Author's Accepted Manuscript

A new approach to solve the multi-product

multi-period inventory lot sizing with suppli-
er selection problem
Leopoldo Eduardo Cárdenas-Barrón, José Luis
González-Velarde, Gerardo Treviño-Garza

www.elsevier.com/locate/caor

PII: S0305-0548(15)00153-7
DOI: http://dx.doi.org/10.1016/j.cor.2015.06.008
Reference: CAOR3808

To appear in: Computers & Operations Research

Cite this article as: Leopoldo Eduardo Cárdenas-Barrón, José Luis González-
Velarde, Gerardo Treviño-Garza, A new approach to solve the multi-product
multi-period inventory lot sizing with supplier selection problem, Computers &
Operations Research, http://dx.doi.org/10.1016/j.cor.2015.06.008

This is a PDF file of an unedited manuscript that has been accepted for
publication. As a service to our customers we are providing this early version of
the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting galley proof before it is published in its final citable form.
Please note that during the production process errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal
pertain.
A new approach to solve the multi-product multi-period inventory lot sizing with supplier selection
problem

Leopoldo Eduardo Cárdenas-Barrón

School of Engineering and Sciences
Tecnológico de Monterrey.
E. Garza Sada 2501 Sur, C.P. 64849, Monterrey, Nuevo León, México

José Luis González-Velarde

School of Engineering and Sciences
Tecnológico de Monterrey.
E. Garza Sada 2501 Sur, C.P. 64849, Monterrey, Nuevo León, México

Gerardo Treviño-Garza
BNSF Railway Company
2650 Lou Menk Drive
Fort Worth, TX 76131-2830, USA
*
• Corresponding author. Tel. +52 81 83284235, Fax +52 81 83284153.
E-mail address: lecarden@itesm.mx (L.E. Cárdenas-Barrón).

Abstract

This research work deals with the multi-product multi-period inventory lot sizing with supplier selection
problem. Formerly, this kind of problem was formulated and solved using an exhaustive enumeration
algorithm and a heuristic algorithm. In this paper, a new algorithm based on a reduce and optimize
approach and a new valid inequality is proposed to solve the multi-product multi-period inventory lot
sizing with supplier selection problem. Numerical experiments ratify the success of the proposed heuristic
algorithm. For the set of 150 benchmark instances, including 75 small-sized instances, 30 medium-sized
instances, and 45 large-sized instances, the algorithm always obtained better solutions compared with
those previously published. Furthermore, according to the computational results, the developed heuristic
algorithm outperforms the CPLEX MIP solver in both solution quality and computational time.

Keywords. Inventory, lot sizing, multi-period, multi-products, supplier selection, mixed integer linear
programming, reduced costs, reduce and optimize approach (ROA).

1. Introduction
Supplier selection and lot sizing are essential activities in supply chain management. Typically, in any
echelon of a supply chain it is important to select the best suppliers from which to acquire the products,
the lot sizes, and the time to place the orders during a finite planning horizon.

The single-product multi-period inventory lot sizing problem has its origins at the end of the 1950s and
was first proposed by Wagner and Whitin (1958). This problem can be represented as a mixed integer
linear programming (MILP) model. Moreover, Zangwill (1969) has shown that this problem is a fixed
charge network problem. Wagner and Whitin (1958) have solved this type of problem optimally with a
dynamic programming algorithm. Furthermore, several dynamic lot size rules have been proposed to
solve the single-product multi-period inventory lot sizing problem. For instance, see Simpson (2001) who
conducted a scrutinized and comprehensive study of nine well-known published rules. Since its
introduction, it became one of the most studied and extended problems in inventory management. A
number of excellent reviews have been published in order to provide an introduction to the lot sizing
problem and its extensions. For example, De Bodt et al. (1984), Bahl et al. (1987), Kuik et al. (1994) and
Wolsey (1995) are the first reviews on the history of the single product lot sizing problem. In addition to
well-known heuristic rules, the lot sizing problem has been also solved using other approaches, i.e. Hop
and Tabucanon (2005) have developed a new and original approach to solve the lot sizing problem using
an adaptive genetic algorithm. Later, Cárdenas-Barrón (2010) discusses some features of the adaptive
genetic algorithm in Hop and Tabucanon (2005). He concludes that it is convenient to solve the lot sizing
problem with Wagner and Whitin (1958) algorithm because this always obtains the optimal solution. In
addition, Jans and Degraeve (2007) have provided a comprehensive and complete review of
metaheuristics for the lot sizing problem. One can see from the reviews mentioned above that there exist
many approaches to solve the lot sizing problem that have been developed for distinct applications.

Another interesting research direction is on the inventory lot sizing with supplier selection problem which
combines lot sizing and supplier choice decisions. The research works of Kasilingam and Lee (1996) and
Jayaraman et al. (1999) address this problem. Both papers contain mixed integer programming models to
select suppliers and determine the lot size of the products. Also, Dahel (2003) develops a multi objective
mixed integer programming approach to select the number of suppliers to use and the lot size of each
product to place the orders to suppliers for a multiproduct, multi-supplier competitive sourcing
environment.
Basnet and Leung (2005) deal with the multi-product multi-period inventory lot sizing with supplier
selection problem. First they solved the problem with an exhaustive enumerative search algorithm. Since
this algorithm cannot obtain a solution within 2 hours of computation for instances with 60 or more
binary variables, they proposed a heuristic algorithm to solve the problem in a reasonable computational
time. On the other hand, Moghadam et al. (2008) propose a new mathematical formulation for the multi-
period inventory lot sizing problem with supplier selection and solve it using a hybrid intelligent
algorithm based on fuzzy neural networks and genetic algorithms. Aissaoui et al. (2007) present an
excellent and comprehensive review on supplier selection and lot sizing. They mainly focus on papers
that deal with the determination of the best combination of suppliers and allocation of the orders in order
to cover different purchasing requirements. In the same year, Wadhwa and Ravindran (2007) model the
supplier selection problem as a multi-objective optimization problem in which one or more purchasers
order multiple products from different suppliers in a multiple sourcing network. Later, Ustun and
Demirtas (2008) propose a model that integrates the well-known analytic network process (ANP) and
achievement scalarizing functions to select the best suppliers and establish the optimal lot sizes between
the chosen suppliers by considering tangible–intangible criteria and time horizon. Afterwards, Ho et al.
(2010) review the literature related to the multi-criteria decision making approaches for supplier selection
and evaluation. Tsai et al. (2010) develop an attribute-based ant colony system (AACS) which is a
framework to study the critical factors to select the best suppliers for purchasing decisions. Ware et al.
(2012) present an exhaustive and well documented state of the art literature review. Basically, Ware et al.
(2012) critique the papers that deal with the supplier selection problem. Afterwards, Ruiz-Torres et al.
(2013) propose a mathematical optimization model to determine the optimal demand allocation over a set
of suppliers considering the risk of supplier failures. Ware et al. (2013) study the interrelationship of
supplier selection criteria using the interpretive structural modeling (ISM) approach. Choudhary and
Shankar (2013) propose an integer linear programming optimization model to jointly determine the timing
of procurement, lot-sizes, and the selection of suppliers and carriers in order to minimize the total cost
thru a finite planning horizon. In a subsequent paper, Choudhary and Shankar (2014) develop a multi-
objective integer linear programming model to make the best decisions on inventory lot-sizing, supplier
selection, and carrier selection. Ware et al. (2014a) propose a methodology for the flexible supplier
selection problem in which both qualitative and quantitative factors are considered jointly. They use the
analytical hierarchy process (AHP) and interpretive ranking process (IRP) for the qualitative model, and a
mixed non-integer linear programming problem for the quantitative model. In the same year, Ware et al.
(2014b) analyze the impact of demand variation on a multi-product, multi-source, multi-period model for
the supplier selection problem. Ware et al. (2014c) develop a mixed-integer non-linear program to
address the dynamic supplier selection problem. More recently, Karsack and Dursun (2015) present a
fuzzy multi-criteria group decision making methodology which uses the well-known quality function
deployment (QFD), and a fusion of fuzzy information and a 2-tuple linguistic representation model for
supplier selection.

This research paper mainly improves the work of Basnet and Leung (2005). It is of fundamental
importance to mention that the inventory lot sizing with supplier selection problem in Basnet and Leung
(2005) is a complex combinatorial NP hard optimization problem. Heuristics or approximation algorithms
play a vital role in solving NP hard problems with the hope that a near optimal solution in a short amount
of time can be found. Consequently, this research presents a heuristic algorithm that provides a quality
solution for the multi-product multi-period inventory lot sizing with supplier selection problem in a
reasonable amount of computational time.

Basnet and Lueng (2005) have left as a fertile area for future research the improvement of their solutions.
Therefore the main goal of this research paper is to propose a new approach to solve the multi-product
multi-period inventory lot sizing with supplier selection problem heuristically. This new approach is
based on the reduce and optimize approach (ROA). It is possible to accelerate the solution process of an
optimization problem by solving it on a small set of variables. Within this context, the ROA approach
always solves the problem over a small feasible space that contains a near optimal solution. Treviño-
Garza (2009) has proved that the ROA is capable of solving other types of binary integer problems (i.e.
single machine total weighted tardiness problem, set covering problem and set partitioning problem) in a
reasonable time. The noteworthy results achieved by Treviño-Garza (2009) have encouraged us to apply
ROA to multi-product multi-period inventory lot sizing with supplier selection problem in the present
research article.

The rest of this paper is organized as follows. Section 2 presents the formal definition of the multi-product
multi-period inventory lot sizing with supplier selection problem and its mathematical formulation.
Section 3 proposes a heuristic algorithm to solve the multi-product multi-period inventory lot sizing with
the supplier selection problem and also provides the results of the extensive computational experiments.
Finally, Section 4 gives some conclusions and future research directions.
2. The multi-product multi-period inventory lot sizing with supplier selection problem.

2.1 Problem definition.

Basnet and Leung (2005) formally define the multi-product multi-period inventory lot sizing with
supplier selection problem as follows. Consider the situation where the deterministic demand of multiple
discrete products is known over a given finite planning horizon. Each product can be supplied from a set
of suppliers, for example one or more suppliers could be chosen in each period for the procurement of a
product. There is a supplier ordering cost that is incurred in each period when an order is placed to the
supplier. Also, there is a product holding cost per period that applies to each product in inventory when it
is carried through a period in the finite planning horizon. There are no capacity constraints and shortages
are not allowed. The decision maker (manager) seeks to decide what products to buy in what lot sizes
from which suppliers and in which periods to place the orders. Basically, Basnet and Leung (2005)
propose a mixed integer linear programming (MILP) formulation which is presented in Subsection 2.2.

2.2. Mixed Integer Linear Programing Model (MILP) for the multi-product multi-period inventory
lot sizing with supplier selection problem

Basnet and Leung (2005) propose the following MILP formulation for the multi-product multi-period
inventory lot sizing with supplier selection problem (the notation and the mathematical model are placed
here just for self-completeness of this paper).

Indices
i = 1,2,3,.., I index for products.
j = 1,2,3,.., J index for suppliers.
t = 1,2,3,.., T index for periods.

Parameters
Dit = demand for product i in period t.
Pij = purchase price for product i from supplier j.
Hi = holding cost for product i per period.
Oj = ordering cost for supplier j.
Decision variables
Xijt = lot size for product i ordered from supplier j in period t.
Yjt = 1 if an order is placed to supplier j in period t, 0 otherwise.

Mixed integer linear programming (MILP) formulation

Then, the problem can be formulated as follows:

Model:

 t t 
Min ∑ ∑ ∑ Pij X ijt + ∑ ∑ O jY jt + ∑ ∑ H i  ∑ ∑ X ijk − ∑ Dik  (1)
t j i j t i t  k =1 j k =1 
Subject to :
t t
∑ ∑X −∑ D ≥0 ∀i,t (2)
k =1 j ijk k =1 ik
 T∑ D Y − X ≥ 0
  ∀ i, j , t (3)
 k =t ik  jt ijt

Y jt = {1,0} ∀ j ,t (4)

X ijt ≥ 0 ∀ i, j , t (5)

The objective function (1) is the total cost that is incurred by the buyer and comprises of the total
purchase cost of the products, the total ordering cost, and the total holding cost for carrying inventory in
each period. Constraints (2) ensure that all demand is satisfied in the period in which it occurs; these
constraints also guarantee avoiding shortages. Constraints (3) establish that is not possible to place an
order without charging a corresponding ordering cost. Constraints (4) define the binary variables.
Constraints (5) impose non-negativity conditions on the remaining decision variables. This mathematical
formulation (1) to (5) is very difficult to solve since it is an NP hard problem. Furthermore, the problem
becomes unsolvable when the instances are of large size. It is important to remark that the mathematical
formulation (1) to (5) can only be solved to optimality by commercial integer linear programming solvers
for very small instance sizes. For instances of large size that are commonly found in the real world, the
problem cannot be solved optimally.
3. Solving the multi-product multi-period inventory lot sizing with supplier selection problem based
on the reduce and optimize approach (ROA)

This section presents a heuristic algorithm for solving the mixed integer linear programming (MILP) for
the multi-product multi-period inventory lot sizing with supplier selection problem, along with, the
computational results for the 150 instances solved in Basnet and Leung (2005).
One way to strengthen the mathematical formulation is to add valid inequalities to the model. For that
reason, the mathematical formulation (1) to (5) just presented above is further reinforced by including the
following valid inequality (VI): ∑Y
j
jt ≤ J , ∀ t ∈ T . This inequality imposes a boundary on the total

of number of suppliers used in any period t. Even though the valid inequality (VI) is redundant to the
mathematical formulation, this is valuable in aiding CPLEX to create some new cuts.

It is important to mention that Basnet and Leung (2005) do not provide any valid inequality because their
solution approaches do not require this. It was detected that the valid inequality (VI) has a positive effect
on the optimization process, particularly on the following two issues: the solution quality of the instances
and the computational time.

3.1 Reduce and Optimize Approach (ROA)

Here, the Reduce and Optimize Approach (ROA) is explained in a brief manner as follows. ROA is based
on establishing a reduced feasible region based on the original problem and optimizing it. Although, in
most of the cases, ROA does not ensure optimality, it can be used because it obtains near optimal
solutions in a reasonable time. Nonetheless, if the optimal solution of the problem at hand is feasible for
the reduced problem then ROA guarantees the optimal solution. Fundamentally, ROA is laid out in the
following statement:

Assume that a mathematical formulation has n binary variables (i.e. y1,y2,…,yn) and BC is the set of all
binary variables of the whole mathematical formulation. Let BO be the set that has the binary variables
that are equal to 1 at the unknown optimal solution. Let BR be a reduced set that contains a portion of the
binary variables. Then it is required that the before mentioned sets satisfy the following condition: BO ⊆
BR ⊆ BC.
Clearly, the challenge of ROA is to obtain a set BR such that BR is a subset of BC and further BR also
covers the set BO. Then ROA repeatedly optimizes the problem over a reduced feasible space BR using a
commercial MIP solver (i.e. CPLEX). In other words, the problem is solved after fixing the binary
variables not in set BR equal to 0 (i.e. yi=0 if and only if yi ∈ BC-BR).

It is crucial to mention that there are some situations in which ROA does not find good solutions. These
situations are: 1) when the set BR contains only a portion of set BO and 2) when the sets BR and BO are
disjoint. Furthermore, it is worth mentioning that the set BO with the optimal solution is always
unidentified.

Conversely, in ROA it is required that the set BR be of small size such that the optimization problem can
be handled by any commercial optimization software (i.e. CPLEX) within a reasonable amount of
computational time. Likewise, finding a practical size for the set BR that includes the binary variables
that are equal to 1 in the optimal solution is the most challenging task in ROA. Therefore, the first set BR
where ROA initiates is sometimes not as good as desired. Therefore the set BR needs to be improved.
Basically, ROA consists of four phases. The first phase of ROA constructs a reduced set of variables
(initial set BR). The second phase optimizes the mathematical model within the reduced set of variables.
The third phase selects new variables to be included in the reduced set of variables. The fourth phase
refines the reduced mathematical model and then goes to the second phase.

Here, there are some ways to construct the initial set BR:

i. create the initial set BR randomly,

ii. solve the full complete model of the problem for a short time and then use its incumbent solution
(i.e. the variables equal to 1 in the incumbent solution form the set BR),
iii. solve an LP-relaxation of the problem and use its optimal solution
(i.e. the variables equal to 1 in the optimal LP-relaxation solution form the set BR),
iv. solve the problem with a heuristic to construct the initial set BR,
v. or with a combination of the above ways.

In this paper option iv is preferred using the Wagner-Whitin (1958) algorithm. The third phase of ROA is
to enhance the set BR iteratively. This task can be performed by taking information from the solution of
the reduced problem previously solved. Here, it is proposed to obtain the reduced costs of the variables
that are not in set BR in order to enhance BR. The problem is then re-optimized over BR until a stop
condition is met or when no more negative reduced costs (in a minimization problem) are found.

It is important to mention that the reduced costs of binary variables cannot be determined. Nevertheless,
an LP-relaxation can be solved easily and then the reduced costs of the variables that are not in the set BR
can be obtained. Afterwards, the variables with negative reduced costs need to be included in the set BR.
In order to prevent that set BR reach an unmanageable size giving its continuous growth in each
improvement iteration, the variables within set BR with a reduced cost greater than or equal to zero are
dropped from the set. Additionally, a maximum of number of variables to be added to set BR is also
established.

It is worth mentioning that several popular algorithms exist that are used to solve problems similar to the
multi-product multi-period inventory lot sizing with supplier selection problem. Some of them are based
on genetic algorithms (GA), particle swarm optimization (PSO), ant colony optimization (ACO), Scatter
Search (SS), Differential Evolution (DE), among others. These algorithms have a population of candidate
solutions and they improve the set of solutions iteratively. In contrast, the reduce and optimize approach
(ROA) always considers a small set of binary variables and optimizes the problem over this set. This is
the main difference between other popular algorithms (GA, PSO, ACO, SS, DE) and ROA.

To the best of our knowledge this paper is the first to apply the ROA to the inventory lot sizing with
supplier selection problem.

3.2 Proposed heuristic algorithm based on ROA

In this subsection, a heuristic algorithm is proposed. The following abbreviations: RedMILP,

RelaxMILP, max-nva, nvnrc , LBvar and UBvar are used in the heuristic algorithm. RedMILP is the
MILP model but when the optimization problem is solved over a reduced set of binary variables (Yjt).
RelaxMILP is a relaxed MILP model in which all binary variables are relaxed. The abbreviation max-
nva refers to the maximum number of variables to add to the set BR. The abbreviation nvnrc represents
the number of variables with negative reduced cost. LBvar and UBvar denote lower bound and upper bound
for the decision variable Yjt. Also, YWWjt represents the solution found by the Wagner-Whitin (1958)
algorithm. The superscript WW corresponds to the initials of Wagner-Whitin. The superscript S in Ysjt
 refers to “solution” and the superscript C in Ycjt corresponds to “continuous”; this is when the binary
variable is relaxed to take continuous values. Finally, RCYjt represents the reduced cost of variable Ycjt.

It is important to point out that the formulation models always contain constraints (2)-(5) and the
proposed valid inequality (VI). The solution procedure is described by the pseudo-code shown in the
following algorithm:

Algorithm: Proposed heuristic algorithm based on ROA. ROA’s phases

Let nvnrc the number of variables with negative reduced cost.
Parameters: β , τ , max-nva
Step 1. Obtain the initial set BR.
Set clock=0
Apply the Wagner-Whitin (1958) algorithm for each product for each supplier and select the best supplier for each product Build a reduced set
Let Y WWjt be the solution after applying the Wagner-Whitin (1958) algorithm
of variables
Set Ysjt = YWWjt. Set LBvar Ysjt = YWWjt ; UBvar Ysjt = YWWjt

Step 2. Solve RedMILP model during β time units

If clock ≥τ time units then Return (solution) Optimize the mathematical
Else model within the reduced set of
variables
Step 3. Obtain the value of binary variables Ysjt,
Step 4. Construct the continuous variables set as follows:
Set LBvar Ycjt= Ysjt ; UBvar Ycjt= Ysjt Select new variables
Step 5. Solve RelaxMILP.
Step 6. Obtain the reduced cost for each variable Yjt: RCYjt. to be included in
If there are no more variables with negative reduced costs to include in BR then Return (solution). the reduced set of variables
Otherwise

Step 7. Refine set BR.

If ((Yjt = 0) and (UBvar Yjt =1))
If RCYjt ≥ 0 set LBvar Yjt =0; UBvar Yjt =0
If (nvnrc < max-nva)
If (UBvar Yjt =0)
If RCYjt < 0 set LBvar Yjt =0; UBvar Yjt =1 Refine the reduced
Else mathematical model
If (UBvar Yjt =0)
If ((RCYjt is one of the max-nva most negative reduced costs))
Set LBvar Yjt =0; UBvar Yjt =1
Update clock
Go to Step 2

The above heuristic algorithm works as follows.

Step 1. The initial set is constructed applying the Wagner-Whitin (1958) algorithm. In this step the
objective is to obtain from which supplier j and in which period t the product must be purchased, and
there is no need to determine the lot sizes. In other words, the Wagner-Whitin (1958) algorithm is used
simply to determine which binary variables (i.e. Yjt) must be included in the initial set BR.
Step 2. Starting with an initial feasible set BR for the optimization problem the RedMILP model is
solved for β time units. If clock < τ then go to Step 3. Otherwise the algorithm stops and returns the
incumbent solution.

Step 3. Once the RedMILP is solved the value of each binary variable (Ysjt) of the incumbent solution is
determined and will be used in Step 4.

Step 4. With the values of all binary variables the continuous variables set is constructed setting the lower
and upper bounds to the corresponding values obtained in Step 3.

Step 5. Here, a surrogate problem called RelaxMILP is solved. This action refers to a tactic used to
determine the reduced costs for each variable defined in Step 4.

Step 6. In this step, if there are no more variables with negative reduced costs to add to BR then the
algorithm stops and returns the incumbent solution. Otherwise, the set BR can be enhanced in Step 7.

Step 7. Now, the set BR is improved as follows: If nvnrc<max-nva then add the nvnrc variables to set
BR. Else add the max-nva variables with most negative reduced cost to set BR. Return to Step 2.

As is typical for the majority of heuristic algorithms, one has to define some parameters. In this algorithm
it is required to determine the initial set BR and the value for the following parameters: β , τ and max-
nva. The values used for β , τ and max-nva are 8 minutes, 30 minutes, and 400, respectively. It is clear
that for very large instances the number of variables with negative reduced costs is also a large number. It
is therefore crucial that the maximum number of variables (max-nva) be established considering the size
of the largest instances which have 10,000 binary variables. The value of max-nva was set to 4% of the
total number of binary variables in the largest instances.

3.3 Computational results

This section presents the comparison of results of the proposed heuristic algorithm with the existing
results in the literature reported by Basnet and Leung (2005). The computational results are presented and
discussed in an effort to evaluate the efficiency of the proposed heuristic algorithm.

Hardware and Software. Here, the technical details related to the computational experiments that were
utilized to evaluate the performance of the proposed heuristic algorithm are specified below. The heuristic
algorithm just explained in section 3.2 was coded in C++ using Microsoft Visual Studio 2008. The
experiments were executed on an Hp Elite 8300 with the following technical details: Intel(R) Core (TM)
i5-3470M CPU @ 3.20GHz 3.30 GHz, 8.00GB RAM and 64-bit Operating system with Windows 7. The
IBM(R) ILOG(R) CPLEX(R) Interactive Optimizer 12.5.1.0 was used to solve the instances.
In order to identify if the algorithm is in fact capable of finding good solutions it is required to disable
some special features of CPLEX. In this case the following parameters were disabled: cliques, covers,
disjoint cuts, lift and project cuts, flow covers, flow paths, fractional cuts, generalized upper bound cover
cuts, implied bound cuts, zero half cuts, and mixed integer rounding cuts.

Benchmark of instances. To evaluate the performance of this new approach, the same instances for the
multi-product multi-period inventory lot sizing with supplier selection problem proposed by Basnet and
Leung (2005) were used. The benchmark set contains 150 instances. These can be classified as small,
medium, and large. The 75 small instances contain between 30 and 100 binary variables. The 30 medium
instances have between 500 and 1500 binary variables. The 45 large instances have between 2000 to
10000 binary variables.

The medium and large instances are considerably hard since not all of them could be solved to optimality
in a reasonable time with the exhaustive enumerative algorithm proposed by Basnet and Leung (2005) or
CPLEX. Those instances were solved with CPLEX during 1 hour. Conversely, all instances were solved
by the proposed algorithm with a time limit of 30 minutes.

The instance size can be coded as follows (l, m, n), where l indicates number of suppliers, m denotes the
number of products, and n means the number of periods. There are 15 instances of each instance size (l,
m, n). In the optimization problem, the number of the binary variables is given by the product l*m and
the number of constraints including the valid inequality is given by m*n + l*m*n + n. For example, for
any instance of size (50, 50, 200), the optimization problem has 10,000 binary variables and 510,200
constraints (See Table 1).
 Table 1. Characteristics of the instances tested.
Constraints
Binary Continuous including
Category Size of the variables variables proposed
instance Yjt Xijt valid
l,m,n inequality
3,3,10 15 45 65
3,3,15 45 135 195
Small 4,4,10 40 160 210
4,4,15 60 240 315
5,5,20 100 500 620
10,10,50 500 5000 5550
Medium 15,15,100 1500 22500 24100
20,20,100 2000 40000 42100
Large 20,20,200 4000 80000 84200
50,50,200 10000 500000 510200

Computational results. In Basnet and Leung (2005) there are two algorithms: enumerative and
heuristic. The results obtained by the proposed heuristic algorithm are compared with the best results
obtained with the heuristic algorithm of Basnet and Leung (2005). The computational results are shown in
Table 2. In detail, Table 2 shows Basnet and Leung (2015)’s solutions, the solutions of the proposed
algorithm, the difference, the percent of improvement, the computational time, the number of optimal
solutions achieved, and the gap. Each value in each line of Tables 2 and 3 is an average over 15 instances.
It is important to note that CPLEX is able to prove optimality before the time limit for only the small
instances.
Table 2. Results of the computational experimentation

Total average Total average

Category Size of the cost cost Difference Percent of Average No. of Average
instance Basnet and Proposed improvement Time Optimal GAP
Leung (2005) heuristic (minutes) Solutions %
Algorithm algorithm

3,3,10 102584 101953.6 630.4 0.614520783 0.009757778 14/15 0.199984542

3,3,15 147887 147162.9 724.0667 0.489608057 0.089163333 15/15 0
Small 4,4,10 126345 124526 1819 1.439708734 0.016915556 15/15 0
4,4,15 187320 185699.3 1620.667 0.865186134 3.440541111 14/15 0.179920572
5,5,20 303529 300900.2 2628.8 0.866078694 4.40799 12/15 0.071243571
10,10,50 1357190 1337662 19527.87 1.438845458 30 * 0.854599987
Medium 15,15,100 3856800 3810899 45900.87 1.190128258 30 * 2.068323929
20,20,100 5048826 4975149 73677.47 1.459298987 30 * 3.861037554
Large 20,20,200 10026074 9915138 110935.5 1.106470323 30 * 9.180299397
50,50,200 25373121 23534629 1838492 7.245823904 30 * 6.993766063
*In these instances CPLEX cannot prove optimality.
Table 3. Results of computational results comparing CPLEX and the proposed heuristic algorithm
solutions

CPLEX Total average

Category Size of the 1 hour cost Difference Percent of
instance Proposed improvement
heuristic
algorithm
30 minutes
3,3,10 101939.8 101953.6 -13.8 -0.01354
3,3,15 147162.9 147162.9 0 0
Small 4,4,10 124526 124526 0 0
4,4,15 185681.3 185699.3 -18.0667 -0.00973
5,5,20 300866.4 300900.2 -33.8 -0.01123
10,10,50 1336996 1337662 -666.467 -0.04985
Medium 15,15,100 3816286 3810899 5387.067 0.14116
20,20,100 5084377 4975149 109228.1 2.148309
Large 20,20,200 10195845 9915138 280706.3 2.753144
50,50,200 27632948 23534629 4098318 14.83127

In line with the results of Table 2 one can conclude that the proposed heuristic algorithm was able to find
on average better solutions than those reported by Basnet and Leung (2005) for all instance sizes. It is
observed that for the very large instances (50, 50, 200) with 10,000 binary variables, a very significant
percent of improvement of 7.245% was obtained. This indicates that the performance of the Basnet and
Leung (2005) algorithm deteriorates for large instances.

With regard to small instances, the proposed heuristic algorithm based on ROA is shown to be an
effective and efficient method because it obtains 93.34% (70/75) of the optimal solutions (see the
penultimate column of Table 2).

It can be observed from Table 3 that when the results of the proposed heuristic algorithm are compared
with CPLEX with a run time of 1 hour, a more impressive improvement is observed for the large
instances of (50,50,200) which contain 10,000 binary variables. In this case the average percent
improvement is 14.831%.
 Table 4. Results for each of the large instances comparing CPLEX and proposed heuristic algorithm

Total cost of Total cost of

Category Instance Size of the CPLEX % GAP proposed heuristic % GAP % GAP
number instance 1 hour algorithm improvement
30 minutes
106 20,20,100 5101378 6.8205 4949737 3.645207 3.175293
107 20,20,100 5112809 6.5253 4973406 3.620841 2.904459
108 20,20,100 4936862 4.3332 4881671 3.166821 1.166379
109 20,20,100 5170381 6.5458 5035871 3.773959 2.771841
110 20,20,100 5045316 6.8011 4914496 4.031854 2.769246
111 20,20,100 4954312 5.4348 4893951 4.150232 1.284568
112 20,20,100 5143246 4.9235 5100410 4.049635 0.873865
113 20,20,100 5141382 5.8799 5027762 3.540048 2.339852
114 20,20,100 5270327 7.4353 5137684 4.73138 2.70392
115 20,20,100 5244504 7.3144 5114953 4.663494 2.650906
116 20,20,100 4953208 6.4777 4837091 3.981566 2.496134
117 20,20,100 5130331 5.7772 5058205 4.290106 1.487094
118 20,20,100 5150935 6.5836 5030712 4.095935 2.487665
119 20,20,100 4905368 4.8005 4786268 2.255994 2.544506
120 20,20,100 5005291 6.4772 4885011 3.918492 2.558708
121 20,20,200 10617676 13.2956 10352356 10.46451 2.831089
122 20,20,200 9793130 9.2344 9706282 8.265681 0.968719
123 20,20,200 10066510 11.7528 9847077 9.316777 2.436023
124 20,20,200 10374157 13.0547 10161970 10.74234 2.312355
125 20,20,200 10323373 12.257 10166683 10.55314 1.703857
126 20,20,200 10915254 18.1395 9950331 7.695811 10.44369
127 20,20,200 9972145 12.6676 9692877 9.512366 3.155234
128 20,20,200 10063046 10.9896 9897932 9.168488 1.821112
Large 129 20,20,200 9866199 10.8207 9699241 8.945368 1.875332
130 20,20,200 10145424 11.4844 9901540 8.804447 2.679953
131 20,20,200 10101287 12.2981 9642607 7.198859 5.099241
132 20,20,200 9930787 11.9457 9766838 10.09757 1.84813
133 20,20,200 10391155 13.4413 10130116 10.59151 2.849789
134 20,20,200 10156716 11.1047 9962665 8.981969 2.122731
135 20,20,200 10220813 11.4238 9848562 7.365647 4.058153
136 50,50,200 27770956 26.0174 23678697 7.447789 18.56961
137 50,50,200 26396570 22.3583 23352424 8.247507 14.11079
138 50,50,200 27479244 25.9933 23102756 5.926949 20.06635
139 50,50,200 27444473 24.7807 23746045 7.965204 16.8155
140 50,50,200 27931490 26.8685 23436186 6.450238 20.41826
141 50,50,200 27960998 26.5155 23764888 7.529305 18.98619
142 50,50,200 28174415 26.3748 23592379 5.822328 20.55247
143 50,50,200 27047419 25.3755 22918244 6.235138 19.14036
144 50,50,200 27697870 25.9144 23819901 8.285169 17.62923
145 50,50,200 27677086 25.7188 23288453 5.784126 19.93467
146 50,50,200 27627311 26.2596 22996596 5.096765 21.16283
147 50,50,200 27699334 25.2242 23984561 8.430313 16.79389
148 50,50,200 28261279 26.4413 23834520 6.635927 19.80537
149 50,50,200 27356126 24.8628 23493593 7.232866 17.62993
150 50,50,200 27969642 25.5966 24010197 7.816865 17.77974
In order to demonstrate the power of the proposed heuristic algorithm, Table 4 presents the results for
each one of the large instances. From results of Table 4, it was calculated that the average % gap of
improvement for the large instances 106-121, 121-136 and 136-150 are 2.280%, 3.080%, and 17.921%,
respectively. Furthermore, the maximum percent improvements are 3.175%, 10.44%, and 21.162% for
the large instances 106-121, 121-136, and 136-150 respectively. These results demonstrate that the
heuristic algorithm performs very well. Moreover, the heuristic algorithm is noteworthy from both
business and computational perspectives.

Integer Linear Programming Model (ILP)

Since constraints (3) do not satisfy the integrality property, the optimization problem can be formulated as
an integer linear programming model (ILP). Thus, if we consider the optimization problem (1) to (4) and
the lot sizes as discrete variables then the constraints (5) are restated as follows:

X ijt ≥ 0 and Integer ∀ i, j , t (5b)

Then the problem becomes in ILP model. Obviously, constraints (5b) significantly increase the
complexity of the optimization problem. The number of integer variables in the ILP model for the
instances are given in column fourth of Table 1. For example, the large instances (50,50,200) have
500,000 integer variables and 10,000 binary variables.
Table 5. Results of computational results comparing CPLEX and the proposed heuristic algorithm
solutions for the case when Xijt is discrete.

CPLEX Total average

Category Size of the 1 hour cost Difference Percent of No. of Optimal
instance Proposed improvement Solutions
heuristic
algorithm
30 minutes
3,3,10 101939.8 101953.6 -13.8 -0.01354 14/15
3,3,15 147162.933 147162.9 0 0 15/15
Small 4,4,10 124526 124526 0 0 15/15
4,4,15 185681.267 185699.3 -18.0667 -0.00973 14/15
5,5,20 300866.4 300910.4 -44 -0.01462 9/15
10,10,50 1337295.93 1340292 -2996.33 -0.22406 *
Medium 15,15,100 3838747.6 3827811 10936.13 0.284888 *
20,20,100 5134674.67 5000178 134496.9 2.619384 *
Large 20,20,200 11280595.6 9930991 1349605 11.96395 *
50,50,200 100020161 23457449 76562712 76.54728 *
*In these instances CPLEX cannot prove optimality.

Each line of Table 5 shows an average over 15 instances. From Table 5 one can conclude that the
proposed heuristic algorithm has an impressive performance for the large instances of (20,20,200) and
(50,50,200) which contain 4,000 and 10,000 binary variables. In these cases, the average percent
improvements are 11.963% and 76.547% for large instances of (20,20,200) and (50,50,200),
respectively. Furthermore, with regard to small instances, the proposed heuristic algorithm based on ROA
also has an effective and efficient performance because it obtains 89.333% (67/75) of the optimal
solutions (see the penultimate column of Table 5).
 Table 6. Results for each of the large instances comparing CPLEX and proposed heuristic algorithm for
the case when Xijt is discrete.

Total cost of Total cost of

Category Instance Size of the CPLEX % GAP proposed heuristic % GAP % GAP
number instance 1 hour algorithm improvement
30 minutes
106 20,20,100 5123888 6.7 4975005 3.599656 3.100344133
107 20,20,100 5100036 5.9 5002062 3.865613 2.034386934
108 20,20,100 5156031 8.07 4904076 2.789043 5.280956777
109 20,20,100 5210239 7.08 5060789 4.008527 3.071472537
110 20,20,100 5050624 6.79 4934160 4.327494 2.462505734
111 20,20,100 5025172 6.55 4917616 4.269463 2.280537223
112 20,20,100 5257103 6.45 5117854 3.630376 2.81962443
113 20,20,100 5256075 8 5054967 3.867703 4.132297199
114 20,20,100 5283281 7.03 5169690 4.728846 2.301154288
115 20,20,100 5213989 6.33 5136432 4.748363 1.581636595
116 20,20,100 4965828 6.45 4866706 4.325171 2.124829313
117 20,20,100 5213949 6.8 5081610 4.089232 2.710767827
118 20,20,100 5166390 6.74 5049583 4.326714 2.413286488
119 20,20,100 4952720 6.37 4825053 3.628085 2.741915309
120 20,20,100 5044795 7.01 4907064 4.088455 2.921544743
121 20,20,200 11828390 21.85 10364667 6.771477 15.07852274
122 20,20,200 10462578 14.59 9706702 6.311368 8.27863179
123 20,20,200 11806726 24.09 9871906 3.754827 20.33517283
124 20,20,200 11981344 24.24 10174116 5.500032 18.7399683
125 20,20,200 11716363 22.2 10205389 6.440756 15.75924396
126 20,20,200 11336472 20.42 9976380 5.972624 14.44737645
127 20,20,200 10881768 19.43 9711031 6.580882 12.84911789
128 20,20,200 11214475 20.1 9889069 5.905732 14.19426773
Large 129 20,20,200 10498725 15.61 9729708 7.141728 8.468271659
130 20,20,200 11151468 20.11 9911019 6.749398 13.36060233
131 20,20,200 11236852 21.28 9666902 4.335438 16.94456205
132 20,20,200 11026251 20.54 9774891 6.860016 13.67998374
133 20,20,200 11202863 19.06 10139403 7.757929 11.30207052
134 20,20,200 11833552 23.85 9986048 4.514016 19.33598385
135 20,20,200 11031107 18.78 9857631 6.144325 12.63567467
136 50,50,200 27404834 25.06 23561601 7.521681 17.53832
137 50,50,200 249310062 91.78 23109566 7.121767 84.65823
138 50,50,200 26674181 23.87 23036138 6.975596 16.8944
139 50,50,200 257458338 92 23692418 7.72138 84.27862
140 50,50,200 27630837 26.11 23413680 6.862459 19.24754
141 50,50,200 27184137 24.32 23628153 8.057577 16.26242
142 50,50,200 27456091 24.58 23580648 6.995462 17.58454
143 50,50,200 26390723 23.46 22878502 7.029271 16.43073
144 50,50,200 27113797 24.32 23727141 8.791777 15.52822
145 50,50,200 222679663 90.77 23242631 5.575987 85.19401
146 50,50,200 239997786 91.51 22912704 4.71337 86.79663
147 50,50,200 27240694 24.01 23905620 8.827475 15.18253
148 50,50,200 27741416 25.19 23801140 7.40853 17.78147
149 50,50,200 26678982 23.11 23389453 7.930488 15.17951
150 50,50,200 259340873 91.99 23982346 7.691801 84.2982
Table 6 shows the results for the large instances for the integer linear programming model. From results
of Table 6, it was computed that the average % gap of improvement for the large instances 106-121, 121-
136, and 136-150 are 2.798%, 14.360%, and 39.523%, respectively. Additionally, the minimum and the
maximum percent improvements are (1.581%, 5.280%), (8.278%, 20.335%), and (15.179%, 86.796%)
for the large instances 106-121, 121-136, and 136-150, respectively. These results demonstrate that the
heuristic algorithm also performs very well for the integer linear programming model.

4. Conclusions

In this paper, for the first time, the simple and powerful reduce and optimize approach (ROA) was used to
solve the multi-product multi-period inventory lot sizing with supplier selection problem. This problem is
a new variant of the lot sizing inventory problem and is very difficult to solve to optimality. A new valid
inequality also was added to reinforce the mathematical model. It was observed that the new valid
inequality has a beneficial impact on both solution and time issues. Solving the 150 benchmark instances
including 75 small-sized instances, 30 medium-sized instances, and 45 large-sized instances revealed that
the proposed heuristic algorithm yields high quality solutions within a reasonable CPU time. The
performance of the heuristic algorithm is very satisfactory because it found better solutions than those
reported by Basnet and Leung (2005). Additionally, the results from detailed computational experiment
confirm the superiority of the proposed heuristic algorithm over the CPLEX MIP solver. For example,
compared with CPLEX with time limit of 1 hour, the heuristic algorithm improved the solutions of all
large instances in 30 minutes. These results ratify the success of the proposed heuristic algorithm.
Moreover, one could conclude that ROA has become a prominent methodology for tackling complex
optimization problems.

There are several future research directions that may be pursued after this paper. Immediate extensions
are incorporating important issues like allowing shortages and capacity constraints such as space and
budget. However, this will significantly increase the complexity of the optimization problem. Another
future direction could be to apply the proposed algorithm to other NP hard problems such as distinct
variants of the lot sizing inventory problem or any optimization problem with binary variables such as the
vehicle routing problem (VRP) and the inventory routing problem (IRP), just to name two possibilities.
Also, the development of metaheuristics to solve the multi-product multi-period inventory lot sizing with
supplier selection problem and comparing their results with our solutions possibly could be another
interesting research opportunity.
Acknowledgements

This research was supported by the Tecnológico de Monterrey Research Group in Industrial Engineering
and Numerical Methods 0822B01006 and the Mexican National Council for Science and Technology
(Grant SEP-CONACYT CB-2011-01-166397). The authors thank Dr. Chuda Basnet for kindly sharing
with us the set of benchmark instances that make possible this research paper. A special gratitude
is given to Dr. Neale R. Smith for his valuable feedback that greatly enhanced this research
work. Finally, the authors thank the anonymous reviewers for their valuable suggestions that
improved this paper.

References

Aissaoui, N., Haouari, M., Hassini, E., (2007). Supplier selection and order lot sizing modeling: A
review, Computers and Operations Research, 34(12), 3516 – 3540.

Bahl, H.C., Ritzman, L.P., Gupta, J.N.D., (1987). Determining lot sizes and resource requirements: A
review, Operations Research, 35(3), 329–345.

Basnet C., Leung, J.M.Y. (2005). Inventory lot-sizing with supplier selection, Computers and Operations
Research, 32(1), 1-14.

Cárdenas-Barrón, L.E., (2010). Adaptive genetic algorithm for lot-sizing problem with self-adjustment
operation rate: A discussion, International Journal of Production Economics, 123(1), 243-245.

Choudhary, D., Shankar, R., (2013). Joint decision of procurement lot-size, supplier selection, and carrier
selection, Journal of Purchasing and Supply Management, 19(1), 16–26.

Choudhary, D., Shankar, R., (2014). A goal programming model for joint decision making of inventory
lot-size, supplier selection and carrier selection, Computers and Industrial Engineering, 71, 1–9.

Dahel, N.E. (2003). Vendor selection and order quantity allocation in volume discount environments,
Supply Chain Management, 8(4), 335-342.
De Bodt, M.A., Gelders, L.F., Van Wassenhove, L.N., (1984). Lot sizing under dynamic demand
conditions: A review, Engineering Costs and Production Economics, 8 (3), 165–187.

Ho, W. Xu, X., Dey, P.K., (2010). Multi-criteria decision making approaches for supplier evaluation and
selection: A literature review, European Journal of Operational Research, 202(1), 16–24.

Hop,N.V., Tabucanon, M.T., (2005). Adaptive genetic algorithm for lot-sizing problem with self-
adjustment operation rate, International Journal of Production Economics, 98(2), 129–135.

Jans, R., Degraeve, Z. (2007). Meta-heuristics for dynamic lot sizing: A review and comparison of
solution approaches, European Journal of Operational Research, 177(3), 1855–1875.

Jayaraman, V., Srivastava, R., Benton, W.C. (1999). Supplier selection and order quantity allocation: A
comprehensive model, Journal of Supply Chain Management, 35(2), 50-58.

Karsak, E. E., Dursun, M., (2015). An integrated fuzzy MCDM approach for supplier evaluation and
selection, Computers and Industrial Engineering, 82, 82-93.

Kasilingam, R.C., Lee, C.P. (1996). Selection of vendors - A mixed-integer programming approach,
Computers and Industrial Engineering, 31(1-2), 347 – 350.

Kuik, R., Salomon, M., Van Wassenhove, L.N., (1994). Batching decisions: Structure and models,
European Journal of Operational Research, 75(2), 243–263.

Moghadam, M.R.S. Afsar,A. Sohrabi, B. (2008). Inventory lot-sizing with supplier selection using hybrid
intelligent algorithm, Applied Soft Computing, 8(4), 1523-1529.

Ruiz-Torres, A.J., Mahmoodi, F., Zeng, A.Z., (2013). Supplier selection model with contingency
planning for supplier failures, Computers and Industrial Engineering, 66(2), 374–382.

Simpson, N.C., (2001). Questioning the relative virtues of the dynamic lot sizing rules, Computers and
Operations Research, 28(9), 899–914.
Treviño-Garza, G. (2009). A heuristic approach to general 0-1 integer programming, PhD dissertation,
Arizona State University, UMI-Number: 3354471.

Tsai, Y.L., Yang, Y.J., Lin, C.-H., (2010). A dynamic decision approach for supplier selection using ant
colony system, Expert Systems with Applications, 37(12), 8313–8321.

Ustun, O., Demirtas, E.A., (2008). Multi-period lot-sizing with supplier selection using achievement
scalarizing functions, Computers and Industrial Engineering, 54(4), 918–931.

Wadhwa, V., Ravindran, A.R., (2007). Vendor selection in outsourcing, Computers and Operations
Research, 34(12), 3725 – 3737.

Wagner, H.M., Whitin, T.M., (1958). Dynamic version of the economic lot size model, Management
Science, 5(1), 89–96.

Ware, N.R., Singh, S.P., Banwet D.K. (2012). Supplier selection problem: A state-of-the-art review,
Management Science Letters, 2(5), 1465–1490.

Ware, N.R., Singh, S.P., Banwet D.K. (2013). A Study of Inter-Relationship of Supplier Selection
Criteria Using ISM Approach, Proceedings of the 7th National Conference; INDIACom-2013 Computing
For Nation Development.

Ware, N.R., Singh, S.P., Banwet D.K. (2014a). Modeling flexible supplier selection framework, Global
Journal of Flexible Systems Management, 15(3), 261-274.

Ware, N.R., Singh, S.P., Banwet D.K. (2014b). Analyzing the Effect of demand variation on multi-
product, multi-source, multi-period model for supplier selection problem, Industrial Engineering Journal,
7(2), 13-18.

Ware, N.R., Singh, S.P., Banwet D.K. (2014c). A mixed-integer non-linear program to model dynamic
supplier selection problem, Expert Systems with Applications, 41(2), 671–678.

Wolsey, L.A., (1995). Progress with single-item lot-sizing, European Journal of Operational Research,
86(3), 395–401.
Zangwill, W.I., (1969). A backlogging model and a multi-echelon model of a dynamic economic lot size
production system—a network approach, Management Science, 15(9), 506–527.
This research work deals with the multi-product multi-period inventory lot sizing with supplier selection
problem.

An algorithm based on a reduce and optimize approach (ROA) is proposed to solve the multi-product
multi-period inventory lot sizing with supplier selection problem.

Numerical experiments ratify the success of the proposed heuristic algorithm.

On the set of the benchmark instances, the algorithm always obtained better solutions compared with
those formerly published.

According to the computational results the developed heuristic algorithm outperforms CPLEX MIP solver
in both solution quality and computational time.