See

discussions, stats, and author profiles for this publication at:

https://www.researchgate.net/publication/220471312

Aggregate production planning for a

continuous reconfigurable manufacturing
process

Article in Computers & Operations Research · May 2005

Impact Factor: 1.86 · DOI: 10.1016/j.cor.2003.11.001 · Source: DBLP

CITATIONS READS

22 171

2 authors, including:

Udatta S. Palekar
University of Illinois, Urbana-Champaign
35 PUBLICATIONS 451 CITATIONS

SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Udatta S. Palekar
letting you access and read them immediately. Retrieved on: 15 June 2016
 Computers & Operations Research 32 (2005) 1213 – 1236
www.elsevier.com/locate/dsw

Aggregate production planning for a continuous

recon$gurable manufacturing process
Akhil Jain, Udatta S. Palekar∗
Department of Mechanical and Industrial Engineering, University of Illinois at
Urbana-Champaign, Urbana, IL 61801, USA

Abstract

Most of the research on aggregate production planning has been focused on discrete parts manufacturing
models. In environments where intermediate inventory cannot be stored, and multiple products are produced
simultaneously using complex con$gurations of production machines, these models may produce erroneous
results. In this paper, we present a con$guration-based formulation for one such manufacturing environment,
where production may involve dissimilar machines performing similar operations at di0erent rates and equip-
ment can be connected together to form di0erent production lines. The production process is continuous and
no in-process inventory can be kept. We present and compare several heuristics to generate input data to
solve the aggregate production-planning problems using the con$guration-based formulation. Computational
experiments show that large-scale real-world problems we encountered can be solved in reasonable time using
our heuristics and commercial optimization software like CPLEX.
? 2003 Elsevier Ltd. All rights reserved.

Keywords: Aggregate planning; Integer programming; Continuous production

1. Introduction

The aggregate production-planning problem has been studied extensively, since it was $rst for-
mulated in the 1950s. Interest in the problem stems from the ability such models provide to control
production and inventory costs. The costs associated with inventory management and production
planning are a substantial portion of the total expenditure of the manufacturing companies. On av-
erage, 44.2 percent of net $xed assets of US manufacturing $rms and 16.3 percent of their total

∗
Corresponding author. Fax: +1-217-244-6534.
E-mail addresses: akhjain@hotmail.com (A. Jain), palekar@uiuc.edu (U.S. Palekar).

0305-0548/$ - see front matter ? 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cor.2003.11.001
1214 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236

assets were invested in inventory in 1986 (US Federal Trade Commission 1986, pp. 4 –5). The
value of inventories in US $rms was about $1.1 trillion in 1993. The number for Canada was
approximately $135 billion in 1994 [1]. IIE Solutions [2] reports that in only a few years, typical
spare parts inventory had increased from $2–10 million to $5 –15 million, an increase of at least
50 percent. Signi$cant savings can be realized by correctly modelling and solving the aggregate
production-planning problem.
Most of the research on this problem has been focused on models for discrete parts manufac-
turing. These models typically permit storage of in-process inventory between di0erent stages of
manufacturing. Moreover, parts are assumed to be independently processed on dedicated machines
at every stage of manufacturing. However, such models do not accurately reGect the true production
capacity when multiple products are co-manufactured using complex con$gurations of production
machines. The problem is further exacerbated when the production process involves multiple stages
and in-process inventory is not permitted.
In this paper, we consider one such situation, in which a base product is di0erentiated into multi-
ple products as it passes through several production stages. The process is a continuous production
process and no in-process inventories may be kept. Several machines are available for each stage of
production and these machines can be combined in di0erent ways to create machine con$gurations
capable of co-manufacturing groups of products at various rates. These machines may be recon$g-
ured, incurring a setup cost, into di0erent con$gurations, to produce di0erent sets of products. The
rate at which a machine produces a product therefore depends on the con$guration as well as the
products involved. We call this problem the continuous process with dissimilar machine (CDM)
aggregate planning problem.
The CDM problem arises in a number of contexts. Our model was designed for a large food-
processing company. The company has numerous production facilities of varying sizes. The number
of machines at each facility ranges between 3 and 57 machines and each facility produces between 2
and 300 stock-keeping units (SKUs). Our model can also $nd extensive application in the chemical
industry, where in-process inventories cannot be stored because of cost considerations or short shelf
life of the intermediate products. The general aggregate production-planning problem has been studied
since the 1950s. Dzielinski and Gomory [3], Lasdon and Terjung [4] and Newson [5] have developed
mathematical programming models for the problem of aggregate production planning, but for the
case where these production stages can be handled independently, i.e., storage of intermediates is
allowed. McClain et al. [6], Nahmias [7] and Silver et al. [1] also consider the problem of aggregate
production planning for production processes where the production stages can be decoupled. Nam
et al. [8] present a survey of the literature on aggregate production planning since early 1950s.
They categorize the various existing techniques into a framework depending upon their abilities to
produce an exact optimal or near-optimal solutions. Note that all these models allow the storage of
intermediates and all machines are assumed to be identical.
Although there is a vast amount of literature discussing production-planning models (e.g., see
Thomas and McClain [9] for an overview of production planning and Shapiro [10] for mathematical
programming models for production planning), few models address the CDM problem. Leachman
and Carmon [11] develop a model for identifying capacity limitations on the machines for the
above problem. Bradley and Arntzen [12] address the problem of simultaneous planning of capacity
investment, inventory and production schedule, in a scenario where all machines do not perform
similar operations. Bermon and Hood [13] developed a model for IBM where unrelated machines
 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236 1215

may perform similar operations at di0erent rates. However, both Bradley and Arntzen [12], and
Bermon and Hood [13] allow the storage of intermediate goods. These models formulate the problem
based on the available resources on each machine and the production requirements of an SKU over
di0erent machines.
In this paper, we are concerned with developing optimized and implementable production plans
for CDM. In Section 2, we present a fuller description of the problem, show that traditional models
are not adequate, and propose a new Con$guration-based model in Section 3. The details of imple-
mentation of our model follow in Section 4. In Section 5, we discuss the computational results, and
conclusions and discussions follow in Section 6.

2. Problem statement and model formulation

We consider the aggregate production-planning problem in the context of a continuous production

process in which the product is di0erentiated at various stages of manufacturing. In general, there may
be K stages of manufacturing. At each stage, the product is modi$ed by processing or mixing with
additives. Downstream machines produce at a rate which is limited by the rate at which an upstream
machine supplies products (plus any additives) and the maximum rate at which it can process the
speci$ed product. Usually the $rst or the processing stage is the most expensive machine and it must
be run at its maximum production rate. The downstream stages are then coordinated by grouping
machines such that the entire output from the upstream stages can be processed. Fig. 1 shows a
typical example of the machine arrangement that may result. Notice that the machines are con$gured
such that upstream output can be processed without intermediate storage. The base product from the
processor is mixed with the additives and converted to product 1 at the mixer. The base product
may also be used unmodi$ed, as in the case of product 2. Product 1 is further di0erentiated at the
packaging stage.
At each stage of manufacturing, there may be several dissimilar machines available to perform
the operation. These machines are selectively grouped to form con$gurations such as the one shown
in Fig. 1. The choice of machine is restricted by the availability of material-handling links between
di0erent machines, and is governed by the demand for end products.
The problem of aggregate capacity planning for such manufacturing environments is not well
researched. Most of the existing formulations assume that the production process is not continuous.
The other few models that address the CDM problem, formulate the problem based on the resource
availability at each machine. The assumptions made in such models have some limitations, which
we discuss in the next section.

2.1. Resource-based formulation

Leachman and Carmon [11] develop a model for the CDM problem, which is applicable at a
more strategic level. At the tactical level, their model may lead to infeasible plans. Besides a few
variations, their basic model in the Step-Separated Formulation is similar to the formulation we
present below. We assume the production process to be a continuous process. Hence, the lead times
from one production stage to another are assumed to be zero. Leachman and Carmon [11] do not
consider setup times, which are modelled in this formulation. Also, we assume that production of
1216 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236

Fig. 1. A typical con$guration of machines.

some SKUs may require some other constituent SKUs as an input. To account for the di0erent
production requirements of the SKUs, the resource-based formulation models the problem based on
the availability of resources on each machine. It takes into account the resource requirements of the
SKU on di0erent machines and the total resource availability on each machine. The resource-based
formulation is primarily concerned with ensuring that adequate machine capacity is available.
Let N be the number of SKUs, M be the total number of machines, T be the length of planning
horizon and xit , the amount of SKU i produced in period t; bik , production requirements, in fraction
of available time, per unit amount of SKU i on machine k;


1 SKU i is scheduled for production on machine k in period t;
yikt =
0 otherwise;

Sikt , setup time for production of SKU i on machine k in period t; dit , demand for SKU i in period
t; Iit , inventory level of SKU i at the end of period t; wij , amount of SKU i required to produce
one unit of aggregate SKU j; Cit , production cost per unit amount of SKU i in period t; Hit ,
inventory-holding cost per unit amount of SKU i in period t.
 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236 1217

2.2. The resource-based model

Minimize
N T N 
 T
Cit xit + Hit Iit (1)
i=1 t=1 i=1 t=1
s.t.
N
 N

bik xit + Sikt yikt 6 1; ∀k ∈ M ; t ∈ T; (2)
i=1 i=1
xit 6 yikt ; ∀i ∈ N ; k ∈ M ; t ∈ T; (3)
N

wij xjt 6 Iit −1 + xit ; ∀i ∈ N ; t ∈ T; (4)
j=1;j
=i
N

Iit −1 + xit − dit − wij xjt = Iit ; ∀i ∈ N ; t ∈ T;
j=1;i
=j

xit ; Iit ¿ 0; ∀i ∈ N ; t ∈ T;
yikt ∈ (0; 1); ∀i ∈ N ; k ∈ M ; t ∈ T: (5)
In the above formulation, the objective is to minimize the total costs of production and inventory
storage. The $rst term of Eq. (1), i.e., the objective function, represents the production costs incurred
by the facility, while the other term in the equation gives the total holding costs over the planning
horizon. Constraint (2) limits the production on a machine to its available production capacity.
Constraint (3) relates the production with setup times. Constraint (4) ensures the availability of
constituent SKUs for the production of the aggregate SKUs. We assume that the SKUs produced
in a period can be utilized for production of other SKUs in the same period. Constraint (4) is only
important if i is an aggregate SKU. Otherwise, the wij are all equal to zero and the constraint is
redundant since the both inventory and production quantities  are required to be non-negative. Eq. (5)
is the typical inventory balance constraint. The extra term Nj=1;i
=j wij xjt denotes the consumption
of SKU i for the production of aggregate SKUs.
Inherent to this model are several assumptions, which include:

(1) It is assumed that each machine works at its full production capacity. In production processes,
where the production stages can be decoupled due to the presence of work in process inventory,
this assumption holds. But in CDM processes, the output rates of the machines may have to
be Gexed to ensure the continuity of the Gow of the material. In these situations, this model
may give an infeasible detailed production plan.
(2) The idea of associating setup times with each machine–SKU combination does not correctly
model the setup time. Since the setup time actually involves connecting machines together in
addition to clean-up and maintenance, the machines share a common setup time. This time
depends on the particular machines being connected as well as the SKUs being produced. This
joint setup dependence distinguishes the problem from other more traditional setup times.
1218 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236

Fig. 2. Con$guration for production of SKU1/SKU2.

To understand the impact of assumption 1, consider the following example: Consider a con$guration
consisting of two machines, machine A and machine B, as shown in Fig. 2.
Production of both SKU1 and SKU2 requires operations $rst on machine A and then subsequent
operations on machine B. The production processes are assumed to be continuous. The processing
rate of machine A for SKU1 is 40 units=day and for SKU2 is 60 units=day. The processing rates for
machine B are 60 and 45 units=day, respectively. For simplicity, assume that the setup time on all
the machines is zero. Assume that the demand is 20 units for SKU1 and 30 units for SKU2. The
above model will signal suNcient capacity for production, and present the following results:

• Schedule machine A for half day for production of SKU1 and the rest of the day for production
of SKU2.
• Schedule machine B for one-third day for production of SKU1 and the rest of the day for pro-
duction of SKU2.

However, continuity of the production processes will impose a restriction on the production rate of
machine B for production of SKU1. Machine B will process SKU1 at lower rate of 40 units=day
instead of the rate of 60 units=day. This in turn makes the available production capacity insuNcient
to meet the demand of the SKUs.
This problem leads to the idea of generating an aggregate production plan taking into account
the con$gurations of machines instead of individual machines, and to develop a plan based on the
actual processing rates of the machines rather than their maximum processing rates.
In the next section, we propose a Con$guration-based model that overcomes these limitations.

3. Conguration-based formulation

As mentioned earlier, a major limitation of the resource-based formulation is the assumption that
all machines produce at their maximum capacity. Moreover, it is possible that a machine performing
a particular operation cannot be linked to all the machines capable of performing the next operation.
This creates a restriction on the availability of machines for the production of the SKU, which is
not accounted for in the resource-based formulation. To overcome these limitations, our formulation
relies on the net outputs of machine con$gurations instead of the production capacities of individual
 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236 1219

machines. An example of a con$guration is shown in Fig. 1. Note that this con$guration may
produce more than one combination of SKUs. We term each such combination of SKUs as a nest
of SKUs. This terminology of con$guration–nest combination was introduced by Bhatia and Palekar
[14], in the context of lot-sizing for jointly manufactured parts. This methodology not only allows us
to model the production process as a continuous process, but also to model the production process
based on the actual production rates of the machines instead of their maximum production rates.

3.1. Assumptions

To model the CDM problem, we make the following assumptions:

(1) Raw materials are always available.

(2) Demand of the end products/SKUs is deterministic.
(3) Processing time of the machines and their yields are deterministic.
(4) Lead times for shipment of products/SKUs are deterministic.
(5) Shortages are permitted but are discouraged. The shortage demand is shipped from a dummy
warehouse, that has a very high shipment costs.
(6) Intermediate SKUs produced in a particular time period are available in the same period to
other processing units, for subsequent operations.
(7) In the model, we distinguish two kinds of setup, i.e., major and minor setup. If the change-over
from one production setup to another means only a change in packaging machines and the set
of SKUs, but the base item to be produced remains the same, the setup is called a minor setup.
On the other hand, if the change involves the production of di0erent base item, i.e., the change
of the processing machine, the setup is called a major setup.
(8) We assume that the setup times are not sequence dependent.

3.2. Model formulation

To elaborate the concept of con$guration and for the sake of clarity, we will $rst present the
constraints associated with a single con$guration and then extend the model to a multi-con$guration
problem.
Consider the con$gurations I and II shown in Fig. 3. Con$guration I consists of $ve machines,
i.e., processing machine A, mixing machine B and processing machines C, D and E. Let rIi be the
production capacity of machine i ∈ {A; B; C; D; E}. The output from the processing machine A is
distributed across the packaging machine E and the mixing machine B. The machine B also receives
an additional external input at the rate . It mixes the two inputs and feeds the packaging machines
C and D. Note that, in the entire process, there is continuous Gow of material and no inventory
is being stored at any intermediate stage. There are three SKUs that are produced simultaneously
by this con$guration. Let their output amounts be variables oI1 ; oI2 and oI3 , i.e., the outputs from
machine C, D and E, respectively. These amounts are dependant on the production rates of the
machines and the production time allotted to this con$guration.
Con$guration II consists of three machines, i.e., processing machine A and the packaging machines
C and D. Let rIIi be the production capacity of machine i ∈ {A; C; D}. The output from the processing
machine A is distributed across packaging machines C and D. Let the output amounts from these
1220 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236

Fig. 3. A multi-con$guration framework.

machines be the variables oII1 and oII2 , respectively. Assume that the SKUs produced by both the
con$gurations belong to the same product family. The constraints related to these con$gurations can
be classi$ed into the following categories:

(1) Capacity constraints: The time for which a machine can be utilized, i.e., the sum of the setup
times and the production times on the machine, cannot exceed the available production time
on the machine. Let yI and yII be the fraction of the available time in period t allotted for
production to con$guration I and con$guration II, respectively. Let WI and WII be the minor
setup times for the above con$gurations, and wI and wII be the {0; 1} minor setup variable for
the above con$gurations I and II for the given set of SKUs. Let Z and z be the major setup
times and {0; 1} variables, respectively, for these con$gurations. Note that there is only one
major setup for the two con$gurations. The capacity constraints for the above con$gurations
are:
Machine A : yI + yII + WI wI + WII wII + Zz 6 1; (6)

Machine B : yI + WI wI 6 1; (7)

Machine C : yI + yII + WI wI + WII wII 6 1; (8)

Machine D : yI + yII + WI wI + WII wII 6 1; (9)

 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236 1221

Machine E : yI + WI wI 6 1; (10)

yI 6 wI ; (11)

yI 6 zI ; (12)

yII 6 wII ; (13)

yII 6 zII : (14)

Constraints (6)–(10) limit the production time on a con$guration to the available time on the
constituent machines. Constraints (11)–(14) relate the production in period t to setting up in
that period.
(2) Production constraints: The production constraints relate the output amounts from a con$gu-
ration to the production times and the output rates of the con$guration. These constraints are
simply the mass balance constraints equating the outputs from each machine of a con$guration
to the $nal outputs of the con$guration, i.e., oI1 ; oI2 and oI3 in the case of con$guration I and
oII1 and oII2 in the case of con$guration II. Let Lm be the set of the leaves that are children
of machine m. The production constraint equates the output of machine m to its $nal children
outputs. The sign in this constraint is an equality if the output rate of machine m cannot be
Gexed, and an inequality otherwise. Assuming that the output rate of machine A cannot be
Gexed while that of other machines may be varied, the constraints are:
For Con$guration I:
oI1 6 rIC yI ; (15)

oI2 6 rID yI ; (16)

oI3 6 rIE yI ; (17)

oI1 + oI2 6 rIB yI ; (18)

oI1 + oI2 + oI3 = (rIA + )yI : (19)

For Con$guration II:
oII1 6 rIIC yII ; (20)

oII2 6 rIID yII ; (21)

oII1 + oII2 = rIIA yII : (22)

Next, we present the formulation for the multi-con$guration problem.

1222 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236

3.2.1. Notation
N total number of SKUs, including the intermediate SKUs
M total number of machines available at the production facility
T total length of the planning horizon
P total number of possible con$guration–nest combinations at the production facility
(In our model, we denote each con$guration–nest combination with a unique index.)
F total number of product families
yik fraction of time allotted to con$guration–nest i in the production schedule for period

k
1 con$guration–nest i is scheduled for production in period k
xik =

0 otherwise
1 product family f is scheduled for production on processor p in period k
zfpk =
0 otherwise
Cik
production cost per unit time for con$guration–nest i for period k
1 processor p is present in the element list of con$guration–nest i
aip =
0 otherwise
tik minor setup time for con$guration–nest i in period k
t1fpk major setup time for product family f on processor p in period k
mik minor setup cost per unit time for scheduling con$guration–nest i in period k
Mfpk major setup cost for scheduling product family f on processor p in period k
Oijk jth output leaf of con$guration i in period k
L1ip set of output leaves that are children to processor p, where p is in the element list
of con$guration i
ij jth external input to con$guration–nest i
L2ip set of external inputs on the paths from the processor p to its children leaves L1ip ,
where p is in the element list of con$guration–nest i
L3is set of output leaves in con$guration–nest i that produce SKU s
Rip maximum output rate of processor p on con$guration–nest i
COsk total production of SKU s in period k
L4s set of SKUs required for the production of SKU s
Wtss
amount in weight of SKU s required to produce one unit weight of SKU s
S set of intermediate SKUs
Isk inventory of SKU s at the end of period k
Hsk holding cost for SKU s in period k
Dsk demand of SKU s in period k

3.3. The con$guration-based model

Minimize
 T
P   T
P   M 
F  T  T
N 
Cik yik + mik xik + Mfpk zfpk + Hsk Isk ; (23)
i=1 k=1 i=1 k=1 f=1 p=1 k=1 s=1 k=1
 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236 1223

s.t.
P

aip (yik + tik xik + t1fpk zfpk ) 6 1 ∀f ∈ F; p ∈ M ; k ∈ T; (24)
i=1

yik 6 xik ∀i ∈ P; k ∈ T ; (25)

yik 6 zfpk ∀i|i ∈ f; ∀f ∈ F; k ∈ T; (26)

 
 
Oijk 6 Rip yik +  ij  yik ∀p|aip = 1 ∀i ∈ P; k ∈ T; (27)
j |j ∈L1ip j |j ∈L2ip

P
 
Oijk = COsk ∀s ∈ N ; k ∈ T; (28)
i=1 j |j ∈L3is


COsk Wtss
6 Is
(k −1) + COs
k ∀s ∈ S; k ∈ T; (29)
s|s
∈L4s

Is(k −1) + COsk − COs
k Wts
s − Dsk = Isk ∀s ∈ N ; k ∈ T; (30)
s
|s∈L4s

yik ; Isk ¿ 0 ∀i ∈ P; s ∈ N ; k ∈ T;

xik ; xfpk ∈ (0; 1) ∀i ∈ P; f ∈ F; p ∈ M ; k ∈ T;

In the above mixed integer-programming formulation, the objective is to minimize the total pro-
duction and inventory-carrying costs. The $rst three terms in objective function (23), represent the
production costs. The $rst term in calculates the cost of production, while the next two terms com-
pute the setup costs. The last term in the equation determines the inventory-holding costs. Constraints
(24)–(26) play similar roles as constraints (6)–(14). Constraint (24) ensures that no machine is uti-
lized more than its availability. Constraints (25) and (26) relate the production in a period to the
setup times.
Constraint (27) is similar to constraints (15)–(22). It relates the outputs from a con$guration to
the processing rates of the machines in the con$guration and the external inputs. Note that the sign
in constraint (27) will be an equality if the rate of the machine p in con$guration i in period
k cannot be Gexed and will be an inequality otherwise. Eq. (28) calculates the total output of a
SKU from di0erent con$gurations in a production facility. Constraint (29) ensures the availability
of intermediate SKUs for producing the aggregate SKU s. The left-hand side of the constraint gives
the amount of intermediate SKU s required to produce the aggregate SKUs in period k, while the
right-hand side gives the total amount of SKU s available for the production of the aggregate SKU s
in the given period. Eq. (30) maintains the inventory balance between the inventory at the beginning
of the period, the amount produced in the period, the demand ful$lled, and the inventory at the end
of the period.
1224 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236

Now let us reconsider the example mentioned in Section 2.1 where we have shown how the
resource-based formulation may lead to infeasible plans. Resolving this problem using the above
formulation will signal insuNcient capacity as the rate of machine B will be limited to 40 units=day
for SKU1 because of the low processing rate of machine A.
As seen from the above example, the con$guration-based formulation produces better results
than resource-based formulation. However, a major drawback of the con$guration-based formula-
tion is the size of the problem, which may render this formulation impractical to use because
of large computation times involved. Consider a production facility having one processing ma-
chine and 10 packaging machines. The facility produces 10 di0erent SKUs. Assume that each
packaging machine has similar production capacities and can be linked to the processing ma-
chine. Also, each packaging machine can pack any of the 10 SKUs. If we model this prob-
lem using the resource-based formulation, the number of yikt variables (Section 2.1) for period
t is 110. In comparison, the number of yit variables for period t for the above formulation will
be (( 101
) × ( 10
1
) + · · · + ( 10
10
) × ( 10
10
)) = 184; 755. But a close inspection of this number reveals
that a lot of these variables are unnecessary. For example, in the organization for which this
model was developed, physical considerations limited the number of packaging machines that can
be linked to a processing machine at any time to between 4 and 5. Also, we can exploit a
symmetry in the structure of the problem and further limit the number of variables. If we as-
sume that production costs for each packaging machine are the same, the solution to this prob-
lem will remain unchanged if we assign a unique SKU to every packaging machine, i.e., each
packaging machine can pack only a unique SKU. This will reduce the number of nest of SKUs
for each con$guration to one, reducing the number of variables from 184,755 to (( 10 1
) + ··· +
( 10
10
) = 1023), keeping the solution value same. While pruning, the number of con$guration–nest
combinations, we need to ensure that we do not omit any necessary con$guration resulting in
heavy burdening of some machines. In the next section, we will present di0erent techniques to
reduce the number of con$guration–nest combinations to generate a “good” set of con$guration–nest
combinations.

4. Limiting conguration–nest combinations

As mentioned in the previous section, generation of a “good” set of con$guration–nest combina-

tions is very important for the e0ectiveness of our formulation. A very large set of con$guration–
nest combinations may be computationally infeasible. At the same time, omitting some essential
con$guration–nest combinations may result in infeasible solutions or heavy burdening of some ma-
chines. Hence, the aim while generating this “good” set of con$guration–nest combinations is that
the set should include all the combinations that are likely to be selected in the production plan. A
closer inspection of the problem shows that we can generate con$guration–nest combinations based
on the raw production data and can re$ne the set depending on the demands in various periods.
For example, initially we may generate a large combination of nests of SKUs for a particular con-
$guration. But we can remove the nest of SKUs that have no demand or a nest of SKUs where
high demand SKUs are combined with some low demand SKUs. In the next section, we discuss the
methods for obtaining this “large” set, and in the subsequent section we present a methodology to
further prune this set with reference to the demand.
 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236 1225

Fig. 4. SKUs and machines to con$guration–nest combinations.

4.1. Limiting con$guration–nest combinations—Step 1

Fig. 4 shows the possible links between the SKUs and the machines, any subset of which will
result in a feasible con$guration–nest combination. The aim is to remove some of the links without
materially altering the solution. This process can be accomplished in multiple stages, where each
stage will include elimination of links at di0erent level. The possible levels include:

(1) Elimination of links between machines, resulting in limiting the number of con$gurations of
machines. There are several methods that can be employed to limit the number of con$gurations.
Restricting the maximum number of machines in a con$guration is one simple way to limit the
number of con$gurations. As stated earlier, the company for which this model was developed
preferred to have at most $ve packaging machines to be attached to a processing machine at
any time. This preference was a result of various physical limitations in linking machines.
A second method to restrict the number of con$gurations is by limiting the capacity under-
utilization allowed for any con$guration. As described in Section 2, a con$guration is a set of
machines where each machine belongs to a speci$c stage of manufacturing. Capacity under-utiliza-
tion may be de$ned as the ratio of maximum production capacity to the minimum production
capacity of all the stages of manufacturing. In a con$guration, it is possible that all the stages
1226 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236

of manufacturing are not present on all the paths from the initial manufacturing stage to the
$nal manufacturing stage. For example, consider three stages of manufacturing, i.e., processing,
mixing and packaging. In a con$guration, it is possible that some paths include all the stages of
manufacturing, while another path may include only the processing and the packaging machine.
In such scenarios, we assume that a dummy mixing stage is present in this path, between the
processing and the packaging machines. The output rate of this dummy is assumed to be the
minimum of the rates of the processing and the packaging machines. For example, con$guration
in Fig. 1 consists of three stages, i.e.,
• Stage 1: processing machine: production capacity = 100;
• Stage 2: mixing machines: production capacity = 60 + 60 + 30 = 150;
• Stage 3: packaging machines: production capacity = 50 + 20 + 30 + 30 + 30 = 160.
Note that in Stage 2, the production rate of 30 is the dummy production rate. Hence the capacity
under-utilization for this con$guration is production capacity at Stage 3
production capacity at Stage 1
= 160
100
= 1:6.
(2) Elimination of links between nests of SKUs and the con$guration of machines, resulting in
limiting the number of con$guration–nest combinations. One simple way of limiting this number
is by constraining the number of nests for each con$guration. We can specify the number of
nests that may have a speci$c SKU for a particular con$guration. This technique for restricting
the problem size is especially e0ective when the total number of nests that can be produced on
a con$guration is very large.
Another way is to limit the number of con$gurations on which a particular nest of SKUs can
be produced. To solve this problem, consider the bipartite graph in Fig. 4, with all the possible
nests of SKUs as nodes on one side and the con$guration of machines as the nodes on the other
side. There exist a link between a nest and a con$guration if that particular nest of SKUs can
be produced on that con$guration. We can model the problem as a network Gow problem, as
formulated below:
Let N be the total number of nests and C be the total number of con$gurations and ci the
number of nests that can be produced on con$guration i; nj the number of con$gurations that
can produce the nest j; Mj the minimum number of con$gurations that should be retained for
the production for nest j. This number is the minimum of the number of links available to nest
j and a used de$ned number;


1 there exists a link between con$guration i and nest j;
aij =
0 otherwise:


1 if the link between the con$guration i and nest j is retained;
xij =
0 otherwise:

Minimize

 N
C 
ci nj xij (31)
i=1 j=1
 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236 1227

s.t.
N

aij xij ¿ 1; ∀i ∈ C; (32)
j=1
N

aij xij ¿ Mj ; ∀j ∈ N: (33)
i=1

In the above formulation, we intend to spread the nests across the con$gurations as uniformly as
possible so that none of the con$gurations is over-burdened during the production process. The
objective function of the above formulation, i.e., Eq. (31) presents one such way to achieve this
aim. Each link is assigned a weight that is equal to the product of the number of nests ci that can
be produced on the linking con$guration and the number of con$gurations nj that can produce
the corresponding nest. Constraint (32) ensures that each con$guration is retained during the
elimination process. Constraint (33) ensures that there are suNcient number of con$gurations
available to produce a nest. Eq. (31), in conjunction to constraints (32) and (33), distributes
the nests across di0erent con$gurations, while trying to keep the number of nests for each
con$guration same.

4.2. Limiting con$guration–nest combinations—Step 2

The methods in the previous section only consider the combinatorial possibilities for nests and
con$gurations, but ignore product demands. For example, the above methods will not eliminate a
combination, just because there is no demand for the output SKUs of the combination. Hence, we
need to devise a method to $lter out these combinations before actually solving the multi-period
aggregate production-planning problem. Here we can exploit the fact that the solution time of a
mixed integer programming (MIP) increases exponentially with the increase in the number of vari-
ables. Hence, solving n single-period problems is computationally much easier than solving a single
n-period problem. Therefore, one way to eliminate the con$guration–nest combinations is to solve
a single-period problem for each period and use only those combinations that were allotted positive
production time in these problems. This procedure of solving single-period problems does not accom-
modate the fact that inventory can be transferred across the periods. Hence, it may eliminate those
con$gurations that may cater to the demand of the current period and some future periods by hold-
ing inventory for the future periods. One way to overcome this limitation is to solve “aggregated”
single-period problems over rolling horizon. Here we can aggregate the demand of k continuous
periods, where k can be the average shelf life of SKUs, and solve a single period problem with this
value of demand. Note that this “aggregation” of demand tends to smoothen the variation in demand
over the periods. Hence the combinations generated may not accommodate the variation in demand
over the periods. This may result in omission of some combinations that describe the demands of
individual periods. For example, consider the following combinations, with same production costs
and setup times:
(1) Combination 1: produces SKU1 at the rate of 100 lb=period.
(2) Combination 2: produces SKU2 at the rate of 100 lb=period.
(3) Combination 3: produces SKU1 and SKU2 simultaneously, both at the rate of 50 lb=period.
1228 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236

Table 1
Maximum-packaging rates of the SKUs

SKU Packaging rate (units/day)

A 2
B 20
C 40
D 60

Assume that the demand is 100 lb for SKU1 in period 1 and 100 lb of SKU2 in period 2. If
we aggregate the demand of the two periods and solve the problem as a single-period problem,
combination 3 will be selected for production of the SKUs which is not the optimal solution for
the multi-period problem. Hence, this method of “aggregation” of demand may omit some essential
con$gurations. This problem can be solved by designing a proper aggregation scheme, as we discuss
in next section.
In the next section, we present computational results for solving the CDM problem based on the
above methods of elimination of con$guration–nest combinations.

5. Computational results

In order to test the e0ectiveness of the methods discussed in Section 4 for limiting the number of
con$guration–nest combinations, we perform computational experiments on real-industry data sets
and on randomly generated data sets. For the randomly generated data sets, we perform experiments
on the following data:
For our experiments, we consider a facility with two processing machines and six packaging
machines. All six packaging machines are identical and can be linked to either of the two processing
machines. The facility produces four di0erent SKUs that belong to the same product family. Each
of the four SKUs can be produced at the two processing machines and can be packaged at any of
the four packaging machines. The processing machines can produce at a $xed rate of 60 units=day
and the rates of the packaging machines are dependent on the SKUs packed. Also, we assume that
the rates of the packaging machines can be varied. The maximum packaging rates of the SKUs are
given in Table 1.
We consider a six-period capacity-planning problem. We perform experiments on four di0erent
kinds of data sets, where the demand in each problem is a fraction of the total processing rate.
For our analysis, we vary the fractions from 0.2 to 0.8 with an increment of 0.2. This demand
is distributed across the SKUs in the ratio of the maximum packaging rate of the SKU to the
2 2
total packaging rate for the SKUs. For example, for SKU A this ratio will be 2+20+40+60 = 122 .
We uniformly perturb the demand of these SKUs to introduce randomness in the data. We create
problems at two levels of perturbation, high, where the demand of the SKU is 1 ± 0:2U times the
original demand, and low, where the demand of the SKU is 1 ± 0:05U times the original demand.
Here U is a Uniform (0, 1) random variable. For each experimental condition, we generate three
such problem instances. Table 2 gives the restricting values we choose for the methods discussed
 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236 1229

Table 2
Restricting con$guration–nest combinations

Con$guration–Nest elimination technique Restricting value

Number of machines in a con$guration 4

Capacity under-utilization of a con$guration–nest combination 2.0
Number of nests in a con$guration for a SKU 5
Number of con$gurations for a nest 4

in Section 4. All the problems are solved using CPLEX v7.0 on a Pentium III Xeon Processor
550 MHz machine with 1 GB of RAM. A 2-h CPU time limit was imposed for each solution of the
problem. To present the computational results in Tables 3 and 4, we use the following notation:
Conguration–Nest elimination technique Notation in the results
Number of machines in a con$guration M/C
Number of nests in a con$guration–nest combination SKU
Capacity under-utilization of a con$guration–nest combination Capacity
Number of con$gurations for a nest Network
Selection of combinations based on the demand Aggregation

To eliminate the con$guration–nest combinations using the aggregation method, we need not solve
the MIP problems to optimality. The aim is to generate a good representation for the demand. For
our experiments, we solve the linear relaxation to the aggregate single period problems with rolling
horizons.
Table 3 presents computation times for the above problems and Table 4 lists the number of
con$guration–nest combinations used for solving the “$nal” multi-period capacity-planning problems.
Note that the computation times for the demand-based elimination techniques are included in the
times presented in Table 3. The reason being that these elimination techniques need to be executed
with every change in demand, unlike the other techniques, which need to be implemented only
once. For the experiments where we could not solve all the three problem instances, we present the
number of problems solved in the brackets, next to the average results for other experiments of the
kind. We state NS (not solved) if none of the three problem instances could be solved.
The results obtained from the runs were similar for the case where the demand perturbation is
high and the case where the demand perturbation is low. In both scenarios, both demand intensity
as well as the number of con$guration–nest combinations together determine the computation times
for solving the problems. As documented in Table 3, none of the problems were solved when no
elimination technique was used. In contrast, the computational times were {0:3; 0:0; 0:3; 26:3} for
high demand perturbation and {0:7; 0:3; 0:7; 0:0} for low demand perturbation, for demand intensity
{0:2; 0:4; 0:6; 0:8}, respectively, when all the elimination techniques were implemented. The reason
becomes obvious from Table 4, which presents the number of combinations for each problem.
For the case where no elimination technique was exploited, there were 29,874 combinations. For
the all-techniques case, there were 82 combinations prior to the implementation of aggregation
method. Note that these 82 combinations were used to solve only the single-period problems (i.e.,
the aggregation method) and the set was reduced to {7; 7; 26; 12} for high demand perturbation and
1230 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236

Table 3
Computation times in CPU secs

High demand perturbation Low demand perturbation

Aggregation Aggregation

No Yes No Yes

D Yes 0.7 0.3 1.0 0.7

E Capacity Yes Network No 338.7 1.0 869.0 1.0
M Yes No Network Yes 0.7 0.3 1.0 0.3
A Yes SKU No 337.7 1.0 866.7 1.0
N Yes 1.0 0.0 0.7 0.3
D No Capacity Yes Network No 4535.3 (1) 163.7 4670.0 163.0
I No Network Yes 0.7 0.0 0.7 0.7
N M/C No NS 225.7 2443.0 222.0
T Yes 1.7 0.0 2.0 0.0
E Capacity Yes Network No 1667.0 83.0 1259.0 91.7
N Yes No Network Yes 1.3 0.3 1.7 0.0
S No SKU No 1381.0 115.3 1054.0 (2) 127.3
I Yes 4.3 0.3 3.7 0.0
T No Capacity Yes Network No NS NS NS 3297.0 (2)
Y No Network Yes 23.7 1.0 32.7 0.7
0.2 No NS NS NS NS

D Yes 1.7 0.0 1.3 0.3

E Capacity Yes Network No 461.3 1.3 744.7 1.7
M Yes No Network Yes 1.3 0.3 1.3 0.3
A Yes SKU No 488.0 1.7 745.0 1.0
N Yes 1.0 0.0 1.3 0.0
D No Capacity Yes Network No NS 145.0 NS 142.7
I No Network Yes 1.3 0.7 1.0 0.0
N M/C No NS 202.0 NS 193.7
T Yes 6.3 1.0 1.6 0.0
E Capacity Yes Network No 1142.3 95.3 2104.0 87.0
N Yes No Network Yes 2.0 0.7 3.3 0.0
S No SKU No 6132.0 128.3 136.0 (2) 113.0
I Yes 4.3 0.3 4.0 0.3
T No Capacity Yes Network No NS NS NS NS
Y No Network Yes 63.7 1.0 15.3 1.0
0.4 No NS NS NS NS

D Yes 5.0 0.3 2.0 0.7

E Capacity Yes Network No 4713.0 (1) 85.0 NS 1.7
M Yes No Network Yes 5.0 0.7 2.3 0.0
A Yes SKU No 4714.0 (1) 89.3 5794.0 1.7
N Yes 3.0 0.7 5.7 0.3
D No Capacity Yes Network No NS 4296.0 NS 6035.5 (2)
I No Network Yes 2.0 0.0 2.7 0.7
N M/C No NS 1611.7 NS 2033.7
T Yes NS 0.7 22.0 1.0
E Capacity Yes Network No 4475.0 (1) 104.3 5406.0 (1) 123.0
N Yes No Network Yes 0.3 0.3 3.5 0.3
 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236 1231

Table 3 (continued)

High demand perturbation Low demand perturbation

Aggregation Aggregation

No Yes No Yes

S No SKU No NS 194.3 NS 315.0

I Yes 14.7 1.0 21.0 1.0
T No Capacity Yes Network No NS NS NS NS
Y No Network Yes 2522.0 1.3 3140.7 1.7
0.6 No NS NS NS NS

D Yes 11.7 26.3 9.7 0.0

E Capacity Yes Network No NS 71.7 NS 1.7
M Yes No Network Yes 64.7 27.0 9.7 0.3
A Yes SKU No NS 29.3 NS 2.0
N Yes 76.0 2.0 13.0 0.0
D No Capacity Yes Network No NS NS NS 3491.5 (2)
I No Network Yes 147.3 12.0 12.3 1.6
N M/C No NS 4890.0 (1) NS 251.0 (1)
T Yes 346.0 (2) 0.7 40.0 0.7
E Capacity Yes Network No NS 2168.3 NS 394.7
N Yes No Network Yes 609.7 2.3 16.7 0.3
S No SKU No NS 478.7 NS 829.0
I Yes 4000.0 13.7 54.7 1.3
T No Capacity Yes Network No NS NS NS NS
Y No Network Yes NS 21.0 NS 1.3
0.8 No NS NS NS NS

{7; 7; 19; 10} for low demand perturbation, for demand intensity {0:2; 0:4; 0:6; 0:8}, respectively, for
solving the multi-period problem. A larger number of combinations imply more integer variables and
typically a harder MIP to solve. For a given demand intensity, the computation time increases with
increase in the number of con$gurations. Consider the example of demand intensity 0.4 and high
demand perturbation, the computational times for the problem are ranged from a few seconds to no
problems being solved in the allotted time. This is easily explained by the number of con$guration–
nest combinations which were as few as 82 for the easy problems and as many as 29,874 for the
hard problems. Table 4 gives the average number of con$guration–nest combinations for each case.
Our experiments show that demand intensity plays an important role in determining the solution
times. A more capacitated problem is likely to require higher computational time for the same set of
con$guration–nest combinations. For example, consider the elimination technique where all, but Net-
work and aggregation technique were implemented. There were 852 combinations remaining from
this elimination technique. The computational times for high demand perturbation with demand inten-
sity {0:2; 0:4; 0:6; 0:8} were {338:7; 461:3; 4713:0(1); NS}. Note that these times are increasing with
increase in demand intensity for the same number of combinations. In fact, our experiments show
that the e0ect of capacity is more pronounced compared to the e0ect of number of con$guration–
1232 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236

Table 4
Number of con$guration–nest combinations

High demand perturbation Low demand perturbation

Aggregation Aggregation

No Yes No Yes

D Yes 82 7 82 7
E Capacity Yes Network No 852 11 852 7
M Yes No Network Yes 82 7 82 7
A Yes SKU No 852 11 852 7
N Yes 86 7 86 7
D No Capacity Yes Network No 1952 215 1952 242
I No Network Yes 106 13 106 13
N M/C No 2472 200 2472 196
T Yes 148 7 148 6
E Capacity Yes Network No 1234 234 1234 240
N Yes No Network Yes 178 6 178 6
S No SKU No 1428 211 1428 213
I Yes 240 7 240 8
T No Capacity Yes Network No 8240 1471 (2) 8240 1604 (2)
Y No Network Yes 752 6 752 6
0.2 No 29874 29874 29874 29874

D Yes 82 7 82 7
E Capacity Yes Network No 852 11 852 12
M Yes No Network Yes 82 7 82 7
A Yes SKU No 852 11 852 12
N Yes 86 7 86 7
D No Capacity Yes Network No 1952 225 1952 226
I No Network Yes 106 13 106 13
N M/C No 2472 186 2472 199
T Yes 148 6 148 6
E Capacity Yes Network No 1234 234 1234 238
N Yes No Network Yes 178 6 178 6
S No SKU No 1428 210 1428 212
I Yes 240 7 240 8
T No Capacity Yes Network No 8240 1477 (2) 8240 1492 (2)
Y No Network Yes 752 6 752 6
0.4 No 29874 29874 29874 29874

D Yes 82 26 82 19
E Capacity Yes Network No 852 122 852 27
M Yes No Network Yes 82 26 82 19
A Yes SKU No 852 122 852 27
N Yes 86 7 86 7
D No Capacity Yes Network No 1952 1179 1952 994
I No Network Yes 106 14 106 14
N M/C No 2472 690 2472 644
T Yes 148 6 148 18
E Capacity Yes Network No 1234 342 1234 329
N Yes No Network Yes 178 19 178 15
 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236 1233

Table 4 (continued)

High demand perturbation Low demand perturbation

Aggregation Aggregation

No Yes No Yes

S No SKU No 1428 450 1428 433

I Yes 240 16 240 18
T No Capacity Yes Network No 8240 3372 8240 3421
Y No Network Yes 752 8 752 7
0.6 No 29874 29874 29874 29874

D Yes 82 12 82 10
E Capacity Yes Network No 852 21 852 20
M Yes No Network Yes 82 12 82 10
A Yes SKU No 852 21 852 20
N Yes 86 14 86 11
D No Capacity Yes Network No 1952 596 1952 490
I No Network Yes 106 16 106 15
N M/C No 2472 700 2472 669
T Yes 148 16 148 12
E Capacity Yes Network No 1234 345 1234 364
N Yes No Network Yes 178 10 178 9
S No SKU No 1428 325 1428 329
I Yes 240 12 240 17
T No Capacity Yes Network No 8240 2567 8240 2868
Y No Network Yes 752 11 752 11
0.8 No 29874 29874 29874 29874

Table 5
E0ect of elimination techniques (CPU secs.)

M/C SKU Capacity Network Aggregation All None

Applied 1561.8 1209.0 2135.9 294.2 1211.4 3.58 7200

Not applied 2734.4 3087.2 2160.3 4002.0 3084.8 2217.28 1985.14

nest combinations. Compare the computational times for demand intensity 0.6 to demand intensity
0.8. The computation times for demand intensity 0.8 are typically larger than those of 0.6 demand
intensity even though the number of con$guration–nest combinations are greater for 0.6 demand
intensity.
A summary table of the e0ect of each elimination technique on computation time in seconds is
given in Table 5. Since the maximum time allotted for a problem was 7200 s, any problem that was
not solved in that much time was assigned a time of 7200 s. The column titled All gives the times
1234 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236

when all techniques are used versus when some techniques are not used. Similarly, the column titled
None shows the times when none of the techniques are applied (second row) versus when at least
one technique is applied (third row). Note that all methods do help improve the computational times.
The most pronounced bene$t is obtained by using the network technique. The bene$t of applying
the techniques is most dramatic when we consider the average solution times when all techniques
are used (3:57 s) as opposed to the average time when none of the techniques is applied (7200 s).
The last number indicates that no single problem could be solved in the allotted time!
Given the e0ectiveness of the elimination techniques in reducing computation times, it is important
to consider the e0ect on solution quality. In all instances, the linear-programming solution when the
elimination techniques was applied was found to be equal to the optimal solution! In the case of
the IP solution, we used a 1 percent threshold for fathoming. The solutions obtained when the
elimination techniques were used were, in all instances, optimal within the 1 percent threshold! This
is remarkable considering the huge reduction in problem size and computational times achieved by
the elimination techniques.
We also tested our models on real-world data obtained from a large food products company.
Currently, the company uses a commercial package which also solves only the linear relaxation of
the problem but with the traditional resource-based formulation. The company usually sets aside a
fraction of capacity to adjust for the inaccurate way that resource based formulation deals with set-up
costs and to round o0 fractional setups. The solution obtained is manually adjusted by a team of
two human schedulers working over several days to make it workable. As a result the cost obtained
is very di0erent from the optimizer generated cost. The company wanted to compare our solutions
in terms of the amount of demand that could be met and the overall solution cost.
For the larger problems given to us by the company, we found that solution of the MIP was
computationally infeasible and the smaller problems were trivial to solve. For the larger problems,
we found that even solving the LP relaxation without the elimination schemes was computationally
infeasible.
Since these models are used for aggregate planning, it is common to consider linear program-
ming relaxations. We solved the linear programming relaxations of these large problems using our
con$guration-based model and with all elimination techniques applied. We found that the e0ect of
aggregation is more pronounced in the cases where the production capacity is very tight. To further
examine the e0ects of aggregation, we performed an experiment on six di0erent data sets for a single
facility problem. We solved a 12 period problem and solve the linear programming relaxation of
these problems. The production facility in consideration has 27 di0erent machines and can produce
29 di0erent SKUs. The set-up costs are zero although the set-up times are provided. In these ex-
periments we focus our study on the e0ects of aggregation. As mentioned in Section 4.2, a poor
scheme of aggregation may result in omission of some vital combinations. For these experiments,
we included three variations of aggregation schemes, i.e.,
• Type 1: Single period aggregation.
• Type 2: Three period aggregation with rolling horizon.
• Type 3: k period aggregation with rolling horizon; k = 1; 2; : : : ; 12.
All these problems are solved using CPLEX v7.0 on a Pentium III Xeon Processor 550 MHz machine
with 1:5 GB of RAM. Table 6 presents the amount of demand that could not be satis$ed for these
experiments for di0erent aggregation schemes.
 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236 1235

Table 6
E0ect of aggregation on demand satisfaction

Data set no. Total demand (Klbs) Unsatis$ed demand for di0. aggregation types (klbs)

Type 1 Type 2 Type 3

1 164,186 189.995 26.415 1.441

2 163,515 182.634 55.171 0.992
3 166,683 196.227 54.53 1.499
4 172,074 166.106 45.167 1.734
5 168,344 96.853 73.946 1.075
6 170,736 85.293 52.531 1.523

As can be seen from the table, it is possible that none of the aggregation schemes could satisfy
the entire demand, but the amounts of unsatis$ed demand are small enough to be ignored. The
type III aggregation scheme performs much better than the other two types of aggregation schemes.
Our experience is that smaller problems with suNcient capacity can usually be solved using type
II aggregation scheme. Larger problems with tight capacity, on the other hand, require type III
aggregation scheme.
The solutions obtained from our model required some rounding up because of fractional setups,
but were well within the times the company reserved for setup adjustments. The rounding up could
be done by the computer and the resulting solutions were workable schedules. This eliminated the
need for human scheduler intervention. The unmet demand was lower than the company’s model
and the overall costs were also lower by about 3 percent when compared to the costs obtained by
the optimizer.

6. Conclusions

In this paper, we study the problem for aggregate production planning. We present the limitations
of traditional models that are based on resource-based formulations and propose a con$guration-based
model for these problems. In our model, production process is assumed to be continuous and no
in-process inventory is allowed. We also assume that dissimilar machines may be combined in dif-
ferent ways to co-manufacture groups of products. One of the key consideration to the successful
implementation of the con$guration-based formulation is the generation of input data. A bad im-
plementation of the model, resulting in generation of large number of variables, may render this
formulation computationally intractable. In this paper, we have presented some heuristics to re-
strict the number of variables. Our computational experiments demonstrate the e0ectiveness of these
heuristics in solving the real size problems in reasonable time.

Acknowledgements

The authors would like to thank two anonymous referees for their comments and suggestions,
which helped improve the readability of the paper.
1236 A. Jain, U.S. Palekar / Computers & Operations Research 32 (2005) 1213 – 1236

References

[1] Silver EA, Pyke DF, Peterson R. Inventory management and production planning and scheduling, 3rd ed. New York:
Wiley; 1998.
[2] IIE Solutions. Application—Ford saves $1 million with parts inventory system, July 1996.
[3] Dzielinski BP, Gomory RE. Optimal programming of lot sizes, inventory and labor allocations. Management Science
1965;11:874–90.
[4] Lasdon LS, Terjung RC. An eNcient algorithm for multi-item scheduling. Operations Research 1971;19:946–69.
[5] Newson EFP. Multi-item lot size scheduling by heuristic Part II: with variable resources. Management Science
1975;21:1194–203.
[6] McClain JO, Thomas LJ, Mazzola JB. Operations management: production of goods and services, 3rd ed. Englewood
Cli0s, NJ: Prentice-Hall; 1992.
[7] Nahmias S. Production and operations analysis, 3rd ed. Homewood, IL: Richard D. Irwin; 1996.
[8] Nam, Sang-jin, Logendran R. Aggregate Production Planning—A Survey of Models and Methodologies. European
Journal of Operational Research 1992;61:255 –72.
[9] Thomas LJ, McClain JO. An overview of production planning. Graves SC, Rinnoy Khan AHG, Zipkin PH, editors.
Handbook in operations research and management science, Vol. 4, Logistics of production and inventory. Amsterdam,
The Netherlands: North-Holland Publishing Company; 1993 [chapter 7].
[10] Shapiro JF. Mathematical programming models and methods for production planning and scheduling. Graves SC,
Rinnoy Khan AHG, Zipkin PH, editors. Handbook in operations research and management science, Vol. 4, Logistics
of production and inventory. Amsterdam, The Netherlands: North-Holland Publishing Company; 1993 [chapter 8].
[11] Leachman RC, Carmon TF. On capacity modeling for production planning with alternative machine types. IIE
Transactions 1992;24(4):62–72.
[12] Bradley JR, Arntzen BC. The simultaneous planning of production, capacity, and inventory in seasonal demand
environments. Operations Research 1999;47:795–806.
[13] Bermon S, Hood SJ. Capacity optimization planning system (CAPS). Interfaces 1999;29(5):31–50.
[14] Bhatia M, Palekar US. A variable rede$nition approach for the lot sizing problem with strong set-up interactions.
IIE Transactions 2001;33(5):357–70.