8326 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO.

8, AUGUST 2022

Multiobjective Optimization-Aided
Decision-Making System for Large-Scale
Manufacturing Planning
Zhenkun Wang , Member, IEEE, Hui-Ling Zhen, Jingda Deng , Qingfu Zhang, Fellow, IEEE, Xijun Li ,
Mingxuan Yuan, and Jia Zeng, Senior Member, IEEE

Abstract—This work is geared toward a real-world manufac- Meanwhile, a new SOP solving strategy is used in the MP system
turing planning (MP) task, whose two objectives are to maximize to further reduce the computational cost. It utilizes two sequen-
the order fulfillment rate and minimize the total cost. More tial easier SOPs as the approximator of the original complex
important, the requirements and constraints in real manufac- SOP for optimization. As part of the MP system, TSMOA and
turing make the MP task very challenging in several aspects. the SOP solving strategy are demonstrated to be efficient in real-
For example, the MP needs to cover many production compo- world MP applications. In addition, the effectiveness of TSMOA
nents of multiple plants over a 30-day horizon, which means that is also validated on benchmark problems. The results indicate
it involves a large number of decision variables. Furthermore, the that TSMOA as well as the MP system are promising.
MP task’s two objectives have extremely different magnitudes,
and some constraints are difficult to handle. Facing these uncom- Index Terms—Integer programming, manufacturing plan-
promising practical requirements, we introduce an interactive ning (MP), multiobjective optimization, real-world application,
multiobjective optimization-based MP system in this article. It two-stage.
can help the decision maker reach a satisfactory tradeoff between
the two objectives without consuming massive calculations. In the
MP system, the submitted MP task is modeled as a multiobjective I. I NTRODUCTION
integer programming (MOIP) problem. Then, the MOIP problem
ANUFACTURING planning (MP) is critical to any
is addressed via a two-stage multiobjective optimization algo-
rithm (TSMOA). To alleviate the heavy calculation burden,
TSMOA transforms the optimization of the MOIP problem into
M industrial company, as it directly affects production
efficiency and supply capacity. The MP aims to seek effi-
the optimization of a series of single-objective problems (SOPs). cient coordination and scheduling of all production activities
during the planning period to achieve the company’s objec-
Manuscript received 7 June 2020; revised 18 October 2020; accepted
3 January 2021. Date of publication 2 February 2021; date of current ver- tives [1], [2]. Real-world MP tasks usually consider more than
sion 19 July 2022. This work was supported in part by the SUSTech start-up one objective, and they are often regarded as multiobjective
funding; in part by the National Science Foundation of China under Grant optimization problems [3], [4]. For example, [5] has two objec-
61876163, Grant 11991023, Grant 62076197, Grant 62072364, and Grant
61721002; in part by the ANR/RGC Joint Research Scheme through the tives, namely: 1)maximizing the customer satisfaction and
Research Grants Council of the Hong Kong Special Administrative Region, 2)minimizing the total cost. In [5], the three objective functions
China, and France National Research Agency under Project A-CityU101/16; related to economy, environment, and society are optimized
in part by the Major Project of National Science Foundation of China
under Grant U1811461; and in part by the Key Project of National Science simultaneously.
Foundation of China under Grant 11690011. This article was recommended This article investigates a real-world MP task with the two
by Associate Editor F. J. Cabrerizo. (Corresponding author: Zhenkun Wang.) objectives: 1) maximizing the order fulfillment rate and 2) min-
Zhenkun Wang is with the School of System Design and Intelligent
Manufacturing, Southern University of Science and Technology, Shenzhen imizing the total cost. Nevertheless, the requirements and
518055, China, also with the Department of Computer Science and constraints in real manufacturing make it very challenging.
Engineering, Southern University of Science and Technology, Shenzhen First, the MP task involves many decision variables because
518055, China, and also with the Department of Computer Science, City
University of Hong Kong, Hong Kong (e-mail: [email protected]). it needs to cover multiple plants’ numerous production items
Hui-Ling Zhen, Mingxuan Yuan, and Jia Zeng are with Huawei over a 30-day horizon. Besides, its two objectives have badly
Noah’s Ark Lab, Hong Kong (e-mail: [email protected]; different magnitudes, and the value ranges have large varia-
[email protected]; [email protected]).
Jingda Deng is with the School of Mathematics and Statistics, Xi’an tions with MP data. In addition, the definition of the order
Jiaotong University, Xi’an 710049, China (e-mail: [email protected]). fulfillment rate contains some intractable constraints.
Qingfu Zhang is with the Department of Computer Science, City Many kinds of optimization algorithms have been
University of Hong Kong, Hong Kong, and also with the Shenzhen Research
Institute, City University of Hong Kong, Shenzhen 518057, China (e-mail: developed, whether for specific problems [6], [7] or general
[email protected]). frameworks [8], [9]. But to the best of our knowledge, none
Xijun Li is with the Department of Electronic Engineering and Information of them can handle such MP tasks well. For instance, some
Science, University of Science and Technology of China, Hefei 230052,
China, and also with the Huawei Noah’s Ark Lab, Shenzhen 518173, China multiobjective optimization algorithms [10], [11] proposed in
(e-mail: [email protected]). our previous work have rarely considered these real-world
This article has supplementary material provided by the challenges (e.g., integer variables and complex constraints).
authors and color versions of one or more figures available at
https://doi.org/10.1109/TCYB.2021.3049712. The development of algorithms or systems for such MP tasks
Digital Object Identifier 10.1109/TCYB.2021.3049712 has economic value and research significance since many other
2168-2267 
c 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: MULTIOBJECTIVE OPTIMIZATION-AIDED DECISION-MAKING SYSTEM FOR LARGE-SCALE MP 8327

real-world applications (e.g., portfolio optimization [12] and

logistic scheduling [13]) also share similar properties.
This article presents an interactive multiobjective
optimization-based MP system, to help the dispatcher or deci-
sion maker (DM) achieve a satisfactory compromise between
the two objectives without causing severe computational
burden. Compared with noninteractive algorithms [14]–[17],
interactive algorithms [8], [18], [19] require less computation
and are more suitable for these MP tasks [14]. Specifically, Fig. 1. Illustration of the feasible set of a MOIP in the objective space.
the submitted MP task is first modeled as a multiobjective
integer programming (MOIP) problem [20], [21]. Thereafter,
the MOIP problem is solved by a two-stage multiobjective f1 , . . . , fm together define the mapping from the decision space
optimization algorithm (TSMOA). Finally, the MP system to the objective space, and f(x) is the objective vector corre-
outputs the solution of the submitted MP task. sponding to x. Accordingly, F = {f(x) : x ∈ X} ⊂ Z m is called
In the design of TSMOA, the problem’s main challenges the feasible set in the objective space.
and the algorithm’s practicability are fully taken into account. Definition 1: f(x1 ) is said to strictly dominate (weakly dom-
TSMOA adopts a two-stage interaction. Its first stage aims to inate) f(x2 ), if fi (x1 ) < fi (x2 ) (fi (x1 ) ≤ fi (x2 )) for each
globally approximate the Pareto front of the MOIP problem with i ∈ {1, . . . , m}.
a small number of Pareto-optimal solutions. For this purpose, Definition 2: f(x1 ) is said to dominate f(x2 ), if fi (x1 ) ≤
a series of well-designed single-objective problems (SOPs) is fi (x2 ) for each i ∈ {1, . . . , m} and fj (x1 ) < fj (x2 ) for at least
constructed for optimization. These SOPs’ optimal solutions one j ∈ {1, . . . , m}.
can give the DM a rough but comprehensive understanding of Definition 3: f(x∗ ) is said to be nondominated, if there is
the entire MP, thus enabling the DM to determine a region of no f(x) ∈ F that can dominate it. Correspondingly, x∗ is called
interest (RoI) in the objective space. Within the RoI, the second an efficient or Pareto-optimal solution.
stage of TSMOA conducts a fine-grained approximation. Here, Definition 4: The set of all efficient solutions is represented
a modified -constraint method [22] is used to construct SOPs by XE , and its image in the objective space (denoted by FN )
for optimization. From these SOPs’ optimal solutions, the DM is called the nondominated front or Pareto front.
can choose the most suitable one as the final solution of the Definition 5: For a given f(x∗ ) ∈ FN , f(x∗ ) is called an
MP task. Compared with the noninteractive MOIP algorithm, unsupported nondominated objective vector if and only if
TSMOA only needs to obtain fewer Pareto-optimal solutions it can be strictly dominated by any convex combination of
and, thus, requires less computational effort. the objective vectors in FN . Otherwise, f(x∗ ) is a supported
Along with TSMOA, a new SOP solving strategy is intro- nondominated objective vector. Accordingly, x∗ is called an
duced in the MP system to lessen the computational bur- unsupported efficient solution or a supported efficient solution.
den further. In this strategy, we use two sequential easier Definition 6: For any fi ∈ {f1 , . . . , fm }, we define its lower
SOPs as the approximator of the original complex SOP for bound and upper bound concerning the Pareto front as fimin =
optimization. As part of the MP system, TSMOA and the minx∈XE fi (x) and fimax = maxx∈XE fi (x). For a subregion
new SOP solving strategy are demonstrated to be efficient in [a, b] ⊆ [fimin , fimax ], we use {f(xB1 ), . . . , f(xBm )} to denote
real-world MP applications. Beyond that, in order to illustrate its boundary-objective vectors, where
TSMOA’s applicability to other problems, we also validate its
performance on the benchmark problems. xBk = argmin fk (x) ∀k ∈ {1, . . . , m}. (2)
The remainder of this article proceeds as follows. Section II a≤fi (x)≤b
x∈XE
presents some concepts of the MOIP problem. The MP task
is described mathematically in Section III. In Section IV, the Definition 7: f(x∗ ) is said to be weakly nondominated, if
proposed MP system is introduced in detail. Section V applies there is no f(x) ∈ F that can strictly dominate it. x∗ is called
the MP system to some real-world MP tasks. In Section VI, a weakly efficient solution.
TSMOA is performed and compared on benchmark problems. It is interesting to note that the terms efficient and
Some discussions are provided in Section VII. Finally, this nondominated are counterparts of each other. A solution
article concludes in Section VIII. x is said to be efficient if and only if the objective
vector f(x) is nondominated. Likewise, unsupported effi-
II. BACKGROUND cient/nondominated, supported efficient/nondominated, and
weakly efficient/nondominated are defined in the same man-
The MOIP problem can be written as
ner. More detailed coverage of these definitions can be found
minimize f(x) = (f1 (x), . . . , fm (x)) in [23] and [24].
subject to x ∈ X (1) From the example in Fig. 1, we can observe that point C
strictly dominates point G, and point F is a weakly nondom-
where X ⊂ Z n denotes the set of all feasible solutions. For a inated objective vector. The Pareto front (FN ) is represented
feasible solution x = (x1 , . . . , xn ) , all its decision variables by the solid points. It can be divided into two subsets: points
must take non-negative integers. The m objective functions A–E are supported nondominated objective vectors, while the

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.
8328 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 8, AUGUST 2022

TABLE I
other solid points are unsupported nondominated ones. Points D ESCRIPTIONS OF THE MP DATA
A and E are the boundary-objective vectors of the Pareto front.
As stated in [25], supported nondominated objective vectors
have some useful properties. They are located at the “lower
left boundary” of the convex hull of the feasible set [26].
Therefore, good global insights about the Pareto front can
be gained by finding supported nondominated objective vec-
tors. Furthermore, for any supported nondominated objective
vector, its corresponding Pareto-optimal solution must be the
optimal solution of (3) with a specific weight vector
m
minimize g(x|w) = wi fi (x)
i=1
subject to x ∈ X (3)
)
where w = (w1 , . . . , wm is aweight vector that satisfies
wi ≥ 0 for all i = 1, . . . , m and m
i=1 wi = 1.

III. P ROBLEM D ESCRIPTION

In this section, the real-world MP task is described math-
ematically. Here, we adopt “items” to generally refer to all
kinds of production components (e.g., raw materials, semi fin-
ished products, and finished products). With the items, each
plant has the following production activities.
1) Store the items transported from other plants or pur-
chased from the market in the warehouse.
2) Transport the items from the warehouse to other plants.
3) Supply customers with the items in the warehouse to
fulfill their orders.
4) Manufacture products using the items in the warehouse.
Newly manufactured products are treated as inbound
items to be stored in the warehouse first.
All these production activities serve two objectives: 1) max- and it is expressed as
imizing the order fulfillment rate and 2) minimizing the  T
total cost. i∈I t=1 ui,t
maximizefsat = (4)
Beyond that, the constraints in real manufacturing are also |I|T
taken into account. For example, the maximum production where ui,t is an intermediate variable that calculates the fill
capacity of each plant cannot be exceeded. Meanwhile, each rate of the order demand concerning item i on the t-th day. It
plant has its minimum production lot-size for each item. These is defined as
constraints are all related to the MP data described in Table I.  
Each MP data is a constant provided by the company. The p∈P zi,p,t − mi,t−1 + 1
ui,t = max ,0
indices i ∈ I, p ∈ P and t = 1, 2, . . . , T are used to denote Di,t + 1
item i, plant p, and the t-th day, where T = 30 and I and P ∀i ∈ I, t = 1, 2, . . . , T. (5)
are the sets of all the items and all the plants.
There are four types of decision variables in the MP task, In (5), adding 1 to both numerator and denominator is to avoid
which are described as follows. the denominator being 0. The adoption of the max function
zi,p,t : The quantity of item i supplied to the customers by is to prevent ui,t from being negative. mi,t−1 represents the
plant p on the t-th day. backordered quantity of item i in the previous t − 1 days. Its
p
si,p,t : The quantity of item i transported from plant p to definition is given as
⎧
plant p on the t-th day. ⎨ mi,0 = Mi 
xi,p,t : The number of batches of item i produced by plant mi,t = mi,t−1 + Di,t − p∈P zi,p,t (6)
⎩
p on the t-th day. mi,t ≥ 0 ∀i ∈ I, t = 1, 2, . . . , T.
i : The quantity of item i used to replace item i for the
ri,p,t
production in plant p on the t-th day. The second objective function of the MP task calculates
the total cost, which includes the production cost (Pcost),
A. Objective Functions transportation cost (Tcost) and holding cost (Hcost). It is
written as
The first objective function of the MP task defines the order
fulfilment rate (i.e., average satisfaction of the order demands), minimize feco = Pcost + Tcost + Hcost (7)

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: MULTIOBJECTIVE OPTIMIZATION-AIDED DECISION-MAKING SYSTEM FOR LARGE-SCALE MP 8329

where Pcost and Tcost are defined in (8) and (9), respec- the t-th day. As shown in (18), the inbound of each item con-
tively. While (10) defines Hcost as the sum of the holding sists of three parts: 1) the production inbound (i.e., PIBi,p,t );
costs incurred during inventory (IHcost), production (PHcost), 2) the transportation inbound (i.e., TIBi,p,t ); and 3) the pur-
and transportation (THcost). Their corresponding definitions chase inbound (i.e., BIBi,p,t ). Note that 1 and 2 herein
are given in (11)–(13), respectively represent the time delays caused by public holidays
⎧
 
T ⎪ ini,p,t = PIBi,p,t + TIBi,p,t + BIBi,p,t
⎪
⎪ PIB
Pcost = PCi,p x̂i,p,t . (8) ⎨ i,p,t = x̂i,p,t−PT i,p −1
 p (18)
i,p,t =
i∈I p∈P t=1
⎪
⎪
TIB p ∈P s
  
T ⎪
⎩
p
i,p,t−LT p −2
Tcost = TCi,p
p p
si,p,t . (9) ∀i ∈ ∀p ∈ P, t = 1, 2, . . . , T.
i∈I p ∈P p∈{P\p } t=1 Equation (19) calculates the outbound quantity of each item
Hcost = IHcost + PHcost + THcost, (10) in each plant on each day. It is the sum of the production
 
T outbound (i.e., POBi,p,t ), the transportation outbound (i.e.,
IHcost = HCi,p vi,p,t , (11) POBi,p,t ), the replacement outbound (i.e., ROBi,p,t ), and the
i∈I p∈P t=1 supply quantity to customers (i.e., zi,p,t )
⎧
 
T
⎪
⎪ out = POBi,p,t +TOBi,p,t + ROBi,p,t + zi,p,t
PHcost = HCi,p PTi,p x̂i,p,t , (12) ⎪ i,p,t
⎪  
⎪
⎨ POB i,p,t = ∈I Bii,p,t x̂i ,p,t − i ∈I ri,p,t
i
i∈I p∈P t=1  i
p
TOBi,p,t = p ∈P si,p ,t (19)
  
T ⎪
⎪ 
p ⎪
⎪ ROB = r i
THcost = HCi,p LTpp si,p,t . (13) ⎪
⎩
i,p,t i ∈I i ,p,t
i∈I p∈P p ∈{P\p} t=1 ∀i ∈ I ∀p ∈ P, t = 1, 2, . . . , T.
It reveals in the definition of POBi,p,t that the consumption of
x̂i,p,t is an intermediate variable [defined in (14)] that repre-
one item can be reduced if it is substituted with another item
sents the quantity of item i produced in plant p on the t-th day.
in the production. This kind of substitution often occurs when
The intermediate variable vi,p,t denotes the inventory quantity
some items are insufficient.
of item i in plant p on the t-th day, and its definition is given
The constraints defined in (21) and (21) are to prevent
in (17).
unnecessary and excessive substitution
B. Constraints POBi,p,t + zi,p,t ≥ 0 ∀i ∈ I ∀p ∈ P, t = 1, 2, . . . , T.
In the production of each item in each plant, the lot-size (20)

and minimum production constraints are considered, that is i
ri,p,t ≤ Ri,p,t ∀i ∈ I ∀p ∈ P, t = 1, 2, . . . , T. (21)
i ∈I
0, if xi,p,t PMi < MLSi,p,t
x̂i,p,t = Equation (21) indicates that the consumption of one item
xi,p,t PMi , otherwise
∀i ∈ I ∀p ∈ P, t = 1, 2, . . . , T. (14) should be higher than the use of its substitutes. While (21)
suggests that the replacement of one item cannot exceed a
It indicates in (14) that the production is executed only when specific number.
x̂i,p,t is greater than the minimum production quantity MLSi,p,t .
Moreover, the capacity constraints of each plant are taken into C. Challenges
account, as shown in the following equation: As described previously, the real-world MP task is modeled
 as a MOIP problem, where f1 = −fsat and f2 = feco . The
Ui,p x̂i,p,t ≤ CAPp,t ∀p ∈ P, t = 1, 2, . . . , T. (15)
i∈I
problem’s main challenges can be summarized as follows.
1) Dissimilar and Changeable Magnitudes: The two objec-
In this MP, the paired production is also considered. The tive functions have extremely different magnitudes. The
production of one item needs to match a certain number of maximum value of fsat is not greater than 1, while
production of another item, such as one charger should match the maximum value of feco can be millions or billions.
one phone. The corresponding constraints are given as Moreover, the two objective functions’ value ranges vary
x̂i ,p,t = PAIRii,p x̂i,p,t ∀i , i ∈ I ∀p ∈ P, t = 1, 2, . . . , T. greatly depending on the MP data.
2) Many Variables and Constraints: The real-world MP
(16)
needs to cover many items of multiple plants over 30
The inventory level of each item in each plant on each day days. As a result, the MOIP problem has to include a
is calculated as large number of decision variables (up to millions) and
⎧ constraints.
⎨ vi,p,0 = Vi,p
3) Complex Constraints: The max function is adopted
vi,p,t = vi,p,t−1 + ini,p,t − outi,p,t (17)
⎩ in (5), which may cause the MOIP problem to be
vi,p,t ≥ 0 ∀i ∈ I ∀p ∈ P, t = 1, 2, . . . , T
multimodal and nonsmooth. As we know, general MOIP
where the intermediate variables ini,p,t and outi,p,t represent problems are much easier to solve than multimodal or
the inbound and outbound quantity of item i in plant p on nonsmooth ones.

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.
8330 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 8, AUGUST 2022

Algorithm 1: TSMOA-stage1
1 Input: the MOIP problem and the maximal number of
the optimized SOPs SOPmax ;
2 Output: the approximate set S1 ;
Fig. 2. Workflow of the proposed MP system. 3 Achieve the boundary solutions xB1 and xB2 via
lexicographic optimization (cf., Eq. (22));
4 Let S1 = {(xB1 , f(xB1 )), (xB2 , f(xB1 ))}, U = {f(xB1 )f(xB2 )}
IV. P ROPOSED MP S YSTEM and SOPs = 4;
To deal with such an MP task, we propose a multiobjective 5 while (SOPs < SOPmax ) and (U = φ) do
optimization-based MP system. Fig. 2 illustrates the workflow 6 Utilize (23) to select the longest line segment from U
of the MP system. First, the system converts the MP task and denote it as f(xa )f(xb );
into a MOIP problem based on the submitted MP data. Then, 7 Construct a SOP using (25), and optimize it to obtain
a MOIP algorithm is employed to solve the MOIP problem. a solution x∗ ;
Finally, the system outputs the solution of the MP task. 8 if f(x∗ ) is different from any objective vector in S1
The details of converting the MP task into a MOIP problem then
has been presented in Section III. To solve the MOIP problem 9 S1 ← S1 ∪ {(x∗ , f(x∗ ))},
more efficiently, we adopt a newly designed interactive MOIP U ← U ∪ {f(xa )f(x∗ ), f(x∗ )f(xb )};
algorithm (i.e., TSMOA) in the MP system. The adoption of an 10 end
interactive algorithm in the system is based on the following
11 U ← U \ {f(xa )f(xb )}, SOPs ← SOPs + 1;
two considerations.
12 end
1) Compared with the noninteractive algorithm, the
interactive algorithm generally only needs to obtain
fewer Pareto-optimal solutions. Therefore, the
interactive algorithm consumes less computational 3) TSMOA globally approximates the Pareto front in its
time, especially when addressing such large-scale and first stage. This is to cope with the challenge caused by
complex MOIP problems. the two objective functions’ dissimilar and changeable
2) The dispatcher or DM needs to participate in the deci- magnitudes. The yielded solutions can give the DM a
sion making process of the MP. The interactive algo- holistic perception of the value ranges of two objective
rithm communicates with the dispatcher or DM in a functions, thereby enabling the DM to determine a RoI.
timely manner, thus better avoiding some unexpected
situations or industrial disasters.
B. TSMOA-Stage1
A. TSMOA As mentioned in Section II, supported nondominated objec-
tive vectors have particular advantages in providing global
As shown in Fig. 3, TSMOA adopts a two-stage interaction. insights about the Pareto front. Therefore, the objective vectors
In the first stage, TSMOA aims to globally approximate the obtained by TSMOA-stage1 are best to be supported nondom-
entire Pareto front with a small number of Pareto-optimal solu- inated ones. Moreover, the more uniformly these objective
tions. With the knowledge provided by these solutions, the vectors are distributed, the better global information they
DM can determine an RoI. In the second stage, TSMOA con- could offer. Last but not least, the Pareto front’s boundary-
ducts a fine-grained approximation for the RoI. The obtained objective vectors have special physical meanings in this MP
approximate solutions are provided to the DM, and the DM task. They reveal the production and supply limits, which are
can choose the most suitable one among them. indispensable for the DM to make overall considerations.
In the design of TSMOA, the problem’s main challenges As shown in Algorithm 1, TSMOA-stage1 first uses the lex-
and the algorithm’s practicability are sufficiently considered. icographic optimization method [27] to achieve the boundary
1) TSMOA has only two loops of interaction. It is because solutions/objective vectors. More specifically, each boundary
the calculation time for such a large-scale problem in solution xBk for k ∈ {1, 2} is obtained by optimizing the
each loop is quite long, and too many interactive loops following two SOPs:
are a heavy burden for the DM. In this MP task as well
⎧ B
as many other real-world scenarios, two interactive loops ⎨ x k = argminfk (x)≤fk (x̂) fl (x)
are sufficient for the DM to reach a final decision. x̂ = argmin fk (x) (22)
2) In terms of information interaction, TSMOA adopts a ⎩
k, l ∈ {1, 2} and k = l.
straightforward manner. The algorithm displays the dis-
tribution of all the obtained objective vectors to the Here, we adopt the lexicographic optimization method because
DM. The DM determines an RoI by giving upper and it can not only avoid the weakly nondominated objective vec-
lower bounds on one objective function. This intu- tors but also overcome the difficulties caused the dissimilar
itive interaction is to reduce the DM’s cognitive burden objective function magnitudes. Note that the constraints intro-
and facilitate the DM to make decisions or express duced in Section III are also included in each SOP, but they
preferences. are not mentioned here and below for brevity.

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: MULTIOBJECTIVE OPTIMIZATION-AIDED DECISION-MAKING SYSTEM FOR LARGE-SCALE MP 8331

Fig. 3. Illustration of the working principle of TSMOA.

After xB1 and xB2 are obtained, the approximate set S1 is Algorithm 2: TSMOA-stage2
initialized as in line 4 of Algorithm 1. The set U is initialized 1 Input: the MOIP problem, the upper bound f2 and the
up
to {f(xB1 )f(xB2 )}, where f(xB1 )f(xB2 ) indicates the line segment lower bound f2low , the number of intervals Imax and the
between the two objective vectors f(xB1 ) and f(xB2 ). Since four approximate set obtained by TSMOA-stage1 S1 ;
SOPs are already optimized in obtaining the two boundary 2 Output: the approximate set S2 and the number of the
solutions, the number of optimized SOPs (i.e., SOPs) is set optimized SOPs SOPs;
to 4. TSMOA-stage1 then iterates until SOPs reaches a given 3
up
Estimate f1 and f1low with (27), achieve xup and x̂ by
number SOPmax . optimizing the SOPs in (26) and (28), and set SOPs = 2;
In each iteration, the algorithm selects the longest line seg-
4 if f2 (x̂) ≥ f2low then
ment from U (line 4 of Algorithm 1). Here, we use f(xa )f(xb )
5 Set xlow to x̂;
to represent it, where f(xa ) and f(xb ) are its two endpoints.
6 else
The process can be expressed as
7 Obtain xlow with (29), and SOPs ← SOPs + 1;
f(xa )f xb = argmax f̄ x − f̄ x 2
(23) 8 end
Let S2 = {(xup , f(xup )), (xlow , f(xlow ))},  = f2 (x )−f 2 (x )
up low
f(x )f(x )∈U 9
Imax
where f̄(x ) and f̄(x ) are two normalized objective vectors and i = 2;
of x and x , respectively. For a solution x, its normalized 10 while i ≤ Imax do
objective vector f̄(x) is defined as 11 ei = f2 (xup ) − (i − 1) · ;
  12 Construct a SOP using (30), and optimize it to obtain
f1 (x) − f1 xB1 f2 (x) − f2 xB2 a solution x∗ ;
f̄(x) = , . (24)
f1 xB2 − f1 xB1 f2 xB1 − f2 xB2 13 S2 ← S2 ∪ {(x∗ , f(x∗ ))}, SOPs ← SOPs + 1;
14 Update i with (31);
With the selected line segment f(xa )f(xb ), we employ the 15 end
Aneja and Nair method [28], [29] to construct an SOP, that is
x∗ = argmin(w1 f1 (x) + w2 f2 (x)) (25)
where w1 = f2 (xa ) − f2 (xb ) and w2 = f1 (xb ) − f1 (xa ). After to be distributed as uniformly as possible. To achieve this,
the SOP’s optimal solution x∗ is obtained, it is compared to as well as cope with the problem’s particular challenges, we
the previously gained solutions. If its objective vector f(x∗ ) is adopt a modified -constraint method [22] in TSMOA-stage2.
different from any objective vector in S1 , (x∗ , f(x∗ )) is added As shown in Algorithm 2, the first step of TSMOA-stage2 is
to S1 . Accordingly, the two new line segments f(xa )f(x∗ ) and to obtain the boundary solutions/objective vectors of the RoI.
f(x∗ )f(xb ) are added to U, f(xa )f(xb ) is removed from U, and The upper boundary solution xup is achieved by optimizing
SOPs is added one. If f(x∗ ) is the same as any existing objec- the following SOP:
tive vector, U and SOPs are updated as shown in line 11 of
Algorithm 1. After all iterations are conducted, the algorithm
xup = argmin (f1 (x) + γ1 f2 (x)) (26)
outputs the approximate set S1 . up
f2 (x)≤f2
In TSMOA-stage1, the Aneja and Nair method is adopted
to ensure that the corresponding SOP’s optimal solution is a
up up up
supported efficient solution. The selection of the longest line where γ1 = ([0.1(f1 − f1low )]/[f2 − f2low ]). f1 and f1low
segment in each iteration is to make the obtained objective are two values that estimated based on the objective vectors
vectors distributed as uniformly as possible. achieved by TSMOA-stage1. Assuming that f and f are the
up/low
two closest objective vectors above and below f2 , we can
low/up
C. TSMOA-Stage2 estimate f1 as
In our MP system, the DM determines an RoI by specifying
up
the upper and lower bounds on f2 (denoted by f2 and f2low ).  
up/low
f2 − f2 f 1 − f1
Then, TSMOA-stage2 conducts a fine-grained approximation low/up
f1 = f1 − . (27)
for this RoI. Thus, the obtained objective vectors are required f 2 − f2

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.
8332 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 8, AUGUST 2022

To obtain the lower boundary solution xlow , we first attain an and then solve them using well-developed mathematical pro-
intermediate solution x̂ with the following equation: gramming methods (e.g., simplex method [31] and sequential
quadratic programming method [32]). Non-MPS strategies
x̂ = argmin (f1 (x) + γ1 f2 (x)) (28) mainly include evolutionary algorithms [33], tabu search [34],
f2 (x)≤f̂2
memetic algorithms [35], and so on. They are usually time
up consuming and cannot guarantee to find the optimal solu-
where f̂2 = f2low + ([f2 − f2low ]/2Imax ). If f2 (x̂) ≥ f2low , xlow is
set to x̂; otherwise, xlow is gained by optimizing another SOP, tion.
that is We adopt the MPS strategy in our MP system for two rea-
sons. On one hand, we fail to find a non-MPS algorithm or
xlow = argmin (f2 (x) + γ2 f1 (x)) (29) solver that is efficient enough to handle such complex SOPs.
f1 (x)<f1 (x̂) On the other hand, the system is ultimately to serve real-world
where γ2 = (0.01/γ1 ). MP applications. The MPS strategy is more stable and robust,
After xup and xlow are obtained, the approximate set S2 which is very important to the company. When solving each
and the interval size  are set as in line 9 of Algorithm 2. SOP, we use two modified easier SOPs as its approximator,
With , the RoI is divided into Imax intervals between f2 (xup ) and utilize the commercial optimizer Gurobi [36] to solve them
and f2 (xlow ). For each interval, TSMOA-stage2 aims to obtain sequentially. Gurobi has good performance on this kind of
a nondominated objective vector that is closest to its upper regular SOPs, as it not only integrates plentiful heuristic and
bound. Since f(xup ) is already the best choice for the first mathematical programming methods but also can adaptively
interval, we omit it and set the interval index i = 2. match the most suitable solving method for each problem.
The ith interval’s upper bound ei can be calculated as in As summarized in Section III, the main difficulty in opti-
line 11 of Algorithm 2. Based on ei , the SOP is defined as mizing the SOP stems from the definition of fillrate [i.e., (5)].
In (5), the max function is used to prevent the variables from
x∗ , y∗ = argmin (f1 (x) + γ3 · y) (30) being negative. But the max function can cause the SOP to be
f2 (x)+y=ei extremely multimodal and nonsmooth, especially it acts on so
where γ3 = ([0.1(f2 (xup ) − f2 (xlow ))]/[f1 (xlow ) − f1 (xup )]), many variables. Our strategy is to replace the max function by
and y ≥ 0 is a slack variable. The SOP’s optimal solution identifying potential variables and adding linear constraints to
x∗ is added to S2 and the interval index i is updated as them. To achieve this, we first modify the SOP by relaxing its
 ∗ definition in (5) as
y 
i=i+1+ . (31) p∈P zi,p,t − mi,t−1 + 1
 ui,t = ∀i ∈ I, t = 1, 2, . . . , T.
Di,t + 1
Equation (31) indicates that based on the value of y∗ , the (32)
intervals containing no nondominated objective vector can be
skipped. Then, we utilize Gurobi to optimize the new SOP to obtain the
TSMOA-stage2 adopts a similar framework to value of feco (denoted as feco ) and the value of ui,t (denoted
AUGMECON2 [15], a well-known algorithm based on as ui,t ). If ui,t > 0, we believe the variable ui,t is very likely
the -constraint method. In terms of approximating a RoI to satisfy the non-negative constraint. Therefore, we can use
of the Pareto front, TSMOA-stage2 has two advantages (33) as an approximation of (5) to define the fillrate
⎧ 
compared to AUGMECON2. TSMOA-stage2 elaborately ⎪ zi,p,t −mi,t−1 +1
⎨ ui,t = p∈P Di,t +1
achieves the boundary solutions and interval division of
ui,t ≥ 0 (33)
the RoI, thereby can obtain objective vectors with better ⎪
⎩  
distribution. Moreover, TSMOA-stage2 carefully estimates ∀(i, t) ∈ (i, t)|ui,t > 0 for i ∈ I, t = 1, 2, . . . , T .
the parameters (i.e., γ1 , γ2 and γ3 ) in the SOPs. Thereby, Also with the constraint feco ≤ feco , we define the other
it can be less troubled by weakly nondominated objective new SOP to be optimized by Gurobi. The obtained value
vectors, especially when the two objective functions have of ui,t (denoted as ûi,t ) may still not satisfy the non-negative
badly different magnitudes. constraint. So, we further repair it as
0, if ûi,t < 0
D. SOP Solving Strategy u∗i,t = (34)
ûi,t , otherwise.
TMSOA converts the optimization of the MOIP problem
into the optimization of several SOPs. Ideally, each SOP Finally, the repaired value u∗i,t is used to calculate the value
solver can serve the purpose. But for such a complex problem of the objective function fsat .
in our MP task, the SOP solving strategy needs to be
elaborately designed. As stated in [30], the SOP solving V. R EAL -W ORLD MP A PPLICATION
strategies can be roughly divided into two categories: 1) In this section, the proposed MP system is applied to real-
mathematical programming solving (MPS) strategies and 2) world MP tasks. Table II describes the MP data adopted.
nonmathematical programming solving (non-MPS) strategies. Columns Item and Plant reveal the number of plants and the
MPS strategies typically reform a complex SOP into multiple number of items involved in the MP task. Columns Planning
standard or quasi-standard mathematical programming SOPs, horizon shows the planning period. Column Scale indicates

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: MULTIOBJECTIVE OPTIMIZATION-AIDED DECISION-MAKING SYSTEM FOR LARGE-SCALE MP 8333

Fig. 4. Plots of the nondominated points obtained by TSMOA on each MP data. (a) Data-I-1. (b) Data-I-2. (c) Data-I-3. (d) Data-II-1. (e) Data-II-2.
(f) Data-II-3. (g) Data-III-1. (h) Data-III-2. (i) Data-III-3.

TABLE II
D ESCRIPTION OF THE N INE R EAL -W ORLD MP DATA [f2min , f2max ]. The parameters SOPmax in TSMOA-stage1 and
Imax in TSMOA-stage2 are set to 7 and 3, respectively.
A. Experimental Results
The experimental results concerning the nine real-world MP
data are exhibited in Fig. 4. The approximate objective vectors
obtained at the first stage are indicated by pentagrams, while
those yielded at the second stage are shown with plus symbols.
From these figures, we can observe that the objective vectors
yielded by the first stage are distributed widely and uniformly.
These objective vectors not only reveal the maximum produc-
tion capacity and the maximum order supply capacity of the
plants but also roughly show the tradeoff between the total
cost and the average order fillrate. With this comprehensive
knowledge, the DM is capable of making the most appropriate
preference for the RoI.
the scale of the decision variables of the corresponding MOIP For the selected RoI, a fine-grained approximation is con-
problem. In the experiment, we assume that the RoI has a size ducted at the second stage. As shown in Fig. 4, the obtained
of 0.1(f2max − f2min ) and is randomly selected within the range objective vectors have a uniform spread in the RoI. The DM

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.
8334 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 8, AUGUST 2022

TABLE III
S TATISTICS OF D OMINATION R ELATIONSHIP B ETWEEN THE T WO S ETS OF A PPROXIMATE P OINTS ON E ACH MP DATA

TABLE IV
can choose the most suitable one from them as the final solu- AVERAGE C OMPUTATIONAL T IME OF TSMOA- STAGE 1 AND
tion of the MP task. It is worth mentioning that one of the TSMOA- STAGE 2 ON T HREE MP DATA OF D IFFERENT S IZES
objective vectors achieved in the RoI related to Data-II-1 is not
a nondominated one. This is caused by the insufficient accu-
racy of the commercial solver. Nevertheless, our experiments
indicate that this phenomenon rarely happens.

B. Effectiveness of the SOP Solving Strategy

To demonstrate the effectiveness of the new SOP solv- of the MP task, such a calculation time is acceptable to the
ing strategy, we compare two algorithms: 1) TSMOA and dispatcher or DM.
2) TSMOA-RLX on the MOIP problems corresponding to
the nine real-world MP data. In TSMOA, the proposed SOP VI. VALIDATION ON B ENCHMARK P ROBLEMS
solving strategy is employed. Considering that the commer-
As a newly proposed interactive multiobjective optimization
cial solver cannot solve the original problem, we use the SOP
algorithm, TSMOA is by no means only applicable to our MP
solving strategy of straightforwardly replacing the complex
tasks. It can also work for other MOIP problems with similar
constraints with the relaxed ones [i.e., (32)] in TSMOA-RLX.
properties. To illustrate this, we validate the effectiveness of
The other parameter settings of the two algorithms are the
TSMOA on the benchmark problems.
same as introduced earlier.
The bi objective assignment problem (BOAP) [37] is
Fig. 8 in the supplementary material plots the objective
adopted in our investigation. The BOAP is also a MOIP
vectors achieved by the two algorithms on the nine MOIP
problem, and its scale (dimension of decision variables) can
problems. Besides, Table III reveals the domination relation-
be easily controlled via a parameter. Furthermore, all the sup-
ships between the approximate objective vectors obtained by
ported nondominated objective vectors (denoted as the set
TSMOA-RXL and those yielded by TSMOA. In addition,
FSN ) of a BOAP can be enumerated using the Hungarian
Table VI in the supplementary material lists the objective func-
method [38] and the Fukuda algorithm [39]. They can be used
tion values of these obtained objective vectors. These results
as a reference when comparing algorithms, which is of great
indicate that TSMOA significantly outperforms TSMOA-RXL
importance. The BOAP can be written as
on all data except the problem related to Data-I-1. The objec- ⎛ ⎞
tive vectors achieved by TSMOA are able to dominate most of n  n 
n 
n
those attained by TSMOA-RXL. On Data-I-1, TSMOA gains minimize f(x) = ⎝ c1i,j xi,j , c2i,j xi,j ⎠
two better objective vectors and eight identical objective vec- i=1 j=1 i=1 j=1
tors compared with TSMOA-RXL. The reason behind is that 
n
these solutions’ fsat values are very close to the upper bound subject to xi,j = 1, i = 1, 2, . . . , n
1, and the enhancements raised from our SOP solving strategy i=1
are not obvious. n

All in all, the experimental results indicate that the proposed xi,j = 1, j = 1, 2, . . . , n
SOP solving strategy is effective on the MOIP problems j=1

related to real-world MP tasks. xi,j ∈ {0, 1}, i, j = 1, 2, . . . , n (35)

where xi,j is the decision variable, and c1i,j and c2i,j are the
C. Computational Time objective function coefficients. As in [26], the coefficients of
The experiments in terms of the MP task are conducted on each objective function are integers generated from the range
a standalone development server with dual Intel Xeon Gold [0, 20]. In this study, three BOAPs with n = 100, 150 and 200
6254 3.10-GHz 18-core CPU and 192-GB RAM. The algo- are employed.
rithm is implemented in Python. Table IV shows the average
computational time of TSMOA-stage1 and TSMOA-stage2 on A. Overall Comparison
the three MP data of different sizes. It can be observed that It is known that the comparison of interactive algorithms
each stage of TSMOA still needs a certain amount of com- is troublesome, especially when they are designed for specific
putational time, especially when the scale of the MP data is applications. Nevertheless, as stated in [14], the effectiveness
very large. Nevertheless, considering the complexity and scale of interactive algorithms can be demonstrated by showing

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: MULTIOBJECTIVE OPTIMIZATION-AIDED DECISION-MAKING SYSTEM FOR LARGE-SCALE MP 8335

TABLE V
N UMBER OF SOP S O PTIMIZED BY E ACH A LGORITHM
AND THE N UMBER OF N ONDOMINATED O BJECTIVE V ECTORS
I T O BTAINS IN THE RO I OF E ACH BOAP

advantages over noninteractive algorithms in reducing com-

putational load and cognitive burden. Toward this direction, Fig. 5. Plot of the approximate objective vectors obtained by the three
we adopt two classic noninteractive algorithms: 1) MOEA/D- algorithms with SOPs = 15 on the BOAP with n = 100.
WS [40] 2) and AUGMECON2 [15] for comparison.
1) In MOEA/D-WS, the SOPs are generated by the optimize fewer SOPs than the other two noninteractive algo-
weighted sum approach [i.e., (3)] using a set of uni- rithms. So, we can conclude that TSMOA can indeed save a
formly distributed weight vectors [41]. certain amount of computational effort.
2) In AUGMECON2, each SOP is constructed by adding Since TSMOA consists of two parts: 1) TSMOA-stage1 and
the constraint on each interval’s upper boundary [i.e., 2) TSMOA-stage2, we also want to validate their effectiveness
(30)]. But the parameter γ3 in the AUGMECON2 is set separetly.
manually, and we set it to (10−6 /[f2 (xB1 ) − f2 (xB2 )]) as
recommended by [15].
B. Comparison on the First Stage
For fair comparisons, we employ Gurobi as the SOP solver in
all the three algorithms. As mentioned, the first stage of the algorithm aims at giv-
In this experiment, the parameters of TSMOA and the RoIs ing the DM a holistic perception of the Pareto front using
are set to be the same as in Section V and each BOAP’s exper- fewer nondominated objective vectors. The objective vectors
imental data are randomly generated. The results of TSMOA are preferably supported nondominated ones and distributed as
on the three BOAPs are illustrated in Fig. 9(a) in the supple- uniformly as possible. In view of this, we adopt the following
mentary material. It can be seen that TSMOA-stage1 yields two indicators to evaluate the performance of the algorithms.
five supported nondominated objective vectors on each BOAP. 1) SN rate : Let S indicate the approximate set obtained by
They are widely and uniformly distributed along the Pareto an algorithm, SNrate is defined as
front, thus can give the DM a rough but comprehensive holis- SNs
tic perception. Thereafter, the DM selects the RoIs as marked SNrate (S) = (36)
SOPs
by the black boxes in Fig. 9 in the supplementary material.
Within each RoI, TSMOA-stage2 achieves five uniformly dis- where SNs represents the number of supported effi-
tributed nondominated objective vectors. From them, the DM cient objective vectors in S, while SOPs indicates the
can choose the most suitable one as the final solution of the number of SOPs optimized in the algorithm. Since the
corresponding problem. calculation is mainly spent on optimizing each SOP,
Since the computational effort is mainly consumed in the SOPs can reflect the computational cost consumed by
optimization of SOPs, we want to know how many SOPs the each algorithm. A larger SNrate value means the algo-
two noninteractive algorithms have to optimize to achieve such rithm achieves more supported nondominated objective
an approximation for the RoI of each BOAP. For MOEA/D- vectors with the same computational cost.
WS and AUGMECON2, we gradually increase the number 2) SP [42]: Let S indicate the approximate set, the spacing
of SOPs in each algorithm and repeatedly execute it on the metric SP is calculated as

BOAP, until it can obtain five objective vectors in the RoI or  |S|
 1 
SP(S) =  
SOPs reaches a maximum of 500. Fig. 9(b) and (c) in the sup- 2
d − di (37)
plementary material shows the approximated objective vectors |S| − 1
i=1
gained by MOEA/D-WS and AUGMECON2, respectively. It
is worth noting that MOEA/D-WS fails to achieve enough where di = minxi ,xj ∈S, xi =xj f(xi ) − f(xj )1 , and 
d rep-
nondominated objective vectors in the RoI of the BOAP with resents the mean of di for each xi ∈ S. The SP metric is
n = 100. It is because MOEA/D-WS can only find supported widely used to assess the uniformity of the approximate
nondominated objective vectors, while that RoI contains only set [43]. The smaller the SP value, the more uniform the
four supported nondominated objective vectors. In addition, distribution of the approximate objective vectors.
Table V presents the number of SOPs optimized (denoted as To be fair, we let each algorithm utilize the same proce-
SOPs) and the number of the nondominated objective vectors dure [see (22)] to obtain the boundary solutions of the Pareto
(represented by NDs) by each algorithm in the RoI of each front (i.e., xB1 and xB2 ), and let each algorithm generate the
BOAP. From these results, we can find that TSMOA needs to same number of SOPs for optimization. We perform the three

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.
8336 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 8, AUGUST 2022

TABLE VI
M EAN SNrate AND SP VALUES O BTAINED BY TSMOA- STAGE 1, MOEA/D-WS, AND AUGMECON2 ON E ACH BOAP W ITH 20 I NDEPENDENT DATA

algorithms until SOPs reaches 10, 15, and 20 on each BOAP method, that is
with 20 independent random experimental data.   m  

Table VI shows the mean SNrate and SP values obtained g(x|w) = max wi fi (x) − fimin + ρ fi (x) − fimin
by the three algorithms along with the statistical tests. It 1≤i≤m
i=1
clearly indicates that TSMOA-stage1 significantly outperforms (39)
MOEA/D-WS and AUGMECON2 in terms of the SP metric.
As for the SNrate metric, TSMOA-stage1 and MOEA/D-WS where ρ is set to 10−6 , as suggested in [45].
are remarkably better than AUGMECON2. TSMOA-stage1 For each interval of the three BOAPs with 20 independent
achieves better mean SNrate values than MOEA/D-WS when data, the three algorithms are performed until SOPs reaches
SOPs is large. But when SOPs is small, TSMOA-stage1 and 5, 7, and 9, respectively. Their performance is evaluated with
MOEA/D-WS have similar performance concerning the SNrate the hypervolume (HV) [46] metric.
metric. When SOPs is small, the SOPs of MOEA/D-WS are 1) HV: Let S be the approximate set obtained by an algo-
widely dispersed, and the optimization of each SOP can lead rithm, and is a reference objective vector. The HV value
to a different supported nondominated objective vector. This calculates the volume between the approximate objective
means that MOEA/D-WS can attain the best SNrate value on vectors and z∗ = (1.2, . . . , 1.2) , that is
each instance, just like TSMOA-stage1.  
Fig. 5 shows the objective vectors attained by the three ∗
! ∗
" ! ∗
"
HV S, z = vol f̄1 (x), z1 × . . . × f̄m (x), zm
algorithms on a BOAP with n = 100. It can be seen that
x∈S
the approximated objective vectors gained by TSMOA-stage1 (40)
have a more uniform distribution than those generated by
MOEA/D-WS and AUGMECON2. where f̄(x) = (f̄1 (x), . . . , f̄m (x)) is the normalized
objective vector of x [see (24)]. vol(·) is the Lebesgue
C. Comparison on the Second Stage
measure. In general, the higher the HV value, the better
In the second stage, the algorithm needs to approximate the performance of the algorithm.
the RoI at a fine-grained level. However, the RoI is deter- Table VIII in the supplementary material shows the mean
mined by the DM, and the performance of the algorithms HV values achieved by the three algorithms with different
may depend on its location. Taking these into consideration, SOPs settings in approximating each interval of the three
we generate ten equal-sized successive intervals (denoted as BOAPs with 20 independent data. It can be observed that
{RGN1 , . . . , RGN10 }) from f2min to f2max , and the ith interval TSMOA-stage2 outperforms AUGMECON2 and MOEA/D-
RGNi for i = 1, . . . , 10 can be expressed as NUMS on almost all the test instances. Only on the intervals
 
i − 1  max  i  max RNG8 and RNG10 , TSMOA-stage2 is slightly worse than
f2 +
min
f − f2
min
, f2 +
min
f − f2min
. AUGMECON2, but the statistical significance tests indicate
10 2 10 2
(38) that the performance differences are not remarkable.
Fig. 6 plots the objective vectors obtained by the three algo-
We take each interval as an RoI and compare algorithms on. rithms with SOPs = 5 on the interval RNG8 of a BOAP with
MOEA/D-WS is not compared here, as it cannot search in n = 150. It can be found that the objective vectors obtained
a specific subregion of the Pareto front. We include another by TSMOA-stage2 have a more uniform distribution than
preference-based MOIP algorithm, MOEA/D-NUMS [44], for those gained by AUGMECON2. MOEA/D-NUMS only attains
comparison. MOEA/D-NUMS maps the uniformly distributed one objective vector inside the RoI. Besides, AUGMECON2
weight vectors to the RoI determined by a local reference vec- yields two weakly nondominated objective vectors; whereas,
tor zr = (f1 (xup ), f2 (xlow )) . The SOP in MOEA/D-NUMS the objective vectors achieved by TSMOA-stage2 are all
is constructed with the augmented weighted Tchebycheff nondominated ones.

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: MULTIOBJECTIVE OPTIMIZATION-AIDED DECISION-MAKING SYSTEM FOR LARGE-SCALE MP 8337

B. TSMOA-Stage2 Versus MOEA/D-NUMS

The differences between TSMOA-stage2 and MOEA/D-
NUMS can be summarized as follows.
1) TSMOA-stage2 uses the upper and lower bounds of one
objective to integrate the DM’s preference for the RoI.
In comparison, MOEA/D-NUMS incorporates the DM’s
preference through a local reference objective vector.
2) TSMOA-stage2 carefully achieves the boundary-
objective vectors of the RoI, while MOEA/D-NUMS
may obtain objective vectors outside of the RoI.
3) TSMOA-stage2 employs the -constraint method to con-
struct SOPs. In contrast, MOEA/D-NUMS builds SOPs
Fig. 6. Plot of the objective vectors obtained by the three algorithms with using the augmented weighted Tchebycheff method. The
SOPs = 5 on the interval RNG8 of a BOAP with n = 150. latter integrates the max function in the SOPs, thus
making them difficult to solve via MPS strategies.
4) The control parameters in the SOPs of MOEA/D-NUMS
are fixed values given by experience. Whereas, those
parameters of TSMOA-stage2 are estimated from objec-
tive vectors achieved previously, making it less troubled
by weakly nondominated objective vectors.

C. Parameter Sensitivity Analysis

SOPmax and Imax are the control parameters of TSMOA-
stage1 and TSMOA-stage2, respectively. To investigate how
they affect the performance of the two algorithms, we have
tried to cover a wide range of values for each parameter. 10,
15, 20, 25, and 30 are considered for SOPmax , while 3, 5, 7,
9, and 11 are considered for Imax .
Fig. 7. Example of SOPs constructed by TSMOA-stage1 and MOEA/D-WS. We have performed TSMOA-stage1 with the five SOPmax
settings on the three BOAPs of Section VI. MOEA/D-WS
and AUGMECON2 have also been performed on the three
VII. D ISCUSSION BOAPs for comparison. Fig. 10 in the supplementary material
A. TSMOA-Stage1 Versus MOEA/D-WS plots the mean SNrate and SP values achieved by the three
The SOPs of TSMOA-stage1 are constructed differently algorithms on each BOAP with 20 independent data. We can
from those of MOEA/D-WS. This is mainly due to the observe that TSMOA-stage1 obtains better SP results than the
following two concerns. other two algorithms with any parameter setting. As for the
1) As shown in Fig. 7, the optimization of the boundary SNrate metric, TSMOA-stage1 always outperforms MOEA/D-
SOPs of MOEA/D-WS may lead to weakly nondomi- WS and AUGMECON2 except when SOPmax is set to a very
nated objective vectors (e.g., I and J) rather than the small value. As explained earlier, when SOPmax is very small,
real nondominated ones (e.g., A and H). TSMOA-stage1 both TSMOA-stage1 and MOEA/D-WS can achieve the best
can overcome such a drawback, as it uses the lexico- mean SNrate values.
graphic optimization method is employed to deal with As for the investigation of Imax , we have taken the three
the boundary SOPs. As stated in [44] and [45], the BOAPs with RNG2 , RNG5 , and RNG8 as the test instances.
achievement of boundary nondominated objective vec- The three algorithms: 1) TSMOA-stage2; 2) AUGMECON2;
tors is very important for interactive or preference-based and 3) MOEA/D-NUMS are tested with five different Imax
MOIP algorithms. settings (i.e., 3, 5, 7, 9, 11) on each of them. Their mean HV
2) For the nonboundary SOPs, MOEA/D-WS constructs values over 20 independent data have been shown in Fig. 11 in
them using the uniformly distributed weight vectors. the supplementary material. It clearly indicates that TSMOA-
However, these SOPs’ optimal objective vectors (e.g., C, stage2 works better than the other two algorithms on almost all
E, and G) have a poor distribution when dealing with the cases. Only on RNG8 of the BOAP with n = 100, TSMOA-
the problem whose Pareto front is extremely convex or stage2 is slightly worse than AUGMECON2 for Imax = 3.
of dissimilar objective function magnitudes. In contrast,
TSMOA-stage1 uses the Aneja and Nair method to gen- VIII. C ONCLUSION
erate SOPs for optimization, thereby can yield nondom- A real-world MP task has been investigated in this arti-
inated objective vectors with better distribution (e.g., B, cle. The MP task aims to maximize the order fulfillment rate
D, and F). The experimental study in Section VI-B has and minimize the total cost by making each plant’s production
also demonstrated this viewpoint. plan, deploying the items (raw materials or semi manufactured

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.
8338 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 8, AUGUST 2022

products) between the plants, and supplying the items (man- [6] C. Purcaru, R.-E. Precup, D. Iercan, L.-O. Fedorovici, R.-C. David,
ufactured products) to customers over a 30-day horizon. The and F. Dragan, “Optimal robot path planning using gravitational search
algorithm,” Int. J. Artif. Intell., vol. 10, no. S13, pp. 1–20, 2013.
MP task involves a large number of decision variables, as it [7] G. Wen, S. S. Ge, C. P. Chen, F. Tu, and S. Wang, “Adaptive tracking
covers many items of multiple plants over 30 days. Moreover, control of surface vessel using optimized backstepping technique,” IEEE
its objectives have extremely different magnitudes, and the Trans. Cybern., vol. 49, no. 9, pp. 3420–3431, Sep. 2019.
[8] W. Liu, Y. Dong, F. Chiclana, F. J. Cabrerizo, and E. Herrera-Viedma,
value ranges change drastically depending on the MP data. In “Group decision-making based on heterogeneous preference relations
addition, some particular constraints in the order fulfillment with self-confidence,” Fuzzy Optim. Decis. Making, vol. 16, no. 4,
rate definition make the MP task hard to solve. pp. 429–447, 2017.
[9] B. Xu, Y. Zhang, D. Gong, Y. Guo, and M. Rong, “Environment
To address the MP task, we have proposed an interactive sensitivity-based cooperative co-evolutionary algorithms for dynamic
multiobjective optimization-based MP system. The system first multi-objective optimization,” IEEE/ACM Trans. Comput. Biol. Bioinf.,
formulates the submitted MP task as a MOIP problem whose vol. 15, no. 6, pp. 1877–1890, Nov./Dec. 2017.
[10] Z. Wang, Q. Zhang, A. Zhou, M. Gong, and L. Jiao, “Adaptive replace-
two minimization objective functions are negative order ful- ment strategies for MOEA/D,” IEEE Trans. Cybern., vol. 46, no. 2,
fillment rate and total cost. Then, the system solves it using pp. 474–486, Feb. 2015.
a TSMOA. [11] Z. Wang, Y.-S. Ong, J. Sun, A. Gupta, and Q. Zhang, “A generator for
multiobjective test problems with difficult-to-approximate Pareto front
To alleviate the computational cost and the DM’s cognitive boundaries,” IEEE Trans. Evol. Comput., vol. 23, no. 4, pp. 556–571,
burden as much as possible without compromising problem- Aug. 2019.
solving, TSMOA adopts a two-stage interaction. TSMOA-stage1 [12] Q. Zhang, H. Li, D. Maringer, and E. Tsang, “MOEA/D with NBI-style
uses the lexicographic optimization method and the Aneja Tchebycheff approach for portfolio management,” in Proc. IEEE Congr.
Evol. Comput. (CEC), 2010, pp. 1–8.
and Nair method to construct the boundary SOPs and a [13] X. Li, M. Yuan, D. Chen, J. Yao, and J. Zeng, “A data-driven three-layer
few widely distributed nonboundary SOPs, respectively. These algorithm for split delivery vehicle routing problem with 3D container
SOPs’ optimal solutions can give the DM a holistic perception loading constraint,” in Proc. 24th ACM SIGKDD Int. Conf. Knowl. Disc.
Data Min., 2018, pp. 528–536.
of the entire MP, thereby enabling the DM to determine a RoI [14] M. J. Alves and J. Clímaco, “A review of interactive methods for
in the objective space. Within the RoI, TSMOA-stage2 adopts multiobjective integer and mixed-integer programming,” Eur. J. Oper.
a modified -constraint method to generate a uniformly dis- Res., vol. 180, no. 1, pp. 99–115, 2007.
[15] G. Mavrotas and K. Florios, “An improved version of the augmented
tributed SOPs. These SOPs’ optimal solutions can approximate -constraint method (AugmeCon2) for finding the exact Pareto set in
the RoI at a fine-grained level. From them, the DM chooses multi-objective integer programming problems,” Appl. Math. Comput.,
the most suitable one as the final solution of the MP task. vol. 219, no. 18, pp. 9652–9669, 2013.
[16] M. Visée, J. Teghem, M. Pirlot, and E. L. Ulungu, “Two-phases method
Along with TSMOA, a specifically designed SOP solver has and branch and bound procedures to solve the bi–objective knapsack
been adopted in the MP system to reduce the calculation work- problem,” J. Global Optim., vol. 12, no. 2, pp. 139–155, 1998.
load significantly. In the strategy, we use two modified easier [17] A. Przybylski, X. Gandibleux, and M. Ehrgott, “A two phase method for
multi-objective integer programming and its application to the assign-
SOPs as the approximator of the original complex SOP and uti- ment problem with three objectives,” Discr. Optim., vol. 7, no. 3,
lize Gurobi to solve them sequentially. We have demonstrated pp. 149–165, 2010.
this strategy is efficient on real-world MP applications. [18] S. C. Narula and V. Vassilev, “An interactive algorithm for solving
multiple objective integer linear programming problems,” Eur. J. Oper.
Although TSMOA is oriented to real-world applications, Res., vol. 79, no. 3, pp. 443–450, 1994.
its development is based on the theoretical analysis. We first [19] D. Gong, J. Sun, and X. Ji, “Evolutionary algorithms with preference
carefully summarized the main challenges (e.g., many decision polyhedron for interval multi-objective optimization problems,” Inf. Sci.,
vol. 233, pp. 141–161, Jun. 2013.
variables and dissimilar and changeable magnitudes) of the MP
[20] M. Ehrgott and M. M. Wiecek, “Multiobjective programming,” in
task and reviewed the existing technologies accordingly. Based Multiple Criteria Decision Analysis: State of the Art Surveys. Cham,
on these technologies, the algorithm with general applicabil- Switzerland: Springer, 2005, pp. 667–708.
ity for similar problems was then proposed. The experimental [21] J. Shi, Q. Zhang, and J. Sun, “PPLS/D: Parallel Pareto local search based
on decomposition,” IEEE Trans. Cybern., vol. 50, no. 3, pp. 1060–1071,
studies on benchmark problems and the application of portfo- Mar. 2020.
lio optimization (see Section V in the supplementary material) [22] M. Laumanns, L. Thiele, and E. Zitzler, “An efficient, adaptive parameter
have demonstrated the effectiveness of TSMOA. variation scheme for metaheuristics based on the -constraint method,”
Eur. J. Oper. Res., vol. 169, no. 3, pp. 932–942, 2006.
[23] R. E. Steuer, Multiple Criteria Optimization: Theory, Computation, and
Application, vol. 233. New York, NY, USA: Wiley, 1986.
R EFERENCES [24] M. Ehrgott, Multicriteria Optimization, vol. 491. New York, NY, USA:
[1] F. R. Jacobs, W. L. Berry, D. C. Whybark, and T. E. Vollmann, Springer, 2005.
Manufacturing Planning and Control for Supply Chain Management. [25] Ö. Özpeynirci and M. Köksalan, “An exact algorithm for finding
New York, NY, USA: McGraw-Hill, 2011. extreme supported nondominated points of multiobjective mixed integer
[2] D. Gong, Y. Han, and J. Sun, “A novel hybrid multi-objective artificial programs,” Manag. Sci., vol. 56, no. 12, pp. 2302–2315, 2010.
bee colony algorithm for blocking lot-streaming flow shop scheduling [26] A. Przybylski, X. Gandibleux, and M. Ehrgott, “Two phase algorithms
problems,” Knowl. Based Syst., vol. 148, pp. 115–130, May 2018. for the bi-objective assignment problem,” Eur. J. Oper. Res., vol. 185,
[3] A. Cheraghalikhani, F. Khoshalhan, and H. Mokhtari, “Aggregate pro- no. 2, pp. 509–533, 2008.
duction planning: A literature review and future research directions,” Int. [27] K.-H. Chang, “Multiobjective optimization and advanced topics,”
J. Ind. Eng. Comput., vol. 10, no. 2, pp. 309–330, 2019. Dec. 2015, pp. 325–406.
[4] D. Gong, B. Xu, Y. Zhang, Y. Guo, and S. Yang, “A similarity- [28] Y. P. Aneja and K. P. Nair, “Bicriteria transportation problem,” Manag.
based cooperative co-evolutionary algorithm for dynamic interval Sci., vol. 25, no. 1, pp. 73–78, 1979.
multiobjective optimization problems,” IEEE Trans. Evol. Comput., [29] J. L. Cohon, Multiobjective Programming and Planning. New York, NY,
vol. 24, no. 1, pp. 142–156, Feb. 2020. USA: Academic, 1978.
[5] C. P. Tautenhain, A. P. Barbosa-Povoa, and M. C. Nascimento, “A multi- [30] B. Xin, L. Chen, J. Chen, H. Ishibuchi, K. Hirota, and B. Liu,
objective matheuristic for designing and planning sustainable supply “Interactive multiobjective optimization: A review of the state-of-the-
chains,” Comput. Ind. Eng., vol. 135, pp. 1203–1223, Sep. 2019. art,” IEEE Access, vol. 6, pp. 41256–41279, 2018.

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: MULTIOBJECTIVE OPTIMIZATION-AIDED DECISION-MAKING SYSTEM FOR LARGE-SCALE MP 8339

[31] J. A. Nelder and R. Mead, “A simplex method for function Jingda Deng received the B.Sc. and M.Sc. degrees
minimization,” Comput. J., vol. 7, no. 4, pp. 308–313, 1965. in mathematics from Xi’an Jiaotong University,
[32] P. T. Boggs and J. W. Tolle, “Sequential quadratic programming,” Acta Xi’an, China, in 2012 and 2015, respectively, and
Numerica, vol. 4, pp. 1–51, 1995. the Ph.D. degree in computer science from the City
[33] L. Costa and P. Oliveira, “Evolutionary algorithms approach to the University of Hong Kong, Hong Kong, in 2020.
solution of mixed integer non-linear programming problems,” Comput. He is currently an Assistant Professor with the
Chem. Eng., vol. 25, nos. 2–3, pp. 257–266, 2001. School of Mathematics and Statistics, Xi’an Jiaotong
[34] F. Glover and M. Laguna, “Tabu search,” in Handbook of Combinatorial University. His main research interests include evo-
Optimization. New York, NY, USA: Springer, 1998, pp. 2093–2229. lutionary computation, multiobjective optimization,
[35] F. Neri, C. Cotta, and P. Moscato, Handbook of Memetic Algorithms, machine learning, and their applications.
vol. 379. Berlin, Germany: Springer, 2011.
[36] Gurobi Optimizer Reference Manual 2015, Gurobi Optim., Inc.,
Houston, TX, USA, 2014.
[37] E. L. Ulungu and J. Teghem, “Multi-objective combinatorial
optimization problems: A survey,” J. Multi Criteria Decis. Anal., vol. 3,
Qingfu Zhang (Fellow, IEEE) received the B.Sc.
no. 2, pp. 83–104, 1994.
degree in mathematics from Shanxi University,
[38] H. W. Kuhn, “The Hungarian method for the assignment problem,”
Taiyuan, China, in 1984, and the M.Sc. degree
Naval Res. Logist. Quart., vol. 2, nos. 1–2, pp. 83–97, 1955.
in applied mathematics and the Ph.D. degree in
[39] K. Fukuda and T. Matsui, “Finding all the perfect matchings in bipartite
information engineering from Xidian University,
graphs,” Appl. Math. Lett., vol. 7, no. 1, pp. 15–18, 1994.
Xi’an, China, in 1991 and 1994, respectively.
[40] Q. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary algorithm
He is a Chair Professor of Computational
based on decomposition,” IEEE Trans. Evol. Comput., vol. 11, no. 6,
Intelligence with the Department of Computer
pp. 712–731, Dec. 2007.
Science, City University of Hong Kong, Hong
[41] I. Das and J. E. Dennis, “Normal-boundary intersection: A new method
Kong. His main research interests include evolution-
for generating the Pareto surface in nonlinear multicriteria optimization
ary computation, multiobjective optimization, neural
problems,” SIAM J. Optim., vol. 8, no. 3, pp. 631–657, 1998.
networks, data analysis, and their applications.
[42] J. R. Schott, “Fault tolerant design using single and multicriteria genetic
Dr. Zhang has been a Web of Science Highly Cited Researcher in Computer
algorithm optimization,” Cell. Immunol., vol. 37, no. 1, pp. 1–13, 1995.
Science for five consecutive years since 2016. He is an Associate Editor of the
[43] S. Jiang, Y.-S. Ong, J. Zhang, and L. Feng, “Consistencies and contra-
IEEE T RANSACTIONS ON E VOLUTIONARY C OMPUTATION and the IEEE
dictions of performance metrics in multiobjective optimization,” IEEE
T RANSACTIONS ON C YBERNETICS.
Trans. Cybern., vol. 44, no. 12, pp. 2391–2404, Dec. 2014.
[44] K. Li, M. Liao, K. Deb, G. Min, and X. Yao, “Does preference always
help? A holistic study on preference-based evolutionary multi-objective
optimisation using reference points,” IEEE Trans. Evol. Comput.,
vol. 24, no. 6, pp. 1078–1096, Dec. 2020.
[45] K. Li, R. Chen, G. Min, and X. Yao, “Integration of preferences Xijun Li received the M.E. degree from Shanghai
in decomposition multiobjective optimization,” IEEE Trans. Cybern., Jiao Tong University, Shanghai, China, in 2018.
vol. 48, no. 12, pp. 3359–3370, Dec. 2018. He is currently pursuing the Ph.D. degree with the
[46] E. Zitzler and L. Thiele, “Multiobjective optimization using evolution- University of Science and Technology of China
ary algorithms—A comparative case study,” in Proc. Int. Conf. Parallel (HUAWEI-USTC Joint Ph.D. Program), Hefei,
Problem Solving Nat. (PPSN), 1998, pp. 292–301. China.
He is currently a Senior Research Engineer with
Huawei Noah’s Ark Lab, Hong Kong. He has
published several papers on top peer-reviewed con-
ferences (KDD, CIKM, and ICDCS). His research
interests focus on learning to optimize combinatorial
optimization problem and machine learning for computer systems.
Zhenkun Wang (Member, IEEE) received the
Ph.D. degree in circuits and systems from Xidian
University, Xi’an, China, in 2016.
From 2017 to 2020, he was a Postdoctoral
Research Fellow with the School of Computer Mingxuan Yuan received the Ph.D. degree from the
Science and Engineering, Nanyang Technological Hong Kong University of Science and Technology,
University, Singapore, and the Department of Hong Kong.
Computer Science, City University of Hong Kong, He is a Principal Researcher with Huawei Noah’s
Hong Kong. He is currently an Assistant Professor Ark Lab, Hong Kong. He has led several projects
with the School of System Design and Intelligent in telecommunication data mining and supply chain
Manufacturing and the Department of Computer optimization. His research interests include spa-
Science and Engineering, Southern University of Science and Technology, tiotemporal data analytics and enterprise operation
Shenzhen, China. His research interests include evolutionary computation, optimization models.
multiobjective optimization, machine learning, and their applications.
Dr. Wang is an Associate Editor of the Swarm and Evolutionary
Computation.

Jia Zeng (Senior Member, IEEE) received the

B.Eng. degree from the Wuhan University of
Hui-Ling Zhen received the B.S. degree in numer- Technology, Wuhan, China, in 2002, and the Ph.D.
ical mathematics and the Ph.D. degree in applied degree from the City University of Hong Kong,
mathematics from the Beijing University of Posts Hong Kong, in 2007.
and Telecommunications, Beijing, China, in 2011 He is currently the AI Chief Scientist of Enterprise
and 2016, respectively. Intelligence (e.g., supply chain management) with
She is currently working with Huawei Noah’s Huawei Noah’s Ark Lab, Hong Kong, where he is
Ark Lab, Hong Kong. Her current research interests also the Director of the AI Planning and Cooperation
include large-scale optimization, constraint program- Department. His research interests are machine
ming, as well as their applications in supply chain learning and big data applications.
management and chip design. Dr. Zeng is a member of CCF and ACM.

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 03,2024 at 08:05:41 UTC from IEEE Xplore. Restrictions apply.