European Journal of Operational Research 205 (2010) 98–105

Contents lists available at ScienceDirect

European Journal of Operational Research

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

Production, Manufacturing and Logistics

An efficient tabu algorithm for the single row facility layout problem
Hamed Samarghandi *, Kourosh Eshghi
Sharif University of Technology, Tehran, Iran

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

Article history: The general goal of the facility layout problem is to arrange a given number of facilities to minimize the
Received 25 September 2008 total cost associated with the known or projected interactions between them. One of the special classes of
Accepted 29 November 2009 the facility layout problem is the Single Row Facility Layout Problem (SRFLP), which consists of finding an
Available online 4 December 2009
optimal linear placement of rectangular facilities with varying dimensions on a straight line. This paper
first presents and proves a theorem to find the optimal solution of a special case of SRFLP. The results
Keywords: obtained by this theorem prove to be very useful in reducing the computational efforts when a new algo-
Facilities planning and design
rithm based on tabu search for the SRFLP is proposed in this paper. Computational results of the proposed
Linear ordering problem
Tabu search
algorithm on benchmark problems show the greater efficiency of the algorithm compared to the other
Integer programming heuristics for solving the SRFLP.
Ó 2009 Elsevier B.V. All rights reserved.

1. Introduction the work of Heragu and Kusiak (1991), Amaral (2006, 2009), Anjos
and Vannelli (2008).
Generally, Facility Layout Problem (FLP) is concerned with the The paper structure is summarized as follows. Section 2 consists
arrangement of a given number of facilities in order to minimize of a literature review of the related research in this area. Section 3
the total cost associated with the known or projected interactions focuses on a special case of the SRFLP and proves a theorem to ob-
between them. Applications of FLP occur in many real environments tain its optimal solution. Section 4 develops a new tabu algorithm
such as service center layout, hospital layout or library layout. that explores near-optimal solutions of the SRFLP by using initial
The Single Row Facility Layout Problem (SRFLP) is categorized solutions generated by the proven theorem. Details of the algo-
as a particular case of FLP. This problem consists of a number of rithm are presented in Section 5. In Section 6 computational results
rectangular facilities to be sorted on a straight line. SRFLP is de- are given. Finally, Section 7 contains conclusions.
fined as follows: n rectangular departments are to be arranged
on a straight line in a given direction. The length li of each depart-
ment i, and a n  n matrix C ¼ ½cij  where cij is the flow between 2. Literature review
departments i and j, are also given. The distance between two
departments is calculated as the distance between their centroids. The theoretical attractiveness and practical applications of the
The objective of the problem is to arrange the departments so as to problem have created a rich and growing literature for the SRFLP,
minimize the weighted sum of the distances between all depart- which has proven to be a NP-complete problem (Suresh and Sahu,
ment pairs. 1993). Since SRFLP is NP-complete, the application of exact meth-
The SRFLP has numerous practical applications including the ods to large instances of the problem is too time consuming; there-
following: the arrangement of departments on one side of a corri- fore, heuristic methods have been developed to obtain a near-
dor in an office building; a hospital or a supermarket (Simmons, optimal solution of the problem.
1969); the arrangement of books on a shelf in a library; the assign- Neghabat (1974) presented a procedure in which a complete
ment of disk cylinders to files (Picard and Queyranne, 1981); and solution is attained by adding one machine at a time to the end
the arrangement of machines on one side of a straight line path of the current solution. Drezner (1987) developed a heuristic meth-
traveled by an automated guided vehicle (Heragu and Kusiak, od based on the eigenvectors of a transformed flow matrix. An-
1988; Solimanpur et al., 2005). Several exact methods have been other heuristic has been proposed by Heragu and Kusiak (1988)
proposed for the SRFLP; the reader is referred, for example, to in which each iteration adds a facility to the right or to the left
of the developing layout.
Heragu and Kusiak (1991) introduced a linear mixed-integer
* Corresponding author. Address: Department of Mechanical and Manufacturing
Engineering, University of Manitoba, 75A Chancellors Circle, Winnipeg, Manitoba,
formulation of the SRFLP and solved it using a penalty technique.
Canada R3T 5V6. Tel.: +1 204 899 1362. A simulated annealing algorithm was proposed by Heragu and Alfa
E-mail address: [email protected] (H. Samarghandi). (1992) to solve the SRFLP. Braglia (1996) developed a combination

0377-2217/$ - see front matter Ó 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2009.11.034
 H. Samarghandi, K. Eshghi / European Journal of Operational Research 205 (2010) 98–105 99

of simulated annealing and genetic algorithms. In this method, it is dij

assumed that the machines have equal size and the locations are
predefined. A constructive greedy heuristic was developed by Ku-
mar et al. (1995). This heuristic does not consider the dimensions j
i
of the equipment in the construction of the single row layout. Dij
Ficko et al. (2004) introduced a genetic algorithm. Solimanpur li lj
et al. (2005) formulated the SRFLP as a non-linear 0–1 program- 2 2
ming model, which was solved by an ant algorithm. Anjos et al.
Fig. 1. Distance between facilities i and j.
(2005) constructed a semi-definite programming relaxation pro-
viding a lower bound on the optimal value of the SRFLP. Recently
proposed meta-heuristic methods include a scatter search algo-
The special case occurs when 8i; j : cij ¼ c which means that the
rithm (Kumar et al., 2008), a hybrid ACO/PSO algorithm (Teo and
flow between each pair of facilities is identical. Therefore, the
Ponnambalam, 2008), and genetic algorithm (Lin, 2009). Later in
objective function of the above model can be written as:
this paper, the performance of these algorithms will be compared
with the tabu algorithm proposed. A quick review of related re- X
n1 X
n

search reveals that the former attempts suffer from at least one min z ¼ c dij : ð5Þ
i¼1 j¼iþ1
of the following drawbacks:
The following theorem can be proved in this case.
(a) The dimensions of the machines either are not considered or
Theorem 1. Suppose that 8i; j : cij ¼ c; Sij ¼ 0. Now, sort the facilities
are assumed to be equal (Braglia, 1996).
in non-descending order such that the shortest facility is denoted by 1
(b) The locations of facilities are predetermined (Kumar et al.,
and the longest facility by n. Then, Fig. 2 shows the optimum solution
1995; Braglia, 1996).
when n is an odd number, and Fig. 3 shows the optimum solution
(c) The size of the machines is only considered in the physical
when n is an even number.
layout of the machines rather than in the computation pro-
cess. Therefore, in this approach, the LOP is assumed to be
Proof. Without loss of generality, consider a case where n is odd
completely equal to the SRFLP (Heragu and Kusiak, 1988).
and assume that its objective function value is z. Exchanging the
(d) The method requires too much time to construct a layout,
locations of the facilities i and j-facility i is in the left side of the
especially when applied to large instances of the SRFLP
facility j-results in another layout with different objective function,
(Anjos et al., 2005).
z0 . We assume that z0 < z. Since the flow between all pairs of facil-
ities is equal, the difference of z and z0 originates from the change
This paper presents a particular case of the SRFLP in order to
in the distance between the facilities i and j, and other facilities.
find its optimal solution theoretically. This optimal solution is used
This can be written as:
to trigger a tabu algorithm that aims to find near-optimal solutions
for the SRFLP. Tabu Search (TS) is a meta-heuristic approach that X
n
0 0
has been widely used in solving different combinatorial optimiza- z0 ¼ z þ ðdik þ djk  dik  djk Þ ð6Þ
k¼1
tion problems (Glover and Laguna, 1997). The proposed approach
0
aims to overcome the limitations of the existing methods as the in which dik is the distance between the facilities i and k in the new
proposed algorithm is very efficient and requires only a short time layout. If z0 < z, then:
to construct a good layout. X
n X
n
0 0
ðdik þ djk Þ > ðdik þ djk Þ: ð7Þ
k¼1 k¼1
3. Special case of the SRFLP
From Eq. (4) we know that:
The special case of the SRFLP, which is introduced and solved
li þ lk
theoretically in this section, makes it possible for the proposed dik ¼ þ Dik ;
2
algorithm to start its initial search from high-quality solutions in
lj þ lk
feasible space. As a result, the proposed algorithm takes less time djk ¼ þ Djk ;
2 ð8Þ
to reach better solutions in the feasible space compared to all other
0 li þ lk
algorithms. In 1991, Heragu and Kusiak presented the well-known dik ¼ þ D0ik ;
ABSMODEL. The ABSMODEL is as follows: 2
0 lj þ lk
X
n1 X
n djk ¼ þ D0jk ;
min z¼ cij dij ð1Þ 2
i¼1 j¼iþ1 Putting Eq. (8) in (7) results in:
1 X
n X
n
s:t: dij P ðli þ lj Þ þ sij ; i ¼ 1; 2 . . . ; n  1; j ¼ i þ 1; . . . ; n; ð2Þ
2 ðDik þ Djk Þ > ðD0ik þ D0jk Þ: ð9Þ
dij P 0; i ¼ 1; 2 . . . ; n  1; j ¼ i þ 1; . . . ; n: ð3Þ k¼1 k¼1

In this model, dij is the distance between the centers of the facil-
ities i and j where sij is the necessary gap between the two facilities,
and it is given as one of the parameters of the model. Heragu and
Kusiak defined dij as follows:
Fig. 2. Optimal layout when n is odd.
li þ lj
dij ¼ þ Dij ; ð4Þ
2
where Dij is the space between facilities i and j. Note that Dij is not
necessarily equal to sij . Fig. 1 depicts that Dij ¼ lk when facility k is
placed between facilities i and j with sij ¼ 0. Fig. 3. Optimal layout when n is even.
100 H. Samarghandi, K. Eshghi / European Journal of Operational Research 205 (2010) 98–105

... t i ... 1 ... j k ...

Set M Set B Set R

Fig. 4. Definitions of M, B and R.

According to Fig. 4, it is possible to partition all facilities in the Case 1:

layout to three distinct sets:
m ¼ r ) T ¼ 0 ) T ¥ 0; z0 ¥ z: ð17Þ
1. Set M: includes all facilities located at the left side of facility i in Case 2:
Fig. 2. Suppose that there are m facilities in set M.
m > r; lj P li ) T ¥ 0; z0 ¥ z: ð18Þ
2. Set B: includes all facilities located at the right side of facility i
and at the left side of facility j in Fig. 2. Suppose that there are b Case 3:
facilities in set B.
m > r; lj 6 li : ð19Þ
3. Set R: includes all facilities located at the right side of facility j
in Fig. 2. Suppose that there are r facilities in set R. Eq. (19) implies that in Fig. 2 the number of facilities on the left side
of facility i is more than the number of facilities on the right side of
Based on the above classification, Eq. (9) can be rewritten as: facility j. This means that facility i is closer to the center of the lay-
X X X out in Fig. 2 than facility j. Thus, according to the theorem’s assump-
ðDik þ Djk Þ þ ðDik þ Djk Þ þ ðDik þ Djk Þ tions, lj i li . However, if lj ¼ li , then T ¥ 0 and z0 ¥ z.
k2M k2B k2R Case 4:
X 0 X 0 X 0
> ðDik þ D0jk Þ þ ðDik þ D0jk Þ þ ðDik þ D0jk Þ: ð10Þ m < r; lj P li : ð20Þ
k2M k2B k2R
The application of the same rationale used in case 3, establishes the
Now, consider the layout in Fig. 2 and the new layout created by
desired result.
exchanging the locations of facilities i and j. For each facility k 2 M,
Case 5:
according to Fig. 5, we have:
m < r; lj 6 li ) T ¥ 0; z0 ¥ z ð21Þ
D0ik ¼ Djk þ lj  li ;
ð11Þ
D0jk ¼ Dik
The preceding five cases prove the optimality of the layout
and for each facility k 2 B, based on Fig. 6, the following relations introduced in Fig. 2. The optimality of the layout of Fig. 3 derives
results: from the same logic and procedure. h
D0ik ¼ Djk ;
ð12Þ
D0jk ¼ Dik :
4. Tabu search algorithm for the SRFLP
Finally, for each facility k 2 R, according to Fig. 7, we know that:
D0ik ¼ Djk ; Tabu Search (TS) is a meta-heuristic approach developed by
ð13Þ Glover (1986). TS moves from a current solution xt in the feasible
D0jk ¼ Dik  lj þ li :
space to the best solution in its neighborhood Nðxt Þ, say xtþ1 , at
By putting Eqs. (11)–(13) into (10), we have the following each iteration t. Since there is no guarantee that xtþ1 is a better
result:
X X
ðlj  li Þ þ ðli  lj Þ < 0: ð14Þ
k2M k2R

Considering the number of the members in sets M and R, Eq. (14)

can be simplified to:
T<0 ð15Þ
in which:
T ¼ ðm  rÞ  lj  ðm  rÞ  li : ð16Þ
Fig. 6. Distances regarding to facility k 2 B before and after the exchange of
The following five possibilities can be derived from Eqs. (15) facilities i and j.
and (16):

Fig. 5. Distances regarding to facility k 2 M before and after the exchange of Fig. 7. Distances regarding to facility k 2 R before and after the exchange of
facilities i and j. facilities and i and j.
 H. Samarghandi, K. Eshghi / European Journal of Operational Research 205 (2010) 98–105 101

solution than xt , a tabu mechanism is used in TS to prevent repeti- Finding a better solution in the neighborhood of an existing
tion of the same solutions. Although preventing a return to all pre- solution can be very time consuming or even impossible. Thus,
viously produced solutions is the simplest way to prevent the algorithm limits improving attempts to a times. After a futile
repetition, the memory required to save these solutions is exces- efforts, the algorithm gives up probing in the neighborhood of
sive. A smarter approach is to save some specifications of the pro- the current solution and selects a new solution from AM to ad-
duced solutions into the memory and banning the existence of the vance. a is a parameter defined by the user. The defined neighbor-
same specifications in newly produced solutions for the next h iter- hood structure is very simple yet effective and makes it possible for
ations of the algorithm. This method is called Short Term Memory the algorithm to search a large number of solutions quickly.
(STM).
TS also uses mechanisms such as diversification and intensifica- 4.3. Intensification, long-term memory, and termination criterion
tion. Diversification assures that the algorithm has searched vast
areas of the solution space before converging to the final solution; Intensification is the search in the neighborhood of the good
intensification forces the algorithm to search the neighborhood of quality solutions for constructing better solutions closer to opti-
the solutions with desirable specifications more comprehensively mum. For intensification phase, we use a technique called Adaptive
(Glover and Laguna, 1997). TS can also exploit an aspiration crite- Memory (AM). AM is probably the strongest approach for intensi-
rion. Being prior to STM, the aspiration criterion allows for the fication (Jean-Francois and Gilbert, 2005). AM was introduced by
acceptance of solutions that have certain specifications existing Rochat and Taillard (1995) and consists of the set of the best solu-
in STM. For more information regarding the TS sub-procedures, tions constructed during the search procedure. In other words, AM
the reader is referred to Glover and Laguna (1997). is a list capable of storing best L layouts of the search procedure
with their objective functions as labels. The algorithm uses the la-
bels to sort the solutions in ascending order. Therefore, the layout
4.1. Initial solutions
with the best objective function stands in the first position of the
list while the layout with the worst objective function remains in
The proposed TS algorithm needs a number of initial solutions
the last place. As it was mentioned in Section 4.1, the algorithm
to start exploration. Fortunately, finding feasible solutions for the
first constructs L initial solutions using Theorem 1 and sorts them
SRFLP is a straightforward routine because each permutation allo-
in the AM according to their objective functions as labels. In the
cating n rectangular facilities to a straight line is a feasible solution
following iterations, algorithm constructs new layouts, stores them
for the SRFLP. However, our algorithm does not use randomly gen-
into the AM, and triggers the sorting process again.
erated solutions since the quality of these solutions makes them
undesirable compared to the solutions produced by using Theorem
4.3.1. Selecting a layout from the AM
1, as we will see in Section 6. The use of these initial solutions pro-
The process of selecting a solution stored in the AM and search-
duced by Theorem 1 increases the efficiency of the proposed
ing for a better layout in its neighborhood fulfills diversification as
algorithm.
well as intensification. The selection process is done through a
It should be noted that Eqs. (17), (19) and (20) are also useful in
probabilistic procedure, tending to select better solutions with
finding other optimum solutions regarding the assumptions of
more frequency. In order to diversify the solutions, the process
Theorem 1. Moreover, Eq. (17) implies that the layouts of Figs. 2
does not negate the possibility of choosing worse solutions stored
and 3 are still optimum by exchanging the locations of facilities i
in the AM. The diversification procedure is performed by selecting
and j, provided that the same number of facilities are at the left
less competitive solutions; the selection of good solutions per-
and right sides of i and j, respectively. Therefore, we can construct
forms intensification. As the exploration continues, the algorithm
at least 2½2 optimum solutions for a problem valid in Theorem 1.
n

fills the AM with better solutions and removes uncompetitive lay-

Nevertheless, the number of required initial solutions is a parame-
out plans from the AM. This gradually reduces the significance of
ter defined by the user in the proposed algorithm. If this parameter
diversification while increases the importance of intensification.
is set to a number greater than 2½2 , it is still possible to use ran-
n

It can be inferred from the above explanations that the procedure

domly generated solutions to initiate the algorithm. Eventually,
is extremely sensitive to the probability function applied to the
the algorithm calculates the objective function of all the initial
selection procedure.
solutions and stores them in a memory called Adaptive Memory
We employ the probability function introduced by Rochat and
(AM), as it will be discussed in Section 4.3. 2i
Taillard (1995). This probability function assigns jLjjLþ1j as the
selection probability to the ith worst solution stored in the AM.
4.2. Neighborhood structure Therefore, the algorithm selects the first solution of the AM list
2jLj
with the probability of jLjjLþ1j and the worst solution with the prob-
2
Defining a proper neighborhood structure is one of the most ability of jLjjLþ1j. The proposed algorithm stops after performing k
important steps in developing a meta-heuristic algorithm. Neigh- iterations. The use of the AM and the described selection procedure
borhood structure in the proposed algorithm is based on binary in the proposed TS structure contributes to the efficiency of the
movements. The algorithm randomly chooses two facilities and ex- algorithm.
changes their locations. This process leads to a neighborhood solu-
tion in feasible space. Then, the algorithm accepts this new 4.4. Final intensification
solution if it improves the objective function of the ABSMODEL.
The algorithm then inserts the solution into the AM and adds the Final intensification guarantees more comprehensive search in
indices of the exchanged facilities to the tabu list for the next h the neighborhood of the best found layout plan. When the TS stops,
iterations. Otherwise, the algorithm rejects the exchange. the algorithm will apply the final intensification procedure to the
If the indices of facilities i and j are entered in the tabu list for best solution stored in the AM by applying the binary exchange
the next h iterations, the algorithm is prevented from exchanging of Section 4.2 to adjacent facilities. If the exchange of two facilities
them unless either the exchange is the only one that improves is accepted, then the final intensification phase will restart; other-
the objective function of ABSMODEL, or if the new solution gener- wise, the best current solution is considered as the final layout. In
ated by the exchange has the best objective function obtained so the proposed algorithm, the mandatory clearance between facili-
far. ties is not considered since there is no test problem with clearance
102 H. Samarghandi, K. Eshghi / European Journal of Operational Research 205 (2010) 98–105

reported in the literature. However, applying the clearance con- Step 3: Use a probability function to choose one of the solutions
straints in our algorithm is an easy task. stored in the AM for further search in its neighborhood.
th 2i
Probability of choosing the i worst solution is jLjjLþ1j .
5. Stepwise procedure Step 4: Randomly choose two facilities such as i and j from the
solution selected in Step 3 and exchange their locations
The stepwise procedure of the proposed TS is as follows: in the layout. If this exchange reduces the objective func-
tion, accept it and add (i, j) to the tabu list. Exchanging i
Step 1: Set the values of the control parameters: L (length of the and j will be prevented in the next h iterations unless it
AM list), a (number of attempts to find a better solution in constructs a solution better than all the current solutions
the neighborhood of the current solution), h (tabu list stored in AM, or if this exchange is the only exchange that
length) and k (number of iterations of the algorithm). improves the objective function. Then proceed to Step 5.If
Step 2: Find L initial solutions based on the procedure described the exchange does not improve the current solution,
in Section 4.1, store and sort them in the AM according repeat this step for a times. Go to Step 3 if no successful
to their objective functions in non-descending order. exchange is found.

Table 1
Problems with optimum solution.

Problem Reference No. of Best non-exact solution so far Proposed method Optimum solution Gap
no. dept. (%)
OFVa Time Reference OFV Time PIBFc OFV Reference
1 Beghin-Picavet and 4 78.00 0.00 Solimanpur et al. (2005) 78.00 0.00 1.00 78.00 Beghin-Picavet and 0
Hansen (1982) Hansen (1982)
2 Amaral (2006) 4 -b – – 638.00 0.00 0.50 638.00 Amaral (2006) 0
3 Love and Wong 5 – – – 151.00 0.00 0.40 151.00 Amaral (2006) 0
(1976)
4 Nugent et al. (1968) 5 1.100 0.00 Solimanpur et al. (2005) 1.100 0.00 0.40 1.100 Anjos and Vannelli 0
(2008)
5 Nugent et al. (1968) 6 1.990 0.00 Solimanpur et al. (2005) 1.990 0.00 2.00 1.990 Anjos and Vannelli 0
(2008)
6 Nugent et al. (1968) 7 4.73 0.00 Solimanpur et al. (2005) 4.73 0.00 5.14 4.73 Anjos and Vannelli 0
(2008)
7 Nugent et al. (1968) 8 6.295 0.00 Solimanpur et al. (2005) 6.295 0.00 0.50 6.295 Anjos and Vannelli 0
(2008)
8 Simmons (1969) 8 – – – 801.00 0.00 5.00 801.00 Amaral (2006) 0
9 Simmons (1969) 8 2324.50 0.00 Solimanpur et al. (2005) 2324.50 0.00 2.50 2324.50 Amaral (2006) 0
10 Simmons (1969) 9 – – – 2469.50 0.00 2.44 2469.50 Amaral (2006) 0
11 Simmons (1969) 9 – – – 4695.50 0.00 5.56 4695.50 Amaral (2006) 0
12 Simmons (1969) 10 2781.50 0.01 Solimanpur et al. (2005) 2781.50 0.00 5.60 2781.50 Amaral (2006) 0
13 Simmons (1969) 11 6933.50 0.03 Solimanpur et al. (2005) 6933.50 0.016 10.18 6933.50 Amaral (2006) 0
14 Heragu and Kusiak 12 – – – 18,190.0 0.017 16.67 18140.23 Anjos and Vannelli 0.27
(1991) (2008)
15 Heragu and Kusiak 15 44,599.9 0.18 Solimanpur et al. (2005) 44,600.0 0.016 43.20 44,600.00 Anjos and Vannelli 0
(1991) (2008)
16 Amaral (2006) 15 – – – 6305.00 0.031 21.73 6305.00 Amaral (2006) 0
17 Heragu and Kusiak 20 119.710 1.8 Solimanpur et al. (2005) 119.710 0.034 14.10 119.710 Anjos and Vannelli 0
(1991) (2008)
18 Heragu and Kusiak 20 15,549.0 2.3 Solimanpur et al. (2005) 15,549.0 0.020 11.20 15,549.00 Anjos and Vannelli 0
(1991) (2008)
19 Anjos and Vannelli 25 – – – 4631.00 0.038 36.88 4618.00 Anjos and Vannelli 0.28
(2008) (2008)
20 Anjos and Vannelli 25 – – – 37,116.5 0.039 30.24 37,116.50 Anjos and Vannelli 0
(2008) (2008)
21 Anjos and Vannelli 25 – – – 24,560.0 0.047 69.20 24,301.00 Anjos and Vannelli 1.06
(2008) (2008)
22 Anjos and Vannelli 25 – – – 48,291.5 0.040 44.40 48,291.50 Anjos and Vannelli 0
(2008) (2008)
23 Anjos and Vannelli 25 – – – 15,623.0 0.041 53.84 15,623.00 Anjos and Vannelli 0
(2008) (2008)
24 Anjos and Vannelli 30 – – – 8247.00 0.052 69.13 8247.00 Anjos and Vannelli 0
(2008) (2008)
25 Anjos and Vannelli 30 – – – 21,582.5 0.051 53.33 21,582.50 Anjos and Vannelli 0
(2008) (2008)
26 Anjos and Vannelli 30 – – – 46,212.0 0.062 54.80 45,449.00 Anjos and Vannelli 1.68
(2008) (2008)
27 Anjos and Vannelli 30 – – – 58,297.5 0.059 46.07 56,873.50 Anjos and Vannelli 2.5
(2008) (2008)
28 Anjos and Vannelli 30 – – – 115,826 0.068 45.93 115,268.0 Anjos and Vannelli 0.48
(2008) (2008)
29 Heragu and Kusiak 30 334.8968 37.3 Solimanpur et al. (2005) 334.870 0.077 59.33 334.870 Anjos and Vannelli 0
(1991) (2008)
30 Heragu and Kusiak 30 44,466.5 91.8 Kumar et al. (1995) 45,126.0 0.066 43.07 44,965.00 Anjos and Vannelli 0.36
(1991) (2008)
a
Objective function value.
b
Sign ‘‘–” is used for problems not previously solved by a non-exact algorithm.
c
PIBF: percentage of iterations in which the best solution appears for the first time.
 H. Samarghandi, K. Eshghi / European Journal of Operational Research 205 (2010) 98–105 103

Step 5: Store the improved layout into the AM, and sort the AM in the following tables are not comparable since different com-
again. Remove the worst layout stored in the AM to limit puter settings lead to different CPU times. Furthermore, in this
its length to L. Then proceed to Step 6. column, the percentage of iterations in which the best solution
Step 6: Select the best solution stored in the AM and proceed to appears for the first time is shown as PBIF. As the results show,
Step 7 provided that the above procedure has been the best solution can normally be found in early stages of the
repeated for k times. Otherwise, go to Step 3. algorithm, and the remaining running time verifies that a better
Step 7: Apply the final intensification procedure of Section 4.4 to solution cannot be found. The next two columns show the opti-
the solution generated by Step 6. mum solution and its reference. The gap between the optimum
solution and the solution constructed by our proposed algorithm
As seen in Step 1, the proposed TS algorithm includes four is reported in the last column. Table 1 shows that the deepest
parameters, L, a, h and k, which affect the algorithm’s performance. gap is only 2.5%, and in 23 problems out of 30 problems
We performed extensive sensitivity analysis on these parameters (77%), there is no gap.
to find the effect of the different values of parameters on the per- Table 2 indicates the key role of Theorem 1 in the proposed
formance of the algorithm. Based on these observations, the fol- algorithm. In this table, the gap between the optimal solution,
lowing experimentally derived values are proposed to set the and the solutions generated by two different methods for produc-
parameters: ing initial solutions are shown. In the first method, an initial solu-
j nk jn k j nk tion was generated by Theorem 1; in the second method, the initial
L¼ 2 ; h¼ ; a¼ ; k ¼ 50n: ð22Þ solution was obtained randomly. Table 2 demonstrates that the use
3 3 2
of the first and second methods resulted in an average gap of
25.40% and 49.67%, respectively. It can be noted that the first
6. Computational results method is always better than the second method, and the differ-
ence between these two methods is more than 60% in many cases,
The proposed algorithm was coded in Microsoft visual C++ as is evident in problem number 17.
and run on a PC equipped with a 2 GHz Intel Centrino Duo Furthermore, we have also tested the algorithm on another set
CPU and 1 GB of RAM. To evaluate the performance of the algo- of problems with no proven optimal solution reported in the liter-
rithm, we have tested it by all the problems available from the ature. These problems have larger instances compared to those
literature. Test problems can be found in Anjos and Kong presented in Table 1. This set contains 20 problems, all of them
(2007). All the problems from the literature can be divided into introduced and solved by a non-exact algorithm in Anjos et al.
two major sets, those with a proven optimal solution and those (2005). This set of problems is considered in Table 3. The best
without a proven optimal solution. The set of problems with an found solutions are printed in bold. The proposed algorithm has
optimal solution includes 30 problems and are shown in Table 1. improved the objective function of all the problems in this set.
The first column of Table 1 is the problem number. The refer- The CPU time for solving each problem is another issue that makes
ence where the problem has been introduced for the first time the proposed algorithm much more appropriate when coping with
is in the second column. The third column specifies the number large instance problems.
of facilities of the corresponding problem, and the next column For example, in problem number 7, the proposed algorithm
has information about the best non-exact solution found so far. takes only 2 second to find a solution which indicates a signifi-
The fifth column shows the results obtained by the proposed cant improvement in CPU time compared with the 7 hour taken
algorithm. As mentioned before, the proposed algorithm follows by the algorithm proposed by Anjos et al. (2005). Also, the objec-
a probabilistic approach. Thus, we have run it five times for each tive function is improved by 3.5% compared with Anjos et al.
problem. The information shown in the fifth column reflects the (2005). Moreover, the average CPU time of the proposed algo-
best objective function value found by the five independent runs rithm for the problems of Table 3 is 2.29 second while the aver-
and its relative CPU time. It should be mentioned that CPU times age CPU time for the same set of problems in Anjos et al.

Table 2
Initial solution for problems in Table 1.

Problem no. Initial solution generated by the Average of three random Problem no. Initial solution generated by the Average of three random
Theorem 1 permutations Theorem 1 permutations
OFV Gap (%) OFV Gap (%) OFV Gap (%) OFV Gap (%)
1 85.0 8.97 100.333 28.63 16 8348.0 32.40 10,976 74.08
2 718.0 12.54 750.6667 17.66 17 154.240 28.84 230.24 92.33
3 172.0 13.91 174.333 15.45 18 18,207.0 17.09 26,083.67 67.75
4 1450.0 31.82 1643.33 49.39 19 6580.0 42.49 6595.33 42.82
5 3140.0 57.79 3793.33 90.62 20 45,190.5 21.75 60,380.83 62.68
6 5950.0 25.79 7366.66 55.74 21 29,043.0 19.51 36,629.67 50.73
7 6.935 10.17 9.295 47.66 22 58,048.5 20.20 77,088.5 59.63
8 1015.0 26.72 10,66.0 33.08 23 20662.0 32.25 23,669 51.50
9 2408.5 3.61 2487.5 7.01 24 11,590.0 40.54 12,680.33 53.76
10 2898.5 17.37 3286.833 33.10 25 26,956.5 24.90 31,188.83 44.51
11 4756.5 1.30 4996.833 6.42 26 55,549.0 22.22 69,095 52.03
12 3889.5 39.83 4436.167 59.49 27 66,563.5 17.04 92,439.17 62.53
13 8581.5 23.77 10,592.5 52.77 28 146,568.0 27.15 192,534 67.03
14 30,655.0 68.99 30,666.33 69.05 29 407.230 21.61 502.927 50.19
15 57,540.0 29.01 66,713.33 49.58 30 55,013.0 22.35 64,193.67 42.76

Average
Initial solution generated by the Theorem 1 Random solutions
25.40% 49.67%
104 H. Samarghandi, K. Eshghi / European Journal of Operational Research 205 (2010) 98–105

Table 3
Problems without optimum solution.

Problem no. No. of dept. Proposed algorithm Anjos et al. (2005) Objective function Anjos et al. (2005) CPU time (hour)
a
Objective function CPU time (second) PIBF
1 60 1,477,834.00 0.820 59.23 1,493,704.00 5
2 60 841,792.00 0.980 66.67 843,644.00 5
3 60 648,337.00 0.900 69.27 656,272.50 5
4 60 398,511.00 0.913 40.00 405,433.00 5
5 60 318,805.00 0.762 72.07 319,501.00 5
6 70 1,529,197.00 1.499 84.91 1,543,098.00 7
7 70 1,441,028.00 1.940 82.14 1,494,182.00 7
8 70 1,518,993.50 1.761 99.06 1,524,171.50 7
9 70 969,130.00 1.233 82.49 974,856.00 7
10 70 4,218,230.00 1.570 78.46 4,230,912.50 7
11 75 2,393,483.00 2.010 98.16 2,399,583.50 10
12 75 4,321,190.00 2.198 80.53 4,348,544.00 10
13 75 1,248,551.00 2.912 77.92 1,295,085.00 10
14 75 3,942,013.00 2.516 83.52 3,949,276.50 10
15 75 1,791,408.00 2.098 82.37 1,816,455.00 10
16 80 2,069,097.00 3.975 84.38 2,138,083.50 10
17 80 1,921,177.00 5.641 82.23 1,939,938.00 10
18 80 3,251,413.00 4.797 87.25 3,332,421.00 10
19 80 3,746,515.00 3.453 83.03 3,773,429.00 10
20 80 1,589,061.0 3.769 74.50 1,611,495.00 10
Average – – 2.29 78.41 – 8
a
Shows the percentage of iterations in which the best solution appears for the first time.

Table 4
Initial solution for problems in Table 3.

Problem no. Initial solution generated by the Average of three random Problem no. Initial solution generated by the Average of three random
Theorem 1 solutions Theorem 1 solutions
OFV Gap (%) OFV Gap (%) OFV Gap (%) OFV Gap (%)
1 1,735,327.0 17.42 2,388,903.66 61.65 11 3,000,673.5 25.05 3,767,459.83 57.00
2 1,012,610.0 20.14 1,304,481.66 54.77 12 5,245,888.0 21.10 6,573,847 51.75
3 854,242.5 30.93 1,044,207.16 60.04 13 1,587,037.0 25.99 2,109,107.66 67.44
4 534,873.0 33.11 682,381.33 69.81 14 4,736,870.5 19.97 6,037,172.16 52.90
5 428,699.0 34.18 570,166 78.46 15 2,108,904.0 17.72 2,663,209.66 48.67
6 2,011,976.0 30.39 2,432,134.66 57.61 16 2,719,720.5 31.01 3,525,657.16 69.84
7 1,775,328.0 22.78 2,462,942 70.34 17 2,512,973.0 30.43 2,979,287.33 54.63
8 1,880,161.5 23.36 2,298,266.5 50.79 18 3,919,589.0 19.98 5,655,295 73.11
9 1,206,906.0 24.15 1,550,659.33 59.51 19 4,546,235.0 20.58 6,270,303.33 66.31
10 5,156,473.5 22.18 6,847,776.83 62.25 20 2,012,963.0 25.66 2,424,835 51.37

Average
Initial solutions generated by the Theorem 1 Random solutions
24.81% 60.91%

(2005) is 8 hour which further confirms the greater efficiency of 7. Conclusion

the proposed algorithm.
Table 4 demonstrates the important role of Theorem 1 in the This paper discusses the Single Row Facility Layout Problem
proposed algorithm. This table shows the gaps in two cases. In (SRFLP) in which the size of the machines is assumed to be differ-
the first case, an initial solution is found by Theorem 1, and in ent. The SRFLP is proven to be NP-Complete, and solving its large
the second case, an initial solution is the best of three random lay- instances is not possible in a reasonable amount of time even by
outs. The gaps between the best solution and the initial layouts in modern computers. Although exact and heuristic algorithms have
both cases are shown in Table 4. Table 4 demonstrates that the been developed in recent years to find optimal and near-optimal
average gap between the objective function of the initial solutions solutions for the SRFLP, we still need to have a fast algorithm that
generated by Theorem 1 and the objective function of the best can obtain a near-optimal solution for large instances of the
found solutions is 24.81% while the average gap in the second case SRFLP.
is 60.91%. This result, compared to those of Table 2, indicates that In this paper, we first concentrated on a special case of the SRFLP
the larger the instance of the general SRFLP, the better the quality and obtained its optimal solution theoretically. Then, this optimal
of the initial layout generated by Theorem 1. It can be noted that all solution was utilized to generate a number of initial solutions for
the results obtained by Theorem 1 are better than those generated a tabu algorithm, which was developed to find solutions of the gen-
by the random method. The maximum gap between the objective eral SRFLP. The proposed tabu algorithm uses a technique called
function of the layout generated by Theorem 1 and of the best adaptive memory for diversification as well as intensification. We
found solution is related to the problem number 5 (34.18%) while provided experimentally derived rules to set the values of the
the minimum gap between the best objective function of three parameters in different problems. The performance of the proposed
randomly generated layouts and of the best solution is related to tabu algorithm was tested over all of the problems available from
the problem number 15 (48.66%). the literature (50 problems) and also compared to many other
 H. Samarghandi, K. Eshghi / European Journal of Operational Research 205 (2010) 98–105 105

algorithms existing in the literature. Computational results of the Heragu, S.S., Kusiak, A., 1991. Efficient models for the facility layout problems.
European Journal of Operational Research (53), 1–13.
algorithm on benchmark problems show the efficiency of the algo-
Jean-Francois, C., Gilbert, L., 2005. Tabu search heuristics for the vehicle routing
rithm compared to other heuristics for solving the SRFLP. problem. In: Cesar, A.B.R. (Ed.), Metaheuristic Optimization via Memory and
Evolution, vol. 1. Kluwer Academic Publishers, Boston/Dordrecht/London, pp.
References 145–164.
Kumar, K.R., Hadjinicola, G.C., Lin, T.L., 1995. A heuristic procedure for the single
row facility layout problem. European Journal of Operational Research (87), 65–
Amaral, A.R.S., 2006. On the exact solution of a facility layout problem. European 73.
Journal of Operational Research (173), 508–518. Kumar, S., Asokan, P., Kumanan, S., Varma, B., 2008. Scatter search algorithm for
Amaral, A., 2009. A new lower bound for the single row facility layout problem. single row layout problem in FMS. Advances in Production Engineering and
Discrete Applied Mathematics 157, 183–190. Management 3 (4), 193–204.
Anjos, M.F., Kong, C., 2007. FLP Database, <http://flplib.uwaterloo.ca/>. Lin, M.T., 2009. The single-row machine layout problem in apparel manufacturing
Anjos, M.F., Vannelli, A., 2008. Computing globally optimal solutions for single-row by hierarchical order-based genetic algorithm. International Journal of Clothing
layout problems using semidefinite programming and cutting planes. INFORMS Science and Technology 21 (1), 31–43.
Journal on Computing 20 (4), 611–617. Love, R.F., Wong, J.Y., 1976. On solving a one-dimensional allocation problem with
Anjos, M.F., Kennings, A., Vannelli, A., 2005. A semidefinite optimization approach integer programming. Information Processes and Operation Research (INFOR)
for the single-row layout problem with unequal dimensions. Discrete (14), 139–143.
Optimization 2, 113–122. Neghabat, F., 1974. An efficient equipment layout algorithm. Operations Research
Beghin-Picavet, M., Hansen, P., 1982. Deux problems d’affectation non lineaires. (22), 622–628.
RAIRO Recherche Operationelle 16 (3), 263–276. Nugent, C.E., Vollman, T.E., Ruml, J., 1968. An experimental comparison of
Braglia, M., 1996. Optimization of a simulated-annealing-based heuristic for single techniques for the assignment of facilities to locations. Operations Research
row machine layout problem by genetic algorithm. International Transactions 16, 150–173.
in Operational Research 1 (3), 37–49. Picard, J.C., Queyranne, M., 1981. On the one-dimensional space allocation problem.
Drezner, Z., 1987. A heuristic procedure for the layout of a large number of facilities. Operations Research (29), 371–391.
Management Science 7 (33), 907–915. Rochat, Y., Taillard, E., 1995. Probabilistic diversification and intensification in local
Ficko, M., Brezocnik, M., Balic, J., 2004. Designing the layout of single- and multiple- search for vehicle routing. Journal of Heuristics 1, 147–167.
rows flexible manufacturing system by genetic algorithms. Journal of Material Simmons, D.M., 1969. One dimensional space allocation, An ordering algorithm.
Processing Technology, 150–158. Operations Research (17), 812–826.
Glover, F., 1986. Future paths for integer programming and links to artificial Solimanpur, M., Vrat, P., Shankar, R., 2005. An ant algorithm for the single row
intelligence. Computers & Operations Research 13, 533–549. layout problem in flexible manufacturing systems. Computers & Operations
Glover, F., Laguna, M., 1997. Tabu Search. Kluwer Academic Publishers, Boston/ Research 32, 583–598.
Dordrecht/London. Suresh, G., Sahu, S., 1993. Multiobjective facility layout using simulated annealing.
Heragu, S.S., Alfa, A.S., 1992. Experimental analysis of simulated annealing based International Journal of Production Economics 32, 239–254.
algorithms for the facility layout problem. European Journal of Operational Teo, Y.T., Ponnambalam, S.G., 2008. A hybrid ACO/PSO heuristic to solve single row
Research (57), 190–202. layout problem. In: 4th IEEE Conference on Automation Science and
Heragu, S.S., Kusiak, A., 1988. Machine layout problem in flexible manufacturing Engineering. Key Bridge Marriott, Washington DC, USA.
systems. Operations Research (36), 258–268.