Você está na página 1de 16

European Journal of Operational Research 261 (2017) 1–16

Contents lists available at ScienceDirect

European Journal of Operational Research


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

Invited Review

Mathematical optimization approaches for facility layout problems:


The state-of-the-art and future research directions
Miguel F. Anjos a,∗, Manuel V. C. Vieira b
a
GERAD & Department of Mathematics and Industrial Engineering, Polytechnique Montreal, Montreal, QC, Canada H3C 3A7
b
Departamento de Matemática, Faculdade de Ciências e Tecnologia & CMA, Universidade Nova de Lisboa, Portugal

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

Article history: Facility layout problems are an important class of operations research problems that has been studied for
Received 2 February 2016 several decades. Most variants of facility layout are NP-hard, therefore global optimal solutions are diffi-
Accepted 30 January 2017
cult or impossible to compute in reasonable time. Mathematical optimization approaches that guarantee
Available online 3 February 2017
global optimality of solutions or tight bounds on the global optimal value have nevertheless been suc-
Keywords: cessfully applied to several variants of facility layout. This review covers three classes of layout problems,
Facilities planning and design namely row layout, unequal-areas layout, and multifloor layout. We summarize the main contributions
Unequal-areas facility layout to the area made using mathematical optimization, mostly mixed integer linear optimization and conic
Row layout optimization. For each class of problems, we also briefly discuss directions that remain open for future
Mixed integer linear optimization research.
Semidefinite optimization
© 2017 Elsevier B.V. All rights reserved.

1. Introduction turbine runners (Laporte & Mercure, 1988), algorithm initialization


in numerical analysis (Brusco & Stahl, 20 0 0), VLSI fixed-outline
Facility layout problems (FLPs) are a general class of operations floorplanning (Luo, Anjos, & Vannelli, 2008), and optimal data
research problems concerned with finding the optimal arrange- memory layout generation for digital signal processors (Wess &
ment of a given number of nonoverlapping indivisible departments Zeitlhofer, 2004).
within a given facility. The objective is to minimize the total ex- FLPs have been extensively studied in the literature since the
pected cost of inter-departmental flows inside the facility, where 1960s. Numerous variations on the basic problem described above
the cost incurred for each pair of departments is equal to the have been considered, and different models have been proposed
rectilinear distance between the centroids of the departments mul- for each variation. Examples of such variations are: specially struc-
tiplied by their pairwise cost. This cost, generally non-negative, ac- tured instances of the problem (e.g. layouts on rows or on loops);
counts in the aggregate for adjacency preferences as well as costs dynamic FLPs with time-dependencies; FLPs under uncertainty in
that may arise from transportation, the construction of a material- the data; and multi-objective FLPs. We refer the reader to the
handling system, or connection wiring. The facility and the books (Heragu, 2008; Kusiak, 1990) and survey papers (Meller &
departments are rectangular, and the area of each department is Gau, 1996; Singh & Sharma, 2006) for more information about the
specified, but if the department’s dimensions can vary, then deter- FLP and its variations. A growing collection of FLP benchmark in-
mining them is also part of the FLP. stances is available online (Anjos, 2015).
FLPs have a variety of applications. Much of the work was mo- The FLP is NP-hard in general, so solving it to global optimality
tivated by the physical organization of manufacturing systems, see in reasonable time is generally difficult. Indeed the restricted ver-
e.g. Meller and Gau (1996). The FLP is particularly relevant in flexi- sion where the dimensions of the departments are all equal and
ble manufacturing systems that produce an array of different parts. fixed, and the optimization is taken over a fixed set of possible lo-
The layout of the production components has a significant impact cations for the departments, is known as the quadratic assignment
on the costs and the productivity of these systems, see e.g. Hassan problem, a combinatorial optimization problem well known for
(1994). Other applications of FLPs include balancing hydraulic its computational difficulty, see e.g. Loiola, de Abreu, Boaventura-
Netto, Hahn, and Querido (2007).
∗ The constraints of the basic FLP can be grouped into two sets:
Corresponding author.
E-mail addresses: anjos@stanfordalumni.org (M.F. Anjos), mvcv@fct.unl.pt
(M.V.C. Vieira). • Department shape requirements include the required area, and
URL: http://www.miguelanjos.com (M.F. Anjos) restrictions on the dimensions (height and width) such as

http://dx.doi.org/10.1016/j.ejor.2017.01.049
0377-2217/© 2017 Elsevier B.V. All rights reserved.
2 M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16

bounds on the ratios height/width and width/height, called as-


pect ratios. These requirements generally lead to convex con-
straints but still pose some challenges. In particular, requiring
small aspect ratios, while desirable in real-world applications,
generally makes the problem harder. On the other hand, while Fig. 1. SRFLP along the path of an AGV.
the area constraint traditionally required a careful linearization
approach, it can be modeled exactly using conic optimization,
see e.g. Anjos and Liers (2012). number of rows, and a pairwise non-negative weight for each pair
• Department location requirements include the nonoverlap of de- of departments, determine (i) an assignment of departments to
partments, fitting every department within the facility, assign- rows, and (ii) the positions of the departments in each row, so that
ing certain departments to, or forbidding them from, particular the total of the weighted center-to-center distances is minimized.
locations within the facility. The main challenge here are the Row FLPs arise in practical contexts where the departments are to
nonoverlap constraints that are inherently nonconvex and com- be placed in rows with a predetermined separation between the
binatorial. rows due to factors such as the material-handling system or the
flows of people. Moreover, within each row, a minimum clearance
This review is focused on FLPs with the following properties: between departments is needed to satisfy safety and operational
1. the departments have different areas requirements. We assume that this clearance is included in the
2. the facility can be one-, two-, or three-dimensional. lengths of the departments. We also assume that the rows and the
departments all have the same height, that any department can be
The different dimensions lead to the three broad classes of FLPs
assigned to any row, and that the distances between adjacent rows
covered in this review, namely row FLPs (Section 2), unequal-areas
are equal. Under these assumptions, solving an instance of the row
FLPs (Section 3), and multifloor FLPs (Section 4).
FLP means resolving three questions:
One-dimensional facilities lead to row FLPs, and we categorize
them in terms of the number of rows: single-row, double-row, or 1. Assign each department to exactly one row;
multi-row. Single-row and double-row problems commonly occur 2. Express mathematically the center-to-center distance between
in practical applications, as we discuss in Sections 2.1 and 2.2 re- departments (that may or may not be in the same row);
spectively. Multi-row problems are a natural extension of the prob- 3. Handle possible empty space between departments.
lem to three or more rows, and are considered in Section 2.3.
Unequal-areas FLPs have two-dimensional facilities with a sin- Section 2.1 is concerned with the simplest version of row FLP,
gle floor, and we assume that the facility is rectangular and that namely the single-row FLP. Section 2.2 covers the double-row FLP,
all the departments fit inside the facility. Unlike in the case of row and Section 2.3 extends the coverage to the general multirow FLP.
layouts, not only the position but also the dimensions of each de-
partment are optimized. After discussing models and approaches 2.1. The single-row FLP
for the basic two-dimensional problem in Sections 3.1–3.4, we con-
sider in Section 3.5 the special case of flexible bay layouts, a type An instance of the Single-Row FLP (SRFLP) consists of n one-
of layout that resembles row FLPs but with the fundamental differ- dimensional departments with given positive lengths 1 , . . . , n and
ence that the width of the bays can vary, depending on the total pairwise costs cij . The problem is to find a permutation of the de-
area of the departments in each bay. partments that minimizes the weighted sum of the pairwise dis-
Three-dimensional facilities give rise to multifloor FLPs in tances. Fig. 1 provides an illustration of the SRFLP in the context
which departments are to be placed over two or more floors. This of placing the departments along the path of an automated guided
is the focus of Section 4. The survey in Section 4.1 shows that vehicle (AGV) transporting material between the departments; in
most of the literature proposes models for specific applications this context the objective is to minimize the distance travelled by
rather than for the general problem. For this reason we propose the AGV. The SRFLP is the most studied of the row FLPs. Some-
in Section 4.2 a formulation for a generic form of the problem that times called the one-dimensional space allocation problem, it has
we hope will motivate further research into multifloor FLP. interesting connections to well-known combinatorial optimization
Regarding the choice of methodologies, we limit the scope of problems such as maximum-cut, quadratic linear ordering, and lin-
this review to mathematical optimization-based approaches. These ear arrangement (see Anjos & Liers, 2012).
include exact methods, but as the problems increase in difficulty Because there is only one row, there is no need to assign de-
very rapidly, we also include heuristic methods that use math- partments to rows. Moreover, cij ≥ 0 ensures that there is no
ematical optimization approximations and/or relaxations. While empty space between departments at optimality. Hence the re-
there is a rich literature on heuristic algorithms for FLPs (see maining question is to express the center-to-center distance be-
e.g. Kothari & Ghosh, 2012; Meller & Gau, 1996; Singh & Sharma, tween departments.
2006), our focus here is on mathematical optimization approaches, A key observation, first made by Simmons (1969), is that the
primarily mixed integer linear optimization (MILO), often referred SRFLP can be expressed as
to as mixed integer programing or MIP, semidefinite optimization
 1  
(SDO), also called semidefinite programing or SDP, and nonlinear min ci j i +  j + Dπ (i, j ) ,
optimization. Because of their importance to the success of these π ∈ n 2
i< j
approaches, we also include brief discussions of symmetry break-
ing (Section 5) and valid inequalities (Section 6) as these are es- where n denotes the set of all permutations of {1, 2, . . . , n}, and
sential ingredients for solving the resulting relaxations efficiently. Dπ (i, j) is the center-to-center distance between departments i and
We conclude with a summary of directions for future research j under permutation π .
in Section 7. A first observation here that if π  denotes the permuta-
tion symmetric to π , defined by πi = πn+1−i , i = 1, . . . , n, then
2. Row FLPs Dπ (i, j ) = Dπ  (i, j ). In other words, the order of the departments
in a particular layout can be reversed without changing the value
Row FLPs share the following common problem statement: of the objective function. Hence, it is possible to simplify the prob-
given a set of rectangular departments each of a given length, a lem by considering only the permutations that have a particular
M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16 3

facility, say facility 1, in the left half of the arrangement. Alterna- 2.1.2. SDO model
tively, we can require that a specific facility be to the left of an- To present an SDO-based relaxation, we begin by introducing {
other; this is known as the position p − k method, see Section 5. ± 1} binary variables as in customary in SDO (see Anjos & Liers,
This type of symmetry-breaking strategy can help reduce the com- 2012). For each pair of departments ij with 1 ≤ i < j ≤ n, define
putational cost of a mathematical optimization algorithm for SRFLP 
and for other types of layout problems, see Section 5. One aspect 1, if i is to the right of j,
Ri j :=
unique to the SDO-based approach is that it implicitly accounts for −1, otherwise.
these symmetries, and thus does not require the use of additional In this definition, the order of the subscripts matters, and Ri j =
explicit symmetry-breaking constraints, see Section 2.1.2. −R ji .
A second observation is that it is not necessary to know the posi- For an assignment of ± 1 values to the Rij variables to represent
tion of each department; it suffices to know for each pair of depart- a permutation, it is necessary to enforce the transitivity condition:
ments which departments are between them. Hence the key here if i is to the right of j and j is to the right of k, then i is to the
is the concept of betweenness. right of k.
There is a large amount of literature on the SRFLP. For more Equivalently, if Ri j = R jk then Rik = Ri j . This condition can be
detailed expositions on the state-of-the-art for the SRFLP, includ- formulated using quadratic constraints:
ing extensions, meta-heuristics, and exact approaches, we refer the
reader to Kothari and Ghosh (2012) and to the recent review paper Ri j R jk − Ri j Rik − Rik R jk = −1 for all triples 1 ≤ i < j < k ≤ n.
(Keller & Buscher, 2015) in this journal. (6)
To give the reader a sense of the mathematical optimization
approaches to the SRFLP, we present here two different ways to Using the Rij variables, it is straightforward to express between-
model betweenness. One is based on MILO and the other based on ness after observing that Rki Rk j = −1 if and only if facility k is be-
SDO. tween i and j. Hence the objective function can be expressed as


2.1.1. MILO model
 1   1 − Rki Rk j
ci j i +  j + k ,
The approach sketched here was originally proposed in Amaral 2 2
i< j k=i, j
(2009b). Other MILO models for SRFLP include, in chronological or-
der, (Amaral, 20 06; 20 08; Heragu & Kusiak, 1991; Love & Wong, and the consequent formulation of SRFLP is:
1976).

 ci j   
For three distinct departments i, j, k, define the betweenness min K− k Rki Rk j − k Rik Rk j + k Rik R jk
variables ζ ijk as: 2
i< j k<i i<k< j k> j
 s.t. (7)
1, if department k lies between departments i and j,
ζi jk = Ri j R jk − Ri j Rik − Rik R jk = −1 for all triples i < j < k
0, otherwise.
R2i j = 1 for all i < j
Using these variables, the objective function of the SRFLP is ex-
c
pressed as: where K := ( i< j 2i j )( nk=1 k ).
 Applying standard techniques from SDO, this formulation leads
 1   to the following SDO relaxation (Anjos, Kennings, & Vannelli,
ci j i +  j + k ζi jk
2 2005):
i< j k=i, j

and this is optimized subject to the following constraints:  ci j   
min K− k Xki,k j − k Xik,k j + k Xik, jk
ζi jk + ζik j + ζ jki = 1, for all {i, j, k} ⊆ {1, . . . , n}, (1) 2
i< j k<i i<k< j k> j
s.t.
(8)
ζi jd + ζ jkd − ζikd ≥ 0, for all {i, j, k, d} ⊆ {1, . . . , n}, (2) Xi j, jk − Xi j,ik − Xik, jk = −1 for all triples i < j < k
Xii = 1, for i = 1, . . . , n
ζi jd + ζ jkd + ζikd ≤ 2, for all {i, j, k, d} ⊆ {1, . . . , n}, (3) X 0, X ∈ S ( 2 )
n

where X 0 denotes that X is symmetric positive semidefinite, and


ζi jk ∈ {0, 1}, for all {i, j, k} ⊆ {1, . . . , n}. (4) 
S (2 ) is the set of symmetric matrices of dimension n2 . The inter-
n

A polyhedral study concerning this formulation can be found in pretation of the entries of X is that X pi ,p j = R pi R p j for any two pairs
Sanjeevi and Kianfar (2010). When Eq. (4) is relaxed to 0 ≤ ζ ijk ≤ pi , pj .
1, the resulting linear optimization (LO) relaxation is weak. Thus Note that if every Rij variable is replaced by its negative, then
an additional class of valid inequalities that improve the relaxation there is no change whatsoever to the formulation. In this way, the
is proposed in Amaral (2009b). formulation (7) and the corresponding SDO relaxation (8) implic-
itly account for the symmetry of the SRFLP.
Proposition 1 Amaral (2009b). Let β ≤ n be a positive even integer
Subsequent improvements to the relaxation (8) were given in
and let S ⊆ {1, . . . , n} such that |S| = β . For each r ∈ S, and for any
Hungerländer and Rendl (2013). We refer the reader to that paper
partition (S1 , S2 ) of S\{r} such that |S1 | = 12 β , the inequality
and to Keller and Buscher (2015) for more details.
  
ζtqr + ζtqr − ζmin{t ,q},max{t ,q},r ≤ 0 (5)
t <q,t ∈S1 ,q∈S1 t <q,t ∈S2 ,q∈S2 t∈S1 ,q∈S2 2.2. The double-row FLP
is valid for the above formulation of the SRFLP.
The Double-Row FLP (DRFLP) is an extension of the SRFLP
It is straightforward to check that for β = 4, (5) is of the form in which departments can be placed on both sides of a central
(2). it is shown in Amaral (2009b) that the size of the LO relax- corridor. The distance between the two rows is assumed to be
ation can be reduced by projecting the feasible set into a lower- negligible, and thus the center-to-center distance between two
dimensional space. departments is measured parallel to the corridor. Fig. 2 illustrates
4 M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16

s.t. di j ≥ xi − x j , di j ≥ x j − xi , 1 ≤ i < j ≤ n (13)

 +  
i j
xi + ≤ x j + L(1 − yi j ), 1 ≤ i, j ≤ n, i = j (14)
2
Fig. 2. DRFLP with a corridor for an AGV.  +    +  
i j i j
di j − yi j − y ji ≥ 0, 1 ≤ i < j ≤ n (15)
2 2
the DRFLP with the corridor as the operating space for an AGV.
Another application for the DRFLP is the arrangement of rooms in
buildings, see e.g. Ahonen, de Alvarenga, and Amaral (2014). To the y ∈ Qn (16)
best of our knowledge, the first reference to double-row layouts
is in Heragu and Kusiak (1988) where a nonlinear optimization
model is proposed and used to find locally optimal solutions. Most yi j ∈ {0, 1}, 1 ≤ i, j ≤ n, i = j (17)
of the subsequent mathematical optimization approaches in the
literature use either MILO (with the first model introduced in i i
Chung and Tanchoco, 2010 and a recent new model in Amaral, ≤ xi ≤ L − , 1 ≤ i ≤ n (18)
2 2
2013a) or SDO (Hungerländer & Anjos, 2015).
Unlike for the SRFLP, there is in the DRFLP a need to address where we use the continuous variables
all three questions for row FLPs. The assignment of departments to
rows is somewhat simplified by the fact that there are only two • xi representing the position of the center of i (1 ≤ i ≤ n) along
rows: it suffices to determine which departments are placed in the the corridor,
first row, because the remaining departments must be in the sec- • dij representing the distance between (the centers of) i and j (1
ond row. On the other hand, betweenness no longer suffices to de- ≤ i < j ≤ n) measured parallel to the corridor.
termine center-to-center distances, and the optimal layout may in- n
volve some empty space between departments. Also L = i=1 i , and

2.2.1. MILO models Qn = {y ∈ Rn(n−1) : (9 ), (10 )(11 ), 0 ≤ yi j ≤ 1, 1 ≤ i, j ≤ n, i = j}.


In this section we describe two approaches that extend in dif-
The integral points of the polytope Qn are precisely the y-incidence
ferent ways the MILO models proposed for the SRFLP. Both ex-
vectors of interest (Amaral, 2013a; Coll, Ribeiro, & de Souza, 2006).
tensions involve a combination of discrete and continuous vari-
Constraints (13) give the rectilinear distance between each
ables, where the former represent the assignment of departments
pair of departments. Constraints (16) and (17) characterize the y-
to rows and the relative position of two departments, and the lat-
incidence vectors, and constraints (18) are bounds on the x vari-
ter give the positions of the department centers with respect to a
ables. Constraints (14) ensure that departments assigned to the
fixed origin. Without loss of generality the corridor is placed along
same row do not overlap.
the x-axis, and the origin is at the left end of the corridor.
Constraints (15) ensure that if department i is placed in the
A model with O(n2 ) binary variables
same row as department j, then the distance between their cen-
Consider the binary vector y = (yi j )1≤i, j≤n such that
ters is at least (i +  j )/2. Note that constraints (15) are redundant

1, if department i is to the left of department j in the presence of constraints (13) and (14), but they may be help-
yi j = and both i and j are in the same row; ful for a branching algorithm.
0, otherwise. A model with O(n3 ) binary variables
For this model, we define two sets of binary variables:
The following inequalities are valid for all y-incidence vectors rep-
resenting a partition of the n departments into two ordered sub- 
1, if department i is assigned to row k
sets: yik =
0, otherwise.
yik + yki + y jk + yk j − yi j − y ji ≤ 1, 1 ≤ i, j, k ≤ n, i < j, k = i, j
1, if department j is placed to the right of
(9)
zki j = department i in row k
yik + y ji + yk j − yki − yi j − y jk ≤ 1, 1 ≤ i, j, k ≤ n, i, k < j, k = i 0, otherwise.
(10)
As in the previous model, we use continuous variables to deter-
yi j + yik + y jk + y ji + yki + yk j ≥ 1, {i, j, k} ⊂ {1, . . . , n}. (11) mine the location of the departments. Specifically we let xik denote
Constraints (9) are transitivity constraints with respect to row the absolute location of department i in row k, and set it to zero if
assignments. They ensure that if i and k are in the same row i is not assigned to row k.
(yik + yki = 1) and k and j are in the same row (y jk + yk j = 1), then These definitions support the model proposed in Chung and
1 + 1 − (yi j + y ji ) ≤ 1, implying yi j + y ji ≥ 1, i.e., i and j are in the Tanchoco (2010). This model explicitly accounts for clearances be-
same row. tween departments. As corrected in Zhang and Murray (2012), the
Constraints (10) are three-cycle constraints. They forbid a solu- model is:
tion where k is placed to the right of i, i is to the right of j, and j n−1 
 n  
is to the right of k (thus forming an impossible cycle). min ci j v+i j + v−i j
Constraints (11) require that at least two of i, j, k must be in the i=1 j=i+1
same row. It also ensures that no more than two rows are used.  
s.t. xik − x jk + v+
ij
− v−
ij
= 0, i ∈ I1 , j ∈ I2 (19)
We now state the MILO model of Amaral (2013a):
k∈K k∈K
n−1 
 n
min ci j di j (12)
xik ≤ Myik , i = 1, . . . , n, k ∈ K (20)
i=1 j=i+1
M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16 5


yik = 1, i = 1, . . . , n (21) In terms of approaches using MILO and SDO, as noted in
k∈K Section 2.2.1, the O(n3 ) MILO formulation of Zhang and Murray
(2012) for the DRFLP can be easily extended to the MRFLP (this
i yik +  j yik
+ aik zk ji ≤ xik − x jk + M (1 − zk ji ) (22) was not specifically done in that paper). More recently, an SDO-
2 based approach was introduced in Hungerländer and Anjos (2015),
and it is this approach that we present here. To the best of our
i ∈ I1 , j ∈ I2 , k ∈ K
knowledge, this is the only global optimization approach for the
i yik +  j yik general row FLP with more than two rows.
+ aik zki j ≤ x jk − xik + M (1 − zki j ), (23)
2
2.3.1. SDO model
i ∈ I1 , j ∈ I2 , k ∈ K The SDO model presented in Hungerländer and Anjos (2015) for
1 the MRFLP is based on the SDO formulation for the SRFLP pre-
zki j + zk ji ≤ (yik + y jk ), i ∈ I1 , j ∈ I2 , k ∈ K (24)
2 sented in Section 2.1.2. The idea is to first assume that the assign-
ment of departments to rows is fixed and that no spaces are al-
zki j + zk ji + 1 ≥ yik + y jk , i ∈ I1 , j ∈ I2 , k ∈ K (25) lowed between departments in the same row. This restricted ver-
sion of the MRFLP is called the k-Parallel Row Ordering Problem
(k-PROP), see Section 2.3.2 and the references therein for more de-
xik ≥ 0, i ∈ I, k ∈ K
tails.
v+i j , v−i j ≥ 0, i ∈ I1 , j ∈ I2 Consider the k-PROP with n departments and m rows, and
yik ∈ {0, 1}, i ∈ I, k ∈ K let the assignment of departments to rows be specified by r :
{1, . . . , n} → {1, . . . , m}. Define the binary variables Rij as in
zki j ∈ {0, 1}, i ∈ I, j ∈ I \ {i}, k ∈ K (26)
Section 2.1.2, and let dij represent the center-to-center distance be-
where aij is the required clearance between departments i and j, tween i and j measured parallel to the rows. If i and j are assigned
I1 = {1, . . . , n − 1}, I2 = {i +1, . . . , n}, K = {1, 2} is the set of rows, to the same row, i.e., if r (i ) = r ( j ), then

and the constant M = i∈I i + max j∈I ai j is analogous to L in the 1  1 − Rki Rk j  1 + Rik Rk j
previous model but also includes the clearances. di j = ( i +  j ) + k + k
2 2 2
Constraints (19) compute the distances between departments. k∈N, k<i k∈N, i<k< j
r ( k )= r ( i ) r ( k )= r ( i )
Constraints (20) set xik = 0 when department i is not assigned to
 1 − Rik R jk
row k. Constraints (21) ensure that a department is assigned to just + k , (27)
one row. Constraints (22) and (23) prevent departments from over- 2
k∈N, k> j
lapping if they are located in the same row. r ( k )= r ( i )
Constraints (24) and (25) ensure consistency between the vari- while if r (i ) = r ( j )
ables y and z as follows: if yik = 1 and y jk = 1 then (24) and ⎡⎛ ⎞
(25) together ensure that exactly one of zkij and zkji is equal to one.
⎢⎜  j  1 + Rk j  1 − R jk ⎟
Otherwise, i.e., if at least one of yik and yjk is equal to zero, then d i j = R i j ⎣⎝
2
+ k
2
+ k
2

(24) sets both zkij and zkji to zero. Constraints (25) force either zkij k∈N, k< j k∈N, k> j
or zkji to be 1 if i and j are both in row k. r ( k )= r ( j ) r ( k )= r ( j )
⎛ ⎞⎤
Note that the O(n2 ) model has significantly fewer variables than
the O(n3 ) model, and that the meaning of the continuous variables ⎜ i  1 + Rki  1 − Rik ⎟⎥
xik differs between the two models. Finally, it is important to ob- −⎝ + k + k ⎠⎦. (28)
2 2 2
k∈N, k<i k∈N, k>i
serve that while the O(n2 ) model is specific to the DRFLP, the O(n3 ) r ( k )= r ( i ) r ( k )= r ( i )
model can be applied directly to the MRFLP by increasing the car-
The above relations, plus the triangle inequalities relating the dis-
dinality of K.
tances between every triplet of departments i, j, k:
2.2.2. SDO model zi j + zik ≥ z jk , zi j + zik ≥ z jk , zik + z jk ≥ zi j , 1 ≤ i < j < k ≤ n,
An SDO-based approach for the MRFLP was developed in (29)
Hungerländer and Anjos (2015) and also applied to the DRFLP. This
are used in Hungerländer and Anjos (2015) to extend the SDO for-
approach is presented in Section 2.3.1.
mulation for the SRFLP to an SDO formulation for the k-PROP. For
the sake of brevity here, we refer the reader to Hungerländer and
2.3. The multirow FLP Anjos (2015) for the technical details.
Once an SDO formulation of the k-PROP is obtained, the possi-
The MRFLP is a natural extension of row layout to three or more bility of spaces is handled using the following results:
rows. An instance of the MRFLP has a given number of rows to
which the departments can be assigned, the departments all have Theorem 1 (Hungerländer & Anjos, 2015). If all the department
the same height (equal to the row height), the distances between lengths i are integer, then there is always an optimal solution to the
adjacent rows are equal, and departments can in general be as- MRFLP on the half-integer grid.
signed to any row.
Corollary 1 (Hungerländer & Anjos, 2015). If all the department
The MRFLP has received very limited attention in the operations
lengths i are integer, then for each instance of the MRFLP, we ob-
research literature to date. In terms of practical applications, it cap-
tain an equivalent instance of the k-PROP by adding spacing depart-
tures the basic structure of contexts where the departments are to
ments of length 0.5 such that the length of each row becomes equal
be arranged in well-defined rows because the separation between
to M := ni=1 i .
the rows is predetermined. It is thus a problem that is discrete in
one dimension and continuous in the other. Heuristic algorithms The strategy is thus to add spacing departments of length 0.5
were proposed in Heragu and Kusiak (1988), and a nonlinear op- and with all involved connectivities equal to zero, and then ap-
timization formulation was given in Gen and Cheng (1997) and ply the SDO approach for k-PROP. Because the number of spac-
solved using a genetic algorithm (GA). ing departments needed will normally be too large for practical
6 M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16

computation, several results are proved in Hungerländer and Anjos For the DRFLP, the O(n2 ) model was used in Amaral (2013a) to
(2015) to reduce the number of spacing departments needed. obtain solutions of instances with up to 12 departments within one
Finally, to remove the restriction that the assignment of de- hour. The O(n3 ) model was also tested in Amaral (2013a) but was
partments to rows is fixed, Hungerländer and Anjos (2015) ob- unable to solve instances with more than 10 departments within
tain global optimal solutions (respectively bounds) by using this three hours. The corrected model of Zhang and Murray (2012) was
approach for all possible assignments (respectively for a subset of used in Murray, Smith, and Zhang (2013) for asymmetric flows. The
them). constraints are (20)–(26), and the objective function is
  
2.3.2. Special cases of the MRFLP (ci j + c ji ) v+i j + v−i j .
The difficulty in solving the general MRFLP has motivated the i∈I1 j∈I2

study of a number of special cases with simplifying assumptions The conclusion of the computational tests is that with a time limit
and/or specific structure that allow for more effective modeling of 10 minutes, most of the heuristic algorithms perform better
and solution approaches. than CPLEX on instances with more than 20 departments.
Finally for the MRFLP, tight global bounds were computed in
The equidistant MRFLP
Hungerländer and Anjos (2015) for instances with up to 12 de-
A first such special case is the equidistant version of the MR-
partments. The authors adapted an approach originally proposed
FLP, denoted MREFLP, in which all departments have the same
in Fischer, Gruber, Rendl, and Sotirov (2006) for the max-cut prob-
length. This structure makes it possible to prove many interest-
lem and several ordering problems. The SDO-based approach was
ing results. The single-row case is known in the literature as
applied to instances with up to 5 rows and up to 8 departments.
the linear arrangement problem, see e.g. Liu and Vannelli (1995),
The results show that the SDO approach is most effective for 4 or
Amaral, Caprara, Letchford, and Salazar (2008), Amaral and Letch-
5 rows. There may be an intuitive explanation for this: as an ex-
ford (2011), Palubeckis (2012), Amaral (2009a), and is well known
treme example, note that it is easier to partition 5 departments
to be NP-hard even if all the pairwise costs are binary (Garey, John-
into 5 rows than into 2 rows. This is in part because the model
son, & Stockmeyer, 1974).
does not take into account the distance between rows, so assign-
The double-row case was considered in Amaral (2011) where a
ing department 1 to row 1 is exactly the same as assigning it to
MILO formulation based on the quadratic assignment problem is
row 4. Accounting for the distances between rows may change the
given.
nature of the results, but has not yet been done to the best of our
For the general MREFLP, it is shown in Anjos, Fischer, and
knowledge.
Hungerländer (2015) that the problem has an optimal solution on
the integer grid (although the lengths of the spaces are in general
3. Unequal-areas FLP
continuous quantities). This implies that only spaces of unit length
need to be used when modeling the MREFLP, and hence that the
The Unequal-Areas FLP (UA-FLP) is concerned with finding the
problem can be formulated as a purely discrete optimization prob-
optimal arrangement of a given number of nonoverlapping indivis-
lem, as is the case for the SRFLP in Section 2.1. Moreover, exact
ible departments with varying areas so as to minimize the total
results were proved in Anjos et al. (2015) for the minimum num-
expected cost of flows inside the facility. Unlike in the row FLPs,
ber of spaces that must be added so as to preserve at least one
the dimensions of each department are optimized (subject to the
optimal solution. These results lead to both MILO and SDO models
area requirement).
for the MREFLP.
The UA-FLP, sometimes called the single-floor FLP, has received
The space-free MRFLP much attention in the literature. It was first stated in Armour and
Another important special case of the MRFLP is the Space-Free Buffa (1963), and one of the first MILO formulations was proposed
MRFLP (SF-MRFLP) in which no spaces are allowed within the in Montreuil (1991) using binary variables to prevent overlap.
rows, all rows have a common left origin, and the leftmost de- We begin with an exact formulation of the UA-FLP in
partment in each row is flush with the left end of the row. When Section 3.1. This allows us to establish notation, and more impor-
there is only one row, the SF-MRFLP is equivalent to the SRFLP. tantly to explicitly show where the main difficulties are for solving
Where there are two rows, the SF-MRFLP is also called the Space- UA-FLP. Exact MILO models are covered in Section 2.2.1. This in-
Free DRFLP or the Corridor Allocation Problem, for which a MILO cludes sequence-pair formulations in Section 3.2.1, one of which
formulation was proposed in Amaral (2012), and an SDO approach solved instances with up to 11 departments to global optimality,
in Hungerländer and Anjos (2012). the largest such results to date (Meller, Chen, & Sherali, 2007).
A special case of the SF-MRFLP that has attracted attention is Most of the approaches reviewed here are two-stage frame-
the k-PROP introduced in Section 2.3.1. Because the assignment of works, where the first stage determines the relative location of
departments to rows is given, and no spaces are allowed within the departments, and the second stage obtains a final layout via
the rows, the k-PROP reduces to finding the optimal permuta- a mathematical optimization model. Two-stage approaches are
tion of the departments within each row. An SDO approach for k- mathematical optimization-based heuristics that are not guaran-
PROP was mentioned in Section 2.3.1, and another was given in teed to find the global optimal layout but they seem to be the
Hungerländer (2014). When the number of rows equals two, this most promising for handling large-scale instances of UA-FLP. The
problem is simply called PROP, and a MILO formulation for it was main differences between the approaches are in the first-stage al-
given in Amaral (2013b). gorithms. We present in Section 3.3 approaches that are entirely
based on nonlinear optimization models, one of which was re-
2.4. Computational performance of the models cently shown to be able to compute layouts for instances with up
to 100 departments in less than 15 minutes of computation time
Row FLPs remain highly challenging problems. We summarize (Anjos & Vieira, 2015). Other two-stage approaches are summa-
here the state-of-the-art in terms of the computational perfor- rized in Section 3.4.
mance of the approaches preserved above. A MILO formulation for the important special case of flexible
For both the SRFLP and the single-row MREFLP, the largest in- bay UA-FLP is discussed in Section 3.5
stances solved to optimality had 42 departments, see Hungerländer A number of heuristics for the UA-FLP make use of a slicing-
and Rendl (2013) and Hungerländer (2014) respectively. tree structure. This is a binary tree that represents the floor plan
M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16 7

after applying a recursive partitioning process. Each node of the these constraints are disjunctive and nonconvex, and are the hard-
tree contains either a department or a cut operator, thus each est ones to handle. If the relative position of each pair of depart-
slicing tree corresponds to a particular layout. This strategy was ments is known, then the constraints (37) can be written as linear
first used in Otten (1982) in the context of VLSI design and inequalities, and the formulation becomes a (convex) conic opti-
later extended to the UA-FLP in Tam (1992). It was also used in mization problem that is straightforward to solve. This observation
Shayan and Chittilappilly (2004), Diego-Mas, Santamarina-Siurana, motivates the two-stage philosophy in several of the approaches in
Cloquell-Ballester, and Alcaide-Marzal (2008), Scholz, Petrick, and the literature; we present the most prominent in Sections 3.3 and
Domschke (2009), Komarudin and Wong (2010), Chang and Ku 3.4.
(2013). Note that this formulation locates the center of the facility at
the origin, while some of the models below locate the origin at the
3.1. An exact formulation of the UA-FLP bottom left-hand corner of the facility. This difference is otherwise
of no consequence.
We begin by presenting an exact formulation that uses only
continuous variables. The reasons for doing so are two-fold: we 3.2. MILO models
establish some notation that will be common for the remainder
of this section, and we explicit point out where the difficulties lie We begin with the MILO model introduced by Meller,
in solving UA-FLP, thus motivating the solution approaches subse- Narayanan, and Vance (1999) and enhanced in Sherali, Fraticelli,
quently presented. and Meller (2003). Define the binary variables
We assume that we are given the height and width of the fa- 
1 if i must precede j horizontally,
cility as hF and wF respectively, and that for each department i we zihj =
0 otherwise,
have lower and upper bounds wmin i
and wmax
i
on its width, and 
min
hi and himax on its height. We also assume that β i , an upper zivj =
1 if i must precede j vertically,
bound on the aspect ratio of department i, is given for each de- 0 otherwise.
partment i. It is necessary that β i ≥ 1, and the closer β i is to unity,
The MILO formulation is as follows:
the closer the shape of department i will be to a square. 
With this notation, the UA-FLP can be formulated as follows min ci j ( u i j + vi j ) (38)
(see van Camp, Carter, & Vannelli, 1991): 1≤i< j≤n

min ci j ( |xi − x j | + |yi − y j | ) (30) ui j ≥ xi − x j and ui j ≥ x j − xi , 1 ≤ i < j ≤ n
xi ,yi ,hi ,wi s.t. (39)
1≤i< j≤n vi j ≥ yi − y j and vi j ≥ y j − yi , 1 ≤ i < j ≤ n
s.t. wmin ≤ wi ≤ wmax , for i = 1, . . . , n (31) 1 1
i i wi ≤ xi ≤ wF − wi , i = 1, . . . , n
2 2
1 1
hmin
i ≤ hi ≤ hmax
i , for i = 1, . . . , n (32) hi ≤ yi ≤ hF − hi , i = 1 , . . . , n (40)
2 2

w i hi = Ai , for i = 1, . . . , n (33) wmin


i ≤ wi ≤ wmax
i , i = 1, . . . , n

  hmin
i ≤ hi ≤ hmax
i , i = 1, . . . , n (41)
w i hi
max , ≤ βi , for i = 1, . . . , n (34)
2
hi w i
λ
ai wi + 4 wmin + ( wmax −) hi
wmin
1 1 1 1 i
−1 i i
xi + wi ≤ wF and wi − xi ≤ wF , for i = 1, . . . , n (35)

2 2 2 2 λ
≥ 2ai wmin + (wmax − wmin ) , λ = 0, 1, . . . ,  − 1 (42)
1 1 1 1
i
−1 i i
yi + hi ≤ hF and hi − yi ≤ hF , for i = 1, . . . , n (36)
2 2 2 2
zihj + zhji + zivj + zvji = 1, 1 ≤ i < j ≤ n (43)
1 1
|xi − x j | ≥ (wi + w j ) or |yi − y j | ≥ ( hi + h j ),
2 2 1 1
xi + wi ≤ x j − w j + wF (1 − zihj ), i = j
for all 1 ≤ i < j ≤ n. (37) 2 2
1 1
The first four sets of constraints enforce the shape requirements. yi + hi ≤ y j − h j + hF (1 − zivj ), i = j (44)
Constraints (31) and (32) enforce the bounds on the width and 2 2
height of each department. Constraints (33) enforce the area re-
zihj , zivj ∈ {0, 1}, i, j ∈ N. (45)
quirement for each department. Note that these constraints can be
relaxed to wi hi ≥ Ai . This relaxed form has the advantage of being Constraints (39) provide a linearization of the objective function
convex, and in fact it can be formulated as a conic constraint (see (30) above. Constraints (40) ensure that each department is within
Section 3.2). Because the optimization will push this relaxed form the facility; they differ from (35) and (36) because this formula-
towards equality, in general wi hi will equal Ai at optimality. More- tion places the origin at the bottom left-hand corner of the facility.
over Theorem 3.1 in Takouda, Anjos, and Vannelli (2005) states Constraints (41) are lower and upper bounds for the widths and
n
that if i=1 Ai = hF wF then the constraints (33) must hold at ev- heights of the departments.
ery feasible solution. Constraints (34) enforce the maximum aspect Constraints (42) are the polyhedral outer approximation on 
ratio; it is straightforward to write them as two linear inequality points of (33). This approximation was introduced in Sherali et al.
constraints. (2003) and also used in Meller et al. (2007) and Liu and Meller
The last two sets of constraints enforce the location require- (2007) (see Section 3.2.1 below). This approximation is effective
ments. Constraints (35) and (36) ensure that the departments are in practice but less efficient that using the aforementioned convex
inside the facility. Finally, constraints (37) prevent overlapping; conic relaxation that is supported by most current MILO solvers.
8 M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16

This is because wi hi ≥ Ai is equivalent to a second-order cone con- 1 1


xi + wi ≤ x j − w j + wF (2 − zi+j − zi−j ), i, j ∈ N, i = j
straint: 2 2
  
1 1
− hi
wi yi + hi ≤ y j − h j + hF (1 + zi+j − zi−j ), i, j ∈ N, i = j
hi Ai
0 ⇔ w i + hi ≥ . (46) 2 2
(49)
Ai wi 2 Ai
2
zi+j , zi−j ∈ {0, 1}, i, j ∈ N. (50)
Finally, constraints (43)–(45) use the relative-location variables
zihj and zivj to prevent overlapping: depending on which of the vari- Constraints (47) ensure that every department appears exactly
ables zihj , zivj , zhji , zvji is set to 1, i is to the left of, to the right of, once in each sequence, and constraints (48) are transitivity con-
below, or above j. straints for the two sequences. Together these constraints ensure
An alternative MILO representation of the relative positions of that the binary variables represent valid sequences. Constraints
departments is given in the next section. (49) express nonoverlapping in terms of the sequence-pair vari-
ables.
Using the MILO model above with additional valid inequalities,
3.2.1. Sequence-pair formulations
including p − k symmetry-breaking constraints (see Section 5), in-
Sequence-pair approaches determine the relative positions of
stances of UA-FLP with up to 11 departments were solved to global
the departments using the so-called sequence-pair representation,
optimality in Meller et al. (2007). The computational time reached
and combine this representation with a MILO model similar to the
almost 17 hours for the 11-department instance.
one above to obtain the optimal layout. The sequence-pair rep-
Castillo and Westerlund (2005) proposed a MILO model that
resentation was first used for VLSI design in Murata, Fujiyoshi,
satisfies the area requirements within a given accuracy ε using
Nakatake, and Kajitani (1995)), and for the FLP in Meller et al.
cutting planes. We omit the details for this because the area con-
(2007) and Liu and Meller (2007).
straints can be handled more effectively using conic optimiza-
A sequence-pair is a pair of sequences of departments, denoted
tion, as mentioned above. We point out however that Castillo and
+ and − , that together encode the relative location of the de-
Westerlund (2005) used the following alternative formulation of
partments. The following theorem indicates how to translate a
nonoverlap that is essentially the sequence-pair representation:
sequence-pair into a layout.
1
Theorem 2 (Meller et al., 2007). Given a sequence-pair ( + , − ) (wi + w j ) − (xi − x j ) ≤ wF (Xi j + Yi j ), 1 ≤ i < j ≤ n (51)
2
and two departments i and j in ( + , − ), i and j satisfy the following
1
horizontal/vertical relationship in the FLP: (wi + w j ) − (x j − xi ) ≤ wF (1 + Xi j − Yi j ), 1 ≤ i < j ≤ n (52)
2
• if j succeeds i in both + and − , then j is to the right of i;
1
• if j precedes i in both + and − , then j is to the left of i; (hi + h j ) − (yi − y j ) ≤ hF (1 − Xi j + Yi j ), 1 ≤ i < j ≤ n (53)
• if j precedes i in + and succeeds i in − , then j is above i; 2
• if j succeeds i in + and precedes i in − , then j is below i. 1
(hi + h j ) − (y j − yi ) ≤ hF (2 − Xi j − Yi j ), 1 ≤ i < j ≤ n (54)
The sequence-pair structure can be incorporated in a MILO
2
model as follows. Given a sequence-pair ( + , − ) and departments
Xi j , Yi j ∈ {0, 1}, 1 ≤ i ≤ n, (55)
i and j, define the binary variables:
 where the variables Xij and Yij are a binary codification of the
1 if i precedes j in + , relative position of departments. A GA was implemented in Liu
zi+j =
0 otherwise, and Meller (2007) where each sequence-pair is a chromosome; the
 sequence-pair gives the relative position of the departments (first
1 if i precedes j in − ,
zi−j = stage), and then a LO model is solved to find the best layout (sec-
0 otherwise.
ond stage). They achieved the best results up to then for instances
with up to 35 departments; the computational time was 26 hours
These definitions lead to Theorem 3:
for the 35-department instance.
Theorem 3 (Meller et al., 2007). For any two departments i and j, A branch-and-bound algorithm that uses the sequence-pair
the following hold: representation was presented in Xie and Sahinidis (2008). A
minimum-cost network flow problem is solved to obtain a feasible
• if zi+j = 1 and zi−j = 1, then i precedes j horizontally; layout from the sequence-pair representation of the relative po-
• if zi+j = 0 and zi−j = 0, then j precedes i horizontally; sition layout. However, this is only valid for a restricted version
• if zi+j = 0 and zi−j = 1, then i precedes j vertically; of the UA-FLP in which the department widths and heights are
• if zi+j = 1 and zi−j = 0, then j precedes i vertically. fixed, and the facility has no limitations. Such an UA-FLP with just
the nonoverlapping constraints can be transformed into a network
A different MILO model (see Meller et al., 2007) for the FLP can flow problem. The advantage is that network flow algorithms can
now be formulated: be several orders of magnitude faster than general LO algorithms.
 It was observed in Meller et al. (2007) and Liu and Meller
min ci j ( u i j + vi j ) (2007) that the MILO model with constraints (43) and (45) has
1≤i< j≤n 2n(n−1 ) possible combinations of the binary variables, while the
s.t. (39 )–(42 ) sequence-pair-based MILO formulation has (n!)2 sequences gen-
zi+j + z+ji = 1, 1 ≤ i < j ≤ n erating the same set of relative-position combinations. The au-
thors claim that this difference is key to the effectiveness of the
zi−j + z−ji = 1, 1 ≤ i < j ≤ n (47) sequence-pair approach. The difference is indeed significant: using
Stirling’s approximation, we have that (n! )2 = θ (e2(n ln(n )−n ) ) while
2n(n−1 ) = e0.693(n −n ) . However, a comparison of the two formula-
2
+
zik + zk+j − zi+j ≤ 1, 1 ≤ i < j ≤ n
tions shows that another important difference is the presence of

zik + zk−j − zi−j ≤ 1, 1 ≤ i < j ≤ n (48) the transitivity constraints (48) in the sequence-pair model. It is
M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16 9

not entirely clear to what extent each of these differences con- In practice, attaining a solution with Dij ≈ Tij is not easy. The
tributes to the efficiency of the sequence-pair approach. It would approach in Anjos and Vannelli (2006) sacrifices convexity and
be interesting to carry out a computational study to clarify this proposes a modified AR model with the objective:
question.  
 Di j
min fi j (xi , x j , yi , y j ) − K log , (58)
3.3. Two-stage approaches using nonlinear optimization Ti j
1≤i< j≤n

A two-stage approach based on the attractor-repeller (AR) tech- where



nique for VLSI floorplanning was introduced in Anjos and Vannelli ⎨ci j z + ti j − 1, z ≥ Ti j
(2002). The first stage uses the AR technique to establish the rel- f i j ( xi , x j , yi , y j ) = z
ative positions of the departments, and the second stage finds a ⎩ 
feasible layout satisfying the relative positions specified by the so-
2 ci j ti j − 1 , 0 ≤ z < Ti j ,
lution to the first stage. The objective of this approach is not to and the logarithmic term steers the optimization away from solu-
achieve global optimality but rather to efficiently compute com- tions with Dij ≈ 0. Indeed the minima of this (nonconvex) func-
petitive solutions to large-scale instances of UA-FLP. tion satisfy Dij ≈ Tij . This can be viewed as a compromise in the
The AR model approximates each department by a circle with sense that convexity is lost, but computational efficiency is gained
radius ri proportional to the square root of Ai . The model places because a suitable choice of starting point and nonlinear optimiza-
the circles inside the facility while allowing some overlapping. The tion solver makes it possible to compute a solution close to these
amount of overlapping is controlled via a so-called target distance: known minima.
given α > 0, the target distance tij for circles i and j is set as ti j = In the second stage, the nonoverlapping constraints (37) are for-
α (ri + r j )2 . The AR model is: mulated as complementarity constraints. For each pair i, j, we intro-

duce new variables Xij and Yij satisfying
 Di j
min ci j Di j + f 1
x,y ti j Xi j ≥ (wi + w j ) − |xi − x j |, Xi j ≥ 0,
1≤i< j≤n
2
1 1
s.t. xi + ri ≤ wF , i = 1, . . . , n Yi j ≥ (hi + h j ) − |yi − y j |, Yi j ≥ 0,
2 2
1
xi − ri ≥ − wF , i = 1, . . . , n Xi jYi j = 0
2
1 This last constraint enforces nonoverlap by requiring that at least
yi + r i ≤ hF , i = 1 , . . . , n
2 one of Xij and Yij equal zero. Using the coordinates of the centers
1 of the circles in the optimal solution of the modified AR model to
yi − ri ≥ hF , i = 1, . . . , n,
2 initialize the nonlinear optimization solver, this approach improved
where (xi , yi ) is the center of circle i, Di j = (xi − x j )2 + (yi − y j )2 on the then-best-known solutions for large instances, in particular
and f (z ) = 1z − 1 is a penalty function. The constraints keep the for the Armour–Buffa 20-department instance.
circles inside the facility whose center is at the origin. The objec- There are some challenges with this approach so far. First a
tive function is a trade-off between the attractor term cij Dij and the nonconvex model with the repeller function 1z − 1 was proposed
D in Anjos and Vannelli (2002); this model was then modified in
repeller term f ( t i j ). Anjos and Vannelli (2006) to achieve convexity, but then the ad-
ij
While the constraints are linear, the objective function is non- dition of a new penalty term resulted again in a loss of convexity,
linear and nonconvex. It was convexified in Anjos and Vannelli though in a more controlled manner. Moreover, the optimization
D
(2002) by replacing the term ci j Di j + f ( t i j ) with the following problem with complementarity constraints is difficult to solve for
ij
piecewise function: large-scale instances.
⎧ % This motivated the significant improvements to this approach

⎪ ti j ti j carried out in Jankovits, Luo, Anjos, and Vannelli (2011). For the

⎨ci j z + z − 1, z≥
ci j first stage, fij (xi , xj , yi , yj ) is replaced by a more complicated expres-
f i j ( xi , x j , yi , y j ) = % (56) sion that also integrates information about the aspect ratio con-

⎪  ti j

⎩2 ci j ti j − 1, 0≤z< straints. We refer the reader to Jankovits et al. (2011) for details on
ci j the first stage, and instead present below the recent further im-
provements in Anjos and Vieira (2015).
where z = (xi − x j )2 + (yi − y j )2 , and it is assumed that cij > 0.
The more significant contributions in Jankovits et al. (2011) are
Note that the second branch of fij is constant, and that by con-
their improved second stage, and the linking of the two stages.
struction, fij attains its minimum whenever the positions of i and j
 They introduce the following convex second-stage model that can
satisfy Di j ≤ ti j /ci j . This includes the case where Di j = 0, i.e., the
be solved efficiently:
two circles completely overlap. Of course, such a placement  is un- 
desirable. The ideal arrangement of the circles has Di j ≈ ti j /ci j , min ( ui j + vi j ) (59)
(xi ,yi ),wi ,hi
i.e., close to the boundary of the flat portion of fij . At these points, 1≤i< j≤n

the minimum of fij is attained at the same time as the overlap is


s.t. ui j ≥ xi − x j , for 1 ≤ i < j ≤ n (60)
minimized. This motivates the introduction in Anjos and Vannelli
(2006) of a generalized target distance:
% ui j ≥ x j − xi , for 1 ≤ i < j ≤ n (61)
ti j
Ti j = , 1 ≤ i, j ≤ n, (57)
ci j +
 vi j ≥ yi − y j , for 1 ≤ i < j ≤ n (62)
where > 0 is chosen sufficiently small so that Ti j ≈ ti j /ci j . This
modification also removes the need for the assumption that cij >
vi j ≥ y j − yi , for 1 ≤ i < j ≤ n (63)
0.
10 M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16

wmin
i ≤ wi ≤ wmax
i , for 1 ≤ i < j ≤ n (64) only approach entirely based on mathematical optimization mod-
els that has been able to reach such large-scale instances of UA-
hmin ≤ hi ≤ hmax , for 1 ≤ i < j ≤ n (65) FLP.
i i

wi hi ≥ Ai , for 1 ≤ i < j ≤ n (66) 3.4. Other two-stage approaches

A heuristic based on a graph-pair representation and a simu-


βi wi − hi ≥ 0, for 1 ≤ i < j ≤ n (67)
lated annealing technique was proposed in Bozer and Wang (2012).
One of the graphs represents the horizontal separation of the de-
βi hi − wi ≥ 0, for 1 ≤ i < j ≤ n, (68) partments, and the other represents the vertical separation. Their
plus appropriately chosen linear inequality constraints to ensure results are generally good, but the authors make changes in the
nonoverlap. areas of the facilities or departments. These modifications are rea-
These nonoverlap constraints are obtained as follows. Consider sonable from a practical point of view, but they make it difficult to
the coordinates of the centers of the circles in the optimal solution compare with other techniques.
to the first stage as a set of points on the plane, and compute their A LO-based GA approach, which differs from Liu and Meller
Delaunay triangulation. One of the properties of this triangulation (2007) in the chromosome coding, was introduced in Kulturel-
is that it maximizes the minimum angle over all the angles of the Konak and Konak (2013). The idea is that the GA searches for
triangles; in practice this means that thin triangles are less likely. the relative locations of the departments, and the LO model de-
The edges of the Delaunay triangulation are taken to represent the termines their exact locations and shapes. In particular, a new lo-
relative positions of the departments, and these positions are then cation/shape representation is proposed to encode the relative lo-
enforced by the appropriate linear constraints. For example, if the cations. Specifically, the relative location of department i is rep-
centers of i and j are connected in the triangulation and j is to the resented as (xi , yi , α i ), where αi = hi /wi . For each (xi , yi , α i ), de-
right of i, then the constraint x j − xi ≥ 12 (wi + w j ) is added to the fine two straight lines, one passing through (xi , yi ) and the upper
model. The result is a second stage model that is a conic optimiza- right corner (xi + wi /2, yi + hi /2 ), and the other passing through
tion problem and can be solved efficiently. (xi , yi ) and the upper left corner (xi − wi /2, yi + hi /2 ). These lines
The overall approach in Jankovits et al. (2011) provided fur- split the facility into four regions with reference to department i,
ther improved layouts for the classical Armour–Buffa instance, and so every other department is above or below or left or right of i.
computed high-quality layouts for several 30-department instances Like the sequence-pair representation, the location/shape represen-
in 5 minutes or less of computation time. tation always generates a consistent assignment of the binary de-
Most recently, Anjos and Vieira (2015) further developed the AR cision variables. Note that while the MILO-model does not contain
concept. As the second stage of Jankovits et al. (2011) is highly ef- the transitivity constraints for the integer variables, this encoding
fective, the novelty in Anjos and Vieira (2015) is the formulation of (based on continuous variables) encapsulates transitivity. The re-
the first stage. Specifically they propose a more precise formulation sults in Kulturel-Konak and Konak (2013) show that this approach
that models the departments as rectangles instead of approximat- outperforms previous techniques: the cost function is reduced and
ing them by circles. The aspect ratio constraints can therefore be the computational time is lower.
exactly enforced at the first stage, instead of being approximated A similarly structured approach was proposed in Gonçalves and
as in Jankovits et al. (2011). They still forego convexity and use the Resende (2015) using a random-key GA in the first stage and a LO
simple objective function model in the second stage. The authors report results on several in-
θi2j stances of UA-FLP from the literature, and find slightly better solu-
ci j D2i j + K − 1, tions in a considerably shorter computational time, in comparison
D2i j with the other GA approaches. They also applied their approach to
  larger instances with up to 125 departments, but without restric-
where θi2j = 14 (wi + w j )2 + (hi + h j )2 . Note that Dij /θ ij ≈ 1 indi-
tions on the dimensions of the facility. The computational time is
cates that some of the borders of the rectangles are close, regard-
reduced because they do not solve all the LO problems originating
less of whether the rectangles are overlapping (by a small amount)
from the relative-position solutions: they only solve the problems
or not.
that provably yield a feasible solution with a cost not exceeding
The resulting first-stage model is:

40% of that of the previous best solution.
 θi2j
min ci j D2i j + K −1
xi ,yi ,hi ,wi D2i j 3.5. Flexible bay structure
1≤i< j≤n

1 1 1 1 A flexible bay structure is a continuous layout where the de-


s.t. xi + wi ≤ wF and wi − xi ≤ wF , for i = 1, . . . , n,
2 2 2 2 partments are located in parallel bays with flexible widths. This
1 1 1 1 special case of the UA-FLP arises in manufacturing facilities (Meller,
yi + hi ≤ hF and hi − yi ≤ hF , for i = 1, . . . , n,
2 2 2 2 1997). The bay structure is similar to the row structure in row FLPs,
wi hi ≥ Ai , for i = 1, . . . , n, but a fundamental difference is that the width of each bay depends
on the total area of the departments in that bay, whereas in row
β wi − hi ≥ 0, for i = 1, . . . , n,
FLPs, the heights of the rows and of the departments are equal and
β hi − wi ≥ 0, for i = 1, . . . , n, fixed. The bays have straight aisles on both sides, and departments
wmin
i ≤ wi ≤ wmax
i , for i = 1, . . . , n, are not allowed to span multiple bays. This structure restricts the
set of feasible solutions, but it has advantages in practice: the bay
hmin
i ≤ hi ≤ hmax
i , for i = 1, . . . , n. boundaries form the basis of an aisle structure that facilitates the

where K = α 1≤i< j≤n ci j , and 0 < α ≤ 1. By solving for different transfer of the layout solution to an actual facility design.
choices of α (and hence of K), the authors of Anjos and Vieira A MILO formulation for this problem was proposed in Konak,
(2015) improved on the best solutions by earlier techniques. Fur- Kulturel-Konak, Norman, and Smith (2006). The continuous vari-
thermore, they computed layouts for instances with up to 100 de- ables xi , yi represent the location of department i, and hik repre-
partments in less than 15 min of computation time. This is the sent the height of department i in bay k. The binary variables are
M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16 11

defined as follows: the bay is used, and zero if the bay is empty. Constraints (75) are
 bounds on the department heights. They also enforce hik = 0 when
1, if department i is assigned to bay k
zik = department i is not located in bay k. Constraints (76) define the
0, otherwise;
heights of the departments. Constraints (78) and (79) ensure that
1, if department i is above department j department i is either above or below department j. Constraints
ri j = in the same bay (77) prevent departments in the same bay from overlapping. Con-
0, otherwise; straints (80) ensure that the departments are inside the facility.
 By adding symmetry-breaking constraints (see Section 5) and
1, if bay k is occupied
δk = valid inequalities (see Section 6), instances with up to 14 depart-
0, otherwise.
ments were solved to optimality in Konak et al. (2006). The 14-
The MILO model is: department instance needed around 120 hours of computational
 time.
min ci j ( u i j + vi j )
1≤i< j≤n 4. Multifloor FLP
s.t. ui j ≥ xi − x j and ui j ≥ x j − xi , 1 ≤ i < j ≤ n
vi j ≥ yi − y j and vi j ≥ y j − yi , 1 ≤ i < j ≤ n The multifloor FLP (MF-FLP) involves finding the optimal ar-
 rangement of departments in a facility with multiple floors. Prac-
zik = 1, i ∈ N (69) tical applications include production facilities, hotels, office build-
k∈K ings, and hospitals. This problem has added complexity in com-
1 parison to the UA-FLP because we must also consider the interac-
wk = hF i∈N zik Ai , k∈K (70)
tions between departments on different floors. Furthermore, eleva-
tors and/or stairwells are required to transfer people and/or mate-
wmin
i zik ≤ wk ≤ wmax
i + wF (1 − zik ), k ∈ K, i ∈ N (71) rial between the floors, and these need to be placed at coherent
locations in every floor that they reach.

xi ≥ w j − 0.5wk − (wF − wmin
i )(1 − zik ), k ∈ K, i ∈ N Globally optimal algorithms for MF-FLP work in general only
j≤k for small instances (Hahn, Smith, & Zhu, 2010). The problem was
 first investigated in Johnson (1982) and later in Meller and Bozer
xi ≤ w j − 0.5wk + (wF − wmin
i )(1 − zik ), k ∈ K, i ∈ N (72) (1997), but most of the subsequent models in the literature are
j≤k
designed for specific types of MF-FLP, as the literature survey in
  Section 4.1 shows. Indeed there is no commonly agreed definition
hik h jk max min
j
− − max i
, (2 − zik − z jk ) ≤ 0, i < j of the MF-FLP because different authors make their own assump-
Ai aj Ai aj
tions about the structure of the problem. This lack of a common
 
hik h jk max min
j definition makes it hard to compare the approaches. We therefore
− + max i
, (2 − zik − z jk ) ≤ 0, i < j (73) propose in Section 4.2 a general formulation for the MF-FLP that
Ai aj Ai aj
 we hope will gain acceptance as a standard formulation, and will
hik = hF δk , k ∈ K (74) lead to increased research activity on this problem.
i∈N
4.1. Survey of the literature
hmin
i zik ≤ hik ≤ hmax
i zik , i ∈ N, k ∈ K (75)
Some approaches first distribute the departments over the

hik = hi , i ∈ N (76) floors, minimizing the vertical interaction costs. This is essentially
i∈N the first stage of a two-stage approach, where the second stage
then optimizes the layout of each floor independently; see Meller
yi − 0.5 hi ≥ y j + 0.5 h j − wH ( 1 − ri j ), i = j (77) and Bozer (1997) and Bernardi and Anjos (2013). Specifically the
following MILO formulation is used in Meller (1997) to assign de-
ri j + r ji = 1, 1 ≤ i < j ≤ n (78) partments to floors:

min ci j divj (81)
ri j + r ji ≥ zik + z jk − 1, 1 ≤ i < j ≤ n, k ∈ K (79) 1≤i< j≤n


p

0.5hi ≤ yi ≤ wH − 0.5hi , i ∈ N. (80) s.t. zik = 1, 1 ≤ i ≤ n (82)


k=1
where K is the set of bays, and N is the set of departments.

p
Constraints (69) ensure that each department is assigned to divj ≥ δ k(zik − z jk ), 1 ≤ i < j ≤ n (83)
a single bay. Constraints (70) calculate the width of each bay as k=1
the total area of the departments assigned to that bay divided by

p
the facility height. Note that under the assumption that i∈N Ai ≤ divj ≥ δ k(z jk − zik ), 1 ≤ i < j ≤ n (84)

wF hF , we have k∈K wk ≤ wF . Constraints (71) impose bounds on k=1
the bay widths, based on the width bounds of the departments as-

n
signed to each bay. Constraints (72) determine the horizontal loca- Ai zik ≤ wF hF , 1 ≤ k ≤ p (85)
tions of the department centroids. In this model, the x-coordinate i=1
is located at the middle of the bays. Therefore, if department i

is assigned to bay k, xi is calculated as xi = kj=1 w j − 0.5wk . This
zik ∈ {0, 1}, 1 ≤ i ≤ n, 1 ≤ k ≤ p, (86)
agrees with constraints (72) with zik = 1. If i and j are in the same where p is the number of floors. The variable zik equals 1 if de-
bay k, then constraints (73) ensure that the widths of the two de- partment i is assigned to floor k, and equals 0 otherwise. Con-
partments are the same and equal to the bay width. Constraints straints (82) assign each department to exactly one floor. Con-
(74) set the total heights of the departments in a bay equal to hF if straints (83) and (84) compute the vertical distance divj between
12 M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16

each pair i, j of departments, where δ is the floor height. Note Let δ denote the ceiling height, p the number of floors, and e
that the number of elevators. Let also M = wF + hF + δ p. Define the fol-
& & lowing variables:
& p &
& &
& k(zik − z jk )& zik = 1 if department i is assigned to floor k, 0 otherwise;
& k=1 &
Zi j = 1 if departments i and j are allocated to the same floor, 0
otherwise;
is precisely equal to the number of floors separating i and j. Con-
Xij , Yij : nonoverlapping binary variables;
straints (85) ensure that the departments assigned to each floor fit
(xi , yi ): coordinates of the centroid of department i;
into that floor.
divj : vertical distance between i and j;
Each floor then becomes an instance of UA-FLP with some ad-
ditional constraints to ensure coherence in the location of the el- dihj : horizontal distance between i and j located on the same
evators. Computing the vertical costs still remains a challenge and floor;
was addressed in Bernardi and Anjos (2013). diej : horizontal distance between i and j located on different
Another possible simplification is to restrict all the departments floors, where the path includes an elevator.
to have the same shape and to require that they be assigned to
specific locations in the building. This reduces the problem to a Note that the indices n + 1, . . . , n + e correspond to the eleva-
quadratic assignment problem. Such a formulation was used in tors.
Hahn et al. (2010), and was solved using the RLT linearization tech- The formulation is as follows:
nique (Adams & Sherali, 1986; Sherali & Adams, 1990) within a

min ci j (diej + divj )
branch-and-bound algorithm.
1≤i< j≤n
A mathematical formulation of MF-FLP for process plant lay-
out was presented in Patsiatzis and Papageorgiou (2002). Its ob- 
p
s.t. zik = 1, 1 ≤ i ≤ n (87)
jective function considers the construction and land costs to de-
k=1
cide the number of floors and the floor area. Another model for & &
a processing plant was proposed in Defersha and Chen (2006); it & p &
& &
incorporates many structural and operational issues, but becomes divj = δ & k(zik − z jk )&, 1 ≤ i < j ≤ n
unwieldy.
& k=1 &
A GA is also used in Lee, Roh, and Jeong (2005) to find a lay- dihj = |xi − x j | + |yi − y j |, 1 ≤ i < j ≤ n
out with inner walls and passages. The connections between the
departments, passages, and elevators are represented as an adja- diej ≥ dihj , 1 ≤ i < j ≤ n
cency graph, and the distances are calculated using Dijkstra’s al- diej ≥ |xi − x | + |yi − y | + |x j − x | + |y j − y | − MZi j ,
gorithm. This representation allows the measurement of the dis-
tances of paths that use corridors and elevators. The bi-objective 1 ≤ i < j ≤ n, n + 1 ≤  ≤ n + e (88)
model minimizes the total cost of transporting the materials and
maximizes the adjacency achieved. It is applied to a multideck ship Zi j ≥ zik + z jk − 1, 1 ≤ i < j ≤ n, k = 1, . . . , p
layout with inner walls.
Zi j ≤ 1 − zik + z jk , 1 ≤ i < j ≤ n, k = 1, . . . , p
Another bi-objective model is proposed in Hathhorn, Sisikoglu,
and Sir (2013) for a MF-FLP formulation that minimizes not only Zi j ≤ 1 + zik − z jk , 1 ≤ i < j ≤ n, k = 1, . . . , p (89)
the material handling costs (as usual) but also the facility con-
struction costs. This model is similar to the one we present in 1 1 1 1
xi + wi ≤ wF , xi − wi ≥ − wF , 1 ≤ i ≤ n + e
Section 4.2, but a major difference is that the length and width 2 2 2 2
of the facility, the number of elevators, and the number of floors 1 1 1 1
yi + hi ≤ hF , yi − hi ≥ − hF , 1 ≤ i ≤ n + e (90)
are decision variables. 2 2 2 2
For completeness, we also mention the robust model in
Izadinia, Eshghi, and Salmani (2014) in which some of the usual w i hi = Ai , 1 ≤ i ≤ n
parameters are considered to be uncertain, and the model in Park, w i − β hi ≤ 0 , hi − β w i ≤ 0, 1 ≤ i ≤ n (91)
Koo, Shin, Lee, and Yoon (2011) that takes into account safety dis-
tances in the event of an explosion.
1
xi − x j ≥ (wi + w j ) − wF (1 − Zi j + Xi j + Yi j ), 1 ≤ i < j ≤n+e
2
4.2. A MF-FLP formulation 1
x j − xi ≥ (wi + w j ) − wF (2 − Zi j − Xi j + Yi j ), 1 ≤ i < j ≤n+e
2
We assume that the following parameters are given: the num- 1
yi − y j ≥ (hi + h j ) − hF (2 − Zi j + Xi j − Yi j ), 1 ≤ i < j ≤n+e
ber of departments and their areas, the number of floors, the di- 2
mensions and height of the floors, the interconnection costs, and 1
the number and size of the elevators. We consider the elevators to y j − yi ≥ (hi + h j ) − hF (3 − Zi j − Xi j − Yi j ), 1 ≤ i < j ≤n+e
2
be a general system (incorporating elevators, stairs, pipes, etc.) for (92)
vertical movement. We want to determine the locations of the el-
evators and the locations and dimensions of the departments. The
horizontal distance is the rectilinear distance (which is a reliable zik = 1, n + 1 ≤ i ≤ n + e, 1 ≤ k ≤ p
measure, as in the single-floor case), and the vertical distance will Zi j = 1, n + 1 ≤ i < j ≤ n + e (93)
be measured using the elevators. This makes the formulation com-
plex. The number of floors and elevators is assumed to be fixed;
Xi j , Yi j , Zi j , zik ∈ {0, 1}, 1 ≤ i < j ≤ n + e, 1 ≤ k ≤ p (94)
if necessary, we could run the model for several different options.
The floor dimensions are fixed, but they could easily be treated as
decision variables. hi , wi , ≥ 0, 1 ≤ i ≤ n. (95)
M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16 13

Constraints (87) allocate each department to exactly one floor. the symmetry-breaking constraints, they are mostly specific to the
Constraints (88) compute the distances between each pair of de- problem at hand. A noteworthy exception are the transitivity con-
partments; if two departments are on different floors, the distance straints, often called triangle inequalities, introduced for the first
depends on the elevator position. Constraints (89) set Zi j = 1 if i in this review in the form (1)–(4), and mentioned subsequently
and j are on the same floor, and 0 otherwise. Constraints (92) pre- throughout, see e.g. constraints (6), (9) and (48). Transitivity can
vent the overlapping of departments and elevators on the same be applied to nearly every variant of the FLP.
floor. Constraints (89) and (92) have been taken from Patsiatzis For row layout problems, some valid inequalities have been
and Papageorgiou (2002). Constraints (93) ensure that each eleva- proposed for the SRFLP. Proposition 1 contains a description of
tor covers all the floors and every pair of elevators shares the same valid inequalities, and Amaral and Letchford (2013) presented sev-
floor. eral large classes of valid inequalities. For the DRFLP, the inequal-
ities (15) in Section 2.2.1 are redundant but may be helpful for a
5. Symmetry-breaking constraints branching algorithm; hence they can be viewed as valid inequali-
ties. However, very little is known with respect to valid inequalities
Many versions of the FLP have symmetric solutions. For exam- for DRFLP and MRFLP.
ple, it is clear that flipping a solution to UA-FLP by 180 degrees Meller et al. (1999) were the first to investigate valid inequali-
gives exactly the same solution. This matters because the pres- ties for the UA-FLP. The inequalities reduced the number of nodes
ence of symmetry is often problematic when solving mixed inte- in the branch and bound tree but increased the computational
ger optimization problems. We briefly summarize here the main time. Sherali et al. (2003) determined that the best results were
symmetry-breaking strategies in the literature, primarily from the obtained by incorporating only the B2 and V2 constraints of Meller
point of view of the UA-FLP because this is the problem for which et al. (1999). Using the notation of model (38)–(45), these inequal-
they have most been used. However the strategies can be extended ities are
in a straightforward manner to many of the MILO models dis-
cussed in this review. (B2) ui j ≥ (wmin
i + wmin
j )(zihj + zhji )
One way to break the symmetry in the UA-FLP (Meller et al., (B2) vi j ≥ (hmin
i + hmin
j )(zivj + zvji )
1999) is to require some department k to be located in a spe-
cific quarter of the facility by adding the pair of constraints xk ≤ (V2) ui j ≥ (wi + w j ) − min{wmax
i + wmax
j , wF }(1 − zihj − zhji )
0.5wF , yk ≤ 0.5hF (where it is assumed that the origin is at the (V2) vi j ≥ (wi + w j ) − min{hmax
i + hmax
j , wF }(1 − zivj − zvji ).
bottom left corner of the facility). This is called the position k
method in Sherali et al. (2003). However, if department k has its These constraints do not reduce the feasible set of the relaxed
centroid located at the facility centroid, then this method does not MILO model because they are redundant, and they do not enforce
work. It is straightforward to extend this method to multirow and the separation of the departments. They are useful in branch and
multifloor layouts. bound algorithms because they improve the lower bounds. In the
An alternative strategy is the position p − k method (Sherali linear relaxation, if
et al., 2003) that considers a given pair of departments p and k zihj , zhji = (wF − wi − w j )/wF and zivj , zvji = (hF − hi − h j )/hF
and requires the centroid of p to be below and to the left of the
centroid of k by adding the following four constraints: then (44) leads to xi ≈ xj , yi ≈ yj , i.e., departments i and j overlap.
h v = 0, and Thus the root lower bound of (38)–(45) is typically zero.
x p ≤ xk , y p ≤ yk , zkp = zkp
Taking this into account, Sherali et al. (2003) model the con-
straint ui j = |xi − x j | in an unusual way. Define the variables
(xk − x p ) + (yk − y p ) ≥ min{wmin + wmin min
p , hk p }.
+ hmin
k 
The departments p and k can be chosen in different ways; a com- 1 if xi ≤ x j ,
tihj =
mon criterion is to choose them to satisfy c pk = maxi, j∈N ci j . It is 0 if xi ≥ x j ,
claimed in Castillo and Westerlund (2005) that simply choosing
where the choice of 0 or 1 is inconsequential when xi = x j . An
departments 1 and 2 works just as well, and there the constraints
x1 − x2 ≥ 0 and y2 − y1 ≥ 0 are used. From Sherali et al. (2003), it upper bound on uij is Ui j = wF − wmin i
− wmin
j
, and it is proved in
is not clear whether the position k method or the position p − k Sherali et al. (2003) that ui j = |xi − x j | can be modeled by
method is better. For the DRFLP, the p − k method was used in
0 ≤ ui j + xi − x j ≤ 2Ui j (1 − ti j ), i < j
Amaral (2013a) with p and k chosen such that c pk = mini, j∈N ci j .
Finally, several classes of hierarchical constraints that are ap- 0 ≤ ui j − xi + x j ) ≤ 2Ui j ti j , i < j
plicable to general symmetric MILO problems were considered in ti j ∈ {0, 1}, i < j.
Sherali and Smith (2001) for the UA-FLP. Those authors study the
A similar set of inequalities exists for vi j = |yi − y j |.
effect of one such class of constraints:
An alternative set of nonoverlapping constraints is proposed in

n 
n
Sherali et al. (2003):
4 ixi ≤ n(n + 1 )wF , 4 yi ≤ n ( n + 1 )hF .
i=1 i=1 xi − x j ≥ wi + w j − Mi j (1 − zihj )
They find that these constraints can break the symmetry effec- yi − y j ≥ hi + h j − Mi j (1 − zivj )
tively, but their dense structure renders the CPLEX enumeration
procedure relatively ineffective. − (wF − wmin
i − wmin
j ) ≤ xi − x j ≤ wF − wmin
i − wmin
j

− (hF − hmin
i − hmin
j ) ≤ yi − y j ≤ hF − hmin
i − hmin
j
6. Valid inequalities
wmin
i + wmin
j ≤ wi + w j ≤ wmax
i + wmax
j
As already mentioned, valid inequalities are essential for solv-
hmin
i + hmin
j ≤ hi + h j ≤ hmax
i + hmax
j
ing mathematical optimization models efficiently in practice, espe-
cially MILO problems. In this Section we gather a number of valid zihj + zhji + zivj + zvji = 1
inequalities used in the literature to improve MILO models. As in
zihj(v ) ∈ {0, 1}.
Section 5, these results are mostly about the UA-FLP, but unlike
14 M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16

They construct a convex-hull representation of the above con- Section 4.2 that we hope will gain acceptance in the community
straint set in a higher dimensional space. This convex hull can also and will motivate further research into this most challenging
be derived using the reformulation-linearization technique (RLT) version of facility layout. At the very least, we hope that this
of Sherali, Adams, and Driscoll (1998). Because of the size of this review prompts a discussion of the assumptions that should be
representation, they use it for just one pair of departments, the made in defining a standard version of the problem. This could
positively interacting (nonfixed) pair with the largest total area. then lead not only to the development of novel models and
Using this, together with constraints (B2) and (V2), symmetry- solution techniques, including classes of valid inequalities, but also
breaking constraints (see Section 5), and a new branching priority to more effective comparisons of them, which is essential to help
rule, Sherali et al. (2003) solve instances with up to 9 departments the research community make further progress on this difficult
to global optimality. Meller et al. (2007) use inequalities (B2) and but important problem.
(V2) in the context of the sequence-pair representation formula- Finally, symmetry remains a key issue for the computational so-
tion, plus symmetry-breaking constraints, and the same branch- lution of FLPs. General methods for handling symmetry in MILO,
ing priority rule to solve instances with 11 departments within 24 such as isomorphism pruning (Margot, 20 02; 20 03) and orbital
hours. branching (Ostrowski, Linderoth, Rossi, & Smriglio, 2009), have
Finally, symmetry-avoidance constraints and a tightening of the proven advantageous for general problems with general symme-
nonoverlapping constraints via try groups. Problem-specific techniques have also been proposed,
e.g. orbitopal fixing (Kaibel, Peinhardt, & Pfetsch, 2011; Kaibel &
1
(wi + w j ) − ui j ≤ wF Xi j , 1 ≤ i < j ≤ n Pfetsch, 2008) is an efficient way to break symmetry in bin packing
2
problems, and modified orbital branching was show to be effective
1
(hi + h j ) − vi j ≤ hF (1 − Xi j ), 1 ≤ i < j ≤ n for for problems with structured symmetry via the unit commit-
2 ment problem (Anjos, 2013; Ostrowski, Anjos, & Vannelli, 2015). It
are used in Castillo and Westerlund (2005). Note that these are the remains to be seen how these results may have an impact for the
same as (B2) and (V2). solution of certain classes of facility layout.

7. Directions for future research


Acknowledgment
Facility layout continues to be the focus of much research, as
evidenced for example in the bibliography of this review that The authors are grateful to the anonymous reviewers for their
includes more than 30 research articles published since 2010. careful reading of an earlier version of this paper, and for their
We covered three classes of layout problems, namely row lay- helpful feedback that allowed us to markedly improve the paper.
out, unequal-areas layout, and multifloor layout, primarily from the Anjos’s research was partially supported by the Natural Sciences
perspective of mathematical optimization techniques, mostly MILO and Engineering Research Council of Canada (Discovery Grant
and SDO. 312125). Vieira’s research was partially supported by the Fundação
Within row layout problems, the single-row FLP is the one most para a Ciência e a Tecnologia (Portuguese Foundation for Science
studied in the literature, and major progress has been achieved in and Technology) through the project PEstOE/MAT/UI0297/2014
recent years since the first papers that modeled and solved this (Centro de Matemática e Aplicações).
problem using SDO (Anjos et al., 2005; Anjos & Vannelli, 2008).
The recent review in this journal by Keller and Buscher (2015) pro- References
vides a detailed exposition of this progress. The double-row and
multirow cases have also attracted attention recently, but have Adams, W., & Sherali, H. (1986). A tight linearization and an algorithm for zero-one
generally been less studied. The multirow FLP in particular sets a quadratic-programming problems. Management Science, 32(10), 1274–1290.
Ahonen, H., de Alvarenga, A., & Amaral, A. (2014). Simulated annealing and tabu
challenge to the research community in terms of providing new search approaches for the corridor allocation problem. European Journal of Op-
ideas for models and algorithms, including new classes of valid in- erational Research, 232(1), 221–233.
equalities, to compute global solutions. Amaral, A. R. S. (2006). On the exact solution of a facility layout problem. European
Journal of Operational Research, 173(2), 508–518.
Unequal-area layout is by far the most studied class of FLPs.
Amaral, A. R. S. (2008). An exact approach to the one-dimensional facility layout
Nevertheless only instances with up to 11 departments have been problem. Operations Research, 56(4), 1026–1033.
solved to global optimality using MILO formulations. Other math- Amaral, A. R. S. (2009a). A mixed 0–1 linear programming formulation for the exact
ematical optimization-based research has focused on two-stage solution of the minimum linear arrangement problem. Optimization Letters, 3,
513–520.
heuristics that can provide good solutions for instances with up Amaral, A. R. S. (2009b). A new lower bound for the single row facility layout prob-
to 100 departments. Moreover, unlike for row FLPs, little work has lem. Discrete Applied Mathematics, 157(1), 183–190.
been done with respect to applying semidefinite optimization to Amaral, A. R. S. (2011). On duplex arrangement of vertices. Brazil: Departamento de
Informática, Universidade Federal do Espírito Santo (UFES) Technical report.
this class of problems beyond the observation that the department Amaral, A. R. S. (2012). The corridor allocation problem. Computers & Operations
area constraint can be relaxed in a conic form (see Eq. (46)) that Research, 39(12), 3325–3330.
is handled efficiently by conic optimization software. Indeed, to Amaral, A. R. S. (2013a). Optimal solutions for the double row layout problem. Op-
timization Letters, 7(2), 407–413.
the best of our knowledge, the only approach entirely based on Amaral, A. R. S. (2013b). A parallel ordering problem in facilities layout. Computers
semidefinite optimization was given in Takouda et al. (2005) where & Operations Research, 40(12), 2930–2939.
it is applied to obtain global bounds for benchmark instances in Amaral, A. R. S., Caprara, A., Letchford, A. N., & Salazar, J. J. (2008). A new lower
bound for the minimum linear arrangement of a graph. Electronic Notes in Dis-
the area of VLSI floorplanning. The important question of comput-
crete Mathematics, 30, 87–92.
ing global bounds for instances of UA-FLP thus remains open. Amaral, A. R. S., & Letchford, A. N. (2011). Linear arrangement polytopes. UK: Lan-
Multifloor layout has received the least attention in the litera- caster University Working paper.
Amaral, A. R. S., & Letchford, A. N. (2013). A polyhedral approach to the single row
ture. While there has been an increased interest in it in different
facility layout problem. Mathematical Programming, 141(1–2), 453–477.
contexts, most of the models are motivated by application-specific Anjos, M. F. (2013). Recent progress in modeling unit commitment problems. In
assumptions that are not commonly used in the literature. This L. Zuluaga, & T. Terlaky (Eds.), Modeling and optimization: Theory and applica-
state of affairs makes it difficult to compare the performance tions: Vol. 84. Springer. Selected contributions from the MOPTA 2012 conference,
springer proceedings in mathematics & statistics.
of different approaches. Because there is no commonly agreed Anjos, M. F. (2015). FLPLIB: Facility layout database. URL http://www.miguelanjos.
definition for the MF-FLP, we proposed a general formulation in com/flplib Accessed 04.03.17.
M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16 15

Anjos, M. F., Fischer, A., & Hungerländer, P. (2015). Solution approaches for equidistant Jankovits, I., Luo, C., Anjos, M. F., & Vannelli, A. (2011). A convex optimisation frame-
double- and multi-row facility layout problems. Cahier du GERAD G-2015-06, 33 work for the unequal-areas facility layout problem. European Journal of Opera-
pages. tional Research, 214(2), 199–215.
Anjos, M. F., Kennings, A., & Vannelli, A. (2005). A semidefinite optimization ap- Johnson, R. V. (1982). Spacecraft for multi-floor layout planning. Management Sci-
proach for the single-row layout problem with unequal dimensions. Discrete Op- ence, 28(4), 407–417.
timization, 2(2), 113–122. Kaibel, V., Peinhardt, M., & Pfetsch, M. E. (2011). Orbitopal fixing. Discrete Optimiza-
Anjos, M. F., & Liers, F. (2012). Global approaches for facility layout and VLSI tion, 8(4), 595–610.
floorplanning. In Handbook on semidefinite, conic and polynomial optimization Kaibel, V., & Pfetsch, M. (2008). Packing and partitioning orbitopes. Mathematical
(pp. 849–877). Springer. Programming, 114, 1–36.
Anjos, M. F., & Vannelli, A. (2002). An attractor-repeller approach to floorplanning. Keller, B., & Buscher, U. (2015). Single row layout models. European Journal of Oper-
Mathematical Methods of Operations Research, 56(1), 3–27. ational Research, 245(3), 629–644.
Anjos, M. F., & Vannelli, A. (2006). A new mathematical-programming framework Komarudin, K., & Wong, K. Y. (2010). Applying ant system for solving unequal
for facility-layout design. INFORMS Journal on Computing, 18(1), 111–118. area facility layout problems. European Journal of Operational Research, 202(3),
Anjos, M. F., & Vannelli, A. (2008). Computing globally optimal solutions for sin- 730–746.
gle-row layout problems using semidefinite programming and cutting planes. Konak, A., Kulturel-Konak, S., Norman, B. A., & Smith, A. E. (2006). A new mixed
INFORMS Journal on Computing, 20(4), 611–617. integer programming formulation for facility layout design using flexible bays.
Anjos, M. F., & Vieira, M. V. (2015). An improved two-stage optimization-based Operations Research Letters, 34(6), 660–672.
framework for unequal-areas facility layout. Optimization Letters, 10(7), Kothari, R., & Ghosh, D. (2012). The single row facility layout problem: State of the
1379–1392. art. Opsearch, 49(4), 442–462.
Armour, G., & Buffa, E. (1963). A heuristic algorithm and simulation approach to Kulturel-Konak, S., & Konak, A. (2013). Linear programming based genetic algorithm
relative location of facilities. Management Science, 9(2), 294–309. for the unequal area facility layout problem. International Journal of Production
Bernardi, S., & Anjos, M. F. (2013). A two-stage mathematical-programming method Research, 51(14), 4302–4324.
for the multi-floor facility layout problem. JORS, 64(3), 352–364. Kusiak, A. (1990). Intelligent Manufacturing Systems p. 448. 200 Old Tappan Road,
Bozer, Y. A., & Wang, C. (2012). A graph-pair representation and MIP-model-based Old Tappan, NJ 07675, USA: Prentice Hall Press.
heuristic for the unequal-area facility layout problem. European Journal of Oper- Laporte, G., & Mercure, H. (1988). Balancing hydraulic turbine runners: A
ational Research, 218(2), 382–391. quadratic assignment problem. European Journal of Operational Research, 35(3),
Brusco, M., & Stahl, S. (20 0 0). Using quadratic assignment methods to generate ini- 378–381.
tial permutations for least-squares unidimensional scaling of symmetric prox- Lee, K., Roh, M., & Jeong, H. (2005). An improved genetic algorithm for multi-floor
imity matrices. Journal of Classification, 17(2), 197–223. facility layout problems having inner structure walls and passages. Computers &
Castillo, I., & Westerlund, T. (2005). An epsilon-accurate model for optimal un- Operational Research, 32, 879–899.
equal-area block layout design. Computers & OR, 32, 429–447. Liu, Q., & Meller, R. D. (2007). A sequence-pair representation and MIP-model-based
Chang, M.-S., & Ku, T.-C. (2013). A slicing tree representation and QCP-model-based heuristic for the facility layout problem with rectangular departments. IIE Trans-
heuristic algorithm for the unequal-area block facility layout problem. Math- actions, 39(4), 377–394.
ematical Problems in Engineering, 2013, 19. doi:10.1155/2013/853586. Article ID Liu, W., & Vannelli, A. (1995). Generating lower bounds for the linear arrangement
853586 problem. Discrete Applied Mathematics, 59(2), 137–151.
Chung, J., & Tanchoco, J. M. A. (2010). The double row layout problem. International Loiola, E., de Abreu, N., Boaventura-Netto, P., Hahn, P., & Querido, T. (2007). A survey
Journal of Production Research, 48(3), 709–727. for the quadratic assignment problem. European Journal of Operational Research,
Coll, P. E., Ribeiro, C. C., & de Souza, C. C. (2006). Multiprocessor scheduling under 176(2), 657–690.
precedence constraints: Polyhedral results. Discrete Applied Mathematics, 154(5), Love, R. F., & Wong, J. Y. (1976). On solving a one-dimensional allocation problem
770–801. with integer programming. INFOR, 14, 139–143.
Defersha, F. M., & Chen, M. (2006). A comprehensive mathematical model for the Luo, C., Anjos, M., & Vannelli, A. (2008). Large-scale fixed-outline floorplanning de-
design of cellular manufacturing systems. International Journal of Production Eco- sign using convex optimization techniques. In Proceedings of the 2008 IEEE de-
nomics, 103(2), 767–783. sign automation conference (ASPDAC) (pp. 198–203).
Diego-Mas, J. A., Santamarina-Siurana, M. C., Cloquell-Ballester, V. A., & Alcaide– Margot, F. (2002). Pruning by isomorphism in branch-and-cut. Mathematical Pro-
Marzal, J. (2008). Slicing tree’s geometric potential: An indicator for layout gramming, 94, 71–90.
problems based on slicing tree structure. International Journal of Production Re- Margot, F. (2003). Exploiting orbits in symmetric ILP. Mathematical Programming, Se-
search, 46(4), 1071–1087. ries B, 98, 3–21.
Fischer, I., Gruber, G., Rendl, F., & Sotirov, R. (2006). Computational experience with Meller, R. (1997). The multi-bay manufacturing facility layout problem. International
a bundle approach for semidefinite cutting plane relaxations of max-cut and Journal of Production Research, 35(5), 1229–1237.
equipartition. Mathematical Programming, 105(2–3), 451–469. Ser. B Meller, R., & Bozer, Y. (1997). Alternative approaches to solve the multi-floor facility
Garey, M. R., Johnson, D. S., & Stockmeyer, L. (1974). Some simplified NP-complete layout problem. Journal of Manufacturing Systems, 16(3), 192–203.
problems. In Proceedings of the sixth annual ACM symposium on theory of com- Meller, R., & Gau, K.-Y. (1996). The facility layout problem: Recent and emerging
puting (STOC ’74) (pp. 47–63). trends and perspectives. Journal of Manufacturing Systems, 15(5), 351–366.
Gen, M., & Cheng, R. (1997). Genetic algorithms and engineering design. New York: Meller, R. D., Chen, W., & Sherali, H. D. (2007). Applying the sequence-pair repre-
Wiley. sentation to optimal facility layout designs. Operations Research Letters, 35(5),
Gonçalves, J., & Resende, M. (2015). A biased random-key genetic algorithm for the 651–659.
unequal area facility layout problem. European Journal of Operational Research, Meller, R. D., Narayanan, V., & Vance, P. H. (1999). Optimal facility layout design.
246(1), 86–107. Operations Research Letters, 23(3–5), 117–127.
Hahn, P. M., Smith, J. M., & Zhu, Y. (2010). The multi-story space assignment prob- Montreuil, B. (1991). A modelling framework for integrating layout design and
lem. Annals OR, 179(1), 77–103. flow network design. In J. A. White, & I. W. Pence (Eds.), Material handling. In
Hassan, M. (1994). Machine layout problem in modern manufacturing facilities. In- Progress in material handling and logistics: vol. 2 (pp. 95–116). Springer-Verlag.
ternational Journal of Production Research, 32(11), 2559–2584. Murata, H., Fujiyoshi, K., Nakatake, S., & Kajitani, Y. (1995). Rectangle-packing-based
Hathhorn, J., Sisikoglu, E., & Sir, M. Y. (2013). A multi-objective mixed-integer pro- module placement. In Proceedings of the 1995 IEEE/ACM international conference
gramming model for a multi-floor facility layout. International Journal of Produc- on computer-aided design (pp. 472–479).
tion Research, 51(14), 4223–4239. Murray, C. C., Smith, A. E., & Zhang, Z. (2013). An efficient local search heuristic
Heragu, S. (2008). Facilities design (3rd). CRC Press. for the double row layout problem with asymmetric material flow. International
Heragu, S. S., & Kusiak, A. (1988). Machine layout problem in flexible manufacturing Journal of Production Research, 51(20), 6129–6139.
systems. Operations Research, 36(2), 258–268. Ostrowski, J., Anjos, M. F., & Vannelli, A. (2015). Modified orbital branching for struc-
Heragu, S. S., & Kusiak, A. (1991). Efficient models for the facility layout problem. tured symmetry with an application to unit commitment. Mathematical Pro-
European Journal of Operational Research, 53(1), 1–13. gramming, 150(1), 99–129.
Hungerländer, P. (2014). Single-row equidistant facility layout as a special case Ostrowski, J., Linderoth, J., Rossi, F., & Smriglio, S. (2009). Orbital branching. Mathe-
of single-row facility layout. International Journal of Production Research, 52(5), matical Programming, 126(1), 147–178.
1257–1268. Otten, R. H. J. M. (1982). Automatic floorplan design. In J. S. Crabbe, C. E. Radke, &
Hungerländer, P. (2014). A semidefinite optimization approach to the parallel row or- H. Ofek (Eds.), Proceedings of the ACM/IEEE nineteenth design automation confer-
dering problem. Alpen-Adria Universitt Klagenfurt, Mathematics, Optimization ence (DAC ’82), 14–16 June (pp. 261–267). Las Vegas, Nevada, USA,.
Group Technical report tr-aruk-m-o-14-05. Palubeckis, G. (2012). A branch-and-bound algorithm for the single-row equidistant
Hungerländer, P., & Anjos, M. F. (2012). A semidefinite optimization approach to facility layout problem. OR Spectrum, 334, 1–21.
space-free multi-row facility layout. Cahier du GERAD G-2012-03, 20 pages. Park, K., Koo, J., Shin, D., Lee, C. J., & Yoon, E. S. (2011). Optimal multi-floor plant
Hungerländer, P., & Anjos, M. F. (2015). A semidefinite optimization-based approach layout with consideration of safety distance based on mathematical program-
for global optimization of multi-row facility layout. European Journal of Opera- ming and modified consequence analysis. Korean Journal of Chemical Engineer-
tional Research, 245(1), 46–61. ing, 28(4), 1009–1018.
Hungerländer, P., & Rendl, F. (2013). A computational study and survey of meth- Patsiatzis, D., & Papageorgiou, L. (2002). Optimal multi-floor process plant layout.
ods for the single-row facility layout problem. Computational Optimization and Computers & Chemical Engineering, 26(4–5), 575–583.
Applications, 55(1), 1–20. Sanjeevi, S., & Kianfar, K. (2010). A polyhedral study of triplet formulation for
Izadinia, N., Eshghi, K., & Salmani, M. H. (2014). A robust model for multi-floor lay- single row facility layout problem. Discrete Applied Mathematics, 158(16),
out problem. Computers & Industrial Engineering, 78, 127–134. 1861–1867.
16 M.F. Anjos, M.V.C. Vieira / European Journal of Operational Research 261 (2017) 1–16

Scholz, D., Petrick, A., & Domschke, W. (2009). STaTS: A slicing tree and tabu search Singh, S., & Sharma, R. (2006). A review of different approaches to the facility lay-
based heuristic for the unequal area facility layout problem. European Journal of out problems. The International Journal of Advanced Manufacturing Technology,
Operational Research, 197(1), 166–178. 30(5–6), 425–433.
Shayan, E., & Chittilappilly, A. (2004). Genetic algorithm for facilities layout prob- Takouda, P., Anjos, M., & Vannelli, A. (2005). Global lower bounds for the VLSI
lems based on slicing tree structure. International Journal of Production Research, macrocell floorplanning problem using semidefinite optimization. In Proceedings
42(19), 4055–4067. of the 2005 IWSOC (pp. 275–280).
Sherali, H. D., & Adams, W. (1990). A hierarchy of relaxations between the continu- Tam, K. (1992). Genetic algorithms, function optimization, and facility layout design.
ous and convex hull representations for zero-one programming problems. SIAM European Journal of Operational Research, 63(2), 322–346.
Journal on Discrete Mathematics, 3(3), 411–430. van Camp, D. J., Carter, M. W., & Vannelli, A. (1991). A nonlinear optimization ap-
Sherali, H. D., Adams, W., & Driscoll, P. (1998). Exploiting special structures in con- proach for solving facility layout problems. European Journal of Operational Re-
structing a hierarchy of relaxations for 0–1 mixed integer problems. Operations search, 57, 174–189.
Research, 46(3), 396–405. Wess, B., & Zeitlhofer, T. (2004). On the phase coupling problem between data
Sherali, H. D., Fraticelli, B. M. P., & Meller, R. D. (2003). Enhanced model formula- memory layout generation and address pointer assignment. In H. Schepers (Ed.),
tions for optimal facility layout. Operations Research, 51(4), 629–644. Software and compilers for embedded systems. In LNCS: vol. 3199 (pp. 152–166).
Sherali, H. D., & Smith, J. C. (2001). Improving discrete model representations via Springer Berlin Heidelberg.
symmetry considerations. Management Science, 47(10), 1396–1407. Xie, W., & Sahinidis, N. V. (2008). A branch-and-bound algorithm for the continuous
Simmons, D. M. (1969). One-dimensional space allocation: An ordering algorithm. facility layout problem. Computers & Chemical Engineering, 32(4–5), 1016–1028.
Operations Research, 17(5), 812–826. Zhang, Z., & Murray, C. C. (2012). A corrected formulation for the double row layout
problem. International Journal of Production Research, 50(15), 4220–4223.