You are on page 1of 5

Neural Comput & Applic (2014) 25:67–71

DOI 10.1007/s00521-013-1448-1

ORIGINAL ARTICLE

Network site optimization of reverse logistics for E-commerce


based on genetic algorithm
Dawei Liu

Received: 30 August 2012 / Accepted: 28 June 2013 / Published online: 10 July 2013
Ó Springer-Verlag London 2013

Abstract Product returns have been viewed as an customers’ satisfaction and cost controls. Costs of reverse
unavoidable cost for online sales, forfeiting any chance of logistics account for about 16 % of total logistical cost of
cost savings. As competition pressure continues to mount Taobao [1]. It is essential that electronic companies be able
in E-commerce, B2B or B2C, E-commerce corporations to physically handle returns. Optimization of network site
have begun to explore the possibility of managing product of reverse logistics for E-commerce can help firms
returns in a more cost-efficient manner by rescheduling the decreasing the cost great without decreasing the service
locations of recycling centers. However, few studies have quality.
addressed the problem of optimization of the numbers and In this paper, a genetic algorithm-based model is pro-
location of collecting points, recycling centers and posed for the design of reverse logistics network site
remanufacturing centers for E-commerce based on genetic handling product returns. The model considers the supply
algorithm. To fill the void in such a line of research, this of returned products from online buyers. It considers
paper proposes a genetic algorithm-based model that can storing, reprocessing, remanufacturing facilities and new
solve the location-allocation optimization of reverse module suppliers in the network site. We consider demand
logistics for E-commerce. The usefulness of the proposed for maintaining customers’ satisfaction level while keeps
model and algorithm was validated by its application to an the cost reasonable for stockholders. The design of such a
illustrative example dealing with logistics sites allocation network sites is strategic as it involves a decision on the
from online sales. number, location and capacities of various facilities and
allocation of material flows between them [2–4] and is one
Keywords Reverse logistics  E-commerce  Genetic of the most challenging elements of managing reverse
algorithm  Network site logistics operations [3]. A properly designed network sites
can also enhance dealing with remanufacturing activities
[5] and competitive advantage.
1 Introduction

Because of the value involved and the great affects on 2 Model assumption
customer relationships, reverse logistics and logistics-
related operational capabilities should be regarded as a To optimize the network sites of reverse logistics, many
managerial priority. Famous E-commerce enterprises, such scholars take mixed integer linear programming and its
as eBay and Taobao, are seeking balance between algorithm improvement as modeling and solving functions.
Total cost minimization is usually set as the final objective,
and constraint conditions include logistics balance, manu-
D. Liu (&) facture capabilities, parameters of nonnegative constraints,
School of Management, Hangzhou Dianzi University,
etc. Location and numbers of network sites are delivered
No. 1st, 2nd Ave. Xiasha Education Zone, Hangzhou 310018,
Zhejiang, People’s Republic of China based on calculation and simulation [6].
e-mail: hduldw@163.com To simplify the model, it is assumed that:

123
68 Neural Comput & Applic (2014) 25:67–71


1. The model only considering single product single cycle 1 remanufacturing center k
Zk ¼ k2H
cost, which means measurable economic cost in operation 0 others
process, do not consider time cost, social efficiency, etc.
Xij, Yjk, Lj, Zi, Zj and Zk are decision variables.
2. Reverse objects from collect point, all sent to the
recycling center.
2.2 Model expression
3. Cost of recycling and transportation, a certain item is
known.
We have the objective function and constraint condition as:
4. Functions of the collecting points, remanufacturing XX XX X
centers and recycling centers are clear. min S ¼ Mij Xij þ Njk Yjk þ Aj Lj
i2G j2T j2T k2H j2T
X X X
2.1 Description of variables þ Oi Zi þ Pj Zj þ Qk Zk
i2G j2T k2H

In our model, collecting point, remanufacturing center and Constraint conditions (i [ G, j [ T, k [ H) are:
recycling center are three key points, whose relationship is X
shown as: Xij  Vi ð1Þ
j
X
Yik  Dk ð2Þ
j

1 collectiong point i
Zi ¼ i2G ð3Þ
0 others

1 recycling center j
Zj ¼ j2T ð4Þ
0 others

1 remanufacturing center k
Zk ¼ k2H ð5Þ
0 others
Based on above assumptions, we have:
Xij  0 ð6Þ
G Set of recycling centers;
T Set of collecting points; Yjk  0 ð7Þ
H Set of remanufacturing centers;
Equation 1 means that items sent to recycling center
Vi Processing capability of collecting point i;
j are not larger than its capability of processing;
Dj Recycling capability of recycling center j;
Equation 2 means that items sent to remanufacturing
Rk Handling capability of remanufacturing center k;
center k are not larger than its capability of processing;
Aj Cost of per item of recycling center;
Equations 3–5 means whether the requirement is
Oi Fixed cost of collecting point i;
accepted or not;
Pj Fixed cost of recycling center j;
Equations 6 and 7 means that the distances between
Qk Fixed cost of remanufacturing center k;
each center are meaningful;
Mij Transportation cost of per item from collecting point
The model includes two parts of cost: the costs of trans-
i to recycling center j;
portation between each center and the fixed cost of all centers.
Njk Transportation cost of per item from recycling center
j to remanufacturing center k;
Xij Product quantity from collecting point i to recycling 3 Model solving based on genetic algorithm
center j;
Yij Product quantity from recycling center j to The penalty function is the genetic algorithm for solving
remanufacturing center k; the constrained optimization problem, the most commonly
Lj Actual product quantity of recycling center j; used technique. Penalty function method is used to deal
with constraint problem by penalizing infeasible solution,
the constrained problem into the unconstrained problem

1 collecting point i [7]. How to determine reasonable penalty function is the
Zi ¼ i2G treatment difficulty lies, because it should not only consider
0 others
 the solution of constrained condition, which is not satisfied
1 recycling center j degree, but also consider the genetic algorithm in the cal-
Zj ¼ j2T
0 others culation of efficiency requirements.

123
Neural Comput & Applic (2014) 25:67–71 69

Optimization model of general expression for: 3.1 Encoding


min f ð X Þ
Bit-string encoding is adopted in encoding process that
s.t. Gi  0 ði ¼ 1; 2; . . . M Þ each decision variable is lined up in a certain order. Based
Hj ð xÞ ¼ 0 ðj ¼ M þ 1; M þ 2; . . . N Þ on our model, the whole encoding can be divided into two
X 2 Rn parts, each part includes eight bits. The first bit of each part
is the processing capability of each recycling center (less
with penalty function, the model is changed to optimization than the maximum capability). The second bit is the scale
problem with no subjection as: of the collecting point. The third bit is whether or not set up
eval ð X Þ ¼ f ð X Þ þ Pð X Þ the recycling center (0 or 1). The fourth bit is whether or
not set up the collecting point (0 or 1). The fifth and sixth
P(X) is the penalty function, P(X) is consisted with two bits are transport quantity between collecting point to
parts, which are: recycling center (less than the second bit). The seventh and
1 XN eighth bits are ratio of items transported to recycling center
PðXÞ ¼ d2 ðXÞ divided by the whole collecting items. Here, we have two
2a i¼1 i
sites (Table 1).
where a is the variable penalty, di(X) is:
X
8 X
8 3.2 Fitness function
S ¼ 1;100 Hi þ Hi  ti
i¼1 i¼1 Fitness function indicates the merits of the solution. To
deal with the minimum optimization, initial target value
þ 200; 000 ð1  i  M; M þ 1  j  NÞ should be replaced with fitness to assure the best item
shows the best fitness. Fitness function is the product of
Determination of penalty is difficult, and the model total cost and penalty, where if vk is the chromosome k of
solution quality depends greatly on the penalty value. the current population, g(vk) is the fitness function, Z(vk) is
When the penalty is not proper, the algorithm may the target value and g(vk) = M * Z(vk). M is the penalty.
converge to the infeasible solution. On the one hand, the When the chromosome is not proper, the penalty is larger
penalty is too small, fails to give the feasible solution to than 1, or it is 1. Z(vk) is calculated with target function, we
penalty enough pressure; on the other hand, the penalty is have:
too large when the method is equivalent to the refusal XX XX X X
strategies [8]. Therefore, when a penalty a is explored, S¼ Mij Xij þ Nij Yjk þ Aj Lj þ Oi Zi
i2G j2T j2T k2H j2T i2G
simulated annealing is adopted to determine the penalty a X X
by article n generation, which is: þ Pj Zj þ Qk Zk
j2T k2H
a0

1þt
and 1/S is chose as the fitness.
where a0 is the initial coefficients.
We can get the expression of penalty function P(X) as: 3.3 Genetic operation
XX XX X X
Mij Xij þ Njk Yjk þ Aj L j þ Pj Zj Hybridization and mutation are conducted in the operation.
i¼1 j¼1
Single point hybridization is adopted. Based on hybrid-
The fitness function is F and we set a larger Z and have: ization rate, two chromosomes are chosen from the popu-
lation and hybridize on a randomly chosen point. We can
F ¼ Z  eval ðeval\ Z Þ
find the hybridization data in Table 2.
With fitness function, item with less eval value always has To maintain the diversity of the population, consecutive
a higher fitness in the function. and multiple trading of mutation strategy is taken to assure

Table 1 Bit-string encoding


210 250 1 1 113 139 0.98 0.99 320 450 0 0 0 0 0.02 0.01
180 300 0 1 202 248 0.76 0.38 240 380 1 1 230 178 0.24 0.62
Roman values are for site one, italicized values are for sit two

123
70 Neural Comput & Applic (2014) 25:67–71

Table 2 Hybridization data


210 250 1 1 11 13 0.98 0.99 320 450 0 0 0 0 0.02 0.01
180 300 0 1 20 24 0.76 0.38 240 380 1 1 23 17 0.24 0.62
210 250 1 1 11 13 0.98 0.99 240 380 1 1 23 17 0.24 0.62
180 300 0 1 20 24 0.76 0.38 320 450 0 0 0 0 0.02 0.01
Column 8 is the randomly chosen point

Table 3 Result of cross and aberrance mutation


210 250 1 1 11 13 0.98 0.99 320 450 0 0 0 0 0.02 0.01
210 250 1 1 13 11 0.98 0.99 320 450 0 0 0 0 0.02 0.01
Column 5 and 6 are the randomly chosen points

Table 4 Result of single point mutation


210 250 1 1 11 13 0.98 0.99 320 450 0 0 0 0 0.02 0.01
210 250 1 1 11 13 0.76 0.99 320 450 0 0 0 0 0.02 0.01
Column 7 is the randomly chosen points

each individual has evident transform in position [9]. The Since the sample electronic corporation has limited
research takes two kinds of mutations. If there are two optional locations, approximate solving encoding is adop-
mutation points in 5th and 6th position within the same ted. Decision of the recycling center location is made on
chromosome, two points cross and aberrance algorithm location chosen odds. Therefore, we turn the discrete
will be adopted, otherwise single point mutation is adopted. variables into continuous variables to enlarge the popula-
We have results of cross and aberrance mutation in Table 3 tion scale. Location chosen odds can be deducted from
and single point mutation in Table 4. transformation formula. In encoding, anti-trigonometric
function is chosen as the transformation formula, and
power methods chosen are shown in Fig. 1.
4 Program construction It is found that y = 2a sin(x.^4)/pi is satisfactory, based
on exchange, y is approximately to the expectation of 0 and
We chose recycling center J1 based on the location of the col- 1, when x is between 0 and 0.8, y is close to 0; when x is
lecting points and all optional recycling center locations. The between 0.8 and 1, y is close to 1. When y is larger than 0.3,
other three recycling centers will be chosen on model pro- we accept the location as the recycling center. Figure 2
gramming, which can control the searching cost effectively. shows that when x [ 0.8, y [ 0.3.

Fig. 1 Comparison of anti-trigonometric functions Fig. 2 Analytical diagram of transformation formula

123
Neural Comput & Applic (2014) 25:67–71 71

We concluded that J2 is the best solution, and it is shown


in tracking map of genetic algorithm as Fig. 3.
When the genetic generation is more than 10, the vari-
ances of the solutions and population average are steady as
shown in Fig. 3. Therefore, the model of the paper is
effective with a ideal fitness.

Acknowledgments The research on which this paper reports has


been financially supported by Natural Science Foundation of Zhejiang
Province with Project Y6110648 and Hangzhou Social Science Fund
by Project 2011JD23.

References

1. Shi H et al (2012) Association between PPAR-c and RXR-a gene


polymorphism and metabolic syndrome risk: a case-control study
of a Chinese han population. Arch Med Res 43(3):233–242
Fig. 3 Tracking map of genetic algorithm 2. Lee CKM, Lam JSL (2012) Managing reverse logistics to
enhance sustainability of industrial marketing. Ind Mark Manag
41(4):589–598
5 Conclusion 3. Ilgin MA, Gupta SM (2010) Environmentally conscious manu-
facturing and product recovery (ECMPRO): a review of the state
Genetic algorithm generally can lead to approximate opti- of the art. J Environ Manag 91(3):563–591
4. Aras N, Aksen D, Gönül Tanuğur A (2008) Locating collection
mal solution, not absolute optimal solution, and its con- centers for incentive-dependent returns under a pick-up policy
vergence is affected by the initial inputs [10]. To solve the with capacitated vehicles. Eur J Oper Res 191(3):1223–1240
problem, the calculation should be carried out multiple 5. Alagador D, Cerdeira JO (2007) Designing spatially-explicit
times, which means different initial individuals should be reserve networks in the presence of mandatory sites. Biol Con-
serv 137(2):254–262
calculated to approach the optimal solution. In our case, the 6. Brandão SN, Sauer J, Schön I (2010) Circumantarctic distribution
calculation is conducted by 10 times. Location odds of X1, in Southern Ocean benthos? A genetic test using the genus
X2 and X3 (for optional location J1, J2 and J3) are as follows: Macroscapha (Crustacea, Ostracoda) as a model. Mol Phylogenet
Evol 55(3):1055–1069
7. Kerh T, Gunaratnam D, Chan YL (2010) Neural computing with
0.8923 0.1145 0.2174 genetic algorithm in evaluating potentially hazardous metropoli-
tan areas result from earthquake. Neural Comput Appl 19(4):
0.9723 0.2974 0.3053 521–529
0.8877 0.53ew 0.6226 8. Acampora G, Gaeta M, Loia V (2011) Hierarchical optimization
0.9122 0.4281 0.3677 of personalized experiences for e-Learning systems through
evolutionary models. Neural Comput Appl 20(5):641–657
0.9037 0.3183 0.5745
9. Dash NB et al (2010) Hybrid neural modeling for groundwater
0.8920 0.1758 0.3365 level prediction. Neural Comput Appl 19(8):1251–1263
0.9295 0.2107 0.1101 10. Dehuri S, Cho SB (2010) A hybrid genetic based functional link
0.9822 0.1935 0.1639 artificial neural network with a statistical comparison of classi-
fiers over multiple datasets. Neural Comput Appl 19(2):317–328
0.9704 0.2073 0.4483
0.9345 0.4933 0.2084

123

You might also like