A Two-Stage Algorithm for the Closed-Loop Location-Inventory Problem Model Considering Returns in E-Commerce

Facility location and inventory control are critical and highly related problems in the design of logistics system for e-commerce. Meanwhile, the return ratio in Internet sales was significantly higher than in the traditional business. Focusing on the existing problem in e-commerce logistics system, we formulate a closed-loop location-inventory problem model considering returned merchandise to minimize the total cost which is produced in both forward and reverse logistics networks. To solve this nonlinear mixed programming model, an effective two-stage heuristic algorithm named LRCAC is designed by combining Lagrangian relaxation with ant colony algorithm (AC). Results of numerical examples show that LRCAC outperforms ant colony algorithm (AC) on optimal solution and computing stability. The proposed model is able to help managers make the right decisions under e-commerce environment.


Introduction
The increasing progress of information and prevalence of internet in the 21st century has forced the e-commerce to develop in a world-wide range.In 2012, B2C e-commerce sales grew 21.1% to top $1 trillion for the first time in history in the whole world [1].Comparing with traditional commerce, customers are more liable to return goods under e-commerce environment.Note that many customer returns online account for 35% of original orders [2,3].Therefore, logistics systems as an important support system in ecommerce need to be adjusted and improved.To adapt to the reality of e-commerce market environment, it is critical to conduct the research on the reverse logistics network and highly integrated logistics process.
Facility location and inventory control are critical problems in the design of logistics system.There is much previous work in these two areas.In fact, there is a mutually dependent relationship among these problems in logistics system.Comprehensive optimizing and logistics activities management should be based on this relationship [4].According to this idea, besides location allocation problem and inventory optimization, the location-inventory problem (LIP) starts to be researched.
Many papers about the LIP are studied deeply and have made some abundant achievements.In recent years, intelligent algorithms and heuristic algorithm have been used to solve LIP model [5][6][7].In the reverse logistics research field, LIP attracts researchers' attention.Lieckens and Vandaele [8] applied a queuing mode in reverse logistics network to solve the facility location problem while considering the impact of inventory costs.Srivastava [9] established a reverse logistics network optimization model to optimize the location-allocation problem and capacity decisions, and he used heuristic algorithm to solve the model.Wang et al. [10] proposed a location-inventory policy in Chinese B2C electronic market as a bilevel programming model.Tancrez et al. [11] studied the LIP in three-level supply chain networks including reverse logistics; they developed an iterative heuristics approach to solve the model.Diabat et al. [12] built a mixed integer nonlinear programming (MINLP) model to minimize the total reverse logistics cost by finding out the number and location of initial collection point and centralized return center considering the inventory cost.Two solution approaches, namely, genetic algorithm (GA) and artificial immune system, are implemented and compared.However, research on the LIP of closed-loop logistics system is limited.Sahyouni et al. [13] designed three generic facility location models that account for the integrated distribution and collection of products in the closed-loop supply chain networks; the authors described a Lagrangian relaxationbased solution algorithm to solve the models.Easwaran and Üster [14] offered a mixed integer linear programming model to optimize the total cost that consists of location, processing, and transportation costs of the multimerchandise in closedloop supply chains; they introduced a heuristics solution approach that combines Benders decomposition and tabu search to solve the model.Abdallah et al. [15] presented the uncapacitated closed-loop location-inventory model; a sensitivity analysis for different parameters of the model reveals that the value of recovered products is a major factor in the economic feasibility of the closed-loop network.For dealing with returned merchandise without quality problems in e-commerce, Li et al. [16] developed a practical LIP model with considering the vehicle routing under e-supply chain environment and provided a new hybrid heuristic algorithm to solve this model.
Previous researches on the closed-loop logistics system optimization mainly focus on the minimization of the total cost of the network.To our best knowledge, few researches on manufacturing/remanufacturing system consider returns and concept of green logistics recycling in logistics network.Since customers may be dissatisfied with merchandise and return it, the cost of processing returns, the cost of inventory and shipping, order time, and size are changed.Furthermore, research on the LIP with return of closed-loop logistics system is limited.
The aim of this study is to develop a practical LIP model with the consideration of returns in e-commerce and provide a new two-stage heuristics algorithm.To our best knowledge, this work is the first step to introduce returns into the LIP in e-commerce, which makes it become more practical.We also provide an effective algorithm named Lagrangian relaxation combined with ant colony algorithm (LRCAC) to solve this model.Lagrangian relaxation algorithm (LR) can obtain a near-optimal solution by analyzing the upper bound and lower bound of objective function.But its effectiveness mainly relies on the performance of subgradient optimization algorithm.On the other hand, AC has great ability of local searching.If there is an appropriate initial solution, the performance of AC will be good.To adopt their strong points while overcoming their weak points, we combine the two algorithms.Results of numerical examples show that LRCAC outperforms ant colony algorithm (AC) on optimal solution and computing stability.
The remainder of the paper is structured as follows.In Section 2, a nonlinear integrated programming model about LIP considering returns in e-commerce is designed.Section 3 proposes the heuristic algorithm named LRCAC based on Lagrangian relaxation and ant colony algorithm.Section 4 shows and analyzes the results of different experiments.Section 5 concludes this paper and discusses the future research directions.

Problem and Mathematic Model
2.1.Problem Description.In e-commerce, some returned merchandise has a high integrity, which makes it usually not in need of being repaired and can reenter the sales channels after simply repackaging [17].Some returned goods have quality problems; they have to be sent back to factory for repair.Therefore, we merge the recycling center with distribution center as merchandise centers (MCs) with an additional inspection function.MC is responsible for distributing normal goods to the sale regions; meanwhile the returned goods are collected to MCs.After inspecting at MCs, the returned goods with quality problems are sent back to factory, and the other returned goods are resalable as normal goods after simply repackaging.Customers can choose to return goods in e-commerce, and quantity of the returns is uncertain [18].However, for a certain sale region (SR), the quantity of the returns can be usually seen as stochastic variable.
The objective of this paper is to determine the quantity, locations, order times, and order size of MCs in the closedloop logistics network in e-commerce.The final target is to minimize the total cost and improve the efficiency of logistics operations.The involved decisions are as follows: (1) location decisions, the optimal number of MCs, and their locations; (2) allocation decisions, the corresponding service relationship between MCs, and sale regions; (3) inventory decisions, the optimal order times, and order size.

Assumptions. (1)
There is a single type of merchandise; (2) the capability of factory is unlimited; (3) the capability of MCs is unlimited; (4) the demand and return of each sale region comply with the normal distribution, whose parameters are fixed; (5) the demands of regions are mutually independent; (6) returned merchandise is inspected and repackaged at MCs.   : Optimal order times at MC  .

Model Formula
(1) Location Cost.The construction cost of MC  is given by ∑  =1     .
(2) Inventory Cost.The total annual inventory cost consists of ordering cost and inventory holding cost, according to literatures [19,20]; it is given by     + ℎ  (  /2  ).
( (4) Transportation Cost.The transportation cost consists of cost from factory to MC, cost from MC to customer region for forward logistics, cost from customer region to MC, and cost from MC to factory for reverse logistics.So, the transportation cost is given by (5) Repacking Cost.The returned goods without quality problems need to be repacked before reentering to sale channel, so the repacking cost is given by To sum up, the location-inventory model with returned merchandise (RLIP) is The objective function ( 1) is to minimize the system's total cost.Constraint (2) ensures that each sale region must be assigned to a MC.Constraint (3) stipulates that the assignment can only be made to the selected MC.Constraints (4) and ( 5) are standard integrality constraints.Constraints ( 6) and ( 7) are nonnegative constraints.

Solution Approach
On the one hand, Lagrangian relaxation algorithm (LR) is used to solve the complex optimization problem very often.It can obtain a near-optimal solution by analyzing the upper bound and lower bound of objective function.But its effectiveness mainly relies on the performance of subgradient optimization algorithm.The speed of convergence becomes more and more slow with the increasing of the number of iterations.On the other hand, AC has great ability of local searching.If there is an appropriate initial solution, the performance of AC will be good.To adopt their strong points while overcoming their weak points, we design a two-stage algorithm.In the first stage, we use LR algorithm to get a near-optimum solution.In the second stage, let the solution obtained from the first stage be the initial solution; we use AC to further improve it.
The abstract idea of solution approach is described as follows.Firstly, we give the formula for solving optimal order quantity   and optimal order times   , which also rely on the decision variables   and   .Secondly, we use LR algorithm to get a near-optimal solution by computing the lower bound and upper bound of objective function.Then, let the near-optimum solution obtained from LR be the initial solution; we use AC to further improve it.

Finding the Optimal Order Quantity and Optimal Order
Times.In the model ( 1)-( 7), the decision variable   only has appeared in the objective function.Also, the objective function is convex for   > 0. Consequently, we can obtain the optimal value of   by taking the derivative of the objective function with respect to   as   = √ℎ    /2    , where As we know the optimal order quantity   =   /  , so there is 3.2.Transforming the Objective Function.In order to apply the LR algorithm, we transform the objective function as linear teams and nonlinear teams separately.The objective function can be rearranged as follows: where For fixed values of the Lagrange multipliers   , we want to minimize (10) over the location variables   and the assignment variables   .We separate the linear teams and nonlinear teams.
(1) For each MC  , let   =    + ∑ ∈ min(0,    −   ) + ℎ   , and let If all   values are positive, we identify the smallest positive   and set the corresponding   = 1.The assignment variables are then easy to determine, setting as follows: (2) However, the presence of the nonlinear terms makes finding an appropriate value of   difficult.So, we need to solve a subproblem as the following form for each candidate MC  : where   =    −   ,   =  2   2  ≥ 0.
In (13), we use   to substitute   .The solution of subproblem SP() refers to literature [19]; the solution of ( 10) is the summary of SP() and    .To get the lower bound, we need to find the optimal Lagrange multipliers.We do so using a standard subgradient optimization procedure as illustrated in literatures [21,22].The optimal value of ( 10) is a lower bound of the objective function (1).

Finding an Upper Bound.
We find an upper bound as follows.
We initially fix the MC locations at those sites for which   = 1 in the current Lagrangian solution.Then we assign SR to MCs in a two-phased process. 1.For each SR  , for which ∑ ∈J   ≥ 1, we assign the SR  to the MC  for which   = 1 and that increases the least cost based on the assignments made so far. 2. We process SR  , for which ∑ ∈J   = 0; we assign each SR to the open MC which increases the least total cost based on the assignments made so far.
Hence, for these SRs, we consider all possible assignments to open MCs, and the cost of this stage is the upper bound.to represent the probability of the SR  and clustering center  at time .The formula of    () is shown as follows: where Tub  () = 0 represents that ant  can cluster SR  in next step;    is the amount of pheromone deposited for transition from state  to ;  is the parameter used to control the influence of    ;    is the desirability of state transition  and ;  is the parameter of controlling the influence of    ;   is the distance from  to .
And the following relationship exists:       The flowchart for our algorithm is shown in Figure 1.

Computational Experiments and Algorithm Analysis
4.1.Computational Experiment.We refer to the logistics network of company  in Hubei province of China as an example.We convert the latitude and longitude coordinates  Based on MATLAB 7.0 platform, we programmed the LRCAC algorithm and run it 30 times on the computer (CPU: Intel Core2 P7570 @2.26 GHz 2.27 GHz; RAM: 2.0 GB; OS: Windows 7); the optimal result is in Table 3.
The optimal cost is 224965 yuan, and logistics network is shown in Figure 2.
For comparison, we programmed AC algorithm in the same platform and run 30 times on the same computer.The optimal objective function values of these two algorithms are shown in Table 4.
The two optimization trends of the two algorithms are shown in Figures 3 and 4.
The fluctuation curves of optimal objective function in 30 times are shown in Figures 5 and 6, respectively.
We can see that the LRCAC algorithm can converge more quickly than AC from Figures 3 and 4.Moreover, LRCAC has better stability than AC, which can be easily found from Table 4 and Figures 5 and 6.

Algorithms Analysis.
In this section, all the data in our experiments come from LRP database of the University of Aveiro [23].A series of experiments show that LRCAC is more efficient and stable than AC.Results of numerical example in Section 4.1 show that the related parameters of LRCAC are reasonable.Thus, we employ these parameters in the remainder of this section.Each instance was calculated 30 times by LRCAC and AC, respectively; the results are shown in Table 5.In this table, Srivastava 86 is the name of this instance; 8 × 2 means there are 8 SRs and 2 candidate MCs, so do others.The coordinate of all nodes and the demands of SRs are given by the database.Table 5 shows that LRCAC can obtain better objective function value and stability than AC.

Conclusion
Customers have a higher return rate in the e-commerce environment.Some returned goods have quality problems and need to be sent back to the factory for repair.The others without quality problems can be reentered in the sales channels just after a simple repackaging process.This phenomenon puts forward high requirements to the logistics system that supports the operation of e-commerce.This study handles the above interesting problem and provides an effective heuristic.The main contributions are as follows.
(1) In reality, the cost of processing returned merchandise is produced by considering the condition that customers are not satisfied with products and return them.We firstly design a closed-loop LIP model to minimize the total cost which is produced in both forward and reverse logistics networks.It is able to help managers make the right decision in e-commerce, decreasing the cost of logistics and improving the operational efficiency of e-commerce.(2) The above closed-loop LIP model with returns is difficult to be solved by analytical method.Thus, a two-stage heuristic algorithm named LRCAC is designed by integrating Lagrangian relaxation with AC to solve the model.
(3) Results of our experiments show that LRCAC outperforms AC on both optimal solution and computing stability.LRCAC is a good candidate to effectively solve the proposed LIP model with returns.
However, some extensions should be considered in further work.Considering the dynamic of the demand, a dynamic model should be established.Considering the fuzzy demand of customs or related fuzzy costs, more practical LIP model should be developed.Moreover, differential evolution algorithms (DEs) have turned out to be one of the best evolutionary algorithms in a variety of fields [24,25].In the future, we may use an improved DE to find better solutions for the LIPs.The integration research and practice of the management of e-commerce logistics system can be constantly improved.

3. 4 .
Ant Colony Clustering.According to the clustering behavior of ant colony, we set the clustering probability    () Start Transform the objective function Finding the lower bound of the objective function Finding the upper bound of Let the near-optimum solution as the initial solution of AC Lagrangian relaxation First stage Initialize the tabu search matrix Set the MC j as the ant nest.Ant k selects a i to its ant nest with and taboos the SR. for solving Q j and N j Select the near-optimum solution Improved ant colony algorithm the objective function SR Tub(t) Tub ki Tub ki (t) updates to null matrix, update the amount of pheromone  ij (t + h) =  ij (t) + Δ ij , Δ ij = ∑ m k=1 Δ k ij , and record the optimal solution

Figure 1 :
Figure 1: The working process of the integral two-stage algorithm.

Figure 4 :
Figure 4: Trends of optimal objective function value by AC.

Figure 5 :
Figure 5: Fluctuation curve of total cost by LRCAC algorithm.

Figure 6 :
Figure 6: Fluctuation curve of total cost by AC.
Constants  : Fixed cost (annual) administrative and operational cost of MC  ;   : Shipping cost per unit of merchandise between factory and MC  ;   : The delivering cost per unit of merchandise between SR  and MC  ; ℎ  2  : Variance of annual demand at SR  ;    : Standard normal deviate such that  ( ≤    ) =   ;   : Service level of MC  ;   : The quantity of return at SR  ; Sets I: Set of SR; J: Set of candidate MC.  : The inventory holding cost per unit of merchandise per year at MC  ;   : Ordering cost per time at MC  ;   : Lead time at MC  ;   : Mean of annual demand at SR  ;  : 1, if the candidate MC  is selected as a MC and 0 otherwise;   : 1, if SR  is served by MC  and 0 otherwise;   : Optimal order size at MC  ; ∈I  2    , so the safety stock is    √  ∑ ∈I  2    , and the safety stock cost is given by ℎ     √  ∑ ∈I  2    .

Table 1 :
Parameters of MCs.

Table 2 :
Parameters of SRs.

Table 3 :
Optimal results.We give the formula for solving   and   , which also rely on the decision variables   and   ; 3. Find the upper bound of objective function by using the LR. 4. Select the solution whose value is equal or approximately equal to the average value of the lower bound and upper bound as near-optimum solution.
Set the MC  as the ant nest   .Ant  selects a SR  to its ant nest   with    () and taboos the SR.If Tub  () is full, go to Step 8; else repeat Step 7.  8.If all the SRs are clustered to MC, the Tub  () updates to null matrix and goes to Step 9; else go to Step 6.  9. Update the amount of pheromone   ( + ℎ) =   () + Δ  , Δ  = ∑  =1 Δ   and record the optimal solution. 10.If the conditions of convergence are meeting, terminate the procedure and output the optimal solution; else remove the MC of the least SRs and go to Step 6.

Table 4 :
Statistical results of optimal objective function value of two algorithms.in which the values represent the actual kilometers.And other parameters' values are as follows: randomly generate the values between 100 and 160 as the   , and assume that the  2  is equal to   ,   = 1,   = 2, ℎ  = 1,   = 7 (day),   = 97.5%, = 0.2, and  = 2.

Table 5 :
Optimal objective function values of two algorithms.