ARTICLE IN PRESS

JID: EOR

[m5G;October 1, 2015;11:19]

European Journal of Operational Research 000 (2015) 113

Contents lists available at ScienceDirect

European Journal of Operational Research

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

Production, Manufacturing and Logistics

A disassembly line balancing problem with xed number of workstations

Eda Gksoy Kalayclar a, Meral Azizoglu b,, Sencer Yeralan c
a

ASELSAN, Naval Systems Program Directorate, Ankara 06172, Turkey

Department of Industrial Engineering, Middle East Technical University, Ankara 06800, Turkey
c

Department of Industrial Engineering, Yasar

University, Izmir,
Turkey
b

a r t i c l e

i n f o

Article history:
Received 30 June 2014
Accepted 3 September 2015
Available online xxx
Keywords:
Integer programming
Heuristics
Disassembly lines
Linear programming relaxation

a b s t r a c t
In this study, a Disassembly Line Balancing Problem with a xed number of workstations is considered. The
product to be disassembled comprises various components, which are referred to as its parts. There is a specied nite supply of the product to be disassembled and specied minimum release quantities (possible zero)
for each part of the product. All units of the product are identical, however different parts can be released
from different units of the product. There is a nite number of identical workstations that perform the necessary disassembly operations, referred to as tasks. We present several upper and lower bounding procedures
that assign the tasks to the workstations so as to maximize the total net revenue. The computational study
has revealed that the procedures produce satisfactory results.
2015 Published by Elsevier B.V.

1. Introduction
The environmental regulations, customer awareness and recent
advances in technology all together have shifted the product recovery
process from the act of disposing to the act of remanufacturing and
recycling. Recycling preserves the material content of the discarded
(used) products via some manufacturing and disassembly operations.
Remanufacturing, on the other hand, keeps the functional content of
the used products and improves their quality up to a desired usable
level via disassembly operations and some manufacturing.
Disassembly is the rst important step of product recovery activities (McGovern and Gupta, 2011) and it is methodical extraction of
valuable parts, operations involve the separation of the reusable parts
from the discarded products. The parts are either subject to some remanufacturing process or sold to suppliers. Disassembly operations
are usually performed on a disassembly line that consists of a number
of serial workstations. The rst workstation takes the product to be
disassembled. The cycle terminates, i.e., the product leaves the line,
whenever all its required parts are disassembled.
The Disassembly Line Balancing Problem (DLBP) assigns the set
of tasks to each workstation for each product to be disassembled.
The problem is critical in minimizing the use of valuable resources
(such as time and money) invested in disassembly, and maximizing
the level of automation of the disassembly process and the quality of
the parts or materials recovered.

Corresponding Author. Tel.: +90 3122102281.

E-mail addresses: egoksoy@aselsan.com.tr (E.G. Kalayclar), ma@metu.edu.tr
(M. Azizoglu), sencer.yeralan@yasar.edu.tr (S. Yeralan).

This study considers a DLBP with a xed number of workstations.

It is assumed that there is a specied supply for the products to be
disassembled. For each part, a dened minimum release quantity
must be met. The amount in excess of the minimum release quantity can also be sold in the market. Hence the excess quantity is produced, provided that the part is protable, and that there is enough
supply. The aim is to assign the tasks to the disassembly workstations
so as to maximize the total net revenue of the parts, while meeting
their specied minimum release quantities, and without exceeding
the specied cycle time. To satisfy the minimum release quantities,
different parts may be released from different units of the product.
The challenge is to use different line balances, hence use different
task assignments to the already mounted workstations, while disassembling different units of the product. To the best of our knowledge,
the study is the rst attempt to solve the DLBP with minimum release
quantities and xed number of workstations.
Our study will have a direct impact on the industries that experience continuous advances in their technology. These advances naturally affect the function and fashion oriented expectations of the
consumers. One such application area that we take our motivation
from is the electrical and electronic equipment industry. The consumers of the personal computers, televisions or cellular phones,
replace their products within a few months to a few years, even
when the products (thereby their parts) are functioning properly.
Electrical and electronic equipments consist of many different substances some of which may contain hazardous components and valuable parts (Capraz, Polat, & Gngr, 2015). As the used products have
proper conditions, their valuable parts (like memories, CPUs, valuable
metals) can be used in manufacturing new models. The minimum release quantities for these parts (that we dene in our models) are

http://dx.doi.org/10.1016/j.ejor.2015.09.004
0377-2217/ 2015 Published by Elsevier B.V.

Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004

JID: EOR
2

ARTICLE IN PRESS

[m5G;October 1, 2015;11:19]

E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113

associated to the demands from the third party like the producers of
the new models. Moreover, as we assume, the amount released excess of the minimum release quantities of the valuable parts can be
stored in the inventory for future orders. Another application area is
the automotive industry where the consumers of the luxurious cars
replace their not-so-much-used cars with new brands that make use
of the most recent technology.
The rest of the study is organized as follows. Section 2 reviews the
disassembly process and literature on the DLBP whereas Section 3
denes our problem. In Section 4 the mathematical models and
their use in nding optimal solutions are discussed. In Section 5,
we present upper bounds together with the mechanisms used to
strengthen them and present a heuristic procedure. The computational experiment is discussed in Section 6 and the study is concluded
in Section 7.
2. Disassembly process & the related literature
Gngr and Gupta (2001b) dened disassembly as a systematic
process of separating a product into its constituent parts, components, subassemblies or other groupings. The issues in the area of
disassembly can be classied into two broad categories as design and
operational. Crowther (1999) considered a design for disassembly issue and mentioned that a life cycle model that incorporates the stages
of a disassembly strategy can highlight the environmental advantages of designing for disassembly, showing how it can extend service
life and thereby improve sustainability. As a practical application, he
discussed the construction industry and mentioned that experience
gained from disassemblable buildings can be used to create guidelines for other products. As discussed in Brennan, Gupta, and Taleb
(1994), the design aspects for disassembly have being recognized by
the industries that generate huge amount of ferrous and plastic waste
like motor cars and appliance sectors.
When the old products come to disassembly plant so that their
components can be recovered in the assembly plant or re-used, operational problems arise. The operational problems that are likely to
arise are layout, resource allocation, process sequencing and disassembly line balancing. As mentioned by Brennan et al. (1994), the operational problems have environmental concerns that are even forced
by the governments like the recycling regulations, and limitations
on the energy consumptions. These regulations affect the operating
costs as extra costs are incurred for covering the expenses related to
the environmental matters.
In their review paper McGovern and Gupta (2011) discussed many
aspects of the operational problems with an emphasis on the disassembly sequencing problem and the DLBP. The disassembly line sequencing problem decides on the disassembly process sequence of a
specied disassembly product and the DLBP decides on the assignment of the tasks to each disassembly workstation.
Lambert (2003) provided a review of the disassembly sequencing
literature. Some noteworthy studies on disassembly sequencing are
due to Lambert (1997), Navin-Chandra (1994) and Lambert (2002).
Navin-Chandra (1994) used a modied traveling salesman problem
so as to minimize total costs while meeting obligatory reclamation of
denite parts. Lambert (1997) proposed a graph based method so as
to maximize the economic performance of the disassembly process
within given technical and environmental constraints. Both methods were applicable to the disassembly of complex consumer products like automotive vehicles, consumer electronics and mechanical assemblies and were applied to a headlamp and ballpoint pen
disassembly products. Lambert (2002) proposed a linear programming based solution procedure for the minimum cost disassembly
sequence of an electronic equipment. His method was applicable all
products having hierarchical modular structures.
The DLBP was rst introduced in Gngr and Gupta (2001b).
Gngr and Gupta (2001a) mentioned several complications like

early leaving, self-skipping, skipping, disappearing and revisiting

work pieces that could be faced in disassembly line practices. To reduce the effects of the complications on the disassembly process,
they proposed a shortest path based solution algorithm. McGovern
and Gupta (2007a) showed that the decision version of the DLBP
is NP-complete. McGovern and Gupta (2007b) considered a DLBP
to minimize the number of workstations while balancing the idle
times between the workstations. They proposed an exhaustive search
method that returns optimal solutions for small sized problem instances and a genetic algorithm that nds high quality solutions, for
the large sized problem instances.
Gngr and Gupta (2002) proposed a heuristic procedure for the
DLBP under complete disassembly. They assumed an innite supply
of a single product and considered the ecient utilization of the resources while satisfying the minimum release quantity. They illustrated the proposed heuristic on an eight task personal computer disassembly example. Koc, Sabuncuoglu, and Erel (2009) considered a
complete disassembly DLBP so as to minimize the number of workstations. They introduced AND/OR graphs and proposed Integer and
Dynamic Programming formulations.
Altekin, Kandiller, and zdemirel (2008) studied the DLBP under partial disassembly and an innite supply of a single product.
They formulated their net revenue maximization model as a mixedinteger linear program and used its relaxations to nd lower and upper bounds. They stated that their approach could be used in designing and operating remanufacturing systems where large volumes of
similar products should be disassembled. Altekin and Akkan (2012)
proposed a predictivereactive approach based on a mixed-integer
model to improve the protability of the disassembly line. A predictive balance was created and then given a failure, the line was rebalanced. They stated that their algorithm could be used as a guide
by the disassembly workers about how to proceed in case of a task
failure.
The recent disassembly lines research considered the uncertainty of the task times and product quality. Bentaha, Battaa, and
Dolgui (2014a, 2014b, 2014c, 2015) and Bentaha, Battaa, and Dolgui
(2014d) studied complete and partial disassembly models, respectively. Bentaha et al. (2014a) modeled uncertainty using the notion of
resource cost and proposed a sample average approximation method.
Bentaha et al. (2014b) studied the joint problem of disassembly line
balancing and sequencing. Bentaha et al. 2014c proposed a lagrangian
relaxation approach to maximize the total prot. Bentaha et al. (2015)
considered the problem of minimizing the workstation operation
costs and additional costs for handling the hazardous parts. They developed several lower and upper bounding mechanisms. Bentaha et
al. (2014d) addressed workload balancing problem with xed number of workstations. They developed a stochastic binary program.
Ding, Feng, Tan, and Gao (2010) stated that a successful disassembly line often requires an integrated consideration of many objectives and proposed an ant colony algorithm to generate the ecient set. They tested their algorithm with three objectives: number
of the workstations, workload balancing between the workstations
and demand rating. They illustrated the heuristic on a 25 task cellular
phone disassembly example. Paksoy, Gngr, zceylan, and Hancilar
(2013) mentioned that in real world applications the objectives could
be fuzzy due to incomplete information and proposed fuzzy goal programming and fuzzy multi-objective programming. They considered
three objectives: number of the workstations, workload balancing
between the workstations and cycle time. The approaches were applied on 10 tasks ash light and 30 tasks radio examples. Hezer and
Kara (2014) introduced parallel line dissassembly balancing problem
and proposed a network model based approach for its solution. For
more information on the dissassembly line balancing and planning,
one may refer to Ullerich (2014).
The most closely related published work to ours is Altekin et al.
(2008)s study. Our problem environment differs from theirs in the

Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004

JID: EOR

ARTICLE IN PRESS

[m5G;October 1, 2015;11:19]

E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113

sense that we take the number of workstations and cycle time as parameters whereas they decide on the number of workstations and
on the cycle time. We assume that there is a nite supply of disassembly product and that there is a lower bound on the quantity of
the part releases whereas they assume innite supply and take minimum release quantities as upper bounds on the quantity of the part
releases. Hence their solution considers a single line balance for all
units, whereas, to satisfy the minimum release quantities, we use different task assignments to disassemble different units.

c
Fig. 1. AND precedence relations.

3. Problem denition

f
We consider a disassembly line balancing model with a nite supply of a single product. There are N tasks where task i is specied
by its cost ci and processing time ti . The tasks that release parts are
called part releasing tasks. The minimum release quantity of each
part-releasing task i is di units. di is the associated demand for part
i from the third party like the producers of the new product models. The part-releasing task i has a revenue ri that includes its market value (for example when the part is sold to the third party) and
recycled material value. The parts may have negative revenues that
represent the disposal costs. Ri is the net revenue of task i, Ri = ri ci .
At the end of the period, each unit of part i that is released excess of
di units can be held in the inventory for future orders and generates a
unit revenue ri when it is sold. Such excess releases would be favored
if they lead to increases in the total net revenue. Our problem is to
assign the tasks to the workstations of the disassembly line for each
unit of the product so as to maximize the total net revenue. We make
the following additional assumptions.

The workstations are already set up and there are K workstations.

All workstations are identically equipped and are capable of performing any one of the tasks at the same pace.
There is a single disassembly product with a nite supply of S
units.
There is a single disassembly period of length L, that is sucient
to disassemble all S units. That is, there is a single batch of size S
to be disassembled in L time units.
The cycle time, C, is the time between the completion of two units
consecutively leaving the disassembly line, hence it is the processing workload that can be allocated to each workstation. One may
view the cycle time to be the ratio L/S .
The cycle time, C, is known with certainty and is not subject to
change as the period length L and supply S, are known with certainty and are not subject to change.
All units of the product are identical, hence contain all disassembly parts.
A partial disassembly is done, i.e., a subset of -but not- all parts is
released.
The subset of parts released from one unit of the product may be
different from another unit, hence different task assignments are
possible while disassembling different units of the product.

The network comprises AND and OR type relations. Predecessor

AND (PAND) relations imply that a task cannot start before all of its
predecessor tasks are nished. The following gure illustrates the
AND precedence relations.
According to Fig. 1, tasks a, b, c and d are predecessors of task e.
Task d is the immediate predecessor of task e as there is no task in
between.
We use AND(i) to denote the set of all AND predecessors of task i
and PAND(i) to denote the set of all AND immediate predecessors of
task i.
Predecessor OR (POR) relations imply that at least one task in a
specied set should be complete before the successor task may begin.
Fig. 2 illustrates POR type relations.

h
Fig. 2. POR precedence relations.

According to the above gure, at least one task in set { f, g, h} must

be completed before task i may begin. We let POR(i) denote the set of
OR predecessors of task i.
4. Mathematical model and properties
The mathematical formulation given below determines the optimum task assignments to the workstations while disassembling each
unit that maximizes the total net revenue. The decision variables of
the model are,

if task i is assigned to workstation

k for s th unit of the product
otherwise

Xiks =
0

Constraint set (1) deals with the minimum release quantities. The
amount of each part released should be higher than its minimum release quantity.
S
K 


Xiks di

i = 1, . . . , N

(1)

k=1 s=1

Constraint set (2) states the cycle time limit should not be exceeded at any workstation for any unit.
N


ti Xiks C

k = 1, . . . , K; s = 1, . . . , S

(2)

i=1

Constraint sets (3) and (4) ensure that the precedence requirements are observed. The AND precedence relations should be satised. Thus task i cannot be assigned to a workstation that is before
the workstation of its AND predecessor task l.
K


kXlks

k=1

K


kXiks

i = 1, . . . , N, l PAND(i), s = 1, . . . , S

k=1

(3)
Similarly, the OR precedence relations should be satised. Thus
task i can be assigned to workstation k if and only if at least one of its
OR predecessors is assigned to workstations 1 through k.

Xiks

k



X jhs

i = 1, . . . , N; k = 1, . . . , K; s = 1, . . . , S

h=1 jPOR(i)

(4)

Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004

ARTICLE IN PRESS

JID: EOR
4

[m5G;October 1, 2015;11:19]

E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113

Constraint set (5) ensures that a task can be assigned to at most

one workstation.

K
N 


K


i=1 k=1

Xiks 1

i = 1, . . . , N; s = 1, . . . , S

(5)

k=1

The assignment variables should be nonnegative as by constraint

set (6).

Xiks 0 and integer

i = 1, . . . , N; k = 1, . . . , K; s = 1, . . . , S
(6)

It follows from Constraint Sets (5) and (6) that Xiks is binary.
The model, hereafter referred to as Model I, is dened by objective
function (7) below, together with constraint sets (1) through (6).

Max

S
K 
N 


Ri Xiks

(7)

Model I has N K S binary decision variables.

McGovern and Gupta (2007b) showed that the decision version of
the DLBP is strongly NP-complete. It follows that our DLBP problem is
NP-hard in the strong sense. We thus resort to several approaches to
solve the DLBP approximately or heuristically. We also develop theoretical instruments, more specically, valid constraints, that improve
the solution effort. These topics are discussed in the remainder of the
paper. In this section, we introduce the key insight into reducing the
problem size.
In typical implementations, the model solution yields different
disassembly tasks and different part-releasing assignments for the
batch of S units. Typically, rst a subset of the batch is disassembled to satisfy the minimum release quantity constraints. Recall that
minimum quantities for parts are novel and distinguishing aspect of
the current work. Once the minimum release quantities are satised,
though, the solution understandably favors task assignments that
yield the maximum revenue. Note that, once all of the part minimum
release quantities have been met, the specic task assignments to attain the maximum revenue is independent of the remaining quantity
to be disassembled. In fact, the maximum revenue yielding task assignments can be found quite easily by solving Model I with S = 1. The
maximum revenue yielding task assignments could be replicated as
many times as needed for the remaining parts.
The above insight immediately suggests an approach to make
problems with large S more tractable. If one had access to the number
of units needed, say Y S, to be disassembled in order to satisfy the
minimum release quantities, then Model I could be solved using Y instead of S. Since Y S, and sometimes even Y << S, the computations
are simplied. The remaining task assignments are then computed by
solving Model I with S = 1 and then replicating the solution for the
remaining SY units.
Clearly, the above approach would also work for an upper bound
of Y. Even an approximate value of Y may be useful. If the chosen
value of Y gives an infeasible solution, then we could investigate ways
to increase Y to obtain a feasible solution. In the following sections,
we discuss how the quantity Y may be computed, exactly or approximately, and how bounds for Y may be obtained.
4.1. Computing the number of units to be disassembled to satisfy the
minimum release quantities
The formulation given below, henceforth referred to as Model II,
nds the minimum number of units to be disassembled to satisfy all
minimum release quantities.
Model II can be dened by the following additional variable and
constraint.

Ys =
0

if s th unit is disassembled to satisfy

minimum release quantity
otherwise

s = 1, . . . , S

(8)

Constraint Set (9) assigns Ys to 1, if there is at least once task assigned to any workstation while disassembling sth unit of the prod K
uct. Note that N
k=1 Xiks is bounded from above by . We use
i=1

K
= N K. If N
i=1
k=1 Xiks 1 then Ys = 1 otherwise Ys is zero as we are


minimizing Y = Ss=1 Ys . Y = Ss=1 Ys gives the total number of units,
with at least one task assignment. Hence it is the minimum number
of disassembly units to satisfy all minimum release quantities.
The objective function of Model II is expressed as below.

Min

S


Ys P

s=1

i=1 k=1 s=1

Xiks Ys

S
K 
N 


Ri Xiks

(9)

i=1 k=1 s=1

The above objective primarily minimizes the number of disassembly units to satisfy all minimum release quantities. Among the solutions that attain the minimum number of disassembly units Y, the
model selects the one with the maximum net revenue. To maximize
the total net revenue while keeping the minimum number of disassembly units, some parts are released more than their minimum release requirements. The remaining SY units are handled by Model III
(Section 4.2).
The magnitude of p used in (9) is important in the sense that it

should not increase Y = Ss=1 Ys en route to increasing total net revenue. p should be set small enough to satisfy constraint (10).
S


Ys P Pmin

s=1

S


Ys + 1 P Pmax

(10)

s=1

where Pmin (Pmax ) = minimum (maximum) possible total net revenue.


Hence any solution with a larger Y = Ss=1 Ys value should not be
favored even if it leads to the maximum increase in the total net revenue value. Constraint (10) implies that P Pmax P Pmin 1. This

1
We use Pmax = S N
follows P Pmax P
i=1 Ri di and Pmin = 0. Hence

P =

1
N

min

i=1 Ri di

guarantees the maximum revenue solution among

ones with the minimum number of disassembled units.

Model II is completely dened by the objective function expressed
in (9) together with constraint sets (1)(6) and (8).
The model satises all minimum release requirements by disas
sembling Y = Ss=1 Ys units. For the remaining (SY) units, the same
task assignments are used. To nd these task assignments we modify Model I by setting S = 1, and then replicating the solution for the
remaining SY units. This solution, that is, Model I with S = 1 and
then replicated as needed, is referred to as Model III. In other words,
Model III nds the maximum net revenue without considering minimum number of part releases.
4.2. Maximum revenue solution in the absence of part minimum
release quantities
In the following model, Model III, we drop subscript s from our
decision variable Xiks as we are considering a single unit. The decision
variable becomes Xik where

Xik =

1
0

if task i is assigned to workstation k

otherwise

The objective function is as expressed in (11).

Max

K
N 


Ri Xik

(11)

i=1 k=1

We also drop the minimum release quantity constraint as they are

met by Model II. Model III is dened by the objective function expressed in (11) together with constraint sets (2) through (6) by setting
S = 1.

Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004

JID: EOR

ARTICLE IN PRESS

[m5G;October 1, 2015;11:19]

E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113

4.3. Solution procedure using Model II and Model III

Model I given above yields the optimum solution but suffers from
computational diculties as the problem size increases. Alternatively, Model II and Model III may now be used to nd an optimal
solution as given by the following procedure.
Procedure 1. Optimal Solution via Model II and Model III
Step 1. Use Model II to nd the task assignments for the rst Y =
S
s=1 Ys disassembly units.
Step 2. Use Model III to nd the task assignments for the last (SY)
disassembly units.
Step 3. The optimal total net revenue over all S units is:
Total Net revenue by Model II + (S Y) Total Net revenue by
Model III
Model II is more complex than Model I due to the added Ys variables, hence in nding the optimal solutions, Procedure 1 should not
be favored to Model I. However, Procedure 1 is of importance, since
we use the procedure to derive upper and lower bounds on the optimal total net revenue value.
4.4. Solution procedure using Model I and Model III
Model I and Model III together may be used to nd optimal solu
tions as well. Recall that Y = Ss=1 Ys returned by Model II is the minimum number units to be disassembled to satisfy all minimum part
requirements. If an upper bound on Y, UB(Y), is available, then Model
I can be solved with UB(Y) in place of S units. Thereafter, Model III
optimally solves the remaining SUB(Y) units. Procedure 2, below, is
used to nd an optimal solution using Model I and Model III. The upper bound UB(Y) may be guessed, or sought using a simple algorithm
such as bisection search. We will revisit the task of nding UB(Y) in
the following sections.
Procedure 2. Optimal Solution via Model I and Model III
Step 1. Solve Model I with S = UB(Y).
Step 2. Solve Model III to nd the task assignments for the remaining S UB(Y) units.
Step 3. Find the optimal total net revenue over all S units as:
Total net revenue over UB(Y) units by Model I+(SUB(Y))
units Total net revenue over a single unit by Model III
Note that there are three ways of nding an optimal solution: the
original Model I, Procedure 1, and Procedure 2. Procedure 2 should be
favored to original Model I, in particular when compared to S, a small
UB(Y) is available.
In our experiments, we use Model I to nd the optimal solutions.
We develop two upper bounds to the optimum objective function
value that use the Linear Programming Relaxations (LPRs) of the original model. We benet from the ideas used in Model II and III to nd
a third upper bound and to develop a heuristic solution procedure.
These are presented next.
5. Solution approaches
As mentioned, our DLBP is NP-hard in the strong sense. Hence
there is little chance of nding polynomial even a pseudo-polynomial
algorithm that solves it optimally. In the section we present our solution approaches that provide upper bounds and a heuristic solution
to our DLBP.
5.1. Upper Bounds
Upper Bound 1 (UB1 ): An optimal solution to any relaxation provides an upper bound on the optimal objective function value of our
maximization problem.

Model I is solved to optimality by relaxing the integrality constraints on the Xiks values. The resulting optimal objective function
value, ZLP , provides an upper bound on the optimal total net revenue.
The only difference between the original model and its LPR is the integrality requirement of the Xiks values. Consequently, ZLP is an upper
bound on the optimal objective function value Z .
Upper Bound 2 (UB2 ): UB2 is found by adding valid cuts to the LPR
of Model I. The valid cuts are found by investigating the properties of
the optimal solutions that may not be satised by the optimal LPR.
Cut1Fixing the minimum release quantities
Theorem 1. There exists an optimal solution in which di units of part i
are released from the rst di disassembly units.
Proof. Assume an optimal solution in which the rst di units are released from di units of the product. If the tasks in any two units, say
a and b, are changed, i.e., Xika is set to Xikb and Xikb is set to Xika for
all i , then the optimal net revenue is not affected. Assume an optimal
solution in which part i is released in units a+1, a+ di . Setting Xjkr =
Xjk(a+r) for all j, r dj and Xjk(ta+r) = Xjkr does not change the revenue
value. Then, there is an optimal solution in which dj units of part i is
released in the rst di units. 
Using Theorem 1, dr units for any part r can be obtained from the
rst dr disassembly units. Constraint set (12) below ensures that dr
units of part r are released from the rst dr disassembly units. To increase the power of (12), we select task r such that dr =Maxi {di }.
dr
K 


Xrks = dr

s = 1, . . . , dr

(12)

k=1 s=1

In computational studies, we observe that the constraint set (12)

eliminates many alternate optimal solutions, thereby decreasing the
computational effort. The reduction in computational effort was reected as signicantly reduced CPU times.
Cut2 Lower Bound on the workstation positions
Cut2 provides a lower bound on the workstation index that any
particular task.
Theorem 2. In all optimal solutions, task i cannot beassigned to workstations 1, 2, . . . , Ei 1 where Ei = 

ti +Min jPOR(i) {t j }+
C

jAND(i) t j

.

Proof. Task i cannot start before its predecessors and at least one
of its immediate POR predecessors are completed. Hence it should
wait at least Min jPOR(i) {t j } units for its POR predecessors and

jAND(i) t j units forits AND predecessors to start and at least
ti + Min jPOR(i) {t j } + jAND(i) t j units to complete. When task splitting is allowed and other tasks are ignored, ti + Min jPOR(i) {t j } +


ti +Min jPOR(i) {t j }+ jAND(i) t j
 workstations.
jAND(i) t j units require 
C
When task splitting
is not allowed and other tasks are considered


ti +Min jPOR(i) {t j }+
C

for task i.

jAND(i) t j

 is a lower bound on the earliest station

In the absence of POR relations, Theorem 2 reduces to the one derived in Patterson and Albracht (1975) for the classical assembly line
balancing problem.
Constraint set (13) below supports Theorem 2.
Ei 1

S


Xiks = 0 i = 1, . . . , n

(13)

k=1 s=1

Constraint set (13) ensures that task i is not assigned to workstations 1, 2, , Ei 1. The network in Fig. 3 is used to illustrate Ei
computations.
Note that { t1 + t2 } is the minimum total processing load to start
task A and Min {t3 , t4 }+ t5 is the minimum total processing load to
start task B. Then Min{(t1 + t2 + tA ), (Min {t3 , t4 } + t5 + tB )} is the

Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004

ARTICLE IN PRESS

JID: EOR
6

[m5G;October 1, 2015;11:19]

E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113

cannot allow more task assignments. This is due to the fact that all
tasks in workstations r + 1 through K are successors of task i, hence
cannot be processed in workstations 1 through r, since task i is to be
processed later. Then, there exists an optimal solution in which task i
is processed in the rst r workstations. 

t1 + t2 = SA
Min {SA+tA , SB+tB } = Sj

Theorem 4. If there exists a feasible schedule that processes all tasks

other than task i and its predecessors in the last r workstations, then
there is an optimal schedule in which task i is processed in workstations
K r + 1 through K.

3
B
Min {t3 , t4 } + t5 = SB

4
5

Fig. 3. An example network to illustrate Ei computations.

minimum total processing load to start task j.

Note that Ei =


Hence Ei =

ti + Minimum total processing load to start task i

.
C

ti + Min{(t1 + t2 + tA ), (Min{t3 , t4 } + t5 + tB )}
.
C

The network in Fig. 4 is adapted from Bourjault (1984) to illustrate

Theorem 2.
We take K = 4, C = 63, S = 80 and the task times are as given
below:
Tasks 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
13 7 1 6 7 5 15 8 17 3 14 12 17 4 11 18 15 19 4 18 16 19
ti


E13 =


E13 =

t13 + Min{(t1 + t3 + t4 ), (t1 + t2 )} + t22 + t7

C


 

17 + Min{(13 + 1 + 6 ), (13 + 7 )} + 19 + 15
71
=
=2
63
63

The theorem states that task i cannot be assigned to workstations

1, 2, , Ei 1. Accordingly, task 13 cannot be assigned to workstation
1 (E13 1 = 1).
Cut3Projecting minimum release quantities
Altekin et al. (2008) stated that in the optimal solution, the total
fraction of any task assigned is never greater than the total fraction
of each of its assigned AND predecessors and the total fractional assignment of all of its OR predecessors, and moreover the constraints
may not be satised for the optimal LPR. Their results hold for any
problem that contains PAND and POR relations and an innite supply.
We generalize their results to our problem where the supply is nite
and impose the constraint sets (14) and (15).
S
K 


Xiks

k=1 s=1
S
K 

k=1 s=!

S
K 


X jks

i = 1, . . . , n; j PAND(i).

(14)

k=1 s=1

Xiks

S
K 
 

X jks

i = 1, . . . , n

(15)

jPOR(i) k=1 s=1

Cut4Existence of a feasible solution

Our last cut uses the results of the following two theorems.

Proof. Assume task i is assigned to workstation k such that k K

r + 1 and workstations 1 through K r process the predecessors of
task j. Taking task i from workstation k and placing it to workstation
l such that l K r cannot increase the total net revenue, as such a
replacement cannot allow more Assignments. This is due to the fact
that all tasks in workstations 1 through Kr are predecessors of task i,
hence cannot be processed in workstations K r + 1 through K, since
task i is processed earlier. Then, there exists an optimal solution in
which task i is processed in the last r workstations. 
Note that Theorems 3 and 4 require a feasible solution. To nd
such a solution we develop the following simple rule.
Let A be a set of tasks that should be assigned to K workstations.
Order the tasks in set A according to non-decreasing task times. Take
the tasks from the order starting from the rst feasible task. A task is
feasible if its inclusion to the current workstation does not violate the
precedence relations and cycle time constraints. If no such task exists,
close the current workstation and open a new one. Stop when all jobs
in set A is assigned. Let m(A) be the resulting number of workstations.
Following Theorem 3, we set A = A1 where A1 is the set of all tasks
except the successors of task i. Following Theorem 4, we set A = A2
where A2 is the set of all tasks except the predecessors of task j. After
m(A1 ) and m(A2 ) are obtained, implementing the heuristic for sets A1
and A2 , we include one of the two cuts (16) and (17).
S


K


Xiks = 0

i = 1, . . . , N

(16)

s=1 k=m(A1 )+1

A2 )1
S m(

s=1

Xiks = 0

i = 1, . . . , N

(17)

k=1

We select one of the cuts for the LPR using the following rule.
Rule. Use Cut (16) if m(A1 ) m(A2 ) 1, else use Cut (17)
The rule selects the cut that prevents more assignments, hence
the stronger one.
We hereafter refer to the LPR using the cuts (12) through (17)
as strengthened LPR. UB2 is the total net revenue returned by the
strengthened LPR.
Upper Bound 3 (UB3 ): To obtain UB3 , we use the idea used in
Model II. Recall that Model II aims to satisfy all minimum release


quantities ( i di ) and minimizes Y = Ss=1 Ys . Here in place of Y, we
use LB(Y), to get rid of the Ys binary variables.
First, we nd LB(Y), and then increase it one by one, until a feasible
solution is reached. Theorem 5 states LB(Y).
We let dnewi = Max{di , Max j|iAND j {d j }}. Note thatdnewi is the
minimum release quantity of task i which is modied by its minimum release quantity and the minimum release quantities of all its
AND successors.

Theorem 3. If there exists a feasible schedule that processes all tasks

other than task i and its successors in r workstations then there exists an
optimal schedule in which task i is processed in workstations 1 through r.

i
 is a valid lower bound on the numTheorem 5. LB(Y ) =  i=1KiC
ber of units to satisfy all minimum release quantities.

Proof. Assume task i is assigned to workstation k such that k r and

workstations r +1 through K process the successors of task i. Taking
task i from workstation k and placing to workstation l such that l
r + 1 cannot increase the total net revenue as such a replacement

Proof. The total processing required to produce all minimum re

lease quantity is ni=1 ti dnewi Each unit can be disassembled in no

more than
n K C time units as each workstation is available for C time
t dnewi
 is the number of units disassembled to satisfy
units.  i=1KiC

n

t dnew

Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004

ARTICLE IN PRESS

JID: EOR

[m5G;October 1, 2015;11:19]

E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113

16
3

21

4
11

20

12

10

2
8
15

22
9
5

17
18

19

14

13
Fig. 4. Precedence diagram adapted from Bourjaults ball-point pen example.

all minimum release quantity if task splitting between the disassembly unitsand workstations were allowed. As no task splitting
n

t dnewi

is allowed  i=1KiC
disassembled. 

 is a lower bound on the number of units

We illustrate Theorem 5 and UB3 on the network in Fig. 4 with

K = 4, S = 80 and C = 63. The minimum release quantities, task times
and dnewi values are given in Table 1.

4
 = 3. We rst
Using Theorem 5 we nd LB(Y ) =  13 4+744++19
63
set the number of units to LB(Y) = 3. However a feasible solution
for the LPR is not found. There is no feasible solution with 4 units
as well. The LPR nds a feasible solution with 5 units. Then the minimum number of units to satisfy all minimum release quantities, Y = 5.
Model III is used to nd the line balance for the remaining 75 units.
Formally we let,
ZA = maximum net revenue for the strengthened LPR problem
with Y units
ZB = maximum net revenue for a single unit problem with
Model III.

UB3 = ZA + (SY) ZB

In our example, the strengthened LPR nds ZA as 1626. Model III

nds ZB , as 326. Then UB3 = 1626 + (80 5) 326 = 26.076.
Note that here, UB2 = UB3 . ZA is found much easier than the LPR
with S = 80. Moreover ZB is found easily as it considers a single unit.
Hence one should expect to obtain UB3 much quicker than UB2 .
For this instance we nd that UB1 =26.796 and the optimal net
revenue = 26.076. Note that UB3, improves UB1 by 720 money units
and catches the optimal solution in considerably small solution
time.

5.2. Lower bound, heuristic solution

The lower bound uses UB3 as a stepping stone to nd a feasible solution. The motivation to use UB3 is its satisfactory behavior in terms
of its solution quality and time. The bound rst solves UB3 , then xes
its variables that are assigned to 1 and obtains a reduced problem.
The reduced problem has only the partially assigned or unassigned
tasks of the UB3 solution. As the partial assignments are quite low
compared to the original variables, the reduced problem is very small
in size. Since our MILP solver can handle small-sized problems very
quickly, we prefer to use it to nd an optimal solution for partially
assigned tasks.
After xing the variables to 1, the resulting solution obtained by
MILP may violate the cycle time constraints or the precedence constraints. If the cycle time constraint is violated, then we increase the
number of disassembly units one by one, until we obtain a feasible
solution. If the precedence relations are violated, the tasks whose xings cause violation, are set to zero, and the MILP is solved again.
Below is the stepwise description of our lower bounding
procedure.
n

t dnew

i
 and let T = LB(Y)
Step 0. Let LB(Y ) =  i=1KiC
Step 1. Solve the LPR with cuts for T units of the disassembly
product.
Fix the variables that have value 1 in the LPR solution.
Step 2. Find the optimal solution by Model I for the non-xed tasks.
If the solution is feasible then let ZC be the objective function
value and go to Step 4.
Step 3. If infeasibility is due to the cycle time constraints then let T
= T+1, go to Step 1.
If infeasibility is due to the precedence relations then set the
variable of a task to zero if one of its predecessor tasks is not
assigned to 1, go to Step 2.

Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004

ARTICLE IN PRESS

JID: EOR
8

[m5G;October 1, 2015;11:19]

E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113

Table 1
The data for the Bourjaults ball-point pen problem.
Tasks

10

11

12

13

14

15

16

17

18

19

20

21

22

di (units)
ti (time units)
dnewi (units)
ci (money units)
ri (money units)
Ri (money units)

0
13
4
17
49
32

0
7
4
18
10
-8

0
1
4
14
35
21

4
6
4
7
37
30

0
7
4
14
16
2

4
5
4
8
12
4

0
15
4
8
20
12

3
8
3
15
37
22

4
17
4
16
43
27

4
3
4
7
32
25

4
14
4
15
24
9

4
12
4
12
41
29

4
17
4
6
19
13

3
4
3
17
35
18

2
11
2
17
19
2

1
18
1
6
18
12

0
15
0
9
48
39

4
19
4
13
21
8

2
4
2
12
10
-2

1
18
1
16
40
24

0
16
0
7
16
9

0
19
4
15
11
-4

34
8

6
9
17
2

18
10

32
23
16

31
33

30

1
11
24
13

22

29

3
28
12
14

19

25

21
15

26

20
27

Fig. 5. Precedence diagram adapted from Lamberts radio example.

Step 4. Find an optimal solution to the single unit problem using

Model III.
Let ZD be the objective function value
The heuristic gives a total net revenue of ZC + (S T) ZD .
6. Computational experiment
In this section the data generation scheme is discussed and the
computational results are evaluated. Two networks from the previous
works are used for network generation.
1st Network (22 tasks). Bourjault (1984) illustrated a disassembly
network for a 10-part ballpoint pen. We adapt this network by
replacing the successor OR relations with successor AND relations and give the resulting network in Fig. 4.

2nd Network (34 tasks). Lambert (1997) illustrated a disassembly

network for a 20-part radio. We adapt this network by replacing the successor OR relations with successor AND relations
and give the resulting network in Fig. 5.
Three networks are found by extending the small networks using
additional arcs.
3rd Network (47 tasks). 2nd network and a part of the 1st network are combined with some arcs added.
4th Network (60 tasks). 3rd network and a part of the 1st network
are combined with some arcs added.
5th Network (73 tasks). 4th network and a part of the 1st network
are combined with some arcs added.
The following two sets for the number of workstations K are used.

Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004

ARTICLE IN PRESS

JID: EOR

[m5G;October 1, 2015;11:19]

E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113

Table 2
UB1 performances.
Variables

Set C1

Set C2

Set D1

Set K1 N = 22

3,520

Set K2 N = 22

7,040

Set K1 N = 34

5,440

Set K2 N = 34

10,880

Set K1 N = 47

7,520

Set K2 N = 47

15,040

Set K1 N = 60

19,200

Set K2 N = 60

38,400

Set K1 N = 73

23,360

Set K2 N = 73

46,720

Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.

Set D2
Avg

Max

Avg

Max

Avg

Max

1
722.4
1
2437.5
0
715.9
0
3803.1
0
1184
0
5492.4
-
8749.8
13515.3
10876.3
16730.7

2
1636
3
3643
1
1950
1
4864
1
4011
1
6865
9579
14742
11804
18440

1
1120.7
1
2497.4
0
1150.8
0
4238.7
0
2509.4
0
5910.8
8999.5
13853.1
11249.8
17291.3

2
1778
3
3592
1
2251
1
5136
1
3924
1
6966
9791
15153
11974
18654

1
487.1
1
1628.1
0
786
0
3004.1
0
584.2
0
4041.1
6887.3
11399.3
9137.3
14045.5

2
1356
3
3017
1
2367
1
4307
1
2403
1
5888
8344
13268
10462
15689

1
470.6
2
1900.4
0
741
0
3102.6
0
1152.4
0
4654.9
7751.1
12307.1
9716.3
15019.8

2
1684
3
3306
1
2210
1
4386
1
3391
1
5946
8935
14280
10633
16147

- illustrates that optimal solution for large networks could not be obtained.

 
ti

 

Set C2. C = 1.5

Set D2

Max

Set K1. K = 2 for small networks with 20, 34, 47 tasks

K = 4 for large networks with 60, 73 tasks
Set K2. K = 4 for small networks with 20, 34, 47 tasks
K = 8 for large networks with 60, 73 tasks
The following two sets for the cycle times, C are used.

Set C1. C =

Set D1

Avg

ti

The processing times are generated from a discrete uniform distribution [1, 20]. A task receives a minimum release quantity, hence
called a part-releasing task, according to a random process. We generate a random number between 0 and 1. If the generated number is
below 0.3, we set its minimum release quantity to zero.
Two distributions are used to generate their minimum release
quantities.
Set D1. Minimum release quantity is uniform [1, 5]
Set D2. Minimum release quantity is uniform [5, 10]
Note that Set D1 contains instances with a low minimum release
quantity, and Set D2, with a high minimum release quantity.
The number of units, S, is set to 80. The costs are generated from a
discrete uniform distribution between 5 and 20 and the revenues are
generated from a discrete uniform distribution between 10 and 50.
There are 40 combinations due to 5, 2, 2 and 2 alternatives for
N, K, C and D respectively (3 2 2 2 = 24 for small networks and
2 2 2 2 = 16 for large networks). For each combination, 10 problem
instances are generated. That is, a total of 400 problem instances are
used in our experiments.
The mathematical models are solved by CPLEX 10 and the algorithms are coded in Microsoft Visual C++ 2008. The experiments are
run in Intel Core2 Duo 2.00 gigahertz, 2 gigabyte RAM.
The execution of the MILP (Model I) is terminated if the optimal
solution is not returned in 3600 seconds. The optimal solutions are
found within 3600 seconds when N = 22, 34 and 47. However, optimal
solutions were not found when N = 60 and 73 .
First the performance of our upper bounds is studied. The deviaUBi OPT
) 100 where OPT =
tion of an instance is found as %Dev = ( OPT

the optimal net revenue and UBi = the total net revenue by upper
bound i. The average and maximum deviations and number of fractional variables are reported in Table 2 and Table 3, for UB1 and UB2 ,
respectively.
As can be observed from the tables, the deviations are consistently
low over all problem sets. For UB1 , the average deviations are below
1 percent and almost all maximum deviations are below 3 percent
when N = 22, 34 and 47. Note that the deviations do not deteriorate
with an increase in N. However they deteriorate with an increase in K
due to the ination of the number of binary variables. When N and K
values are xed, the minimum release quantities and cycle times do
not affect the deviations signicantly.
UB2 produces slightly smaller deviations due to the power of
our cuts. For example, the maximum deviation for UB1 is 3 percent,
whereas it is nearly zero for UB2 in sets K1, C1 and D1 for N = 22.
We also give the number of fractional variables in Tables 2, 3 and 4.
The number of the fractional variables produced by the LPR is too
high. For example when N = 22, for sets K2, C2 and D2 (i.e., 4 workstations, high minimum release quantity, and high cycle time cases),
3306 out of 7040 variables are found to be fractional. This value
reduces to 412 when cuts are added. On the average, the cuts reduce the number of fractional variables from 1900.4 to 180.3. The
number of fractional variables is generally affected by the N values. For example, from Tables 24, it can be seen that for sets K2,
C2 and D2, increasing N increases the number of fractional variables of UB2 , on the average. Note that the average number of fractional variables are 1900.4, 3102.6 and 4654.9 for N = 22, 34 and 47
respectively.
It can also be observed from the tables that, for xed N, an increase
in the number of workstations increases the number of the fractional
variables. For example for N = 22 and sets C1 and D1, the average
number of fractional variables is 722.4 when K = 2, and 2437.5 when
K = 4. This is due to the fact that more workstations lead to higher
task splits, leading to more fractional variables. For xed N and K,
we observe that increasing C reduces the number of fractional variables (with few exceptions). For example, when N = 47, K = 2, and
D1 combination, the average number of fractional variables are 1184
and 584.2 for low and high C, respectively. This is due to the fact that
increasing C gives more room for complete task assignments, hence
reducing the number of fractional variables. In our experiments, we

Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004

ARTICLE IN PRESS

JID: EOR
10

[m5G;October 1, 2015;11:19]

E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
Table 3
UB2 performances.
Variables

Set C1

Set C2

Set D1

Set K1 N = 22

3,520

Set K2 N = 22

7,040

Set K1 N = 34

5,440

Set K2 N = 34

10,880

Set K1 N = 47

7,520

Set K2 N = 47

15,040

Set K1 N = 60

19,200

Set K2 N = 60

38,400

Set K1 N = 73

23,360

Set K2 N = 73

46,720

Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.

Set D2
Max

Avg

Max

Avg

Max

Avg

Max

0
65.3
0
368.4
0
19
0
525.2
0
49
0
570.3
-
746
1778.7
657.5
1789.1

0
160
0
570
0
160
0
870
0
160
0
970
1524
2203
922
2326

0
67
0
376
0
22.9
0
517.7
0
50
0
592
765.7
1725.2
649.9
1769

0
160
0.01
596
0
158
0
855
0
160
0
982
1546
2116
859
2244

0
0
0
197.8
0
0
0
271.1
0
0
0
370.4
418.3
1192.1
460.8
1197.1

0
0
0
432
0
0
0
438
0
0
0
800
1068
1523
1027
1931

0
0
0
180.3
0
0
0
259.1
0
0
0
362.2
424.8
1179.2
462.4
1201.8

0
0
0
412
0
0
0
446
0
0
0
815
1068
1434
1045
2008

- illustrates that optimal solution for large networks could not be obtained.

C1

C2

D1

K1

K2

22
34
47
22
34
47

Set D2

Avg

Table 4
The number of units with different line balances (out of 80).
N

Set D1

D2

D1

D2

Avg

Max

Avg

Max

Avg

Max

Avg

Max

5.30
5.10
5.50
5.10
5.00
5.30

6.00
6.00
6.00
4.00
5.00
6.00

11.00
10.00
11.30
11.80
10.90
11.10

12.00
10.00
12.00
13.00
11.00
12.00

5.50
5.90
5.80
5.30
5.10
5.40

6.00
6.00
6.00
6.00
6.00
6.00

11.20
10.00
11.20
10.70
10.00
11.00

12.00
10.00
12.00
11.00
10.00
12.00

do not observe a signicant effect of the minimum release quantity

gures on the number of fractional variables.
Table 4 reports the number of different line balances incurred in a
period, while S units are disassembled, for small networks.
Note that the number of different line balances returned by Model
II is quite small relative to the total number of disassembly units. For
D1, there are about 5 and 6 different line balances. Increasing D from
D1 to D2, almost doubles the number of different balances, as more
units have to be used to satisfy the minimum release quantities. UB3
is based on the idea of the different line balances, hence one should
expect that it runs much faster than UB2 .
The solution times are measured in central processing unit (CPU)
seconds. The average and maximum CPU times for our three upper
bounds are given in Table 5 and Table 6 for small and large networks,
respectively. Table 5 includes the CPU times of the MILP, as the small
networks could be solved to optimality.
As can be observed from the tables, the upper bounds are produced very quickly. Compared to UB1 , UB2 is computed more quickly,
due to the power of the cuts, i.e., their effectiveness in reducing the
solution space.
Moreover, compared to UB2 , UB3 is computed more quickly as relatively fewer units, hence fewer variables, are used by the LPRs. For
example, for N = 22 and the instances in sets K2, C2 and D2, the CPU
time decreases from 0.59 to 0.25 seconds by adding the cuts, and similarly, the CPU time decreases from 0.25 to 0.08 seconds by using up
to 10 supply units instead of 80. The CPU times of nding the upper bounds increase with an increase in N and K. This is due to the

increase in the dimensions of the linear programs. For example, for

sets K1, C2 and D2, as N increases from 60 to 73, the average CPU
times of UB1 , UB2 and UB3 increase from 3.18, 0.72, 0.11 to 4.52, 0.83,
0.13 respectively.
For N = 34 and sets C2 and D2, as K increases, the average CPU
times increase from 0.43, 0.08, 0.06 to 1.18, 0.31, 0.09 seconds for UB1 ,
UB2 and UB3 respectively.
For xed N, K and D values, an increase in the C value decreases the
CPU time. This is due to the fact that for large C, more tasks nd an assignment for a workstation, and hence assignment decisions are produced easier by linear programs. For example, for N = 34 and Sets K2
and D2, increasing C from C1 to C2, decreases the average CPU times
from 1.67, 0.46, 0.11 seconds to 1.18, 0.31, 0.09 seconds for UB1 , UB2
and UB3 , respectively. When the other parameters are xed, the increase in the D value increases the CPU times slightly. This is because
when D increases, the number of units with different assignments
increases. For example, for N = 34 and sets C1 and K2, increasing D
from D1 to D2, increases the average CPU times from 1.27, 0.44, 0.08
seconds to 1.67, 0.46, 0.11 seconds for UB1 , UB2 and UB3 respectively.
The effect of K is more dominant for the MILP, due to the ination of
the number of decision variables.
For xed N, an increase in the K value increases the CPU times remarkably. For example for N = 22 and sets C1 and D1, an increase in
K from K1 to K2 increases the CPU time from 3.83 seconds to 1078.32
seconds. However the effect of N is not as signicant as that of K.
For example, for K = 2, when N increases from 22 to 34 for sets C1
and D1, the CPU time increases from 3.83 to 4.37 seconds. Note from
Table 5 that for xed N and K, an increase in the C value decreases
the CPU times remarkably. This is due to the fact that a workstation
has room to accommodate many tasks, thereby leading to easier decisions. However D values do not have a consistent effect on the solution time of the MILP.
Fig. 6 illustrates the CPU times of UB1 for sets K, C and D when N =
60 and 73.
As can be observed from Fig. 6 the CPU times of UB1 increase with
increases in N and K values and decrease with an increase in C value.
To summarize, our experiments have revealed that high quality
upper bounds could be obtained even for large networks in less than
half a minute. The upper bound deviations are not affected from
the increases in the problem sizes and CPU times to obtain them

Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004

ARTICLE IN PRESS

JID: EOR

[m5G;October 1, 2015;11:19]

E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113

11

Table 5
The CPU times (in second) of the Upper Bounds and MILPsmall networks.
Set C1

Set C2

Set D1

Set K1 N = 22

UB1
UB2
UB3
MILP
UB1
UB2
UB3
MILP
UB1
UB2
UB3
MILP
UB1
UB2
UB3
MILP
UB1
UB2
UB3
MILP
UB1
UB2
UB3
MILP

Set K2 N = 22

Set K1 N = 34

Set K2 N = 34

Set K1 N = 47

Set K2 N = 47

Set D2

Set D1

Set D2

Avg

Max

Avg

Max

Avg

Max

Avg

Max

0.33
0.09
0.05
3.83
0.7
0.41
0.07
1078.32
0.57
0.11
0.06
4.37
1.27
0.44
0.08
1454.94
1.05
0.16
0.07
22.77
2.46
0.82
0.09
784.49

0.5
0.1
0.09
10.67
1.25
0.67
0.1
3599.85
0.86
0.12
0.08
6.04
1.95
0.53
0.1
3600.34
1.36
0.18
0.11
118.85
3.59
1.17
0.11
3600.0

0.42
0.11
0.06
10.65
0.84
0.4
0.09
796.07
0.78
0.11
0.07
5.93
1.67
0.46
0.11
939.75
1.42
0.14
0.08
14.38
2.94
0.79
0.16
759.57

0.54
0.14
0.08
73.21
1.09
0.65
0.11
3600.62
0.92
0.13
0.1
8.15
2.21
0.58
0.12
3417.72
1.63
0.16
0.12
64.45
3.51
1.05
0.2
3600.0

0.22
0.07
0.05
1.76
0.47
0.23
0.06
363.9
0.36
0.09
0.06
3.22
0.91
0.32
0.08
373.65
0.48
0.11
0.06
5.65
1.52
0.55
0.09
376.38

0.28
0.09
0.06
2
0.61
0.35
0.07
3599.87
0.44
0.11
0.08
4.32
1.14
0.39
0.09
3599.81
0.69
0.12
0.07
6.88
1.9
0.83
0.1
3600.0

0.23
0.07
0.06
1.77
0.59
0.25
0.08
365.29
0.43
0.08
0.06
3.41
1.18
0.31
0.09
19.76
0.65
0.12
0.07
5.62
2.17
0.56
0.12
385.74

0.34
0.08
0.08
2.15
0.73
0.36
0.1
3600.3
0.53
0.11
0.07
4.44
1.34
0.38
0.11
27.86
0.87
0.13
0.09
7.02
2.86
0.84
0.17
3600.0

Table 6
The CPU times (in seconds) of the Upper Bounds - large networks.
Set C1

Set C2

Set D1

Set K1 N = 60

Set K2 N = 60

Set K1 N = 73

Set K2 N = 73

UB1
UB2
UB3
UB1
UB2
UB3
UB1
UB2
UB3
UB1
UB2
UB3

Set D2

Set D1

Set D2

Avg

Max

Avg

Max

Avg

Max

Avg

Max

4.27
1.21
0.10
13.87
5.36
0.17
6.27
1.57
0.11
19.99
5.83
0.21

5.03
1.69
0.12
18.6
6.79
0.21
8.8
2.23
0.14
23.34
8.80
0.31

4.89
1.26
0.16
19.59
4.97
0.38
7.59
1.49
0.17
32.14
6.92
0.56

5.82
1.98
0.23
25.51
6.76
0.54
9.42
2.51
0.22
42.58
10.83
0.99

2.47
0.72
0.08
6.82
2.52
0.15
3.85
0.79
0.11
9.35
3.63
0.15

2.94
1.12
0.10
8.67
3.13
0.18
4.37
1.13
0.15
10.88
5.06
0.17

3.18
0.72
0.11
9.44
2.67
0.24
4.52
0.83
0.13
13.59
3.31
0.29

3.66
1.21
0.15
11.55
3.44
0.33
5.21
1.14
0.16
17.4
4.16
0.45

N = 60

73

35
30
25
20
15
10
5
0
C1 D1
K1

C2 D1
K1

C1 D2
K1

C2 D2
K1

C1 D1
K2

C2 D1
K2

C1 D2
K2

C2 D2
K2

Fig. 6. The average CPU times (in seconds) of UB1 .

increase at a linear rate. On the contrary, the solution times of the

MILP model increase at an exponential rate and the model cannot
nd any solution in one hour, for large networks. Hence our upper
bounds could be used in place of the optimal solutions to evaluate
the performances of the lower bounds that are used to nd feasible
solutions.

We nally investigate the performances of the lower bound. For

small networks with 22, 34 and 47 tasks, the optimal solutions
are found by the MILP and the deviations are measured relative to
the optimal solutions, i.e., Percent Dev = ( OPTLB
OPT ) 100. For large
networks, the optimal solutions are not available and the deviations are measured relative to UB2 in place of optimal solutions, i.e.,

Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004

ARTICLE IN PRESS

JID: EOR
12

[m5G;October 1, 2015;11:19]

E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
Table 7
The Lower Bound performances.
Variables

Set C1

Set C2

Set D1

Set K1 N = 22

3,520

Set K2 N = 22

7,040

Set K1 N = 34

5,440

Set K2 N = 34

10,880

Set K1 N = 47

7,520

Set K2 N = 47

15,040

Set K1 N = 60

19,200

Set K2 N = 60

38,400

Set K1 N = 73

23,360

Set K2 N = 73

46,720

Percent DEV
CPU
Percent DEV
CPU
Percent DEV
CPU
Percent DEV
CPU
Percent DEV
CPU
Percent DEV
CPU
Percent GAP
CPU
Percent GAP
CPU
Percent GAP
CPU
Percent GAP
CPU

Set D2
Avg

Max

Avg

Max

Avg

Max

0
0.06
0
0.1
0
0.05
1
0.03
0
0.09
1
0.12
0
0.12
0
0.2
0
0.14
1
0.26

1
0.13
1
0.15
0
0.08
2
0.04
1
0.15
1
0.14
1
0.14
1
0.24
0
0.17
1
0.34

0
0.08
0
0.12
1
0.08
1
0.13
0
0.1
1
0.19
0.01
0.18
1
0.41
0
0.18
1
0.61

1
0.12
1
0.16
2
0.08
2
0.15
1
0.14
1
0.23
1
0.25
2
0.59
1
0.25
2
1.04

0
0.07
0
0.08
0
0.09
1
0.1
0
0.11
1
0.12
0
0.1
1
0.18
0
0.14
1
0.19

0
0.09
1
0.1
0
0.11
2
0.12
0
0.12
1
0.14
1
0.13
1
0.23
1
0.2
1
0.22

0
0.07
0
0.1
0
0.08
1
0.11
0
0.12
0
0.15
0
0.13
1
0.28
1
0.16
1
0.32

0
0.08
0
0.14
0
0.11
1
0.14
0
0.13
1
0.21
1
0.18
1
0.38
1
0.21
2
0.49

80
Average

40

Maximum

20
0
22

34

Set D2

Max

100
60

Set D1

Avg

47

Fig. 7. The optimal total net revenue values.

large networks, the optimal solutions, hence their total net revenue
values are not available.
Note from the above gure that the average optimal total net revenue values are around 30, 45 and 63 money units for 22, 34 and 47
tasks, respectively. The respective maximum total net revenue values
are 37, 54 and 78 money units and the respective maximum lower
bound deviations are around 1 percent, 2 percent and 1 percent.
These altogether imply that the practitioners that use the results of
our study would lose no more than 1000 money units over the optimal total net revenue that are of magnitude of around 50.000 money
units.
7. Conclusions

UB LB

2
Percent Gap = ( UB
) 100. The average and maximum deviations
2
of the lower bounds are reported in Table 7.
Note from the table that the lower bound solutions are quite satisfactory over all problem sets. The solutions have small deviations and
are obtained in negligible times. The performance does not deteriorate with an increase in N. However the deviations slightly increase
with an increase in K. For example for N = 60 and sets C1 and D2, the
average deviations of K1 and K2 are 1 percent, whereas the maximum
deviations are 1 percent and 2 percent for K1 and K2, respectively.
The CPU times to nd the lower bounds increase with an increase in
N and K with only a few exceptions. This is due to the fact that the
linear and mixed-integer linear programs have more decision variables for higher N and K. For example, for sets K1, C2 and D2, as N
increases from 60 to 73, the average CPU times increase from 0.13 to
0.16 seconds. For N = 47 and sets C2 and D2, increasing K from 2 to 4,
increases the average CPU times from 0.12 to 0.15 seconds.
We could not observe any notable effect of C on the lower bound
performances and CPU times. For xed N, K and C, increasing D, does
not change the performance, but increases the CPU times slightly.
This increase can be attributed to the increase in LB(Y). For example
for N = 60 and sets C1 and K2, the average CPU times are 0.20 seconds
and 0.41 seconds for set D1 and set D2, respectively.
To summarize, our computational experiment has revealed the
excellent performances of the bounds that do not deteriorate as the
number of tasks increases. On the other hand, the optimal total revenue values increase signicantly with the increases in the number of
the tasks. Fig. 7 shows the optimal total net revenue values for each
n, over all other parameter combinations, for small networks. For the

This study considers disassembly systems that have gained significant importance in recent years. This heightened importance stems
from the recognition of environmental issues and the advances in
manufacturing technologies. A disassembly line balancing problem,
where the disassembly line is congured with dened workstations,
is studied. The units of the product to be disassembled are identical and the parts have dened minimum release quantity, cost and
revenue. It is assumed that the amount released excess of minimum
release quantity can be sold at the end of the period. The aim is to
assign the tasks to the workstations so that the total net revenue is
maximized.
A MILP model that could solve the problems with up to 50 tasks,
is developed. For larger sized instances, we propose upper and lower
bounds that are motivated by the satisfactory behavior of the LPR. The
LPR is strengthened by same valid cuts.
The experimental results have revealed that the bounding mechanisms produce high quality solutions very quickly. For the maximum
trial size of 73 tasks, there is a gap of less than 5 percent between our
lower and upper bounds.
To the best of our knowledge, the study is the rst attempt to
solve the DLBP with minimum release quantities and xed number of
workstations. The proposed model assumes that all units of a specied part bring the same revenue. The presented methods can be directly applied when the unit revenues of the rst level (minimum
release quantity) and second level (excess of minimum release quantity) are different. Note that as the minimum release quantities are
xed, the contribution of the rst levels total revenue to the objective function is xed, hence irrelevant for the model solutions.

Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004

JID: EOR

ARTICLE IN PRESS
E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113

We hope that the results of our study will trigger some advances
in the disassembly literature. Some extensions of the study may consider the Successor OR type precedence relations, non-identical products (different units of the disassembly product with different parts)
and a stochastic nature of the disassembly outcome (some parts may
be found to be defective or damaged during disassembly).
References
Altekin, F. T., & Akkan, C. (2012). Task-failure-driven rebalancing of disassembly lines.
International Journal of Production Research, 50, 49554976.
Altekin, F. T., Kandiller, L., & zdemirel, N. E. (2008). Prot oriented disassembly line
balancing. International Journal of Production Research, 46, 26752693.
Bentaha, M. L., Battaa, O., & Dolgui, A. (2014a). A sample average approximation
method for disassembly line balancing problem under uncertainty. Computers and
Operations Research, 51, 111122.
Bentaha, M. L., Battaa, O., & Dolgui, A. (2014b). Disassembly line balancing and sequencing under uncertainty. Procedia CIRP, 15, 5660.
Bentaha, M. L., Battaa, O., & Dolgui, A. (2014c). Lagrangian relaxation for stochastic
disassembly line balancing problem. Procedia CIRP, 17, 239244.
Bentaha, M. L., Battaa, O., & Dolgui, A. (2014d). Disassembly line balancing problem
with xed number of workstations under uncertainty. In Proceedings of the 19th
IFAC World Congress.
Bentaha, M. L., Battaa, O., & Dolgui, A. (2015). An exact solution approach for disassembly line balancing problem under uncertainty of the task processing times.
International Journal of Production Research, 53, 18071818.
Bourjault, A. (1984). Contribution to a systematic approach of automatic generation of
task sequences. In PhD Dissertation, Universite de Franche-Comte, Besancon, France.
Brennan, L., Gupta, S. M., & Taleb, K. N. (1994). Operations planning issues in an assembly/disassembly environment. International Journal of Operations and Production Management, 14, 5767.
Capraz, O., Polat, O., & Gngr, A. (2015). Planning of waste electrical and electronic
equipment (WEEE) recycling facilities: MILP modelling and case study investigation. Flexible Services and Manufacturing Journal. doi:10.1007/s10696-015-9217-3.
Crowther, P. (1999). Designing for disassembly to extend service life and increase sustainability. In 8th International Conference on Durability of Building Materials and
Components.

[m5G;October 1, 2015;11:19]
13

Ding, L. P., Feng, Y. X., Tan, J. R., & Gao, Y. C. (2010). A new multi objective and colony algorithm for solving the disassembly line balancing problem. The International Journal of Advanced Manufacturing Technology, 48, 761771.
Gngr, A., & Gupta, S. M. (2001a). Disassembly sequence plan generation using a
branch and bound algorithm. International Journal of Production Research, 39, 481
509.
Gngr, A., & Gupta, S. M. (2001b). A solution approach to the disassembly line balancing problem in the presence of task failures. International Journal of Production
Research, 39, 14271467.
Gngr, A., & Gupta, S. M. (2002). Disassembly line in product recovery. International
Journal of Production Research, 40, 25692589.
Hezer, S., & Kara, Y. (2014). A network-based shortest route model for parallel disassembly line balancing problem. International Journal of Production Research, 53, 1849
1865.
Koc, A., Sabuncuoglu, I., & Erel, E. (2009). Two exact formulations for disassembly
line balancing problems with precedence diagram construction using an AND/OR
graph. IIE Transactions, 41, 866881.
Lambert, A. J. D. (1997). Optimal disassembly of complex products. International Journal
of Production Research, 35, 25092523.
Lambert, A. J. D. (2002). Determining optimum disassembly sequences in electronic
equipment. Computers and Industrial Engineering, 42, 553575.
Lambert, A. J. D. (2003). Disassembly sequencing: A survey. International Journal of Production Research, 41, 37213759.
McGovern, S. M., & Gupta, S. (2011). The disassembly line: Balancing and modeling. New
York: McGraw Hill, New York.
McGovern, S. M., & Gupta, S. M. (2007a). A combinatorial optimization analysis of the
unary NP-complete disassembly line balancing. International Journal of Production
Research, 45, 14851551.
McGovern, S. M., & Gupta, S. M. (2007b). A balancing method and genetic algorithm for
disassembly line balancing. European Journal of Operational Research, 179, 692708.
Navin-Chandra, D. (1994). The recovery problem in product design. Journal of Engineering Design, 5, 6586.
Paksoy, T., Gngr, A., zceylan, E., & Hancilar, A. (2013). Mixed model disassembly line
balancing problem with fuzzy goals. International Journal of Production Research, 51,
60826096.
Patterson, J. H., & Albracht, J. J. (1975). Assembly line balancing: 01 programming with
bonacci search. Operations Research, 179, 166174.
Ullerich, C. (2014). Advanced disassembly planning: Flexible, price-quantity dependent
and multi-period planning approaches. Wiesbaden: Springer Gabler.

Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004