculation (lines 16 to 18)
culate APS, most of the top-rank solutions will be close together,
clustered around the TOS as the solutions near the TOS will mas- To produce fscore1...N , both ATC1...N and TAPS1...N are normalized
sively increase the APD for other solutions nearly the same level as C
into the range of zero to one and symbolized as AT 1...N and
the TOS does.
TˆAPS1...N , respectively. The normalized values are calculated as fol-
Generally, DMs want a variety of options from which to choose.
lows (12), (13).
For this reason, we also consider Penalty Intensity Rate (PIR), a co-

efficient value in the range between zero and one, based on the TAPS j − min TAPS1...N
distance between the dominator and alternative solutions. The far- TˆAPS j =   (12)
max TAPS1...N − min TAPS1...N
ther a solution is from the dominator solution, the lower the value
of PIR. This makes the overall APS for the angle-dominated solu-
tion lower, raising its rank. As a result, the diversity of solutions C = AT C j − min(AT C1...N )
AT j (13)
is promoted. By incorporating PIR, the algorithm provides a larger max(AT C1...N ) − min(AT C1...N )
variety of choices rather than focusing only on near TOS solutions.
The fscore1...N are the final scores with the range between zero
To calculate PIR, we use an analogy to the relationship between
and one. The lower value, the higher rank. The scores can be cal-
distance and light intensity, as shown in Fig. 5. The dominator so-
culated as follows (14).
lution is analogous to the light source; the light intensity is in-
versely proportional to the square of the distance from the light C
fscore j = TˆAPS j × AT (14)
j
source (Voudoukis, 2017). Hence, PIR will be dramatically reduced
if angle-dominator and angle-dominated solutions are in the dis- For the 1st rank solution, fscore = 0, there are two possible
tance. cases: with TOS and with non-TOS. If the solution j is a TOS, then
C = 0 as it cannot be angle-dominated by any
TˆAPS j = 0 and AT
For example, as shown in Fig. 5, if the distance from the light j
source is three units, the illuminated area increases by nine times other solutions. In this case, only the TOS will have the 1st rank.
and the intensity at any point is decreased to 1/9. Mathematically, Conversely, in case of no TOS, the 1st rank may come the different
the intensity is expressed as a function of the distance r as fol- solutions with either the lowest TAPS or ATC then causing TˆAPS or
lows: f (r ) = 1/r 2 . By analogy, in Fig. 4, solution A produces a lower C , respectively, to zero. In this case, RADA ranks them equally.
AT
6 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446

3.6. Computational complexity Table 1

Test problems

In Algorithm RADA, the normalization from F1...N,1...M to ZDT3 min f1 (x) = x1 

Fˆ1...N,1...M in line 2 requires O(MN). While the threshold angles min f 2 (x ) = g(x )(1 − x1
g( x )
− g(x )sinx(110π x ) )
1

calculation line 3 requires a sorting algorithm to order the data 

D
g( x ) = ( 1 + 9
xi )
of each objective, so the computational complexity is O(MN2 ) for D−1
i=2
0 ≤ xi ≤ 1, i = 1, …, D
the worst-case scenario of Quicksort. The complexity of the main
DEB2DK min f 1 (x ) = g(x )r (x1 )sin(π x21 )
loop, the penalty accumulation, in lines 4-14 requires O(MN2 ) min f 2 (x ) = g(x )r (x1 )cos(π x21 )
because every solution must be compared to every other solu- 9 D
g(x ) = (1 + D−1 i=2 xi )
tion. The normalization of final scores in lines 15-17 requires r (x1 ) = 5 + 10(x1 − 0.5)2 + 1k cos(2kπ x1 )
O(N). Hence, the overall worst-case complexity of the RADA is 0 ≤ xi ≤ 1, i = 1, …, D
DEB3DK min f 1 (x ) = g(x )r (x1 , x2 )sin(π x21 )sin(π x22 )
O(MN) + O(MN2 ) + O(MN2 ) + O(N) ≈ O(MN2 ).
min f 2 (x ) = g(x )r (x1 , x2 )sin(π x21 )cos(π x22 )
min f 3 (x ) = g(x )r (x1 , x2 ) cos(π x21 )
9 D
g(x ) = (1 + D−1 i=2 xi )
r (x1 , x2 ) = r1 (x1 )+2 r2 (x2 )
4. Experimental setup and results ri (xi ) = 5 + 10(xi − 0.5 )2 + 2k cos(2kπ xi )
0 ≤ xi ≤ 1, i = 1, …, D
In this section, we describe experiments conducted on well-
known benchmark problems: ZDT3 (Zitzler, Deb, & Thiele, 20 0 0),
DEB2DK and DEB3DK (Branke et al., 2004). DEB2DK and DEB3DK
problems have a parameter called k for customizing the number The optimal solutions are the solutions that uniformly dis-
of knees in the problems. For DEB2DK, parameter k represents the tributed on an optimal Pareto front. This is the ideal input for the
numbers of knees in each problem. We will explore the algorithms’ ranking system of RADA They have been generated by a straight-
performance by varying the numbers of knees from one to nine. forward exact algorithm by using MATLAB. The algorithm used lin-
For DEB3DK, number of knees in a problem is k2 , we will conduct early spaced series of each decision variable as the input; in this
two tests for k is one and two, which makes the number of knees case, the number of decision variables was set to D = 2. The feasi-
in the problems to one and four, respectively. Although ZDT3 does ble solutions were evaluated on the objective space. Then, the al-
not have a knee customization parameter, we can say that it has gorithm performed the pairwise comparison by using Pareto dom-
five knees because of the nature of the Pareto front, which is di- inance concept that described in (2) to eliminate the dominated
vided into five obvious sections. The test functions/problems are solutions. At a result, the non-dominated solutions are the optimal
presented in Table 1. For RADA, we set the number of outstand- solutions. The size of linearly spaced series in each problem is a
ing solutions K to the number of knee solutions plus three. This trail-and-error to produce roughly 50 0 and 1,50 0 optimal solutions
range was sufficient for RADA to discover all the knee solutions in for a two-objective and three-objective optimization problems, re-
most of the problems. In our result tables, the ranking numbers spectively.
are shown beside the outstanding solutions. It is possible for two The approximated solutions are the output solutions from an
solutions to have the same fscore and hence the same rank. MOEA. The benchmark test problems were implemented in the
We will demonstrate the testing of RADA against four pos- PlatEMO (Tian, Cheng, Zhang, & Jin, 2017). Two_Arch2 (Wang, Jiao,
teriori knee finding algorithms: ADA (Leesutthipornchai & Wat- & Yao, 2015) is chosen as the MOEA for solving these test prob-
tanapongsakorn, 2013), ASA (Sudeng & Wattanapongsakorn, 2015), lems. All parameters except the number of solutions and num-
EMUr (Bhattacharjee et al., 2017), and EMU (Branke et al., 2004). ber of decision variables are the default values from PlatEMO. The
The implementation is performed in MATLAB. As none of the al- number of solutions is set to 100 and 400 for two-objective and
gorithms can perform ranking, the number of obtained solutions three-objective optimization problems, respectively. While a num-
and the number of discovered knees are used to judge the perfor- ber of decision variables is set to D = 2 to match the optimal solu-
mance, instead. tions and the number of objective function calls was set to 10,0 0 0
The experiments were executed on a Windows 10 × 64 com- for ZDT3 and DEB2DK, and 10 0,0 0 0 for DEB3DK. The size, genera-
puter with a 2.3 GHz Intel i7-4712MQ CPU, 8GB RAM. Two sets of tion time of Pareto fronts of the optimal set and approximated set
experiments, optimal solutions and approximated solutions, have for each benchmark problem along with RADA ranking times are
been conducted on each benchmark problem. shown in Table 2. From the table, it is worth noting that the gen-

Table 2
Pareto fronts’ size, generation time, and time for RADA used for ranking of optimal and approximated set for benchmark problems:
ZDT3, DEB2DK with number of knees = 1 to 9 and DEB3DK with number of knees = 1 to 2 when the number of decision variables
D = 2.

Benchmark Optimal set Approximated set

problems
PF size Generation time (secs) RADA (secs) PF size Generation time (secs) RADA (secs)

ZDT3 533 2.33 0.14 100 6.12 0.08

DEB2DK 1 500 2.20 1.56 100 6.21 0.08
2 500 2.06 1.52 100 6.09 0.08
3 500 2.10 1.57 100 6.09 0.08
4 500 2.12 1.55 100 6.45 0.08
5 500 2.07 1.52 100 6.29 0.08
6 500 2.06 1.53 100 5.51 0.08
7 500 2.06 1.56 100 6.28 0.08
8 500 2.12 1.57 100 6.12 0.08
9 500 2.10 1.55 100 6.08 0.08
DEB3DK 1 1519 0.63 18.49 400 251.66 1.29
2 1611 0.63 20.21 400 242.78 1.28
 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446 7

Table 3
Parameter setting.

Algorithms Parameters

RADA -
ADA -
ASA tau = 0.94
EMUr W = PF Size / 2, k = number of knees
EMU |λ| = PF Size / 2

eration time for optimal set is greatly lower than the time for the
approximated set. This is because the number of decision variables
is two which is very low. In the real world, the number of deci-
sion variables can be significantly higher than two, and the exact
method will be impractical to use since computational complexity
is exponentially grown by the number of decision variables.
The parameters setting for each algorithm is shown in Table 3.
For RADA and ADA, there is no parameter to set. While ASA, the
level of pruning tau = 0.94, is chosen as it can identify all knees
in each problem. For EMUr , W, the number of reference directions
Fig. 6. The results on the optimal solution set of ZDT3.
is set to the half of PF size the same as was used in the original
paper, and k is set to a number of knees in each problem. For EMU,
the number uniform preference vector, |λ|, is set to half of Pareto-
optimal Front (PF) Size by trial-and-error. The same obtained opti-
mal or approximated solution sets of each algorithm are shown in
cascading style next to each other for knee finding comparison for
ZDT3 and DEB2DK test problems in the following sections.

4.1. ZDT3

ZDT3 is a well-known disconnected two-objective test problem.

Although ZDT3 is not explictly classified as a knee test problem,
its solution has five sections that can be treated as five knees. We
set the top K outstanding solutions to eight, that is, the number of
sections plus three.
Table 4 and Fig. 6 present the top eight outstanding solutions
from RADA against ZDT3 optimal solution set. The figure reveals
that all five knee solutions can be discovered within the top six
ranks. The table reveals that there is one true outstanding solution
(with ATC = 0) in the second section from the left. Interestingly,
the 2nd rank has 59 dominated angles (with ATC = 59), but it has
a higher rank than the 5th rank that has only one dominated angle Fig. 7. The results of the approximated solution set of ZDT3.
(with ATC = 1). However, the distance of the 5th rank is 6.08E−04,
which is much closer to the dominating solution that other better
outstanding solutions. Hence the fscore is dramatically increased. Table 5 and Fig. 7 present the top eight outstanding solutions
Fig. 6 presents RADA in comparison with the referenced algo- from RADA against ZDT3 approximated solution set. The figure re-
rithms. RADA can identify all knees within six outstanding solu- veals that all five knee solutions can be discovered within the top
tions, while ADA has only one obtained solution and can identify five ranks. It has one true outstanding solution (with ATC = 0)
only one knee. EMUr gives 12 solutions and cannot identify the in the second section of the solution set (counting from the left
lower right knee solution. For ASA and EMU, they can identify all hand side) same as in the optimal solution set shown in Fig. 6.
knees solutions with 61 and 70 obtained solutions, respectively. In In addition to the 6th rank solution, the situation is same as the
this case, RADA outperforms all referenced algorithms. These re- 5th rank shown in Table 3 that has only one dominated angle
sults are shown in Table 8. (with ATC = 1) but with a very low distance which in this case

Table 4
The eight outstanding solutions obtained by RADA on the optimal solution set of ZDT3.

APD Distance
rank fscore C
AT TˆAPS ATC
Mean s.d. mean s.d.

1 0 0 0 0 - - - -
2 1.43E−06 1.11E−01 1.29E−05 59 4.42E−02 1.64E−02 2.33E−01 1.76E−02
3 1.48E−06 1.15E−01 1.29E−05 61 4.26E−02 2.01E−02 2.49E−01 3.92E−02
4 1.45E−05 5.30E−01 2.73E−05 282 4.49E−02 3.07E−02 5.52E−01 1.74E−01
5 1.89E−05 1.88E−03 1.00E−02 1 1.32E−02 - 6.08E−04 -
6 2.70E−05 7.80E−01 3.47E−05 415 6.38E−02 3.49E−02 7.24E−01 2.35E−01
7 5.10E−05 3.76E−03 1.36E−02 2 8.85E−03 1.06E−02 8.81E−04 4.15E−04
8 5.56E−05 1.13E−01 4.93E−04 60 4.35E−02 1.73E−02 2.29E−01 3.47E−02
8 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446

Table 5
The eight outstanding solutions obtained by RADA on the approximated solution set of ZDT3.

APD Distance
rank fscore C
AT TˆAPS ATC
Mean s.d. mean s.d.

1 0 0 0 0 - - - -
2 1.51E−05 6.06E−02 2.49E−04 6 2.70E−02 1.44E−02 2.34E−01 1.68E−02
3 1.35E−04 1.82E−01 7.45E−04 18 3.15E−02 2.83E−02 3.62E−01 1.11E−01
4 5.07E−04 4.24E−01 1.19E−03 42 4.81E−02 3.29E−02 5.50E−01 1.80E−01
5 1.67E−03 7.47E−01 2.23E−03 74 6.66E−02 3.91E−02 6.77E−01 2.46E−01
6 2.04E−03 1.01E−02 2.02E−01 1 1.63E−01 - 8.44E−03 -
7 4.17E−03 2.02E−02 2.06E−01 2 9.00E−02 1.16E−01 1.15E−02 4.09E−03
8 5.18E−03 7.58E−01 6.83E−03 75 6.67E−02 4.03E−02 6.62E−01 2.56E−01

Table 6
The outstanding solutions obtained by RADA on the optimal solution sets of DEB3DK (k = 1 and k = 2).

APD Distance
k rank fscore C
AT TˆAPS ATC
mean s.d. mean s.d.

1 1 (2) 0 0 9.91E−03 4 4.91E−02 2.23E−02 9.65E−03 3.75E−03

1 (2) 0 9.26E−01 0 1277 1.87E−01 1.09E−01 7.80E−01 1.68E−01
5 (2) 2.84E−06 3.49E−02 8.14E−05 52 1.74E−02 1.52E−02 3.69E−01 1.69E−01
2 1(16) 0 0 0 0 - - - -
17 3.75E−10 2.52E−03 1.49E−07 2 2.19E−04 0 3.27E−01 0

Note: The numbers in parenthesis represent the number of solutions with the same results.

is 8.44E−3, so the obtained fscore is higher than those of the rank the convention to 6, because the all four solutions have the same
1st to 5th. rank and the first non-zero fscore is worth to be discovered which
Fig. 7 presents RADA in comparison with the referenced algo- are the pair solutions of 5th and 6th ranks. Remarkably, The sec-
rithms. RADA can identify all five knees by using five outstanding ond pairs (5th and 6th) have much better dominated angles than
solutions. The results are analogous to the optimal set with the the first pairs (3rd and 4th) but they have the lowest total angle-
number of solutions for ADA, ASA, EMUr , and EMU are 1, 11, 6, dominated scores (TAPS ), so after the normalization, the TˆAPS be-
and 13. Respectively, as shown in Table 9. comes zero and then causes fscore to zero. For k = 2, the first six-
teen solutions are all true outstanding solutions. They successfully
4.2. DEB2DK k = 1 to 9 span the four knees. Therefore, the number of top outstanding so-
lutions is changed from k + 3 which is 4 + 3 = 7 to 17 as the 17th
We tested the RADA algorithm against DEB2DK for k = 1 to 9. rank is the lowest non-zero fscore .
The number of top K solutions for each problem is set to k + 3. In the comparation of algorithms, the results obtained from
Each optimal solution set and approximated solution set comprise other knee finding algorithms against DEB3DK k = 1 and k = 2
50 0 and 10 0 solutions, respectively. The results of optimal sets and are depicted in Figs. 11 and 12, respectively. From Fig. 11, it is not
approximated sets are depicted in Figs. 8 and 9, respectively which a surprise that ADA cannot identify any knee solution as it has no
show that RADA can discover every knee in each problem within TOS found in the problem. The other algorithms: ASA, EMUr , and
the given range. It is worth noting that some approximated sets EMU can identify the single knee in the middle with 6, 13, and
are not on the optimal Pareto fronts, especially for k = 2 and k = 5. 17 obtained solutions, respectively. For RADA, the 1st rank com-
However, the shapes of Pareto front are well-preserved. prises four solutions; therefore, RADA can outperform other refer-
In comparison, RADA and EMUr outperform all referenced enced algorithms in the test. From Fig. 12, the number of solutions
algorithms. The details of the comparasions are described in of ADA, ASA, EMUr , and EMU are 16, 79, 18, and 10, respectively.
Section 4.4. All algorithms can identify all four knees in the problem, so EMU
leads for the test, as shown in Table 8.
The results from approximated solution sets are shown in
4.3. DEB3DK
Table 7. Fig. 10c displays the results of RADA for DEB3DK with
k = 1 and Fig. 10d for DEB3DK with k = 2. Table 7 with k = 1
DEB3DK is a de facto standard for three-objective knee finding
shows that the top four outstanding solutions as number of knees
test problems. Similar to DEB2DK, DEB3DK also has parameter k
plus three. There are two TOSs for this test at the knee region
for researchers to vary the number of knees. The number of knee
and the other at the middle bottom of the Pareto front. The 3rd
solutions is k2 . We have tested RADA against DEB3DK for k = 1
rank is in the knee region and the 4th rank are in the same re-
and k = 2 problems, which have one and four knees, respectively.
gion as the 2nd rank. For k = 2, the number of knees is k2 = 4.
The outstanding results from optimal solution sets are shown
Therefore, the number of top solutions is set to 7 as the number
in Table 6. Fig. 10a and b display the results of RADA for DEB3DK
of knees plus three. The first sixteen solutions are all true out-
with k = 1 and k = 2, respectively. For k = 1, RADA gives the
standing solutions. They successfully span the four knees. There-
top four outstanding solutions with the same rank. There is no
fore, the number of top outstanding solutions is changed from
true outstanding solution for this test. However, the first two so-
k + 3 which is 4 + 3 = 7 to 17 as the 17th rank is the low-
lutions have the lowest dominated angles (with ATC = 4), which
C zero. This pair of solutions is in the middle of the fig- est non-zero fscore . Table 7 reveals that there are four true out-
makes AT
standing solutions, and Fig. 10d shows that the four solutions can
ure, the knee region. The second pair of top solutions, which are
discover three knees while the 6th rank can discover the last
in the bottom left and right, have the lowest TAPS which makes
knee and the 7th rank discovers the most non-knee outstanding
the TˆAPS to be zero. Therefore, they are also the top rank solutions.
solution.
The number of top outstanding solutions are changed from 4 as
 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446 9

Fig. 8. The results on the optimal solution set of DEB2DK (k = 1 to 9).

Fig. 13 depicts the results from other refererenced algorithms k=1. For RADA, there are 4 TOSs that spans into three knees, which
against DEB3DK k = 1. All algorithms can identify a single knee so- left the other non-TOS knee. Therefore, ADA can discover only
lution. RADA and ADA make a tie for two solutions of TOSs. While three knees, and RADA needs 6th rank to discover the rest knee, it
ASA, EMUr , and EMU produce 9, 16, and 6, respectively. Fig. 14 de- is the winner of the test, although. The other algorithms have no
picts the results from other referenced algorithms against DEB3DK

Fig. 11. The results of other knee finding algorithms on the optimal solution set of DEB3DK (k = 1).
10 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446

Fig. 9. The results of the approximated solution set of DEB2DK (k = 1 to 9).

problem identifying all four knees with 17, 13, and 7 obtained so- EMUr is very impressive for DEB2DK test problem where the per-
lutions for ASA, EMUr, and EMU, respectively, as shown in Table 9. formance is comparable to RADA but not for ZDT and DEB3DK test
problems. EMU is weak for two-objective optimization problems,
but it is better for DEB3DK.
4.4. Performance evaluation The number of the test problems that each algorithm can out-
perform other algorithms against optimal sets is 7, 1, 0, 4, and
In this section, RADA and all four referenced knee-finding 1 for RADA, ADA, ASA, EMUr , and EMU, respectively. For the ap-
algrithms are evaluated in terms of number of obtained solutions proximated sets are 8, 2, 0, 6, and 0 in the same order, re-
and discovered knees. The obtained results discussed previously spectively. RADA, therefore, when considering all 24 test prob-
in Sections 4.1–4.3 are gathered and presented again in Tables 8 lems/cases, RADA outperforms other referenced algorithms by win-
and 9. In summary, only RADA and ASA can identify all knees in ning 15 out of 24 tests when compare to the next best algorithm,
every test problem, although ASA produces too many solutions. EMUr , that wins 10 out of 24 tests. Moreover, RADA can identify
ADA is the worst algorithm here becausem it focuses only on TOSs.

Fig. 12. The results of other knee finding algorithms on the optimal solution set of DEB3DK (k = 2).
 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446 11

Fig. 10. The outstanding solutions obtained by RADA on DEB3DK (k = 1, 2).

Table 7
The outstanding solutions obtained by RADA on the approximated solution sets of DEB3DK (k = 1 and k = 2).

APD Distance
k rank fscore C
AT TˆAPS ATC
mean s.d. mean s.d.

1 1 (2) 0 0 0 0 - - - -
3 8.49E−05 3.72E−03 2.28E−02 1 2.53E−02 - 3.58E−02 -
4 1.18E−04 3.72E−03 3.16E−02 1 3.53E−02 - 3.59E−02 -
2 1 (4) 0 0 0 0 - - - -
5 4.45E−07 4.08E−03 1.09E−04 1 9.16E−03 - 3.08E−01 -
6 2.12E−06 4.08E−03 5.20E−04 1 4.46E−03 - 9.85E−02 -
7 1.22E−05 1.63E−02 7.50E−04 4 1.67E−02 6.48E−03 3.14E−01 2.02E−02

Note: The numbers in parenthesis represent the number solutions with the same results.
12 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446

Fig. 13. The results of other knee finding algorithms on the approximated solution set of DEB3DK (k = 1).

Fig. 14. The results of other knee finding algorithms on the approximated solution set of DEB3DK (k = 2).

Table 8 all knees in every test problems with small number of outstanding
The number of solutions resulted from the knee finding algorithms solutions.
against optimal sets.

Algorithm k RADA ADA ASA EMUr EMU 5. System design and optimization problem

Problem = ZDT3 - 5 1 61 12 70
DEB2DK 1 1 1 22 6 48
5.1. Problem formulation
2 2 1 40 6 38
3 3 1 46 5 22 This section further explores the effectiveness of the RADA al-
4 5 1 52 4 18 gorithm by applying it to a well-known real-world application:
5 6 1 56 5 16
a system design and optimization problem called a system reli-
6 7 1 60 6 12
7 10 1 64 6 12 ability design and optimization (Coit & Smith, 1996; Coit, Tong-
8 10 1 64 8 12 dan, & Wattanapongsakorn, 2004; Khalili-Damghani, Abtahi, & Ta-
9 9 1 70 8 14 vana, 2013; Kuo & Wan, 2007). This NP-hard, mixed-integer op-
DEB3DK 1 4 0 6 13 17 timization problem aims to archive a system design that maxi-
2 16 16 79 18 10
mizes the system reliability, by using highly reliable components
Note: The underlines denote the tests that cannot identify all knees which is normally expensive or choosing redundant components
and the bold font values denote the algorithm that outperforms all with the lowest weights, costs and volumes. This kind of system
other algorithms.
is a basic requirement for mission critical applications such as
aircrafts, data centers, power plants, etc. (Rausand, 2014). There
Table 9
The number of solutions resulted from the knee finding algorithms are many algorithms designed to solve this problem such as Co-
against approximated sets. evolutionary Differential Evolution with Harmony Search (CDEHS)
(Wang & Li, 2012), heuristic survival signature based technique
Algorithm k RADA ADA ASA EMUr EMU
(Huang, Coolen, & Coolen-Maturi, 2019), Fuzzy Adaptive Particle
Problem = ZDT3 - 5 1 11 6 13 Swarm (FAPSO) (Samanta & Basu, 2019), Stochastic Perturbation
DEB2DK 1 1 1 3 2 5
2 2 1 9 3 6
Particle Swarm Optimization (SPPSO) (Ouyang, Liu, Ruan, & Jiang,
3 3 1 12 3 8 2019).
4 5 1 13 4 7 The system of interest is derived from Fyffe, Hines, and
5 5 1 17 6 7 Lee (1968). It comprises of several subsystems chaining in series.
6 6 1 18 6 7
Fig. 15a. shows an example of a small system with three sub-
7 8 1 20 7 6
8 10 1 21 8 6 systems connected in series. Each subsystem has its own level
9 9 1 24 9 7 of reliability: 0.95, 0.80, and 0.99, respectively. The reliability
DEB3DK 1 2 2 9 16 6 of a series-system is n(i=1 ) Ri ,. In this example, n = 3 repre-
2 6 4 17 13 7
sents the number of subsystems in the series and Ri = 0.95,
Note: The underlines denote the tests that cannot identify all knees 0.80, 0.99 represents the reliability of each subsystem, respec-
and the bold font values denote the algorithm that outperforms all tively. Therefore, the overall reliability of the series system is
other algorithms.
RT = 0.95 × 0.80 × 0.99 = 0.7524.
 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446 13

Fig. 15. Component diagram of example systems.

Fig. 17. The top six of outstanding solutions obtained by RADA (highlighted and
zoom in).
Fig. 16. The top six of outstanding solutions obtained by RADA.

where
A common technique for increasing the reliability is to ap-   yi
ply redundant components in parallel in an existing subsystem. Ri = 1 − 1 − Xi j Ri j
The reliability of a parallel-system is R = 1 − (1 − Riˆ)yi where Riˆ j  mi

represents the reliability of each component in subsystem i and


n 
yi represents the number of identical components in the subsys- CT = Xi jCosti j
tem i. In Fig. 15b, increasing the reliability of subsystem 2 by in- i=1 j mi
stalling a redundant component that makes yi = 2 and Riˆ = 0.80,
so Rsub2 = 1 − (1 − 0.80)2 = 0.96, which makes the overall relia- 
n 
WT = Xi jW eighti j
bility rise to RT = 0.95 × 0.96 × 0.99 = 0.9028.
i=1 j mi
Although component redundancy can increase the reliability of
a system, it also increases the costs of the system with extra hard-
Xi j = 0, 1; yi = 1, 2, 3, 4; i = 1, 2, . . . , n; m = A, B, C, D
ware, electricity, maintenance, space, weight, and so forth. In this
paper, we consider two objectives; reliability (f1 ) and cost of the where n = 7 is the number of subsystems, WT or A, B, C, D and
system (f2 ) with weight as a constraint. Ri are the selectable component types and reliability of subsystem
i, respectively. j is the component type, Costij and Weightij are the
5.2. Problem formulation and experimental design cost and weight of subsystem i using component type j that has yi
identical components allocated. The aim of the design is to max-
The problem includes seven subsystems; each subsystem may imize the system reliability (RT ) while minimizing the total sys-
have one component chosen from three (comp type A, B or C) or tem cost (CT ) with the total system weight (WT ) less than 150 as
four (comp type A, B, C or D) component types. Once the com- a constraint. The input parameters for the problem are shown in
ponent type is chosen, each subsystem may install one or up to Table 10.
three identical components for redundancy purposes. The model
of multi-objective optimization is as follows. 5.3. Results

n
Max f1 = RT = Ri We evaluated the Pareto front for the problem using an exact
i=1 method implemented in C# with total 84,934,656 possibilities. The
Min f2 = C T
front includes 58 Pareto-optimal solutions, and the generating time
Subject to was 192.51 seconds. Table 11 represents the single objective opti-
 mization for systems. We can notice that the extreme optimization
Xi j = 1, i = 1, 2, . . . , n for a single objective give a very poor result for the other objective.
j  mi Then, we use RADA to rank all systems. The Pareto front and
the top six outstanding systems based on RADA are shown in
W T < 150 Fig. 16 and all highlighted outstanding solutions are shown in
14 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446

Table 10
Component input data (Fyffe et al., 1968).

Design Alternative (j)

Comp type A Comp type B Comp type C Comp type D

Sub j
R C W R C W R C W R C W

1 0.90 1 3 0.93 1 4 0.91 2 2 0.95 2 5

2 0.95 2 8 0.94 1 10 0.93 1 9 - - -
3 0.85 2 7 0.90 3 5 0.87 1 6 0.92 4 4
4 0.83 3 5 0.87 4 6 0.85 5 4 - - -
5 0.94 2 4 0.93 2 3 0.95 3 5 - - -
6 0.99 3 5 0.98 3 4 0.97 2 5 0.96 2 4
7 0.91 4 7 0.92 4 8 0.94 5 9 - - -

Table 11
The solutions that are optimized by one objective at a time, RT or CT .

Subsystem Objective
optimize
1 2 3 4 5 6 7 RT CT WT

R 4B 4A 4D 4B 4C 3B 4C 0.9996 81 148
C 1B 1B 1C 1A 1A 1C 1B 0.5295 14 42

Table 12
The top six outstanding solutions obtained by RADA.

Subsystem Objective
Rank
1 2 3 4 5 6 7 RT CT WT

1 3B 3B 3C 3A 2A 2C 2B 0.9816 34 109
2 3B 3B 4C 3A 2A 2C 2B 0.9834 35 115
3 3B 3B 4C 3A 3A 2C 2B 0.9868 37 119
4 3B 3B 3C 3A 3A 2C 2B 0.9849 36 113
5 3B 3B 4C 3A 3A 2C 2C 0.9896 39 121
6 3B 3B 4C 3A 3A 2C 3B 0.9926 41 127

Table 13
The specification of the top six of outstanding solutions.

APD Distance
rank fscore C
AT TˆAPS ATC
Mean s.d. Mean s.d.

1 0 0 0 0 - - - -
2 1.01E−03 1.75E−02 5.77E−02 1 6.48E−02 - 1.49E−02 -
3 6.15E−03 7.02E−02 8.76E−02 4 6.79E−02 2.73E−02 3.73E−02 1.93E−02
4 6.93E−03 5.26E−02 1.32E−01 3 7.87E−02 5.26E−02 2.99E−02 1.49E−02
5 1.32E−02 1.23E−01 1.08E−01 7 8.52E−02 3.96E−02 5.97E−02 3.22E−02
6 1.86E−02 1.58E−01 1.18E−01 9 9.35E−02 2.68E−02 7.46E−02 4.09E−02

Fig. 18. Component Diagram of the first ranked solution of RADA.

Fig. 17. To be able to depict the knee region in the problem, we The top two outstanding solutions are depicted as component
need to convert a maximization problem of RT to a minimiza- diagrams in Figs. 18 and 19. They are nearly identical except from
tion problem − RT . The figure represents all non-dominated sys- the 1st system except for the subsystem three, the 1st rank system
tem along with the top six outstanding systems. All of them are focuses on the cost, so only three redundant subsystems are used
clustered at the same knee region as there is only one knee in the while the 2nd rank system focuses on the reliability in which four
problem. Tables 12 and 13 show the outstanding systems in detail. redundant subsystems are used. Consequently, the 1st rank system
Table 12 reveals that the system reliabilities of the top six out- has a lower total cost of 34 vs. 35 units while the 2nd rank system
standing systems are between 0.9816 and 0.9926, while the sys- has a better total reliability of 0.9834 vs. 0.9816. Moreover, from
tem costs range between 34 to 41 units, and the system weights Table 13, the 1st rank system has ATC = 0, so it is the true out-
are between 109 and 127. The weights are high but satisfied with standing solution (TOS).
the specified system constraint. From the results above, it can be observed that the applica-
tion problem has one knee region. All top six outstanding solutions
 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446 15

Fig. 19. Component Diagram of the second ranked solution of RADA.

from RADA are successfully ranked and located at the knee region. Acknowledgements
The first solution is the true outstanding solution which the DMs
should consider first while the rest are nonetheless appeals to the This research is supported by Petchra Pra Jom Klao Doctoral
DMs as other alternatives. Scholarship, grant number 15/2560, King Mongkut’s University of
Technology Thonburi.
6. Conclusion
References
In this paper, we propose RADA, a novel method for pruning
and ranking solutions from an existing set of non-dominated so- Bechikh, S., Said, L., & Ghedira, K. (2011). Searching for knee regions of the Pareto
front using mobile reference points. Soft Computing, 15(9), 1807–1823. doi:10.
lutions. This parameter-free algorithm is based on an angle-based
10 07/s0 050 0- 011- 0694- 3.
penalty technique for filtering out less-outstanding solutions and Bhattacharjee, K. S., Singh, H. K., Ryan, M., & Ray, T. (2017). Bridging the gap: Many-
a penalty intensity measure that favors more distant solutions for objective optimization and informed decision-making. IEEE Transactions on Evo-
promoting diversity. RADA can rank every solution in a single run lutionary Computation, 21(5), 813–820. doi:10.1109/tevc.2017.2687320.
Branke, E., Deb, K., Dierolf, H., & Osswald, M. (2004). Finding knees in multi-objec-
by using round-robin tournament comparison. The decision mak- tive optimization. Parallel Problem Solving from Nature - Ppsn Viii, 3242, 722–731
ers can choose to see the top K outstanding solutions without the Retrieved from <Go to ISI>://WOS:0 0 022410330 0 073.
prior knowledge about the problem. To demonstrate the perfor- Branke, J., Kaußler, T., & Schmeck, H. (2001). Guidance in evolutionary multi-
objective optimization. Advances in Engineering Software, 32(6), 499–507. https:
mance, we have tested RADA against ADA, ASA, EMUr , and EMU //doi.org/10.1016/S0965-9978(0 0)0 0110-1.
by using well-known benchmark problems; ZDT3, DEB2DK and Braun, M., Shukla, P., & Schmeck, H. (2017). Angle-based preference models in multi-
DEB3DK. We tested the DEB2DK problem for number of knees objective optimization. In Angle-based preference models in multi-objective opti-
mization (pp. 88–102). doi:10.1007/978- 3- 319- 54157- 0_7.
from one to nine. The problem sets are generated by an exact Cheng, R., Jin, Y. C., Olhofer, M., & Sendhoff, B. (2016). A reference vector guided evo-
method which produces optimal solution sets and Two_Arch2 algo- lutionary algorithm for many-objective optimization. IEEE Transactions on Evolu-
rithm which produces approximated solution sets. The results are tionary Computation, 20(5), 773–791. doi:10.1109/tevc.2016.2519378.
Coit, D. W., & Smith, A. E. (1996). Reliability optimization of series-parallel sys-
encouraging. All tests can successfully discover the knee solutions
tems using a genetic algorithm. IEEE Transactions on Reliability, 45(2), 254–260.
within the top k + 3 outstanding solutions excepts for DEB3DK doi:10.1109/24.510811.
k = 2 where there are 16 solutions that share the 1st rank for Coit, D. W., Tongdan, J., & Wattanapongsakorn, N. (2004). System optimization with
component reliability estimation uncertainty: A multi-criteria approach. IEEE
the optimal set. RADA discovers the same outstanding regions for
Transactions on Reliability, 53(3), 369–380. doi:10.1109/TR.2004.833312.
both solution sets but the ordering might be different due to the Cui, Y., Geng, Z., Zhu, Q., & Han, Y. (2017). Review: Multi-objective optimization
lower quality of the approximated sets. RADA outperforms all ref- methods and application in energy saving. Energy, 125, 681–704. https://doi.org/
erenced algorithms, as it can always find all knee solutions and the 10.1016/j.energy.2017.02.174.
Das, I. (1999). On characterizing the “knee” of the Pareto curve based on Normal-
obtained solutions are also in the knee area. Boundary Intersection. Structural and Multidisciplinary Optimization, 18(2-3),
Finally, the algorithm has been successfully applied to solve a 107–115. doi:10.1007/BF01195985.
real-world problem involving system reliability design with redun- Deb, K., & Gupta, S. (2010). Towards a link between knee solutions and preferred
solution methodologies. In Vol. 6466 LNCS. Lecture Notes in Computer Science (in-
dant components. The top two solutions are significantly different, cluding subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioin-
so that the decision makers can realize different outcomes and can formatics (pp. 182–189).
choose the preferred solutions to implement. Deb, K., & Gupta, S. (2011). Understanding knee points in bicriteria problems and
their implications as preferred solution principles. Engineering Optimization,
One area for further improvement of the algorithm is to inte- 43(11), 1175–1204. doi:10.1080/0305215X.2010.548863.
grate the post-pruning approach with an MOEA. Consequently, the Deb, K., Pratap, A., Agarwal, S., & Meyarivan, T. (2002). A Fast and Elitist Multiobjec-
computational complexity can be greatly reduced as the integrated tive Genetic Algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation,
6(2), 182–197. doi:10.1109/4235.996017.
algorithm will focus only on outstanding solutions rather than the
Dhiman, G., & Kumar, V. (2019). KnRVEA: A hybrid evolutionary algorithm
whole Pareto front. This is the future work. based on knee points and reference vector adaptation strategies for many-
objective optimization. Applied Intelligence, 49(7), 2434–2460. doi:10.1007/
s10489- 018- 1365- 1.
Declaration of Competing Interest Fyffe, D. E., Hines, W. W., & Lee, N. K. (1968). System reliability allocation and a
computational algorithm. IEEE Transactions on Reliability, R-17(2), 64–69. doi:10.
1109/TR.1968.5217517.
The authors declare that they have no known competing finan- Hu, J., Yu, G., Zheng, J., & Zou, J. (2017). A preference-based multi-objective evo-
cial interests or personal relationships that could have appeared to lutionary algorithm using preference selection radius. Soft Computing, 21(17),
influence the work reported in this paper. 5025–5051. doi:10.10 07/s0 050 0- 016- 2099- 9.
Huang, X. Z., Coolen, F. P. A., & Coolen-Maturi, T. (2019). A heuristic survival signa-
ture based approach for reliability-redundancy allocation. Reliability Engineering
& System Safety, 185, 511–517. doi:10.1016/j.ress.2019.02.010.
Credit authorship contribution statement
Khalili-Damghani, K., Abtahi, A.-R., & Tavana, M. (2013). A new multi-objective par-
ticle swarm optimization method for solving reliability redundancy allocation
Supoj Choachaicharoenkul: Methodology, Investigation, Soft- problems. Reliability Engineering & System Safety, 111, 58–75. https://doi.org/10.
ware, Visualization, Validation, Writing - original draft. Naruemon 1016/j.ress.2012.10.009.
Kuo, W., & Wan, R. (2007). Recent advances in optimal reliability allocation. IEEE
Wattanapongsakorn: Validation, Methodology, Supervision, Con- Transactions on Systems, Man, and Cybernetics Part A:Systems and Humans, 37(2),
ceptualization, Writing - review & editing. 143–156. doi:10.1109/TSMCA.2006.889476.
16 S. Choachaicharoenkul and N. Wattanapongsakorn / Expert Systems With Applications 158 (2020) 113446

Leesutthipornchai, N., & Wattanapongsakorn, N. (2013). Extended dominance angle Sudeng, S., & Wattanapongsakorn, N. (2016a). A decomposition-based approach for
algorithm for filtering pareto optimal solutions in multi-objective optimization knee solution approximation in multi-objective optimization. Paper presented
problems. KMUTT Technical Report, 2013(01), 1–31. at the Proceeding of the 2016 Congress on Evolutionary Computation, Vancouver,
Li, K., Deb, K., & Yao, X. (2018). R-Metric: Evaluating the performance of preference- Cananda.
based evolutionary multi-objective optimization using reference points. IEEE Sudeng, S., & Wattanapongsakorn, N. (2016b). A knee-based multi-objective evolu-
Transactions on Evolutionary Computation 1-1. doi:10.1109/TEVC.2017.2737781. tionary algorithm: an extension to network system optimization design prob-
Li, H., & Zhang, Q. (2009). Multiobjective optimization problems with complicated lem. Cluster Computing-the Journal of Networks Software Tools and Applications,
pareto sets, MOEA/D and NSGA-II. IEEE Transactions on Evolutionary Computation, 19(1), 411–425. doi:10.1007/s10586-015-0492-2.
13(2), 284–302. doi:10.1109/TEVC.2008.925798. Tian, Y., Cheng, R., Zhang, X., & Jin, Y. (2017). PlatEMO: A MATLAB platform for
Liu, R., Li, J., Feng, W., Yu, X., & Jiao, L. (2018). A new angle-based preference selec- evolutionary multi-objective optimization. IEEE Computational Intelligence, 12(4),
tion mechanism for solving many-objective optimization problems. Soft Com- 73–87. doi:10.1109/MCI.2017.2742868.
puting, 22(19), 6311–6327. doi:10.10 07/s0 050 0- 017- 2978- 8. Von Neumann, J., & Morgenstern, O. (1944). Theory of games and economic behavior.
Ouyang, Z. Y., Liu, Y., Ruan, S. J., & Jiang, T. (2019). An improved particle swarm op- Princeton, NJ, US: Princeton University Press.
timization algorithm for reliability-redundancy allocation problem with mixed Voudoukis, N. (2017). Inverse square law for light and radiation: A unifying edu-
redundancy strategy and tor heterogeneous components. Reliability Engineering cational approach. Engineering Research and Science, 2(11), 23–27. doi:10.24018/
& System Safety, 181, 62–74. doi:10.1016/j.ress.2018.09.005. ejers.2017.2.11.517.
Ramirez-Atencia, C., Mostaghim, S., & Camacho, D. (2017). A knee point based evo- Wang, H., Jiao, L., & Yao, X. (2015). Two_Arch2: An Improved Two-Archive Algorithm
lutionary multi-objective optimization for mission planning problems. In Paper for Many-Objective Optimization: 19.
presented at the Proceedings of the Genetic and Evolutionary Computation Confer- Wang, L., & Li, L.-P. (2012). A coevolutionary differential evolution with harmony
ence, Berlin, Germany. search for reliability–redundancy optimization. Expert Systems with Applications,
Rausand, M. (2014). Reliability of Safety-Critical Systems: Theory and Applications (1st 39(5), 5271–5278. https://doi.org/10.1016/j.eswa.2011.11.012.
ed.). John Willey & Sons. Wang, K. Z., Yuan, B., Zhao, M. T., & Lu, Y. W. (2020). Cooperative route planning
Samanta, A., & Basu, K. (2019). A novel particle swarm optimization with fuzzy for the drone and truck in delivery services: A bi-objective optimisation ap-
adaptive inertia weight for reliability redundancy allocation problems. Intelli- proach. Journal of the Operational Research Society, 1–18. doi:10.1080/01605682.
gent Decision Technologies-Netherlands, 13(1), 91–99. doi:10.3233/idt-190357. 2019.1621671.
Setoguchi, Y., Narukawa, K., & Ishibuchi, H. (2015). A knee-based EMO algorithm with Weiszer, M., Chen, J., Stewart, P., & Zhang, X. J. (2018). Preference-based evolutionary
an efficient method to update mobile reference points. Cham. algorithm for airport surface operations. Transportation Research Part C-Emerging
Shalan, S., & Ykhlef, M. (2015). Solving multi-objective portfolio optimization prob- Technologies, 91, 296–316. doi:10.1016/j.trc.2018.04.008.
lem for Saudi Arabia stock market using hybrid clonal selection and particle Ye, X., Liu, S. H., Yin, Y. L., & Jin, Y. C. (2017). User-oriented many-objective cloud
swarm optimization. Arabian Journal for Science and Engineering, 40(8), 2407– workflow scheduling based on an improved knee point driven evolutionary al-
2421. doi:10.1007/s13369-015-1744-4. gorithm. Knowledge-Based Systems, 135, 113–124. doi:10.1016/j.knosys.2017.08.
Shukla, P. K., Braun, M., & Schmeck, H. (2014). On the interrelationships between knees 006.
and aggregate objective functions. Paper presented at the GECCO 2014 - Companion Zhang, X., Tian, Y., & Jin, Y. (2015). A knee point-driven evolutionary algorithm
Publication of the 2014 Genetic and Evolutionary Computation Conference. for many-objective optimization. IEEE Transactions on Evolutionary Computation,
Sudeng, S., & Wattanapongsakorn, N. (2015). Post Pareto-optimal pruning algo- 19(6), 761–776. doi:10.1109/tevc.2014.2378512.
rithm for multiple objective optimization using specific extended angle domi- Zitzler, E., Deb, K., & Thiele, L. (20 0 0). Comparison of multiobjective evolutionary
nance. Engineering Applications of Artificial Intelligence, 38, 221–236. doi:10.1016/ algorithms: Empirical Results. Evolutionary Computation, 8(2), 173–195. doi:10.
j.engappai.2014.10.020. 1162/106365600568202.