Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

Evolutionary Intelligence (2019) 12:253–272

https://doi.org/10.1007/s12065-019-00222-9

RESEARCH PAPER

A gradual weight‑based ant colony approach for solving


the multiobjective multidimensional knapsack problem
Imen Ben Mansour1 · Ines Alaya1 · Moncef Tagina1

Received: 9 May 2018 / Revised: 14 December 2018 / Accepted: 6 March 2019 / Published online: 25 March 2019
© Springer-Verlag GmbH Germany, part of Springer Nature 2019

Abstract
The multiobjective multidimensional knapsack problem (MOMKP) is an extension of the multiobjective knapsack problem
that consists in selecting a subset of items in order to maximize m objective functions. The MOMKP creates an additional
difficulty than the monodimensional version caused by the fact of respecting more than one constraint simultaneously. In
this paper, we propose to solve the MOMKP with an ant colony optimization approach based on a gradual weight generation
method, named Gw-ACO. Here, the weight vectors are gradually distributed in the objective space and change relatively to
the optimization process. This enables ants to target, at each cycle, different regions in order to try to achieve almost all solu-
tions covering the Pareto front. To evaluate the suggested Gw-ACO approach, a set of experiments is performed on MOMKP
benchmark instances and compared with well-known state-of-the-art metaheuristic approaches. The obtained experimental
results show that Gw-ACO is significantly better and able to achieve a well distribution all over the Pareto-optimal front.

Keywords Ant colony optimization · Multiobjective multidimensional knapsack problem · Weight-based method · Weight
vector generation method

1 Introduction To solve KP variants, several approaches based on


metaheuristics were proposed in the literature. Zitzler et al.
Nature-inspired metaheuristics are widely used to solve NP- [10] introduced the well-known SPEA2 which is an elit-
hard problems [1–5], since the exact methods require high ist algorithm, based on a ranking dominance procedure
implementation time. The multiobjective multidimensional to solve the MOMKP. In [11], the NSGA-II is proposed,
knapsack problem (MOMKP) is a variant of the knapsack another well-known multiobjective evolutionary algorithm
problem (KP). The MOMKP is one of the hardest multiob- that uses also an elitist approach and a different ranking
jective combinatorial optimization problems (MOP) [6]. It dominance procedure for the same problem variant. A
modelizes many real world problems such as resource allo- memetic algorithm was described in [12], integrating tabu
cation [7] and budget allocation [8]. Moreover, MOMKP search called MEMOTS. To select parents located in region
can be modeled as a subproblem of other problems, such of minimal density for recombination, the algorithm uses a
as the flight crew choice [9]. The main goal of this problem dynamic hyper grid in the objective space, for the offspring
is to choose a subset of items in order to optimize several a tabu search method was applied. In [13], a guided local
objectives while not exceeding a set of capacity constraints. search algorithm was developed named GPLS, this method
combines a Pareto local search algorithm with a guided
local search method based on penalization and applied to
* Imen Ben Mansour MOMKP.
imen.benmansour@ensi‑uma.tn The ant colony optimization (ACO) and its multiobjective
Ines Alaya version have shown their efficiency in many optimization
ines.alaya@ensi‑uma.tn problems [14–17], it has been as well successfully applied
Moncef Tagina to the different variants of KP. In [18], a hybrid monobjec-
moncef.tagina@ensi‑uma.tn tive ant colony optimization was proposed (MMACS) to
solve the Strongly Correlated Knapsack Problem (SCKP).
1
ENSI‑COSMOS, University of Manouba, 2010 Manouba, The method integrates an Ant Colony System component
Tunisia

13
Vol.:(0123456789)
254 Evolutionary Intelligence (2019) 12:253–272

into Max–Min Ant System metaheuristic by modifying its called MOSA. A set of weights is uniformly generated and
functionality. MMACS uses a proposed 2-opt algorithm to assigned to each solution in order to find a good approxi-
improve the best solutions found by ants. In [19], the authors mation of the Pareto front. Paquete and Stützle [24] intro-
define a monobjective ant colony optimization method duced TPLS, a Two-Phase Local Search method, to solve
with several start strategies for the monobjective multi- the biobjective quadratic assignment problem. The idea of
dimensional Knapsack problem. The proposed approach TPLS is to collect several local optima to a sequence of
employs different start strategies when an ant chose first scalarizations of the objective function vector. Each solu-
node, from which to start to create a solution. The author tion returned for one scalarization, using one weight vector,
in [20] presents an ant colony system based heuristic for serves as the starting solution of the local search for the
the multidimensional 0–1 knapsack problem. The approach next scalarization in the sequence. More recently, in [25] a
was adapted to the mono and multiobjective variants of multiobjective evolutionary algorithm based on decomposi-
KP. In [21], a generic multiobjective ACO algorithm called tion, named MOEA/D, was presented. MOEA/D employs
m-ACO, is proposed and instantiated with four variants. This weighted aggregation approaches where the multiobjective
algorithm is parameterized by the number of ant colonies problem is decomposed into a number of single objective
and the number of the pheromone structures. In [22], an optimization subproblems. In MOEA/D the neighborhood
ACO algorithm named IBACO was introduced. The IBACO relations among the subproblems are based on the distances
algorithm employs binary indicators. It uses hypervolume among their aggregation weight vectors. In [26], authors
and epsilon indicators to guide the artificial ants to find the proposed the two-Phase Pareto Local Search (2PPLS). As
best solutions by laying pheromone trails relatively to the indicated by its name, 2PPLS is composed of two phases. In
indicator values. the first phase, an initial population of potentially efficient
The above mentioned approaches assume that, during the solutions is produced, then as a second phase the Pareto-
optimization phase, the relative importance of every objec- based local search (PLS) [27, 28] is applied to this popula-
tive function in the Pareto front is generally equal. This tion. 2PPLS uses a very-large scale neighborhood (VLSN)
may place the exploration of distributed solutions cover- [29], based on uniformly distributed weight vectors to define
ing the Pareto-optimal front in question. To solve this prob- the neighborhood function.
lem one may force ants to search in different regions of the Several proposed MOACO approaches deal with design-
Pareto front using weight vectors. Based on the ant colony ing weight-based methods. Iredi et al. proposed in [30]
metaheuristic and inspired by the weight-based methods, we the BicriterionAnt algorithm, which is probably the first
propose in this paper Gw-ACO. In this proposed algorithm, MOACO algorithm using the weighted objectives method to
the relative importance of every objective is weighted dif- guide ants to search in different regions of the Pareto front.
ferently during the search, the ant colony is orientated to In this algorithm, the weight vectors are defining according
look in different regions of the non-dominated front thanks to the ants of the colony. Different ant groups target differ-
to a proposed Gradual weight generation method, called Gw. ent points on the non-dominated front and try to approxi-
This method generates a set of weight vectors, gradually mate the Pareto set in each cycle. In [31], the Crowding
distributed in the first orthant of the objective space and Population-based ACO algorithm (CPACO) was defined.
changes according to the optimization process. Our motiva- The algorithm uses a random weight generation method to
tions for using a weighted ACO are to promote diversity of determine weightings for each objective. It assigns each ant a
the search space and to speed up the convergence of ants unique weight vector to explore the search space. In [32], an
towards the Pareto-optimal set. Ants explore new areas in a ant colony optimization approach based on decomposition
cooperative way and try to approximate a part of the Pareto method called MOEA/D-ACO is introduced. It convert a
front at each generation. Furthermore, we use, in our study, multi-objective problem into a set of single objective prob-
a random weight-based variant of the proposed ACO algo- lems by means of uniformly distributed weight vectors. In
rithm in order to investigate the impact of the weight vector this approach, each ant targets a particular point in the Pareto
generation method and the exploration strategy. front and each group of ants tries to approximate a particular
The weight vector generation is the first step in decom- part of the Pareto front. MOEA/D-ACO uses neighborhood
position/aggregation based approaches. These weights rep- that enables collaboration between groups and between the
resent search directions in the objective space on a decom- members of the same group.
position context or a weighted aggregation of the different The main contributions of this paper could be summa-
objectives. Those approaches have been widely studied in rized as follows. The first contribution is the weight vec-
the literature, adopting the idea of the weight vector vari- tor generation method Gw, hence the setting of the search
ation during the optimization process: in [23] authors pro- directions and their distribution on the Pareto front. This
posed a multiobjective simulated annealing method based on method aims to maintain diversity of the population. The
scalarizing function for the multiobjective knapsack problem second is its application to ACO by defining the exploration

13
Evolutionary Intelligence (2019) 12:253–272 255

strategy adopted by ants, i.e. the ants’ decision mechanism The goal of a MOP is to find the set of all non-dominated
and the pheromone update rules. Indeed, the major differ- solutions called Pareto set. When using a metaheuristic
ences between the proposed Gw-ACO approach, the Bic- approach, the goal is to find a Pareto set approximation.
riterionAnt, CPACO and MOEA/D-ACO lie on these two
aspects: the Gw method and the exploration strategy. In Bic- 2.2 MOMKP formulation
riterionAnt, CPACO and MOEA/D-ACO the weights are
generated according to ants, the Pareto front is decomposed The multiobjective multidimensional knapsack problem
into the number of ants. During the same generation, each could be formulated as follows:
ant targets one particular point in the Pareto front. While in
Gw-ACO, weights change with the optimization process. ∑
n
Maximize pkj xj k = 1, … , m (2)
They are generated gradually according to each generation, j=1
the Pareto front is divided under the maximum number of
generations, each weight vector is assigned to one generation
and all ants target the same direction during this generation. ∑
n
Subject to wij xj ≤ bi i = 1, … , q
The remaining of the paper is organized as follows. In j=1 (3)
Sect. 2, we define briefly the multiobjective optimization
xj ∈ {0, 1} j = 1, … , n
problems, we formulate the studied problem and we present
the multiobjective ant colony optimization. In Sect. 3, we
detail the proposed approach Gw-ACO and we define its where n is the number of items, for each item Ij is assigned
different functions. The parameters of the algorithm are set, a decision variable xj equal to 1 if the item is selected, 0
the experimental studies and comparisons are provided in otherwise. Each item Ij has a profit pkj relatively to the objec-
Sect. 4. Finally, we end the paper in Sect. 5 by conclusion tive k and a weight wij relatively to the resource i. The aim of
and our future works. the problem is to select a subset of items in order to maxi-
mize m objective functions while not exceeding q resource
constraints. bi is the total quantity available for the resource
i.
2 Preliminaries In this paper, the considered instances assume that the
number of objective functions is equal to the number of
2.1 Multiobjective optimization problem resources (m = q).

The purpose of a multiobjective optimization problem 2.3 Multiobjective ant colony optimization


(MOP) is to optimize a vector F of m objective functions
simultaneously. More formally, a MOP can be defined as The ant colony optimization (ACO) has been success-
follows: fully applied to many optimization problems [33, 34]. This
{ metaheuristic is inspired from the behavior of real ants,
max f (x) = (f1 (x), … , fm (x))
(MOP) =
s.t. x ∈ X (1) where ants move from their nest to a food source using the
shortest path. To find this path, ants follow a chemical sub-
Without loss of generality, we assume that the m ≥ 2 objec- stance called pheromone laid by other ants.
tive functions have to be maximized. Let X denote the set The first ACO algorithm: Ant System for the trave-
of solutions in decision space, and x = ( x1,..., xn ) ∈ X is a ling salesmen problem, was proposed by Dorigo [35]. In
vector representing the decision variables. For each solution the algorithm, an artificial ant starts with a random solu-
x ∈ X is assigned one objective vector z ∈ Z on the basis tion component, then adds iteratively solution components
of a vector function f ∶ X ⟶ Z such as z = ( z1,...,zm ) = regarding to a probability transition depending on phero-
f (x) = (f1 (x), … , fm (x)), where Z = f (X) represents the set mone trails in the solution component accumulated previ-
of feasible points of the objective space. To evaluate the ously by the colony, and a heuristic information that gives an
quality of a solution in a MOP, we define the following order idea about the quality of the solution component.
relations: These two factors, pheromone trails and heuristic infor-
mation are weighted with two parameters that specify their
• A decision vector x ∈ X is said to dominate another deci- relative influence in the probability transition. After all
sion vector x� ∈ X , noted x ≻ x′ , iff ∀ i ∈ {1, … , m}, fi(x) ants have constructed their solutions, the pheromone trails
≥ fi (x� ) and ∃ j ∈ {1, … , m}, fj(x) > fj (x� )(. are updated. To prevent unlimited accumulation of phero-
• x ∈ X is Pareto-optimal or non-dominated, iff there is not mone, the pheromone trails decrease by applying evapora-
any solution that dominates it; iff ∀x� ∈ Xx�, ⊁ x. tion ratio. The pheromone trails reflect the desirability to

13
256 Evolutionary Intelligence (2019) 12:253–272

visit a search area. In the monobjective case, these trails are in every direction. The purpose of this search strategy is
defined according to the function to optimize. In the multi- to allow a significant number of ants to explore the same
objective case, we can consider two strategies. A first one is region. So, at each cycle of the algorithm, every generation
to consider a single pheromone trail as used in [14, 22, 31, of ants tries to approximate a particular part of the Pareto
36–38] in this case, the pheromone trails are defined as an front. As a first step, the pheromone trails are initialized to
aggregation of the objectives to optimize. The second one is an initial amount then, in a second step, the set 𝛬 contain-
to consider several pheromone trails as in [39–43], usually in ing all weight vectors of size L is generated according to the
this case the researchers use several colonies where each one Gw method. At each cycle (generation) all ants select one
uses its own pheromone trail. To construct their solutions, search direction according to the generation index from the
ants use probability transition which is defined according 𝛬 set. Each ant builds its solution according to a probabilistic
to the pheromone factor and heuristic factor. In the multi- function. Then, the archive is updated with the non-domi-
objective case, the pheromone factor is frequently defined nated solutions found during the current generation and the
relatively to the considered strategy of the pheromone trails. pheromone trails of the archive member are updated with an
The heuristic factor is defined, as for the pheromone trails, amount of pheromone. The algorithm stops iterating when
according to two strategies. Either an aggregation of differ- a maximum number of generations gmax is attained and the
ent objectives in one heuristic information or associating for non-dominated set of all the generations of the algorithm
each objective a heuristic information. is returned. More formally, Gw-ACO can be described in
In this paper, to generate new solutions, ants in Gw-ACO Algorithm 1. The details of the Gw method, solution con-
share a single pheromone structure which contains the struction and pheromone update steps are provided in the
amount of pheromone previously deposited by other ants following subsections.

Algorithm 1 Gw-ACO algorithm


Input: N bants (number of ants)
gmax (maximum number of generations)
m (number of objectives)
Output: P (Pareto approximation set)
Step 1 - initialization: Initialize the pheromone trails τinit ← 1 and g ← 0
Step 2 - weight vector generation: Generate the weight vectors set Λ of size L
according to Gw method
Step 3 - main loop: While g < gmax do:
1) Archive initialization: A ← ∅
2) Weight vector selection: Select weight vector λ(g)=(λ1 (g),...,λm (g)) ∈ Λ
3) Solutions construction: For h from 1 to N bants: Ant h constructs a solution Sh
with probability pSh (Ij ) defined in Eq.(8)
4) Archive update: If no solution in A  Sh add Sh to A and remove from A all
solutions dominated by Sh
5) Pheromone update: Update pheromone trails for each member of A with τ (Ij )
defined in Eq.(11)
6) Pareto approximation set update: P ← non-dominated solutions of P ∪ A
7) Generation index incrementation: g ← g + 1
Step 4 - Termination: Return the Pareto approximation set P

on the selected items. Whereas, for each objective is associ- 3.2 Weight vector generation methods
ated one heuristic information matrix.
The weight vector generation is a crucial step in our proposed
approach. Like any weight-based method, the efficiency of the
suggested Gw-ACO approach depends strongly on the gen-
3 The proposed gradual weight‑based ant erated weight vectors. In this section, first, we describe the
colony optimization approach proposed gradual weight generation method Gw. Then, we
present a random weight generation method which will be part
3.1 Algorithm description of our experimental study.

In Gw-ACO ants use weights to search in different regions


of the Pareto front, they perform a strongly intensified search

13
Evolutionary Intelligence (2019) 12:253–272 257

3.2.1 Gradual weight generation method

In Gw-ACO, ants try to reach different Pareto-optimal points


by appropriately varying the search directions in the objective
space. For this purpose, all ants search in the same direction
at the same time. They change this direction generation by
generation during the search process which allow them, first,
to explore new search areas by trying to reach any point on the
Pareto front, second to exploit almost all possible directions by
carrying out a strong intensified search.
To specify these search directions, we propose the Gw
method which generate a set of weight vectors, located in a
surface in the first orthant of the objective space and gradually
distributed in [0,1].
Let 𝜆(g) = (𝜆1 (g), … , 𝜆m (g)) be a weight vector, such that
∑m
𝜆 (g) ≥ 0 for all k = 1, … , m and k=1 𝜆k (g) = 1. For the bi-
k
Fig. 1  The weights change within 200 generations and FQ set to 800
objective case (m = 2), the weights are calculated as follow:
[( ) ( )]
4×g×e 2×𝜋×g
(4) the last generation. Figure 1 shows 𝜆1 (g) and 𝜆2 (g) values
1
𝜆 (g) = ln + cos
FQ FQ
change within 200 generations during the evolution process.
By setting the maximum number of generations larger than
𝜆2 (g) =1.0 − 𝜆1 (g) (5) the number of generated weight vectors, gmax > L, the Pareto
where g is the generation index and FQ is the weight change front is explored again using the same weight vectors. For
frequency. The variation of 𝜆(g) in [0,1] is adjusted by the clarity, taking a bi-objective problem as an example, if gmax
relative value of FQ which is has to be tuned to achieve is set to 400 and L = 200 , 𝜆 will change from 𝜆 = (0, 1) to
optimal performance. That is to say that, FQ should not be 𝜆 = (1, 0) two times.
too high so that the weight vector can move slowly from The extension of Gw method in Eqs. (4) and (5) to
one point to another trying to explore almost all regions. In optimize problems with more than two objectives is theo-
(m−1)
addition, FQ allows to set the size of 𝛬, where L = FQ 4
. retically straightforwardly. For example, if there are three
For instance, if FQ is set to 800 with m = 2, 200 weight vec- objectives, Gw proceeds as shown in algorithm 2. If there
tors are generated ( L = 200). In fact, in the first generation, are m objectives, Gw holds m − 1 loops for and is easily
𝜆1 (g) and 𝜆2 (g) are set respectively to 0 and 1. These values extended as done with three objectives. Gw is executed
change gradually from generation to generation by increas- offline and only once before the main loop of Gw-ACO as
ing 𝜆1 and decreasing 𝜆2 until 𝜆1 (g) get 1 and 𝜆2 (g) get 0 at shown in Algorithm 1.

Algorithm 2 Gradual weight generation method: Gw


Input: F Q (the frequency change)
L (the number of weight vectors to be generated)
Output: Λ (the weight vectors set)
Λ←∅
1
 
For i from0 to L m−1 do:
 
λ1 = ln 4∗i∗e
FQ
+ cos 2∗π∗i
FQ
1
 
For j from 0 to L 
m−1 do:
λ2 = (1 − λ1 ) ln 4∗j∗e + cos 2∗π∗j
  
FQ FQ
λ3 = 1 − λ1 − λ2
Λ ← Λ ∪ {λ}
end for
end for

13
258 Evolutionary Intelligence (2019) 12:253–272

The idea behind the Gw method is simple and well estab- optimization CPACO [31], adopt a random weight genera-
lished. For a clear explanation, let us consider the bi-objective tion method (Rw) which distributed the weights uniformly
case. In the first stage, 𝜆1 is generated based on the natural among the ants in the population. In other words, a weight
logarithm function ln(x), this choice was guided by the fact vector generated by a random function is assigned to each
that ln(x) is a continuous and a strictly increasing function in ant in the population when constructing its solution. In order
]0; +∞[. As it is well-known ln(1) = 0 and ln(e) = 1, the to study the efficiency of the Gw method, we will apply the
weights in the proposed Gw method are generated with respect Rw method as a variant of our proposed ACO approach and
to these two proprieties. In fact, the principal role of the param- we compare it thereafter to the Gw method. If there are two
eters in Eq. (4) is to define the weights in [0,1] and to smoothly objectives, the random weights are generated as follow:
change these weights in this same range. Equation (4) could
be divided in two main parts. Let ( 4×g×e ) be the first part
𝜆1h (g) =random[0, 1] (6)
FQ
denoted as p and cos( 2×𝜋×g
FQ
) is the second part denoted as p′, 𝜆2h (g) =1.0 − 𝜆1h (7)
where: where h = 1, … , Nbants is the ant index and random[0, 1] is
⎧ 0 if g = 0 a random function generates a uniformly distributed number
⎪ between 0 and 1. To calculate m weights for Nbants ants in
p = ⎨ e if g = FQ 4 the population, the Rw proceeds as in algorithm 3.
⎪ [0, 1] otherwise

Algorithm 3 Random weight generation method: Rw


Input: m (number of objectives)
N bants (maximum number of ants)
Output: Λ (weight vectors set)
For h from 1 to N bants do:
For k from 1 to m − 1 do:
choose randomly a number Wk in [0, 1]
end for
sort W in ascending order, such that: W1 ≤ · · · ≤ Wm−1
λ1 ← W1
For k from 2 to m − 1 do:
λk = Wk − Wk−1
end for
λm = 1 − Wm−1
Ant h ← λh = (λ1h , ..., λm
h )
end for

and As previously mentioned, in Rw the weights are uni-


formly distributed among the ants since these weights are
⎧ 1 if g = 0 randomly regenerated in every generation. While in Gw the

p� = ⎨ 0 if g = FQ 4
weights are distributed among the generations, they change
⎪ [0, 1] otherwise gradually from one generation to another through the search

process. All ants use the same weight vector during one
generation. Therefore, these two weight generation meth-
In such way, p slowly grows from 0 to e and p′ slowly ods differ not only in the weights generation, but also in
approaches 0 from 1 as g increases in the range from 0 to FQ4
. the exploration strategy. Figure 2 illustrates these two main
In the second stage, 𝜆2 is defined as the rest of 1−𝜆1. There- differences between Rw and Gw.
fore, a set of weight vectors is generated in [0,1] adjusted
according to the frequency FQ. With small modification in 3.3 Solution construction step
Gw (e.g. parameter g) it is also possible to explore a spe-
cific regions of the decision space defined by a preference Let us consider that ants construct their solutions within a con-
directions. struction graph G = (V, E), a complete graph where the set of
vertices V represents the set of items and the set of edges E
3.2.2 Random weight generation method represents the connection between the vertices V i.e. path and
where the pheromone trails are associated with vertices of this
Some multiobjective weighted ACO approaches from the graph. At each generation, each ant h constructs one feasible
literature, such as Crowding population-based ant colony

13
Evolutionary Intelligence (2019) 12:253–272 259

Fig. 2  Illustration of the two weight generation methods. a Gradual weight generation method Gw. b Random weight generation method Rw

solution, it stores the previously selected paths and prevents importance of the heuristic information denoted by 𝜂Sh (Ij ).
that each vertex is selected more than once. Ant h starts by The pheromone information measures the desirability of the
choosing randomly an item/vertex, then add next items Ij to the item Ij learned from the previous search and the heuristic
solution Sh among a set of feasible items Feas (candidate verti- information measures its utility. More formally, The solution
ces). The set Feas is updated by including items not yet added construction step is described in Algorithm 4.
and that do not violate any constraint. In order to guide ants in

Algorithm 4 Solution construction procedure


Input: n (number of items)
Output: Sh (solution of ant h)
choose randomly a first item I1 in [0, 1]
Sh ← {I1 }
Initialize(F eas)
while F eas is not empty do
Choose item Ij relatively to pSh (Ij )
Sh ← Sh ∪ {Ij }
U pdate(F eas)
end while

different directions of the Pareto front when constructing their Gw-ACO assigns to each objective a weighted heuristic
solutions, ants weight the relative influence of the m objectives information matrix, allowing to ants to explore almost all
differently, using the weights created by the Gw method. To possible directions while sharing the same pheromone trails.
choose the next item to add to the solution Sh from the feasible The heuristic information was inspired from [44]. It cor-
set Feas, a probability pSh is applied, that is defined as follows: responds to the pseudo-utility factor of the k-th objective and
is set as follows:
∏m 𝛽𝜆 k
[𝜏Sh (Ij )]𝛼 ⋅ k=1
[𝜂Sh (Ij )k ]
pSh (Ij ) = (8) pkj
∑ 𝛼 ∏m k 𝛽𝜆
k
𝜂Sh (Ij ) = (9)
[𝜏Sh (Il )] ⋅ k=1
[𝜂Sh (Il ) ] RSh (Ij )
Il ∈Feas

where 𝛼 determines the relative importance of the phero- where pkj is the profit of the item Ij relatively to the k-th
mone information denoted by 𝜏Sh (Ij ), and 𝛽 is the relative objective and RSh (Ij ) is the ratio which measures the

13
260 Evolutionary Intelligence (2019) 12:253–272

tightness of the item Ij on all the constraints i ∈ 1 … q rela- As a second set of experiments, first, Gw-ACO is evalu-
tively to the constructed solution Sh which is defined as ated against the state-of-the-art MOACO approaches: Bic-
follows: riterionAnt, IBACO, m-ACO4 and MOEA/D-ACO. Bicri-
terionAnt is one of the best existing MOACO, which was

q wij initially applied to the Single Machine Total Tardiness Prob-
RSh (Ij ) = (10) lem (SMTTP) in [30] and was thereafter applied to several
i=1
dSh (i)
bi-criteria Traveling Salesman Problem (TSP) instances [39,
∑ 45]. IBACO, an indicator-based ACO that was successfully
with dSh (i) = bi − t∈Sh rit representing the remaining quan-
tity of the resource i when the ant h has constructed its solu- applied to MOMKP in [22]. The best variant of the generic
tion Sh. ACO algorithm, m-ACO4 presented in [21] and MOEA/D-
ACO described in [32], which is one of the best MOACO
3.4 Pheromone update step approaches suggested for MOMKP and TSP and is consid-
ered as reference approach. Finally, other than MOACO,
In Gw-ACO, the pheromone trails are laid on the vertices Gw-ACO is compared against two among the best state-of-
of the construction graph (i.e. items). In order to exploit the-art multiobjective metaheuristics applied to MOMKP.
the most promising parts, ants that are allowed to update MOEA/D [25] a powerful multiobjective evolutionary
the pheromone trails are only those that have found non- approach and 2PPLS, a two-Phase Pareto Local Search,
dominated solutions during the actual cycle. To give every introduced in [26].
generation of ants the same influence in a particular part of
the Pareto front, ants that are authorized to update, lay the 4.1 Test instances and parameter settings
same amount of pheromone.
Once all ants have constructed their solutions, first, all The benchmark instances used for these experiments are
amounts of pheromone are decreased in order to simulate defined in [46] which are widely used in testing several
evaporation by multiplying the quantity of pheromone lay- multi-objective heuristics. The experiments are based on
ing on each vertex of the construction graph by a pheromone nine instances with 2, 3 and 4 objectives, in combination
persistence rate (1 − 𝜌), with 0 ≤ 𝜌 ≤ 1. Then, a quantity with 250, 500 and 750 items. The used parameter settings
of pheromone 𝛥𝜏(Ij ), equal to the current archive set size for Gw-ACO are adjusted experimentally and presented as
|A|, is laid on the non-dominated solutions SND of the cycle. follow: we have set the number of ants to 20. The weight of
More formally, the pheromone update step can be defined pheromone factor 𝛼 to 1 and the weight of heuristic factor
as follows: 𝛽 to 10. The evaporation ratio 𝜌 is set to 0.02 and 𝜏init to 1.
The maximum number of generations gmax and the weight
𝜏(Ij ) ← (1 − 𝜌) × 𝜏(Ij ) + 𝛥𝜏(Ij ) (11) change frequency FQ of Gw-ACO for all instances are given
{ in Table 1. The chosen maximum number of generations is
|A| if Ij ∈ SND
𝛥𝜏(Ij ) = justified by the fact that for every instance, the weight vector
0 otherwise
changes twice from 0 to 1.
After each pheromone update step, the Pareto set P is In order to get fair comparison, all algorithms have been
updated with the non-dominated solutions found by ants in run and tested on the same machine and we have repro-
each search direction. In this way, the non-dominated solu- duced the same experiments in literature. The source codes
tions of almost all search directions are stored in the set P. of the compared algorithms have been obtained either by
downloading them from the web sites or by asking them

4 Experimental study
Table 1  The setting of Instance gmax FQ
the maximum number of
In our experimentation, we aim to evaluate the efficiency of generations and the weight 2_250 400 800
the proposed algorithm Gw-ACO. To this end, we perform change frequency accorded to
2_500 400 800
two sets of experiments. In the first set, we study the influ- Gw-ACO for all instances
2_750 400 800
ence of the mainly parameter FQ of the Gw method, which
3_250 800 80
allows us to extract useful conclusions about this method.
3_500 800 80
Then, in order to evaluate both, the weight vector generation
3_750 800 80
method and the exploration strategy. We carry out a com-
4_250 1024 32
parative study between the proposed gradual weight genera-
4_500 1024 32
tion method Gw and the random weight generation method
4_750 1024 32
Rw, derived from the algorithm CPACO [31].

13
Evolutionary Intelligence (2019) 12:253–272 261

Table 2  The number of Algorithms


constructed solutions
by Gw-ACO, Rw-ACO, Instance Gw-ACO Rw-ACO Bicriterion Ant IBACO m-ACO4 MOEA/ D-ACO MOEA/D 2PPLS
BicriterionAnt, IBACO,
m-ACO4, MOEA/D-ACO, 2_250 8000 8000 9600 8000 300,000 45,000 75,000 20,000
MOEA/D and 2PPLS for all 2_500 8000 8000 9600 8000 300,000 60,000 100,000 10,000
instances 2_750 8000 8000 9600 8000 300,000 75,000 125,000 10,000
3_250 16,000 16,000 – 16,000 300,000 105,300 175,500 –
3_500 16,000 16,000 – 16,000 300,000 105,300 175,500 –
3_750 16,000 16,000 – 16,000 300,000 105,300 175,500 –
4_250 20,480 20,480 – 20,480 300,000 136,500 227,500 –
4_500 20,480 20,480 – 20,480 300,000 136,500 227,500 –
4_750 20,480 20,480 – 20,480 300,000 136,500 227,500 –

personally to the different authors. The parameters of these good overview of the performance of each approach, the
algorithms were set as recommended by their authors [21, average hypervolume difference for each approach and for
25, 26, 32]. each instance is computed.
For IBACO and BicriterionAnt algorithms, we always The summary attainment surface To compare the
allow an equal or a larger number of ants and maximum approaches behaviors in a graphical way, we use the
number of generations in order to construct a number of summary attainment surface [47] which is defined by the
candidate solutions superior or equal to the solutions con- objective vectors that have been attained by at least one
structed by Gw-ACO for each instance. Table 2 gives the approximation set of the tested algorithm. In the follow-
number of constructed solution by each algorithm for each ing, we compare the median (50%) attainment surface of
instance. Since this is the first application of the Bicrite- the non-dominated solutions found for 30 runs of all the
rionAnt algorithm to MOMKP, we propose to use for this compared algorithms.
algorithm the same heuristic information and the same Moreover, we use the non-parametric Mann–Whitney
pheromone update rule than Gw-ACO while respecting the statistical test described in [48]. In order to verify if the
number of the pheromone and the heuristic information difference between the tested algorithms is statistically
matrices defined in the original BicriterionAnt algorithm significant, the results of the hypervolume difference are
[30]. The rest of the parameters are tuned as suggested in analyzed through the Mann–Whitney statistical test. We
[22, 30]. Note that, as pointed out in [26], 2PPLS could not reject the null-hypothesis, if the P-value is equal to the
solve instances with more than 2 objectives in a reasonable lowest significance level. We said an algorithm A out-
time. Therefore, in the following, 2PPLS is only tested on performs an algorithm B if the P value obtained by the
the bi-objective instances. Mann–Whitney statistical test is lower than the confi-
dence level which is set to 5%.
4.2 Performance metrics In [47–50] authors have shown that these metrics
are sufficient to measure the difference in performance
In order to evaluate the efficiency of a proposed approach in between algorithms. All the computational results were
a multi-objective context, it is important to use performance obtained by the performance assessment software pack-
metrics that consider three aspects: minimal distance to the age PISA [48].
Pareto-optimal front, the spread across the Pareto-optimal
front and a good distribution of the obtained solutions along 4.3 Experimental studies on Gw‑ACO
the Pareto front. In our experimental analysis, we use the
following performance metrics: The Gw method is designed to enhance the quality of the
The hypervolume difference Let be PO∗ the set of non- generated solutions in terms of diversification and intensifi-
dominated solutions extracted from the union of all solu- cation. This first set of experiments mainly aims to study the
tions obtained from each algorithm and from the different FQ sensitivity, the weight vector generation method and the
30 runs and Zref a reference point corresponding to the exploration strategy and their impact on these two aspects.
worst value of each objective function in PO∗. To evaluate
a set Ai of solutions, we compute the difference between 4.3.1 Influence of frequency
Ai and PO∗ in terms of hypervolume [46]. The hypervol-
ume difference has to be as close as possible to zero to Gw-ACO uses the Gw method to guide ants during solu-
prove the efficiency of the algorithm. In order to have a tion construction step in different directions of the objective

13
262 Evolutionary Intelligence (2019) 12:253–272

Fig. 3  Illustration of the median


attainment surfaces obtained by
Gw-ACO with FQ = 800, 400
and 200 for 2_250 instance

Fig. 4  Illustration of the median


attainment surfaces obtained by
Gw-ACO with FQ = 800, 400
and 200 for 2_500 instance

space. This proposed method is parameterized by the fre- with FQ = 200, 400 and 800 can be visually detected. As
quency FQ which allows to adjust the weight variations. In expected, the surfaces provided by Gw-ACO with FQ = 800
order to investigate the effect of this parameter on the per- are more interesting than those provided with FQ = 200 and
formance of the proposed approach Gw-ACO we have set, FQ = 400 on all the tested instances. As it can be seen from
FQ to 200, 400 and 800 while keeping the same number of these figures, with FQ = 200, Gw-ACO does not converge
constructed solutions of Gw-ACO given in Table 2. In fact, sufficiently and it did not give solutions at the extreme ends
the purpose of these experiments is to study the impact of of the Pareto front. With FQ = 400, the surfaces are better
FQ not only on the convergence towards the optimal Pareto than those provided by Gw-ACO with FQ = 200, Gw-ACO
front, but also on the diversification of the generated solu- gives a better distribution, but still not cover all the extreme
tions. To this end and to clearly visualize the difference ends of the Pareto front especially with the 2_500 and 2_750
between the resulting Pareto fronts, only the bi-objective instances (see Figs. 4, 5). It is also clear that Gw-ACO with
instances are tested. FQ = 800 outperforms the other variants where solutions
Analysis of the median attainment surface Figures 3, 4 obtained by Gw-ACO with FQ = 800 dominate an impor-
and 5 show the distribution of the median attainment sur- tant part of those found by Gw-ACO with FQ = 200 and
faces of the approximation sets for the 3 values of FQ for the FQ = 400. These observations confirm our suggestion in
bi-objective instances 2_250, 2_500 and 2_750 respectively. Sect. 3.2.1. The weight vector should change slowly from
From Fig. 3, the difference between the 3 variants: Gw-ACO one point to another. Indeed, the degradation in the quality

13
Evolutionary Intelligence (2019) 12:253–272 263

Fig. 5  Illustration of the median


attainment surfaces obtained by
Gw-ACO with FQ = 800, 400
and 200 for 2_750 instance

Table 3  Average hypervolume Instance FQ = 800 FQ = 400 FQ = 200


difference, results of Mann–
Whitney statistical test and Avg CPU Avg M–W.T CPU Avg M–W.T CPU
average CPU time (in s) of FQ
= 800, FQ = 400 and FQ = 200 2_250 2.23E−01 13.2 3.89E−01 > 17.3 3.93E−01 > 18.4
on bi-objective instances 2_500 2.45E−01 56.6 4.31E−01 > 74.7 4.32E−01 > 83.3
2_750 2.47E−01 126.1 4.96E−01 > 164.3 4.37E−01 > 179.7

of the results found by Gw-ACO with FQ = 200 and with that the two compared variants are not statistically different
FQ = 400 is due to the fact that ants jump from one direction (i.e., statistically insignificant).
to another which should be avoided. In fact, directions have From this table, one can see that the hypervolume values
to change smoothly in order to target almost all points on returned by the variant FQ = 800 are better than the hyper-
the Pareto front. Another important observation that can be volume values returned by the two other variants FQ = 400
extracted from Figs. 3, 4 and 5 is that, although we have kept and FQ = 200. In fact, these numerical values confirm the
the same maximum number of generations for all variants, results obtained with the median attainment surface. Moreo-
Gw-ACO with FQ = 800 achieves the best performances. ver, Table 3 shows that the variant FQ = 800 statistically
That is to say that the adjustment and the variation of the outperforms the two other variants on all the bi-objective
search directions in such weight-based approach is very sig- instances. From the same table and by analyzing the CPU
nificant and improve strongly the performance of the algo- time, it is clear that the CPU time consumed by the variant
rithm probably better than the exploration process itself. In FQ = 200 is more important than this consumed by the vari-
the case where the maximum number of generations is set ant FQ = 400 and the CPU time of the variant FQ = 400 is
to 400, with FQ = 800 the whole Pareto front is explored 2 higher than this of the variant FQ = 800 which is expected
times, with FQ = 400 it is explored 4 times and with FQ = since the exploration process over the Pareto front is time-
200 the Pareto front is explored 8 times. consuming step.
Analyses of the average hypervolume difference, the
Mann–Whitney statistical test and CPU time Table 3 gives 4.3.2 Gw‑ACO vs Rw‑ACO
the average hypervolume difference values (Avg) of FQ =
800, FQ = 400 and FQ = 200, the results of the Mann–Whit- In Sect. 3.2, we have presented the Gw method as well as the
ney statistical test (M–W.T) returned by FQ = 800 with FQ random weight generation method Rw. In this section, we
= 400 and FQ = 200 and the average execution time (CPU) provide an experimental study in order to evaluate the two
consumed by the three variants on the bi-objective instances. weight vector generation methods.
The notation “>” indicates that the performance of FQ = Analysis of the median attainment surface Figures 6, 7
800 is significantly better than the compared variant, the and 8 plot the median attainment surface of the approxima-
notation “<” indicates that the results of FQ = 800 are sig- tion sets (non-dominated points) from the 30 runs of Gw-
nificantly worse than the compared variant and “≈” implies ACO and Rw-ACO for the bi-objective test instances 2_250,

13
264 Evolutionary Intelligence (2019) 12:253–272

Fig. 6  The median attainment


surfaces obtained by Gw-
ACO and Rw-ACO for 2_250
instance

Fig. 7  The median attainment


surfaces obtained by Gw-
ACO and Rw-ACO for 2_500
instance

Fig. 8  The median attainment


surfaces obtained by Gw-
ACO and Rw-ACO for 2_750
instance

13
Evolutionary Intelligence (2019) 12:253–272 265

Table 4  Average hypervolume difference, results of Mann–Whitney front which implies that several directions might be never
statistical test and average CPU time (in s) of Gw-ACO and Rw-ACO reached contrary to Gw method that tries to reach almost
on all instances
all directions even those on the extremities: 𝜆 = (0, … , 1)
Instance Gw-ACO Rw-ACO and 𝜆 = (1, … , 0); (2) even if in Rw the weight vectors are
Avg CPU Avg M–W.T CPU uniformly distributed that cannot guarantee the equispaced
or/and the uniformity of the optimal solutions on the Pareto
2_250 2.23E−01 13.2 4.20E−01 > 14.2 front [51, 52]; (3) in Rw-ACO, each direction is explored by
2_500 2.45E−01 56.6 5.03E−01 > 56.2 only one ant which could place in question the convergence
2_750 2.47E−01 126.1 5.48E−01 > 127.0 of the algorithm towards the Pareto-optimal set since, on the
3_250 3.86E−01 34.1 4.51E−01 > 34.3 one hand, weak search efforts are allocated to explore one
3_500 3.34E−01 135.9 5.12E−01 > 136.9 direction of the Pareto front, on the other hand, the solu-
3_750 3.02E−01 310.3 5.71E−01 > 310.5 tion found by ant could not be always non-dominated. As
4_250 3.33E−01 51.0 5.11E−01 > 51.9 a matter of fact, all these factors have contributed to the
4_500 3.36E−01 206.0 4.99E−01 > 206.2 performance degradation of Rw-ACO.
4_750 3.32E−01 471.8 5.60E−01 > 473.1
4.4 Comparisons with state‑of‑the‑art approaches

2_500 and 2_750 respectively. From these figures, it is clear This part of experiments is designed to evaluate the effec-
that the points obtained by Rw-ACO are dominated by those tiveness of Gw-ACO against several state-of-the-art mul-
obtained by Gw-ACO on instances 2_250, 2_500 and 2_750 tiobjective approaches. At first, Gw-ACO is compared to
and there are no solutions returned by Rw-ACO that domi- MOACO approaches. Performances of Gw-ACO are evalu-
nate any one returned by Gw-ACO. Moreover, solutions ated with other multiobjective metaheuristics afterwards.
achieved by Gw-ACO are well-distributed on the trade-off
surface and covering all the Pareto front on all the bi-objec- 4.4.1 Comparison with MOACO approaches
tive instances. Contrary to Gw-ACO, solutions returned by
Rw-ACO are not covering all the Pareto front, they covered We compare in this section the suggested Gw-ACO approach
only the middle of the Pareto front. with MOACO approaches from the literature: Bicriterion-
Analyses of the average hypervolume difference, the Ant, IBACO, m-ACO4 and MOEA/D-ACO.
Mann–Whitney statistical test and CPU time Table 4 shows Analysis of the median attainment surface Figures 9, 10
the average hypervolume difference values returned by Gw- and 11 show the median attainment surface of the approxi-
ACO and Rw-ACO, the results of the Mann–Whitney statis- mation sets returned by Gw-ACO, BicriterionAnt, IBACO,
tical test of Gw-ACO with Rw-ACO and the average execu- m-ACO4 and MOEA/D-ACO for the bi-objective test
tion time consumed by the two algorithms on all instances. instances 2_250, 2_500 and 2_750, respectively. As previ-
By analyzing the table, we can observe that Gw-ACO ously mentioned, Gw-ACO produces a very well-distributed
achieves the best results. In fact, it strongly outperforms Rw- Pareto front where the found solutions covered all the Pareto
ACO in all the tested instances even if both of Gw-ACO front. It is evident from these figures that the solutions found
and Rw-ACO construct the same number of solutions (see by Gw-ACO dominate almost all the solutions found by Bic-
Table 2). The hypervolume difference values returned by riterionAnt and there is no solutions returned by Bicriterion-
Gw-ACO are always superior than the results returned by Ant that dominate any one returned by Gw-ACO and this is
Rw-ACO. Moreover, the results obtained by Gw-ACO are for all the bi-objective tested instances.
always significantly better than the results obtained by Rw- In fact, the solutions attained by BicriterionAnt are on
ACO and those of Rw-ACO never statistically outperform a small region of the Pareto front nearly on the extremity,
the results returned by Gw-ACO. It seems that the results especially with the 2_500 and 2_750 instances (see Figs. 10,
of Rw-ACO decrease according to the size of the problem. 11). They do not cover all the middle of the Pareto front,
For the largest instances, the quality of the results returned which is probably caused by the fact that BicriterionAnt
by Rw-ACO is clearly worst. Thus, the difference between optimizes one criterion much more than the other one, by
the two algorithms became much more important when the promising research and by guiding ants in search areas more
instance is larger and harder. We note that both of the algo- than others. From Fig. 9, the surfaces are confused therefore
rithms take almost the same CPU time. it is difficult to visually distinguish the algorithms. How-
The poor results obtained by the random weight genera- ever, we could see that Gw-ACO and MOEA/D-ACO are
tion method are caused by (1) the random change of the more distributed. On the instances with 500 and 750 items
weight vector during optimization makes the search process represented in Figs. 10 and 11 respectively, the surfaces pro-
less efficient because ants jump randomly over the Pareto vided by IBACO and m-ACO4 are slightly above the surface

13
266 Evolutionary Intelligence (2019) 12:253–272

Fig. 9  The median attainment


surfaces obtained by Gw-
ACO, BicriterionAnt, IBACO,
m-ACO4 and MOEA/D-ACO
for 2_250 instance

Fig. 10  The median attainment


surfaces obtained by Gw-
ACO, BicriterionAnt, IBACO,
m-ACO4 and MOEA/D-ACO
for 2_500 instance

Fig. 11  The median attainment


surfaces obtained by Gw-
ACO, BicriterionAnt, IBACO,
m-ACO4 and MOEA/D-ACO
for 2_750 instance

13
Evolutionary Intelligence (2019) 12:253–272 267

provided by Gw-ACO. This is could be explained by the fact example, on the instance 3_500, the hypervolume difference
that, both IBACO and m-ACO4 algorithms favor intensifi- values in Gw-ACO are about 49.7%, 55.9% and 84.5% of
cation. Ants focus research in only one region of the search those obtained by IBACO, m-ACO4 and MOEA/D-ACO
space during the evolution process. In fact, ants improve respectively. On the largest instance with 4 objectives and
one search area: they favor the reproduction of the best 750 items, the Gw-ACO values are about 41.7% of IBACO,
components without exploration of new search areas. The 45.1% of m-ACO4 and 84.4% of MOEA/D-ACO. By analyz-
obtained solutions by the two algorithms are laying only on ing the table according to the statistical test, we can confirm
the central part of the Pareto front where they do not cover that the difference between MOEA/D-ACO and Gw-ACO is
all the Pareto front, thereby they are not well-distrusted on not statistically significant on the bi-objective instances and
the surface. The surfaces returned by MOEA/D-ACO and 3_250 instance. Moreover, Gw-ACO statistically outperforms
Gw-ACO in Fig. 9 are very close. On the two other Figs. 10 MOEA/D-ACO on the 4 largest instances and it is never sta-
and 11, the surfaces of MOEA/D-ACO are slightly above tistically worse than any compared algorithm.
those found by Gw-ACO. But, these difference between the Analysis of the average CPU time Table 6 presents the
surfaces provided by MOEA/D-ACO and Gw-ACO is not average execution time consumed by every algorithm on all
statistically significant. So, for the bi-objective instances, instances. The table shows that BicriterionAnt, IBACO and
the results of the two algorithms are close. However, for m-ACO4 consume more CPU time than Gw-ACO. On the other
the hardest instances with 3 and 4 objectives, the surfaces hand, MOEA/D-ACO consumes less CPU time than Gw-ACO.
of Gw-ACO are better than those of MOEA/D-ACO since
the numerical hypervolume difference results significantly 4.4.2 Comparison with other state‑of‑the‑art
outperform MOEA/D-ACO results according to Table 5. multiobjective metaheuristics
Analyses of the average hypervolume difference and the
Mann–Whitney statistical test Table 5 presents the average To have an overview of the efficiency of the proposed
hypervolume difference values and results of the Mann–Whit- approach, in this section, we compare the proposed Gw-ACO
ney statistical test obtained by Gw-ACO, BicriterionAnt,
IBACO, m-ACO4 and MOEA/D-ACO on all instances. The
first important result which can be extracted from this set of Table 6  Average CPU time (in s) of Gw-ACO, BicriterionAnt,
experiments is that Gw-ACO performs much better than Bic- IBACO, m-ACO4 and MOEA/D-ACO on all instances
riterionAnt, IBACO and m-ACO4 on all the tested instances. Instance Gw-ACO Bicriterion- IBACO m-ACO4 MOEA/D-
For example, the hypervolume difference values in Gw-ACO Ant ACO
are about 33.4%, 50.7% and 54.9% of those obtained by Bic-
2_250 13.2 35.6 19.6 408.1 5.1
riterionAnt, IBACO and m-ACO4 respectively on instance
2_500 56.6 152.9 76.8 1450.2 15.2
2_250. Moreover, Gw-ACO statistically outperforms Bicrite-
2_750 126.1 327.7 175.1 3114.3 24.7
rionAnt, IBACO and m-ACO4 for all instances. By compar-
3_250 34.1 – 54.8 441.5 8.7
ing Gw-ACO and MOEA/D-ACO, we can observe that the
3_500 135.9 – 185.5 1511.0 20.7
average hypervolume difference values obtained by MOEA/
3_750 310.3 – 410.8 3695.8 37.9
D-ACO are better than those obtained by Gw-ACO on the
4_250 51.0 – 100.2 1300.6 11.8
smallest instances with 2 objectives and on instance 3_250.
4_500 206.0 – 293.6 1941.9 29.9
However, on the other instances with 3 and 4 objectives the
4_750 471.8 – 636.4 4389.0 50.5
best results are achieved by our proposed approach. As an

Table 5  Average hypervolume Instance Gw-ACO BicriterionAnt IBACO m-ACO4 MOEA/D-ACO


difference and results of Mann–
Whitney statistical test of Avg Avg M–W.T Avg M–W.T Avg M–W.T Avg M–W.T
Gw-ACO with BicriterionAnt,
IBACO, m-ACO4 and MOEA/ 2_250 2.23E-01 6.66E−01 > 4.39E−01 > 4.06E−01 > 1.21E−01 ≈
D-ACO on all instances 2_500 2.45E−01 7.42E−01 > 4.83E−01 > 4.45E−01 > 1.87E−01 ≈
2_750 2.47E−01 8.44E−01 > 5.70E−01 > 5.44E−01 > 1.76E−01 ≈
3_250 3.86E−01 – – 6.05E−01 > 5.14E−01 > 3.18E−01 ≈
3_500 3.34E−01 – – 6.71E−01 > 5.97E−01 > 3.95E−01 >
3_750 3.02E−01 – – 7.25E−01 > 6.13E−01 > 3.08E−01 ≈
4_250 3.33E−01 – – 7.06E−01 > 6.72E−01 > 4.18E−01 >
4_500 3.36E−01 – – 7.85E−01 > 7.16E−01 > 4.06E−01 >
4_750 3.32E−01 – – 7.96E−01 > 7.35E−01 > 3.99E−01 >

13
268 Evolutionary Intelligence (2019) 12:253–272

Fig. 12  The median attainment


surfaces obtained by Gw-ACO,
2PPLS and MOEA/D for 2_250
instance

Fig. 13  The median attainment


surfaces obtained by Gw-ACO,
2PPLS and MOEA/D for 2_500
instance

with two state-of-the-art multiobjective metaheuristics: CPU time of Gw-ACO, MOEA/D and 2PPLS and the results
MOEA/D and 2PPLS. of Mann–Whitney statistical test of Gw-ACO with MOEA/D
Analysis of the median attainment surface Figures 12, and 2PPLS on all instances. The table shows the efficiency of
13 and 14 plot the median attainment surfaces obtained by Gw-ACO algorithm, which obtains a good results, in terms
Gw-ACO, 2PPLS and MOEA/D on instances 2_250, 2_500 of hypervolume, on all instances against MOEA/D. The
and 2_750, respectively. Figure 12 shows that the surfaces table also shows that the hypervolume difference values of
obtained by the three algorithms are similar. They are well- 2PPLS are better than those of Gw-ACO on the bi-objective
distributed along the Pareto front and are very close. In instances. By analyzing the table according to the statistical
Figs. 13 and 14, one can see that the surfaces of Gw-ACO results, Table 7 confirms the obtained hypervolume values
and MOEA/D are slightly below than those obtained by where Gw-ACO statistically outperforms MOEA/D on all
2PPLS. By comparing the approximation sets of Gw-ACO instances. It is clear that the efficiency of MOEA/D seems to
and MOEA/D, we can observe that the solutions returned increase according to the size of the problem. We can notice,
by Gw-ACO are better than those obtained by MOEA/D on from the same table, that 2PPLS statistically outperforms Gw-
several parts of the Pareto front. ACO on the instances 2_250 and 2_750 but the difference
Analyses of the average hypervolume difference, the between the two algorithms is not statistically significant on
Mann–Whitney statistical test and the average CPU time the 2_500 instance. The table suggests also that MOEA/D and
Table 7 gives the average hypervolume difference, the average 2PPLS consume less CPU time than Gw-ACO.

13
Evolutionary Intelligence (2019) 12:253–272 269

Fig. 14  The median attainment


surfaces obtained by Gw-ACO,
2PPLS and MOEA/D for 2_750
instance

Table 7  Average hypervolume Instance Gw-ACO MOEA/D 2PPLS


difference, results of Mann-
Whitney statistical test and Avg CPU Avg M–W.T CPU Avg M–W.T CPU
average CPU time (in s) of
Gw-ACO, MOEA/D and 2PPLS 2_250 2.23E−01 13.2 3.14E−01 > 4.8 1.00E−01 < 3.1
on all instances 2_500 2.45E−01 56.6 3.26E−01 > 14.5 1.17E−01 ≈ 14.8
2_750 2.47E−01 126.1 3.63E−01 > 28.6 1.06E−01 < 25.1
3_250 3.86E−01 34.1 4.51E−01 > 9.2 – – –
3_500 3.34E−01 135.9 4.30E−01 > 24.5 – – –
3_750 3.02E−01 310.3 4.14E−01 > 48.8 – – –
4_250 3.33E−01 51.0 4.74E−01 > 22.1 – – –
4_500 3.36E−01 206.0 4.36E−01 > 64.6 – – –
4_750 3.32E−01 471.8 4.59E−01 > 130.8 – – –

4.5 Additional experimental comparison a given CPU time for all instances. From the table, we can
say that MOEA/D-ACO, MOEA/D and 2PPLS could obtain
In the previous experiments we have found that the sug- a better hypervolume difference values if a more computing
gested Gw-ACO approach statistically outperforms MOEA/ time is accorded. However, these approaches fail to obtain
D-ACO and MOEA/D on almost all instances and that Gw- better results than those found by Gw-ACO on the largest
ACO performs the same as 2PPLS on the 2_500 instance. and hardest instances. Gw-ACO still obtains the best results
By comparing the CPU time consumed by our approach, on these instances. From the same table and by analyzing
we have found that MOEA/D-ACO, MOEA/D and 2PPLS the statistical results, we notice that the difference between
are faster than Gw-ACO. We recall that we have set the the algorithms is statistically insignificant on almost all
parameters as recommended by their authors for the same instances except on 2_250 and 3_750 instances, where
problem and for the same instances. In this section, we car- 2PPLS statistically outperforms Gw-ACO and on 3_250
ried out another experiment tests where MOEA/D-ACO, instance, where MOEA/D-ACO and MOEA/D statistically
MOEA/D and 2PPLS are given the same execution time as outperform Gw-ACO. However, when considering the hard-
for Gw-ACO to verify if these approaches could outperform est instance 4_750, Gw-ACO success to statistically outper-
Gw-ACO if more CPU time will be accorded. Note that the forms MOEA/D-ACO and MOEA/D.
median attainment surface is not a part of these experiments
since the compared algorithms provide almost the same sur- 4.6 Analysis and discussion
faces as the previous experiments.
Table 8 shows the average hypervolume difference of Gw- The comparison of Gw-ACO and Rw-ACO shows that the
ACO, MOEA/D-ACO, MOEA/D and 2PPLS and the results results provided by Gw-ACO are more efficient than those
of the statistical test of Gw-ACO with these approaches for provided by Rw-ACO whatever the considered performance

13
270 Evolutionary Intelligence (2019) 12:253–272

Table 8  Average hypervolume Instance CPU Gw-ACO MOEA/D-ACO MOEA/D 2PPLS


difference and results of
Mann-Whitney statistical test Avg Avg M–W.T Avg M–W.T Avg M–W.T
for a given CPU time (in s) of
Gw-ACO, MOEA/D-ACO, 2_250 15 2.20E−01 1.01E−01 ≈ 2.92E−01 > 9.85E−02 <
MOEA/D and 2PPLS on all 2_500 60 2.40E−01 1.37E−01 ≈ 1.73E−01 ≈ 9.36E−02 ≈
instances 2_750 130 2.45E−01 1.54E−01 ≈ 2.51E−01 ≈ 9.00E−02 <
3_250 35 3.86E−01 1.07E−01 < 1.88E−01 < – –
3_500 135 3.34E−01 3.45E−01 ≈ 3.65E−01 ≈ – –
3_750 300 3.00E−01 2.01E−01 ≈ 2.51E−01 ≈ – –
4_250 50 3.33E−01 3.15E−01 ≈ 3.08E−01 ≈ – –
4_500 200 3.37E−01 3.48E−01 ≈ 3.67E−01 ≈ – –
4_750 500 3.20E−01 3.69E−01 > 3.87E−01 > – –

metric. Moreover, Gw-ACO shows its efficiency with small directions towards different parts of the Pareto front. Dur-
and large instances. By comparing the MOACO approaches ing the optimization process, at each generation of the algo-
with our proposed approach, we have found that Gw-ACO rithm, all ants construct their solutions according to one
obtains the best results on almost all the tested instances. direction generated by Gw. Gw-ACO has been compared
The attainment surfaces have confirmed the numerical to different powerful and reference algorithms of the state
results where Gw-ACO achieves a good approximation set. of the art on the well-known multiobjective multidimen-
Furthermore, the larger the number of the items is, the well- sional knapsack problem knapsack problem. Based on this
distributed the Pareto front is. The experiment tests carried study, some conclusions can be drawn. First, the proposed
out with MOEA/D-ACO, MOEA/D and 2PPLS have shown approach proves its ability to find an approximation set well
that Gw-ACO performs slightly less than or about the same spread all over the Pareto frontier. The results show the
as these algorithms in terms of intensification. high efficiency of Gw-ACO in terms of diversity. Second,
In summary, even if Gw-ACO constructs less solutions thanks to the exploration strategy adopted by ants, a sig-
and consumes less running time than most of the compared nificant research effort has been dedicated to each region of
algorithms, it gives better results for almost the nine bench- the objective space. The set of computational experiments
mark instances and the statistical test showed that this differ- demonstrates that the proposed approach can achieve a good
ence is significant. In fact, the performance of the proposed convergence and outperforms different algorithms from the
algorithm Gw-ACO is reached thanks to its ability to deter- literature on almost all the benchmark instances. Another
mine a good balance between intensification and diversifica- advantage of Gw-ACO is its high level of generality, mainly
tion. Intensification is achieved by the fact that ants carried due to the proposed Gw method which require a single
out an intensive search in every search region and by favor- parameter FQ to set. In fact, this proposed method could be
ing the rewarding of the best solutions, which allow them to easily adapted to a wide range of multiobjective optimiza-
converge quickly towards the locally optimal solutions. To tion problems. This work opens many future research works.
counterbalance the intensification, Gw-ACO achieves diver- In our experiments, we have noticed that Gw-ACO whose
sification by changing the search direction during optimiza- run time basically depends on the size of the weight vectors
tion, ants explore these new regions due to the proposed Gw set is more time consuming than MOEA/D-ACO, MOEA/D
method that enables to reach almost all parts of the Pareto and 2PPLS especially on the large instances. Indeed, we can
front. The experimental studies have confirmed that the sug- integrate an intensification mechanism, such as local search
gested Gw-ACO approach for the multiobjective multidi- techniques, instead of exploring the Pareto front several
mensional knapsack problem performs a good convergence, times. This will allow us to both, decrease the computational
gives good sampling of points along the Pareto front and time and improve the quality of the generated solutions par-
finds points at the extreme ends of the Pareto front. ticularly in terms of exploitation. An interesting direction of
future research would be to investigate Gw-ACO on other
multiobjective problems with more objectives and higher
5 Conclusion and future works complexity. Another perspective of this work would be to
develop a self-adaptive version of Gw-ACO. Here the goal
In this paper, we have proposed a Gradual weight-based is to improve the performance of Gw-ACO. In particular,
ACO approach for solving the MOMKP, called Gw-ACO. it should be interesting to adopt a self-adaptive pheromone
This approach uses a proposed gradual weight generation evaporation mechanism to get a better exploitation of the
method (Gw), in order to guide ants in multiple search search space and speed up the convergence. On the other

13
Evolutionary Intelligence (2019) 12:253–272 271

hand, the purpose of this self-adaptive version is that Gw- 17. Chica M, Cordón Ó, Damas S, Bautista J (2015) Interactive
ACO would be able to effectively address a new optimiza- preferences in multiobjective ant colony optimisation for assem-
bly line balancing. Soft Comput 19:2891–2903
tion problem with minimal preliminary input. 18. Zouari W, Alaya I, Tagina M (2017) A hybrid ant colony algo-
rithm with a local search for the strongly correlated knapsack
problem. In: Computer systems and applications (AICCSA),
2017 IEEE/ACS 14th international conference on IEEE, pp
527–533
References 19. Roeva O, Fidanova S, Paprzycki M (2018) Comparison of differ-
ent ACO start strategies based on intercriteria analysis. Recent
1. AbdelBasset M, ElShahat D, ElHenawy I, Sangaiah AK (2018) advances in computational optimization. Springer, Cham, pp
A modified flower pollination algorithm for the multidimensional 53–72
knapsack problem: human-centric decision making. Soft Comput 20. Shelokar P (2018) Ant colony system: application for the single
22(13):4221–4239 and multi-objective multidimensional knapsack problems, 2017.
2. AbdelBasset M, ElShahat D, Mirjalili S (2018) A hybrid whale Math Res Summ 2:27
optimization algorithm based on local search strategy for the per- 21. Alaya I, Solnon C, Ghédira K (2007) Ant colony optimiza-
mutation flow shop scheduling problem. Future Gener Comput tion for multi-objective optimization problems. In: 19th IEEE
Syst 85:129–145 international conference on tools with artificial intelligence
3. AbdelBasset M, Manogaran G, AbdelFatah L, Mirjalili S (2018) (ICTAI’07), pp 450–457
An improved nature inspired meta-heuristic algorithm for 1-D bin 22. Ben Mansour I, Alaya I (2015) Indicator based ant colony
packing problems. Person Ubiq Comput 2018:1–16 optimization for multi-objective Knapsack problem. Procedria
4. AbdelBasset M, ElShahat D, ElHenawy I (2018) Solving 0–1 Comput Sci KES Singapore 60:448–457
knapsack problem by binary flower pollination algorithm. Neural 23. Ulungu EL, Teghem J, Fortemps Tuyttens D (1999) MOSA
Comput Appl 2018:1–19 method: a tool for solving multiobjective combinatorial opti-
5. AbdelBasset M, ElShahat D, Sangaiah AK (2017) A modified mization problems. J Multi-Criteria Decis Anal 8(4):221–236
nature inspired meta-heuristic whale optimization algorithm 24. Paquete L, Stützle T (2006) A study of stochastic local search
for solving 0–1 knapsack problem. Int J Mach Learn Cybern algorithms for the biobjective QAP with correlated flow matri-
2017:1–20 ces. Eur J Oper Res 169(3):943–959
6. Martello S, Toth P (1990) Knapsack problems: algorithms and 25. Zhang Q, Li H (2007) MOEA/D: a multiobjective evolutionary
computer implementations. Wiley, New York algorithm based on decomposition. IEEE Trans Evol Comput
7. Shih H (2005) Fuzzy approach to multilevel knapsack problems. 6(11):712–731
Comput Math Appl 49(7–8):1157–1176 26. Lust T, Teghem J (2012) The multiobjective multidimensional
8. Smeraldi F, Malacaria P (2014) How to spend it: optimal invest- knapsack problem: a survey and a new approach. Int Trans Oper
ment for cyber security. In: Proceedings of the 1st international Res 19(4):495–520
workshop on agents and CyberSecurity, ACySe 27. Angel E, Bampis E, Gourvs L (2004) A dynasearch neighbor-
9. Ehrgott M, Ryan DM (2002) Constructing robust crew sched- hood for the bicriteria traveling salesman problem. In: Gandi-
ules with bicriteria optimization. J Multi-Criteria Decis Anal bleux X, Sevaux M, Srensen K, Tkindt V (eds) Metaheuristics
11(3):139–150 for multiobjective optimisation. Lecture notes in economics and
10. Zitzler E, Laumanns M, Thiele L (2001) SPEA2: improving the mathematical systems, vol 535. Springer, Berlin
strength pareto evolutionary algorithm for multiobjective optimi- 28. Paquete L, Chiarandini M, Sttzle T (2004) Pareto local opti-
zation. In Giannakoglou K (eds) Evolutionary methods for design, mum sets in the biobjective traveling salesman problem: an
optimisation and control with application to industrial problems experimental study. In: Gandibleux X, Sevaux M, Srensen K,
(EUROGEN 2001). International Center for Numerical Methods Tkindt V (eds) Metaheuristics for multiobjective optimisation.
in Engineering (CIMNE), pp 95–100 Lecture notes in economics and mathematical systems, vol 535.
11. Deb K, Pratap A, Agarwal S, Meyarivan T (2002) A fast and elit- Springer, Berlin
ist multi-objective genetic algorithm: NSGA-II. IEEE Trans Evol 29. Ahuja RK, Ergun Ö, Orlin JB, Punnen AP (2002) A survey of
Comput 6(2):181–197 very large-scale neighborhood search techniques. Discret Appl
12. Lust T, Teghem J (2008) Memots: a memetic algorithm integrat- Math 123:75–102
ing tabu search for combinatorial multiobjective optimization. 30. Iredi S, Merkle D, Middendorf M (2001) Bi-criterion opti-
RAIRO Oper Res 42(1):3–33 mization with multi colony ant algorithms. In: First interna-
13. Alsheddy A, Tsang EPK (2009) Guided pareto local search and tional conference on evolutionary multi-criterion optimization
its application to the 0/1 multi-objective knapsack problems. In: (EMO01). Lecture notes in computer science, pp 359–372
Proceedings of the eighth metaheuristic international conference 31. Angus D (2007) Crowding population-based ant colony optimi-
(MIC09). Hamburg sation for the multi-objective travelling salesman problem. In:
14. Barán B, Schaerer M (2003) A multiobjective ant colony system IEEE symposium on computational intelligence in multi-criteria
for vehicle routing problem with time windows. In: Proceedings decision-making (MCDM), pp 333–340
of twenty first IASTED international conference on applied infor- 32. Liangjun K, Qingfu Z, Battiti R (2013) MOEA/D-ACO: a mul-
matics, Insbruck, Austria, pp 97–102 tiobjective evolutionary algorithm using decomposition and
15. Ippolito MG, Morana G, Sanseverino ER, Vuinovich F (2005) AntColony. IEEE Trans Cybern 43(6):1845–59
Ant colony search algorithm for optimal strategical plan- 33. Dorigo M, Maniezzo V, Colorni A (1991) Positive feedback
ning of electrical distribution systems expansion. Appl Intell as a search strategy. Technical report 91–016. Dipartimento di
23:139–152 Elettronica, Politecnico di Milano, Milan, Italy
16. Angus D, Woodward C (2009) Multiple objective ant colony 34. Stützle T, Hoos H (1998) MAX-MIN ant system and local
optimisation. Swarm Intell 3:69–85 search for combinatorial optimization problems, meta-heuris-
tics: advances and trends in local search paradigms for optimi-
zation. Kluwer Academics, Boston

13
272 Evolutionary Intelligence (2019) 12:253–272

35. Dorigo M (1992) Optimization learning, natural algorithms (in 45. Garcia-Martinez C, Cordon O, Herrera F (2007) A taxonomy
Italian). PhD thesis, Dipartimento di Elettronica, Politecnico di and an empirical analysis of multiple objective ant colony
Milano, Italy optimization algorithms for bi-criteria tsp. Eur J Oper Res
36. Mariano C E, Morales E (1999) A multiple objective ant-q algo- 180(1):116–148
rithm for the design of water distribution irrigation networks. 46. Zitzler E, Thiele L (1999) Multiobjective evolutionary algo-
Technical report n HC-9904. Instituto Mexicano de Tecnologa rithms: a comparative case study and the strength pareto
del Agua, Mexico approach. IEEE Trans Evol Comput 3(4):257–271
37. Gravel M, Price WL, Gagne C (2002) Scheduling continuous 47. Grunert da Fonseca V, Fonseca C M, Hall A O (2001) Infer-
casting of aluminium using a multiple objective ant colony opti- ential performance assessment of stochastic optimisers and the
mization metaheuristic. Eur J Oper Res 143(1):218–229 attainment function. In: 1st international conference on evolu-
38. McMullen PR (2001) An ant colony optimization approach to tionary multi-criterion optimization (EMO (2001). Lecture note
addressing a JIT sequencing problem with multiple objectives. in computer science. Springer, Berlin, pp 213–225
Artif Intell Eng 15(3):309–317 48. Knowles J D, Thiele L, Zitzler E (2005) A tutorial on the per-
39. López-Ibáñez M, Stützle T (2010) The impact of design choices formance assessment of stochastive multiobjective optimizers.
of multi-objective ant colony optimization algorithms on per- Technical report TIK-Report No. 214. Computer Engineering
formance: an experimental study on the biobjective TSP. In: and Networks Laboratory, ETH Zurich
GECCO 2010. ACM press, New York, pp 71–78 49. Zitzler E, Brockhoff D, Thiele L (2007) The hypervolume indi-
40. Dörner K, Gutjahr WJ, Hartl RF, Strauss C, Stummer C cator revisited: on the design of pareto-compliant indicators via
(2004) Pareto ant colony optimization: a metaheuristic weighted integration. In: International conference on evolution-
approach to multiobjective portfolio selection. Ann Oper Res ary multi-criterion optimization. Springer, Berlin, pp 862–876
131(1–4):79–99 50. Emmerich M, Beume N, Naujoks B (2005) An EMO algorithm
41. Gambardella L, Taillard ED, Agazzi G (1999) MACS-VRPTW: using the hypervolume measure as selection criterion. In: Inter-
a multiple ant colony system for vehicle routing problems with national conference on evolutionary multi-criterion optimiza-
time windows. In: Corne FGD, Dorigo M (eds) New ideas in tion. Springer, Berlin, pp 62–76
optimization. McGraw Hill, London, pp 63–76 51. Das I, Dennis JE (1997) A closer look at drawbacks of mini-
42. Bullnheimer B, Hartl RF, Strauss C (1999) An improved ant mizing weighted sums of objectives for Pareto set generation in
system algorithm for the vehicule routing problem. Ann Oper multicriteria optimization problems. Struct Optim 14(1):63–69
Res 89:319–328 52. Vahidinasab V, Jadid S (2010) Normal boundary intersection
43. Doerner K, Hartl RF, Teiman M (2003) Are COMPETants more method for suppliers’ strategic bidding in electricity markets:
competent for problem solving? The case of full truckload trans- an environmental/economic approach. Energy Convers Manag
portation. Cent Eur J Oper Res 11(2):115–141 51(6):1111–1119
44. Alaya I, Solnon C, Ghédira K (2004) Ant algorithm for the
multi-dimensional knapsack problem. In: Proceedings of inter- Publisher’s Note Springer Nature remains neutral with regard to
national conference on bioinspired optimization methods and jurisdictional claims in published maps and institutional affiliations.
their applications (BIOMA), pp 63–72

13

You might also like