Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Keywords: A main issue in supply chain management is coordinating production and distribution decisions. To
Supply chain coordination achieve effective logistics scheduling, it is critical to integrate these two functions and plan them in a
Air transportation coordinated way. The problem is to determine both production schedule and air transportation allocation
Single machine scheduling of orders to optimize customer service at minimum total cost. In order to solve the given problem, two
Genetic algorithm
genetic algorithm (GA) approaches are developed. However, the effectiveness of most metaheuristic algo-
Taguchi experimental design
rithms is significantly depends on the correct choice of parameters. Hence, a Taguchi experimental design
method is applied to set and estimate the proper values of GAs parameters to improve their performance.
For the purpose of performance evaluation of proposed algorithms, various problem sizes are utilized and
the computational results of GAs are compared with each other. Moreover, we investigate the impacts of
the rise in the problem size on the performance of our algorithms.
Ó 2010 Elsevier Ltd. All rights reserved.
0957-4174/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2010.05.060
8256 M. Rostamian Delavar et al. / Expert Systems with Applications 37 (2010) 8255–8266
Li et al. (2005, 2006a, 2006b, 2007, 2008, 2004) and Zandieh and
2. Mathematical model and descriptions Molla-Alizadeh-Zavardehi (2008) used two type capacities in each
ordinary flight with two different transportation cost. For exten-
In this section, we provide a mathematical programming model sion of their work and generating more realistic schedule, similar
for integrated production and air transportation scheduling prob- to Zandieh and Molla-Alizadeh-Zavardehi (2009) we assumed that
lem. The aim is to determine an optimal allocation of orders to in many industries, we may have only one or even more than two
the existing transportation capacities and also specify sequence type capacities in each flight. Therefore, we considered T type
and completion times of these orders in production in such a capacities by Captf in the notation. However, if we have h (h < T)
way that the total cost of supply chain is minimized. The model type capacities for an ordinary flight f, then the Captf, i.e.
is based on the model developed by Zandieh and Molla-Alizadeh- t = h + 1, . . . , T will be fixed to zero. In addition, considering charter
Zavardehi (2009) for synchronized production and air transporta- flights and delivery tardiness, the orders can be transported by
tion scheduling problem. The assumptions used in this problem charter flights. With respect to the problem defined above, a math-
are: ematical programming model is formulated as follows:
!
1. The plant is treated as a single machine. X
T X
N X
F
min 0; ci Df LN 1
min 1
Tctf qtif
2. No idle time is allowed. t¼1 i¼1 f ¼1
c i D f LN
3. There are multiple flights in the planning period with different
þ a0i ðDf ci Þ qtif þ ai maxð0; di Af Þ qtif
transportation specifications such as cost, capacity, etc. !!
4. Business processing time and cost, together with loading time
and loading cost for each flight are included in the transporta-
þ bi maxð0; Af di Þ qtif
tion time and transportation cost. 1
!!
5. Local transportation transfers products from the plant to the min 0; ci Df LN 0
þ 1 1
bi qtif
airport. Local transportation time is assumed to be included in ci Df LN
transportation time.
þ min a0i ; ai maxð0; MDi ci Þ qtif
6. Local transportation can transfer an order to the airport when
N
the order is produced completely. X
þ bi maxð0; ci MDi Þ qtif þ b0i qðTþ1Þi
7. Orders released into plant for the planning period are delivered i¼1
within the same planning period, which means there are no
þ minða0i ; ai Þ maxð0; MDi ci Þ qðTþ1Þi
production backlogs.
þ bi max ð0; ci MDi Þ qðTþ1Þi ð1Þ
The notations that will be used to describe the problem and
algorithm are as follows:
s:t:
!
Indices X
T
i, i0 0
order/job index, i, i = 1, 2, . . . , N qtif ðDesi desf Þ ¼ 0; i ¼ 1; . . . ; N; f ¼ 1; . . . ; F ð2Þ
t¼1
f, f0 ordinary flight index, f, f0 = 1, 2, . . . , F
k destination index, k = 1, 2, . . . , K X
N
qtif 6 Captf ; t ¼ 1; . . . ; T; f ¼ 1; . . . ; F ð3Þ
p, p0 position or sequence of order i, p, p0 = 1, . . . , N
i¼1
M. Rostamian Delavar et al. / Expert Systems with Applications 37 (2010) 8255–8266 8257
X
T X
F lue is assigned to each chromosome according to its performance.
qtif þ qðTþ1Þi ¼ Q i ; i ¼ 1; . . . ; N ð4Þ The more desirable the chromosome, the larger the fitness value.
t¼1 f ¼1
The population evolves by a set of operators until some stopping
X
N
criterion is met. A typical iteration of a GA, a generation, proceeds
uip ¼ 1; i ¼ 1; . . . ; N ð5Þ
p¼1
as follows. The best chromosomes of the current population are
copied directly to the next generation (reproduction). A selection
X
N
uip ¼ 1; p ¼ 1; . . . ; N ð6Þ mechanism chooses chromosomes of the current population in
i¼1 such a way that the chromosome with the higher fitness value
!! has a greater probability of being selected (roulette wheel). The se-
X
N X
p1 X
N
uip pi Q i þ ui0 p0 Pi0 Q i0 ¼ ci ; i ¼ 1; . . . ; N ð7Þ lected chromosomes mate and generate new offspring (crossover).
p¼1 p0 ¼1 i0 ¼1 After the mating process, each offspring might mutate by another
uip 2 f0; 1g; i ¼ 1; . . . ; N; p ¼ 1; . . . ; N ð8Þ mechanism called mutation. The new population is then evaluated
qtif ¼ Non-negative integer variable ð9Þ again and the whole process is repeated (Goldberg, 1989).
We design two kinds of genetic approaches. The first approach
qðTþ1Þi ¼ Non-negative integer variable ð10Þ
(GA1) is similar to the other previous researches in the synchroni-
zation of air transportation and production scheduling. In these re-
The model considers both transportation allocation and production
searches, considering the mathematical model, at first a
scheduling (orders allocations to ordinary flights and charter flights,
transportation allocation is obtained and then with having that
determination of sequence and also completion time of orders). Due
allocation in mind, a heuristic or a metaheuristic is applied to
to the completion times, some orders might not reach to their allo-
achieve a production sequence and their completion times. In this
cated flights and must be transported by charter flights. Thereupon,
approach, at first a GA is used to solve the transportation problem
the objective function includes three parts. The first part is about
and find a solution with minimum transportation cost and then an-
the allocated orders which can reach to their ordinary flights and in-
other GA is applied to the achieved solution to generate the pro-
cludes transportation cost of ordinary flights, departure time earli-
duction sequence. Stopping criterion is set to be the CPU time
ness penalties, and total delivery earliness tardiness penalties. The
(CT) which is dedicated to the entire algorithm. To obtain the stop-
second part describes the allocated orders which cannot reach to
ping criterion for each of the transportation allocation and the pro-
their ordinary flights. Finally, the allocated charter flights are pre-
duction scheduling, the time is divided as follows:
sented in the last part. Since these two last parts should be trans-
ported by charter flights, their costs are the same. These two parts
CPU time for transportation’s GA (TCT) = CPU time time per-
include transportation cost of charter flights, minimum of departure
centage of transportation (TPT).
time and delivery earliness penalties, and delivery tardiness penal-
CPU time for production’s GA (PCT) = CPU time time percent-
ties. Those orders which are supposed to be transported by charter
age of production = CPU time (1 TPT).
flights and their completion time is less than their MD, if a0i 6 ai ,
departure time of charter flight i will be MDi and if a0i P ai , the
Other input parameters in the GA1 algorithm are population
departure time will be ci. This happens because the departure time
size of transportation (PST), population size of production (PSP),
of charter flights are arbitrary. So we consider minða0i ; ai Þ as the ear-
reproduction percentage of transportation (TPr), reproduction per-
liness penalty. Constraint sets (2) ensures that if order i and ordin-
centage of production (PPr), mutation probability of transportation
ary flight f have different destinations, order i cannot be allocated to
(TPm), mutation probability of production (PPm). The algorithm is
ordinary flight f. Constraint sets (3) states that the capacity 1 to T of
depicted in Fig. 1.
ordinary flight f is not exceeded. Constraint set (4) ensures that or-
Since the transportation cost is just a part of the total cost, hav-
der i is completely allocated. Constraint sets (5) and (6) state that
ing a minimum transportation cost does not ensure that the
each job has to be assigned to a position, and each position has to
respective solution has the minimum total cost. So in the second
be covered by a job. Constraint set (7) calculates completion time
approach (GA2), we perform a production schedule’s GA to some
of jobs.
more transportation allocations. To more clarify, let us consider a
simple example of two orders (n = 2) and two flights (F = 2) with
3. Proposed genetic algorithms a sufficient capacity, same destination and different departure
times. The orders essential information is given in Tables 1 and
The use of conventional tools for solving the mathematical pro- 2, and in Fig. 2.
gramming models problem is limited due to the complexity of the
problem and the large number of variables and constraints, partic- Solution I: X12 = 10, X22 = 10.
ularly for realistically sized problems (Li, Hendry, & Teunter, 2007). Transportation cost = (10 + 10) 18 = 360.
Regarded as the time complexity function and also a class of com- Production earliness penalty = (12 2) 2 + (12 5) 2 = 34.
binational optimization problems known as nondeterministic Delivery earliness penalty = 0.
polynomial-time hard (NP-hard), two genetic algorithms are ap- Delivery tardiness penalty = (13 8) 8 + (13 9) 8 = 72.
plied to solve the problem. Total cost = 360 + 34 + 0 + 72 = 466.
The basic concepts of GAs, introduced by Holland (1975), have Solution II: X11 = 10, X21 = 10.
been successfully applied to solve many combinatorial optimiza- Transportation cost = (10 + 10) 20 = 400.
tion problems. The Genetic algorithms are search algorithms based Production earliness penalty = (6 2) 2 + (6 5) 2 = 8.
on Darwin’s ‘survival of the fittest’ concept, meaning that good par- Delivery earliness penalty = (8 7) 3 + (9 7) 3 = 9.
ents produce better offspring. Nowadays, GAs are considered to be Delivery tardiness penalty = 0.
one of the typical metaheuristic approaches for tackling both dis- Total cost = 400 + 8 + 9 = 417.
crete and continuous optimization problems. GAs have been
shown as a robust optimization technique to solve many real world As shown above, transportation cost in the first solution is
problems (Gen & Cheng, 1997, 2000). much less than the second one. But since other types of costs (pro-
The GA searches a problem space with a population of chromo- duction earliness penalty and delivery earliness tardiness penal-
somes, each of which represents an encoded solution. A fitness va- ties) in the latter are less than the equivalent costs in the former,
8258 M. Rostamian Delavar et al. / Expert Systems with Applications 37 (2010) 8255–8266
Start
Production’s Mutation
NO Output the best transportation allocation
Evaluate fitness of new population (Cost of Transportation Allocation)
Transportation’s Crossover
End
Transportation’s Mutation
Fig. 1. Proposed GA1’s flowchart. TCT: transportation CPU time, PCT: production CPU time.
Table 1 structural difference of GA1 and GA2, the stopping criterion of each
Orders essential parameters. production sequence is set to be the number of iterations. By
i Qi pi di a0i ai bi increasing the problem’s size, due to larger solution space, more
iteration should be devoted to the production sequence algorithm.
1 10 2 8 2 3 8
2 10 3 9 2 3 8 Hence, the number of iterations in the production sequence algo-
rithm of GA2 is set to be ‘‘constant coefficient n”. The constant
coefficient is tuned in the parameter tuning section. This total pro-
cedure continues until the termination time is over. In the GA2, the
Table 2
Flights essential parameters. stopping criterion of the production sequence algorithm is differ-
ent from the GA1, which is the number of iterations instead of time
f Df Af Cf
percentage of production. The other input parameters are the same
1 6 7 20 in both GAs.
2 12 13 18
In the following subsections we describe all parameters and
operators used in the proposed GAs.
it has lower total cost, so it is better. Thus, it is obvious that a solu- 3.1. Encoding scheme
tion with minimum transportation cost may not have the lowest
total cost. Since orders and flights have different destinations, allocating
In the second approach (GA2), as shown in Fig. 3, only if a better an order to each available ordinary flight is not possible; so the or-
transportation solution is obtained at the end of each generation of ders should only be allocated to the flights with the same destina-
transportation allocation’s GA, the production schedule’s GA is ac- tions. For instance, in Fig. 4, the order 1 which is for destination 3,
quired to achieve the production sequence. Similar to GA1, CPU cannot be allocated to the flights 1, 2, 5, and 6, flying to destina-
time is considered as the stopping criterion. But considering the tions 1 and 2. Therefore the transportation matrix is modified:
D1 A1 d1 d2 D2 A2
1 2
1 2 3 4 5 6 7 8 9 10 11 12 13
Start
YES TBTA=BTA
Is ITGA=1
TBPS=BPS
Set parameters
NO
Transportation’s Mutation
NO
Is BTA better than
Evaluate fitness of new pop (Total cost) TBTA?
YES
Fig. 3. Proposed GA2’s flowchart. CT: CPU time, ITGA: iteration of transportation’s GA, BTA: best transportation allocation, BPS: best production sequence, TBTA: total best
transportation allocation, TBPS: total best production sequence.
5 5 5 5
1 4
3.2. Initial procedure
7 7 7
2 1
The significant role of the initial solution on the quality of the
3 3 3
final result of a search procedure has already been recognized 2 6
and emphasized by many researchers in recent years (Jungwatta- 4 4 4
nakita, Reodechaa, Chaovalitwongsea, & Werner, 2009; M’Hallah, 3 2
2007; Naderi, Zandieh, Ghoshe Balagh, & Roshanaei, 2009). What 7 7 7
3 5
has been utilized so far by the majority of researchers to generate
the initial solution for their algorithms has been the random gen-
Fig. 5. Representation encoding for transportation allocation.
Desi 3 1 2 3 2 1 1 2 3 1
eration of the initial solution, which has led them to poor quality
Desf
i 1 2 3 4 5 6 7 8 9 10 solutions. Therefore, in order to acquire a satisfactory level of solu-
f tion quality for such a hard combinatorial problem, meticulous
7 7 7 7 7 7 7 7 7 7
2 1 considerations should be given to the intelligent selection of their
4 4 4 4 4 4 4 4 4 4 initialization procedure. In this work, several types of initial proce-
3 2
dures have been used. In the following subsections we describe the
1 3
9 9 9 9 9 9 9 9 9 9 initial procedures which used for each of air transportation and
production scheduling.
5 5 5 5 5 5 5 5 5 5
1 4
Step 1: If the transportation problem is unbalanced, it is chan- Quantity 200 100 300 300 900
ged into a balanced one by introducing a dummy flight or a
dummy order. Fig. 7. Balanced and allocated transportation matrix.
Step 2: Find the row with minimum cost. If there are more than
one row with minimum cost, select one of them randomly.
Step 3: Randomly select a column. DD: 3 5 4 2 6 1
Step 4: Allocate the maximum possible value which is equal to EMD: 5 3 4 2 1 6
minimum value of flight’s capacity and order’s quantity to the SPT: 2 5 3 1 6 4
crossing position (cell) of selected row and column. LPT: 4 6 1 3 5 2
Step 5: Subtract the allocated value from both flight’s capacity
(selected row) and order’s quantity (selected column).
Step 6: If the capacity of a flight or the quantity of an order
becomes zero, do not consider that row or column in the next 3.3. Selection mechanism
steps, and when both of them become zero spontaneously,
eliminate both row and column from the remaining In the model, we want to minimize the objective function. Be-
considerations. cause in Roulette-Wheel we give more chance to the solution
Step 7: Iterate steps 2–6, until sum of updated flight’s capacity which has greater fitness value, we consider the fitness value as
and order’s quantity becomes zero. follow:
Fitness Value ¼ 1=Objective Function
In order to understand this algorithm better, we design the fol-
lowing simple example with three flights and three orders, in Considering inverse objective function as a fitness value, the greater
which a dummy order is added to balance the allocation matrix fitness value a solution has, the more chance it has to be selected.
(see Figs. 6 and 7).
Since in Northwest Corner Method, cost does not have any roll 3.4. GA operators
in generating the initial solution, so it is directly applied in our
solution algorithm. But this method produces just one initial solu- 3.4.1. Reproduction
tion and to complete the initial generation the mutation operator With more probability, better parents can generate better off-
(explained in Section 3.4.3) is used to generate different solutions, spring, so it would be necessary to transfer the best solutions of
up to the number of specified population size. each generation to the next one. Therefore chromosomes with
higher fitness values are more desirable, so the pr% of the chromo-
somes with the greater fitness values are automatically copied to
3.2.2. Initial procedures used in production scheduling the next generation (elite strategy).
We use five initial procedures in our work; SPT, LPT, EDD, EMD,
and also generating randomly. In SPT and LPT, a table is con- 3.4.2. Crossover
structed to list the jobs which are arranged in shortest and longest Crossover operates on two chromosomes at a time and gener-
processing time order, respectively. In EDD, jobs are sorted on the ates offspring by combining both chromosomes’ features. The pur-
basis of earliest due date. In EMD, arranging the jobs is based on pose is to generate ‘better’ offspring, i.e. to create better sequences
maximum departure time, and jobs are listed in descending way. after combining the parents. As we assigned pr% of the chromo-
Because each order’s destination is different, and has a different somes of generation to reproduction, the (1 pr)% remaining chro-
transportation time, arranging with EMD method is not similar mosomes are generated through crossover operator.
to EDD. For more details we consider the following example: Because our proposed model, considers two sections; air trans-
Assume that we have six orders, each of which have different portation and production scheduling, we specified a crossover
due dates and transportation times (Table 3). As it is obvious, operator for each one:
arranging these orders with EMD and EDD is different.
3.4.2.1. Crossover operator used in air transportation. We explain the
Orders algorithm of the operator presented by Vignaux and Michalewicz
1 2 3 Capacity
Flights (1991) in a nutshell. For more details the reader can see Gen and
10 10 10
1 400
Table 3
5 5 5
2 300 Essential parameters for the example with six orders.
Orders 1 2 3 4 5 6
10 10 10
3 200 Processing times 3 1 2 4 1 3
Due dates 18 15 10 13 11 17
Quantity 200 100 300
Transportation times 4 2 3 1 5 1
Maximum departure times 14 13 7 12 6 16
Fig. 6. Unbalanced transportation allocation matrix.
M. Rostamian Delavar et al. / Expert Systems with Applications 37 (2010) 8255–8266 8261
Step 3: Divide matrix R into two matrices, R1 and R2 such that: tioned earlier about crossover operator, we need to use two
mutation operators.
R ¼ R1 þ R2
X n Xn
1X n
3.4.3.1. Mutation operator used in air transportation. Assigning a
r 1ij ¼ r 2ij ¼ r ij ; i ¼ 1; 2; . . . ; F predefined number for both rows (p) and columns (q) in each de-
j¼1 j¼1
2 j¼1
signed problem, we select as many rows and columns as the spec-
X
F X
F
1X F
ified numbers randomly. According to Vignaux and Michalewicz
r 1ij ¼ r 2ij ¼ r ij ; j ¼ 1; 2; . . . ; n
i¼1 i¼1
2 i¼1 (1991) the following algorithm is presented:
Step 4: Then generate two offspring, X 01 and X 02 , as follows: Step 1: Make a submatrix from a parent matrix using prede-
fined numbers of p (2 < p < F 1) and q (2 < q < n 1). Ran-
X 01 ¼ D þ R1
domly select {i1, . . . , ip} rows and {j1, . . . , jq} columns to create a
X 02 ¼ D þ R2 (p q) submatrix.
Step 2: Copy the value of the element in the crossing position of
In order to elaborate the upcoming algorithm obviously, we design
selected row i and column j from the parent matrix to the same
the following example as shown in Fig. 8:
position in the new created submatrix.
Step 3: Reallocate the values of the elements of the submatrix
randomly in such a way that the sum of each row and column
3.4.2.2. Crossover operator used in production scheduling. In this
remain identical to the one before reallocation.
operator, one point is randomly selected to divide both selected
Step 4: Replace the respective elements of the parent matrix
parents into two separate parts. If n is the number of genes in a
with new elements from the reallocated submatrix.
chromosome, there are (n 1) crossover points. One of these
points is selected with equal probability. The jobs on one side (each
In Fig. 10, we specified two random rows (3 and 4) and columns
side is chosen with the same probability) of a parent are inherited
(1 and 3) to generate submatrix.
by the offspring, and the other side’s jobs are placed in the order
they appeared in the other parent (Murata & Ishibuchi, 1994). After
changing the roles of parents, the same procedure is applied to pro- 3.4.3.2. Mutation operator used in production scheduling. We choose
duce the second offspring. This crossover is illustrated in Fig. 9. ‘‘swap” operator for this section. In this operator, two genes are se-
lected randomly and are swapped. The example of this operator is
shown in Fig. 11.
3.4.3. Mutation
The mutation operator is used to rearrange the structure of a
chromosome and to slightly change the sequence, i.e. generating 4. Experimental design
a new but similar sequence. The main purpose of applying muta-
tion is to avoid convergence to a local optimum and diversify the 4.1. Taguchi parameter design
population. The mutation operator can also be considered as a sim-
ple form of a local search. The probability of mutating an offspring The effectiveness of genetic algorithms greatly depends on the
is called the probability of mutation, pm, which is usually a small correct choice of parameters (Naderi, Zandieh, & Fatemi Ghomi,
number. After generating an offspring by a crossover operator, a 2008). In this section, we study the behavior of the different
random number from uniform [0, 1] is dedicated to the offspring. parameters of the proposed GAs. The full factorial design, which
If this random number was less than or equal to pm, then the muta- tests all possible combinations of factors, is the method widely
tion operator would be performed on that offspring. As we men- used in the most researches (Al-Aomar & Al-Okaily, 2006; Kim,
X1 R1 X'1
8 1 0 0 1 0 0 1 5 0 1 3
0 7 1 0 D R 0 1 0 0 0 6 1 1
0 0 7 0 4 0 1 2 1 1 1 1 0 0 1 0 1 0 6 0
0 0 0 9 0 5 1 1 0 1 0 1 0 1 0 0 2 3 0 5
X2 1 0 5 0 0 1 1 0 R2 X'2
1 0 3 5 2 1 0 5 1 1 0 0 0 1 1 0 4 1 2 2
0 4 1 3 0 0 0 1 0 5 1 2
2 1 4 0 0 1 0 0 1 1 5 0
5 3 0 1 1 0 0 0 3 1 0 5
Table 4
5 2 6 1 3 4 Factors and their levels.
Table 6
Experimental design used in random problems generation.
the GAs were based on the L27 orthogonal array, therefore 27 dif- the mean RPD is calculated for each trial. To do according Taguchi
ferent combinations of control factors were considered. Due to parameter design instructions, these mean RPDs, are transformed
having stochastic nature of GAs, three replications were performed to S/N ratios. The S/N ratios of trials are averaged in each level
for each trial to achieve the more reliable results. We implement and the value is shown in Fig. 12. As shown in Fig. 12, in GA1, best
the algorithms in C++, and run on a PC with 2.0 GHz Intel Core 2 parameters of factors A, B, E, G, and H are obviously 2, 2, 3, 3, and 3,
Duo and 1 GB of RAM memory. Because the scale of objective func- respectively. But, determining best parameters of other factors
tions in each instance is different, they could not be used directly. needs more investigations. Thus, in addition to S/N ratios, another
To solve this problem, the relative percentage deviation (RPD) is measurement, the RPD, is used. The results of RPD for each
used for each instance. parameter level are demonstrated in Fig. 13. As can be seen in
Fig. 13, the RPD illustrates the best parameters of factors C, D, F
Algsol Minsol and I as 2, 2, 1, and 3, respectively which confirms the same results
RPD ¼ 100
Minsol as S/N ratios.
where Algsol and Minsol are the obtained objective value for each Similar to GA1, considering Figs. 14 and 15, all the best param-
replication of trial in a given instance and the obtained best solu- eters for GA2 are defined as 2, 2, 2, 3, 2, 2, 2, 1 and 3, respectively,
tion, respectively. After converting the objective values to RPDs, according to their alphabetical order.
-14.5
-15
-15.5
-16
S/N
-16.5
-17
-17.5
-18
A B C D E F G H I
Fig. 12. Mean S/N ratio plot for each level of the factors in GA1.
8.1
7.7
7.3
6.9
RPD
6.5
6.1
5.7
5.3
A B C D E F G H I
Fig. 13. Mean RPD plot for each level of the factors in GA1.
-14.6
-14.9
-15.2
-15.5
S/N
-15.8
-16.1
-16.4
-16.7
A B C D E F G H I
Fig. 14. Mean S/N ratio plot for each level of the factors in GA2.
M. Rostamian Delavar et al. / Expert Systems with Applications 37 (2010) 8255–8266 8265
6.8
6.6
6.4
6.2
RPD
5.8
5.6
5.4
A B C D E F G H I
Fig. 15. Mean RPD plot for each level of the factors in GA2.
Table 7
Average relative percentage deviation (RPD) for the algorithms.
Algorithms 20j4f2d 30j6f2d 40j8f3d 50j10f3d 60j12f3d 70j14f4d 80j16f4d 90j18f4d 100j20f5d Average
GA1 5.15 4.07 4.25 4.87 4.93 6.74 5.56 6.92 6.26 5.41
GA2 1.25 1.34 0.86 0.6 0.43 0.72 0.44 0.62 0.33 0.73
8
7
6
5 GA1
RPD
4 GA2
3
2
1
0
20j4f2d
30j6f2d
40j8f3d
50j10f3d
60j12f3d
70j14f4d
80j16f4d
90j18f4d
100j20f5d
Fig. 17. Means plot for the interaction between each algorithm and problem size.
Fig. 16. Means plot and LSD intervals for the algorithms.
orthogonal array). Thus, total number of experiments with two ge- Li, H., Hendry, L., & Teunter, R. (2007). A strategic capacity allocation model for a
complex supply chain: Formulation and solution approach comparison.
netic algorithms, nine different problem sizes and three replica-
International Journal of Production Economics. doi:10.1016/j.ijpe.2007.02.033.
tions, i.e. 2 9 3 3^8 5^1 = 1771470, was reduced to 2 Li, K. P., Ganesan, V. K., & Sivakumar, A. I. (2005). Synchronized scheduling of
9 3 27 = 1458 experiments. Setting the parameters on the opti- assembly and multi-destination air transportation in a consumer electronics
mal values, the proposed genetic algorithms are compared to- supply chain. International Journal of Production Research, 43, 2671–2685.
Li, K. P., Ganesan, V. K., & Sivakumar, A. I. (2006a). Scheduling of single stage
gether. As a result, GA2 generates better solutions than GA1. We assembly with air transportation in a consumer electronic supply chain.
examine the behaviors of the algorithms within the several prob- Computers and Industrial Engineering, 51, 264–278.
lem sizes, and the results have shown that when the problem size Li, K. P., Ganesan, V. K., & Sivakumar, A. I. (2006b). Methodologies for synchronized
scheduling of assembly and air transportation in a consumer electronic supply
increases, GA2 represents better results and also keeps its robust chain. International Journal of Logistic System Management, 2, 52–67.
performance. Li, K. P., Sivakumar, A. I., Fu, Q., & Jin, X. (2007). A case study for synchronized
There are potentially unlimited opportunities for research in scheduling and manufacturing and air transportation in consumer electronics
supply chain. In IEEE conference on industrial engineering and engineering
coordinated production and air transportation scheduling. For fu- management (IEEM) Singapore (pp. 1629–1633).
ture researches, it is possible to investigate and develop new algo- Li, K. P., Sivakumar, A. I., & Ganesan, V. K. (2008). Complexities and algorithms for
rithms based on other metaheuristics such as Tabu Search, synchronized scheduling of parallel machine assembly and air transportation in
consumer electronic supply chain. European Journal of Operational Research, 187,
Simulated Annealing, Variable Neighborhood Search, Scatter 442–455.
Search, and Ant Colonies. Furthermore, we may take into account Li, K. P., Sivakumar, A. I., Mathirajan, M., & Ganesan, V. K. (2004). Solution
more realistic aspects of the problem such as dynamic environ- methodology for synchronizing assembly manufacturing and air transportation
of consumer electronics supply chain. International Journal of Business, 9,
ment and stochastic demands.
361–380.
M’Hallah, Rym (2007). Minimizing total earliness and tardiness on a single machine
using a hybrid heuristic. Computers and Operations Research, 34, 3126–3142.
References Michalewicz, Z., Vignaux, G. A., & Hobbs, M. (1991). A nonstandard genetic
algorithm for the nonlinear transportation problem. Informs Journal on
Al-Aomar, R. (2006). Incorporating robustness into genetic algorithm search of Computing, 3, 307–316.
stochastic simulation outputs. Simulation Modelling Practice and Theory, 14, Murata, T., & Ishibuchi, H. (1994). Performance evaluation of genetic algorithms for
201–223. flowshop scheduling problems. In International conference on evolutionary
Al-Aomar, R., & Al-Okaily, A. (2006). A GA-based parameter design for single computation (pp. 812–817).
machine turning process with high-volume production. Computers and Naderi, B., Zandieh, M., & Fatemi Ghomi, S. M. T. (2008). Scheduling sequence-
Industrial Engineering, 50, 317–337. dependent setup time job shops with preventive maintenance. International
Chen, Z. L. (1996). Scheduling and common due date assignment with earliness– Journal of Advanced Manufacturing Technology. doi:10.1007/s00170-008-1693-0.
tardiness penalties and batch delivery costs. European Journal of Operational Naderi, B., Zandieh, M., Ghoshe Balagh, A. K., & Roshanaei, V. (2009). An improved
Research, 93, 49–60. simulated annealing for hybrid flowshops with sequence-dependent setup and
Erenguc, S. S., Simpson, N. C., & Vakharia, A. J. (1999). Integrated production/ transportation times to minimize total completion time and total tardiness.
distribution planning in supply chain: An invited review. European Journal of Expert Systems with Applications, 36, 9625–9633.
Operational Research, 115, 219–236. Park, S. H. (1995). Robust design and analysis for quality engineering. London:
Gen, M., & Cheng, R. (1997). Genetic algorithms and engineering design. NY: John Chapman & Hall.
Wiley and Sons. Phadke, M. S. (1989). Quality engineering using robust design. NJ: Prentice-Hall.
Gen, M., & Cheng, R. (2000). Genetic algorithms and engineering optimization. NY: Sarmiento, A. M., & Nagi, R. (1999). A review of integrated analysis of production–
John Wiley and Sons. distribution systems. IIE Transaction, 31, 1061–1074.
Goetschalckx, M., Vidal, C. J., & Dogan, K. (2002). Modeling and design of global Taguchi, G. (1986). Introduction to quality engineering. White Plains: Asian
logistics systems: A review of integrated strategic and tactical models and Productivity Organization/UNIPUB.
design algorithms. European Journal of Operational Research, 143, 1–18. Tsai, J. T., Ho, W. H., Liu, T. K., & Chou, J. H. (2007). Improved immune algorithm for
Goldberg, D. E. (1989). Genetic algorithms in search, optimization and machine global numerical optimization and job shop scheduling problems. Applied
learning. Reading, MA: Addison-Wesley. Mathematics and Computation, 194, 406–424.
Holland, J. (1975). Adaptation in natural and artificial systems. Ann Arbor: University Vidal, C., & Goetschalckx, M. (1997). Strategic production–distribution models: A
of Michigan Press. critical review with emphasis on global supply chain models. European Journal
Jung, H., & Jeong, B. (2005). Decentralised production–distribution planning system of Operational Research, 98, 1–18.
using collaborative agents in supply chain network. International Journal of Vignaux, G. A., & Michalewicz, Z. (1991). A genetic algorithm for the linear
Advanced Manufacturing Technology, 25, 167–173. transportation problem. IEEE Transactions on Systems Man and Cybernetics, 21,
Jungwattanakita, J., Reodechaa, M., Chaovalitwongsea, P., & Werner, F. (2009). A 445–452.
comparison of scheduling algorithms for flexible flow shop problems with Yang, K., & El-Haik, B. S. (2003). Design for six sigma; a roadmap for product
unrelated parallel machines, setup times, and dual criteria. Computers and development. NY: McGraw-Hill.
Operations Research, 36, 358–378. Zandieh, M., & Molla-Alizadeh-Zavardehi, S. (2008). Synchronized production and
Kim, S. J., Kim, K. S., & Jang, H. (2003). Optimization of manufacturing parameters for distribution scheduling with due window. Journal of Applied Sciences, 8,
a brake lining using Taguchi method. Journal of Material Processing Technology, 2752–2757.
136, 202–208. Zandieh, M., & Molla-Alizadeh-Zavardehi, S. (2009). Synchronizing production and
Kolischa, R., & Padman, R. (2001). An integrated survey of deterministic project air transportation scheduling using mathematical programming models. Journal
scheduling. Omega, 29, 249–272. of Computational and Applied Mathematics. doi:10.1016/j.cam.2008.12.022.