Abstract
In the area of reverse logistics, remanufacturing has been proven to be a valuable option for product recovery. In many industries, each step of the products' recovery is carried out in lot sizes which leads to the assumption that for each of the different recovery steps some kind of fixed costs prevail. Furthermore, holding costs can be observed for all recovery states of the returned product. Although several authors study how the different lot sizes in a remanufacturing system shall be determined, they do not consider the specificity of the remanufacturing process itself. Thus, the disassembly operations which are always neglected in former analyses are included in this contribution as a specific recovery step. In addition, the assumption of deterministic yields (number of reworkable components obtained by disassembly) is extended in this work to study the system behavior in a stochastic environment. Three different heuristic approaches are presented for this environment that differ in their degree of sophistication. The least sophisticated method ignores yield randomness and uses the expected yield fraction as certainty equivalent. As a numerical experiment shows, this method already yields fairly good results in most of the investigated problem instances in comparison to the other heuristics which incorporate yield uncertainties. However, there exist instances for which the performance loss between the least and the most sophisticated heuristic approach amounts to more than 6%.
Keywords:
reverse logistics; remanufacturing; lot sizing; disassembly; random yieldIntroduction
The reuse field has grown significantly in the past decades due to its economical benefits and the environmental requirements. Remanufacturing which represents a sophisticated form of reuse (see, for instance, Atasu et al. [1]) focusses on valueadded recovery and has been introduced for many different products ranging from car engines (as has been reported in [2]) over photocopiers (as in [3]) to water pumps for diesel engines (as in [4]). Within the process of remanufacturing, products that are returned by the customers to the producer are disassembled to obtain functional components. The obtained components are afterwards cleaned and reworked until a "goodasnew" quality is assured. Having met the required quality standards, these components can be used for the assembly of a remanufactured product that is delivered to the customers with the same warranty as a newly produced one. In addition to the economic profitability, as a part of the embedded economic value can be saved by remanufacturing, there is an increasingly legislative restriction that assigns the producers the responsibility for their used products, for instance the Directive 2002/96/EC related to Waste Electrical and Electronic Equipment and the Directive 2002/525/EC related to End of Life Vehicles. Because of that, remanufacturing has become an important industry sector to achieve the goal of sustainable development (see, for instance, [5]). Therefore, the management and control of inventory systems that incorporate joint manufacturing and remanufacturing options has received considerable attention in recent literature contributions.
One of the main topics in these contributions is the assessment of joint lot sizing decisions for remanufacturing and manufacturing which has been thoroughly investigated in recent years. One of the first authors who established a basic modeling approach was Schrady [6] who developed a simple heuristic procedure for determining the lot sizes of repair and manufacturing lots. He assumes in his work that a constant and continuous demand for a single product has to be satisfied over an infinite planning horizon. Furthermore, a constant return fraction is established that describes the percentage of used products that return to the producer. By using that assumption a constant and continuous return rate is ensured. Presuming fixed costs for remanufacturing and manufacturing as well as different holding costs for repairable and newly manufactured products, a simple EOQtype formula (with EOQ being the economic order quantity) is proposed that minimizes the sum of fixed and holding costs per time unit. As a result, an efficient cyclic pattern is established which is characterized by the fact that within each repair cycle a number of repair lots of equal size succeed exactly one manufacturing lot. By solving the proposed EOQformula which can be applied because an infinite production and repair rate is presumed as well, the number of repair lots and the length of a repair cycle can be determined. Teunter [7] generalized the results of Schrady in a way that he examined different structures of a repair cycle. His analysis concludes that it is not efficient if more than one repair lot and more than one manufacturing lot are established in the same repair cycle. This result extends the efficient cycle patterns by a cycle in which several manufacturing lots of equal size are followed by exactly one repair lot. The assumption of equal lot sizes is among other aspects critically studied in the contribution of Minner and Lindner [8]. They show that a policy with nonidentical lot sizes can outperform a policy with identical lot sizes. However, the structure of an efficient repair cycle prevails also when the assumption of equal lot sizes is lifted.
Next to the analysis of the basic model context several extensions have been proposed that relax some of the assumptions made so far. Teunter [9], for instance, relaxes the assumption of an instantaneous manufacturing and repair process in order to derive more general expressions for the number of manufacturing and repair lots and their corresponding lot sizes. Since only a heuristic procedure was introduced on how to determine these values, Konstantaras and Papachristos [10] extended Teunter's work by developing an algorithm that leads to the optimal policy for certain parameter classes. By incorporating stochastic lead times and thereby including the possibility of backorders, Tang and Grubbstrom [11] extend the basic model. Two general options are recommended on how such a system can be dealt with, a cycle ordering model and a dual sourcing ordering policy. Both approaches are compared in a numerical study that indicates certain parameter specifications under which one approach outperforms the other. Furthermore, several papers have been published by Richter and Dobos (e.g. [12] and [13]) that relax the assumption of a constant rate of return. In their work they derived for several situations that a so called pure strategy is always optimal. In this context, a pure strategy means that either every returned product is repaired or everything is disposed of immediately. Therefore, a mixed strategy in which a part of the returned products is repaired and the rest is disposed of is always dominated by one of the pure strategies. Finally, the assumption of continuous demand and return rates has been relaxed by several authors. Consequently, the formerly EOQtype model becomes a dynamic lot sizing problem. The contribution of Teunter et al. [14] extended wellknown dynamic lot sizing heuristics such as the Silver Meal or the Part Period algorithm in order to test their performance in a remanufacturing environment. In their work, the adapted Silver Meal approach revealed an average percentage deviation of around 8% compared to the optimal solution. Schulz [15] extended among other things their approach by incorporating ideas known from the static environment and could reduce the average error to around 2%.
Common to all contributions is that they do not consider the remanufacturing process explicitly. Although some authors speak of remanufacturing, they analyze a remanufacturing system in the same way as a repair system. This may lead to wrong conclusions as it is not regarded that the remanufacturing process itself consists of two different subprocesses, a disassembly process in which the returned products are disassembled and a rework process in which the obtained components are brought to an asgoodasnew quality (for a definition see Thierry et al. [3] as well as Atasu et al. [1] for a more recent one). By explicitly incorporating both subprocesses in this contribution, the decisions that need to be made regarding disassembly and rework are decoupled which generalizes the basic models used so far.
Next to this generalization, this contribution will further relax the assumption of a deterministic yield, i.e. the number of components obtained by disassembly is not known with certainty beforehand. To present a practical application of this problem, the remanufacturing process for a car engine can be analyzed. When a batch of returned engines is disassembled, the remanufacturer does not know in advance how many remanufacturable components can be obtained. This is because the quality of the returns cannot always be assessed before disassembly. Hence, such a process can only be analyzed thoroughly when both processes disassembly and rework are evaluated separately. Considering stochastic yields has attained significant interest in the scientific literature as the basic work of Yano and Lee [16] as well as the overview of GrosfeldNir and Gerchak [17] present. However, most of the contributions presented in [17] describe purely manufacturing environments which cannot be entirely translated to a remanufacturing system as this inherits greater risks to be dealt with [18]. Nevertheless, stochastic yields have also been studied in a remanufacturing environment. Inderfurth and Langella, for instance, have concentrated their analysis specifically on the yield risk within the disassembly process [19]. Yet, they focussed on a multiproduct multicomponent problem setting in which a given discrete demand for components needs to be satisfied by either disassembling used products or manufacturing new components. The authors develop in their contribution heuristic methods on how to deal with such a problem in which they neglected the presence of fixed costs for the disassembly and the remanufacturing process. In another work, Ferrer [20] evaluates four different scenarios in a single period remanufacturing environment that differ in their process capabilities. For each scenario the optimal policy has been derived. In a numerical study, all four scenarios have been tested and compared.
After this short introduction, the problem assumptions and the nomenclature used in the remainder of the paper are illustrated in the subsequent section. Thereafter, two solution procedures are presented to find the optimal solution in the deterministic yield scenario before the scope is widened to a stochastic yield problem. For this problem setting, three heuristic approaches are introduced that facilitate the decision making process in such an environment. Next, a numerical experiment is conducted in order to test the heuristics' performance. Finally, a conclusion and an outlook are given in the last section.
Problem setting and model formulation
A company engaged in the area of remanufacturing that remanufactures several used products (e.g. engines) coming back from their customers shall be the background for the problem setting. To keep the analysis simple, the focus shall be restricted to only one specific remanufactured product named A. Figure 1 presents the general structure of this simplified system which is modeled as a multilevel inventory system containing three stages. Further simplifications are made regarding the fact that there are neither lead nor processing times. Furthermore, no disposal option is included in the problem setting.
Figure 1. Inventory system in a remanufacturing environment.
The customers' demand for the final product A is assumed to be constant and depletes the finished goods inventory continuously by a constant rate of λ units per time unit. In order to satisfy that demand, the company manufactures the final product by using component C which represents the most important component of the product. For the sake of simplicity, only the most important component C is included in the analysis. However, the proposed model could be easily extended to a multicomponent setting. The assembly process is supposed to be a flow line process at which the final product is assembled continuously and immediately delivered to the customers. When the customers have no further use for their product A (e.g. it is broken or its leasing contract ends) they have the opportunity to return the product to the company. However, only a fraction (named α) of those products in the market returns to the producer. For the subsequent analysis, the return flow of used products (which are denoted A') fills the used product inventory by the constant and continuous rate of λα. By disassembling A' the worn component C' is obtained. Although the process of disassembly typically consists of manual work, fixed costs prevail for setting up required disassembly tools and/or measuring devices that allow an improved assessment of the reusability of components before disassembly. Within this model K_{d }represents the fixed costs for a disassembly batch while h_{d }is the holding cost incurred for storing one unit of A' for one time unit. Due to different stages of wear, not all returned products contain a reworkable component C'. The ratio of the number of reworkable items obtained from the disassembly of A' to the rate of product returns λα is denoted by β. Assuming that at most one reworkable component C' can be obtained by disassembling one unit of A' the ratio β must not exceed one while being nonnegative. As the released components C' cannot be used directly for the assembly of the final product A since they usually do not meet the designated quality standards, these components have to be remanufactured. Since the remanufacturing process incurs fixed costs of K_{r }for setting up the cleaning and mechanical rework tools, a batching of reworkable components takes place as well. Hence, some reworkable components need to be stored before the next remanufacturing batch is started resulting in costs of h_{r }per unit and time unit. It is furthermore assumed that each component that is remanufactured is brought to an asgoodasnew condition. All successfully reworked components are held in a serviceables inventory at a cost of h_{s }per unit and time unit. In order to secure the final product assembly of A, some components of C have to be manufactured in addition (as α and β are usually smaller than one). The decision relevant fixed costs are denoted by K_{m }representing the cost for setting up a manufacturing lot for component C. Newly manufactured components are held in the same serviceables inventory as remanufactured ones and it is supposed that the holding costs do not differ between both sourcing options. A detailed discussion on the topic on how to set the holding cost parameters can be found in [21]. In general, the holding costs (when interpreted as costs for capital lockup) of all levels are connected by the following inequality since more value is added to the component on each level, i.e. h_{d }< h_{r }< h_{s}.
Balancing fixed and holding costs shall be achieved by applying an average cost approach to this model. This is commonly done for onelevel inventory systems as for the well known EOQmodel formulation but can be easily extended to a multilevel environment by respecting the stipulated assumptions of the EOQmodel (e.g. infinite planning horizon with constant costs over time). As a result, an optimal cyclic pattern is obtained by minimizing the average cost per time unit. In order to control the entire system, three decision variables are required. Firstly, the length of the disassembly cycle T determines the lot size of each disassembly batch (λαT ) under the assumption that there is only one disassembly lot per cycle. This assumption is made for the sake of simplicity as an additional decision variable (number of disassembly lots per cycle) would complicate the analysis significantly. However, if we consider high fixed costs of disassembly, we conjecture that this assumption of one disassembly lot per cycle assures the optimality of the introduced deterministic policy. Furthermore, by fixing the number of remanufacturing lots R per disassembly cycle, their equal lot size can be computed by λαβT/R. Finally, the number of manufacturing lots M per disassembly cycle determines the lot sizes of the manufacturing lots to be λ (1  αβ)T/M. The subsequent chapter presents the optimal solution of a completely deterministic setting in which all parameters are known with certainty.
Deterministic yields
In this section, a model is introduced that permits the evaluation of the optimal number of manufacturing and remanufacturing lots in a disassembly cycle. Before expanding the scope to stochastic yields from disassembly which represents the core issue of this contribution, the deterministic setting is studied in order to gain insight into the interrelations of the whole system. Figure 2 illustrates the behavior of the relevant inventory levels for three consecutive disassembly cycles. As a matter of fact, the optimal decision variables (T, R, and M) remain constant over time in a deterministic environment. As shown in the figure below the manufacturing lots are positioned always after the remanufacturing lots in the serviceables inventory. This is obvious as this strategy strictly dominates the strategy of starting a cycle on the serviceables level with a manufacturing lot due to the increased holding costs on the remanufacturables level.
Figure 2. Used product, remanufacturables, and serviceables inventory in a deterministic yield environment (with R = 2 and M = 1).
By minimizing the total average cost per time unit, this specific example shows the optimal cycle length T for two remanufacturing lots (R = 2) which split the remanufacturables inventory inflow equally and one manufacturing lot (M = 1) which satisfies the remaining demand of the assembly process for component C. To analyze the total cost function (TC^{D}) only two main types of costs have to be considered, the fixed costs F^{D }and the holding cost H^{D }in which the index D indicates the deterministic setting. A detailed discussion on how this formula can be obtained is presented in [22]. In addition, this contribution proves that equal lot sizes are optimal in this setting. The total cost function in the deterministic setting can be formulated as follows:
with F^{D }= K_{d }+ RK_{r }+ MK_{m }and
In order to minimize the total cost function which is a mixedinteger nonlinear optimization problem two procedures can be applied. The first procedure is a simple enumerative procedure. Since R and M need to be integer valued only a finite number of calculations (in which R and M are set to integer values) have to be compared if R and M are restricted to certain intervals. The original objective function simplifies for given values of R and M to a nonlinear convex function that only depends on T. Such a problem can be solved easily by using the subsequent equations:
The formulas presented above are comparable to the determination of the economic order interval. However, the optimality of this solution approach can only be guaranteed if the optimal total cost is determined for every combination of realization of R and M which leads to a quite large number of calculations. Nevertheless, a good solution can be obtained in a fast manner by restricting the number of possible realizations.
After introducing an enumerative procedure another promising approach will be presented next. By relaxing the original objective function (1) such that R and M need not to take on integer values, one can prove that the total cost function has only a single local minimum in the relevant area (for T, R, M > 0). Appendix C of [22] focusses on this specific aspect. Yet, by evaluating the Hessian matrix in this area, it can be shown that the total cost function is not entirely convex in all variables. This leads to the significant problem that a simple rounding procedure cannot be used to obtain the optimal solution for the integer valued number of remanufacturing and manufacturing lots. Therefore, a solution algorithm could be implemented that can globally determine the minimum cost of this mixedinteger nonlinear optimization problem. The BARON algorithm, as implemented in the GAMS software package, proved to be a valuable tool for this problem setting. In general, BARON implements deterministic global optimization algorithms of the branchandreduce type in order to determine the optimal solution for a mixedinteger nonlinear optimization problem. For a detailed description of the algorithm please refer to [23].
The subsequent chapter extends the deterministic model of this section to incorporate stochastic yields.
Stochastic yields
One of the main problems for many practical applications in the area of remanufacturing is that they have to deal with stochastic yields which means that the amount of remanufacturable components obtained from disassembling used returned products is not known with certainty (see also [19]). Due to the significance of that problem in a remanufacturing planning environment, we will now put forth the extension of the deterministic model that was introduced in the last section to incorporate stochastic yield fractions resulting from the disassembly process. Although being uncertain, it can be assured that the lowest possible yield fraction β_{l }cannot be smaller than zero as negative yields would not be reasonable. The largest possible yield fraction β_{u}, however, cannot exceed the value of one since this describes the situation that from every disassembled used product more than one remanufacturable component is obtained which is ruled out by the assumptions made. Within the range from β_{l }to β_{u }a specific distribution function can be defined which will be denoted in the following analysis by φ(β ). As the number of returned products disassembled per cycle corresponds to λαT, the independence of φ(β ) with respect to T reflects the fact that the subsequent analysis assumes stochastic proportional yields (for a definition see [16]). Therefore, the formerly used total cost function for a deterministic yield scenario (formula (1)) has to be extended in order to incorporate any possible yield outcome. Hence, the total cost of a given stochastic yield scenario can only be formulated as an expected total cost (which will be further on denoted as TC^{S}) that is presented in the following equation:
with and
The fact that for any possible yield realization β an integer number of R and M has to be defined complicates the analysis of the total cost function TC^{S }significantly. In this setting, R(β) describes the optimal number of remanufacturing lots for a given yield fraction β. Likewise, M(β) represents the optimal number of manufacturing lots if the yield fraction β is fixed. Due to the fact that β is not known with certainty the total cost per time unit can only be formulated as an expectation over all different yield realizations. In contrast to the total cost function of the deterministic case (1), F^{S }and H^{S }can be regarded as an expectation of their corresponding deterministic equivalents F ^{D }and H^{D}. As finding the optimal solution for any problem setting cannot be guaranteed, which will be shown later in this chapter, three different heuristic policies will be presented in the succeeding paragraphs that differ in their degree of sophistication. The first and least complex policy is introduced in the following:
Policy I
The easiest option on how to handle a stochastic problem is to neglect the underlying stochastics in order to derive a deterministic equivalent of the stochastic problem. The first policy introduced proceeds exactly in this manner as it neglects the fact that R and M depend on the yield realization β. Thus, only one value for R and M needs to be derived that is valid for every yield realization between β_{l }and β_{u}. To obtain these values, one can insert a specific yield fraction into the deterministic total cost function of the last section (1) and apply the recommended solution procedures to obtain R and M. As any yield fraction can be inserted that lies in the range of possible yield realizations and therefore many different combinations of R and M may prevail, we limit the focus of policy I on inserting only the mean yield fraction into the deterministic model since the mean yield is one of the most important characteristics of the underlying yield distribution. As a result we obtain the values of R^{D }and M^{D }that replace R(β ) and M(β ) for every possible yield realization β in formula (4). The expected total cost of the first policy (TC_{I }) can therefore be easily calculated by the subsequent equation:
Since policy I is a very simple approach, the decision maker can improve the expected total cost by incorporating the underlying stochastics in the decision making process which is introduced in policy II.
Policy II
Contrary to the first policy, the second policy does not neglect the dependence of R and M on the realization of the random yield fraction β. Nevertheless, in order to keep this policy simple, the disassembly cycle length T is kept constant which reduces the policies' complexity significantly. For the sake of simplicity the length of the disassembly cycle T will be set to the optimal deterministic cycle length obtained by formula (2) assuming that the mean yield fraction has been inserted as the deterministic equivalent for the underlying yield distribution. The assumption of fixing the cycle length to a specific value can be further used to draw some basic conclusions that can only be drawn for a given cycle length. The stochastic yield realization β determines for each disassembly cycle the number of remanufacturable items. As the number of remanufacturable and manufactured components per cycle always adds up to the value of λT, the number of manufactured items depends as well on the yield realization. However, for both options of demand fulfillment it can be observed that if more items are processed (either by manufacturing or remanufacturing) the number of respective lots in a cycle does not decrease. Therefore, when comparing two different yield realizations with all other parameters set equally it can be said: For the larger yield realization the number of remanufacturable components increases which means that the number of remanufacturing lots per cycle does not decrease. On the other hand, the number of newly manufactured components decreases with larger yield realizations which means that the number of manufacturing lots per cycle does not increase. Figures 3 and 4 compare both heuristic policies introduced so far for three consecutive disassembly cycles. On the left hand side (Figure 3), it can be observed for policy I that regardless of the realized yield fraction the same number of R and M is applied in every cycle (R = 2 and M = 1). Figure 4 on the right hand side shows policy II that reacts for the same cycle length T on the different realizations of β which is supported by the fact that for a small yield realization the number of remanufacturing lots is smaller than for a large yield realization (R = 1 in the first cycle compared to R = 3 in the third cycle). An opposing behavior can be observed for the number of manufacturing lots per cycle that does not increase the smaller the yield realization is.
Figure 3. Inventory system in a stochastic yield environment applying policy I (R = 2 and M = 1).
Figure 4. Inventory system in a stochastic yield environment applying policy II.
These general conclusions cannot only be formulated verbally but also in a mathematical form by introducing socalled transition yield fractions which have the property that either the number of remanufacturing lots or the number of manufacturing lots changes when optimizing the deterministic equivalent problem. For the calculation of the specific yield fraction that is characterized by a switch of the optimal policy from R to R+1 remanufacturing lots, one needs to equate the deterministic total cost functions for R and R+1 as presented in the following equation:
with F^{D}(R) = K_{d }+ RK_{r }+ M K_{m }and
This equation can be solved with respect to β in order to obtain the transition yield fraction β (R) at which the optimal decision in the deterministic case switches from R to R+1 for a given cycle length T :
Not only is this function monotonously increasing in R which corresponds to the findings above that the number of remanufacturing lots does not decrease for larger values of β but it also does not depend on the number of manufacturing lots per cycle M. Thus, the same analysis can be carried out independently for the transition from M to M1 manufacturing lots per cycle by equating both total cost functions in order to obtain the transition yield fraction β (M):
Because this function is monotonously decreasing in M, the insight that a larger yield fraction does not lead to less manufacturing lots in a cycle is approved. Consequently, the lowest and highest values for R and M can be determined by exploiting the two formulas given above. Thus, for the lowest possible yield fraction β_{l }R_{min }and M_{max }can be computed by the following procedure (analogously R_{max }and M_{min }can be computed for the highest possible yield fraction β_{u}):
As the disassembly cycle length is fixed to a given value, the distribution function of the stochastic yield fraction can be subdivided into several intervals. Each interval j contains all yield realizations between its lower bound l_{j }and its upper bound u_{j}. The main characteristic of such an interval is the fact that within this interval only one number of remanufacturing and manufacturing lots induces the optimal solution for any possible yield fraction within this interval. The optimal number of remanufacturing and manufacturing lots in a certain interval j are furthermore denoted by R_{j }and M_{j}, respectively. For the identification of the respective interval bounds the following pseudocode can be used:
Start j = 1, l_{j }= β_{l}, R_{j }= R_{min}, M_{j }= M_{max}, β(0) = ∞
While min{β(R_{j }+ 1), β(M_{j } 1)} < β_{u }do
if β(R_{j }+ 1) < β(M_{j } 1) then
u_{j }= β(R_{j }+ 1)
j = j + 1, l_{j }= u_{j1}, R_{j }= R_{j1 }+ 1, M_{j }= M_{j1}
else
u_{j }= β(M_{j } 1)
j = j + 1, l_{j }= u_{j1}, R_{j }= R_{j1}, M_{j }= M_{j1 } 1
end if
end do
u_{j }= β_{u}, J = j
end
After the initialization in which the first interval j = 1 is opened (l_{1}=β_{l}) and given the values R_{min }and M_{max }the procedure evaluates if the transition to R_{min}+1 or M_{max}1 is closer to β_{l}. For the lower of these two values, the upper bound of the first interval u_{1 }is fixed to the transition rate and the next interval is opened (l_{2 }= u_{1}). This procedure stops when both next transitions to R+1 and M1 are larger than the highest possible yield fraction β_{u}. At this point the total number of intervals into which the yield distribution can be separated is determined by the index j which is set to the number of intervals J. As a result, the total yield distribution is separated into several intervals which is depicted for an example in Figure 5. In this example (with β_{l }= 0 and β_{u }= 1) it can be observed that the solution of policy I would have been R = 3 and M = 4 as this would solve the deterministic equivalent to optimality for β = 0.5.
Figure 5. Exemplary separation of a yield distribution according to policy II.
As the interval bounds vary with a changing disassembly cycle length T, the expected total cost function for policy II can be formulated as follows using the optimal disassembly cycle length T^{D* }obtained by inserting the mean yield fraction into equation (3):
with and
In comparison to formula (4) only a finite number of R and M have to be considered in order to determine the solution using policy II. The formerly required R(β ) which represents the optimal number of remanufacturing lots for any given yield fraction β can be replaced with R_{j }after separating the yield distribution into intervals in which only one R is optimal for each yield realization. Consequently, the same simplification holds for the number of manufacturing lots M. However, this solution can be further improved by varying the disassembly cycle length T which shall be done in the most sophisticated heuristic approach of this contribution.
Policy III
As the convexity of the expected total cost function of policy II (9) regarding the only remaining variable T cannot be proven for any possible yield distribution we face the fact that obtaining the optimal solution for this system cannot be guaranteed. Nevertheless, a simple local search heuristic can be implemented that alters the disassembly cycle length T from its initial value of policy II in order to check whether the expected total cost increases or decreases. The expected total cost function is evaluated by applying the procedure of policy II for any chosen parameter T. The local search procedure stops when both an increase or a decrease of T results in an increasing expected total cost meaning that at least a local minimum has been found that improves the solution of policy II at the expense of an increased complexity.
The following chapter elaborates a numerical experiment in which all three introduced heuristic approaches are tested in order to evaluate their performance in a stochastic yield environment.
Numerical experiment
The main objective of the numerical experiment conducted in this section is to evaluate the error that can be made when the simplest approach (policy I) is used compared to the more complex ones (policies II and III). In order to estimate the error, several numerical tests have been conducted using randomly generated instances. To our knowledge, no scientific contribution contains reliable and complete real life data for this specific problem setting. As the number of adequate test instances cannot be guaranteed in this case, Rardin and Uzsoy [24] recommend to create an experimental design based on random test instances. Although they discuss the pitfalls of random test instances in detail, we have applied this procedure to provide a first insight into each policy's performance. All parameters required for the test instances were drawn from a discrete uniform distribution DU (a, b) with a as the lower bound and b representing the upper bound of the distribution. Some parameters were multiplied after the random draw with a constant term in order to obtain reasonable values. Table 1 lists all parameters that were randomly drawn in this experiment:
Table 1. Parameters generated randomly in numerical experiment
The return fraction α, for instance, can take on values between 30% and 90%, only limited by the fact that the percentage must be an integer multiple of 5%. Regarding the fixed costs, we restricted the possible region on integer values between 0 and 50 for the disassembly process and 1 to 100 for setting up a remanufacturing or a manufacturing lot. For the disassembly lot, we established smaller values as these processes are done manually in some industrial applications and do not necessarily require a specific setup. With respect to the holding costs we implicitly assumed that the holding cost increase from level to level as more effort has been put into the components. This means that each randomly generated instance has to fulfill the presumed inequality h_{d }< h_{r }< h_{s}. From these probability distributions, 1,000 instances were drawn and tested for different yield distributions. Generally, the yield distribution followed a symmetric betadistribution within the limits β_{l }= 0 and β_{u }= 1. The parameter that altered the yield distribution was the coefficient of variation ρ that was changed in the limits between 0.05 and 0.55 which is motivated by our experience with an automotive remanufacturer regarding its yield fractions. While a ρ of 0.05 indicates that almost the entire probability mass is centered around the distribution's mean, a coefficient of variation of 0.55 indicates for a betadistribution within the interval 0 to 1 an approximately uniform yield distribution.
All three introduced heuristic approaches were tested for all instances. Figure 6 illustrates, for instance, the percentage deviation of the total costs of policies I and II. Δ_{I>II }denotes this percentage deviation and is calculated by Δ_{I>II }=TC_{I }/TC_{II }1. In detail, this deviation shows the percentage loss in performance if policy I (at which only the mean yield fraction is considered to represent the entire yield distribution) is applied instead of policy II. The deviation with respect to the coefficient of variation of the underlying yield distribution is presented with the aid of box plots that do not only show the maximum and minimum deviation but also where half of the deviations are located inside the shaded area around the median.
Figure 6. Percentage deviation of policy I compared to policy II.
For very small coefficients of variation that are characterized by the fact that almost the entire probability mass is centered around the mean yield, the deviation between policy I and policy II is almost negligible. The reason for that is easy to be found. Although the yield distribution is defined in the interval between 0 and 1, the range of realizations that have a significant probability is very small. If the optimal number of remanufacturing and manufacturing lots per cycle that is determined by policy I is also optimal for a wide range of yield fractions around the distribution's mean both policies arrive at the same result. However, if the coefficient of variation grows larger the deviations increase as well. For an approximately uniformly distributed yield, for instance, the maximum deviation between policy I and II is around 5.4%. On the other hand, the minimum deviation is 0% which means that the optimal cycle pattern of policy I is still optimal for every yield realization between 0 and 1 even for such a widespread distribution. Although many instances have been tested, the effect of every parameter on the deviation cannot be observed without doubt. Yet, some general trends can be derived from the experiments. For instance, it seems to be the case that the percentage gap in the total cost between policy I and II increases in most scenarios for instances with an increasing return rate α. Additionally, the different fixed costs seem to influence this gap as well. For high fixed costs for disassembly and remanufacturing (K_{d }and K_{r}) as well as for small fixed costs for manufacturing (K_{m}) the observed percentage gap increases for a large coefficient of variation of the yield distribution. The same analysis can be conducted for the different holding cost parameters, too. The percentage gap between policy I and II increases if the holding costs h_{d}, h_{r}, and h_{s }deviate significantly. Furthermore, it can be said as larger the difference between R_{min }and R_{max }as well as the difference between M_{min }and M_{max }is as larger is the percentage gap. Finally, no considerable influence on the percentage gap can be observed for the demand per time unit λ.
Figure 7 presents the deviation of policy II from policy III which means that the cycle length T is varied in order to decrease the total cost function even further. By Δ_{II>III }this deviation is represented. Regarding the coefficient s of variation the same can be observed as for the first examined deviation. For small coefficients of variation there is almost no improvement possible by changing the cycle length. On the other hand, for larger coefficient s the percentage gap grows larger which means that an adaption of T can improve the total cost function. However, these improvements are relatively small (in 97.4% of all cases smaller than 1% for ρ = 0.55). Regarding the cost deviation between policy II and III, it is even more difficult (in comparison to the deviation between policy I and II) to define parameter areas at which the deviation is typically high or low. Yet, two general trends can be noticed. The largest deviations can be observed for instances with a large α and a wide spread of the holding cost levels. However, this observation cannot be generalized for all instances with this parameter constellation.
Figure 7. Percentage deviation of policy II compared to policy III.
Another interesting question that can be analyzed with this numerical experiment is whether the optimal cycle length increases or decreases in comparison to the cycle length of policy I and II that remains constant for all coefficients of variation. In 69.1% of all instances the cycle length decreased while it increased in the remaining 31.9%.
Therefore, no general conclusion can be drawn regarding this aspect as no specific parameter constellation can be identified that increases or decreases the cycle length in general.
Conclusion and outlook
This contribution outlined an approach on how to handle deterministic and stochastic yield fractions within a multilevel remanufacturing system that considers the disassembly process explicitly. While being restricted to a single disassembly lot per cycle, simple derivations are made with respect to the three necessary parameters, the optimal disassembly cycle length as well as the optimal number of remanufacturing and manufacturing lots per disassembly cycle. By examining both the stochastic and the deterministic case, the error that can be made by neglecting the underlying stochastics is evaluated. The numerical experiment in section 5 has confirmed a quite straightforward assumption. The less variability of the random yield fraction is faced, the smaller is the error that is made by using the mean yield policy I instead of the more sophisticated ones. However, there exist situations in which using the simple policy I results in performance losses of more than 5%. Nevertheless, in most cases the decision maker will obtain fairly good results if he neglects the underlying yield distribution and follows the deterministic mean yield fraction approach of policy I. In this sense, a problem setting has been identified in which the influence of stochastic yields does not complicate the decisions to be made significantly.
Next, an outlook regarding future research efforts shall be given. The proposed model can be extended in several ways. For both the stochastic and the deterministic one, the option of allowing more than one disassembly lot per disassembly cycle is a promising extension of the model presented in this contribution. Especially for instances showing a small fixed cost of disassembly this might provide a valuable option to decrease the average costs per time unit. Furthermore, it can be studied how a multiproduct multicomponent setting affects the decision making process in both environments since aspects like multiplicity (one component can be obtained by the disassembly of different product types) have to be incorporated. Another interesting topic that can be included in the analysis is a disposal option. This might be a worthwhile option if the fixed costs of remanufacturing are quite high and the yield realization is very small. In the proposed model context, at least one remanufacturing process has to be set up in such a disassembly cycle. However, if there is a disposal option, the obtained components can be disposed of and the total customer demand will be satisfied by newly manufactured components, i.e. the optimal number of remanufacturing lots R can be 0. As a last extension, all heuristic approaches can be tested not only for proportional stochastic yields but also for nonproportional yields. In order to achieve this objective, the yield fraction distribution cannot be modeled as a betadistribution any more but needs to be modeled for instance with a binomial distribution.
As discussed above, a number of different uncertainties can be found in a real life remanufacturing system. This contribution has revealed that a possible yield uncertainty can be neglected in a multitude of problem instances when considering joint remanufacturing and manufacturing lot sizing decisions. This is a rather untypical result when real life industrial applications such as the remanufacturing of car engines face stochastic yields in their process. To obtain good solutions for the lot sizing problem presented above, the only required information regarding the yield distribution is its mean value. By planning the lot sizes with this mean value, the error of neglecting stochastic yields can be reduced in most cases to less than 2%.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
TS elaborated the deterministic and stochastic solutions and carried out the numerical study. IF defined the problem setting, participated in the elaboration of the general method and validated the results by implementing a simulation model. All authors read and approved the final manuscript.
Acknowledgements
The authors wish to express their thanks to the Ministry of Education, University and Research of Italy (MIUR) for the support given for the development of this work under the Interlink program (Supply Chain Sustainability project). Furthermore, the authors wish to express their thanks to the helpful comments of the referees.
References

Atasu A, Guide VDR Jr, van Wassenhove LN: So what if remanufacturing cannibalizes my new product sales?

Seitz M, Wells PE: Challenging the implementation of corporate sustainability: The case of automotive engine remanufacturing.
Business Process Management Journal 2006, 12(6):822836. Publisher Full Text

Thierry M, Salomon M, van Nunen J, van Wassenhove LN: Strategic issues in product recovery management.

Tang O, Teunter R: Economic lot scheduling problem with returns.

Webster S, Mitra S: Competitive strategy in remanufacturing and the impact of takeback laws.
Journal of Operations Management 2007, 25(6):11231140. Publisher Full Text

Schrady DA: A deterministic inventory model for reparable items.
Naval Research Logistics Quarterly 1967, 14:391398. Publisher Full Text

Teunter RH: Economic ordering quantities for recoverable item inventory systems.
Naval Research Logistics 2001, 48(6):484495. Publisher Full Text

Minner S, Lindner G: Lot sizing decisions in product recovery management. In Reverse Logistics  Quantitative models for closedloop supply chains. Edited by Dekker R, Fleischmann M, Inderfurth K, van Wassenhove LN. Springer; 2004:157179.

Teunter RH: Lotsizing for inventory systems with product recovery.
Computers & Industrial Engineering 2004, 46(3):431441. PubMed Abstract  Publisher Full Text

Konstantaras I, Papachristos S: A note on: Developing an exact solution for an inventory system with product recovery.
International Journal of Production Economics 2008, 111(2):707712. Publisher Full Text

Tang O, Grubbstrom RW: Considering stochastic lead times in a manufacturing/remanufacturing system with deterministic demands and returns.
International Journal of Production Economics 2005, 9394:285300.

Richter K: The extended EOQ repair and waste disposal model.
International Journal of Production Economics 1996, 45(13):443448. Publisher Full Text

Richter K, Dobos I: Analysis of the EOQ repair and waste disposal problem with integer setup numbers.
International Journal of Production Economics 1999, 59(1):463467. Publisher Full Text

Teunter RH, Bayindir ZP, van den Heuvel W: Dynamic lot sizing with product returns and remanufacturing.
International Journal of Production Research 2006, 44(20):43774400. Publisher Full Text

Schulz T: A new silvermeal based heuristic for the singleitem dynamic lot sizing problem with returns and remanufacturing.
International Journal of Production Research 2011, 49(9):25192533. Publisher Full Text

Yano CA, Lee HL: Lot sizing with random yields.
Operations Research 1995, 43(2):311334. Publisher Full Text

GrosfeldNir A, Gerchak Y: Multiple lotsizing in production to order with random yields: Review of recent advances.

Inderfurth K, Langella IM: Heuristics for solving disassembletoorder problems with stochastic yields.
OR Spectrum 2006, 28(1):7399. Publisher Full Text

Ferrer G: Yield information and supplier responsiveness in remanufacturing operations.
European Journal of Operational Research 2003, 149(3):540556. Publisher Full Text

Teunter RH, van der Laan E, Inderfurth K: How to set the holding cost rates in average cost inventory models with reverse logistics.
Omega 2000, 28(4):409415. Publisher Full Text

Schulz T, Ferretti I: On the alignment of lot sizing decisions in a remanufacturing system in the presence of random yield. Technical report, Faculty of Economics and Management, OttovonGuericke University Magdeburg; 2008.

Sahinidis NV, Tawarmalani M: Accelerating branchandbound through a modeling language construct for relaxationspecific constraints.
Journal of Global Optimization 2005, 32(2):259280. Publisher Full Text

Rardin RL, Uzsoy R: Experimental evaluation of heuristic optimization algorithms: A tutorial.
Journal of Heuristics 2001, 7:261304. Publisher Full Text