You are on page 1of 23

International Journal of Production Research

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tprs20

Parameterisation of demand-driven material


requirements planning: a multi-objective genetic
algorithm

David Damand, Youssef Lahrichi & Marc Barth

To cite this article: David Damand, Youssef Lahrichi & Marc Barth (2022): Parameterisation of
demand-driven material requirements planning: a multi-objective genetic algorithm, International
Journal of Production Research, DOI: 10.1080/00207543.2022.2098074

To link to this article: https://doi.org/10.1080/00207543.2022.2098074

Published online: 14 Jul 2022.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=tprs20
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH
https://doi.org/10.1080/00207543.2022.2098074

Parameterisation of demand-driven material requirements planning:


a multi-objective genetic algorithm
David Damand, Youssef Lahrichi and Marc Barth
HuManis laboratory at EM Strasbourg Business School, Strasbourg, France

ABSTRACT ARTICLE HISTORY


Demand-Driven Material Requirements Planning (DDMRP) is a recent inventory management Received 24 December 2021
method that has generated considerable interest in both academia and industry. Many recent papers Accepted 27 June 2022
have demonstrated the superiority of DDMRP over classical methods like MRP or Kanban, an observa- KEYWORDS
tion confirmed by companies that have implemented DDMRP. However, DDMRP depends on many DDMRP; multi-objective;
parameters that affect its performance. Only general rules are given by the authors of the method to metaheuristic; NSGA-II; OTD;
fix these parameters but no algorithm. The present paper aims to fill this gap by proposing a multi- on-hand inventory; lot sizing
objective optimisation algorithm to fix a set of eight identified parameters. The suggested genetic
algorithm is coupled with a simulation algorithm that computes the objective functions. Two oppos-
ing objective functions are considered: first, the maximisation of orders delivered on-time to the
customer and, second, the minimisation of on-hand inventory. A set of data instances was generated
to test the suggested method. Fronts of non-dominated solutions are found for all these instances.

1. Introduction
a planning system that lacks visibility with regards to
Due to new consumption trends, global competition and execution. Kanban, which was widely adopted in the
fluctuating energy prices, companies now face grow- West in the 1980s, attempts to correct this weakness.
ing volatility and uncertainty in terms of demand and It is less dependent on forecasting which is only used
replenishment. This complex environment is referred to for sizing the Kanban system. Day-to-day replenishment
as VUCA (Bennett and James Lemoine 2014). VUCA orders are placed based on real-time observation of on-
stands for: Volatility, Uncertainty, Complexity and Ambi- hand inventory and demand (Ohno and Bodek 2019).
guity, four factors that indeed characterise today’s indus- Kanban is inspired by a JIT (Just-In Time) philosophy
trial environment. Volatility materialises in frequent which involves ordering only what is needed and only
changes that require rapid process adaptation. Uncer- when needed. JIT is a strength of Kanban but it can
tainty in lead times, costs, supply and demand results also constitute a weakness in the event of unforeseen
in inaccurate forecasting. Complexity of processes makes peaks in demands: Kortabarria et al. (2018). Ptak and
decision-making harder. Ambiguity is often encountered Smith (2011) and Ptak and Smith (2016) thus present
due to the rise in the number of stakeholders and diffi- DDMRP as a viable alternative. This new method com-
culties in interpreting market trends. The VUCA envi- bines both forecasts and real-time observation of the
ronment affects the companies’ KPIs (Key Performance system. The inaccuracy of forecasts is corrected by a
Indicators). In addition, the encountered KPIs are often day-by-day observation of inventory and demand. Based
adversarial, like when companies try to maximise their on comparative studies, several authors have shown
service level while minimising their costs. DDMRP to be efficient in terms of cost and service level
Companies rely on efficient inventory management (El Marzougui et al. 2020; Azzamouri et al. 2021). Service
systems to deal with the VUCA environment. MRP, level is often measured by the OTD (On-Time Delivery)
which is based on forecasts, has been widely used since indicator (Miclo 2016), in other words, the percentage of
the 1960s. Because forecasts fail to give accurate pre- orders that are delivered on-time to the customer (APICS
dictions of demand, MRP is considered as inefficient in dictionary). Many authors measure costs by the aver-
today’s uncertain environment (Kortabarria et al. 2018). age on-hand inventory (Kortabarria et al. 2018; Velasco
Besides, according to Huq and Huq (1994), MRP is Acosta, Mascle, and Baptiste 2020). Indeed, the on-hand

CONTACT Youssef Lahrichi youssef.lahrichi.contact@gmail.com HuManis laboratory at EM Strasbourg Business School, 61 avenue de la Forêt Noire,
Strasbourg F-67000, France

© 2022 Informa UK Limited, trading as Taylor & Francis Group


2 D. DAMAND ET AL.

inventory is stored by the company and generates oper- in 1964 by Joseph Orlicky and was implemented
ating costs. the same year at the company Black & Decker
DDMRP is parameter-dependent, relying on a num- (Orlicky 1974). Based on demand and initial on-hand
ber of parameters that are fixed by the manager. The inventory data, MRP computes production and pur-
authors of DDMRP (Ptak and Smith 2016) give some chasing schedules (Dolgui and Prodhon 2007). This
rules to fix these parameters. However, these rules do computation is based either on mathematical models
not give an exact value for each parameter, instead or on heuristics like the silver meal heuristic (Silver
offering only intervals. In addition, the rules strongly and Harlan 1973). MRP II, which stands for Material
depend on the manager’s judgment. For instance, Ptak Resource Planning, is an extension of MRP that first
and Smith (2016) recommend setting a given parameter appeared in the 1980s. It is a comprehensive system
of DDMRP between 20% and 40% if some data is ‘high’. that contains many more modules alongside with the
It is quite subjective to judge whether the data in ques- MRP module, including the MPS (Master Production
tion is ‘high’ or not. Many authors, like Velasco Acosta, Schedule), which is the process of planning the quan-
Mascle, and Baptiste (2020) and Kortabarria et al. (2018), tity of goods to be produced in each period (typically a
point out the dependence of DDMRP on its parameters year) in order to adjust the system capacity. MRP (and
as one of its possible shortcomings. Other authors, like MRP II) can be described as a push system. This type
Bahu, Bironneau, and Hovelaque (2019), draw attention of systems uses demand forecasting to build the pro-
to the lack of automatic procedures to fix the parameters duction and purchasing schedules. Flow and products
of DDMRP. are pushed from production to customers. The major
We suggest in this paper a multi-objective optimi- drawback of MRP is that it relies on forecast data that
sation algorithm based on NSGA-II to parameterise can potentially be inaccurate.
DDMRP. The maximisation of OTD and the minimi- • Kanban: Kanban means sign or card in Japanese. It
sation of average on-hand inventory are considered was first introduced by Toyota in 1953, before gaining
simultaneously. Fronts of non-dominated solutions are in popularity in the 1980s (Ohno and Bodek 2019).
obtained for randomly generated instances. The paper Kanban, also called JIT (Just-In Time), is a lean man-
is organised into seven sections. Section 2 is devoted agement system, which means that it is designed to
to the related literature. Section 3 describes the main minimise costs by reducing waste. Kanban is a reorder
steps of DDMRP. Section 4 sets out the optimisation point system which means that a replenishment order
problem considered and gives the notations. Section 5 (production or purchase) is launched when the on-
describes the suggested resolution approach. In Section 6, hand inventory and the scheduled receipts reach a
computational experiments are performed on randomly- given level. In this situation, a (physical) Kanban card
generated data instances. Concluding remarks and per- containing the order information is released. The Kan-
spective research directions are given in the last section ban card must contain at least the reference of the
of the paper. article, the order amount, the name of the workstation
where the order is produced and the location where it
is stored. Unlike MRP which relies on forecasts, Kan-
2. Literature review
ban relies on on-time observation of the system, which
In this section, DDMRP (Demand-Driven Material is why it is classified as a pull system. It means that Kan-
Requirements Planning) is positioned in relation to ban is demand-driven: flow is pulled from demand
MRP, Kanban and theory of constraints. Then, DDMRP- (customer) to the production (company). Kanban can
related literature is presented. Finally, the literature gap is be less efficient, however, when unexpected peaks in
explained to justify our contribution. demand occur (Kortabarria et al. 2018).
DDMRP is the most recent inventory management • Theory of constraints: The Theory Of Constraints
method. It was first introduced in Ptak and Smith (2011) (TOC) was first introduced in 1997 by Eliyahu M.
then in Ptak and Smith (2016). The method is intended to Goldratt. This management paradigm tends to first
address the shortcomings of the previous inventory man- identify the constraints of the system (Goldratt 2017).
agement methods in terms of overstocks and stock-outs, A constraint is any limiting resource or process that
thereby reducing operating costs and increasing the level affects the system’s performance. Once identified, the
of customer service. DDMRP combines the best practices constraints must be exploited to achieve the desired
of three main methods that underpin the system: performance. Everything in the system must be sub-
ordinated to the constraints in order to elevate them.
• MRP and MRP II: MRP stands for Material Require- TOC introduces buffers as a powerful tool to deal with
ments Planning. The method was first introduced constraints. Buffers are placed before the constraints
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 3

to help absorb shocks and prevent the constraints Acosta, Mascle, and Baptiste (2020) apply DDMRP
from being violated. The identification of the con- on fictitious company data that are characterised by
straint (combinatorial decision) is decisive for a good complex product structures. The results are analysed
application of the TOC approach. The buffers, which in terms of stock levels and lead times. Shofa and
are used to remove these constraints, must then be Widyarto (2017) present a real case study from car
sized. TOC has been shown to reduce inventory and manufacturing industry. Due to the implementation
lead times (Wu, Lee, and Tsai 2014) but the replen- of DDMRP, the authors observe a reduction of stock
ishment policy is undermined in TOC which makes levels and lead times.
its application somewhat arbitrary and difficult (Ptak • Buffer positioning: Velasco Acosta, Mascle, and Bap-
and Smith 2016). tiste (2020) point out that the performance of DDMRP
is quite dependent on the positioning of buffers within
DDMRP is the most recent evolution in inventory the BOM (Bill Of Materials). Buffer positioning is the
management systems. It is based on the three earlier sys- first step of DDMRP. It is situated at the strategical
tems. DDMRP is often categorised as a push and pull sys- level while the following steps are situated either at
tem (Azzamouri et al. 2021). It relies on both forecasting the tactical or at operational levels. Buffering a ref-
(push system, MRP) and real-time observation of inven- erence leads simultaneously to a reduction in lead
tory and demand (pull system, Kanban). DDMRP also times and an increase in operating costs. Jiang and
uses buffers for critical items (TOC) to account for pos- Rim (2017) deal with the problem of buffer position-
sible fluctuations in demand. All non-buffered references ing by means of a genetic algorithm. The suggested
are managed by means of a pull strategy while buffered approach decides which references in the BOM should
references are handled by what looks more like a push be buffered. It is intended to minimise the inven-
strategy. For this reason, DDMRP is described as a hybrid tory operating costs while satisfying the required lead
system taking advantage simultaneously from opposing times.
approaches (Velasco Acosta, Mascle, and Baptiste 2020). • Parameterisation of DDMRP: As several authors have
As DDMRP is a recent inventory management system, noted, DDMRP is highly parameter-dependent (Bahu,
the literature on the subject is relatively scarce. A com- Bironneau, and Hovelaque 2019; Velasco Acosta, Mas-
prehensive state of the art can be found in El Marzougui cle, and Baptiste 2020). Its parameters have a real
et al. (2020). The latter lists 23 publications on DDMRP impact on a system’s performance. The parameter-
appeared between 2011 and 2019. We have selected 17 isation of DDMRP is usually performed by prac-
publications from El Marzougui et al. (2020) and added titioners based on their own experience. Ptak and
7 further publications appeared between 2020 and 2021. Smith (2016) provide some general rules to fix the
We classify the resulting 24 publications into six major parameters. These rules are approximate and largely
classes: depend on the manager’s judgment. Bahu, Bironneau,
and Hovelaque (2019) points out the absence of an
• Comparative studies: In this category of papers, automatic and objective method to fix these param-
authors compare DDMRP with other inventory man- eters. Miclo (2016) attempts to address this issue in
agement systems thanks to simulation tools and con- some part of his thesis. The author use a simula-
clude that DDMRP performs relatively well. For exam- tion software to optimise three parameters thanks to
ple, Miclo et al. (2019) compares DDMRP with MRP a simulated annealing algorithm. The objective func-
II and Kanban via a structured experimental design. tion considered is a weighted sum of OTD and aver-
The authors use a well-known product structure called age stock. Lee and Rim (2019) suggest a heuristic
the kanban game and generate demand data ran- formula to fix two parameters of DDMRP, namely
domly. Based on the experiments, the authors show the lead time factor and the variability factor. The
that DDMRP achieves better performance in terms formula is extracted from a simulation based on a
of OTD (On-Time Delivery) and WIP (Work In- case study. We note that this class of papers con-
Progress). We note that this class of papers contains tains only papers that suggest algorithms to parame-
only papers that compare DDMRP with other plan- terise DDMRP. Some other papers, like Martin (2020)
ning systems empirically by simulation and not papers and Dessevre et al. (2019), deal with parameterisa-
that perform only a conceptual comparison. tion in a statistical way. The authors set a plan of
• Case studies: Some authors have studied DDMRP by experiments, test different parameterisation scenarios
applying it to company data. This data can be either and derive general rules for parameterisation. These
real and provided by an existing company or ficti- rules are generated thanks to the data set studied and
tious and generated by the authors themselves. Velasco are therefore specific to the case study considered.
4 D. DAMAND ET AL.

Table 1. Literature review.


Comparat. Case Buffer Parame- Review Book
Paper study study position. terisation
Favaretto and Marin (2018)  
Kortabarria et al. (2018)  
Shofa, Moeis, and Restiana (2018)  
Miclo et al. (2019)  
Thürer, Fernandes, and Stevenson (2022)  
Ihme (2015) 
Shofa and Widyarto (2017) 
Bayard and Grimaud (2018) 
Dessevre et al. (2019) 
Martin (2020) 
Velasco Acosta, Mascle, and Baptiste (2020) 
Dessevre et al. (2021) 
Rim, Jiang, and Lee (2014)  
Jiang and Rim (2016)  
Jiang and Rim (2017)  
Abdelhalim, Hamid, and Tiente (2021)  
Miclo (2016)   
Lee and Rim (2019)  
Bahu, Bironneau, and Hovelaque (2019) 
Pekarčíková et al. (2019) 
Azzamouri et al. (2021) 
El Marzougui et al. (2020) 
Ptak and Smith (2011) 
Ptak and Smith (2016) 
The present paper 

For this reason, we classified these papers as case objective (weighted sum) is studied which prevents
studies. from having fronts of Pareto-optimal solutions.
• Reviews: Some recent reviews have been written on • In Lee (2018), the suggested heuristic formula is
DDMRP: Pekarčíková et al. (2019), Bahu, Bironneau, derived from a given case study. No proof is given that
and Hovelaque (2019), El Marzougui et al. (2020) and it is optimal, nor that it yields good results in gen-
Azzamouri et al. (2021). Very often, the method is eral. Besides, it is shown in Damand et al. (2019) that
situated in relation to MRP, Kanban and the The- these formulas are highly contextual and cannot be
ory of Constraints (TOC). Most authors highlight the generalised.
growing interest in DDMRP in both industry and
academia (El Marzougui et al. 2020). Some authors
In an attempt to overcome these gaps, we put forward
also underscore the lack of research with regards to the
several contributions:
parametrisation of DDMRP (Bahu, Bironneau, and
Hovelaque 2019).
• Books: This last category contains the two seminal • Eight parameters affecting the performance of DDMRP
books on DDMRP: Ptak and Smith (2011), and Ptak are identified.
and Smith (2016). • A first multi-objective metaheuristic is suggested.
It is based on NSGA-II (Non-dominated Sorting
Table 1 summarises and classifies the main publi- Genetic Algorithm). It is designed to compute (local)
cations on DDMRP in the six above-mentioned major Pareto-optimal solutions with respect to OTD and
classes. average stock. Unlike the formula suggested in Lee
and Rim (2019), a metaheuristic can adapt to new
instances thanks to operators that search the landscape
Literature gap and contribution of this paper
of solutions.
We note a lack of research work regarding the parameter- • 60 data instances are generated to test the suggested
isation of DDMRP. The two only research items available algorithm. Characteristics of optimal parameters are
in the literature contain some gaps that justify the present identified.
work:
In the following section, we describe DDMRP to iden-
• In Miclo (2016), only three parameters are considered tify the eight parameters to be optimised. This also allows
for optimisation. However, DDMRP performance is to describe in more details the two papers about param-
affected by far more than three parameters. A single eterisation from literature.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 5

3. DDMRP: general description The lead time of subcomponent 4 is not considered in


the above formula since it is buffered. DLT of product
DDMRP works through five consecutive steps.
1 is 16(2 + 14). We can note that, without buffering,
31(2 + 5 + 13 + 11) time units would be needed to pro-
duce product 1 instead of 16 time units. This is known
Step 1: Strategic buffer positioning as lead time compression (Ptak and Smith 2016), which
In the first step, the decision-maker determines which is a major benefit of DDMRP. DLT is defined only for
products in the BOM (Bill Of Materials) must be buffered references. The DLT of subcomponent 4 is 11.
buffered. Buffering a product involves high costs rela-
tive to setting up and managing the buffer stock. For Step 2: Buffer sizing
this reason, only the most ‘critical’ products are chosen
to be buffered. Criticality is contextual and depends on In the second step, the decision-maker sizes the buffers.
the considered industry. For example, the manager could In Ptak and Smith (2016), it is recommended that
choose to buffer products that are more sensitive to delays the buffers contain three (fictitious) zones as shown in
or alternatively products that are less expensive to buffer Figure 2:
(Abdelhalim, Hamid, and Tiente 2021).
An example of BOM is given in Figure 1. Product 1 • The red zone is the bottom level of the stock. In this
needs 2 times units to be assembled provided Component zone, the stock level is critically low. An order, with a
1 and Component 2 are available. In this example, the LT high level of priority, must be generated. The size of
(lead times) are represented in a circle at the top of the the red zone is given by:
reference.
The decision-maker may choose to buffer product 1, TOR = ADU × DLT × FLT
  
component 1 and subcomponent 3 as shown in Figure 1. Redbase
Once the buffers are positioned, Ptak and Smith (2016) Red base
define the DLT (Decoupled Lead Time). DLT is defined as   
+ ADU × DLT × FLT ×FV (1)
‘the longest cumulative coupled lead time chain in a manu-   
factured item’s product structure.’ In other words, the DLT Red safety
associated with a reference is the time needed to produce
this reference and all its non-buffered components. For where ADU stands for Average Daily Usage, in other
instance, the DLT associated with component 1 can be words, the mean over some horizon of the customers’
calculated based on Figure 1 as follows: orders. DLT is the Decoupled Lead Time of the
buffered reference in question. FLT and FV are respec-
DLT[comp. 1] = LT[comp. 1] + Max{LT[sub-comp. 1] tively the lead time factor and the variability factor.
These factors are parameters of the method to be
+ LT[sub-comp. 3], LT[sub-comp. 2]}
fixed by the decision-maker. The variability factor is
= 5 + Max{8 + 10, 13} = 23 a percentage between 0% and 100%. It captures the

Figure 1. Buffer positioning on BOM for product 1. Figure 2. Buffer zones.


6 D. DAMAND ET AL.

Table 2. Recommendations to fix the cause this change in the profile of demand (Ptak and
lead time factor. Smith 2016). Buffers can be adjusted either by directly
Low DLT FLT between 61% and 100% modifying the parameters that come into play in sizing
Medium DLT FLT between 41% and 60% the buffers or by considering some Planned Adjustment
Long DLT FLT between 20% and 40%
Factors (PAF) that handle seasonal changes. No seasonal
changes are considered in the present paper.
Table 3. Recommendations to fix the vari-
ability factor.
High variability FV over 61% Step 4: Order generation
Medium variability FV between 41% and 60%
Low variability FV between 0% and 40% The fourth and the fifth steps of DDMRP deal with the
execution phase of the system following the design phase
(the first three steps). In the fourth step, the decision-
extent to which the customers’ orders vary. The vari-
maker decides if a replenishment order must be gener-
ability factor is close to 100% when the demand is
ated and determines the quantity to be ordered (lot size).
highly volatile, and it is close to 0% when the demand
The decision to generate an order on day t is based on the
is nearly stable. The lead time factor is a percentage
value of the net flow on that day, denoted qflow
t . It is given
between 20% and 100% that should be barely inversely
by the following formula:
proportional to the DLT. Ptak and Smith (2016) make
some recommendations to fix the variability and the
lead time factors (Tables 2 and 3). qflow
t = qton-hand + qon
t
-order − d̃ (net flow equation)
t
• The yellow zone is superposed over the red zone. It
covers the demand over the days that are beyond the where qton-hand is the on-hand inventory, i.e. the available
reach of a replenishment order. Since a replenishment inventory/stock, on day t. qon -order is the on-order inven-
t
order needs DLT days to be processed, the size of the tory, i.e. the quantity that has been ordered but not yet
yellow zone is given by the following formula: received and d̃t is the qualified demand on day t. Qual-
ified demand is given by the sum of the demand on the
TOY
 − TOR
 = ADU × DLT (2)
day and the demand spikes over a given spike horizon. A
Size of the yellow zone
spike is defined as a demand that exceeds a given thresh-
• The green zone is superposed over the yellow zone. old. The threshold is expressed as a percentage of the red
Within this zone, the stock level is considered to be zone. This threshold is recommended to be set at 50% of
‘comfortable’ and no replenishment order is generated. the red zone. The spike horizon is recommended to be set
The size of the green zone is given by the following at DLT. Figure 3 helps to explain the calculation of qual-
formula: ified demands. In Figure 3, the order spike threshold is
assumed to be 50% of the red zone, in other words, 40.
TOG
 − TOY
 = Max{MOQ, ADU × DLT A spike horizon of 3 days is assumed, which means that
Size of the green zone days 5, 6 and 7 must be considered for the calculation of
the qualified demand on day 4. Only demands on days
× FLT , ADU × Horder } (3)
5 and 7 are retained since they exceed 40 which makes
The above formula gives the average quantity of them spikes. Qualified demand on day 4 is finally given
replenishment orders. MOQ stands for the Minimum by:
(replenishment) Order Quantity. Horder represents
the desired order cycle, indicating that the company d̃4 = 35 + 50 + 60 = 145
wishes to launch a replenishment order every Horder
days. The qualified demand is used in the net flow equation.
Based on the value of the net flow, two different scenarios
are possible:
Step 3: Buffer adjustment
The third step deals with the adjustment of buffer size. • If the net flow is inferior to the TOY (Top of yellow),
Indeed, in a VUCA environment, the system is subject a replenishment order is generated. The amount of
to unforeseen changes. These changes affect the buffer the order is the difference between the TOG (Top of
size that must be adjusted. Buffers may also need to be green) and the value of the net flow.
adjusted to ramp up in case of high demand or to ramp • If the net flow is superior to the TOY, no order is
down in case of a drop in demand. Seasonal changes may generated.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7

Figure 3. Illustration of qualified demand.

In other words, the order amount on day t is given as with buffer sizing and order generation that are situated at
follows: the tactical and operational levels (Ptak and Smith 2016).
 Only one objective function (weighted sum) is consid-
TOGt − qflow If qflow ≤ TOYt
ot = t t
ered in Miclo (2016) which prevents from computing
0 Otherwise.
(local) Pareto-optimal solutions.
where TOGt and TOYt are respectively the TOG and the The red zone formula (1) depends on the lead time fac-
TOY on day t. tor and the variability factor. To overcome the problem of
fixing these two parameters, Lee and Rim (2019) suggest
Step 5: Order prioritisation an alternative formula that is independent of these two
factors. As said before, this formula is not proven to be
The last step of DDMRP deals with order prioritisation optimal and is based only on a single case study.
thanks to colour codes and alerts. The considered optimisation problem is described in
This section highlights the dependence of DDMRP on the next section.
many parameters. The DLT is determined by buffer posi-
tioning (step 1). All other parameters should be fixed by
the manager. The variability factor and the lead time fac-
4. Problem statement
tor affect the size of the red and green zones. The order
spike threshold and the order spike horizon affect the In the present paper, we consider that the first step of
qualified demand and, therefore, the net flow. The man- DDMRP has already been addressed, which means that
ager should also choose a horizon for the calculation of buffers have already been positioned. The Decoupled
the ADU (Average Daily Usage). A daily calculation of Lead Times (DLT) are known for all products. We first
the ADU is the broader assumption we can take. It allows describe the made assumptions then we define the prob-
to rapidly account for any significant changes in the size lem recalling the eight parameters considered, and we
of demands. Two additional parameters are introduced: finally introduce notations used in the remainder of the
the ADU future horizon and the ADU past horizon. For paper.
example, if the ADU past horizon is 2 and the ADU future The following assumptions are made:
horizon is 3, then, the ADU is the mean of the demands
on the 2 past days, the 3 following days and the cur- • We consider that the parameters are fixed once a year
rent day. The Minimum Order Quantity (MOQ) and the based the history of the previous year. The consid-
order cycle affect the size of the green zone. This brings ered time horizon is the previous year, denoted by
out 8 parameters to be optimised. H = [1, 2, . . . , 253] (open days in 2020 = 253 = 366
In Miclo (2016), the lead time factor, the variability – official holidays and week-ends). All the demand
factor and the DLT (Decoupled Lead Time) are opti- quantities are known during the time horizon. We are
mised by means of a metaheuristic (simulated annealing). interested in finding the optimal parameters over the
The remaining parameters affecting the performance of period H. Assuming the demand to be known with-
DDMRP are considered known. Besides, Miclo (2016) out uncertainty over the planning horizon may seem
considers the DLT as an optimisation variable while it is a strong assumption, however this can be justified by
most often a given (technological) data. Indeed, the DLT two arguments. First, the literature lacks a first com-
is known as a result of the first step of DDMRP which prehensive deterministic study that sets the ground
is situated at the strategic level unlike the steps dealing for the parameterisation of DDMRP. Second, DDMRP,
8 D. DAMAND ET AL.

thanks to its replenishment policy, is intended to Table 4. Notations.


behave well in the presence of uncertainty (Ptak and Data
Smith 2016). For example, the red zone of the buffer H = [1, 2, . . . , 253] Time horizon.
(red base + red safety) can help to prevent stockouts DLT Decoupled Lead Time.
dt , t ∈ H Demand quantity on day t.
in case where demand is higher than expected. We s0 Initial on-hand stock or inventory.
can reasonably assume, to some extent, that a limited
Variables
uncertainty in the demand won’t have a significant
MOQ Minimum order quantity.
impact on the KPIs that are computed over the entire FLT ,FV Respectively the lead time and the variability
planning horizon. factors.
Horder Desired order cycle horizon.
• We consider no capacity constraints. Therefore, Tspike The order spike threshold.
the parameterisation problem can be solved inde- Hspike The horizon for the calculation of the qualified
pendently for each buffered reference. This allows spikes.
H+ −
ADU , HADU Respectively the future and the past horizons for
us to consider only one buffered reference with the calculation of the ADU. ADUt is calculated on

a given DLT. the following interval {t − HADU , . . . , t, . . . , t +
+
• An initial on-hand inventory is given, and replenish- HADU }.
ment orders are generated with a DDMRP policy. Other (algorithmic) variables
• If a replenishment order is launched on day t, it is qton-hand , t ∈ H On-hand inventory at the beginning of day t.
delivered and immediately available without uncer- qton-order , t ∈ H On-order inventory at the beginning of day t.
t ,t ∈ H
qflow Net flow at the beginning of day t.
tainty at the end of day t+DLT (Decoupled Lead d̃t , t ∈ H Qualified demand quantity on day t.
Time). ot , t ∈ H order amount launched on day t.
• A demand is considered to be satisfied if the sum of õt , t ∈ H order amount arrival on day t.
ADUt , t ∈ H Average daily usage calculated on day t.
the available on-hand inventory and the order arrival TORt , TOYt , TOGt , t ∈ H respectively the top of red, the top of yellow and
of the day is greater than the demand quantity. the top of green levels on day t.
• The Average Daily Usage (ADU) is updated each day. Objectives
Consequently, the buffer zone sizes (TOR, TOY and OTD On-Time Delivery computed over the year horizon.
TOG) are updated on a daily basis. Qon-hand Average on-hand computed over the year horizon.

We define the DDMRP Parameter Optimisation Prob-


lem (DDMRP-POP). It can be stated as follows: • Objectives:
(1) –Minimise the average on-hand stock over the
• Data: days of the planning horizon.
(1) –The planning horizon H. (2) –Maximise the On-Time Delivery (OTD), i.e. the
(2) –The demand quantities over the days of the plan- rate of satisfied demands.
ning horizon.
(3) –The DLT of the considered buffered product. In Table 4, we describe the notations used in the
(4) –The initial on-hand inventory. remainder of the paper.
(5) –The lower bound and the upper bound for To solve DDMRP-POP, we first modeled it mathemat-
each one of the 8 optimisation variables given ically. The resulting model is quadratic and mixes con-
below. tinuous and binary variables. Solely for informative rea-
• Variables: sons, it is given in Appendix where only the objective of
(1) –The Minimum Order Quantity (MOQ). minimising the average on-hand inventory is considered
(2) –The Lead Time (LT) factor. while the OTD is set to 100% by means of a constraint.
(3) –The variability factor. The presence of quadratic constraints and binary vari-
(4) –The order cycle. ables deterred us from using a solver to solve the model.
(5) –The order spike threshold. This justifies the use of the approximate optimisation
(6) –The order spike horizon. method that is described in the next section.
(7) –The Average Daily Usage (ADU) future horizon.
(8) –The Average Daily Usage (ADU) past horizon.
• Constraints:
5. Description of the simulation-optimisation
(1) –The order quantities must be computed using a
approach
DDMRP policy. A DDMRP simulation algorithm The general scheme of the suggested approach is
is given in the next section. depicted in Figure 4. It is intended to optimise the
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 9

front of non-dominated solutions. A good approxima-


tion is characterised by diversified solutions that are close
enough to the front of non-dominated solutions (Talbi
et al. 2012).
Many multi-objective metaheuristics have been sug-
gested in recent years to solve various optimisation
problems (Talbi 2015). Among these methods, genetic
algorithms and notably NSGA-II, have stood out for
their adaptability as well as for their excellent results
(Lee 2018).
The Non-dominated Sorting Genetic Algorithm-II,
Figure 4. General scheme of the optimisation-simulation NSGA-II, was introduced by Deb et al. (2002). It has
approach. since been successfully applied to a wide range of multi-
objective optimisation problems: Katoch, Chauhan, and
Kumar (2021). NSGA-II has also been applied to lot-
eight parameters: HADU + −
, HADU , MOQ, Hspike , Tspike , sizing problems in the context of MRP: Hnaien, Delorme,
FLT , FV , Horder . Initial solutions with initial values for and Dolgui (2010), Amorim, Antunes, and Almada-
these parameters are generated to initiate the optimi- Lobo (2011), Yahia et al. (2015) and Rezaei et al.
sation algorithm. Each time new parameters values are (2016).
generated by the optimisation algorithm, the simula- Like all genetic algorithms, NSGA-II is a bio-inspired
tion algorithm is executed to compute the values of the optimisation method that uses selection, crossover and
two objective functions (On-Time Delivery and aver- mutation within a population of solutions (chromo-
age on-hand inventory). After a given number of iter- somes). Selection is used to select the best individuals for
ations, the optimisation algorithm returns the obtained reproduction, i.e. those with the greatest fitness. Repro-
non-dominated solutions. duction is performed thanks to a crossover operator,
In what follows, we describe the main components of and better diversification can be achieved thanks to a
the suggested approach. mutation operator.
Compared to a classical genetic algorithm, NSGA-II
introduces non-dominated ranking and crowding dis-
5.1. Simulation algorithm tance to rank solutions that are evaluated with more than
one objective function.
The simulation algorithm is described in Algorithm 1. The Non-dominated ranking technique ranks the
It summarises the functioning of DDMRP (based on population within ordered fronts of solutions. The solu-
Section 3 and Ptak and Smith 2016). The algorithm also tions contained within each front are not dominated by
computes the values of the two objective functions. any solution in the following fronts (Figure 5). The first
Initialisation of variables is performed at line 1 of front contains all non-dominated solutions.
Algorithm 1. For each day in the period (line 2), the In Figure 5, solutions S1, S2 and S3 are non-dominated
algorithm computes the on-hand inventory (line 3), the by any other solution, which explains why they belong
on-order inventory (line 4), the average daily usage (line to the first front. S4 is dominated by S2 as it has a worst
5) and the sizes of the buffer zones (lines 6–8). The value for both objectives, and thus belongs to the sec-
algorithm then calculates the qualified demand over ond front. S6 belongs to the third front as it is dominated
the horizon Hspike (lines 9–14). The net flow is then by S4.
expressed thanks to the net flow equation (line 15). If the Crowding distance is an indicator of diversity. It is
net flow is below the top of yellow, an order is generated small when the neighbourhood around a solution is
(lines 16–18). Finally, the simulation algorithm computes dense. Figure 6 helps to understand the calculation of the
the values of the objective functions, namely the on-time crowding distance.
delivery (lines 21–29) and the average on-hand inventory In NSGA-II, reproduction is applied until the popu-
(line 30). lation size is doubled. From one generation to the next,
only the best half of the population is preserved. The non-
dominated sorting then the (highest) crowding distance
5.2. A non-dominated sorting genetic Algorithm-II
are the two used criteria for this purpose.
Multi-objective optimisation by means of metaheuris- The pseudo-algorithm of NSGA-II is given in
tics aims at computing an approximation of the Pareto Algorithm 2.
10 D. DAMAND ET AL.

Algorithm 1 Simulation algorithm


+ −
INPUT H, DLT, HADU , HADU , MOQ, Hspike , Tspike , FLT , FV , Horder , s0 , dt , t ∈ H
OUTPUT ot , OTD, Q on-hand

1: Initialisation: qon-hand
0 := s0 , qon-order
0 := 0, õ0 := 0, ok := 0, k ∈ {1 − DLT, . . . , 0}, OTD-Aux-1 :=0, OTD-Aux-2
:=0
2: for t=1 to 253 do
3: qon-hand
t := qon-hand
t−1 + õt−1 − dt−1
4: qton-order := qon-order
t−1 + ot−1 − õt−1
t+HADU+

k=t−HADU − dk
5: ADUt := + −
HADU + HADU +1
6: TORt := ADUt × DLT × FLT + ADUt × DLT × FLT × FV
7: TOYt := TORt + ADUt × DLT
8: TOGt := TOYt + Max{MOQ ; ADUt × DLT × FLT ; ADUt × Horder }
9: d̃t := dt
10: for k = t + 1 to t + Hspike do
11: if dk ≥ TORt × Tspike then
12: d̃t := d̃t + dk
13: end if
14: end for
15: qflow
t := qon-hand
t + qon-order
t − d̃t
16: if qt ≤ TOYt then
flow

17: ot := TOGt − qflow t


18: end if
19: õt := ot−DLT
20: end for
21: for t=1 to 253 do
22: if dt > 0 then
23: OTD-Aux-1 := OTD-Aux-1 +1
24: if qon-hand
t + õt ≥ dt then
25: OTD-Aux-2 := OTD-Aux-2 +1
26: end if
27: end if
28: end for
OTD-Aux-2
29: OTD := × 100%
OTD-Aux-1
253 on-hand
q
30: Qon-hand := t=1 t
253

+ −
A solution is encoded as an eight-uplet: (HADU , HADU , The cross-over operator (line 5) represents the merg-
MOQ, Hspike , Tspike , FLT , FV , Horder ). It represents the ing of two solutions to generate one child solution.
eight optimisation variables. The value of an optimisation variable in the child solu-
The initial solutions (line 1) are generated by a select- tion is the average of the values of the correspond-
ing randomly and uniformly a value for each variable ing variables from its parents. Cross-over is illustrated
between its lower bound and upper bound. in Table 5.
A deterministic tournament selection (line 4) is The aforementioned cross-over operator tends to cen-
applied. This works by selecting randomly a subset of tre the variables in the interval defined by their upper
individuals and choosing the best individual from the and lower bounds. An adequate mutation operator is
subset to be the first parent. The same process is repeated necessary to correct this effect, allowing a better diver-
for the second parent. sification of solutions. The mutation operator (line 7) is
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 11

Figure 5. Non-dominated sorting.

Figure 6. Crowding distance.

Table 5. Cross-over operator.


Parameters values
+ −
FLT FV HADU HADU Hspike Tspike MOQ Horder
Parent 1 40% 25% 4 0 5 90% 1000 0
Parent 2 60% 40% 3 2 5 50% 500 6
Child 50% 32.5% 4 1 5 70% 750 3

applied by taking, for each variable, the average of its is then averaged with the value of the randomly-drawn
value with either its lower bound or its upper bound. For bound. The mutation operator is shown in Table 6. For
each variable, a coin toss is performed between the lower each variable, the value of the randomly-drawn bound is
bound and the upper bound. The value of the variable framed.
12 D. DAMAND ET AL.

Algorithm 2 NSGA-II
INPUT Data of DDMRP-POP
OUTPUT Pareto local-optimal solutions

1: Generate N solutions to build the initial population .


2: for generation=1 to gmax do
3: while size of population ≤ 2.N do
4: Selection: choose two parents to perform reproduction.
5: Cross-over: apply cross-over operator and add a new solution to the population.
6: end while
7: Mutation: a mutation operator is applied with a given probability to each individual. The individual and its
muted version are kept in the population.
8: Population filtering: keep only the best N individuals in the population with respect to non-dominated sorting
and crowding distance.
9: end for

Table 6. Mutation operator.


Parameters values
+ −
FLT FV HADU HADU Hspike Tspike MOQ Horder

Lower bound 20% 0% 0 0 0 0% 0 0


Upper bound 100% 100% 30 29 30 100% 1000 10
Solution 50% 32.5% 4 1 5 70% 750 3
Muted solution 75% 16.25% 2 15 3 35% 875 2

6. Computational experiments 60 data instances were generated. For each data


instance, we generated demand data over the 253 open
In this section, we present experiments performed to
days of 2020. A normal distribution was chosen with a
evaluate the suggested approach for the parameterisation
mean of 1000 and a standard variation ranging from 100
of DDMRP. 60 data sets were generated. This section
to 1000 in steps of 100. Negative demands are made pos-
is organised within four subsections. The first subsec-
itive by using the absolute value. A demand entry has a
tion presents the characteristics of the different data
given probability to be a spike (spike probability). This
instances and explains why such characteristics are
probability varies according to the instance, it is within
needed for experimentation (design of experiments). The
{0.1, 0.2}. The spike is generated randomly and uniformly
second subsection highlights the results for one par-
within an interval. This interval depends on the mean,
ticular data instance. The third subsection summarises
the standard variation and two parameters named min
results for all the data instances. The last subsection
and max. The interval used to generate the spikes is given
discusses general conclusions that can be drawn from
as follows:
experiments.
Computer experiments were performed on a com- [mean + min × std. var., mean + max × std. var.]
puter equipped with 16Gb in RAM and an Apple
M1 processor. The source code has been written The extremities of this interval are called the spike lower
in JAVA. bound (mean + min × std. var.) and the spike upper
bound (mean + max ×std. var.). The values of min and
max vary according to the instance. Table 7 resumes the
characteristics of the generated data instances.
6.1. Design of experiments and data sets
The 60 data sets were generated by varying the stan-
The aim of the experimental section is to test the sug- dard deviation, the spike probability and the min and
gested algorithm under different scenarios. This involves max values used to calculate the spike lower bound and
varying the standard deviation to test different variability the spike upper bound. There are 10 possible values for
scenarios. It also involves testing different spike profiles the standard variation, 2 possible values for the spike
to evaluate how the suggested approach behaves with low, probability and 3 possible combinations of min-max
medium, or high spikes. which gives 10 × 2 × 3 = 60.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 13

The parameters of the suggested optimisation method days for all the instances. This prevents a stock-out from
are determined experimentally and are given in Table 8. occurring in the first 5 days of the planning horizon.
These values are the result of a compromise between The suggested algorithm converges and computes
the quality of the solutions and the CPU time. Indeed, non-dominated solutions for all the data instances within
increasing the size of the population or the number of at most 10 minutes. The next section is intended to
generations leads to better solutions at the expense of present detailed results for a particular instance then
more CPU time. The mutation probability should be just results are summarised for all the 60 instances in
large enough to give some diversification to the pool of Section 6.3.
solutions.
We first fixed an approximate CPU time. We retained
a CPU time of 10 minutes which is relevant based on 6.2. Detailed results for a given data instance
the literature of genetic algorithms (Katoch, Chauhan, A worst-case scenario is studied in this subsection. High
and Kumar 2021). We then tried to find a combination standard deviation and spikes are considered. Although
(number of generations, number of solutions) that yields such a scenario may not be representative of all indus-
a good and well-distributed front within at most 10 min- tries, it is useful for our study by allowing to push the
utes. We reached the bottleneck combination (number of method to its limit. The characteristics of the considered
generations = 100, number of solutions = 1000) by test- instance are given in Table 10.
ing smaller values and increasing them gradually. Beyond Figure 7 gives the data distribution of demand for
100 generations the improvement on the solutions is neg- the instance under consideration. As expected, it shows
ligible and beyond 1000 solutions the CPU time tends a normal distribution centred in 1000 and a uniform
to grow over 10 minutes. To fix the probability of muta- distribution for spikes between 26,000 and 31,000.
tion, we went gradually from 0% to 30%. A probability of The suggested optimisation algorithm outputs the
mutation within 20% introduces just enough diversifica- solutions of the first front from the last generation of solu-
tion to the pool of solutions and keeps the running time tions. This set of solutions constitutes the non-dominated
in the order of 10 minutes. solutions searched by the algorithm and approximates the
The parameters of DDMRP that are optimised thanks Pareto-optimal solutions. These solutions are shown for
to the optimisation algorithm have lower and upper the considered instance in Table 11.
bounds that must be respected by the optimisation We can make the following remarks based on Table 11:
algorithm. The values of these bounds for each one the
eight parameters are given in Table 9. • The 29 solutions are mutually non-dominated with
We can note that the lower bound of the lead time fac- respect to the two considered objectives.
tor is set to 20% based on a recommendation by Ptak and • Only one solution with 100% OTD is obtained: S1.
Smith (2016). We recall that the order spike threshold is This result is expected since two 100% OTD solutions
expressed as a percentage of the red zone. cannot be mutually non-dominated.
The DLT is constant and equals 5 for all the instances.
This way, any replenishment order would need 5 days
to be processed. For this reason, we fix the initial on-
Table 9. Lower and upper bound on the parameters.
hand inventory to the sum of the demand over the first 5
Parameter Min value Max value
Minimum order quantity 0 1000
Table 7. Design of experiments: data characteristics. Lead time factor 20% 100%
Variability factor 0% 100%
Customer demand data Order cycle 1 20
Order spike threshold 0% 100%
Mean 1000
Order spike horizon 0 20
Standard deviation 100,200,300,400,500,600,700,800,900,1000
ADU future horizon 0 4
Probability for a demand to 0.1, 0.2
ADU past horizon 0 4
be a spike
Min-Max 0–5, 10–15, 25–30

Table 10. Data characteristics.


Table 8. Parameters of the
optimisation algorithm. Customer demand data
Mean 1000
Genetic algorithm
Standard deviation 1000
Size of the population 1000 Probability for a demand to be a spike 0.20
Number of generations 100 Lower bounds for spikes 26,000
Probability of mutation 20% Upper bounds for spikes 31,000
14 D. DAMAND ET AL.

Figure 7. Demand data distribution.

Table 11. Obtained non-dominated solutions.


Parameters values Objective functions
+ −
Solution FLT FV HADU HADU Hspike Tspike MOQ Horder Average on-hand OTD
S1 35 38 3 2 5 86 319 1 74,815.74 100.00
S2 31 89 4 2 2 82 574 1 61,969.43 99.60
S3 25 66 3 0 2 94 493 1 60,393.13 99.21
S4 28 67 3 2 2 87 509 1 59,936.48 98.42
S5 26 71 3 2 2 91 533 1 57,319.72 98.02
S6 27 93 4 1 1 88 716 1 54,887.98 97.63
S7 34 34 4 3 1 47 348 1 52,955.75 97.23
S8 35 8 4 3 1 81 356 2 50,898.05 96.05
S9 24 79 3 2 1 94 708 1 50,527.55 95.65
S10 25 51 3 2 1 88 638 1 49,328.48 95.26
S11 25 47 3 2 1 87 421 1 49,309.34 94.86
S12 25 42 3 2 1 89 645 1 48,126.27 94.47
S13 25 43 3 2 1 89 387 1 47,568.00 94.07
S14 24 44 3 2 1 91 515 1 47,149.83 93.68
S15 25 67 3 2 0 65 504 1 46,845.55 93.28
S16 25 61 3 2 0 75 513 1 46,462.90 92.89
S17 25 60 3 2 0 75 568 1 44,729.00 92.49
S18 24 58 3 2 0 79 519 1 43,875.86 92.09
S19 25 43 3 2 0 85 342 1 42,548.36 91.70
S20 25 31 3 2 0 62 546 1 41,047.90 91.30
S21 25 37 3 2 0 62 539 1 41,012.21 90.91
S22 24 38 3 2 0 73 468 1 40,639.07 90.51
S23 24 28 3 2 0 50 463 1 39,479.14 90.12
S24 24 25 3 2 0 82 466 1 38,977.09 89.72
S25 27 4 4 3 0 40 678 1 35,130.66 89.33
S26 22 8 3 3 0 21 187 1 34,859.71 87.75
S27 21 10 3 3 0 78 646 1 34,170.64 87.35
S28 21 13 3 3 0 93 225 1 33,911.57 86.56
S29 21 7 3 3 0 37 166 1 33,227.12 86.17

• S1 is the only solution with Hspike = 5. This spike hori- average on-hand inventory. The other solutions have
zon is the highest among all the solutions. It allows a a lower spike horizon which deprives from detect-
100% OTD to be reached at the expense of a greater ing all spikes and therefore from satisfying all the
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 15

demands (OTD < 100%). These solutions (S2–S29) • It seems that the lead time factor (FLT ) is overall sta-
have a lower average on-hand inventory. In addition, ble. It is around 27% which is low and corresponds to
it seems that the spike horizon (Hspike ) is proportional a long DLT according to Ptak and Smith (2016). The
to the average on-hand inventory and inversely pro- variability factor is much more volatile. It is of 38% for
portional to the OTD. S1 which is low according to Ptak and Smith (2016).

Figure 8. Front of non-dominated solutions.

Figure 9. Evolution of the on-hand inventory, the red zone, the yellow zone and the green zone in 2020.
16 D. DAMAND ET AL.

Figure 10. Evolution of fronts over generations.

+ −
• The ADU horizon (HADU + HADU + 1) is overall sta- philosophy which introduces buffers to guard against
ble. It is around one week (5 open days). shortages, at the expense of a higher cost. Table 12 gives
the solutions with a 100% OTD that are contained in the
The front of non-dominated solutions is represented last generation of solutions. In this table,  represents
in Figure 8. It shows that the solutions are well distributed the deviation from the best average on-hand inventory
Average on-hand−s∗
among the front. (s∗ = 74815.74):  = s∗ × 100%.
From a decision-maker perspective, the solution with The 100% OTD solutions all have a spike horizon
100% OTD (S1) is of crucial importance since it provides which is at least 5( = DLT). This result is not surprising
a maximum service level. Solution S1 has 62 replenish- since a replenishment order needs 5 days to be deliv-
ment orders. For this solution, Figure 9 shows the evolu- ered. A spike detection horizon of 5 allows orders to
tion of on-hand inventory with respect to the three buffer be launched 5 days in advance and to be delivered
zones. on time. This allows the solutions to reach a 100%
Figure 10 shows the evolution of non-dominated solu- OTD. These solutions also have a high spike threshold
tions through the generations in the genetic algorithm. (Tspike ).
The algorithm starts from only 3 non-dominated solu-
tions (generation 1) and is able to generate better solu-
tions after 50 generations of reproduction and mutation. 6.3. General results for the remaining data
The last generation of solutions is significantly better than instances
the 50th generation of solutions.
Many customer-oriented industries are interested in In this subsection, we summarise the obtained results for
solutions with 100% OTD. These solutions provide a the 60 instances that have been generated. For the sake of
maximum service level, as recommended by the DDMRP succinctness, only the solution with 100% OTD is given
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 17

Table 12. Solutions with 100% OTD.


Parameters values Objective functions
+ −
Solution FLT FV HADU HADU Hspike Tspike MOQ Horder Average on-hand OTD (%)
T1 = S1 35 38 3 2 5 86 319 1 74,815.74 100 0.00
T2 30 68 3 0 5 91 731 1 75,191.24 100 0.50
T3 33 58 3 2 5 75 556 1 78,909.23 100 5.47
T4 33 61 3 0 5 86 586 1 79,277.50 100 5.96
T5 35 44 3 2 5 81 473 2 79,430.43 100 6.17
T6 21 91 3 4 5 26 172 2 80,342.87 100 7.39
T7 36 59 3 2 5 82 431 1 80,362.85 100 7.41
T8 32 75 3 2 5 84 447 2 81,277.80 100 8.64
T9 57 39 3 3 5 84 452 2 81,596.07 100 9.06
T10 40 47 3 2 5 85 387 1 82,124.83 100 9.77
T11 30 50 3 0 5 81 586 1 82,385.57 100 10.12
T12 40 45 3 2 5 80 511 1 82,488.47 100 10.26
T13 26 75 3 2 5 70 819 1 82,503.57 100 10.28
T14 30 46 3 1 5 84 728 1 82,562.79 100 10.35
T15 22 82 4 4 5 49 908 2 82,627.86 100 10.44
T16 41 40 3 2 5 76 567 1 83,398.26 100 11.47
T17 35 62 3 1 5 86 640 1 83,605.48 100 11.75
T18 34 68 3 1 7 9 612 1 83,756.58 100 11.95
T19 42 38 3 2 5 72 648 2 83,964.60 100 12.23
T20 43 35 3 2 5 72 624 2 84,372.31 100 12.77
T21 35 47 2 1 5 83 519 1 84,384.15 100 12.79
T22 28 63 3 1 5 76 702 1 84,391.90 100 12.80
T23 35 47 2 1 5 81 560 1 84,578.87 100 13.05
T24 34 47 2 1 5 83 529 1 84,711.64 100 13.23
T25 36 46 2 1 5 82 569 1 84,787.59 100 13.33
T26 35 48 2 1 5 81 548 1 84,964.47 100 13.56
T27 36 47 2 1 5 81 558 1 84,990.86 100 13.60
T28 22 1 1 3 5 54 102 1 85,069.15 100 13.70
T29 29 65 3 2 5 66 713 1 85,085.91 100 13.73
T30 32 46 2 1 5 83 628 1 85,162.75 100 13.83
T31 35 48 2 1 5 81 577 1 85,201.38 100 13.88
T32 37 53 3 1 5 86 532 1 85,420.22 100 14.17
T33 36 47 2 1 5 83 514 1 85,465.93 100 14.24

for each instance and only 6 parameters are shown. These and the probability of the spikes. Instances with
results are shown in Table 13. larger standard deviation, larger spike probability, or
larger spike intensity have a larger variability factor
(Table 13).
6.4. General discussion (6) There is a correlation between the order spike hori-
The following conclusions can be inferred from the com- zon (Hspike ) on one side and the standard deviation,
puter experiments: the interval and the probability of spikes on the
other side (Table 13). Instances with low standard
(1) The suggested simulation-optimisation approach is deviation, lower spike probability or lower spike
able to find a significant number of non-dominated intensity have a lower order spike horizon. For exam-
solutions (Table 11). ple, instances numbered 1, 7 and 13 all have a null
(2) The front of solutions is well-distributed and diver- spike horizon. This result is expected since spikes are
sified (Figure 8). negligible for these instances which means that the
(3) The genetic operators can generate better solutions ADU is sufficient to represent the demand in these
after dozens of generations which demonstrates cases.
their effectiveness (Figure 10). (7) As expected, the average stock is impacted by the size
(4) The spike horizon and to a certain extent the of demand (Table 13). Instances with larger standard
spike threshold are proportional to the OTD and deviation, larger spike probability, or larger spike
inversely proportional to the average on-hand inven- intensity have a larger demand and therefore a larger
tory (Table 11). average stock.
(5) It seems clear that the variability factor (FV ) is (8) We did not capture any clear correlation between
+ −
impacted by the standard deviation, the interval HADU or HADU on one side and the size of demand,
18 D. DAMAND ET AL.

Table 13. Solutions with 100% OTD for 60 instances.


Instance Parameters values
+ −
Nb. Std. dev. Spike proba. Spike interval FLT FV HADU HADU Hspike Tspike Avg. stock
1 100 0.1 [0,5] 21 1 3 2 0 79 1807
2 100 0.1 [10,15] 24 21 3 2 0 65 2671
3 100 0.1 [25,30] 29 20 3 2 1 49 4789
4 100 0.2 [0,5] 21 1 2 2 0 80 1820
5 100 0.2 [10,15] 24 12 3 2 0 42 2800
6 100 0.2 [25,30] 22 55 4 3 1 69 6032
7 200 0.1 [0,5] 21 1 3 1 0 90 1833
8 200 0.1 [10,15] 36 28 4 2 0 46 4986
9 200 0.1 [25,30] 21 15 4 3 5 66 10,300
10 200 0.2 [0,5] 21 1 3 3 0 25 2004
11 200 0.2 [10,15] 32 55 3 0 4 91 6609
12 200 0.2 [25,30] 44 27 3 2 5 89 13,699
13 300 0.1 [0,5] 22 8 3 0 0 59 2132
14 300 0.1 [10,15] 25 50 3 1 5 83 5311
15 300 0.1 [25,30] 26 26 3 3 5 89 11,799
16 300 0.2 [0,5] 21 70 3 1 0 75 3026
17 300 0.2 [10,15] 34 90 4 3 1 34 9889
18 300 0.2 [25,30] 39 78 3 3 5 87 18,764
19 400 0.1 [0,5] 24 56 3 2 0 85 3101
20 400 0.1 [10,15] 26 20 2 3 5 85 7360
21 400 0.1 [25,30] 27 14 3 3 5 86 14,971
22 400 0.2 [0,5] 25 73 4 1 2 99 3521
23 400 0.2 [10,15] 33 62 3 2 2 87 12,135
24 400 0.2 [25,30] 31 64 3 0 0 79 24,410
25 500 0.1 [0,5] 27 56 2 2 4 88 4369
26 500 0.1 [10,15] 25 38 2 2 5 79 8592
27 500 0.1 [25,30] 26 36 3 3 5 78 22,993
28 500 0.2 [0,5] 29 35 3 3 1 64 5210
29 500 0.2 [10,15] 22 50 3 4 5 67 12,716
30 500 0.2 [25,30] 33 47 3 2 5 90 36,371
31 600 0.1 [0,5] 28 18 3 2 1 38 4996
32 600 0.1 [10,15] 35 53 3 2 5 86 14,718
33 600 0.1 [25,30] 34 38 2 1 0 60 32,983
34 600 0.2 [0,5] 34 46 3 2 4 87 5763
35 600 0.2 [10,15] 57 4 4 1 3 98 18,807
36 600 0.2 [25,30] 21 30 3 3 5 42 35,225
37 700 0.1 [0,5] 29 48 4 2 2 84 4979
38 700 0.1 [10,15] 23 43 3 3 6 69 14,976
39 700 0.1 [25,30] 23 39 3 3 5 94 25,185
40 700 0.2 [0,5] 30 20 2 3 3 87 5701
41 700 0.2 [10,15] 41 24 3 0 0 47 22,352
42 700 0.2 [25,30] 32 44 3 2 5 90 44,934
43 800 0.1 [0,5] 38 15 3 1 0 59 5908
44 800 0.1 [10,15] 22 91 4 4 5 49 15,417
45 800 0.1 [25,30] 22 10 3 3 6 56 30,763
46 800 0.2 [0,5] 33 29 2 2 3 82 7933
47 800 0.2 [10,15] 27 86 3 1 5 94 25,222
48 800 0.2 [25,30] 38 64 3 3 4 86 56,547
49 900 0.1 [0,5] 28 27 3 1 1 75 5003
50 900 0.1 [10,15] 23 15 3 3 5 77 17,621
51 900 0.1 [25,30] 21 13 1 4 7 80 33,581
52 900 0.2 [0,5] 43 30 3 3 0 39 8016
53 900 0.2 [10,15] 27 63 3 1 3 83 28,353
54 900 0.2 [25,30] 26 50 3 2 5 83 52,089
55 1000 0.1 [0,5] 30 54 3 1 2 87 5977
56 1000 0.1 [10,15] 23 52 3 4 5 69 20,678
57 1000 0.1 [25,30] 23 33 3 3 5 58 43,650
58 1000 0.2 [0,5] 27 46 3 3 3 72 7700
59 1000 0.2 [10,15] 27 75 4 4 5 94 29,411
60 1000 0.2 [25,30] 35 38 3 2 5 86 74,816

the standard deviation, the interval or the probability proportional to the variability of the demand. To some
of spikes on the other side (Table 13). extent, experiments confirm also the other rule from Ptak
and Smith (2016) stipulating that the lead time factor
We note that (5) confirms Ptak’s rule (Ptak and should be inversely propositional to the DLT. Indeed, we
Smith 2016) stating that the variability factor should be observed from experiments that increasing the standard
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 19

deviation, the probability or the intensity of spikes leads • It could be interesting to compare the suggested
to increasing the size of demand (ADU). Increasing the approach with a practitioner parametrisation of
size of demand leads in turn to a relative reduction of the DDMRP. To go further, the practitioner experience
DLT and therefore to an increase of the lead time factor. could help the metaheuristic to converge rapidly
For example, the mean lead time factor for instances with towards better solutions. Additionally, the suggested
a standard deviation of 100 is 23.5 while it is of 49.6 for approach could be compared with the DDMRP pro-
instances with a standard deviation of 1000. prietary software available in the market.
Even if Ptak’s rules are confirmed by our exper- • The problem could be tackled by means of a learn-
iments, these rules do not give exact values of the ing approach. Indeed, in case where demand is not
optimal parameters which justifies the relevance of known in advance and cannot be predicted, a learn-
using an optimisation algorithm for the parametrisation ing algorithm can learn how to parameterise DDMRP
of DDMRP. based on demand data that is fed into the algorithm
on a daily basis.
• No capacity constraints are considered in this paper.
It could be an interesting future research direction to
7. Conclusion and perspectives
consider that the resource is limited. In this case, pri-
A new simulation-optimisation method is presented orities between replenishment orders would need to
in this paper to optimise the parameters of DDMRP. be managed.
Eight relevant parameters are identified. NSGA-II (Non- • Other types of costs like ordering or transport costs
dominated Sorting Genetic Algorithm) is coupled with may be considered in addition to inventory costs.
the DDMRP simulation algorithm. Two objective func-
tions are optimised simultaneously: the On-Time Deliv-
ery (OTD) and the average on-hand inventory.
Acknowledgments
The suggested approach is tested on randomly- The authors acknowledge the support received from the indus-
generated instances. It is able to find non-dominated trial chair FM logistic.
solutions for all these instances within a few minutes.
From a managerial point of view, the developed tool
Data availability statement
can be used by the decision-maker on the history of
the previous year to fix the parameters for the following The data that support the findings of this study are avail-
year. Alternatively, the tool can be used on the predicted able from the corresponding author, Y.L., upon reason-
demand data of the following month to fix the parameters able request.
during this same month.
The considered problem is new and offers many
research perspectives: Disclosure statement
No potential conflict of interest was reported by the author(s).
• In the present study, the impact of uncertainty in
demand is neglected. It could be interesting to con-
sider a noisy demand that is given alongside with a Notes on contributors
probability distribution. Stochastic optimisation can David Damand is currently associate pro-
be investigated for this perspective. One can also fessor at EM Strasbourg Buisness School
adapt previous works for MRP in the presence noisy in Strasbourg (France). He is member
of Humanis laboratory. David Damand
demand: Thevenin, Adulyasak, and Cordeau (2021),
heads the FM Logistic research chair and
Kovačić, Usenik, and Bogataj (2017) and Hnaien, the supply chain management master’s
Delorme, and Dolgui (2010). degree at EM Strasbourg. His research
• A mathematical model approach could also be inves- focuses on inventory management and
tigated for the problem which involves a linearisation facility layout.
effort. Only one objective function should be cho- Youssef Lahrichi is currently post-doctoral
sen for the mathematical model, which could be the researcher at EM Strasbourg Buisness
average on-hand inventory while the On-Time Deliv- School in Strasbourg (France). He holds
a Ph.D. degree from Université Clermont
ery could be forced to 100% with a constraint. The
Auvergne (France). His research focuses
mathematical model could help to evaluate the opti- on applications of operational research
mality gap of the 100% OTD solutions presented in models for problems of logistics, transport
this paper. or production.
20 D. DAMAND ET AL.

Marc Barth is currently is associate pro- Literature Review and Research Issues.” In 13ème Conférence
fessor in Industrial Engineering at INSA Internationale de Modélisation, Optimisation et Simulation
of Strasbourg, France. He is member of (MOSIM2020), Agadir, Maroc.
HUMANIS Laboratory of EM Strasbourg. Favaretto, Daniela, and Alessandro Marin. 2018. An Empiri-
His research addresses the applications of cal Comparison Study Between DDMRP and MRP in Mate-
production planning and control, theory rial Management. Department of Management, Università
of inventive problem solving and applica- Ca’Foscari Venezia Working Paper, (15).
tion of ABC and TDABC in the area of Goldratt, Eliyahu M. 2017. Critical Chain: A Business Novel.
design of production systems and warehouse. London: Routledge.
Hnaien, Faicel, Xavier Delorme, and Alexandre Dolgui.
References 2010. “Multi-objective Optimization for Inventory Con-
trol in Two-level Assembly Systems Under Uncertainty of
Abdelhalim, Achergui, Allaoui Hamid, and Hsu Tiente. 2021. Lead Times.” Computers & Operations Research 37 (11):
“Demand Driven Material Requirements Planning Buffer 1835–1843.
Positioning Considering Carbon Emissions.” In IFIP Inter- Huq, Ziaul, and Faizul Huq. 1994. “Embedding JIT in MRP:
national Conference on Advances in Production Management The Case of Job Shops.” Journal of Manufacturing Systems
Systems, 460–468. Springer. 13 (3): 153–164.
Amorim, Pedro, Carlos H. Antunes, and Bernardo Almada- Ihme, Mathias. 2015. “Interpreting and Applying Demand
Lobo. 2011. “Multi-objective Lot-sizing and Scheduling Driven MRP: A Case Study.” PhD diss., Nottingham Trent
Dealing with Perishability Issues.” Industrial & Engineering University.
Chemistry Research 50 (6): 3371–3381. Jiang, Jingjing, and Suk-Chul Rim. 2016. “Strategic Inven-
Azzamouri, Ahlam, Pierre Baptiste, Guillaume Dessevre, and tory Positioning in BOM with Multiple Parents Using
Robert Pellerin. 2021. “Demand Driven Material Require- ASR Lead Time.” Mathematical Problems in Engineering
ments Planning (DDMRP): A Systematic Review and Classi- 2016.
fication.” Journal of Industrial Engineering and Management Jiang, Jingjing, and Suk-Chul Rim. 2017. “Strategic Wip
14 (3): 439–456. Inventory Positioning for Make-to-order Production with
Bahu, Baptiste, Laurent Bironneau, and Vincent Hovelaque. Stochastic Processing Times.” Mathematical Problems in
2019. “Compréhension Du DDMRP Et De Son Adoption: Engineering 2017.
Premiers éléments Empiriques.” Logistique & Management Katoch, Sourabh, Sumit Singh Chauhan, and Vijay Kumar.
27 (1): 20–32. 2021. “A Review on Genetic Algorithm: Past, Present,
Bayard, Stéphanie, and Frédéric Grimaud. 2018. “Enjeux and Future.” Multimedia Tools and Applications 80 (5):
Financiers de DDMRP: Une Approche Simulatoire.” In 12e 8091–8126.
Conférence Internationale de Modélisation, Optimisation et Kortabarria, Alaitz, Unai Apaolaza, Aitor Lizarralde, and
SIMulation-MOSIM’18. Itxaso Amorrortu. 2018. “Material Management Without
Bennett, Nathan, and G. James Lemoine. 2014. “What a Differ- Forecasting: From MRP to Demand Driven MRP.” Jour-
ence a Word Makes: Understanding Threats to Performance nal of Industrial Engineering and Management 11 (4):
in a VUCA World.” Business Horizons 57 (3): 311–317. 632–650.
Damand, David, Ridha Derrouiche, Marc Barth, and Samia Kovačić, Danijel, Janez Usenik, and Marija Bogataj. 2017.
Gamoura. 2019. “Supply Chain Planning: Potential Gener- “Optimal Decisions on Investments in Urban Energy Cogen-
alization of Parameterization Rules Based on a Literature eration Plants–Extended MRP and Fuzzy Approach to the
Review.” Supply Chain Forum: An International Journal 20 Stochastic Systems.” International Journal of Production Eco-
(3): 228–245. nomics 183: 583–595.
Deb, Kalyanmoy, Amrit Pratap, Sameer Agarwal, and T. A. Lee, Carmen Kar Hang. 2018. “A Review of Applications of
M. T. Meyarivan. 2002. “A Fast and Elitist Multiobjective Genetic Algorithms in Operations Management.” Engineer-
Genetic Algorithm: NSGA-II.” IEEE Transactions on Evolu- ing Applications of Artificial Intelligence 76: 1–12.
tionary Computation 6 (2): 182–197. Lee, Chan-Ju, and Suk-Chul Rim. 2019. “A Mathematical Safety
Dessevre, Guillaume, Pierre Baptiste, Jacques Lamothe, and Stock Model for DDMRP Inventory Replenishment.” Math-
Robert Pellerin. 2021. “Visual Charts Produced by Simu- ematical Problems in Engineering 2019.
lation to Correlate Service Rate, Resource Utilization and Martin, Guillaume. 2020. “Contrôle Dynamique du Demand
DDMRP Parameters.” International Journal of Production Driven Sales and Operations Planning.” PhD diss., Ecole des
Research 1–13. Mines d’Albi-Carmaux.
Dessevre, Guillaume, Guillaume Martin, Pierre Baptiste, Miclo, Romain. 2016. “Challenging the” Demand Driven MRP”
Jacques Lamothe, Robert Pellerin, and Matthieu Lauras. Promises: A Discrete Event Simulation Approach.” PhD
2019. “Decoupled Lead Time in Finite Capacity Flowshop: A diss., Ecole des Mines d’Albi-Carmaux.
Feedback Loop Approach.” In 2019 International Conference Miclo, Romain, Matthieu Lauras, Franck Fontanili, Jacques
on Industrial Engineering and Systems Management (IESM), Lamothe, and Steven A. Melnyk. 2019. “Demand Driven
1–6. IEEE. MRP: Assessment of a New Approach to Materials Manage-
Dolgui, Alexandre, and Caroline Prodhon. 2007. “Supply Plan- ment.” International Journal of Production Research 57 (1):
ning Under Uncertainties in MRP Environments: A State of 166–181.
the Art.” Annual Reviews in Control 31 (2): 269–279. Ohno, Taiichi, and Norman Bodek. 2019. Toyota Production
El Marzougui, M., N. Messaoudi, W. Dachry, H. Sarir, and B. System: Beyond Large-scale Production. New York: Produc-
Bensassi. November 12–14, 2020. “Demand Driven MRP: tivity press.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 21

Orlicky, Joseph A. 1974. Material Requirements Planning: The In Design and Modeling of Mechanical Systems-II, 139–149.
New Way of Life in Production and Inventory Management. Springer.
New York: McGraw-Hill, Inc.
Pekarčíková, Miriam, Peter Trebuňa, Marek Kliment, and Jozef
Trojan. 2019. “Demand Driven Material Requirements Plan- Appendix. (Non-linear) mathematical model for
ning. Some Methodical and Practical Comments.” Manage- the paremeterisation of DDMRP
ment and Production Engineering Review 10: 50–59.
Ptak, Carol, and Chad Smith. 2011. Orlicky’s Material Require-
Data
ments Planning. New York: McGraw-Hill Education. • The planning horizon: H = {0, . . . , hmax }
Ptak, Carol, and Chad Smith. 2016. Demand Driven Mate- • The demand quantities over the days of the planning hori-
rial Requirements Planning (DDMRP): Version 2. New York: zon: {dk ; k ∈ H}
McGraw-Hill Education. • The decoupled lead time: DLT
Rezaei, Jafar, Mansoor Davoodi, Lori Tavasszy, and Mohsen • The initial on-hand inventory: s0
Davarynejad. 2016. “A Multi-objective Model for Lot-sizing • A very large number: M
with Supplier Selection for An Assembly System.” Interna-
tional Journal of Logistics Research and Applications 19 (2):
125–142. Continuous variables
Rim, Suk-Chul, Jingjing Jiang, and Chan Ju Lee. 2014. “Strategic • The lead time factor: 0 ≤ FLT ≤ 1
Inventory Positioning for MTO Manufacturing Using ASR • The variability factor: 0 ≤ FV ≤ 1
Lead Time.” In Logistics Operations, Supply Chain Manage- • The order spike threshold: 0 ≤ Tpeak ≤ 1
ment and Sustainability, 441–456. Springer. • The top of red: 0 ≤ TOR
Shofa, Mohamad J., Armand Omar Moeis, and N. Restiana. • The top of yellow: 0 ≤ TOY
2018. “Effective Production Planning for Purchased Part • The top of green: 0 ≤ TOG
Under Long Lead Time and Uncertain Demand: MRP Vs • The on-hand inventory: qon -hand , t ∈ H
t
Demand-Driven MRP.” In IOP Conference Series: Materials -order , t ∈ H
• The on-order inventory: qon t
Science and Engineering, Vol. 337. IOP Publishing.
Shofa, Mohamad Jihan, and Wahyu Oktri Widyarto. 2017. • The net flow: qt , t ∈ H
flow

“Effective Production Control in an Automotive Industry: • The qualified demand: d̃t , t ∈ H


MRP vs. Demand-Driven MRP.” In AIP Conference Proceed- • The peak horizon: Hpeak
ings, Vol. 1855. AIP Publishing LLC. • The order amount: ot , t ∈ H
Silver, Edward A., and C. Harlan. 1973. “Meal. A Heuristic for • The incoming supply at day t: õt , t ∈ H
Selecting Lot Size Quantities for the Case of a Deterministic • The average on-hand inventory: Qon-hand
+ −
Time-varying Demand Rate and Discrete Opportunities for • The ADU future and past horizons: HADU , HADU
Replenishment.” Production and Inventory Management 14 • The desired minimum order quantity: MOQ
(2): 64–74. • The order cycle: Horder
Talbi, El-Ghazali. 2015. “Hybrid Metaheuristics for Multi- • The average on-hand inventory: Qon-hand
objective Optimization.” Journal of Algorithms & Computa-
tional Technology 9 (1): 41–63.
Binary variables
Talbi, El-Ghazali, Matthieu Basseur, Antonio J Nebro, and
Enrique Alba. 2012. “Multi-objective Optimization Using 
1 If demand of day k exceeds the order spike threshold
Metaheuristics: Non-standard Algorithms.” International zk =
0 Otherwise.
Transactions in Operational Research 19 (1-2): 283–305. 
Thevenin, Simon, Yossiri Adulyasak, and Jean-François 1 If a replenishment order should be launched on day t.
rt =
Cordeau. 2021. “Material Requirements Planning Under 0 Otherwise.
Demand Uncertainty Using Stochastic Optimization.” Pro-
duction and Operations Management30 (2): 475–493.
Thürer, Matthias, Nuno O. Fernandes, and Mark Stevenson.
2022. “Production Planning and Control in Multi-stage Objective
Assembly Systems: An Assessment of Kanban, MRP, OPT
(DBR) and DDMRP by Simulation.” International Journal of Min {Qon-hand }
Production Research 60 (3): 1036–1050.
Velasco Acosta, Angela Patricia, Christian Mascle, and Pierre Constraints
Baptiste. 2020. “Applicability of Demand-Driven MRP in a
Complex Manufacturing Environment.” International Jour- -hand = s
qon
0 0 (A1)
nal of Production Research 58 (14): 4233–4245.
Wu, Horng-Huei, Amy H. I. Lee, and Tai-Ping Tsai. 2014. “A qon -hand = qon -hand + õt−1 − dt−1 ∀t ∈ H − {0} (A2)
t t−1
Two-level Replenishment Frequency Model for TOC Supply
qon -order = 0 (A3)
Chain Replenishment Systems Under Capacity Constraint.” 0
Computers & Industrial Engineering 72: 152–159. -order = qon-order + o − õ
qon
t t−1 t−1 t−1 ∀t ∈ H − {0} (A4)
Yahia, Wafa Ben, Houssem Felfel, Omar Ayadi, and Faouzi 
Masmoudi. 2015. “Comparative Study for a Multi-objective k∈H 1[t−HADU
− +
,t+HADU ] (k).dk
MLCSP Problem Solved Using NSGA-II & E-Constraint.” ADUt = + − ∀t ∈ H (A5)
HADU + HADU + 1
22 D. DAMAND ET AL.

TORt = ADUt × DLT × FLT + ADUt qflow = qon -hand + qon-order − d̃ ∀t ∈ H (A15)
t t t t
× DLT × FLT × FV ∀t ∈ H (A6) TOYt − qflow ≤ M.rt ∀t ∈ H (A16)
t
TOYt = TORt + ADUt × DLT ∀t ∈ H (A7)
M.(rt − 1) ≤ TOYt − qflow
t ∀t ∈ H (A17)
TOGt = TOYt + Max{ADUt × DLT
0 ≤ ot ≤ TOGt − qflow
t + M.(1 − rt ) ∀t ∈ H (A18)
× FLT , MOQ, Horder } ∀t ∈ H (A8)
M.(rt − 1) + TOGt − qflow ≤ ot ≤ M.rt ∀t ∈ H (A19)
0 ≤ FLT ≤ 1 (A9) t
ot−DLT = õt ∀t ∈ H; t ≥ DLT (A20)
0 ≤ FV ≤ 1 (A10)
 õ0 = · · · = õDLT−1 = 0 (A21)
d̃t = dt + 1[t+1,t+Hpeak ] (k).zk .dk ∀t ∈ H (A11)
-hand
k∈H dt ≤ qon
t + õt ∀t ∈ H (A22)
hmax on-hand
dk − TORt × Tpeak ≤ zk .M ∀t, k ∈ H (A12) q
Qon-hand = t=0 t (A23)
(zk − 1).M ≤ dk − TORt × Tpeak ∀t, k ∈ H (A13) hmax

0 ≤ Tpeak ≤ 1 (A14) Note: 1A (x) is the indicator function, it equals one if x ∈ A


and 0 otherwise.

You might also like