You are on page 1of 5

Comparing two metaheuristics for the Joint Replenishment Problem

with Quantity Discounts


Orlando Durn
Escuela de Ingeniera Mecnica, Pontificia Universidad Catlica de Valparaso,
Quilpue, V region, Chile
Email: orlando.duran@ucv.cl

Abstract. This work presents the definition and solution of an optimization model using
techniques based on the Particle Swarm Optimization (PSO) algorithm or the Particle Swarm
Optimization and Genetic Algorithms. The model corresponds to the Joint Replenishment
Problem in a system operating with quantity discounts. Results are shown in problems, where
their size makes it difficult or impossible to establish the optimum solution (through exhaustive
search), besides presenting comparisons with the results from Genetic Algorithms.
Keywords:Metaheuristics,ParticleSwarm,JointReplenishmentProblem,QuantityDiscounts,CombinatorialOptimization.

1Introduction

Goyal [5] presented a heuristics to solve the Joint Replenishment problem (JRP), a problem of
the joint replenishment, in situations of multi-products obtained from the same supplier. After
that work, various articles have dealt with the problem using different approaches. A thorough
review of the literature by late eighties may be found in Goyal and Satir [6]. There are solution
strategies for the constant demand, stochastic demand and dynamic demand. One of the
classical heuristics for the problem is known as RAND (Kaspi and Rosenblatt [9]). Recently,
some strategies based on evolutionary algorithms to solve the JRP have emerged. Khouja [10]
developed a genetic algorithm (GA) and compared its performance with the RAND. Under the
JRP point of view under the demand with stochastic behavior, two policies stand out in the
literature. Johansen and Melchiors [8] consider a periodical replenishment policy and a can-
order-type policy. Regarding the dynamic demand case, the work by Boctor [1] shows a
procedure combining different heuristics, which gives excellent results. Other works for special
JRP cases are presented by Klein and Ventura [11] who solve the replenishment problem with
discreet time. Khouja et al. [10] approach the problem under the continuous reduction condition
of the unitary costs of the items. Only one work has approached the JRP with the existence of
discounts per quantity (Cha and Moon, 2005, [2]). That work shows a model that uses two
propositions in order to develop two solution algorithms. The algorithms were proved in 1,600
problems randomly developed.

About the use of evolutionary algorithms for the JRP, we can mention Hong and Kim [7], who
developed a genetic algorithm to solve the JRP problem by easing the restriction that the
replenishment cycle is a multiple of the item cycle, being more frequent among those involved
in the system. Thus, according to the authors, further precision is achieved in the estimation of
the system cost. Leung et al. [12] have presented an extension of the classical JRP problem with
multi-customer and multi-item. They have suggested and tested a simulated annealing
combination (SA) genetic algorithm (GA) called SAGA. The authors have tested the
applicability of the strategy suggested in a Hong Kong bank. Chan et al. (2003) [3] presented
an extension of the classical multi-customer and multi-item JRP problem, thus proposing a
genetic algorithm. Khouja (2000) [10] compared the performance of a genetic algorithm with
JRP with the RAND performance. Olsen (2005) [13] has used a genetic algorithm for the JR
problem, using the direct grouping concept, which, according to the author exceeds the
traditional algorithm (RAND) in situations where the quotient between the fixed cost of order
and the variable cost of order is considered high. In 2008, Olsen [14] deepened in the use of AG
in the JRP problem, approaching the situation where costs of order of an item depend on other
items to be jointly replenished (inter-dependant costs). Dye and Hsieh (2010) [4] tested the
efficiency of using another evolutionary algorithm, the Particle Swarm Optimization in a JR
problem.

2DesignoftheExperiment

The main hypothesis ruling this work is that when metaheuristic techniques are used is then
possible to solve optimization models for the joint replenishment problems with discounts per
quantity when the item quantity results to be significant. Based on this hypothesis, an
optimization model was defined. Through experimentation and using a random data group, the
model was optimized using Particle Swarm optimization intelligence to respond to the
following investigation question: Is it possible to optimize the inventory cost in systems with
joint replenishment in situations where quantity discounts are applied using optimization
techniques based on metaheuristics? The main objectives of this work are as follows: a) To
define an inventory optimization model that considers the joint replenishment problem (JRP)
with suppliers using discounts per quantity. b) To use techniques of combinatory optimization to
solve the optimization model developed. c) To use the algorithms developed in a situation of
large quantity items and to compare the performance of the applied optimization techniques.

2.1 Parameters

Ci = Order cost associated to the item i ($/order)


i = Index showing the item (i = 1,2,3, n).
Di= Demand associated to the item i (uni/time)
Pi = Unitary price associated to the item i ($/unit).
Qi = Quantity to be replenished of the item i[unit].
Ii= Discount rate applied to the item inventory i [/time].
T = Purchase Period [time].
ki = Integer representing the interval number the inventory associated to the item i (ki = 1,2,3,
, n).
i,j =Discount rate applied to Pi in function of Qi [].
j = Index indicating the discount interval for Qi (i = 1,2,3, n).
m = Integer representing the total duration of a purchasing cycle [].

2.2. Definition of the Problem


The problem involves searching for the minimum cost for a system operating with joint
replenishment and where suppliers offer discounts per quantity to be able to define the purchase
system of a large quantity of items. First, a model of costs showing such situation is established.
In order to start the development of the optimization model, some assumptions must be defined,
such as: Known and constant purchase period., Known and constant demand rate for all items;
No delay in the delivery; Only one supplier for each item exists; Finite analysis horizon; The
inventory of any item goes to zero the same day a replenishments occurs; each item involves an
specific order cost; and, there is a single order fixed cost, which is incurred into every time a
replenishment occurs, being just for one or more items.

As one of the assumptions indicates, the inventory of any item falls to zero every ki*T period,
just when they are bought again. The latter may mean the inventory of all items may go
simultaneously to zero after m * T periods, where m is the minimum common multiple of all ki.
Then, the decision variable corresponds to a vector formed by the ki, i.e. every other period the
item ki must be purchased. The, we will assume T corresponds to a time unit. Once the
optimization model has been defined, the metaheuristics will aim to find the values of ki
assuring the lowest global cost of inventory management with joint replenishment with
discounts per quantity. As shown in figure 1, the inventory of the item 1 (blue line) lasts 2
periods of time (2T); therefore, k1 equals 2. The inventory of the item 2 (red line) lasts 3
periods of time (3T), so k2 equals 3. The inventory of the item 3 (green line) lasts 4 periods
of time (4T), thus k3 equals 4. Just as shown by our assumption, the inventory of any item falls
to zero every ki*T periods, just in the moment they are purchased again. The latter allows
concluding the inventory of all items will fall simultaneously to zero after m * T periods, where
m is the minimum common multiple of all ki. In this example, it has a value of 12 periods.

Inventory
[u]

T ............. T T Time [ut]


T T
... T
Cycle=m*T
Fig.1.Purchasecycleanddurationofeachinventory

The total cost incurred during the purchase, order and maintenance in inventory of any item is:

The total cost incurred during the purchase, order and maintenance in inventory of all items is:

However, as such cost represents the total cost of a particular cycle, which duration depends on
the values ki, this cannot be compared to the cost of another cycle, as for different ki we obtain
a different m (representing the total duration in periods T of the cycle). For example, if the
duration of a cycle for certain values ki is of 3 periods of time (m=3), then the total cost will
represent the amount to be paid during the whole cycle, i.e. it is the cost incurred during a cycle
of 3 periods of time. If the duration of the cycle for values ki different to the aforementioned
corresponded to 5 period of time (m=5), the total cost will represent the amount to be paid
during the complete cycle, i.e. the cost incurred during the cycle of 5 periods of time. Then,
such values cannot be compared. A way to eliminate the problem is to take this cost to the unit
of a period dividing the total cost CT by the quantity of periods m the cycle lasts, as follows:

Therefore, no matter the values ki, this cost can be compared and, thus, minimized as while the
period T remains constant (which is the basis of all assumptions) no errors will occur.
Where delta_k represents the fraction of non-void replenishments during the m periods. This
factor may be calculated by:
Figure 2 shows the aspect of the total cost function for a case with two items. The horizontal
axes show the variations of k1 and k2. It may be observed this function has several local
minimums, which may damage the performance of the optimization technique.

Fig.2.Functionoftotalcostfortwostockitems.

3. Results and discussion

Three problems were randomly generated. The first involved 10 items, the second 20 items and
the third 100 items. For each problem, a Genetic Algorithm was applied, based on the Particle
Swarm. Each technique was applied 30 times to each problem.

Table1Comparativeresultsobtainedthroughbothtechniques.

Technique P.S.O A.G


Problems 10 20 100 10 20 100
RelativeWidth%totheMinimum 2,01% 1,59% 8,72% 4,85% 4,25% 8,62%
%ofcases<1%ofdiversion 87% 23% 3% 87% 50% 6%
%ofcases<5%ofdiversion 100% 97% 43% 97% 100% 93%
8 de
MinimumNumbers 30 1 de 30 1 de 30 18 de 30 3 de 30 1 de 30

Table 1 is shown for comparison purposes. There the following indicators or metrics are shown:
Relative Percentage Width to the Minimum reached Value, % of results within a range not
exceeding 1% of the minimum, % of results within a range no higher than 5% of the minimum
and number of times the minimum was obtained. It may be observed from these results both
algorithms show an adequate Repeatability, although the dispersion of the PSO algorithm is
lower, the results with AG are higher in terms of the minimum values obtained. According to
the latter, and considering the impossibility to obtain optimum values for comparison, we
believe both algorithms are capable of obtaining proper solutions in most cases. Table 2 shows
the cost values obtained for each experiment.

4. Conclusions
The greatest difference between this work and previous ones lies in the fact this project
combines the joint replenishment problem with the existence of discounts in the unitary cost of
the items according to the acquired quantity. There is only one work in the literature presenting
a model for this situation. However, the solution mechanism presented is a heuristics that only
considers a limited quantity of items. There are no previous works in the literature tackling with
the use of these techniques in the optimization of situations with a large quantity of items. The
model was evaluated, solved and compared using metrics established by literature. These
problems are used in situations where companies try to make the most of the economies of
scale, optimizing at the same time the order costs. The latter may translate, for instance, in
shipment and processing costs or in difficult access premises, such as mining works or large
civil work construction.
Table2Costsobtainedforbothtechniques

10items 20items 100items


PSO AG PSO AG PSO AG
MIN $3.098.400 $3.042.517 $5.610.813 $5.571.515 $38.493.000 $34.019.000
MEDIA $3.112.410 $3.055.248 $5.665.196 $5.634.189 $40.521.466 $34.850.367
MAX $3.160.800 $3.189.978 $5.700.279 $5.808.135 $41.849.000 $41.849.000

References

[1]. Boctor, G. Laporte and J. Renaud, Models and algorithms for the dynamic demand joint
replenishment problem, International Journal of Production Research 42 (2004), pp. 2667
2678
[2]. Cha B., I. Moon, The joint replenishment problem with quantity discounts, OR Spectrum,
27, pp. 569581, 2005.
[3]. Chan CK, Cheung BKS, Langevin A (2003) Solving the multi-buyer joint replenishment
problem with a modified genetic algorithm TRANSPORTATION RESEARCH PART B-
METHODOLOGICAL Vol: 37, n. 3: 291-299.
[4]. Dye CY, Hsieh TP (2010), A particle swarm optimization for solving joint pricing and lot-
sizing problem with fluctuating demand and unit purchasing cost: COMPUTERS &
MATHEMATICS WITH APPLICATIONS. Vol.60, n.7: 1895-1907
[5]. Goyal S (1974) Determination of optimum packaging frequency of items jointly
replenished. Management Science 23: 436443
[6]. Goyal S, Satir A (1989) Joint replenishment inventory control: deterministic and stochastic
models. European Journal of Operational Research 38: 213
[7]. Hong S-P, Kim, Y-H. A genetic algorithm for joint replenishment based on the exact
inventory cost. Computers & Operations Research, Volume 36, Issue 1, January 2009, Pages
167175
[8]. Johansen, S.G., P. Melchiors, Can-order policy for the periodic review joint replenishment
problem, Journal of the Operational Research Society 54 (2003), pp. 283290.
[9]. Kaspi M, Rosenblatt M (1991) On the economic ordering quantity for jointly replenished
items. International Journal of Production Research 29: 107114.
[10]. Khouja M, Michalewicz Z, Satoskar S (2000) A comparison between genetic algorithms
and the RAND method for solving the joint replenishment problem. Production Planning &
Control 11: 556564
[11]. Klein C. M., J. A. Ventura , An Optimal Method for a Deterministic Joint
Replenishment Inventory Policy in Discrete Time, The Journal of the Operational Research
Society Vol. 46, No. 5 (May, 1995), pp. 649-657
[12]. Leung TW, Chan CK, Troutt MD (2008) A mixed simulated annealing-genetic
algorithm approach to the multi-buyer multi-item joint replenishment problem: Advantages
of meta-heuristics JOURNAL OF INDUSTRIAL AND MANAGEMENT
OPTIMIZATION. Vol: 4, n.:1: 53-66.
[13]. Olsen AL (2005), An evolutionary algorithm to solve the joint replenishment problem
using direct grouping, COMPUTERS & INDUSTRIAL ENGINEERING. Vol: 48, n.: 2:
223-235.
[14]. Olsen AL (2008) Inventory replenishment with interdependent ordering costs: An evolutionary
algorithm solution, INTERNATIONAL JOURNAL OF PRODUCTION ECONOMICS Vol: 113, 1:
359-369.

You might also like