Академический Документы
Профессиональный Документы
Культура Документы
Omega
journal homepage: www.elsevier.com/locate/omega
art ic l e i nf o a b s t r a c t
Article history: An effective planning of lot sizes is a key strategy to efficiently manage a combined manufacturing/
Received 23 November 2012 remanufacturing system in the presence of substantial setup costs. Due to its complex interdependen-
Accepted 24 September 2013 cies, optimal policies and solutions have not been identified so far, but several heuristic approaches have
Available online 3 October 2013
been analyzed in recent contributions. The main heuristic shortcuts are forcing equally sized lot sizes
Keywords: over the planning horizon as well as imposing a specific cycle structure, i.e., a sequence of manufacturing
Inventory management batches is followed by a sequence of remanufacturing batches. We are instead proposing a flexibly
Remanufacturing structured heuristic that allows for differently sized remanufacturing batches. We show in a compre-
Lot sizing hensive numerical study that our approach outperforms other existing approaches in more than half of
EOQ
all instances by up to 17%.
& 2013 Elsevier Ltd. All rights reserved.
0305-0483/$ - see front matter & 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.omega.2013.09.003
22 T. Schulz, G. Voigt / Omega 44 (2014) 21–31
contrast to Schrady's idea, only a single remanufacturing batch is organized as follows. Section 2 introduces the basic modeling
issued per cycle that is succeeded by M equal manufacturing batches assumptions and outlines the solution finding process of the
before the next identical cycle commences. Konstantaras and Skouri above-mentioned policy structures. In Section 3, we introduce
[9] derive sufficient conditions to determine which class of policy, our flexibly structured ðR; MÞflex heuristic which we extensively test
ðR; 1Þ or ð1; MÞ, is optimal for a specific problem instance. in a numerical study with the experimental design of Choi et al.
Next to formulating the (1, M)-policy, Teunter names two [4]. The results of this study can be found in Section 4. Finally,
options to further improve the solution. At first, he proposes that Section 5 presents a short summary and an outlook on future
a more general (R,M)-policy could reduce the total cost per time research opportunities.
unit. This idea has been elaborated by Choi et al. [4] who define
such a policy structure. They prove that there is only one cost
minimizing sequence of equally sized batches in a cycle for any 2. Basic modeling assumptions and current solution
given (R,M) combination. Moreover, they present an algorithmic approaches
procedure to determine the minimum total cost for this policy
structure. In comparison to the underlying approach of this 2.1. Basic modeling assumptions
contribution, however, they do not allow for variable remanufac-
turing lot sizes over time. We show that the performance can be Since a remanufacturing system contains a large number of
considerably increased if this restriction is lifted. Choi et al. [4] different planning tasks, we restrict our attention to a simplified
procedure has been facilitated later by Liu et al. [11] using a model setting that focuses on analyzing a smaller subset of
slightly different experimental design as testing environment, problems in greater detail. In general, one possibility to model a
however, both contributions derive similar results. The best solu- remanufacturing system is to describe its relevant processes and
tion of the ðR; 1Þ and ð1; MÞ policies is only seldom improved by stocking points. Due to their importance, there are many options
using the (R,M)-policy (in about 0.2% of all instances examined). to illustrate the existing interdependencies of the corresponding
Moreover, the actual improvement is also small (always less inventory levels and processes. After conducting a thorough
than 0.5%). literature review, Akçalı and Çetinkaya [2] elaborate 14 different
The second possible improvement option mentioned by Teunter settings to model these interdependencies that can be found in the
[19] is to allow for scheduling differently sized (re)manufacturing literature. Among these settings, one (named 2SP-c) seems to be of
batches in a cycle. This idea has been analyzed at first by Minner special interest as it has been applied in a large number of
and Lindner [12] who show with a Lagrange-multiplier approach scientific contributions. In our work, the basic modeling approach
that initiating equal remanufacturing batches in a cycle does not coincides with their 2SP-c setting. Its relevant stocking points and
have to be optimal. Yet, these authors do not evaluate the actual processes are presented in Fig. 1.
benefit of scheduling differently sized remanufacturing batches. In this simplified model, a remanufacturing firm faces a
Feng and Viswanathan [5] contribute to the discussion by constant and continuous demand (denoted by λ) for a single
considering an (R,M) type policy with differently sized remanu- product. To obtain remanufacturable products, the remanufacturer
facturing batches. As in our approach, a sequence of manufactur- takes back products from his customers when they have no further
ing batches is followed by one remanufacturing batch in each use for it. We assume that only a fraction (denoted by α) of the
subcycle. However, for facilitating the solution finding process, entire customer demand returns to the remanufacturer who keeps
they allow for only two classes of subcycles. In contrast, our all returns in a corresponding used product inventory (at a given
approach is more flexible by not limiting the number of classes holding cost hR per item and time unit).
to two. Thus, we are pushing the idea of the subcycle approach After collecting some returns, the remanufacturer issues a
even further and allow for a flexible sequence of manufacturing/ remanufacturing batch to recover these returns which brings
remanufacturing batches while the size of the remanufacturing them to an as-good-as-new condition. In this work, we omit
batch may vary for every single setup. While Feng and Viswa- different quality levels of remanufactured products and refer
nathan [5] conclude that the total costs can only be slightly interested readers to Mitra [13] for a more detailed analysis on
decreased with two subcycle classes, we find that introducing an this subject. Each remanufacturing run necessitates a specific
even more flexible structure of manufacturing/remanufacturing setup (that incurs a cost of KR), for instance to adjust the required
sequences can decrease costs substantially by up to 17.5% when tools. All successfully remanufactured products are stored in a final
allowing for differently sized remanufacturing batches compared product inventory (at a given holding cost hF per item and time
to the best (R,M)-policy with equally sized (re)manufacturing unit) from which the customers receive their orders.
batches. When interpreting the holding cost hR and hF as cost of capital
The main objective of this contribution is to comprehensively tied up in inventory, the latter is always larger since more effort
analyze the benefits of scheduling differently sized (re)manufac- has been put into a final product than into a remanufactured
turing batches within a flexible cycle structure. We will show that one. A detailed discussion on the topic on how to set the holding
different batch sizes can reduce the total cost substantially for a cost parameters in a remanufacturing environment can be
large number of problem instances. The remainder of this work is found in Teunter et al. [20]. Since α is typically smaller than 1,
Table 2 Table 3
Setup and holding cost terms for the (R, 1), (1, M), and (R, M) policies. Example parameters.
Policy SC HC 1 λ α KR KM hR hF
2
(R, 1) RK R þ K M 1 1 α 1000 50% 200 900 1 2
λ 1þ α 1 αhR þ þ ð1 αÞ2 hF
2 R R
" ! #
(1, M) K R þ MK M 1 ð1 αÞ2
λ αhR þ α þ2
hF
2 M
" ! #
ðR; MÞeq RK R þ MK M 1 ð1 αÞðR 1Þ 1 α2 ð1 αÞ2 remanufacturing batch is not sufficient to satisfy the subcycle's
λ þ αhR þ þ hF
2 RM R R M demand, a certain amount of components (denoted by ΘM;s ) have
to be manufactured in a number of νM;s equally sized manufactur-
ing batches. Each manufacturing batch in a subcycle should be of
batches persists (i.e., both sides adjacent to the right angle in the equal size since deviating from this would increase the holding
well-known holding cost triangles are doubled resulting in four cost incurred (see Appendix C). The individual lot size of a
times larger inventories). manufacturing batch Q M;s is therefore determined by ΘM;s =νM;s .
Lemma 1 allows establishing a total cost function that is However, it is possible that no new component is fabricated in a
independent of the respective cycle length T for a given sequence subcycle, i.e., ΘM;s ¼ 0. To summarize, each subcycle contains
of batches. By defining HC 1 as the total holding cost for a cycle exactly one remanufacturing batch and zero, one, or more man-
length of one time unit (T ¼ 1), the holding cost HC can be ufacturing batches.
reformulated as HC ¼ HC 1 T 2 . Inserting this general condition As in the other heuristic approaches discussed in the literature
into (1), optimizing it with respect to T and inserting the optimal so far, the number of remanufacturing and manufacturing batches
cycle length T n into TC yield the optimal total cost per time unit R and M has to be fixed in advance to obtain the solution of the
TC n as ðR; MÞflex -policy. The best policy is then chosen by enumerating
over R and M. Thus, after fixing R and M, the following three steps
SC þ HC 1 T 2
TC ¼ ð2Þ have to be taken to compute the total cost per time unit. We
T
illustrate each step for a ð3; 2Þflex -policy with respect to the
sffiffiffiffiffiffiffiffiffi parameters summarized in Table 3. The development of the
SC
Tn ¼ ð3Þ inventory positions for this example is presented in Fig. 2.
HC 1
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
TC n ¼ 2 SC HC 1 : ð4Þ 3.1. Step 1: determine the batch sequence
This general dependency can be used to eliminate the cycle
length as a variable from the total cost function.1 Table 2 sum- When both the number of remanufacturing and manufactur-
marizes the relevant setup and holding cost terms for the ing batches in a cycle are larger than one at the same time, a
respective policy type. We refer to Appendix A for a derivation number of different batch sequences can be observed. This
of the cost terms along with numerical examples. In the (M, 1)- problem prevails for the ðR; MÞflex -policy as well. Choi et al. [4]
policy or (1, M)-policy, there is only one variable remaining that have shown that there is only one optimal batch sequence for the
can be easily obtained by differential calculus. For a predetermined case of equal remanufacturing and equal manufacturing batches.
sequence of batches in a (R, M)-policy as the ðR; MÞeq -policy, we can To determine this sequence, Choi et al. introduce the so-called
also establish a total cost function that is independent of the cycle SAIL algorithm. In the following, we apply their SAIL algorithm to
length. This considerably simplifies finding optimal policies by determine the batch sequence for the ðR; MÞflex -policy. In order to
enumerating over only R and M. present our ðR; MÞflex -policy as a whole, we recap the approach.
The SAIL algorithm assumes both equally sized remanufacturing
and manufacturing batches. As the cycle length is set to one time
3. The ðR; MÞflex -policy unit to determine HC 1 (i.e., λα products return to the remanu-
facturer), exactly λα=R items are remanufactured in each batch.
Common to all heuristic policy structures introduced so far is The basic idea of the algorithm is to schedule a remanufacturing
that they fix a number of decisions in advance to facilitate the batch if at least λα=R returns are available in the used product
solution finding process. For instance, the decisions regarding the inventory (denoted by yR ) whenever the final product inventory
batch sequence and lot sizes are always predefined when the is depleted. The result of the algorithm is the sequence of batches
decision maker chooses one of these policies. In the following, we in a cycle that is represented by the set of γ i variables (with
present an alternative approach that determines the optimal cycle i ¼ 1; …; R þ M), where γ i ¼ 1 when a remanufacturing batch is
structure when only the number of (re)manufacturing batches is initiated and γ i ¼ 2 when a manufacturing batch is initiated. To
fixed in advance but the exact structure is flexible. determine all γ i the following pseudocode of the SAIL algorithm
In our flexibly structured heuristic (denoted with the super- can be applied (see [4]):
script flex), the concept of subcycle-oriented optimization is
employed. In this concept, the whole cycle is separated into R start γ 1 ¼ 1, yR ¼ λα2 =R
subcycles (denoted by s ¼ 1; …; R) in which the following pre- for i ¼ 2 to R þ M
sumptions are required to hold. At the beginning of each subcycle, if yR Zλα=R then γ i ¼ 1, yR ¼ yR λα=R þ λα2 =R
the sole remanufacturing batch is initiated. It contains exactly Q R;s else γ i ¼ 2, yR ¼ yR þ λð1 αÞ=M α
items that are remanufactured at once. If the lot size of the end if
end for
1
We thank an anonymous referee who pointed out that this approach is not
necessarily viable for the optimal (R, M)-policy. Since we do not know the optimal
(R, M)-policy structure, we cannot formally proof that the sequence and relative The first batch of a cycle is fixed to be a remanufacturing batch
scheduling of batches do not change with different cycle lengths. comprising λα=R items. After α=R time units, customer demand
T. Schulz, G. Voigt / Omega 44 (2014) 21–31 25
cannot be satisfied from this batch anymore and a new batch has subsequent pseudocode describes a procedure to determine all
to be scheduled. While the first remanufacturing batch satisfied νM;s values:
demand, λα products return per time unit which means that the
inventory contains λα2 =R returns at this point in time. This amount start i ¼ 1, s ¼ 0
is not sufficient to schedule a remanufacturing batch as α o1 and, while i r R þM
thus, the second batch in a cycle is always a manufacturing batch if γ i ¼ 1 then s ¼ s þ1, νM;s ¼ 0
containing λð1 αÞ=M items. Thereafter, the algorithm checks if at else νM;s ¼ νM;s þ 1
least λα=R returns are available as soon as the second batch cannot end if
satisfy demand any more. In this case, a remanufacturing batch is i ¼ i þ1
initiated. Otherwise, a manufacturing batch is set up. The algorithm end while
stops after all batches have been scheduled. After determining the
batch sequence, the second step computes the corresponding batch
sizes. The heuristic commences with the first remanufacturing batch
of a cycle and counts the number of adjacent manufacturing
Example. We present in the following the SAIL algorithm intro- batches before the next remanufacturing batch is initiated.
duced by Choi et al. [4] for the numerical example. The develop- After computing all νM;s values, the amount of manufactured
ment of the inventory positions is depicted for this policy products in subcycle s (denoted by ΘM;s ) can be computed by
in Appendix A.3, Fig. A3. The first batch is by definition a (ΘM;s ¼ νM;s λ ð1 αÞ=M).
remanufacturing batch comprising (1000 x 5)/3 = 166.67 items, By definition, all returns collected in subcycle s are remanu-
i.e., γ 1 ¼ 1. Since only 166.67 0.5 ¼ 83:33 units return during factured at the beginning of the following subcycle. Two factors
the first remanufacturing phase, and this amount is not sufficient determine the size of each subcycle's remanufacturing batch Q R;s :
to schedule another remanufacturing batch comprising 166:67 the number of used products remanufactured at the beginning of
units, a manufacturing batch has to follow (γ 2 ¼ 2). This manufac- the preceding subcycle s 1 as well as the number of manufac-
turing batch contains (1000 x 5)/2 = 250 units. During this tured items in s 1. In general, Q R;s can be computed by
manufacturing phase 250 0:5 ¼ 125 units are returning to Q R;s ¼ ðQ R;s 1 þΘM;s 1 Þ α 8 s ¼ 2…R: ð5Þ
the used product inventory. Thus, the inventory position in the
used product inventory amounts to 83:33 þ 125 ¼ 208:33. Since Due to the cyclical structure, the first subcycle is preceded by the last
208:33 Z 166:67, the next remanufacturing batch can be sched- subcycle of the previous identical cycle. Thus, Q R;1 is obtained by
uled (γ 3 ¼ 1). When the remanufacturing batch is initiated, the Q R;1 ¼ ðQ R;R þ ΘM;R Þ α: ð6Þ
used product inventory drops to 41:67, and 166:67 0:5 ¼ 83:33
units are returning to the used product inventory during the Hence, a system of R linear equations (see Appendix A.4) results that
remanufacturing phase. Thus, after that remanufacturing phase contains altogether R variables Q R;1 to Q R;R . Solving this system of
there are 41:67 þ 83:33 ¼ 125 units in the used product inventory. equations leads to a closed-form expression for each subcycle's
Again, this amount is not sufficient to schedule a remanufacturing remanufacturing batch size as in Appendix A.4. We get
lot (166:67 Z 125), and a manufacturing batch has to follow ∑Ri ¼ 01 ΘM;R i αi þ 1
(γ 4 ¼ 2). After the manufacturing phase, there are 125 þ 250 Q R;1 ¼ and
1 αR
0:5 ¼ 250 units in the used product inventory, and another ∑R 1 ΘM;R i αi þ s s 1
remanufacturing batch will be scheduled (γ 5 ¼ 1). Q R;s ¼ i ¼ 0 þ ∑ ΘM;i αs i 8 s ¼ 2…R: ð7Þ
1 αR i¼1
The used product inventory drops to 250–166.67 ¼83.33, and
another 83:33 units return during this remanufacturing phase. After determining all (re)manufacturing batch sizes of the considered
Finally, there are exactly 166:67 units in the used product cycle, the total holding cost can be determined to compute the
inventory and the next cycle can start with a remanufacturing optimal total cost for the ðR; MÞflex -policy.
batch (γ 1 ¼ 1).
Example. Fig. 2 presents the development of inventory positions
for the ð3; 2Þflex -policy. The heuristic starts with γ 1 ¼ 1, and the first
3.2. Step 2: determine the batch sizes subcycle is initiated. Since γ 2 ¼ 2, a manufacturing batch will be
initiated. From γ 3 ¼ 1 follows that the next subcycle s ¼ 2 starts.
While the previous section recaps the SAIL algorithm as From γ 4 ¼ 2 follows that a manufacturing batch is initiated in the
proposed by Choi et al. [4], we show next how to adapt the second subcycle, i.e., νM;2 ¼ 1. From γ 5 ¼ 1 follows directly that the
approach to unequal remanufacturing batch sizes. We assume for third subcycle s ¼ 3 contains no manufacturing batch. According to
simplifying reasons that all manufacturing lot sizes in a cycle are (7), the remanufacturing lot sizes are Q R;1 ¼ 107:14, Q R;2 ¼ 214:29,
identical.2 As λð1 αÞ products need to be manufactured in a and Q R;3 ¼ 178:57.
cycle with T ¼ 1, each manufacturing batch contains, therefore,
λð1 αÞ=M items. In order to compute the size of all remanufactur- 3.3. Step 3: compute HC 1 and the total cost per time unit TC RMflex
ing batches, the sequence of scheduling plays an important role. In
our approach, each subcycle s begins with a remanufacturing The holding cost for a cycle length of one time unit HC 1 can be
batch that is followed by νM;s identical manufacturing batches. computed by multiplying the region bounded by the inventory
Each cycle contains, thus, exactly R subcycles (i.e., s ¼ 1; …; R) levels with the corresponding inventory holding cost. Using (7) to
while νM;s can take on any non-negative integer value. The determine the remanufacturing lot sizes Q R;s as well as the number
of new products manufactured in each subcycle ΘM;s , we can insert
2 these values into the following equation (see [16], p. 21):
This approach is not necessarily optimal but facilitates the solution finding
" #
process for determining the remanufacturing batches thereafter substantially. Note
1 R ðQ R;s Þ2 ðQ Þ2 ðΘM;s Þ2
that we can prove that all manufacturing lot sizes within a subcycle must be HC 1 ¼ ∑ hR þ R;s hF þ hF : ð8Þ
identical, see Appendix C. That said, since the total amount of manufactured goods 2s¼1 λα λ λ νM;s
within a subcycle may vary, the performance may be improved by considering
different manufacturing lot sizes in different subcycles. However, the merits of such The computed holding cost can be inserted afterwards into (4) to
an approach are left for future research. determine the optimal total cost TC RMflex for the ðR; MÞflex policy
26 T. Schulz, G. Voigt / Omega 44 (2014) 21–31
when both R and M are given. In the following, we present an percentage improvement Δ is computed for each instance by
extensive numerical study in order to gain deeper insights in which
TC nRM TC nRMflex
situations the flexibly structured heuristic outperforms preexisting Δ¼ ð9Þ
approaches. Note, we are benchmarking our results in the following TC nRMflex
against the ðR; MÞeq -policy introduced by Choi et al. [4] which, in The results of the full factorial study are presented in Table 5. For
turn, nests the ð1; RÞ-policy and the ðM; 1Þ-policy. each factor, two results are presented therein: the mean percen-
tage improvement over all instances Δ and the percentage of
Example. As in the previous approaches, the holding cost HC 1 can instances (denoted by P) that yield a lower total cost when
be calculated by summing up the triangles in Fig. 2. Applying (8) initiating the ðR; MÞflex -policy instead of a policy structure with
gives HC 1 ¼ 303:36. equal remanufacturing batches.
Table 5
Results of the numerical study.
α (%) 10 20 30 40 50 60 70 80 90
Δ (%) 0 0 0.1 0.4 1.1 2.4 3.8 5.0 5.7
P (%) 7.8 14.9 25.2 33.6 50.1 74.4 88.4 97.0 98.2
hR =hF 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Δ (%) 0.2 0.7 1.3 1.8 2.2 2.6 2.9 3.2 3.5
P (%) 40.4 54.9 56.9 55.9 56.6 56.0 55.8 56.3 56.9
KR 100 200 300 400 500 600 700 800 900 1000
Δ (%) 3.6 3.0 2.5 2.2 1.9 1.7 1.6 1.4 1.3 1.2
P (%) 65.2 60.4 59.3 55.7 53.5 53.2 50.6 49.6 49.4 47.3
KM 100 200 300 400 500 600 700 800 900 1000
Δ (%) 0.7 1.2 1.6 1.9 2.1 2.3 2.5 2.6 2.8 2.9
P (%) 39.1 47.2 50.9 53.7 55.6 57.5 58.8 59.8 60.7 60.9
T. Schulz, G. Voigt / Omega 44 (2014) 21–31 27
Fig. 3. Comparison of the best ðR; MÞeq and ðR; MÞflex policies for hR =hF ¼ 0:9. (a) ð16; 1Þflex -policy, (b) ð13; 1Þeq -policy.
Fig. 4. Comparison of the best ðR; MÞeq and ðR; MÞflex policies for hR =hF ¼ 0:1. (a) ð15; 1Þflex -policy, (b) ð22; 1Þeq -policy.
which leads to an increase in the importance of remanufacturing holding cost) but lower average inventories in the finished product
decisions. On the contrary, a larger setup cost for manufacturing inventory (with relatively high holding cost). Fig. 4 presents the
puts the focus more on manufacturing new products and, thus, on best ðR; MÞeq -policy and the best ðR; MÞflex -policy.
the final product level. Therefore, the ðR; MÞflex -policy has less
potential in these settings. In Table 5 we see, however, that in those
4.3. Do further lot size adjustments of remanufacturing batches
cases with high potential benefits/losses, the ðR; MÞflex -policy out-
increase performance?
performs the ðR; MÞeq . As an example, the mean improvement for
K R ¼ 100 is 3.6% and the mean improvement for K M ¼ 1000 is 2.9%.
One might argue that the performance of the heuristic can be
The above analysis indicates that the instances with relatively
further improved by combining the ideas of equally sized and
high importance of remanufacturing decisions (i.e., high α and low
flexibly sized remanufacturing batches.3 We assessed the merits of
K R =K M -ratio) are those instances with the highest potential gains/
pursuing such an approach exemplarily with a (3, 2)-policy
losses of our ðR; MÞflex -policy. These effects amplify for a high
parameterized as mentioned in Appendix A. In this example, the
hR =hF -ratio, i.e., when the holding cost of the used product
ðR; MÞeq -policy gives holding cost per period of 333.33 (see
inventory is relatively high. The importance of this ratio can be
Appendix A.3), and our ðR; MÞflex -policy 303.6 (see Section 3).
explained by its influence on both inventory levels. For relatively
Given the ðR; MÞflex -policy as a starting point, we evaluated via
large used product inventory holding cost, the decisions regarding
enumeration a large set of policies in which Q R;1 ¼ Q R;3 ¼
the used product level have a larger impact on the overall total
fxj125 r x r 225 4 x A Ng holds. Interestingly, none of those poli-
cost. Hence, an efficient solution for this inventory level is sought
cies outperformed our ðR; MÞflex -policy. Schulz [16] provides
which can be achieved when all returns on hand are remanufac-
further indications that a mix of equally sized and flexibly sized
tured in a batch and none are kept in stock for a later batch. Since
remanufacturing batches is unlikely to further improve the sys-
neither the (M, 1) nor the ðR; MÞeq -policy can provide such a
tem's performance. He also presents non-linear mixed-integer
solution, the ðR; MÞflex -policy performs better for relatively large
program to the lot sizing problem discussed in this paper, which
hR values. Not surprisingly, we can find our largest improvements
was solved for several instances with the GAMS SBB solver (see
in instances with a high hR =hF ratio. As an example, our largest
[1]). The data indicates that all solutions presented by the GAMS
percentage improvement of 17.55% can be found in instances with
SBB solver have either a ðR; MÞeq -structure or a ðR; MÞflex -structure.
α ¼ 90%, hR =hF ¼ 0:9, K R ¼ 100, and K M ¼ 1000. Fig. 3 compares
However, since no NLP solver can guarantee optimality, this is not
both policy structures for these parameters. We present a detailed
a formal proof that one of both presented policies has to be
analysis of this example in B.1. It can be seen that the ðR; MÞflex
optimal. It seems though an interesting avenue for future research
leads to lower inventory levels in the used product inventory,
if such a conjecture can be formally proven.
because the used product inventory is depleted in every subcycle.
Because of the relatively large holding cost rate for the used
product inventory, we see a superior performance of the
5. Conclusion and outlook
ðR; MÞflex -policy.
In contrast, holding all other parameters constant and setting
Efficient planning of remanufacturing is critical in closed-loop
hR =hF ¼ 0:1 we observe a poor performance with performance
supply chains facing substantial setup costs for (re)manufacturing.
losses of 3.59% (see Appendix B.2. In these settings, it is more
beneficial to apply a ðR; MÞeq -policy that leads to higher average
inventories in the used product inventory (with relatively low 3
We thank an anonymous referee for pointing this out.
28 T. Schulz, G. Voigt / Omega 44 (2014) 21–31
The present work, therefore, aims at improving the cost efficiency depletion is ðα T=RÞ þ ð1 αÞ T. Multiplied with the return frac-
of preexisting heuristics. We review the existing approaches and tion per cycle, α λ T, we get Y R ¼ yR;max =2 ¼ ðλ α T 2 =2Þ
find that all heuristics are somewhat limited in either the ððα=RÞ þð1 αÞÞ ¼ ðλ α T 2 =2Þð1 þ α ðð1=RÞ 1ÞÞ.
flexibility regarding the batch sizes as well as the cycle structure. Y F results from the average cumulated amount of inventory in
The only exception is reported by Feng and Viswanathan [5] who each cycle resulting from manufactured (remanufactured) goods
are assuming a sub-cycle structure in which the sequences as well denoted as Y F;M (Y F;R ).
as sizes of batches might vary between two subcycles. While Feng We know that we need to manufacture λ ð1 αÞ T units in
and Viswanathan [5] report that “the reduction obtained in the each cycle. During the manufacturing phase, we have on an
cost is quite marginal”, we find that substantial cost improvements average ðλ ð1 αÞ TÞ=2 units of finished product inventory. Multi-
can be achieved by not limiting the number of subcycle classes. plied with the depletion time, ð1 αÞ T, we get Y F;M ¼ λ
Our flexibly structured ðR; MÞflex -policy (see Section 3) builds ð1 αÞ2 T 2 .
upon the SAIL algorithm as proposed by Choi et al. [4] that We also know that a total of λ α T units need to be remanu-
provides the optimal batch sequence for the case of equal factured in each cycle. Thus, ðλ α TÞ=R units need to be remanu-
remanufacturing and equal manufacturing batches. We extend factured in each batch and the average inventory during a
the approach by dividing the sequence of manufacturing and remanufacturing phase is ðλ α TÞ=2 R. Multiplying with the time
remanufacturing batches into subcycles, and propose a heuristic in which the remanufacturing batch is depleted, ðα TÞ=R, and
to define the size of the remanufacturing batch within the multiplying with the number of remanufacturing batches, R,
subcycle. We performed an extensive numerical study based on gives Y F;R ¼ R λ α2 T 2 =ð2 R2 ). It follows that Y F ¼ Y F;R þ Y F;M ¼
Choi et al. [4] factorial design and find that 54.4% of all instances ðλ T 2 =2Þðð1 αÞ2 þ α2 =RÞ.
can be improved by applying the ðR; MÞflex approach with a mean The total inventory holding cost result from multiplying the
improvement of around 2% and a maximum improvement of cumulated inventory positions, Y F and Y R , with the respective
17.55%. A detailed analysis reveals that our approach performs holding cost, hF and hR , and summing them up, i.e., HC ¼ 12λ T 2
particularly well in situations when the holding cost rates for ½ð1 þ αðð1=RÞ 1ÞÞαhR þ ððα2 =RÞ þð1 αÞ2 ÞhF . HC 1 results directly
returned items are high. We conclude that our approach can from HC with T ¼ 1.
substantially improve performance in those industrial settings.
This contribution provides an interesting approach for extend-
ing the considered problem setting in future contributions. Example. We present the development of inventories for a (3,1)-
As there are a number of possible extensions (like finite produc- policy with T ¼ 1 (Fig. A1).
tion and recovery rates, a disposal option, etc.), the improvements Used product inventory: The used product inventory reaches the
for introducing differently sized remanufacturing batches in these maximum after 0:5 3 þð1 0:5Þ ¼ 3 time units. The return rate per
2
settings can be analyzed as well. Another interesting task to follow unit time is α λ ¼ 500, and the used product inventory after 23 time
in future would be to analyze optimization approaches as pre- units is therefore Y R;Max ¼ 23 500 ¼ 333:33. The average used pro-
sented in Schulz [16] in order to determine the optimal solution duct inventory is Y R;Max =2 ¼ 166:66.
and policy structures. However, due to limited computational Finished product inventory: We need to manufacture 1000
power this is hardly possible nowadays. Finally, to model a more ð1 0:5Þ ¼ 500 units in each cycle, i.e., the average inventory in
realistic remanufacturing system, a multitude of different subpro- the manufacturing phase is 250. These manufactured items are
cesses (disassembly, rework, manufacturing, reassembly) should depleted over ð1 0:5Þ ¼ 0:5 time units. Thus, the average cumu-
be considered as well as a large number of components that need lated amount of finished product inventory during the manufac-
to be (re)manufactured. turing phase is Y F;M ¼ 500 2 ð1 0:5Þ ¼ 125 units. We need to
remanufacture 1000 0:5 ¼ 500 units in three batches. Thus, the
average inventory during a remanufacturing phase is 500 23 ¼ 83:33.
Acknowledgments
One remanufacturing batch is depleted over 0:5 3 ¼ 0:166 time units.
It follows that the three remanufacturing batches are depleted
We are grateful to two anonymous referees and our colleague over 3 0:166 ¼ 0:5 time units. Thus, the amount of finished
Ian M. Langella for the constructive comments that they provided product inventory cumulating in the remanufacturing phases is
on previous versions of this manuscript. Y F;R ¼ 3 10000:5
2
¼ 41:66. Thus, the sum of average inventories
232
during the manufacturing phase and during the (re)manufacturing
phases amounts to 125 þ 41:66 ¼ 166:66.
Appendix A. Model formulation: details
The total holding costs per unit time are HC 1 ¼ hR 166:66 þ hF
166:66 ¼ 500:
In the following we will derive the cost terms for the respective
policy types. We will present for each policy type an example that
is based on the parameters summarized in Table 3.
A.1. (R,1)-policy
A.2. (1,M)-policy
manufacture λ ð1 αÞ T units in each cycle. In the average facturing batches. In subcycles with ⌈M R ⌉ the used product inven-
amount of inventory in a manufacturing phase is therefore tory will increase relatively to subcycles with ⌊M R ⌋ manufacturing
ðλ ð1 αÞ TÞ=2 M. The depletion time for one manufacturing batches.
batch is ðð1 αÞ TÞ=M. Multiplying the average amount of used When applying the SAIL algorithm, we observe that the last
product inventory during a manufacturing phase with the deple- subcycle is a cycle with ⌊M R ⌋ manufacturing batches. Thus, in order
tion time for one batch and multiplying with the total number of to ensure that (A.1) holds for the last subcycle, we need to have an
manufacturing batches gives Y F;M ¼ ðλ ð1 αÞ2 T 2 Þ=ð2 M). used product inventory at the start of the last subcycle of
The cumulated inventory holding cost in each cycle result from ðM=R ⌊M R ⌋Þ Q M α.
multiplying the cumulated inventory positions (Y R ; Y F ) with the We know that there are M ⌊M R ⌋ R subcycles in which the
respective holding cost (hR ; hF ) and summing them up, i.e., inventory position increases compared to R ðM ⌊M R ⌋ RÞ sub-
HC ¼ 12 λ T 2 ½αhR þ ðα2 þð1 αÞ2 =MÞhF . HC 1 results directly from cycles in which the used product inventory declines. Thus,
HC with T ¼ 1 regardless of the actual scheduling of phases with increasing/
decreasing stock levels, we know that the average inventory build-
Example. We present the development of inventories for a (1,3)- up (denoted as d) in each subcycle must fulfill the condition that at
policy with T ¼ 1 (Fig. A2). the beginning of the last subcycle ðM=R ⌊M R ⌋Þ Q M α units are
Used product inventory: The average amount of used product available.
inventory is Y R ¼ Y R;Max =2 ¼ 12 1000 0:5 ¼ 250. Thus, the following condition must hold:
Finished product inventory: We need to manufacture 1000
M M M
ð1 0:5Þ ¼ 500 units in each cycle, i.e., the average inventory in M ⌊ ⌋ R d ¼ ⌊ ⌋ QM α ðA:2Þ
R R R
each manufacturing phase is 250 3 ¼ 83:33. The depletion time for
one manufacturing batch is ð1 30:5Þ ¼ 0:166. Thus, the cumulated Rearranging the right hand side of (A.2) gives
average amount during all three manufacturing phases is
M 1 M
Y F;M ¼ 83:33 0:166 3 ¼ 41:66. The average amount of finished M ⌊ ⌋ R d ¼ M⌊ ⌋ R QM α ðA:3Þ
R R R
product inventory in the single remanufacturing phase is
Q R =2 ¼ 12 0:5 1000 ¼ 250. Given a depletion time of α ¼ 0:5, it and it follows directly from (A.3) that
follows Y F;R ¼ 125. Thus, the sum of average inventories during the 1
manufacturing phase and during the (re)manufacturing phases d¼ Q M α: ðA:4Þ
R
amounts to 125 þ41:66 ¼ 166:66.
The total holding costs per unit time are HC 1 ¼ hR 250 þ The average inventory build-up occurs in (R–k) subcycles,
hF 166:66 ¼ 583:32. where k denotes the number of subcycles in which the used
product inventory is zero (including the last subcycle). In these
A.3. ðR; MÞeq -policy subcycles, there is neither an increase nor a decrease of the
average inventory position relatively to (A.1). Then, the average
At the end of the cycle we have by definition Q R units in the inventory build-up d takes place over ðR kÞ subcycles and we
used product inventory. Moreover, the ratio of products to be have a maximum amount of used product inventory of Q R þ
remanufactured within a cycle to products manufactured within ð1=RÞ Q M α ðR kÞ. Multiplying the average inventory in the
the cycle must be equal to the ratio of the return rate to the rate of used product inventory with the cycle length T gives TðQ R þ
products not being returned. Thus, within a cycle the condition ð1=RÞ Q M α ðR kÞÞ. Inserting Q R ¼ ðα λ TÞ=R and Q M ¼ ðð1 αÞ
M Q M =R Q R ¼ ð1 αÞ=α must hold. Rearranging the latter term λ TÞ=M gives Y R;Max ¼ λ α T 2 ððð1 αÞðR kÞ=ðR MÞÞ þ 1=RÞ and it
gives follows directly that Y R ¼ Y R;Max =2.
Finally, that for any ðR′; M′Þ-policies with k 41, we can deduct a
M structurally identical ðR; MÞ-policy with k ¼ 1 by dividing R′ and M′
Q M α ¼ Q R ð1 αÞ: ðA:1Þ
R by the least common denominator. This structurally identical
For the ðR; MÞeq it is clear that the maximum amount of stock is Q R policy yields identical per period holding costs. We therefore omit
when M=R ¼ z with z A N þ , because then there is an equal number these structurally identically policies with k 4 1 in our further
of manufacturing batches within a remanufacturing cycle, i.e., analysis and consider only policies with k ¼ 1.
M=R-manufacturing batches fill exactly the amount that was not The stock in the finished goods inventory can be derived
returned during the remanufacturing phase of a subcycle. How- according to the method presented for the (1, M)-policy. HC 1
ever, once M=R a z with z A N þ , we will observe stockkeeping in results directly from HC with T ¼ 1.
30 T. Schulz, G. Voigt / Omega 44 (2014) 21–31
Example. We present the development of inventories for a the first subcycle that only depends on α and the Θ values. We find
ð3; 2Þeq -policy with T ¼ 1.
R1
Used product inventory: We have 2 ⌊23⌋ R ¼ 2 subcycles in Q R;1 ¼ Q R;1 αR þ ∑ ΘM;R i αi þ 1
i¼0
which inventories are build up. At the beginning of the last
subcycle, there are 23 250 0:5 ¼ 83:33 units in the used product ∑R 1 ΘM;R i αi þ 1
Q R;1 ¼ i ¼ 0 ðA:10Þ
inventory. In order to reach this inventory level, there is an average 1 αR
build-up of d ¼ 13 250 0:5 ¼ 41:66 per subcycle, accumulating
over R k ¼ R 1 subcycles. Thus, the maximum used product After formulating the size of the first remanufacturing batch Q R;1 ,
inventory is Q R þð3 1Þ 41:66 ¼ 166:66þ 83:33 ¼ 250. The average we can use (A.6) to calculate Q R;2 . Thereafter, a more general
condition for Q R;s (for 2 o s rR) can be derived.
used product inventory is 250 2 ¼ 125.
Finished product inventory: We need to manufacture 1000 !
∑Ri ¼ 01 ΘM;R i αi þ 1
ð1 0:5Þ ¼ 500 units in each cycle, i.e., the average inventory in Q R;2 ¼ þ Θ M;1 α
1 αR
each manufacturing phase is 500 22 ¼ 125. The depletion time for one
manufacturing batch is ð1 20:5Þ ¼ 0:25. Thus, the cumulated average
∑Ri ¼ 01 ΘM;R i αi þ s s 1
amount during the two manufacturing phases is Y F;M ¼ 125 ⋯Q R;s ¼ þ ∑ ΘM;i αs i 8 s ¼ 3; ‥; R
1 αR i¼1
0:25 2 ¼ 62:5. We need to remanufacture 1000 0:5 ¼ 500 units
in three batches. Thus, the average inventory during a remanu-
facturing phase is 500 23 ¼ 83:33. One remanufacturing batch is
depleted over 0:5 3 ¼ 0:166 time units. It follows that the three Appendix B. Numerical examples: sensitivity analysis
remanufacturing batches are depleted over 3 0:166 ¼ 0:5 time
units. Thus, the amount of finished product inventory accumulat-
2 In this section we provide two examples based on the para-
ing in the remanufacturing phases is Y F;R ¼ 3 10000:5 ¼ 41:66. The
232 meters listed in Table B1. We vary the hR =hF -ratio in order to
average finished product inventory is Y F ¼ 62:5 þ 41:66 ¼ 104:16.
highlight instances in which the ðR; MÞflex -policy outperforms the
The total holding costs per unit time are HC 1 ¼ hR 125 þ
ðR; MÞeq -policy and vice versa. As no closed-form expression exists
hF 104:16 ¼ 333:33.
to evaluate Rn and M n for the ðR; MÞflex -policy, the best values for R
and M have been determined by enumeration (R as well as M can
take on any integer value between 1 and 30).
A.4. ðR; MÞflex -policy
The holding cost term for the ðR; MÞflex policy is derived in B.1. High hR/hF-ratio.
detail in Section 3. For each subcycle s, the size of each remanu-
facturing batch Q R;s can be formulated according to (5) and (6). In this example we set hR ¼ 63 and hF ¼ 70, i.e., hR =hF ¼ 0:9. The
ðR; MÞflex -policy with the lowest total cost has been a ð16; 1Þflex -pol-
Q R;1 ¼ ðQ R;R þ ΘM;R Þ α ðA:5Þ icy. The best ðR; MÞeq -policy is a ð13; 1Þeq -policy. We present the
total unit costs and cycle lengthes for both policies in Table B2.
Note, the lowest achievable cost with a classical (R, M)-policy in
Q R;2 ¼ ðQ R;1 þ ΘM;1 Þ α ðA:6Þ
which M manufacturing setups are followed by R remanufacturing
setups would at best yield total cost of 8211.08 per unit time (see
⋯Q R;R 1 ¼ ðQ R;R 2 þ ΘM;R 2 Þ α ðA:7Þ calculations in [6] with infinite production and manufacturing
rates). One can see that scheduling differently sized remanufactur-
ing batches in a cycle by using the (16,1)flex -policy is able to reduce
Q R;R ¼ ðQ R;R 1 þ ΘM;R 1 Þ α: ðA:8Þ the total cost incurred by around 17.55% when compared to the
best ðR; MÞeq -policy. Fig. 3 in Section 4.2 compares both policy
By inserting (A.8) into (A.5) the following expression can be
structures. It can be seen that the ðR; MÞflex leads to lower inventory
derived:
levels in the used product inventory, because the used product
Q R;1 ¼ ððQ R;R 1 þ ΘM;R 1 Þ α þ ΘM;R Þ α inventory is depleted in every subcycle. Because of the relatively
large holding cost rate for the used product inventory, we see a
¼ Q R;R 1 α2 þΘM;R 1 α2 þ ΘM;R α ðA:9Þ superior performance of the ðR; MÞflex -policy.
λ α KR KM
Table B2
Performance comparison with hR =hF ¼ 0:9.
TC 6985.26 8211.08
T 0.74 0.56
eq
Fig. A3. Development of inventory positions for a ð3; 2Þ policy.
T. Schulz, G. Voigt / Omega 44 (2014) 21–31 31
Table B3 Solving (C.5) for ΘM;s and inserting into (C.3) gives Q M;s;1 ¼
Performance comparison with hR =hF ¼ 0:1. Q M;s;i 8 i ¼ 2; …; n. Thus, it is optimal to schedule equally sized
manufacturing batches in a subcycle. &
Best ðR; MÞflex policy: Best (R, M)-policy:
ð15; 1Þflex policy ð22; 1Þeq policy
TC 5354.99 5162.94
T 0.93 1.24
References
[1] SBB solver manual. In: G. D. Corporation, editor. GAMS – The solver manuals,
2009. p. 513–20.
B.2. Low hR/hF-ratio [2] Akçalı E, Çetinkaya S. Quantitative models for inventory and production
planning in closed-loop supply chains. International Journal of Production
Research 2011;49(8):2373–2407.
The numerical analysis shows that the performance of the [3] Barker T, Zabinsky Z. A multicriteria decision making model for reverse
ðR; MÞflex depends primarily on the hR =hF -ratio. Thus, we consider logistics using analytical hierarchy process. Omega 2011;39:558–573.
now the above example with hR ¼ 7 and hF ¼ 70, i.e., we reduce [4] Choi D, Hwang H, Koh S. A generalized ordering and recovery policy for
the holding cost ratio to hR =hF ¼ 0:1. In this case, the optimal reusable items. European Journal of Operational Research 2007;182
(2):764–774.
ðR; MÞflex -policy is a ð15; 1Þflex -policy and the optimal ðR; MÞeq -policy [5] Feng Y, Viswanathan S. A new lot-sizing heuristic for manufacturing systems
is a ð22; 1Þeq -policy. Table B3 summarizes the performances of both with product recovery. International Journal of Production Economics
policies. In this example, we observe a performance deterioration 2011;133.
of 4.11% when the ðR; MÞflex -policy is used instead of the [6] Feng Y, Viswanathan S. Heuristics with guaranteed performance bounds for a
manufacturing system with product recovery [Technical report]. Singapore:
ðR; MÞeq -policy. That is because the ðR; MÞflex -policy tends to limit Nanyang Business School; 2012.
the used product inventory, which provides little benefits in [7] Francas D, Minner S. Manufacturing network configuration in supply chains
instances with (relatively) low holding cost for used products. with product recovery. Omega 2009;37:757–769.
[8] Guide Jr. VDR. Production planning and control for remanufacturing: industry
This trade off is depicted in Fig. 4 in Section 4.2.
practice and research needs. Journal of Operations Management 2000;18(4):
467–483.
[9] Konstantaras I, Skouri K. Lot sizing for a single product recovery system with
Appendix C. Proof that equally sized manufacturing batches in variable setup numbers. European Journal of Operational Research 2010;203:
a subcycle are optimal 326–335.
[10] Li X, Baki F, Tian P, Chaouch B. A robust block-chain based tabu search
algorithm for the dynamic lot sizing problem with product returns and
In each subcycle with νM;s manufacturing batches ΘM;s units are remanufacturing. Omega 2014;42:75–87.
manufactured. The individual lot size i is denoted as Q M;s;i . [11] Liu N, Kim Y, Hwang H. An optimal operating policy for the production system
with rework. Computers & Industrial Engineering 2009;56(3):874–887.
Lemma 2. Given a fixed amount of units to be manufactured in a [12] Minner S, Lindner G. Lot sizing decisions in product recovery management. In:
subcycle, it is optimal to choose equally sized manufacturing batches. Dekker R, Fleischmann M, Inderfurth K, Van Wassenhove L, editors. Reverse
logistics – quantitative models for closed-loop supply chains. Springer; 2004.
Proof. We minimize the finished product inventory holding cost p. 157–179.
for a fixed amount of items to be manufactured (ΘM;s ) by [13] Mitra S. Revenue management for remanufactured products. Omega 2007;35.
[14] Mitra S. Analysis of a two-echelon inventory system with returns. Omega
1 νM;s 2009;37.
min Z¼ ∑ ðQ Þ2 ðC:1Þ
2 λ i ¼ 1 M;s;i [15] Schrady D. A deterministic inventory model for reparable items. Naval
Research Logistics Quarterly 1967;14:391–398.
[16] Schulz T. Optimal and predefined policies for the static lot sizing problem in a
νM;s
two stage recovery system [Technical Report 7]. Faculty of Economics and
s:t: ∑ Q M;s;i ¼ ΘM;s ðC:2Þ
i¼1 Management, Otto-von-Guericke University Magdeburg; 2011.
[17] Schulz T. A new silver-meal based heuristic for the single-item dynamic lot
Solving (C.2) for Q M;s;1 gives sizing problem with returns and remanufacturing. International Journal of
νM;s Production Research 2011;49(9):2519–2533.
Q M;s;1 ¼ ΘM;s ∑ Q M;s;i ðC:3Þ [18] Srivastava S. Network design for reverse logistics. Omega 2008;36:535–548.
i¼2 [19] Teunter R. Economic ordering quantities for recoverable item inventory
systems. Naval Research Logistics 2001;48(6):484–495.
Inserting (C.3) into (C.1) gives [20] Teunter R, van der Laan E, Inderfurth K. How to set the holding cost rates in
! average cost inventory models with reverse logistics. Omega 2000;28
νM;s νM;s νM;s νM;s
1 (4):409–415.
Z¼ Θ2M;s 2ΘM;s ∑ Q M;s;i þ 2 ∑ Q M;s;i ∑ Q M;s;j þ 2 ∑ Q M;s;i [21] Teunter R, Bayındır Z, van den Heuvel W. Dynamic lot sizing with product
2λ i¼2 i¼2 j ¼ iþ1 i¼2
returns and remanufacturing. International Journal of Production Research
ðC:4Þ 2006;44(20):4377–4400.
[22] Thierry M, Salomon M, van Nunen J, Van Wassenhove L. Strategic issues in
product recovery management. California Management Review 1995;37
Deriving (C.4) w.r.t. Q M;s;i gives (2):114–135.
! [23] Volkswagen. Umweltbilanz Austauschgetriebe, September 2010.
νM;s
∂Z 1 [24] Zhou L, Naim M, Tang O, Towill D. Dynamic performance of a hybrid inventory
¼ 2 ΘM;s þ 2 ∑ Q M;s;i þ 4 Q M;s;i ðC:5Þ system with a Kanban policy in remanufacturing process. Omega
∂Q M;s;i 2λ j ¼ 2\fig
2006;34:585–598.