A Column-Generation Approach for a Short-Term Production Planning Problem in Closed-Loop Supply Chains

We present a new model formulation for a multi-product lot-sizing problem with product returns and remanufacturing subject to a capacity constraint. The given external demand of the products has to be satisfied by remanufactured or newly produced goods. The objective is to determine a feasible production plan, which minimizes production, holding, and setup costs. As the LP relaxation of a model formulation based on the well-known CLSP leads to very poor lower bounds, we propose a column-generation approach to determine tighter bounds. The lower bound obtained by column generation can be easily transferred into a feasible solution by a truncated branch-and-bound approach using CPLEX. The results of an extensive numerical study show the high solution quality of the proposed solution approach.


Introduction
In the last decade, many researchers have been attracted by production planning aspects in closed-loop supply chains (Fleischmann, Bloemhof Ruwaard, Dekker, van der Laan, van Nunen, and van Wassenhove 1997;Guide and Van Wassenhove 2001;Rubio, Chamorro, and Miranda 2008;Sbihi and Eglese 2010). In addition to forward-oriented supply chains from the manufacturer to the customer, the reverse direction also has to be taken into account in the field of closed-loop supply chains. This means that customers return old and used products to the manufacturer at the product's end-of-use. Well-known examples are printer cartridges and single-use cameras (Akcali andCetinkaya 2011: 2376). These returned products can be remanufactured to become products, which are as good as new. Therefore, remanufactured products can also be used to fulfill customers' demand. Returned products are denoted as recoverables, i.e., these items have to be remanufactured. After (re)manufacturing products are denoted as serviceables. The products can be (re)manufactured on a capacity-restricted production system. For (re)manufacturing products a setup is necessary, which causes setup costs and/or setup times. Hence, a multi-product capacitated lot-sizing problem arises, as the respective products compete for the production system's scarce capacity. Most of the established model formulations neglect either the capacity constraints or the interactions of multiple products. Therefore, a new model formulation for the multi-product capacitated lot-sizing problem with product returns and remanufacturing (CLSP-RM) is presented. Furthermore, a solution approach by combining column generation and a truncated branch-andbound approach is proposed. The remainder of this paper is organized as follows: In section 2 we provide a literature review of lot-sizing problems in closed-loop supply chains. Afterwards, we present a new model formulation

BuR --Business Research Official Open Access Journal of VHB German Academic Association for Business Research (VHB) Volume 6 | Issue 1 | May 2013 | 55--75
for a capacitated lot-sizing problem with remanufacturing in section 3. In this section, we also introduce valid inequalities to tighten the lower bounds of the LP relaxation. In section 4 a columngeneration approach is proposed to determine a lower bound close to the optimal solution. To generate a feasible production plan based on this lower bound, a truncated branch-and-bound method is applied. Numerical results are reported in section 5. The paper ends with a summary and an outlook for further research in section 6.

Literature Review
Reviews of the literature on dynamic lot sizing in general are given, e.g., by Karimi, Fatemi Ghomi, and Wilson (2003), Quadt and Kuhn (2008) and Buschkühl, Sahling, Helber, and Tempelmeier (2010). In this section, we focus only on dynamic lot-sizing problems with product returns and remanufacturing. However, the literature on lot-sizing problems in closed-loop supply chains is rather limited. One of the first papers on dynamic lot-sizing problems in closed-loop supply chains was presented by Richter and Sombrutzki (2000). They suggested a model formulation for the reverse Wagner-Whitin problem (Wagner and Whitin 1958) by assuming that the demand can be totally satisfied by remanufacturing returned products. Richter and Weber (2001) extended the reverse Wagner-Whitin problem by introducing variable manufacturing and remanufacturing costs. Richter and Sombrutzki (2000) as well as Richter and Weber (2001) proved that the reverse Wagner-Whitin problem can usually be solved in polynomial time. Teunter, Bayindir, and van den Heuvel (2006) described model formulations for the single-level uncapacitated lot-sizing problem with remanufacturing (SLULSP-RM) based on the Wagner-Whitin problem. Two cases are distinguished: In the first case, the products are manufactured and remanufactured on the same (uncapacitated) production system and only one joint setup is necessary for both remanufacturing and manufacturing. In the second case, it is assumed that manufacturing and remanufacturing take place on different (uncapacitated) production systems. Thus, a separate setup is required on each production system. Teunter, Bayindir, and van den Heuvel (2006) proved that in the case of joint setups the SLULSP-RM can be solved in polynomial time using an adapted version of Wagner and Whitin's dynamic programming approach. Furthermore, van den Heuvel (2004) showed that the SLULSP-RM is N P-hard in the case of separate setups. To solve both cases of the SLULSP-RM, Teunter, Bayindir, and van den Heuvel (2006) adapted wellknown heuristics for the Wagner-Whitin problem, namely the least-unit cost approach, the Silver-Meal heuristic and the part-period-balancing approach. They pointed out that the adapted Silver-Meal heuristic outperforms the other approaches. Schulz (2011) introduced further variants of the Silver-Meal heuristic to solve the SLULSP-RM with separate setup costs. Furthermore, he proposed improvement steps yielding a higher solution quality than the Silver-Meal heuristic alone. To strengthen the lower bound of the LP relaxation of the SLULSP-RM, Retel Helmrich, Jans, van den Heuvel, and Wagelmans (2010) adapted valid inequalities proposed by Barany, van Roy, and Wolsey (1984) for the Wagner-Whitin problem and present reformulations based on the shortestpath problem for the SLULSP-RM. Quariguasi Frota Neto, Walther, Bloemhof, van Nunen, and Spengler (2009) integrated sustainable aspects into the SLULSP-RM by taking the required energy for (re)manufacturing into account. The SLULSP-RM with joint setups was extended by Schwarz, Buscher, andRudert (2009) andSchwarz (2010) by allowing disposals of the returned products. The dynamic programming approach by Teunter, Bayindir, and van den Heuvel (2006) was adapted to solve the SLULSP-RM with disposals without an increase of the numerical effort. Pineyro and Viera (2010) also considered the possibility to dispose returned products. A tabu search heuristic was suggested to solve this kind of lot-sizing problem. Li, Chen, and Cai (2006) presented an uncapacitated multi-product lot-sizing problem. However, the interactions of these products by competing for the scarce capacity are neglected. To the best of our knowledge, only a few papers addressed capacitated lot-sizing problems with product returns and remanufacturing. Li, Chen, and Cai (2007) described a model formulation for a lot-sizing problem with product substitution, i.e., the remanufactured product serves as a substitute for the serviceable. The production capacity is limited, while setup times are neglected. A genetic algorithm is applied to determine a produc- Here, the number of produced, remanufactured and disposed items are restricted in each period. Akcali and Cetinkaya (2011) pointed out that most of the lot-sizing model formulations with product returns and remanufacturing in the existing literature only investigate single-item problems. Therefore, the interactions between several items produced on the same production system are totally neglected. Neglecting these interactions often leads to an infeasible production plan if these items compete for the scarce capacities of the production system. Hence, we present in this paper a new model formulation for the multi-product capacitated lot-sizing problem with product returns and remanufacturing. This CLSP-RM is solved by combining a column-generation approach and a truncated branch-and-bound method.

Model Formulation of the CLSP-RM
The objective of the capacitated lot-sizing problem with product returns and remanufacturing is to determine a production schedule for K products äk ã 1; : : : ; Kå which minimizes the sum of setup, production, remanufacturing and holding costs over a planning horizon of T periods ät ã 1; : : : ; Tå.
This production schedule includes the production quantities Q kt and the remanufacturing quantities Q r kt as well as the end-of-period inventory levels of the serviceables Y kt and of the recoverables Y r kt of product k in period t. For each product k the external demand d kt has to be fulfilled in the respective period t, while backlogging is not allowed. Serviceables of product k can be held in stock (Y kt ) to satisfy the demand in later periods at holding costs hc k per unit and period. The quantities of returned items r kt of product k in period t are also known in advance. The returned items can be remanufactured (Q r kt ) to become as good as new, i.e. serviceables. Furthermore, returned products can also be held in stock as recoverables (Y r kt ) at holding costs hc r k per unit and period to be remanufactured in later periods. The variable production costs are pc k and the remanufacturing costs are pc r k for one unit of product k. The products are (re)manufactured on a production system, which consists of two separated resources: On the first resource, the products can only be manufactured, while remanufacturing is located on the second resource. For this reason, a (separate) setup is required for producing and another setup for remanufacturing. For manufacturing product k, a setup leads to setup costs sc k and setup times ts k . If product k is manufactured in period t, the setup variable kt equals 1, otherwise kt ã 0. For remanufacturing product k, different setup costs sc r k and setup times ts r k are considered.

BuR --Business Research Official Open Access Journal of VHB German Academic Association for Business Research (VHB) Volume 6 | Issue 1 | May 2013 | 55--75
A further setup variable r kt is used to model a setup for the recoverables of product k in period t. Analogously, the setup variable r kt of product k in period t equals 1, if product k is remanufactured in the respective period, otherwise r kt ã 0. The capacity c t of the manufacturing system and the capacity c r t of the remanufacturing system are given in each period t. The production times are tp k and the remanufacturing times are tp r k for product k. The described production process is visualized in Figure 1. The CLSP-RM with separate setups (CLSP-RM-SS) can be stated using the notation in Table 1:

Model CLSP-RM-SS:
( The objective function (1) minimizes the sum of inventory holding, production, remanufacturing and Indices and index sets: K set of products äk ∈ è1; : : : ; Kéå T set of periods ät ∈ è1; : : : ; Téå Parameters: c t available capacity of the resource for manufacturing in period t c r t available capacity of the resource for remanufacturing in period t d kt external demand of product k in period t hc k holding cost of product k per unit and period hc r k holding cost for returns of product k per unit and period M kt big number for product k in period t pc k production cost of product k per unit pc r k remanufacturing cost of product k per unit r kt returns of product k in period t sc k setup cost of product k sc r k setup cost of returned product k tp k production time for one unit of product k tp r k remanufacturing time for one unit of product k ts k setup time of product k ts r k setup time of returned product k Decision variables: Q kt production quantity of product k in period t Q r kt remanufacturing quantity of product k in period t Y kt inventory of product k at the end of period t Y r kt inventory of recoverables of product k at the end of period t kt binary (joint) setup variable of product k in period t r kt binary setup variable for remanufacturing product k in period t setup costs. Equations (2) and (3) represent the inventory balance constraints for the serviceables, respectively the recoverables. Equations (2) ensure that the given demand d kt for product k is satisfied completely in each period t. Analogously, restrictions (3) guarantee that the returned items are either held in stock or remanufactured in the respective period. Constraints (4) and (5) contain the capacity restrictions. In both cases, the production

BuR --Business Research Official Open Access Journal of VHB German Academic Association for Business Research (VHB) Volume 6 | Issue 1 | May 2013 | 55--75
(re)manufacturing times and also the setup times must not exceed the given capacity. Constraints (6) and (7) link the production variables Q kt and Q r kt to the respective binary setup variable kt or r kt for product k in period t. Constraints (8) are the non-negativity constraints. The setup variables are defined to be binary, according to constraints (9). To tighten the linking constraints (6) and (7), the parameter M kt can be determined as follows: which describes the maximum (re)manufacturing quantity of product k that can possibly be (re)manufactured in period t due to the available capacity.
In the next step, we present a small numerical example for the CLSP-RM-SS with 4 products and 5 periods. The underlying parameter settings are given in Appendix 6. The optimal (re)manufacturing quantities are shown in Figure 2. A Gantt chart and the optimal production plan are given in Appendix 6 (Figure 7(a) and Table 10). Production Quantity k=1 Remanufacturing Quantity k=1 Production Quantity k=2 Remanufacturing Quantity k=2 Production Quantity k=3 Remanufacturing Quantity k=3 Production Quantity k=4 Remanufacturing Quantity k=4 The capacity of the resource for manufacturing is almost completely used. With the exception of period 3, products are either manufactured or remanufactured. Only one product is manufactured and remanufactured as well, e.g., product 2 in the third period. The total costs of the production plan amount to 9620. In the case of joint setups, the products are (re)manufactured on the same resource with a given capacity c t in each period t. A (joint) setup is necessary to prepare the production system to manufacture and/or remanufacture product k in period t. This setup leads to setup costs sc k and setup times ts k . If product k is produced and/or remanufactured in period t, the respective setup variable kt equals 1, otherwise kt ã 0.
The CLSP-RM with joint setups (CLSP-RM-JS) can be modeled as follows.
Model CLSP-RM-JS: In the objective function (10) of the CLSP-RM-JS, the setup costs sc r k can be omitted compared to (1). Due to the consideration of a common production system, only one capacity constraint (11) is required. Finally, the linking constraints (6) and (7) can be combined into constraints (12). Production Quantity k=1 Remanufacturing Quantity k=1 Production Quantity k=2 Remanufacturing Quantity k=2 Production Quantity k=3 Remanufacturing Quantity k=3 Production Quantity k=4 Remanufacturing Quantity k=4 Finally, the numerical example is also solved in the case of the CLSP-RM-JS. The optimal (re)manufacturing quantities are illustrated in Figure 3. A Gantt chart and the optimal production

BuR --Business Research Official Open Access Journal of VHB German Academic Association for Business Research (VHB) Volume 6 | Issue 1 | May 2013 | 55--75
plan are also provided in Appendix 6 ( Figure 7(b) and Table 11).
In the case of the CLSP-RM-JS, the total costs amount to 6090 and are, therefore, lower compared to the CLSP-RM-SS. However, please note that the objective function values are not comparable as only one setup is required for (re)manufacturing in the case of the CLSP-RM-JS.

Complexity Results
Theorem 1 The CLSP-RM is N P-hard.
Proof By setting r kt ã 0 for all products k and periods t the CLSP-RM can be reduced to the wellknown CLSP. As Florian, Lenstra, and Kan (1980) showed that the CLSP is N P-hard, the CLSP-RM is also N P-hard.
Theorem 2 For the CLSP-RM, the proof of the existence of a feasible solution is N P-complete in the case of positive setup times äts k > 0å.
Proof As explained above, the CLSP-RM can be easily reduced to the CLSP. Florian, Lenstra, and Kan (1980) also proved that it is N P-complete to prove the existence of a feasible solution for the CLSP in the case of positive setup times. Therefore, the same holds for the CLSP-RM.

Model Extensions of the CLSP-RM
Practical extensions of the CLSP-RM can be, e.g., the use of overtime or allowing back-orders. In each period the available capacity c t of the resource for manufacturing can be extended by the use of overtime (O t ). The overtime costs are oc for each required unit of overtime. Analogously, the amount of overtime required on the resource for remanufacturing is denoted as O r t . Furthermore, in the case of a stock-out, the demand can be back-ordered; in this case a backlog B kt has to be taken into account. The backlog costs are bc k for product k. In the following we present a model formulation for the CLSP-RM-SS with backlogs and overtime by using the additional notation presented in Table 2.

Model CLSP-RM-SS-BO-Ov:
In the objective function (1) the backlog and overtime costs have to be taken into account, see (13). Furthermore, backlogs have to be considered in the inventory balance equation (14). The capacity restrictions (4) and (5) are extended by the use of overtime, see (15) and (16). Finally, backlogs and the amount of overtime are non-negative according to (17). The CLSP-RM-JS can be extended analogously.
It is worth mentioning that the CLSP-RM-SS-BO-Ov is no longer N Pácomplete since backlogging and the use of overtime are not restricted.

Valid Inequalities
As the well-known CLSP is a special case of the CLSP-RM, it is not surprising that first numerical experiments have shown that the LP relaxation of the proposed model formulations for the CLSP-RM yields very poor lower bounds. Thus, including valid inequalities in the model formulation is often promising for accelerating the solution process. Therefore, we use the adapted valid inequalities for the SLULSP-RM proposed by Retel Helmrich, Jans, van den Heuvel, and Wagelmans (2010).
In the case of separate setups, the valid inequalities are defined as follows: M ks ± ä ks à r ks å ∀ k; t ã 1; : : : ; T á 1; p ã 1; : : : The valid inequalities (18) can be interpreted as follows: If we assume no setup activities for product k in the periods t à 1; : : : ; t à p, the end-of-period inventory Y kt must be high enough to fulfill the total demand in the periods t à 1; : : : ; t à p. Otherwise, if the end-of-period inventory Y kt is not sufficient at least one setup is required in the periods t à 1; : : : ; t à p. Analogously, inequalities (19) are defined for the recoverables. If joint setups are considered, the valid inequalities can be defined as follows: M ks ± ks ∀ k; t ã 1; : : : ; T á 1; p ã 1; : : : The interpretation of the valid inequalities (20) and (21) is similar to the explanation of (18) and (19). Obviously, the valid inequalities are also valid for the CLSP-RM, as the demand has to be fulfilled without backlogging and disposals are not allowed. The adaptation of the VI in the case of back-orders can be found in Appendix 6. Due to the N P-hardness of the CLSP-RM it is nearly impossible to solve problem instances of practical size to optimality within a reasonable time. Therefore, we propose an MP-based solution approach to determine high-quality solutions. At first, a column-generation (CG) approach (Haase 2005) is applied to generate tight lower bounds for the CLSP-RM. Afterwards, a truncated branchand-bound approach is used to construct a solution based on the lower bound provided by CG.

Basic Idea
The general idea of a column-generation approach can be described as follows (Bahl 1983;Cattrysse, Maes, and van Wassenhove 1990): A reformulation of the CLSP-RM is used as a so-called master problem. At first, the master problem includes only a very limited number of columns. Iteratively, new columns are generated for the master problem by solving a respective subproblem. If the new column promises a reduction of the objective function value, this new column is included in the master problem. New columns are generated iteratively as long as they reduce the current objective function's value of the master problem.
In the following, we only present the proceeding of the CG approach for the CLSP-RM with separate setups. Analogously, the CG approach can be adapted to the CLSP-RM with joint setups straightforwardly.

The Master Problem --A Set
Partitioning Problem For the CG approach, we use a reformulation of the CLSP-RM, namely the Set Partitioning Problem (SPP, Garfinkel and Nemhauser 1969;Cattrysse, Maes, and van Wassenhove 1990) with capacity restrictions, as the LP relaxation of the SPP provides a substantially tighter lower bound (Haase 2005) than the LP relaxation of the presented model formulation.
The objective of the SPP is to choose exactly one production schedule for each product k at minimal

BuR --Business Research Official Open Access Journal of VHB German Academic Association for Business Research (VHB) Volume 6 | Issue 1 | May 2013 | 55--75
costs so that the capacity constraints are met in each period. Therefore, a set of production schedules S k exists for each product k which has to be generated in advance. For each production schedule n ∈ S k of product k the total costs nc kn (including setup, (re)manufacturing and holding costs) are known. Furthermore, the capacity consumptions tp ktn (including manufacturing and setup times) and tp r ktn (including remanufacturing and setup times) can also be derived. For the choice of a schedule n for product k a binary variable kn is used. Here, kn equals 1, if schedule n is chosen for product k and otherwise kn ã 0.

Table 3: Additional notation used for the SPP
Indices and index sets: S k set of schedules of product k än ∈ è1; : : : ; S k éå Parameters: nc kn cost of schedule n of product k tp ktn capacity consumption of schedule n for manufacturing product k in period t tp r ktn capacity consumption of schedule n for remanufacturing product k in period t Decision variables: kn binary variable for choosing schedule n of product k In the following, we present the model formulation of the SPP using the additional notation given in Table 3.

Model SPP:
The objective function (22) minimizes the total costs. Equation (23) guarantees that exactly one schedule is chosen for each product k. The capacity constraints (24) and (25) ensure that the capacity consumption of all products does not exceed the given capacity in period t. Constraints (26) define the variables kn to be binary. It is worth mentioning that the optimal solution of the SPP is also the optimal solution of the CLSP-RM if all possible production schedules are known in advance.
As the generation of all production schedules is quite time consuming, new schedules are determined iteratively and included in the SPP afterwards. Therefore, in each iteration a subproblem is solved in order to generate a new schedule n for each product k. This new production schedule n is included in the SPP, i.e. in the set S k , if it promises a reduction of the current objective function value, i.e., the reduced costs are negative. However, if the number of schedules increases, the solution time to solve the SPP to optimality also increases substantially. Hence, to reduce the numerical effort, we only solve the LP relaxation of the SPP and thus, the CG approach terminates with a lower bound for the CLSP-RM.

The Subproblem --An SLULSP-RM
As a subproblem in the CG approach, we solve an SLULSP-RM (Teunter, Bayindir, and van den Heuvel 2006) to optimality for each product k.
Numerical investigations have shown that the numerical effort to solve an instance of the SLULSP-RM with separate (or joint) setups to optimality is almost negligible using CPLEX with the VIs (18) and (19), respectively (20) and (21). The SLULSP-RM k can be stated for a respective product k using the dual variables k , t , and r t of the relaxed SPP. Here, k denotes the dual variable which corresponds to equation (23) for product k. The dual variables (or shadow prices) t and r t correspond to the capacity constraint (24), respectively (25) of period t.
Model SLULSP-RM k : (2), (3), (6) -- (9) The objective function (27) of the SLULSP-RM k tends to minimize the reduced costs of the potentially new schedule for product k. The remaining constraints equal those of the CLSP-RM for the respective product k except for the capacity constraints (11). Each SLULSP-RM k is solved to optimality using CPLEX. A new production schedule for product k is only included in the set S k if the reduced costs of this new schedule are negative, i.e., the objective function value of the SLULSP-RM k is negative.

Outline of the Column-Generation Approach
To initialize the column-generation approach we start with a dummy schedule for each product k. Within a dummy schedule, no production and no setups take place, i.e., q kt ã 0 for all products k and all periods t. Hence, there exists no capacity consumption, i.e., tp ktn ã tp r ktn ã 0. However, to avoid the choice of these schedules in the optimal solution, they are penalized at very high costs nc kn . In each iteration, the LP relaxation of the master problem is solved first. Afterwards, the optimal solution of the related SLULSP-RM k is determined for each product k by taking the dual variables k , t , and r t into account (see (27)). If the reduced costs of the corresponding new schedule, i.e., the objective function value of the related SLULSP-RM k , are negative this new schedule is included in the set S k . The algorithm terminates if the reduced costs are non-negative for all products in the current iteration. The LP relaxation of the SPP yields a lower bound for the CLSP-RM. Note that no feasible lower bound exists if any dummy schedule is selected in the solution of the LP relaxation after the column-generation approach terminates. In the next step, the lower bound obtained by the column-generation approach is used to generate a feasible solution for the CLSP-RM by a truncated branch-and-bound approach using CPLEX. To minimize the numerical effort all binary setup variables are fixed for those products k with an integer solution in the LP relaxation of the SPP, i.e., ∃ n ∈ S k with kn ã 1. The respective setup variables kt and r kt of product k are fixed related to the setup pattern, which corresponds to the production schedule n with kn ã 1. After a given time limit the branch-and-bound approach (BîB) terminates with a production plan for the CLSP-RM. A flow chart of the CG approach is presented in Figure 5. In Figure 4, the course of the objective function value of the relaxed SPP is displayed for the numerical example in section 3.1 in the case of the CLSP-RM-SS. After the generation of 9 production schedules for each product the CG approach terminates with a feasible lower bound. In Appendix 6 we present necessary adaptations of the solution approach in the case of back-orders and overtime.

Description of the Test Design
We analyze the quality of our solution approach by defining five problem classes (PC) by varying the number of products and periods. Each problem class consists of 216 test instances (TI) leading to 1080 TI. Table 4 gives an overview of these five problem classes. We vary different parameters to define the TI, e.g., the time between orders (TBO) to determine setup costs, the utilization Util of the production  system, the consideration of setup times ts and the portion of holding costs hc r for returned products compared to the holding costs of the serviceables. A detailed description of the TI including calculation rules is given in Appendix 6. Table 5 shows a list of these parameters and their range of values. We implemented the model formulations and the solution approach in GAMS 23.7. For the solution of the subproblems and for the determination of reference values we used CPLEX 12.3. The MP-

Solution Quality of CPLEX Reference Values
The valid inequalities (18) and (19) are included in the CLSP-RM-SS to speed up the solution process  (20) and (21) are used in the case of joint setups, respectively. In Table 6 the solution quality of the CPLEX reference values is reported, for both the case of joint and separate setups. The CPLEX reference values and the numerical results of the solution approach are provided for all TI online in the supplementary material. A (feasible) solution has been found for all TI. The average solution time in seconds is given iǹ`T CPU CPX ''. As we do not expect to solve all TI to optimality within a reasonable time, each TI is solved by a truncated branch-and-bound approach within a given time limit (column``TLim CPX ''). The average integrality gap is reported in columǹ`A vgGAP CPX '', where GAP CPX TI is derived as follows:

BuR --Business Research Official Open Access Journal of VHB German Academic Association for Business Research (VHB) Volume 6 | Issue 1 | May 2013 | 55--75
For each TI, Sol CPX TI describes the best objective function value found. LowB CPX TI denotes the best lower bound obtained by CPLEX within the given time limit. In the case of joint setups, all TI of PC 1 and a large number of TI of the remaining problem classes could be solved to optimality within the given time limit using the cluster TANE as reported in columǹ`O ptSol CPX ''. However, in the case of separate setups, CPLEX often fails to find the optimal solution within the given time limit due to the larger number of binary setup variables. In the following, the ref-erence solutions obtained by CPLEX are denoted as CPLEX solutions or values.

Numerical Analysis of the Lower
Bounds obtained by Column-Generation To get an impression regarding the quality of the lower bound obtained by the column-generation approach, we first compare this lower bound with the CPLEX reference values. In Table 7, the average solution time in seconds to generate a feasible lower bound for the CLSP-RM is reported in column``TCPU CG ''. For each TI the deviation GAP CG TI of the CPLEX reference solution Sol CPX TI from the lower bound LowB CG TI obtained by CG can be determined as follows: The average deviation is given in columǹ`A vgGAP CG ''. Additionally, column``OptSol CG '' indicates the portion of TI whose lower bound equals the optimal solution. Finally, the portion of products with an integer solution in the lower bound is given in column``KFixed''. Table 7 gives an overview of the numerical results of the CG approach for all five problem classes. For all TI a feasible lower bound is obtained by the column-generation approach. Here, a lower bound is called feasible if the LP relaxation does not include any dummy schedule. As expected, the run time increases with respect to the rising number of products.

BuR --Business Research Official Open Access Journal of VHB German Academic Association for Business Research (VHB) Volume 6 | Issue 1 | May 2013 | 55--75
In the case of joint setups, the average integrality GAP does not exceed 0.5ô, while the average lower bound even amounts to only 0.01ô in the case of the largest PC 5. Thus, the lower bound obtained by the CG approach is very close to the optimal solution. At least 25ô of the TI could already be solved to optimality via CG. The portion of products with an integer solution (∃ n ∈ S k with kn ã 1) is quite high at more than 65ô. The portion even rises to 96ô for PC 5.
In the case of separate setups, the average solution time increases substantially due to the N Phardness of the SLULSP-RM with separate setups. The average integrality gap is almost comparable to the lower bound obtained for the case of joint setups. The portion of products with an integer solution ( kn ã 1) is only slightly lower compared to the portion of products in the case of joint setups. A detailed evaluation of our numerical results shows that an increase of setup costs and, therefore, of the TBO leads to a significant increase of the solution time. The integrality GAP also deteriorates with respect to the increase in setup costs. It is worth mentioning that at least 75ô of the TI with low setup costs (TBO ã 1) could even be solved to optimality with the CG approach in the case of joint setups. In summary, the solution quality decreases slightly if the setup cost rises. The numerical results of the CG approach with respect to the varying setup costs can be found in Table 12.
In the case of scarce capacities, i.e., a higher utilization, the numerical effort rises as the solution times in column``TCPU CG '' shows. Simultaneously, the number of iterations increases as well. The average lower bound also rises in the case of a higher utilization. The numerical results with respect to the varying utilization of the production system are presented in Table 13. We observed that the impact of the remaining parameters on the solution quality is rather limited.

Numerical Analysis of the Upper Bounds obtained by the truncated Branch-and-Bound Method
In the next step, we investigate the solution quality of the production plans obtained by the truncated BîB method. Although a setup decision has already been fixed for a large number of products in the lower bound obtained by the CG approach, we do not expect to solve the remaining problem to optimality within a reasonable time. Therefore, a time limit is used for the truncated BîB approach. In Table 8 the numerical results of the upper bound obtained by the BîB approach are reported. In contrast to Table 7, the column``TCPU CG '' is substituted by``TCPU BîB ''. Here, the entries in column``TCPU BîB '' show the run time in seconds of the truncated BîB approach. For each TI the deviation Dev BîB TI of the CPLEX reference solution Sol CPX TI from the solution Sol BîB TI obtained by the In column``AvgDev BîB '' the average deviation from the CPLEX reference solution is reported. Furthermore, we report the average gap based on the lower bound obtained by the CG approach in column``AvgGAP BîB '', where GAP BîB is defined as follows:

BuR --Business Research Official Open Access Journal of VHB German Academic Association for Business Research (VHB) Volume 6 | Issue 1 | May 2013 | 55--75
In the case of joint setups, the portion of optimally solved TI are given in``OptSol BîB ''. As the number of optimally solved TI is rather small in the case of separate setups, the column``OptSol BîB '' is substituted by the portion of TI which are solved at least as well as by CPLEX or even better. Therefore, the column is named``BetSol BîB ''. The given time limit is reported in``TLim BîB '' Due to the very small number of remaining free binary variables the numerical effort is rather limited. In the case of joint setups, the mean deviation from the CPLEX reference values ranges between 0.01ô and 0.58ô. This can be explained by the large portion of already fixed products after the CG approach (Table 7). More than 39ô of all TI could be solved to optimality.
In the case of separate setups, the average deviation from the CPLEX reference values is also very moderate. For more than 18ô of the TI the truncated BîB method finds a production plan which is at least as good as the CPLEX solution or even better. Furthermore, the truncated BîB determines for all TI a better solution than CPLEX in the case of PC 5. The low average GAP compared to the lower bound obtained by CG also demonstrates the high solution quality of the combined solution approach. The numerical results of the truncated BîB method based on the different TBO patterns are reported in Table 14. Analogous to the observations regarding the CG approach, the solution time rises with respect to increasing setup costs. The same holds true for the solution quality. In the case of low setup costs (TBO ã 1) the average deviation from the CPLEX reference values is lower compared to the average deviation in the case of high setup costs (TBO ã 4). For low setup costs, at least 87.50ô of all TI could be solved to optimality. This number decreases substantially to 0ô with increasing setup costs. Furthermore, the total solution time of both approaches also grows according to rising setup costs. For the sake of completeness, we give the numerical results of the truncated BîB method with respect to the utilization of the production system in Ta-

BuR --Business Research Official Open Access Journal of VHB German Academic Association for Business Research (VHB) Volume 6 | Issue 1 | May 2013 | 55--75
ble 15. The rise of solution times is also noticeable in the case of a higher utilization, while the average deviation from the CPLEX reference values deteriorates only marginally. Worth mentioning are the results of PC 5: Contrary to the results of the other classes, the solution quality does not decline in the cases of higher setup costs and/or higher utilization. This is due to the fact that CPLEX fails to find better solutions within the given time limit. We also investigated the solution quality of the proposed solution approach by allowing overtime and back-orders. The solution quality and the solution time are nearly comparable. Therefore, we abstain from showing these numerical results.

Summary and Outlook
In this paper, we presented a new model formulation for a capacitated lot-sizing problem with product returns and remanufacturing. In literature, only one-product problems have been considered so that the influence of different products on the same production systems are completely ignored. Furthermore, most of these approaches in literature neglect capacity restrictions. Thus, these production plans are of limited value as the capacity restrictions are often violated. Obviously, a rescheduling is necessary in these cases. The proposed solution approach yields a very high solution quality. The lower bounds found by the CG approach are very close to the optimal solution as our numerical investigation shows. The combination of CG and a truncated branch-and-bound method leads to a very high solution quality and the solutions are very close to the CPLEX reference values, but with substantially reduced numerical effort. Furthermore, we have shown that an integrated (re)manufacturing planning can yield a cost reduction. Future research should address the adaptation of further solution approaches, e.g., the Fix-and-Optimize heuristic by Helber and Sahling (2010). Possible extensions of the CLSP-RM will be linked lot sizes where the setup state of one product can be carried over to a subsequent period. Furthermore, different quality levels of the returned products should be taken into account. Different quality levels will result in varying remanufacturing costs and times.

Appendices Appendix A: Description of the Numerical Example
For the numerical example in section 3.1, the setup costs sc k and sc r k equal 500 for each product k. The holding costs are hc k ã 1 and hc r k ã 0:5. The (re)manufacturing times tp k and tp r k are assumed to be 1 and the setup times ts k and ts r k are equal to 20 for each product k. The demand and returns are given in Table 9.  1  20  30  30  20  30  2  50  40  60  40  40  3  40  60  50  30  40  4  30  30  30 50 20 In the case of the CLSP-RM-SS, the capacities c t and c r t amount to 300 capacity units in each period. In Table 10, the optimal production plan is given for the CLSP-RM-SS. In the case of the CLSP-RM-JS, the capacity c t amounts to 600 capacity units in each period. In Table 11, the optimal production plan is given for  For the solution approach, the use of overtime has to be considered only in the master problem. Therefore, the objective function (22), the capacity constraints (24) and (25)