Expert Systems With Applications 45 (2016) 172–184

Contents lists available at ScienceDirect

Expert Systems With Applications

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

An efficient incremental evaluation function for optimizing truck

scheduling in a resource-constrained crossdock using metaheuristics
Mojtaba Shakeri a,∗, Malcolm Yoke Hean Low b, Stephen John Turner c, Eng Wah Lee d
a
Faculty of Computer and Information Technology Engineering, Qazvin Branch, Islamic Azad University, Qazvin 34185-1416, Iran
b
Singapore Institute of Technology, Singapore 13868, Singapore
c
School of Computer Engineering, Nanyang Technological University, Singapore 639798, Singapore
d
Singapore Institute of Manufacturing Technology, Singapore 638075, Singapore

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

Keywords: This paper addresses truck scheduling optimization in a resource-constrained crossdock. The truck schedul-
Optimization ing problem decides on the succession of incoming and outgoing trucks at the dock doors of a crossdocking
Runtime
terminal such that the total crossdocking operation time is minimized. The paper tackles the optimization
Incremental evaluation
from the computational perspective by developing an incremental evaluation of the objective function in
Metaheuristics
Crossdocking the body of single-solution based metaheuristics. It consists in evaluating only the transformation applied to
Truck scheduling the current solution rather than the complete evaluation of the neighbor solution. The proposed incremen-
tal neighborhood evaluation is integrated into two metaheuristics including tabu search (TS) and variable
neighborhood search (VNS). In terms of solution quality vs. runtime, experimental results show that the in-
cremental mechanism helps the two algorithms with dedicating their runtime to solution optimization rather
than spending it on fitness evaluation when compared with a deterministic local search (LS) algorithm that
exploits a simple complete evaluation of the objective function. This is in particular evident for the TS al-
gorithm which obtains comparable results to LS while achieving on average 67.6% reduction in runtime for
huge instances of scheduling 2048 trucks in a 256-door crossdock. Our findings on the efficiency of the pro-
posed incremental evaluation are reinforced when the two metaheuristics are re-assessed with a complete
evaluation of the objective function.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction There are a number of techniques to handle the computational

burden. One technique is to approximate the objective function and
Applications of optimization are countless. There are many chal- then replace the original objective function by its approximation
lenging applications in science and industry that can be formulated function, referred to as a surrogate objective. It is a function corre-
as optimization problems most of which are complex and difficult to lated to the true objective but is less computationally demanding, in
solve. They cannot be solved in an exact manner within a reasonable order to identify a (small) set of promising candidates (potential solu-
amount of time. Using metaheuristics is the main alternative to tions achieving the best values for the surrogate). The true objective
solve this class of problems (Talbi, 2009). An important element in is then computed for this small set of candidates and the best one
designing a metaheuristic is the objective function (also referred to is selected to become the next solution. This approach is known as
as evaluation or fitness function). It guides the search toward “good” meta-modeling. Many meta-modeling techniques may be employed
solutions of the search space. It is well-known that most of the time, for expensive objective functions (Barthelemy & Haftka, 1993; Buche,
in metaheuristics, the time-intensive part is the evaluation of the Schraudolph, & Koumoutsakos, 2005; Crainic, Gendreau, Soriano, &
objective function (Talbi, 2009). Some real-life optimization prob- Toulouse, 1993; Li, 2011). They are based on constructing an approxi-
lems are characterized by a huge computational cost of the objective mate model from a properly selected sample of solutions. The model
function(s). selection problem is however far from being a simple one (Burnham
& Anderson, 2002). There is a trade-off between the complexity of
the model and its accuracy. The reader may refer to Jin (2005) for a
∗
Corresponding author. Tel.: +98 28 33675782.
comprehensive survey.
E-mail addresses: [email protected], [email protected] (M. Shak-
eri), [email protected] (M.Y.H. Low), [email protected] (S.J.
Another technique is the incremental evaluation of the objective
Turner), [email protected] (E.W. Lee). function. By the incremental evaluation we mean evaluating only

http://dx.doi.org/10.1016/j.eswa.2015.09.041
0957-4174/© 2015 Elsevier Ltd. All rights reserved.
 M. Shakeri et al. / Expert Systems With Applications 45 (2016) 172–184 173

the changes to the problem environment rather than the full eval- (Alvarez-Perez, Gonzalez-Velarde, & Fowler, 2009; Bloori Arabani
uation of the whole space. There are a few studies which integrated et al., 2011; Liao et al., 2013; Miao, Cai, & Xu, 2014; Miao et al.,
an incremental mechanism in their proposed solution techniques to 2009; Vahdani & Zandieh, 2010; Van Belle, Valckenaers, Van-
their studied problems. Tang, Zhao, and Liu (2014) addressed a dy- den Berghe, & Cattrysse, 2013), variable neighborhood search (Kuo,
namic scheduling problem in steelmaking-continuous casting pro- 2013; Soltani & Sadjadi, 2010; Vahdani & Zandieh, 2010), greedy
duction. They proposed an improved differential evolution (DE) al- randomized adaptive search procedure (Alvarez-Perez et al., 2009;
gorithm by embedding an incremental mechanism to generate a new Ghobadian, Tavakkoli-Moghaddam, Javanshir, & Naderi-Beni, 2012),
initial population for the DE whenever a real-time event arises, based electromagnetism-like algorithm (Vahdani & Zandieh, 2010), par-
on the final population in the last DE solution process. Kang, Mok- ticle swarm optimization (Bloori Arabani et al., 2011), differen-
bel, Shekhar, Xia, and Zhang (2010) addressed the continuous re- tial evolution (Bloori Arabani et al., 2011; Liao et al., 2013), and
verse nearest neighbor (RNN) queries. They considered the contin- ant colony optimization (Bloori Arabani et al., 2011; Liao et al.,
uous movement of both the query and data objects in their prob- 2014; Liao et al., 2013), among others. Nonetheless, none of the
lem definition and proposed an incremental mechanism in their al- studies explicitly presented any mechanism to tackle their ad-
gorithms for identifying RNNs throughout the lifetime of a contin- dressed problems from the computational perspective. They rather
uous query. The main idea was to initially identify a single region r streamlined their implementation methodology of their proposed
around the query object and a set of objects S such that only r and S metaheuristics.
need to be monitored to trigger subsequent changes of the answer. In this paper we address the optimization of the truck schedul-
Wu, Yu, Jin, Lin, and Schiavone (2004) developed a genetic algorithm ing problem defined in Shakeri et al. (2012) by using TS and VNS
(GA) approach to the problem of task scheduling for multiprocessor metaheuristics. The two algorithms had successful implementations
systems. Because of the complexity of the solutions, they developed in the related work. Both TS and VNS have been applied to other
an incremental fitness function that changes over time. It initially re- variations of the truck scheduling problem and the results recorded
wards for finding short valid sequences of tasks assigned to proces- in terms of both solution quality and runtime have been among the
sors. Over time, the length of the sequences that can be rewarded best ones (see Liao et al., 2013; Vahdani and Zandieh, 2010). We fo-
is increased, encouraging the GA to find and maintain longer valid cus on the runtime of the two algorithms by developing an incre-
sequences. mental evaluation function that evaluates only the transformation
In this paper, we aim at developing an incremental evaluation applied to the current solution rather than the complete evaluation
function in the body of two single-solution based metaheuristics of the neighbor solution, allowing the runtime to be more efficiently
including tabu search (TS) and variable neighborhood search (VNS) dedicated to solution optimization rather than being spent on fit-
to evaluate only the transformation applied to the current solution ness evaluation. We compare the performance of the proposed in-
rather than the complete evaluation of the neighbor solution. This cremental mechanism against a deterministic local search (LS) al-
allows the runtime to be more efficiently dedicated to solution gorithm that uses a simple complete evaluation of the objective
optimization rather than being spent on fitness evaluation. The function.
optimization problem under study is truck scheduling in a cross- The rest of the paper is organized as follows. Section 2 describes
docking terminal. Crossdocking is a warehousing strategy that our optimization model by first introducing the truck scheduling
moves products through flow consolidation centers or crossdocks problem defined in Shakeri et al. (2012) and then the optimization
without putting them into storage. The practice not only brings framework required to develop the two metaheuristics. The incre-
substantial reduction in the distribution cost but also shortens mental evaluation function is proposed in Section 3. Section 4 devel-
the product delivery time (Apte & Viswanathan, 2000). The truck ops the TS and VNS metaheuristics equipped with the proposed in-
scheduling problem decides on the sequence of a number of trailers1 cremental evaluation function. Section 5 sets up our experimental en-
that are to exchange some of their products at the dock doors of a vironment by introducing the test data used for the assessment. The
crossdocking terminal such that the total crossdocking operation experimentation starts with tuning the control parameters of both
time is minimized. The problem was defined and formulated in metaheuristics with the aim of maintaining robustness in terms of
Shakeri, Low, Turner, and Lee (2012) according to a crossdocking solution quality vs. runtime. The two algorithms are deployed to im-
model that has practical applications in the fast moving consumer prove the solutions generated by the two-phase heuristic algorithm
goods (FMCG) industry and military logistics. The paper aimed at proposed in Shakeri et al. (2012). Their results in terms of solution
studying truck scheduling from a practical viewpoint by developing quality and runtime are then compared with those of the LS that
a crossdocking model that emulates the real world crossdocking exploits the complete evaluation of the objective function. Finally,
by considering pallets as the unit of shipments exchanged between Section 6 concludes the paper and highlights the future research
the trailers and imposing constraints on the number of doors and directions.
forklifts as the crossdock critical resources. In the same paper, a
two-phase greedy heuristic algorithm was developed to produce fea-
sible solutions for problem instances of various types and difficulty 2. Optimization model
levels.
There are numerous studies that aimed at optimizing truck Fig. 1 presents the crossdocking model developed in Shakeri et al.
scheduling defined in different crossdocking models using meta- (2012) for the truck scheduling problem under study. According to the
heuristics. Examples are simulated annealing (Bozer & Carlo, 2008; figure, a number of trailers are crossdocked in a multiple door cross-
Liao, Chang, Kuo, & Liao, 2014; Liao, Egbelu, & Chang, 2013; Madani- dock. Incoming trailers are assigned to dock doors on a daily basis
Isfahani, Tavakkoli-Moghaddam, & Naderi, 2014; Soltani & Sadjadi, to exchange some of their products before being dispatched to their
2010; Vahdani & Zandieh, 2010), genetic algorithms (Bloori Ara- customers. (The freight exchange pattern is known in advance.) Prod-
bani, Fatemi Ghomi, & Zandieh, 2011; Lee, Kim, & Joo, 2012; ucts are staged onto staging areas until their destination trailers are
Miao, Lim, & Ma, 2009; Vahdani & Zandieh, 2010), tabu search docked in the crossdock. Without loss of generality, it has been as-
sumed in Shakeri et al. (2012) that the products whose destination
trailers have not yet been assigned to an available door are temporar-
ily staged in front of the doors where they have been unloaded. The
1
In this paper, the terms “trailers” and “trucks” convey the same concept and are formation of staging areas is illustrated in Fig. 1, based on Bartholdi
used interchangeably. and Gue’s observation from real crossdocks (Bartholdi & Gue, 2004).
174 M. Shakeri et al. / Expert Systems With Applications 45 (2016) 172–184

Given the above operational characteristics, a solution to the truck

scheduling problem under study entails the solutions to the following
three decision sub-problems:

1. How are trailers sequenced for the door assignment?

2. To which door are trailers assigned?
3. How is the movement of products to their destination doors
scheduled when one forklift is available at each origin door?

It is important to note that the assumption of non-preemptive

door assignment incurs complexities in solution feasibility of the
aforementioned truck scheduling problem. The reason is that the
problem may be entrapped in deadlock and no feasible solution is
produced. By deadlock, we mean a situation where all doors have
been assigned to trailers and cannot be released by any of them.
This is because the trailers are waiting for some other trailers to re-
ceive their products but no door is available for unloading the re-
quired products. Avoiding deadlock and establishing solution feasi-
bility depends on how trailers are sequenced for the door assign-
ment. This was addressed in Shakeri et al. (2012) by proposing a two-
phase greedy heuristic algorithm – referred to as dependency rank-
ing search (DRS) plus door fitness (DF) or simply the DRS+DF – where
the first phase is merely dedicated to establishing solution feasibil-
ity when a limited number of doors are available. The detailed truck
scheduling procedure is then carried out in the second phase after
a feasible schedule was found in the first phase. Fig. 2 shows how
the above three decision sub-problems are addressed in the DRS+DF
heuristic. As the figure shows, a feasible trailer sequence list is con-
structed in the first phase by the DRS. Next the assigned door of each
Fig. 1. Structure and operation of the crossdocking model proposed in Shakeri et al. sequenced trailer together with the moving order of unloaded prod-
(2012). ucts to their destination doors are determined in the second phase by
the DF to produce the objective value which is the makespan.
In this paper we aim at optimizing the feasible solutions pro-
duced by the DRS+DF heuristic using TS and VNS metaheuristics.
As the figure shows, some space is reserved in front of each door to The focus of the paper, however, is on the computational aspect of
stage freight for that door. the two algorithms by integrating an incremental evaluation func-
Synchronous crossdocking has been employed in Shakeri et al. tion into their structure, allowing the runtime to be more efficiently
(2012) which allows trucks to be available at the crossdocking termi- dedicated to solution optimization rather than being spent on fitness
nal at any time the terminal manager desires. Truck scheduling starts evaluation.
when the first trailer is assigned to a door and ends when the last Fig. 3 (a) illustrates the truck scheduling solution encoding by
trailer leaves its assigned door. As soon as a trailer is assigned to a depicting how a solution is represented and decoded in the search
free door, a worker unloads the exchanged products onto the staging space. According to the figure, a solution s is represented by a permu-
area provided in front of the door. The products are then moved by tation of n trailers. The quality of the solution is assessed by a fitness
forklifts and discharged onto the respective staging areas of the doors function that determines the assigned doors of the trailers (according
to which their destination trailers are assigned. Similar to unloading, to their position in the permutation) and schedules the movement of
a worker is dedicated for loading the exchanged products onto a des- unloaded products to their destination trailers. Given the structure
tination trailer. The objective of the truck scheduling problem is thus of the proposed two-phase heuristic, this fitness function is the DF
to find an optimal schedule of trucks for the assignment to dock doors heuristic. Note that the evaluation measure is the makespan, as the
so that the total crossdocking operation time starting when the first objective function f(s) of solution s.
trailer is assigned to a door till the last trailer leaves its assigned door As seen in Fig. 3, a simple notation is adopted in this paper to refer
is minimized. The objective function, i.e., the schedule length, is also to any element(s) of an array or list in the related figures and pseudo
referred to as makespan. codes. Given an array or list X with k elements, X[i] (1 ≤ i ≤ k) refers
The truck scheduling problem defined in Shakeri et al. (2012) is to the element located at position i of X whereas X[i . . . i ], where 1 ≤
subject to a limited number of dock doors and forklifts. This means i < i ≤ k, refers to the elements stored from position i to position i .
that the number of trucks in the crossdock yard may exceed the num- To generate a neighbor of a given solution s, adjacent pairwise
ber of doors, leading to queues of trucks waiting to be assigned to interchange neighborhood structure – the neighborhood is denoted
doors. For ease of modeling, however, it has been assumed that the by N(s) – is applied. The search operator exchanges the positions of
number of forklifts is equal to the number of doors in which one fork- two consecutive trailers with each other in the current solution. See
lift is dedicated to each door for moving its unloaded products. The Fig. 3(b) where two trailers b and c located at positions i and i + 1 of
allocation of the crossdock resources including dock doors, forklifts, solution s, respectively, are interchanged to form solution s . As the
and workers has been considered non-preemptive. In terms of doors, perturbed solution must be evaluated by the DF heuristic, the change
this means that a trailer does not release its assigned door unless in the assignment order of the two trailers may lead to a better or
it completes its operations (i.e., unloading/loading). Regarding fork- worse makespan. So the neighborhood of a solution s includes all the
lifts and workers, they cannot be interrupted for another service un- solutions that have been produced by interchanging the positions of
til they finish moving and unloading/loading their assigned products, two adjacent trailers in solution s. The number of neighbors is n − 1
respectively. for a solution sequence list of n trailers.
 M. Shakeri et al. / Expert Systems With Applications 45 (2016) 172–184 175

Fig. 2. The solution structure of the two-phase DRS+DF greedy heuristic algorithm proposed in Shakeri et al. (2012) to address the three decision sub-problems of the studied truck
scheduling problem.

Fig. 3. (a) How a solution s is represented and evaluated in the search space (b) a new solution s in neighborhood N(s) is produced by interchanging two adjacent trailers b and c
located at positions i and i + 1 of solution s, respectively.

The proposed neighborhood structure is adopted in accordance

with the nature of the problem. It is a rule of thumb that efficient
scheduling requires that interdependent trailers, i.e., trailers that
send or receive products to or from one another, be sequenced in
nearby positions relative to each other. Though very restrictive, adja-
cent pairwise interchange neighborhood structure encompasses the
following benefits. First, it ensures that no trailer is unreasonably
moved to a position far from its interdependent trailers. Second, the
neighborhood generation complexity is of O(n) for a solution of n
trailers, implying that the neighborhood size is scalable with respect
to the problem size. Last and most important is the capability to de-
velop an incremental evaluation of the neighbor solutions to achieve
significant reduction in runtime. The incremental evaluation of the
objective function is presented in the next section.

3. The incremental evaluation function

The incremental evaluation function proposed in this paper as-

sesses each neighbor of a given solution using the information the
solution already holds. The objective is to reduce runtime by invok-
ing the DF heuristic to evaluate only those portions of the neighbor
solutions whose fitness cannot be retrieved from the fitness of the Fig. 4. Solution s and the three examples of its perturbed solution s .
current solution.
Fig. 4 (a) shows a sample solution s representing a permutation
of eight trailers. Neighbor solution s is produced by interchanging produced for each of the two solutions. This turning point is referred
trailers 8 and 4 located at the fifth and sixth positions of solution s, to as “pivot” (and denoted by v) in this study. The other term is “win-
respectively. Since the door assignment is performed according to the dow”, which is defined as the number of trailers explicitly affected by
order the trailers are positioned, for both solutions s and s , the first a perturbation in the current solution to produce a new solution (it
four trailers are assigned exactly the same, meaning that all the cross- is denoted by w). A window starts from the very first position in the
docking timings are the same before the fifth trailer (i.e., trailer 8 for solution representation whose designated trailer has changed (which
solution s and trailer 4 for solution s ) is assigned. In other words, the is the pivot) and ends at the very last position whose assigned trailer
assignment procedure changes from the fifth trailer and the change has also changed. For the example of Fig. 4(a), the pivot is equal to
may extend to the remaining trailers so that a different makespan is five and the window is two.
176 M. Shakeri et al. / Expert Systems With Applications 45 (2016) 172–184

Depending on the perturbation made, different values for v and Procedure 1 fCheck(n, s, s , v, w).
w may be obtained. See Fig. 4(b) where solution s is produced after 1: Restore door availability after trailers before v are assigned;
arbitrarily re-ordering the four trailers located from the second to the 2: Record its assigned trailer for each unavailable door;
fifth position of s. In this figure, v = 2 and the window covers the four 3: for i = v to n do
affected trailers, i.e., w = 4. According to the definition of the win- 4: if g(s[i] ) = 0 then
5: return false; {feasibility of s is not verified}
dow of affected trailers, if two trailers located at different positions
6: else if i ≥ v + w and g(s[i] ) ≥ g(s[i] ) then
of the solution sequence list are exchanged with each other, the trail- 7: return true; {feasibility of s is verified}
ers positioned in between are also affected. Consider the example of 8: else
Fig. 4(c) where the first trailer in solution s is interchanged with the 9: Assign trailer s[i] to a free door;
last trailer. In this case, the window will cover the entire list of trailers 10: Update door availability if any assigned trailer can release its door;
11: end if
and the pivot is equal to one. 12: end for
The preceding paragraphs help in describing the proposed in- 13: return true; {feasibility of s is verified}
cremental evaluation strategy more clearly. Given a solution and its
neighbor, all the information regarding the assigned doors and cross- Table 2
docking timings of the trailers positioned before the pivot in the Interdependencies of eight trailers scheduled for crossdocking.
neighbor solution can be retrieved from the current solution. For
Trailer Successors Predecessors
those trailers placed after [and including] the pivot, however, the re-
quired information must be obtained by invoking the DF heuristic. As 1 {2, 4, 5, 6} {2, 3, 4}
perturbing the order of trailers for the assignment may distort fea- 2 {1, 4, 5} {1, 3, 4, 5, 8}
3 {1, 2, 5} {5, 6, 7, 8}
sibility, the incremental evaluation proceeds in two stages. First, the
4 {1, 2, 6} {1, 2, 8}
feasibility is checked and then, in the case the solution is feasible, its 5 {2, 3, 7} {1, 2, 3}
quality will be evaluated. 6 {3, 7, 8} {1, 4, 7, 8}
7 {3, 6, 8} {5, 6, 8}
8 {2, 3, 4, 6, 7} {6, 7}

3.1. Incremental feasibility verification

For the incremental feasibility verification of solution s , produced

after perturbing solution s, it is required to record the number of
available doors each time a trailer – picked from the sequence list
of solution s – is assigned. The number of available doors in solution Fig. 5. The moving order of the products unloaded from trailer 1 onto trailers 5 and 6
s remains the same as the number in s when the trailers positioned is different for solutions s and s . This requires that the pivot be displaced backwards
from position 7 to position 6 for the incremental evaluation of solution s .
before the pivot are assigned one after the other. From the pivot, how-
ever, the number of available doors must be re-calculated before the
assignment of each trailer. If in the meantime the number reaches 3.2. Incremental solution evaluation
zero, it shows that solution s is not feasible. On the other hand, if,
after passing the window of affected trailers, the number of avail- The incremental evaluation of a feasible solution s , produced after
able doors before assigning the next trailer is equal or greater than perturbing solution s, involves restoration of the timings from solu-
the one in solution s, it is concluded that s is feasible and the verifi- tion s for the trailers positioned before the pivot and then invoking
cation procedure terminates; otherwise, the procedure continues till the DF heuristic to schedule the remaining trailers (i.e., those posi-
the last trailer is assigned to an available door. In that case, solution tioned after and including the pivot). The crossdocking state just be-
s is also feasible. Given the parameters defined in Table 1, the incre- fore the assignment of the trailer placed at the pivot can be readily
mental feasibility verification of solution s is described in the pseudo retrieved using the information stored for solution s. The information
code of Procedure 1. includes the interdependencies and timings of those trailers located
According to Procedure 1, it is required to restore the avail- before the pivot, together with their associated products. It is used
ability status of the doors after the trailers positioned before the to restore the allocation status of the doors and availability times of
pivot have all been assigned. For each unavailable door, its as- both the doors and forklifts at the time where the trailer positioned
signed trailer (which has made the door unavailable) should then at the pivot is to be assigned.
be recorded. Note that solution s is a neighbor of s when w = 2 or Unfortunately, the restoration procedure is not straightforward as
else it could be any solution produced after arbitrarily perturbing the change in the scheduling order of the trailers may affect the cross-
solution s. docking process for the trailers scheduled before the pivot. Consider
the problem of optimizing the schedule of eight trailers in a cross-
docking terminal. The interdependencies of the trailers, i.e., the list
Table 1 of their successors and predecessors2 are given by Table 2. A local
Parameters defined to describe the fCheck and sEval procedures. search algorithm (based on the adjacent pairwise interchange neigh-
Parameter Description borhood) is used to optimize the schedule. Suppose the algorithm is
currently at solution s and is evaluating its neighbor s (see Fig. 5).
n Total number of trailers
According to the figure, the pivot is seven implying that the first six
s Current solution
s Perturbed solution trailers should be scheduled exactly similarly to solution s. However,
v Pivot the implication is not true when trailer 1 at the sixth position is to be
w Window of affected trailers assigned.
s[1...i] Partial solution from position 1 to i According to Table 2, trailer 1 has some products destined to trail-
sbest Best solution
g(j) Returns the number of available doors at the time of
ers 5 and 6, respectively, which must be moved by a single forklift.
assigning trailer j
h(j) Returns the order of trailers receiving products from trailer j
2
f(s) Objective function (returns the makespan for s) The successors of a trailer are those receiving products from the trailer and the
predecessors are those giving products to the trailer.
 M. Shakeri et al. / Expert Systems With Applications 45 (2016) 172–184 177

Procedure 2 sEval(n, s, s , v, sbest ).

1: v ← v;
2: for i = 1 to v − 1 do
3: if h(s[i] ) = h(s[i] ) then
4: v ← i; {pivot is displaced backwards}
5: end if
6: end for
7: for i = 1 to v − 1 do
Fig. 6. The assessment function detects that the moving order of products unloaded
8: Retrieve the assigned door and x-docking timings from s;
from a trailer b located at a position v of solution s is different from that in solution s.
9: if f (s[1..i] ) ≥ f (sbest ) then
10: return f (sbest ); {s is not better than the current best solution}
11: end if
12: end for
The forklift scheduling function proposed in Shakeri et al. (2012) in 13: for i = v to n do
the DF heuristic decides on the movement of products according to 14: if i < v and door availability timings have not changed then
the completion status of their recipient trailers assessed by the lat- 15: Retrieve the assigned door from s
est position that one of the predecessors of the trailers holds in the 16: Re-calculate x-docking timings when s[i] is assigned;
17: else
solution sequence list. A trailer that acquires a smaller value receives 18: Invoke DF to assign s[i] to the best available door;
its required products earlier than another with a larger value. Accord- 19: end if
ingly, trailer 5 should receive its products earlier than trailer 6 for so- 20: if f (s[1..i] ) ≥ f (sbest ) then
lution s. The last predecessor of trailer 5, which is trailer 2, is located 21: return f (sbest ); {s is not better than the current best solution}
at the seventh position while the last predecessor of trailer 6, which 22: end if
23: end for
is trailer 4, is at the eighth position. For solution s , the moving order 24: sbest ← s {best solution has improved}
is reversed as the positions of trailers 2 and 4 are interchanged. The 25: return f (sbest );
change can lead to different timings which, like dominoes falling, may
affect the next trailers. It is thus required to carry out a quick assess-
ment on the moving order of the products unloaded from the trailers than the best solution found so far. This is accomplished by evaluat-
placed before the pivot to determine the very first trailer in the so- ing the partial sequence list s [1 . . . i] updated each time a new trailer
lution sequence list, if any, whose unloaded products would have a is scheduled per iteration (see Procedure 2). The objective function
different moving schedule (relative to the previous solution). If one is (makespan) of the partial schedule is then compared with that of the
found, the original pivot is displaced backwards to the position of the current best schedule to decide whether to continue the evaluation
trailer. or stop the procedure. It is important to note that for situations in
The idea is better illustrated in Fig. 6. A solution s is generated by which it is required to completely evaluate a solution (e.g., to make a
perturbing a given solution s with the total n trailers. As the incre- non-improving move as it is done in a tabu search-based algorithm),
mental evaluation procedure depends on the value of the pivot and the sbest entry in the sEval procedure is set to a null value.
not the window of affected trailers, without loss of generality, it is No lower bound is imposed on the completion times of the re-
assumed that s belongs to N(s) and is formed by interchanging two maining unscheduled trailers whenever a trailer is scheduled. This
adjacent trailers c and d located at positions v and v + 1 of solution is because obtaining tight lower bounds requires estimating which
s, respectively. The quick assessment function is performed from the door would be assigned to the trailers to incorporate product travel
leftmost trailer in solution s and continues until it finds a trailer b distances. Unfortunately, correct estimations cannot be guaranteed
in a position v < v with a product moving schedule different from or are even achievable because of the high complexity of the truck
that in solution s. (It is possible that not such a trailer is found and all scheduling problem. In the case that product travel distances are ex-
the products unloaded from the trailers positioned before the pivot cluded, the calculated lower bounds would be so poor that they do
in solution s follow a moving schedule similar to the one in solu- not help to reduce the runtime but rather incur considerable over-
tion s.) Given the assessment function is denoted by h(j) for trailer j – head, counteracting the current savings.
see Table 1 – for the example in Fig. 5, we have h(s [6]) = (6, 5) and In the end, it should be noted that in the real implementation of
h(s[6]) = (5, 6) so the pivot should be repositioned from the seventh both fCheck and sEval Procedures, solution s shares its data with so-
to the sixth position. lution s . Put differently, no individual database is considered for the
Having updated the pivot, if necessary, the incremental evalua- solutions, eliminating the laborious process of copying data per eval-
tion procedure uses the information stored in solution s to retrieve uation. The complication is that both procedures may change the ini-
the crossdocking state while the trailers located before the pivot are tial data required for the evaluation of the next solution produced
being assigned. The schedule of the trailers positioned after [and in- after perturbing solution s, if any.3 For this purpose, any changes are
cluding] the pivot is then carried out by invoking the DF heuristic. recorded in a temporary database and will later be used to restore
For situations where the initial pivot has changed, it is possible to re- the initial information for the next evaluation. In the case the search
duce the assessment time by retrieving the assigned door of those makes a move to a new solution, the corresponding information will
trailers located between the new and old values of the pivot directly be updated accordingly.
from solution s. The requirement is that the availability times of the
available doors have not changed relative to those in solution s. In the 4. Applying incremental evaluation to TS and VNS metaheuristics
case of any difference, the evaluation procedure must use the DF to
assign the remaining trailers (till the last trailer) to the best available The proposed incremental evaluation function can be readily in-
door, respectively. The proposed incremental evaluation function is tegrated into both single-solution based and population based evo-
described in the pseudo code of Procedure 2. lutionary metaheuristics developed for the aforementioned truck
The ultimate goal is to integrate the proposed incremental evalua- scheduling problem. This section describes how the technique is ap-
tion function into TS and VNS metaheuristics so that new solutions to plied to two single-solution based metaheuristics: TS and VNS.
the truck scheduling problem are evaluated on the quest for a better
solution. Further improvement in the procedure runtime is achieved 3
For example, if a steepest descent algorithm is to be employed, one must evaluate
by immediately terminating the evaluation process if it is detected the objective function for every element of neighborhood N(s) of a given solution s
in the meantime that the solution being evaluated cannot be better before making the next best move.
178 M. Shakeri et al. / Expert Systems With Applications 45 (2016) 172–184

Table 3
Procedure 3 incrementalTS(n, s0 , r, l, imax ).
Parameters defined to describe the TS procedure.
1: s ← s0 ;
2: Initialize T by invalidating all its entries; Parameter Description
3: i ← 0;
4: sbest ← s0 ; {s0 is initially considered the best solution} n Total number of trailers
5: repeat s0 Initial solution
6: snt_best ← NULL; s Current solution
7: best_improved ← false; s Perturbed solution
8: for all 1 ≤ k ≤ l where T[k] is valid do v Pivot
9: if tabu tenure of T[k] is zero then N (s) A reduced sample of neighborhood N(s) of solution s
10: Invalidate T[k] ; r Number of neighbors in N (s) chosen randomly
11: else T Tabu list
12: Decrease tabu tenure of T[k] by one; l Tabu list size
13: end if imax Maximum number of non-improving iterations
14: end for snt_best Best non-tabu solution recorded per TS iteration
15: Build N (s) by randomly choosing r neighbors; sbest Best solution
16: for k = 1 to r do f(s) Objective function (returns the makespan for s)
17: Initialize s with one neighbor in N (s);
18: Initialize v based on the difference between s and s ;
19: if fCheck(n, s, s , v, 2) = true then
20: if move to s is tabu then tional burden. (The number of random neighbors is determined by a
21: sEval(n, s, s , v, sbest ); control parameter.)
22: else The current solution will then be replaced with either a better so-
23: sEval(n, s, s , v, snt_best );
lution, if any, or the best solution among the non-tabu moves eval-
24: end if
25: if f (s ) < f (sbest ) {aspiration criterion is applied} then uated for that iteration. The two trailers whose positions have been
26: sbest ← s ; {update the best found solution} exchanged with each other in the current move are put in the tabu
27: best_improved ← true; list. The trailer pair stays in the tabu list for a pre-defined number
28: else if move to s is non-tabu and f (s ) < f (snt_best ) then of iterations called the tabu tenure. The length of the tabu tenure is
29: snt_best ← s ; {update the best non-tabu solution}
30: end if
determined randomly for each pair within the interval equal to the
31: end if tabu list size. The termination criterion applied is to stop the TS algo-
32: end for rithm after a pre-specified number of iterations without an improve-
33: if best_improved = true then ment in the objective function value. (The maximum number of non-
34: s ← sbest ; {move to the best solution}
improving iterations is specified by a control parameter.) Aspiration
35: if move to sbest is non-tabu then
36: Add the exchanged trailer pair in sbest to T ; criteria are employed to ensure that moves which are exceptionally
37: Update tabu tenure randomly in [1, l] promising are not ignored because of the tabu status of some com-
38: else ponent elements. The aspiration criterion applied in the implemen-
39: Update the corresponding tabu tenure randomly in [1, l]; tation of the proposed TS algorithm is to allow a move, even if it is
40: end if
41: i ← 0;
tabu, if it results in a solution with an objective function value better
42: else than that of the current best-known solution.
43: s ← snt_best ; {move to the best non-tabu solution} According to the code, the procedure evaluates tabu moves to
44: Add the exchanged trailer pair in snt_best to T ; check whether they lead to solutions better than the overall best
45: Update tabu tenure randomly in [1, l]
solution or not. In this case, the procedure sEval is invoked by having
46: i ← i + 1; {no improvement}
47: end if the best solution (i.e., sbest ) so that it can terminate the evaluation in
48: until i = imax early stages, to reduce the runtime, if it detects the current solution
49: return f (sbest ); cannot be better than the overall best solution. On the other hand,
for a non-tabu move, the evaluation cannot terminate unless it is de-
tected that the solution is not better than the best solution produced
from the previous non-tabu moves (denoted by snt_best ) at the same
4.1. The incremental TS iteration. Finally note that due to the application of adjacent pairwise
interchange neighborhood structure, the window of affected trailers
Originally proposed by Glover (1986), TS is an extension of classi- (which is used as the input parameter to the procedure fCheck) for
cal local search methods. According to Gendreau and Potvin (2010), each perturbation is equal to two.
the basic principle of TS is to pursue LS whenever it encounters a lo-
cal optimum by allowing non-improving moves. This implies that the 4.2. The incremental VNS
adjacent pairwise interchange neighborhood structure can be applied
to TS algorithm as well and the incremental evaluation function can VNS is a metaheuristic, proposed by Mladenovic and Hansen
then be used accordingly. The incremental TS algorithm is described (1997), based on the idea of a systematic change of neighborhood
in the pseudo code of Procedure 3 according to the parameters de- both in a descent phase to find a local optimum and in a perturba-
fined in Table 3. tion phase to get out of the corresponding valley. In other words,
Given an initial solution generated by the two-phase DRS+DF VNS proceeds by a descent method to a local minimum then ex-
heuristic, the TS algorithm starts exploring the search space by evalu- plores, systematically or at random, increasingly distant neighbor-
ating its neighbors. To make a move, a pre-specified number of neigh- hoods of this solution. According to Talbi (2009), neighborhood struc-
bors are generated randomly and evaluated by the DF algorithm. In tures for permutation scheduling problems are position-based like
the regular TS, one must evaluate the objective function for every ele- the insertion (or shift) operator or order-based like the exchange
ment of the neighborhood N(s) of the current solution s. For the truck and inversion operators. In a similar work by Vahdani and Zandieh
scheduling problem addressed in this study, however, the full evalu- (2010), they developed nine neighborhood structures, both position-
ation is extremely expensive from the computational standpoint. As based and order-based, for their proposed VNS algorithm. Our pro-
a consequence, this work deploys probabilistic TS by considering only posed neighborhood structures partly differ from those defined by
a random sample N (s) of N(s) to significantly reduce the computa- Vahdani and Zandieh (2010) as they have been adapted to the
 M. Shakeri et al. / Expert Systems With Applications 45 (2016) 172–184 179

Table 4
Parameters defined to describe the VNS p rocedure.

Parameter Description

n Total number of trailers

s0 Initial solution
s Current solution
s , s Perturbed solutions
v Pivot
w Window of affected trailers
kmax Maximum number of neighborhoods
Nk (s) Neighbor structure k (k = 1 to kmax ) of solution s
N (s) A reduced sample of neighborhood N(s) of solution s
r Number of neighbors in N (s) chosen randomly
imax Maximum number of non-improving iterations
sl_best Best local solution found by LS per VNS iteration
f(s) Objective function (returns the makespan for s)

vent the move from crossing beyond the borders of the solution
sequence list.

The VNS starts searching by generating a solution at random

from neighborhood Nk (k = 1) of the initial solution (the procedure
is referred to as shaking). The shaking is followed by a local search
routine on the quest for better solutions. A move is made to a new
solution if it is better than the current solution. Otherwise, the algo-
Fig. 7. The VNS neighborhood structures.
rithm jumps to Nk (k ← k + 1), to invoke the shaking and local search
procedures from there. If the current solution improves, the search
nature of the problem. The neighborhoods are defined within clus- resumes from the first neighborhood, Nk (k = 1), or else the VNS
ters of interdependent trailers such that random exchanges of trail- will move to the next neighborhood, Nk (k ← k + 1), if any, to repeat
ers are carried out in the cluster and not in the whole trailer list. In the shaking and local search procedures from that neighborhood.
the original implementation of the proposed VNS algorithm, eight The algorithm will stop if after a pre-specified number of iterations,
neighborhood structures were defined. In the preliminary experi- no improvement is made by exploring each neighborhood Nk per
mentations conducted, however, it was noticed that solution quality iteration. (The maximum number of non-improving iterations is
improved in normally three or four of the neighborhoods combined specified by a control parameter in the proposed implementation.)
with the fact that exploring all the eight neighborhoods incurs con- Unlike the basic VNS which employs a deterministic local search
siderable computational overhead. This is consistent with what was by evaluating every element of a neighborhood (Mladenovic &
observed by Hansen, Mladenovic, Brimberg, and Moreno Perez (2010) Hansen, 1997), we adopt reduced local search by randomly choos-
that the best value for the number of neighborhoods is often two or ing a pre-specified number of neighbors to evaluate per iteration
three. Regarding their order in the search, neighborhoods are usu- (by the DF algorithm). This is because similar to the TS algorithm,
ally ranked in such a way that solutions are explored increasingly far the full evaluation is too costly from the computational perspective.
from the current solution. In this way, intensification of the search (The number of random neighbors evaluated per iteration by the lo-
around the current solution is followed naturally by diversification cal search procedure is determined by a control parameter.) Note
(Hansen & Mladenovic, 1997). Given the above explanations, three that the neighborhood structure applied in the LS is adjacent pair-
neighborhood structures are proposed according to the following wise interchange. The incremental VNS algorithm is described in the
order: pseudo codes of Procedures 4–6 according to the parameters defined
in Table 4.
1. Adjacent pairwise interchange (N1 ). In this structure, a trailer is Note that the shake procedure ensures that a feasible solution is
chosen randomly and its position in the solution sequence list is randomly chosen from the current neighborhood. It is also required
exchanged with its subsequent trailer. This neighborhood struc- to fully evaluate the random solution before invoking the local search
ture is illustrated in Fig. 7(a). routine. For this purpose, the entry to the best solution (i.e., sbest ) in
2. Shift between two random points within a cluster (N2 ). Given the sEval procedure has been set to a null value in the code.
two points p1 and p2 , randomly chosen within a cluster of inter-
dependent trailers in the solution sequence list, a forward shift is 5. Experimental setup and results analysis
applied if p2 > p1 or else a backward shift is applied. Fig. 7(b) de-
picts the two situations where the forward and backward shifts The efficiency of the proposed incremental evaluation function in-
are applied, respectively. It is good to note that for both shifts in tegrated in the body of the TS and VNS metaheuristics is assessed
the figure the pivot and window of affected trailers are equal to against a deterministic local search algorithm that uses a simple com-
four. plete evaluation of the objective function. Note that the neighborhood
3. Inverting and moving a subsequence of trailers within a clus- structure applied in the LS is adjacent pairwise interchange. All the
ter (N3 ). Again two points p1 and p2 are randomly chosen within compared algorithms are to optimize the solutions generated by the
a cluster of interdependent trailers to form a trailer subsequence. two phase DRS+DF heuristic algorithm.
The subsequence is first inverted and then moved forward if p2 >
p1 or backward otherwise (see Fig. 7(c)). Note that for the forward 5.1. Test data
move in the figure, the pivot is three while for the backward, it is
equal to two. Nonetheless, the window of affected trailers for both We use the same data set that were generated in Shakeri et al.
moves is six, which is twice the trailer subsequence size. It is also (2012) to assess the efficiency of the two-phase DRS+DF heuristic al-
good to note that the values of p1 and p2 may be adjusted to pre- gorithm in establishing solution feasibility. Each category of the test
180 M. Shakeri et al. / Expert Systems With Applications 45 (2016) 172–184

Table 5
Procedure 4 shake(n, s, k).
Test data characterization.
1: repeat
2: Generate a random solution s ∈ Nk (s); (a) Data sizes defined for the truck scheduling problem
3: Initialize v and w based on the difference between s and s ;
Size Number of doors Number of trailers
4: until fCheck(n, s, s , v, w) = true
5: return s ; α=1 α=2 α=4 α=8

Small 8 8 16 32 64
Procedure 5 LS(n, s, r). Medium 32 32 64 128 256
Large 96 96 192 384 768
1: sl_best ← s;
Huge 256 256 512 1024 2048
2: repeat
3: s ← sl_best ;
(b) FMLs defined for the truck scheduling problem
4: Build N (s) by randomly choosing r neighbors;
5: for i = 1 to r do FML Percentage (%) of trailers
6: Initialize s with one neighbor in N (s); [that receive the exchange volume of each single trailer]
7: Initialize v based on the difference between s and s ;
8: if fCheck(n, s, s , v, 2) = true then Low Less than 10% of the total trailers
9: sEval(n, s, s , v, sl_best ); Medium 10–25% of the total trailers
10: if f (s ) < f (sl_best ) then
High 25–50% of the total trailers with 95% probability
11: sl_best ← s ;
More than 50% of the total trailers with 5% probability
12: end if
13: end if
14: end for
15: until f (s) ≤ f (sl_best )
extreme case, the number of recipient trailers is equal to the number
16: sl_best ← s;
17: return sl_best ;
of exchanged pallets, i.e., each trailer receives one pallet.
To identify the IDs of the recipient trailers for each trailer, i.e., to
form the interdependencies, two approaches have been adopted in
Procedure 6 incrementalVNS(n, s0 , kmax , r, imax ). Shakeri et al. (2012): uncorrelated and correlated. In the uncorrelated
1: s ← s0 ; approach, the IDs are chosen uniformly from the entire group of trail-
2: i ← 0; ers while in the correlated one, the dependency history of the trailer,
3: repeat
if any, is involved in the procedure of specifying the IDs. By the de-
4: k ← 1;
5: repeat pendency history, we mean the list of the trailers which either have
6: s ← shake(n, s, k); sent some products to the trailer or, together with the trailer, have re-
7: Initialize v based on the difference between s and s ; ceived products from the same donor trailer. The correlated approach
8: sEval(n, s, s , v, NULL); assigns higher probability to the trailers in the history list to be cho-
9: s ← LS(n, s , r);
10: if f (s ) < f (s) then
sen as the recipient trailers.
11: s ← s ; {make a move} So, according to the steps introduced for the data characteriza-
12: k ← 1; {first neighborhood} tion, each category of the test data has been represented in a 3-tuple
13: i ← 0; (size:α ; FML; correlation status). Ten random instances have been
14: else
generated for each combination of the problem sizes, trailer-to-door
15: k ← k + 1; {next neighborhood}
16: end if ratios (α ), freight mixed levels (FML), and uncorrelated and corre-
17: until k = kmax lated interdependencies, which equal to 4 × 4 × 3 × 2 × 10 = 960 to-
18: i ← i + 1; tal instances. In case one or more attributes are not reflected in the
19: until i = imax tuple, it means that the reduced tuple encompasses all the values of
20: return f (s);
the missing attributes. For example, the tuple (small:1; low) repre-
sents both (small:1; low; uncorrelated) and (small:1; low; correlated)
tuples.
data characterizes four elements: the size of the crossdocking termi- Rectangular crossdocks have been used in Shakeri et al. (2012) for
nal (formally represented by the number of doors (Bartholdi & Gue, all the categories of the test data. Given d the number of dock doors,
2004)), queue length of waiting trailers (i.e., the ratio of the number the crossdock layout is formed by placing d/8 doors on the shorter
of trailers to the number of doors, which is denoted by α ), product side of the crossdock and 3d/8 ones on the longer side. (Note that d
exchange patterns between trailers, and correlation status of inter- is a multiple of 8.) The distance between doors has been specified
dependent trailers. The values assumed for the number of doors and by assuming that the distance between adjacent doors (and between
trucks are summarized in Table 5(a). The patterns for the exchange each corner and the closest door) is 1 distance unit.
composition of a trailer have been formalized by three freight mixed
levels (FMLs) defined in Table 5(b). A freight mixed level is quantified
5.2. Metaheuristic parameter tuning
by the percentage of the trailers receiving the exchange volume of
each single trailer. This also specifies the level of interdependencies
The proposed TS and VNS algorithms have a number of control
defined between trailers.
parameters (or factors) that must be tuned before the performance
All trailers have a maximum capacity of 60 pallets.4 The initial
of each is to be compared with the LS algorithm. A detailed analysis
number of pallets in each trailer has been uniformly distributed be-
was conducted by Shakeri (2013) to tune the control parameters of
tween 20 and 40 pallets. The amount of exchange for the trailers
both metaheuristics by employing a 2k factorial design. The method-
ranges evenly from half to all of its initial carrying pallets. For each
ology is to choose a reasonable minimum and maximum value for
trailer, a random number is generated uniformly in the interval cor-
each control parameter and then test all the combinations of the low
responding to a particular mixed level to represent the number of
and high values of the parameters.
trailers receiving the exchange volume of that trailer. Note that in the
Table 6 lists the control parameters defined for both TS and VNS
algorithms together with their considered minimum and maximum
4
In this study we simply assume that one pallet contains only one product. In that values. The values of the factors have been chosen in a way to alter the
case the words “pallet” and “product” are interchangeable. performance of the algorithms in terms of their search methodology
 M. Shakeri et al. / Expert Systems With Applications 45 (2016) 172–184 181

Table 6
The TS and VNS factor levels.

(a) The TS factor levels

Factor Description Min. Max.

A Tabu list size (denoted by l) 0 20

B Number of neighbors chosen randomly per iteration (denoted by r) 10 50
C Maximum number of non-improving iterations allowed (denoted by imax ) 20 100

(b) The VNS factor levels

Factor Description Min. Max.

A Number of neighbors chosen randomly per LS iteration (denoted by r) 10 50

B Maximum number of non-improving iterations allowed (denoted by imax ) 1 20

Table 7 optimization is tested on the solutions produced by the two-phase

Values assigned to the TS and VNS control parameters after tuning.
DRS+DF heuristic algorithm. According to the work carried out in
(a) Values assigned to the TS control parameters Shakeri et al. (2012), establishing solution feasibility (by applying the
Problem size l r imax
DRS+DF) is impractical for data instances of α = 8 and the medium
and high mixed levels, and α = 4 and the high mixed level.5 So these
Small 10 if α ≤ 4 10 100 data categories are not included in the experiment. For the remaining
v1 if α = 8
categories, a representative subset of 150 instances are considered for
Medium 10 if α = 1 10 100
v1 if α ≥ 2
optimization by the proposed algorithms.
Large v1 10 if α ≤ 2 100 if α ≤ 2 Table 8 presents the average experimental results for all the com-
20 if α ≥ 4 50 if α ≥ 4 binations of the problem sizes, trailer-to-door ratios, and freight
Huge v1 10 if α ≤ 2 50 mixed levels.6 The second to forth columns of the table include
20 if α ≥ 4
the initial makespan produced by the DRS+DF heuristic and the
(b) Values assigned to the VNS control parameters makespan improved by the LS (that exploits the complete evaluation
of the objective function) together with its runtime, respectively. The
Problem size r imax
stochastic components embedded in the implementation of both TS
Small 50 20 and VNS metaheuristics require that the algorithms be run for a few
Medium 50 20
times per problem instance to establish robustness in solution qual-
Large 50 20
Huge 50 20 ity. The number of times each algorithm is run per instance is set to
five in the experiments to achieve a balance between solution qual-
1
Indicates that the size of the tabu list varies between 0, 10, and 20.
ity robustness and runtime. The remaining four columns of the table
thus include the best makespan obtained within all the five runs to-
and search space. For example, the two values of factor A defined for gether with the total runtime taken for both TS and VNS algorithms,
the TS in Table 6(a) represents two situations where in the former, respectively.
the algorithm uses no tabu list in its search path while in the latter it According to Table 8, as the number of trailers approaches 400
uses a rather long tabu list to orient the search. The two values con- (the test data categories are shown in bold in the table), the TS al-
sidered for the remaining factors are mainly used to balance a trade- gorithm exhibits a better performance in terms of solution quality
off between solution quality and runtime by confining and expanding improvement vs. computational time compared to the other two al-
the search space, respectively, according to the structure of the given gorithms. For example, in the extreme instance of scheduling 2048
algorithm. This results in having different minimum and maximum trailers in a 256-door crossdock, i.e., the (huge:8) data category, the
values for control parameter imax for the TS and VNS algorithms. TS is able to find comparable solutions in nearly 39 h shorter than the
The standard methodology for a parameter setting involves choos- LS. In other words, the average time taken by the TS algorithm to im-
ing a representative subset of the problem instances to train param- prove a solution for this data category is nearly 4 h per run compared
eters on, avoiding over-training the data (Silberholz & Golden, 2010). to 57 h taken by the LS to improve the same solutions. This advan-
For both TS and VNS algorithms, a factorial ANOVA (analysis of vari- tage is mainly due to the incremental evaluation function embedded
ance) has been applied to decide whether and which factors or in- in the body of the algorithm combined with the simple yet efficient
teractions have significant effect (influence) on the response variable search strategy of the TS to escape local optima. (Again note that the
(here, the makespan) – see Shakeri (2013). It should be noted that runtime recorded for both TS and VNS algorithms is the total time
the three assumptions on normality, independence and equality of taken to improve a solution for all the five runs.)
variance are all applicable to the experimental model. Given the gen- On the other hand, even with the incorporation of the incremen-
erated synthetic data and the stochastic structure of the proposed tal evaluation function, the VNS demands a much longer time than
metaheuristics, no biasness was observed questioning the validity of the other two algorithms to terminate its search process and return
the conducted experiments. Table 7 summarizes the values assigned the final solution. Although this has led to relatively better solutions
to the TS and VNS control parameters after being tuned based on the for almost all data categories, the exponential increase in the run-
procedure carried out by Shakeri (2013). time for larger problem instances leaves us no convincing justifica-
tion to implement the VNS in an industrial setting. Consider a large
5.3. Performance evaluation crossdocking terminal hosting over 2000 trucks daily where truck

All the proposed procedures were coded in C++ and run on a 5

Of the total 960 instances generated in Shakeri et al. (2012), feasible solutions have
CoreTM 2 Quad 2.66 GHz computer with 3GB RAM. As already stated, been produced for 644 ones.
the performance of the incremental evaluation mechanism in re- 6
It was noticed that the performance of the algorithms was almost independent of
ducing the runtime of the TS and VNS metaheuristics for solution the correlation status of data instances. So this attribute is not included in the table.
182 M. Shakeri et al. / Expert Systems With Applications 45 (2016) 172–184

Table 8
The performance of the incremental TS and VNS algorithms compared to the LS that exploits the complete evaluation of the objective function.

Data category DRS+DF LS TS VNS

Cmax Cmax time(s) Cmax time(s) Cmax time(s)

(small:1;low) 54 50 0.09 30 34.8 30 29.2

(small:1;medium) 63 47 0.18 34.5 34.9 34.5 41.2
(small:1;high) 55.3 48 0.2 42.7 51.4 41.7 57.7
(small:2;low) 104.5 96.5 0.6 92.5 102.8 93.5 172.5
(small:2;medium) 167.7 139.3 1.05 114 162.7 115 286.4
(small:2;high) 258 198.5 1.3 188.5 160.7 190.5 455.3
(small:4;low) 259.4 236.8 3.9 224.6 236.4 218.4 1077.4
(small:4;medium) 792 772 3.8 666 439 617 2031.3
(small:8;low) 1019.2 916 24.3 876.8 701.6 829.6 6661.3
(medium:1;low) 149 104.2 7.1 80.8 257.1 74.8 885
(medium:1;medium) 170.2 130.7 8.8 114.7 377 108 1517.4
(medium:1;high) 269.5 245.7 14.3 221.7 813 203.2 4379.6
(medium:2;low) 265.2 191.4 26.8 176.4 508.1 164.6 4385.6
(medium:2;medium) 370.8 313.4 26.7 288.2 753.1 278.4 3801.6
(medium:2;high) 1076.7 902 104.2 750.3 2175.9 695.3 24143.2
(medium:4;low) 492 414.5 149.2 404.7 1332.3 387.5 9753.6
(medium:4;medium) 3284.7 2670 323.3 2510 1919.1 2252.3 22264.7
(medium:8;low) 1324.2 1137 606 1131.7 2964 1091.5 28209.9
(large:1;low) 433.3 286.3 221.4 286.7 1425.5 254 10473.7
(large:1;medium) 514.5 385.5 308.5 371.2 2368.2 342.2 15625.9
(large:1;high) 952.5 767 525 711.7 6544.2 648.5 64445.5
(large:2;low) 723.5 524.5 898.6 490.7 3191.2 472.5 19714.5
(large:2;medium) 1063.6 823 732.4 780.2 3789.1 729.4 29878.9
(large:2;high) 2206.5 1829.8 1222.2 1772.3 6866.7 1670.8 55400.9
(large:4;low) 1361.1 1018 2697 1006.4 5702.8 958.6 50729.2
(large:4;medium) 17682.7 13931 11591.2 13262.3 9515.2 11737.7 125782.9
(large:8;low) 2792.8 2403.6 14553.1 2376 14508.1 2326.8 115038.1
(huge:1;low) 1427.5 911.2 4591.9 853.2 6226.7 775.2 70662.1
(huge:1;medium) 1645.5 1204 7797.8 1170 10519.4 1096.5 122688.9
(huge:1;high) 3157.3 2262.7 13990.5 2170.3 23312.7 1948 319741.1
(huge:2;low) 2316.7 1563.9 15274.7 1497.1 11646.4 1389.3 149322.1
(huge:2;medium) 3228.7 2206.5 15851.3 2218.7 9548.1 2053.2 162832.2
(huge:2;high) 5459.5 4665 17130.8 4581.3 17581.5 4389.7 269347.4
(huge:4;low) 3951.8 2728 53355.2 2743 34936.1 2635 246255.9
(huge:8;low) 7794.7 6110.4 205176.4 6058.4 66368.2 6006.9 477024.6

Fig. 8. Experimental results depicting the performance of the incremental evaluation function in reducing the runtime of the TS and VNS for solution optimization compared to
the LS that uses a simple complete evaluation of the objective function. As the diagram shows, the TS is successful in reaching comparable solutions in much lesser time compared
to the other two for huge size problem instances of scheduling over 2000 trucks in a 256-door crossdock.

scheduling must be carried out in a short-term planning horizon of evaluation function with the search strategy of the TS becomes more
one day. Given the results shown in Table 8, even one single run of efficient in reducing the runtime. This, however, is not true for the
the VNS cannot terminate the optimization process within that time VNS algorithm. The runtime dramatically increases far from the LS as
limit. the size of instances increases. It is thus required to combine some ad-
Fig. 8 better illustrates the behavior of the incremental TS and VNS vanced techniques (other than the incremental evaluation) with the
algorithms in terms of solution quality and runtime with respect to basic components of the proposed VNS algorithm so that a good bal-
the LS. In this figure, the results of the LS have been set to one to be ance of exploitation and exploration of the search space is achieved
taken as the reference point. We have normalized the absolute differ- in reasonable running times. Accordingly, we recommend the current
ences in the average values of makespan and runtime obtained by the implementation of the incremental TS metaheuristic to be applied in
TS and VNS from those of the LS in [0,1]. In the case of improvement, an industrial setting.
the normalized results have been subtracted from one and in the case We expand our experiments to assess the direct impact of the in-
of deterioration have been added to one. According to the figure, as cremental mechanism on reducing the runtime of the TS and VNS
the size of instances increases, the combination of the incremental metaheuristics by running them with and without the incremental
 M. Shakeri et al. / Expert Systems With Applications 45 (2016) 172–184 183

Table 9
Computational results for the TS and VNS algorithms with and without the incremental mechanism.

Data category TS VNS

tcmp (s) tinc (s) r. r.(%) tcmp (s) tinc (s) r. r. (%)

(small:8;low) 698.5 433.6 37.92 1273.7 890 30.12

(medium:8;low) 4841.2 2406.3 50.3 6876.3 3475.9 49.45
(large:8;low) 30699.4 10938.1 64.37 40623.2 14017.2 65.49
(huge:8;low) 172363.4 50359.3 70.78 198605.8 76021.3 61.7

6. Concluding remarks and future work

This paper studied truck scheduling optimization in a resource-

constrained crossdock. The truck scheduling problem was defined in
Shakeri et al. (2012) according to a real crossdocking model. The pa-
per tackled the optimization from the computational perspective by
developing an incremental evaluation function that evaluates only
the transformation applied to the current solution rather than the
complete evaluation of the neighbor solution. To the best of our
knowledge, no related papers on truck scheduling optimization in
Expert and Intelligent Systems explicitly presented any mechanism
Fig. 9. The reduction in the TS computational time is linear with the problem size. This to tackle their addressed problems from the computational perspec-
is not true for the VNS. tive. They rather highlighted their implementation methodology of
their proposed metaheuristics. Lee et al. (2012) focused on the solu-
tion representation of their proposed GA by developing three chro-
evaluation of the objective function. The test data used for the ex- mosome types. Kuo (2013) streamlined the development of a VNS
periment consist of those instances where the number of trucks to that accepts new neighborhood solutions with a probability based on
the number of doors is maximum (i.e., α = 8). Both algorithms were their performance, even if the solution is worse similar to the method
run five times per instance on a CoreTM i5-2320 3.00 GHz computer of simulated annealing. Finally, Miao et al. (2014) focused on the de-
with 4GB RAM. As solution quality is not considered in the experi- velopment of an adaptive TS by defining four kinds of moves to al-
ment, without loss of generality, we limit the number of randomly- ternately explore neighborhood space based on a number of adaptive
chosen neighbors that must be evaluated before a move is made per rules.
iteration to 10 for all the instances. Given tcmp , the total time taken The incremental mechanism was embedded in the body of two
with the complete evaluation and tinc , the time with the incremental single-solution based metaheuristics including TS and VNS so that
evaluation, we define r.r. (%) as the relative reduction in the compu- the algorithms could dedicate their runtime to solution optimiza-
tational time of both algorithms when the incremental evaluation is tion rather than spending it on fitness evaluation. The experimental
deployed. The r.r. (%) measure is calculated according to the following results demonstrated that the incremental evaluation of the objec-
formula: tive function is very significant in reducing the computational time
of both metaheuristics. In particular, the combination of the incre-
tcmp − tinc mental mechanism with the search strategy of the TS turned out to
r.r. (%) = × 100 be highly efficient in reducing the runtime. On the other hand, the
tcmp
incremental VNS demanded a much longer time than the other two
Table 9 shows the average computational times of the two fit- algorithms (the TS and LS) to terminate its search process and return
ness evaluation strategies employed for the TS and VNS metaheuris- the final solution. This means that the integration of the incremental
tics, respectively. According to the table, the incremental evalua- mechanism into the VNS as its own would not justify the implemen-
tion is able to significantly reduce the runtime for both algorithms, tation of the algorithm in an industrial setting where a large cross-
more specifically for large and huge size instances where the re- docking terminal hosting over 2000 trucks daily is considered.
duction exceeds half of the time taken with the complete evalua- As a subject of future work, we should investigate the application
tion. Comparing the TS and VNS, it is noticed that the reduction in of some further advanced techniques to the basic components of the
the TS computational time is almost linear with the problem size proposed TS and VNS metaheuristics so that a good balance of ex-
(see Fig. 9). This is, however, not true for the VNS as the amount of ploitation and exploration of the search space is achieved in reason-
reduction slightly degrades from large to huge size instances. The able running times. One idea is to enhance the solution representa-
reason should be sought in the neighborhood structures defined tion of the problem by not only encoding the sequence of trailers for
for the two algorithms. The efficiency of the incremental evalua- the door assignment but also incorporating their assigned doors. This
tion is directly proportional to the degree of the perturbation made can simplify the objective function which in turn would lead to a fur-
to produce a neighbor solution. The lower is the perturbation; the ther reduction in the runtime. Another research topic is to investigate
more efficient is the incremental mechanism. The TS explores the the application of the proposed incremental evaluation function to
search space by accepting non-improving moves within the adja- evolutionary algorithms developed for the same problem. The func-
cent pairwise interchange neighborhood whereas the VNS explores tion is used as a fitness function for evaluating each individual. That
the search space by shaking within two larger neighborhoods. This is, the fitness evaluation of new individuals that have been created
reinforces our observation derived from the first phase of the ex- from old individuals can be carried out by using the information the
periments that the current implementation of the incremental TS old ones already hold. The developed solution approaches can also be
metaheuristic is more appropriate to be applied in an industrial extended to address other related combinatorial optimization prob-
setting. lems defined in different industrial applications.
184 M. Shakeri et al. / Expert Systems With Applications 45 (2016) 172–184

Acknowledgment Kang, J. M., Mokbel, M. F., Shekhar, S., Xia, T., & Zhang, D. (2010). Incremental and gen-
eral evaluation of reverse nearest neighbours. IEEE Transactions on Knowledge and
Data Engineering, 22(7), 983–999.
This research is supported by the collaborative research project Kuo, Y. (2013). Optimizing truck sequencing and truck dock assignment in a cross dock-
(CRP) between NTU and SIMTech and sponsored by NTU Graduate ing system. Expert Systems with Applications, 40(14), 5532–5541.
Scholarship. Lee, K., Kim, B. S., & Joo, C. M. (2012). Genetic algorithms for door-assigning and se-
quencing of trucks at distribution centers for the improvement of operational per-
formance. Expert Systems with Applications, 39(17), 12975–12983.
References Li, M. (2011). An improved kriging-assisted multi-objective genetic algorithm. Journal
of Mechanical Design, 133(7), 1–11.
Alvarez-Perez, G. A., Gonzalez-Velarde, J. L., & Fowler, J. W. (2009). Crossdocking-just Liao, T., Chang, P., Kuo, R., & Liao, C.-J. (2014). A comparison of five hybrid metaheuris-
in time scheduling: an alternative solution approach. Journal of the Operational Re- tic algorithms for unrelated parallel-machine scheduling and inbound trucks se-
search Society, 60(4), 554–564. quencing in multi-door cross docking systems. Applied Soft Computing, 21, 180–193.
Apte, U. M., & Viswanathan, S. (2000). Effective cross docking for improving distribu- Liao, T., Egbelu, P., & Chang, P. (2013). Simultaneous dock assignment and sequencing of
tion efficiencies. International Journal of Logistics Research and Applications, 3(3), inbound trucks under a fixed outbound truck schedule in multi-door cross docking
291–302. operations. International Journal of Production Economics, 141(1), 212–229.
Barthelemy, J. F. M., & Haftka, R. T. (1993). Approximation concepts for optimum struc- Madani-Isfahani, M., Tavakkoli-Moghaddam, R., & Naderi, B. (2014). Multiple cross-
tural design: A review. Structural Optimization, 5(3), 129–144. docks scheduling using two meta-heuristic algorithms. Computers & Industrial En-
Bartholdi, J., & Gue, K. (2004). The best shape for a crossdock. Transportation Science, gineering, 74, 129–138.
38(2), 235–244. Miao, Z., Cai, S., & Xu, D. (2014). Applying an adaptive tabu search algorithm to optimize
Bloori Arabani, A., Fatemi Ghomi, S., & Zandieh, M. (2011). Meta-heuristics implemen- truck-dock assignment in the crossdock management system. Expert Systems with
tation for scheduling of trucks in a cross-docking system with temporary storage. Applications, 41(1), 16–22.
Expert Systems with Applications, 38(3), 1964–1979. Miao, Z., Lim, A., & Ma, H. (2009). Truck dock assignment problem with operational
Bozer, Y. A., & Carlo, H. J. (2008). Optimizing inbound and outbound door assignments time constraint within crossdocks. European Journal of Operational Research, 192(1),
in less-than-truckload crossdocks. IIE Transactions, 40(11), 1007–1018. 105–115.
Buche, D., Schraudolph, N., & Koumoutsakos, P. (2005). Accelerating evolutionary algo- Mladenovic, N., & Hansen, P. (1997). Variable neighborhood search. Computers & Oper-
rithms with Gaussian process fitness function models. IEEE Transactions on Systems, ations Research, 24(11), 1097–1100.
Man, and Cybernetics, Part C: Applications and Reviews, 35(2), 183–194. Shakeri, M. (2013). Truck scheduling optimization in the logistics of crossdocking. Singa-
Burnham, K. P., & Anderson, D. R. (2002). Model selection and multimodel inference. New pore: Nanyang Technological University, Ph.D. thesis.
York: Springer-Verlag. Shakeri, M., Low, M. Y. H., Turner, S. J., & Lee, E. W. (2012). A robust two-phase heuristic
Crainic, T. G., Gendreau, M., Soriano, P., & Toulouse, M. (1993). A tabu search proce- algorithm for the truck scheduling problem in a resource-constrained crossdock.
dure for multicommodity location/allocation with balancing requirements. Annals Computers & Operations Research, 39(11), 2564–2577.
of Operations Research, 41(4), 359–383. Silberholz, J., & Golden, B. (2010), International Series in Operations Research & Manage-
Gendreau, M., & Potvin, J.-Y. (2010). Handbook of metaheuristics, International Se- ment Science (2nd, pp. 625–640). Springer Science+Business Media, LLC.
ries in Operations Research & Management Science (2nd, pp. 41–59). Springer Sci- Soltani, R., & Sadjadi, S. J. (2010). Scheduling trucks in cross-docking systems: A robust
ence+Business Media, LLC. meta-heuristics approach. Transportation Research Part E-Logistics and Transporta-
Ghobadian, E., Tavakkoli-Moghaddam, R., Javanshir, H., & Naderi-Beni, M. (2012). tion Review, 46(5), 650–666.
Scheduling trucks in cross docking systems with temporary storage and dock re- Talbi, E.-G. (2009). Metaheuristics: From design to implementation. Wiley Publishing.
peat truck holding pattern using grasp method. International Journal of Industrial Tang, L., Zhao, Y., & Liu, J. (2014). An improved differential evolution algorithm for
Engineering Computations, 3(5), 777–786. practical dynamic scheduling in steelmaking-continuous casting production. IEEE
Glover, F. (1986). Future paths for integer programming and links to artificial intelli- Transactions on Evolutionary Computation, 18(2), 209–225.
gence. Computers & Operations Research, 13(5), 533–549. Vahdani, B., & Zandieh, M. (2010). Scheduling trucks in cross-docking systems: Robust
Hansen, P., & Mladenovic, N. (1997). Variable neighborhood search for the p-median. meta-heuristics. Computers & Industrial Engineering, 58(1), 12–24.
Location Science, 5(4), 207–226. Van Belle, J., Valckenaers, P., Vanden Berghe, G., & Cattrysse, D. (2013). A tabu search
Hansen, P., Mladenovic, N., Brimberg, J., & Moreno Perez, J. A. (2010), International Se- approach to the truck scheduling problem with multiple docks and time windows.
ries in Operations Research & Management Science (2nd, pp. 61–86). Springer Sci- Computers & Industrial Engineering, 66(4), 818–826.
ence+Business Media, LLC. Wu, A. S., Yu, H., Jin, S., Lin, K.-C., & Schiavone, G. (2004). An incremental genetic al-
Jin, Y. (2005). A comprehensive survey of fitness approximation in evolutionary com- gorithm approach to multiprocessor scheduling. IEEE Transactions on Parallel and
putation. Soft Computing, 9(1), 3–12. Distributed Systems, 15(9), 824–834.