Computers & Industrial Engineering 148 (2020) 106697

Contents lists available at ScienceDirect

Computers & Industrial Engineering

journal homepage: www.elsevier.com/locate/caie

Optimal stock control and procurement by reusing of obsolescences in

manufacturing
Emanuele Frontoni, Fabrizio Marinelli, Roberto Rosetti, Primo Zingaretti ∗
Dipartimento Ingegneria dell’Informazione - Università Politecnica delle Marche, Via Brecce Bianche - 60131 Ancona, Italy

ARTICLE INFO ABSTRACT

Keywords: Production, procurement planning and logistic organization are complex tasks in companies with multiple
Production planning production/stocking sites. For companies operating in fashion markets, these tasks are harder due to the high
Distributed production system demand variability and frequent turnover of trends that make products outmoded rapidly.
Obsolete component reusing
In the field of industry informatics under an Industry 4.0 scenario, this paper presents a data and processing
Decision support system
framework that mix a Decision Support System (DSS) and a Mixed Integer Linear Programming (MILP)
Mixed integer linear programming
Fashion industry
model to maximize the potential revenue of the outmoded products. The solution proposed considers both
the investment for missing components and the transportation costs between hubs and takes into account
constraints on the overall number of items produced, the budget for new component orders and the minimum
lot scheduling thresholds. Moreover, to avoid the increment of obsolescences due to raw material orders, the
MILP model bounds purchased components to a percentage of used ones. A real worldwide dataset provided
by a fashion company, market leader in the design, manufacture and distribution of fashion, luxury, sports
and performance eyewear, was used to test performances and repeatability of the mathematical model under
different configurations. The results obtained show a huge positive impact on the financial results.

1. Introduction be implemented planning the production of outmoded items for paral-

Fashion oriented trading markets lead and govern style tendencies,
apparel trends and all related products affecting the wearing industry.
Fashion markets are not only related to cloths and shoes, but a consid-
erable market share regards the fashion accessories like bags, jewellery,
hats, sunglasses, etc. In particular, eyewears are some of those fashion
products that better combine the engineered side with the stylish side of
an industrial product (Montalto, Graziosi, Bordegoni, & Landro, 2018).
Despite appearances, this kind of accessory can have more than 20
different components in its bill-of-material (BOM), making it a small
and complex engineering product. Eyewear market is inside a ‘‘fast’’
industry with high demand variability and frequent turnover of trends
that makes products outmoded rapidly.
Despite the efforts to forecast production and procurement phases,
other factors, like delays in suppliers’ deliveries (Chung, Talluri, &
Kovács, 2018) and human errors in procurement management, can
overfill the company’s warehouses of components and raw materials
that remain unused. A no longer fashionable product, in fact, drags
a set of components with it, making these ones no more useful for
mainstream production. Components related to a discontinued product
are called obsolescences. The reusing of this obsolete components can
lel markets, such as outlets and second markets, where the products
will be sold at a lower price. The manufacturing of products for
the second markets normally uses only obsolescences, but sometimes
also components originally intended for different products (component
commonality Takai & Sengupta, 2017). Outlets and second markets
work mainly during sale events like Black Fridays, local special events
(openings, local feasts, etc.) and season sales where higher trades and
advantageous prices create a demand that can easily absorb all the
production in the case study proposed.
This work proposes a framework of ’Data Integration’ that aims to
combine the production, procurement and logistics departments of a
company. One of the main objectives of the framework is optimal stock
control. This is directly connected to the maximization of profit and
indirectly to the minimization of obsolescences, which obviously are
the core business of any second market. To achieve both these goals we
chose to adopt a MILP model, which we considered a suitable approach
for the problem treated.
In a preliminary work [anonymous reference], the authors only
considered stored components and finished product BOMs in order
to find the best production and procurement planning to maximize

∗ Corresponding author.
E-mail addresses: e.frontoni@univpm.it (E. Frontoni), fabrizio.marinelli@univpm.it (F. Marinelli), r.rosetti@univpm.it (R. Rosetti), p.zingaretti@univpm.it
(P. Zingaretti).

https://doi.org/10.1016/j.cie.2020.106697
Received 24 July 2019; Received in revised form 22 July 2020; Accepted 24 July 2020
Available online 4 August 2020
0360-8352/© 2020 Elsevier Ltd. All rights reserved.
E. Frontoni et al.
financial revenues. This paper considerably expands that work. First, by
completely introducing a DSS module that allows managers to choose
solutions and simulate their impact before implementing the activities.
Second, by taking into account the competitive prices of suppliers and
transportation handling costs, resulting in a more accurate mathemat-
ical model. Third, a dashboard for visualizing dynamic analytics and
reports to get the trading-off solutions that best fit the needs of the
company.
This project has been developed under the influence of Industry
4.0 guidelines with a seamless integration of the new ‘‘Operational
Research" (OR) module with the existing framework.
In the field of industry informatics, the application of a novel
data and processing framework to one of the most complex scenarios
available in the fashion market is one of the main contributions of the
proposed approach together with the novel OR model to maximize the
potential revenue from reusing obsolete components.
From an academic point of view, the proposed framework and
the global integration of OR and DSS in a real and complex world
production scenario is a new and effective solution, which can be
adapted to many other industrial processes. From a practical point of
view, managers can take advantage of the different results proposed by
the system before taking their decision.
This introduction continues with the description of a worldwide
case study with real data and of ad hoc strategies within an Industry
4.0 scenario.
1.1. The worldwide case study
A real worldwide case study related to eyewear manufacturing is
investigated in this section, focusing on embedding an optimal solution
for obsolete component reusing within the existing framework. The
case study was approached for a prominent company that has an
overall production of 93 millions of prescription glasses and sunglasses
and counts many manufacturing facilities, distribution centres and a
wide network of retail stores spread worldwide. Company hubs can be
used for both the production of glasses and the storage of components
and the orders of parts can be done from each hub using a local
procurement network.
The cost of stocked and ordered components varies according to
component’s type and location. Transfer costs of raw materials have
to be managed in order to enhance profits and to increase savings. The
established management policy for the mainstream production consists
of weekly scheduled transfers of materials between company’s hubs.
The transfer of components proposed in this work is done according to
the scheduled transfers, amortizing the fixed cost for a potential new
cargo transfer.
Data used to solve this problem are complex, semi-structured and,
if we consider a global company test case, big in term of number of
records. The solution to be applied in a real scenario is supposed to be
fast, optimal and able to assist actions and decisions in a worldwide
and multi-user scenario. For these reasons we propose a modern data
architecture in an industrial informatics scenario able to mix data
analytics and OR tool.
In the following the words product, finished good, glasses and item, as
well as component, part and raw material are interchangeable.
1.2. Challenges in fashion markets and industrial strategy
Fashion products follow trends that can be imposed or offered
to customers and, in both cases, the forecast demand cannot be as
expected due to trend variability or bad response to the products
proposed. Since the products have a high turnover, some items go
rapidly out of market, becoming ‘‘out of fashion’’. All the compo-
nents related to outmoded products are called obsolescences. The high
demand variability can lead to an overstock of material giving two
main drawbacks: first, stoked obsolescences increase the management
2
Computers & Industrial Engineering 148 (2020) 106697
costs and secondly they represent frozen assets that cannot be sold
diversely, risking to be lost. A largely adopted strategy in the using
of obsolescences is to produce outmoded products that will be sold
in the secondary market. Indeed, the selling markets can be generally
divided in two main branches: first market and secondary/outlet market.
The former includes all the current products that follow the trend; such
products are considered top sellers and are pushed by the company.
The latter has all the out of fashion products, which companies sell at
a discounted price with respect to the original list price. The use of
secondary markets, a practice commonly adopted also for, e.g., elec-
tronic components and pharmaceutical supply chains (Angelus, 2011),
guarantees the reduction of storage costs, a potential increase in profits
and the appreciation of obsolete components. Nevertheless, the produc-
tion of outmoded articles generally requires the purchasing of some
components of the BOM that are not overstocked as obsolescences.
Besides, a heavy presence on the secondary markets is not the core
business of the company and a flood of out of fashion products can
even damage the image of the brand. As a consequence, the simple
strategy of reusing as far as possible the obsolescences is not viable for
the company that instead is looking for a sort of equilibrium between
a suitable first and secondary production mix and the recovery of
the obsolete components value. The proposed decision support system
aims at maximizing the potential revenue at the net of (purchased
or stocked) component costs and of transfer and manufacturing costs
necessary for the production.
The paper is organized as follow. Section 2 reports a brief literature
review along with discussion on the meaning of the word ‘‘reusability’’.
Section 3 describes the software architecture. Section 4 presents the
issue related to the reusing of obsolete components, the mathematical
formulation of the problem and a discussion about a bounding con-
straint introduced in the MILP. Section 5 describes the data formalism
used in the tests performed and the results obtained. A deep analysis
of the results is reported in Section 5.1 while an assessment of the
mathematical program proposed is presented in 5.2. Conclusions and
future works are presented in Section 6.
2. Literature review
Many commercial softwares are now considered standard for solv-
ing part or all of the many problems that companies have to face. The
logistics staff often can use a warehouse management system (WMS), a
database application software, to improve the efficiency of warehouses,
but in many case custom solution approaches are necessary to fit
particular situations or specific requirements to keep stock levels under
control (Antosz & Ratnayake, 2019).
On the other hand, big companies embed specific tools dedicated to
resource handling into their management software. The case study con-
sidered in this paper requires a novel processing framework including
a dedicated tool, in particular to manage stock levels of components
related to outmoded products, typical of the fashion markets.
In the following we analyse the literature approaches related to our
solution and proposed mainly in the context of production, resource,
and logistics management. At the best of the authors’ knowledge, there
are no works focusing on the reusing or appreciation of obsolescences
as defined in this paper. Besides, none of the following cited papers
is based on a modern big data architecture able to mix data analytics
and result optimization for supporting decision making and action in
complex logistic and procurement systems.
2.1. Reusability and secondary markets
The term reusing, as well as commonality and remanufacturing, is
broadly used with several meanings in the management science lit-
erature. A particular attention has been paid in the recent few years
to reverse logistics and closed-loop supply chains where a flow of
used products (consumer, end-of-use or end-of-life returns, Jena &
E. Frontoni et al.
Sarmah, 2016) are returned from customers to facilities. The reusing in
reverse logistics is mainly conceived as the recycling of raw material,
recovering of spare parts, or remanufacturing, that is the process of
restoring used products to ‘‘likenew" condition by disassembly, re-
placing parts, and reassembly operations. Most contributions describe
heuristic policies (based on mathematical programming) rather than
optimal approaches, since the uncertainty in the flow of returned parts
adds up to the uncertain customer demands and greatly impacts on the
difficulty of the inventory management and assembly systems (Calmon
& Graves, 2017; Fu, Gong, & Liang, 2019; Lei, Zhong, & Chaovalit-
wongse, 2009). The two following papers are those that deal with
problems more similar to our one, despite our meaning of ‘‘reusing"
(as defined in the introduction) is quite different. Both Fang, Lai, and
Huang (2017) and Phuc, Yu, and Tsao (2016) approach the production
planning of industrial commodities considering the appreciation of
discontinued item’s parts. In Fang et al. (2017) the authors analyse the
hybrid production system where new and remanufactured products are
assembled at the same time. Either by reusing parts of old products or
by refurbishing them, the company improves his environment-friendly
public image by enhancing his selling probability and respecting gov-
ernment for recycling. In Phuc et al. (2016) a reverse supply chain
(RSC) is designed to better exploits end-of-life vehicles (ELV). Conform-
ing to the ELV classification, the authors propose a multi-product RSC
modelled as a mixed-integer programming problem able to minimize
the total cost of the RSC network.
The concept of reusability is also used in the context of mass
customization, and in particular in the platform-based product-family
design (Vié, Zufferey, & Cordeau, 2019). The rationale beneath such
practice consists in pursuing at the same time the reduction of de-
sign/production costs and the increase of product variety to better
meet the diverse market niches. Product-family design is generally
implemented by defining a family of products based on a backbone of
hidden common components (the platform) and a set of configurable
parts. Here, the reusing of parts mainly refers to the ability, at the
design stage, of reusing common elements across product variants.
Different meanings of reusability also occur in make-or-buy de-
cisions, e.g., in the selection of Off-the-Shelf software components
(COTS) (Cortellessa, Marinelli, & Potena, 2008) and in manufactur-
ing process optimization, e.g., when leftovers resulting from cutting
operations are stocked back into the raw material warehouse (Cherri,
Arenales, Yanasse, Poldi, & Vianna, 2014) or are reused to assemble
final products (Arbib, Di Iorio, Marinelli, & Rossi, 2002).
Finally, also the term secondary market is frequently used to describe
a different scenario as that considered in this paper. Indeed, it usually
refers to the markets where resellers, after receiving the products from
the manufacturer (primary market), trade among themselves in order
to dispose excess inventories (Kumar & Aouam, 2019; Lee & Whang,
2002).
2.2. Solution approaches
Mathematical programming is a powerful and effective modelling
language for describing complex problems in the wildly broad and
varied context of production planning and scheduling, e.g., see Pochet
and Wolsey (2006) and Pinedo (2009).
Just to give a few examples, in Ramezanian, Rahmani, and Barz-
inpour (2012) the authors model the Aggregate Production Planning
(APP) problem with multi-period, multi-product and multi-machine
systems through a mixed integer linear program (MILP) and propose
a genetic algorithm and a tabu search heuristic to address realistic
instances. Similarly in Wang and Yeh (2014) the authors studied the
APP problem in the general and discrete versions, modelling it by
ILP, but proposing a modified particle swarm optimization algorithm.
In Han, Lee, and Kim (2019), the authors propose an MILP based
heuristic able to minimize the total production costs in two level supply
chain with multiple suppliers and manufacturing plants. In Rasmi,
3
Computers & Industrial Engineering 148 (2020) 106697
Kazan, and Türkay (2019) the authors explore a more complex APP
problem, which considers many social and environmental aspects. The
resulting quadratic objective MILP (QMILP) model has been solved with
the Generator of ND and Efficient Frontier (GoNDEF) method to find
all (or a large part of) the non dominated solutions.
In Yaghin (2020) a fashion-industry case study, where a multi-
period, multi-product, multi-site, multi-sale channel APP problem is
studied. The problem is formulated with a geometric programming
model and then converted in a convex programming model requiring
a non linear solver.
A methodological framework for production and transportation is
presented in Liotta, Kaihara, and Stecca (2016). The authors propose
a framework that combines optimization and simulation techniques to
increase profit and reduce CO2 , showing that random variability of
demand plays an important role on production costs and freight per-
formance. As in this paper, the authors identify the shipment plans and
the transportation schedule according to BOM lists, different suppliers’
prices and production/transportation costs.
In most of the above works, mathematical programming is merely
used as a description language and the problems are eventually solved
by heuristic approaches. Indeed, the chance to optimally solve the
problem by means of general solvers deeply depends from the type of
program worked out. Algorithms for non-linear mathematical programs
(NLP) generally are neither robust nor scalable: local NLP solvers guar-
antee global optimality on the vast majority of practical convex NLP
problems with up to thousands of variables and constraints, whereas
the same solvers typically has many more chances of failures on non-
convex NLPs. On the other hand, the most advanced MILP solvers are
able to solve in acceptable times practical cases with tens of thousands
of variables and/or constraints, although, being based on the Branch-
and-Bound exponential algorithm, it is very hard to make predictions
on the maximal size of manageable MILPs.
Despite each solution obtained via a DSS is peculiar, it is in-
teresting that DSS have also adopted for industrial applications and
specifically for managements problems. For example, in Saha, Aqlan,
Lam, and Boldrin (2016) the authors propose an End-to-End Customer
Order Management System, composed by three integrated tools, which
supports real-time order management, considering limited shared re-
sources and risk of missing orders, or in Galizia, ElMaraghy, Bortolini,
and Mora (2019) the authors propose a DSS for product platform
design.
In this paper, the details related to the implementation of the DSS
proposed have not been discussed and will be presented in a future
work.
3. The system architecture
Fig. 1 represents a simplified scheme of the company’s workflow.
The Extract-Transform-Load (ETL) module extracts data from a SAP
database and feeds the Repository module with pre-processed data,
useful for the revision and decision processes by the DSS and for
visualizing dynamic analytics and reports by a Kibana Dashboard. The
OR module interacts with the ETL module just for the management of
obsolete component reusing by harmonizing, as it will be shown in the
rest of the paper, all the production, procurement and logistic activities.
Main purposes of the Data Integration framework are: (i) data
linking of different sources (e.g., BOM and stock data); (ii) data filtering
process based on pre-defined rules and motivated by data clean-up; (iii)
data preparation for the OR modules.
Next two subsections describe in detail, respectively, the main
functionalities and characteristics of the software modules in the system
and an example of processing operations performed by the ETL module.
E. Frontoni et al.
Fig. 1. Integration of the ‘‘Operational Research’’ module within the data processing
workflow.
3.1. Software modules
Data are imported from a SAP platform, which provides the inte-
gration of information and processes, in addition to industry-specific
functionality and scalability.
The ETL module is a sophisticate toolset for organizing, manipu-
lating and collecting data. This software component makes data more
efficient, effective and reportable. ETL processes are fundamental being
responsible for: (i) the extraction of data from heterogeneous sources;
(ii) their transfer to the specific areas of the data warehouse where they
will be processed; (iii) the data transformation and the computation of
new values for obeying to the data warehouse structure to which they
are targeted; (iv) the isolation and cleaning of problematic tuples to
guarantee that database constraints and business rules are respected.
Most data integration techniques similar to the scenario described
above require significant preprocessing of the voluminous raw data,
which must be cleaned, filtered and organized into a useable format
for querying. Often, this step is a one-time effort because the goal
is to perform an ‘‘offline analysis’’, rather than to provide ongoing
support for interactive querying of ‘‘live’’ and growing online data.
Consequently, researchers often opt to store the processed data in a
well known RDBMS, such as MySQL or SQLite, much simpler to set
up and easier to query. While this kind of approach is likely the
best choice for some tasks, it is not well suited to the needs of Big
Data analysis, which requires supporting much larger volumes of data,
accommodating the real-time nature of incoming data and providing
quick responses to queries. In a traditional RDBMS, large data require
the creation of many indices to reduce the execution time of queries;
at the same time the presence of indices greatly slows the process of
data updating. Since neither of these choices is practical in presence
of Big Data, it is worthwhile to consider what other approaches might
work well in designing a repository. On the basis of our experience with
Elasticsearch (2020), an open source search engine that provides near
real-time search and full-text search capability, as well as a RESTful
API, we have used Elasticsearch to build a tool that collects data coming
from an ETL process and is able to mix structured and unstructured data
in a logistic and procurement scenario of a global brand. The main
characteristics is the ability of Elasticsearch to mix different sources
providing aggregated information about the process and, by a full set of
APIs, to bring back processed data in the production list or, in general,
in the ERP. The improvement in response time was remarkable, so
that it enabled users to interact with our analysis in real time. This
effectively turned our proof-of-concept research tool into a practical
industrial-strength OR and analytics tool.
The Dashboard module allows to review in a synthetic way both the
warehouse status and the part of the analytics with the performance
4
Computers & Industrial Engineering 148 (2020) 106697
Fig. 2. Example of an ETL process. The schema represents operations performed to
filter surplus data, before loading input data to the OR module. Other extracted tables
and fields unnecessary to complete these operations are not displayed.
of the management. The data visualization system is based on Kibana
(2020), an open source data visualization plugin for Elasticsearch. It
provides visualization capabilities on top of the content indexed on an
Elasticsearch cluster. Users can create bars, lines and scatter plots or
pie charts and maps on top of large volumes of data. The Dashboard
displays a set of saved visualizations, such as the projection of the
cost of materials provided and the procurements programmed by the
planner.
As it will be shown in Section 5, the OR module offers a range of
solutions that can be pareto-optimal according to some parameters (as
order cost, production, stock values used, etc.). The Decision Actuator,
exploiting the solutions proposed by the OR module, allows to simulate
how the adoption of each solution will change the situation of the
warehouse. It also generates the corresponding production orders and
the handling of components to be passed to the SAP database. In
other words, this module shows a set of outputs and their related
effects in terms of warehouse value used, order costs, production and
other computed parameters. An example of data analysis is give in
Section 5.1.
3.2. Example of an ETL process
As an example, Fig. 2 depicts the ETL process we implemented to
identify the surpluses in company’s warehouses, with the aim of surplus
reduction by optimizing the reusing of products no longer active.
Initially, the data extracted from the SAP database by the ‘‘Extract"
operation give the image of the warehouse, with all the available prod-
ucts and components and the active and non active products. Stocks
levels are recorded for each component in every warehouse, giving the
real state of the company’s productivity power. Synchronization issues
do not rise because the SAP system is able to provide real data of all
E. Frontoni et al.
the worldwide distributed warehouses which do not have high level
variability due to their nature (obsolete components). Each extracted
instance has real and accurate data.
In the ‘‘Transform" operation no more active products are identified
by Expiry Date (exceeded this date orders for that product are not
accepted) and Estimated End Of Life (date of production end).
The input data prepared for the OR module are lists of:
- all reproducible products in each division;
- BOMs of reproducible products;
- the minimum and maximum lots, profit margins and the index of
precedence for each reproducible product;
- surplus in each division for each useable material;
- production/purchase costs in each division for each useable ma-
terial.
Finally, the ‘‘Load" operation is executed, by which the ETL module
loads results of OR module in the Repository to analyse the impact of
optimization on the warehouse.
4. The OR module
The previous section has exposed in details the company IT struc-
ture and the work done for integrating the new framework within
the existing architecture. The OR module aims to propose a set of
solutions to increase profits, reduce obsolete stocks and improve logis-
tic/procurement management. This section presents the optimization
problem that lies beneath the DSS, formalizing the problem definition
and exposing the mathematical formulation used.
4.1. The obsolete part reusing problem
The problem of reusing obsolete parts can be formalized as follows.
A set 𝐶 of components (e.g., lens and screws) and materials (e.g., glue
and paint) are stored in a set 𝐷 of worldwide production/storage
hubs. Let 𝑀 be the set of outmoded glasses types that can be produced
for the second market with elements in 𝐶 according to their bill-
of-materials (BOM). Although the manufacturing of glasses generally
requires several raw materials, as paint and chemical compounds, in
the following description of the problem we assume for the sake of
conciseness and without loss of generality that all the BOMs consist of
discrete parts only; raw materials are addressed in the solution of the
real case (see Section 5) by adding continuous variables and fractional
parameters to the mathematical program. Also, in the interests of
providing fuller information, it is worth mentioning that all the BOMs
of the problem are 1-layer BOMs.
Every hub is able to produce any type of glasses. The transfer
of material between hubs is allowed as well as the purchase of new
components. Procurement costs and inventory values depend on the
hub and component type: for each component 𝑖 ∈ 𝐶 and hub ℎ ∈ 𝐷,
let 𝑝𝑖ℎ be the cost for buying component 𝑖 at the hub ℎ, and let 𝑐ℎ𝑖 and
𝑔ℎ𝑖 be respectively the inventory value and the quantity of component 𝑖
stocked in the hub ℎ. Since most of the stocked components are obsolete
is supposed to have many parts already in the warehouses making
useless any reference to safety stock levels. Furthermore, production
hubs are surrounded by a network of providers that closely collaborate
with the company according to a providing policy forecast of 3–6
months giving full support to company needs. Nevertheless, second
market production is pretty small compared to the overall production
(one twentieth of it) and scheduled in few events (6–7 in total) in
the whole year. This means that the second market production can
be accomplished almost effortless. Products sold in the second market
have a selling price of 𝑟𝑗 , 𝑗 ∈ 𝑀, that is ranged between 60% to 70% of
the reference retail price, i.e., the revenue of products sold in the main
market.
5
Computers & Industrial Engineering 148 (2020) 106697
The company has cargo flights for the main production, which
regularly move components among all production/storage hubs accord-
ing to a weekly schedule. Due to the quantity and the footprint of
the components the company deals with, no additional cargo flights
are necessary for the transportation of the obsolescences, so the fixed
transportation costs can be neglected. In our approach, indeed, trans-
portation costs are considered proportional to the cost 𝑚ℎ𝑘 for moving
a unit of any component between any pair of hubs ℎ and 𝑘.
For each BOM associated to a glasses type 𝑗 ∈ 𝑀, the parameters 𝑎𝑖𝑗
indicates the number of items of component 𝑖 ∈ 𝐶 needed to assemble
a unit of product 𝑗. The sizes of production lots and order stocks are
defined both for manufacturing and for procurement. Parameters 𝑞𝑗ℎ
and 𝑄𝑗ℎ define, respectively, the minimum and maximum quantity
of model type 𝑗 that can be manufactured at hub ℎ, while 𝑏𝑖ℎ is the
minimum size of the production or purchasing lot of component 𝑖 in
hub ℎ. On the overall, the production for the second market cannot
exceed a maximum number 𝑄 of glasses items and cannot involve a
budget greater than 𝐻 for the procurement or manufacturing of lacking
components.
All the above parameters are available for the OR module after the
extraction and transformation processes performed by the ETL module.
Parameters are summarized in the following list:
• 𝑀: set of outmoded glasses types that can be produced;
• 𝐶: set of components listed in the BOMs related to glasses types
in 𝑀;
• 𝐷: set of hubs;
• 𝑎𝑖𝑗 : number of parts of component 𝑖 ∈ 𝐶 required to produce an
item of glasses type 𝑗 ∈ 𝑀;
• 𝑔ℎ𝑖 : number of parts of component 𝑖 ∈ 𝐶 that are stocked in the
hub ℎ ∈ 𝐷;
• 𝑐ℎ𝑖 : inventory value of the component 𝑖 ∈ 𝐶 stocked in the hub
ℎ ∈ 𝐷;
• 𝑝𝑖ℎ : cost of component 𝑖 ∈ 𝐶 ordered/produced in the hub ℎ ∈ 𝐷;
• 𝑟𝑗 : second market selling price of the item 𝑗 ∈ 𝑀;
• 𝑞𝑗ℎ , 𝑄𝑗ℎ : minimum and maximum sizes of a lot of the glasses type
𝑗 ∈ 𝑀 produced in hub ℎ ∈ 𝐷;
• 𝑏𝑖ℎ : minimum size of a lot of component 𝑖 ∈ 𝐶 purchased and/or
produced by the hub ℎ ∈ 𝐷;
• 𝑚ℎ𝑘 : cost for moving a component (whatever type) from hub/
warehouse ℎ ∈ 𝐷 to hub/warehouse 𝑘 ∈ 𝐷;
• 𝑄: maximum overall number of items that can be produced;
• 𝐻: available budget for component purchasing or manufacturing;
• 𝜇: opportunity ratio between the purchasing/manufacturing cost of
required components and the value of obsolete parts reused.
The last three parameters control the overall production of obso-
lete glasses types addressed to the second market. In particular, the
opportunity ratio 𝜇 limits the value of ordered and manufactured
components (plus the moving cost) to a percentage of the value of
reused components stocked in the hubs, attaining in such a way a trade-
off between the opportunity of reusing obsolete assets and the necessity
of commit further financial resources to achieve that, see Section 4.2
for a detailed discussion.
The proposed integer linear program (1)–(13) works on the follow-
ing decision variables:
• 𝑥𝑗ℎ : integer number of items of type 𝑗 ∈ 𝑀 produced in hub
ℎ ∈ 𝐷;
• 𝑦𝑗ℎ : binary variable equal to 1 if item 𝑗 ∈ 𝑀 is produced in hub
ℎ ∈ 𝐷, and 0 otherwise;
• 𝑧𝑖ℎ𝑘 : integer number of components 𝑖 ∈ 𝐶 ordered/produced in
the hub ℎ ∈ 𝐷 and used for production in hub 𝑘 ∈ 𝐷;
• 𝑢𝑖ℎ𝑘 : integer number of obsolete components 𝑖 ∈ 𝐶 stored in hub
ℎ ∈ 𝐷 and used for production in hub 𝑘 ∈ 𝐷;
• 𝑡𝑖ℎ : binary variable equal to 1 if component 𝑖 ∈ 𝐶 is pro-
duced/ordered in hub ℎ ∈ 𝐷, and 0 otherwise.
E. Frontoni et al. Computers & Industrial Engineering 148 (2020) 106697

The objective function (1) maximizes the total potential revenue by
taking into account the costs for purchasing and moving components
between hubs and the value of the stocked components used. Since one
of the main aims of our approach is to clear obsolete parts as soon as
possible, components that are used to manufacture outmoded glasses
types are considered valueless in the objective function; to this purpose
we introduce the parameter 𝑣𝑖ℎ , which is equal to 0 if component 𝑖 ∈ 𝐶
stored in the hub ℎ ∈ 𝐷 is obsolete, and corresponds to 𝑐ℎ𝑖 otherwise.
Moreover, we do not consider storage costs for obsolescences in the
objective function, because the additional storage/management costs
for obsolescence are negligible, since the proposed framework should
succeed in eliminating them quickly. Constraint (2) equals the number
of components used for production to those purchased or available
among the warehouses. Constraint (3) limits the stocked components
used to the quantity actually available in the hubs. The opportunity con-
straint (4) limits the value of purchased and manufactured components
(summed to the moving cost) to a percentage 𝜇 of the value of stocked
components reused for the production. Overall production of glasses
items is bounded by both the allowed maximum manufacturing volume
and the available budget for component purchasing, see constraint (5)
and (6), respectively. For each manufacturing plant, the minimum size
of manufactured lots of products is imposed by constraint (7), while
constraint (8) limits the overall production of each glasses type among
the whole supply chain. Finally, constraint (9) defines the minimum
and maximum sizes of components’ orders.
a constraint (2) of ≤ allows the fictitious reuse of obsolete parts that
pushes the production/purchase of new parts in the constraint (4).
Program (1)–(13), with 𝑂(|𝐶||𝐷|) constraints and 𝑂(|𝐶||𝐷|2 ) variables,
is not hub-wise decomposable and embeds the knapsack-like constraint
(6). Consequently, it is most likely not solvable efficiently, i.e., in a
CPU time polynomial with the size of the instance. Nevertheless, we
could use a commercial solver embedding a sophisticated implementa-
tion of the Branch-and-Bound algorithm. Such choice has been mainly
motivated by the requirement of fast prototyping and low development
costs. Moreover, the Branch-and-Bound has several nice properties: it
is very robust in general; it is an exact algorithm, i.e., able to guarantee
global optimality; it is of primal type, i.e., it moves within the space of
feasible solutions and therefore can also be used as a heuristic.
4.2. Opportunity constraint
In the following we show in details the meaning of parameter 𝜇
and how the opportunity constraint works. Indeed, the optimal solutions
of the mathematical model we propose represent a trade-off between
a suitable first and secondary production mix, on one side, and the
recovery of the obsolete components’ value, on the other side. This is
exactly what our worldwide prominent partner is looking for. To show
this behaviour consider a very simple instance consisting in just one
hub (|𝐷| = 1) and a single obsolete glasses type (|𝑀| = 1) whose
unitary revenue is 𝑟 = 20. Suppose moreover that:

∑ ∑ • the BOM of the product consists in three components {1, 2, 3} = 𝐶

max 𝑟𝑗 𝑥𝑗ℎ each required in a single part 𝑎1 = 𝑎2 = 𝑎3 = 1 (since |𝑀| = |𝐷| =
𝑗∈𝑀 ℎ∈𝐷 1 we omit the relevant indices for the sake of readability);
∑
− (𝑧𝑖ℎ𝑘 (𝑝𝑖ℎ +𝑚ℎ𝑘 )+𝑢𝑖ℎ𝑘 (𝑣𝑖ℎ +𝑚ℎ𝑘 )) (1) • the costs, inventory values and stock levels of the components are
𝑖∈𝐶 the following:
ℎ,𝑘∈𝐷
1. 𝑝1 = 2, 𝑝2 = 8, 𝑝3 = 3
∑ ∑
𝑎𝑖𝑗 𝑥𝑗𝑘 = (𝑧𝑖ℎ𝑘 +𝑢𝑖ℎ𝑘 ) ∀𝑖 ∈ 𝐶, ∀𝑘 ∈ 𝐷 (2) 2. 𝑐 1 = 2, 𝑐 2 = 3, 𝑐 3 = 2
𝑗∈𝑀 ℎ∈𝐷 3. 𝑔 1 = 0, 𝑔 2 = 10, 𝑔 3 = 5.
∑
𝑢𝑖ℎ𝑘 ≤ 𝑔ℎ𝑖 ∀𝑖 ∈ 𝐶, ∀ℎ ∈ 𝐷 (3) Suppose, moreover, that 𝑄 and 𝐻 are not limited (or however large
𝑘∈𝐷
∑ ∑ ∑ enough), 𝑞 = 0 and 𝑏1 = 𝑏2 = 𝑏3 = 0. The resulting simplified
𝜇 𝑐ℎ𝑖 𝑢𝑖ℎ𝑘 ≥ 𝑝𝑖ℎ 𝑧𝑖ℎ𝑘 + 𝑚ℎ𝑘 (𝑢𝑖ℎ𝑘 +𝑧𝑖ℎ𝑘 ) (4) mathematical program is the following:
𝑖∈𝐶 𝑖∈𝐶 𝑖∈𝐶 ∑
ℎ,𝑘∈𝐷 ℎ,𝑘∈𝐷 ℎ,𝑘∈𝐷 max 𝑟 ⋅ 𝑥 − (𝑧𝑖 𝑝𝑖 + 𝑢𝑖 𝑣𝑖 ) (14)
∑ 𝑖∈𝐶
𝑥𝑗𝑘 ≤ 𝑄 (5)
𝑗∈𝑀 𝑎 𝑥 = 𝑧𝑖 + 𝑢𝑖
𝑖
∀𝑖 ∈ 𝐶 (15)
𝑘∈𝐷
∑ 𝑢𝑖 ≤ 𝑔 𝑖 ∀𝑖 ∈ 𝐶 (16)
𝑝𝑖ℎ 𝑧𝑖ℎ𝑘 ≤𝐻 (6)
𝑖∈𝐶
ℎ,𝑘∈𝐷
plus the opportunity constraint
∑ ∑
𝑦𝑗𝑘 𝑞𝑗𝑘 ≤ 𝑥𝑗𝑘 ≤ 𝑦𝑗𝑘 𝑄𝑗𝑘 ∀𝑗 ∈ 𝑀, ∀𝑘 ∈ 𝐷 (7) 𝜇 𝑐 𝑖 𝑢𝑖 ≥ 𝑝𝑖 𝑧𝑖 (17)
∑ 𝑖∈𝐶 𝑖∈𝐶
𝑥𝑗𝑘 ≤ 𝑄𝑗𝑘 ∀𝑗 ∈ 𝑀 (8)
It is easy to see that the optimal value of (14)–(16) is 0 for 𝑟 ⋅ 𝑥 ≤
𝑘∈𝐷 ∑
∑ 𝐻
𝑖 𝑖 𝑖 𝑖
𝑖∈𝐶 (𝑧 𝑝 + 𝑢 𝑣 ), i.e., when the production of the obsolete glasses is
𝑏𝑖ℎ 𝑡𝑖ℎ ≤ 𝑧𝑖ℎ𝑘 ≤ 𝑖 𝑡𝑖ℎ ∀𝑖 ∈ 𝐶, ∀ℎ ∈ 𝐷 (9) uneconomic, whereas the model is unbounded otherwise, and the same
𝑘∈𝐷 𝑝ℎ
behaviour can be observed for any potential penalizing factor of the
𝑢𝑖ℎ𝑘 , 𝑧𝑖ℎ𝑘 ≥ 0, integer ∀𝑖 ∈ 𝐶, ∀ℎ, 𝑘 ∈ 𝐷 (10) terms 𝑧𝑖 𝑝𝑖 or 𝑢𝑖 𝑣𝑖 . In general, this means that the manufacturing of
products in 𝑀 is only limited by the budget 𝐻 and the maximum pro-
𝑥𝑗ℎ ≥ 0, integer ∀𝑗 ∈ 𝑀, ∀ℎ ∈ 𝐷 (11)
duction volume 𝑄, whereas the main overall objective of the company
𝑡𝑖ℎ ∈ {0, 1} ∀𝑖 ∈ 𝐶, ∀ℎ ∈ 𝐷 (12) is a suitable reuse of obsolescences (that in fact can be achieved by
setting the parameter 𝜇 to a suitable value). In particular, for several
𝑦𝑗ℎ ∈ {0, 1} ∀𝑗 ∈ 𝑀, ∀ℎ ∈ 𝐷 (13)
reasons (e.g., internal marketing policy, limited room for temporary
For each type of outmoded glasses, the solution of program (1)–(13) stocks, etc.) the company may not want to buy (even if it can) more
provides the production volume and location. Moreover, the stocked components than those it is going to use in the manufacturing of
materials can be picked up from any warehouse and the order of new obsolete items. A large production of obsolete items could be itself an
components is decided according to the price offered by each hub. issue: the supply of the secondary market is not the core business of
The movement of materials takes into account the price of labour cost the company; indeed, too many products in the secondary market can
of the hub. Observe that, although it is generally required that the even damages the brand.
number of manufactured and reused parts would at least be equal to When, instead, the opportunity constraint (17) is considered, the
the number of required parts, constraint (2) cannot be a ≤ constraint. In model is never unbounded and the optimal solution describes a trade-
fact, because the obsolete parts are valueless in the objective function, off between the cost of the new components that must be purchased

6
E. Frontoni et al.
Table 1
Parameter values for Configurations 1–3.
𝑄 𝑞𝑗𝑘 a 𝑄𝑗𝑘 a 𝑏𝑖𝑘 a
Configuration 1 unbounded 0 unbounded 0 100.000 units according to second market demand.
Configuration 2 100000 200 unbounded 200
Configuration 3 100000 200 500 200
a
∀𝑗 ∈ 𝑀, ∀𝑘 ∈ 𝐷, ∀𝑖 ∈ 𝐶.
to produce articles in 𝑀 and the recovering of value given by the
reusing of obsolete components. In particular, for 𝜇 = 1 the optimal
solution depletes the obsolete stocks (with a production of 10 items);
for 𝜇 = 0.8 the optimal solution leaves 2 items of component 2 in the
warehouse; for 𝜇 = 0.5 the optimal solution depletes only the stock of
component 2, and finally no new glasses items are produced for 𝜇 < 0.4.
The same effect cannot be achieved by penalizing the purchasing of
new components in the objective function. In particular, for value of 𝜇
greater than 1, that is, when the company is willing to spend more than
the value of reused obsolete parts (for example, in order to unburder
the warehouse), any penalized purchasing in the objective function can
be counter-productive.
Clearly, a single parameter does not catch all the potential trade-offs
arising in a so complex system, but the more parameters are used the
more complicated is the tuning process. A single parameter, instead, is
of very practical utility: in this case, 𝜇 expresses the commitment of
the company to recover the value of obsolete stocks, and this simple
concept is easy to understand and quantify by the management that
therefore appreciated this designing choice.
On the opposite, it is generally unclear on how to set up, for
example, the numerical value of penalizing factors in the objective
function.
5. Experimental setup and mathematical program assessment
The mathematical model has been implemented in AMPL (Takriti,
Fourer, Gay, & Kernighan, 1994) and solved by Cplex 12.8 (CPLEX
Optimizer, 2020). Tests have been performed on a desktop PC, running
Windows 7 Pro and equipped with an Intel Core i7 processor at 3.2 GHz
and 6 GB of ram. Considering the large values of optimal solutions (or-
der of millions of Euros), the absolute optimality gap tolerance (Cplex
option absmipgap) and the relative tolerance of integer variables
(Cplex option mipgap) have been set to 100 and 10−2 , respectively.
Moreover, a cpu time limit of 2 h has been set for each run.
The company has provided a real scenario instance consisting in
957 outmoded glasses types, 3457 components and 4 hubs. The orig-
inal dataset has been modified for privacy reasons and the public
version can be downloaded from Obsolescences dataset (0000). Com-
ponents’ cost and quantity have been provided as well as BOMs for
each outmoded glasses type. The fixed transportation cost has not been
considered due to the company policy, but a cost proportional to the
volume of the transported material has been estimated according to
labour cost of the country where the hub is located. Finally, although
the mathematical model includes the budget constraint (6), the com-
pany has not indicated any budget limit, mainly because its present
foremost objective is to empty the warehouse of the obsolete parts
regardless the budget required.
To evaluate both hypothetical and realistic scenarios, the three
parameter sets reported in Table 1 have been considered.
Configuration 1 refers to a hypothetical scenario where manufac-
turing and procurement issues (lot sizes) as well as secondary market
production volume and assortment are not considered. The aim of such
a scenario is just to evaluate the opportunity/potentiality of implement-
ing a reusing policy of obsolete parts. Configurations 2 and 3 represent
real scenarios where the parameters are set according to company’s
guidelines. For each production phase, the quantity to produce or the
budget for ordering components can vary according to the period of
7
Computers & Industrial Engineering 148 (2020) 106697
the year and these two configurations represent a viable framework
the company works with. As a consequence, many parameters result
in empirical settings. In particular, overall production has been set to
5.1. Data analysis
Each configuration was run 10 times, varying the parameter 𝜇 from
0.1 to 1.0 with a step of 0.1. Results are reported in Tables 2–7 (on
Configuration 2, the optimal solutions do not change for value of 𝜇
greater than 0.4). Columns obj value, opt gap, CPU time and b&b
nodes of Tables 2, 4 and 6 list the performance of the branch-and-
bound: respectively, the value of the objective function (in millions of
Euro), the percentage gap between the best integer solution and the
available upper bound (0.00 indicates optimal solutions or solutions
within the specified Cplex tolerance), the running times in sec, and the
size of the search tree in terms of number of visited nodes. The last
four columns of Tables 2, 4 and 6 indicate (in millions of Euro) the net
revenue 𝑅, the total component production/procurement costs 𝑃 , the
value 𝑈 of the reused inventory of which 𝑉 is the value of the obsolete
parts, and the total transportation costs 𝑇 .
Details of the solutions are reported in Tables 3, 5 and 7: the 3rd col-
umn reports the total number (in thousands) of items manufactured; the
2nd column lists the number of distinct glasses types produced (lots):
some of them, see the 4th column, are small lots (i.e., with less than 200
items); others, see the 5th column, are large lots (with more than 500
items). The same information, but related to the purchased/produced
components, are listed from the 6th to the 9th column. Finally, the last
column indicates the total number (in thousands) of reused inventory
parts.
As expected, the net profit (i.e., the value of the objective function)
always increases with 𝜇 and therefore with the production volumes:
clearly the price established for each product must guarantee a net
positive profit, even if the product is manufactured without any reusing
of parts and sold in the secondary market at a reduced price. Therefore,
a more meaningful assessment of the solution economical leverage can
be done by means of the investment indicator given by the ratio 𝑊 =
𝑅
𝑃 +𝑈 +𝑇
between potential revenue and total costs. The investment indi-
cator 𝑊 works by comparing potential income with total costs, giving a
more accurate measure of the investment and an immediate feedback
on the effort to implement a solution. This index is not the only one
that can be calculated, but it has an immediate understanding and can
be easily presented at any internal company meeting (describing how
many times potential returns are higher than total costs). Despite it does
not consider the difficulty of implementing a production plan, because
it does not take into account implementation details (e.g., amount
of components used, variety of finished products offered, number of
small lots of components ordered, etc.), this index can help in decision
making during the analysis of the solutions proposed.
For each configuration, Figs. 3, 6 and 9 show the value of 𝑊
together with the percentage value of used warehouse (the size of the
balls) calculated with parameter 𝑐ℎ𝑖 .
In Configuration 1 (see Fig. 3) the ratio 𝑊 and the value of the
used warehouse grow according with the opportunity ratio. The value
of the leverage is always greater than 36 and this result is strengthened
by considering that on the average the 93.3% of the value of reused
inventories consists of the value of obsolete components. Although such
solutions are very profitable, they generally are not compliant with
both production organization and market requirements. In fact, the
number of regular (or suitable) lots, i.e., those having a size between 200
and 500 parts, are always very low. In particular, small and large lots
represent on the average respectively the 76.9% and 10.5% of the whole
production of glasses types (see Fig. 4), with the largest lot consisting of
more than one million of pieces. Analogously, the number of small and
large lots of the ordered/manufactured components are on the average
respectively the 61.8% and 37.5% of the total volume (see Fig. 5).
E. Frontoni et al. Computers & Industrial Engineering 148 (2020) 106697

Table 2
Configuration 1: results.
𝜇 Obj value Opt gap CPU time b&b nodes 𝑅 𝑃 𝑈 𝑇 𝑉
(Me) (%) (s) (#) (Me) (Me) (Me) (Me) (Me)
0.1 4.439 0.00 16.02 1490 4.456 0.011 0.112 0.000 0.107
0.2 13.034 0.00 33.8 3388 13.072 0.031 0.159 0.000 0.153
0.3 24.006 0.00 35.15 3274 24.078 0.063 0.210 0.001 0.201
0.4 39.760 0.00 44.61 1649 39.892 0.114 0.287 0.001 0.269
0.5 63.732 2.15 limit 235834 64.005 0.237 0.476 0.001 0.440
0.6 97.353 2.56 limit 626201 97.734 0.339 0.568 0.002 0.528
0.7 134.458 2.50 limit 227556 134.937 0.428 0.614 0.002 0.565
0.8 174.946 2.91 limit 289545 175.569 0.566 0.711 0.002 0.656
0.9 221.889 3.31 limit 266977 222.671 0.712 0.794 0.003 0.727
1.0 275.426 3.23 limit 221360 276.413 0.899 0.903 0.003 0.818

Table 3
Configuration 1: Production details.
𝜇 Production of glasses Purchasing/Production of parts Reused parts
Lots Parts Small lots Large lots Lots Parts Small lots Large lots Parts
(#) (k#) (#) (#) (#) (k#) (#) (#) (k#)
0.1 65 68.3 52 3 126 358.5 83 42 41.8
0.2 101 200.7 81 7 194 1037.3 132 61 67.5
0.3 140 369.8 112 11 301 1906.0 187 114 103.5
0.4 165 612.8 126 17 364 3155.9 227 136 147.9
0.5 188 983.1 142 23 410 5078.5 238 172 264.2
0.6 212 1501.5 163 25 498 7682.6 329 169 295.9
0.7 251 2073.7 194 28 677 10586.6 430 247 336.4
0.8 273 2698.7 206 34 731 13786.0 437 294 403.6
0.9 290 3422.9 216 39 780 17525.1 444 336 447.0
1.0 329 4248.8 241 46 953 21738.0 528 415 591.7

Table 4
Configuration 2: results.
𝜇 Obj value Opt gap CPU time b&b nodes 𝑅 𝑃 𝑈 𝑇 𝑉
(Me) (%) (s) (#) (Me) (Me) (Me) (Me) (Me)
0.1 2.607 0.00 50.39 92 2.616 0.006 0.057 0.000 0.054
0.2 6.534 0.00 50.45 21 6.554 0.016 0.082 0.001 0.079
0.3 6.693 0.00 81.92 1161 6.725 0.028 0.095 0.001 0.091
0.4 6.820 0.00 34.61 0 6.873 0.041 0.105 0.001 0.094

Table 5
Configuration 2: Production details.
𝜇 Production of glasses Purchasing/Production of parts Reused parts
Lots Parts Small lots Large lots Lots Parts Small lots Large lots Parts
(#) (k#) (#) (#) (#) (k#) (#) (#) (k#)
0.1 11 40.0 0 3 25 205.1 0 24 18.4
0.2 18 100.0 0 8 33 489.3 0 32 54.5
0.3 24 100.0 0 10 56 612.2 0 52 77.2
0.4 22 100.0 0 9 52 183.9 0 50 68.4

Table 6
Configuration 3: results.
𝜇 Obj value Opt gap CPU time b&b nodes 𝑅 𝑃 𝑈 𝑇 𝑉
(Me) (%) (s) (#) (Me) (Me) (Me) (Me) (Me)
0.1 0.987 0.00 22.2 3355 0.993 0.004 0.043 0.000 0.041
0.2 1.401 0.00 75.86 3416 1.420 0.014 0.074 0.000 0.071
0.3 1.733 0.00 67.94 3427 1.769 0.027 0.092 0.000 0.083
0.4 2.053 0.00 93.56 2628 2.114 0.047 0.118 0.001 0.104
0.5 2.407 0.00 125.59 2182 2.498 0.068 0.137 0.001 0.116
0.6 2.769 0.00 959.8 9374 2.892 0.096 0.162 0.001 0.136
0.7 3.370 1.27 time limit 226735 3.571 0.172 0.247 0.001 0.219
0.8 4.004 0.00 1442.72 82558 4.266 0.228 0.287 0.001 0.254
0.9 4.521 0.00 2163.48 92381 4.844 0.281 0.314 0.002 0.272
1.0 4.979 0.00 347.05 22102 5.369 0.341 0.343 0.002 0.297

This aspect of the solutions would hinder the production of glasses
and it could represent an infeasible scenario to be accomplished due
to industrial constraints. Moreover, the production involves more than
the 20% of the catalogue, making the production pretty complex to be
managed.
Configuration 2 forbids small lots of products and components. For
𝜇 ≤ 0.2, the ratios 𝑊 are comparable with the corresponding values
of Configuration 1, see Fig. 6. Then 𝑊 decreases for 𝜇 between 0.2
and 0.4, and stabilizes from 𝜇 = 0.4 on (the model (1)–(13) returns the
same optimal solution for 𝜇 ≥ 0.4). On the other hand, the percentages

8
E. Frontoni et al. Computers & Industrial Engineering 148 (2020) 106697

Table 7
Configuration 3: Production details.
𝜇 Production of glasses Purchasing/Production of parts Reused parts
Lots Parts Small lots Large lots Lots Parts Small lots Large lots Parts
(#) (k#) (#) (#) (#) (k#) (#) (#) (k#)
0.1 33 14.7 0 0 32 62.5 0 31 32.1
0.2 49 21.4 0 0 83 109.5 0 78 55.1
0.3 64 26.6 0 0 133 158.5 0 120 64.5
0.4 78 31.7 0 0 184 211.0 0 170 77.0
0.5 94 37.4 0 0 229 264.8 0 210 90.7
0.6 109 43.3 0 0 265 328.9 0 240 108.0
0.7 134 53.6 0 0 328 488.6 0 300 173.5
0.8 162 63.9 0 0 432 607.7 0 390 207.2
0.9 181 72.7 0 0 501 717.1 0 461 235.8
1.0 203 80.5 0 0 556 807.4 0 507 266.4

Fig. 3. Configuration 1 — Revenue–Costs ratio and warehouse used.

Fig. 6. Configuration 2 — Revenue–Costs ratio and warehouse used.

Fig. 4. Configuration 1 — lots of produced glasses types.

Fig. 7. Configuration 2 — lots of produced glasses types.

of used warehouse, though smaller than those in Configuration 1,

Fig. 5. Configuration 1 — lots of ordered/produced components.
progressively grow with 𝜇, and again the vast majority of the used com-
ponent are obsolete (94.2% on average). Differently from Configuration
1 where a dominant strategy is evident, three solutions of Configuration
2 (those for 𝜇 ∈ {0.2, 0.3, 0.4}) are non-dominated. Therefore there
exists a trade-off between the economical leverage 𝑊 and the inven-
tory reusing that must be evaluated by the management according to
warehouse or industrial requirements (reduction of warehouse volume)
or emergency company policy (low budget for the purchase of new
components).
Besides, the number of production lots, i.e., the assortment of
glasses types produced for the secondary market, is much smaller than
that indicated in Configuration 1 (see Fig. 7); yet the number of large
lots is on average more than 40% of the total (see Fig. 8).
In Configuration 3, also the production of large lots of glasses types
is forbidden. As a direct consequence, the production of profitable

9
E. Frontoni et al.
Fig. 8. Configuration 2 — lots of ordered/produced components.
Fig. 9. Configuration 3 — Revenue–Costs ratio and warehouse used.
but not popular glasses is prevented (though the assortment of glasses
types increases) and the more evident effect is the drop-off of the
economical leverage 𝑊 of almost one order of magnitude, see Fig. 9.
The value of the reused inventory components grows up to 11.08% of
the total warehouse value and, similarly to the previous configurations,
the most (about 88.8%) is given by the reusing of obsolete parts. All
the ten solutions of this realistic scenario are non-dominated; they
therefore provide the management an overview for defining a suitable
trade-off between the potential revenues and the reusing of inventory
components.
The choice of the parameter 𝜇, that is the commitment of the
company to recover the value of obsolete stocks, depends from both
the strategy a company manager wants to follow and the current
overall state of the company. So, the 𝜇 adopted will correspond to the
solution with the highest value for the objective function that is also
compatible with all other conditions that fit managers’ needs, even if
they were not considered in the mathematical model, such as workforce
or investment indicator 𝑊 . For example, in Configuration 3, assuming
a workforce limitation that cannot move more than 200k units of
stocked components, solutions with 𝜇 ≥ 0.8 cannot be implemented (see
Table 7). So managers will pick the solution with 𝜇 = 0.7. If managers
also want a 𝑊 ≥ 15 the best solution that fits their needs (and still
respects the workforce constraint) is that with 𝜇 = 0.2 (see Fig. 9).
Actually, there is no generic policy or guideline that works in every
situation. For example, managers should pay attention to warehouse
stock levels, workforce availability, fiscal deadlines (reduce stock for
milder taxation), etc. In general, the managers have their fingers on
the pulse and know where it is appropriate to push their companies
and can always choose the solution that best fits the company status.
This is also the inner purpose of a DSS, which assists the manager for
taking decisions.
10
Computers & Industrial Engineering 148 (2020) 106697
Fig. 10. Transfers of components.
As a final remark, the transfers of components, either ordered or
stocked, are summarized in Fig. 10. For each configuration, the stocked
material shifted among hubs is around the 30% of the components used,
while the ordered parts are almost completely used in the same hub
they are purchased; in any case the transferring costs are always below
the 1% of the total costs. This aspect highlights the importance of the re-
usage of in-house materials even if it involves the transfer of a relevant
part of the overall stocked components used.
5.2. Mathematical program assessment
A preliminary computational experience was done by testing the
model (1)–(13) on a set of smaller instances obtained by randomly
select up to 50 BOMs from the real instance. In all cases an optimal
solution is achieved within seconds or fractions of a second.
The size of the program, populated with real data ranges, after
Cplex preprocessing, was from 23609 variables and 14864 constraints
(Configuration 3, 𝜇 = 0.1) to 84224 variables and 50999 constraints
(Configuration 2, 𝜇 = 0.3). In particular, the instances of Configuration
1 result larger even if constraints (7)–(9) are neglected, because the
bounds on the lot sizes allow Cplex to fix variables and tighten bounds.
All the cases of configurations 2 and 3 were solved to optimality
within the specified precision in about 7 min on average, with the
exception of the instance Configuration 3, 𝜇 = 0.7. However, in all the
cases when Cplex reaches the time limit (almost the totality belonging
to the Configuration 1 hypothetical scenario), the optimality gap is
quite low: 1.27% for Configuration 3, 𝜇 = 0.7, and 2.78% on average
for the instances of Configuration 1.
The relative gap between the best (possibly optimal) integer solu-
tion found and the upper bound provided by the linear relaxation of
(1)–(13) is quite small for Configurations 1 and 2 (on average 4.36% and
0.85%, respectively) and more substantial for Configuration 3 (31.38%
on average). Such gaps indicate a good quality of the formulation,
that is also proved by the ability of solving the challenging real case
provided by the company.
We close this section with a couple of considerations on the opportu-
nity to design a tailored solution approach. The economic investment in
the development of a specific heuristic should be justified by the chance
to overcome the potential limits of the adopted approach. However, the
tool we provided is used to solve a strategic/tactical problem, i.e., to
evaluate potential production scenarios and guide the management
in mid-long term decisions. This means that optimality is not strictly
required (due to the natural variability of input data) and small gaps
do not affect the usefulness of solutions. Moreover, the CPU time could
be largely increased in the attempt to solve the instance to optimality
due the mid-long term nature of decisions. About possible common
characteristics of the instances that prevent to solve to optimality,
E. Frontoni et al.
in general, we do not know how to characterize the complexity of
the instances neither a priori, being the problem NP hard, nor a
posteriori, being Cplex a commercial product of which we do not have
the algorithmic details. In Configuration 1 the difficulties are evident:
increasing the value of 𝜇 gives more combinations of components to
reuse/buy/produce and final products to assemble, i.e. each step of 0.1
increases exponentially the space to search for solutions. In all other
configurations optimal solutions are obtained except in the case 𝜇 = 0.7
of Configuration 3, where there is a small gap, which can be neglected
in our strategic/tactical problem.
6. Conclusions and future steps
Corporate decision-makers must act in such a way that all the
information in their possession is transformed into strategic actions to
support the business. DSSs increase the effectiveness of data analysis,
providing support to all those who need to make strategic decisions. In
the solution proposed in this paper, a DSS is integrated in the company
information system to get a ‘‘Data Integration" framework for deci-
sion making activities in procurements, production, management and
logistics. In particular, we embed in the existing software architecture a
new OR module, a component of our DSS that provides good solutions
to the problem of reusing obsolete components in fashion markets. Our
DSS and the company information system together are able to perform
all the simulations necessary to calculate the impact of the results of
the OR module on the different economic quantitative variables. In
particular, the overall system has different data interfaces for manage-
rial insights that allow to: (i) define model parameters with respect
to maximum batch production, maximum purchase budget, etc.; (ii)
access optimization results based on the previous described parameters
and on the overall economic data derived from the model; (iii) access
the production and component buying list of the solution provided and
move forward in the procurement and production pipeline.
The proposed mathematical model has several elements in common
with production planning models (Pochet & Wolsey, 2006), but it
has been adapted to the issues related to the problem examined in
this paper, i.e., different components prices, multiple production hubs,
multiple warehouses location, production planning, procurement man-
agement. All these features required a deep analysis and tailored design
in connection with the company needs and the specific problem faced.
In particular, we introduce an opportunity ratio 𝜇 that balances the
procurement/production of components, and therefore the production
of glasses for the secondary market, with the value of reused stocked
ones.
The solutions found by the OR module describes:
• quantity and type of glasses to produce and hubs where their
production should take place;
• quantity and type of company-owned (obsolete) components to
use for production and hubs from which these components are
picked up;
• quantity and type of components to be purchased or manufac-
tured and locations where their purchasing or manufacturing
should take place;
• transfers of stocked and purchased components between hubs.
In particular, we underline that: (i) the proposed framework and the
overall integration of OR and DSS in a real and complex worldwide
production scenario is a novel and effective solution and, up to our
knowledge, there are no similar frameworks in the literature; (ii) the
ability to propose good solutions for the manufacturing problem in a
totally automated way and in a fast and easy way is the main contri-
bution and insight of the proposed system; (iii) main contributions are
also on the DSS integration in the overall process, the real worldwide
data used for result evaluation with impressive insight and impact
on the company activities and turnover; (iv) reporting testing on a
real case instance is relevant to prove such complex methods on very
11
Computers & Industrial Engineering 148 (2020) 106697
complex scenarios (i.e., 957 BOMs, 3457 components, 4 international
hubs with related stocks, three configurations with parameters’ values
set according to real and hypothetical scenarios.)
The novel OR module collects a set of solutions that are shown
through the Kibana Dashboard, which also allows to save the results
to share or reload at a later time. Then, the Decision Actuator helps the
production manager and the procurement planner in finding a good
strategy. Through the solutions proposed by the OR module, the DSS
allows to simulate how the adoption of those situations will change
the situation of the warehouse and it also generates the corresponding
production orders and movements of components to be passed to the
SAP database. In such scenarios the company can better exploit the
Dashboard module to get the trading-off solutions that best fit the needs
of the company. For these reasons it is better to have a range of good
and heterogeneous strategies able to offer a range of suitable solutions
instead of only one approach. The proposed problem in a very complex
data scenario, as described in the result section, is not affordable at a
human level: for each configuration reported in the paper, a human
based approach, before the introduction of our method, was extremely
expensive with more than 20 working days necessary to produce a
solution. The overall framework and the final DSS brings the overall
processing time to minutes and the parameter definition based on the
proposed solutions allows to reach almost always the optimal solution
in less than 2 h.
Finally, another fundamental benefit of the approach proposed in
this paper is that it results performing, effective and widely doable with
a huge economic impact for the manufacturer that provided the case
study (a glasses production leader company operating in 150 different
countries). Furthermore, the mathematical model can be easily adapted
to many other manufacturing processes and product types.
The approach presented is tailored for the problem proposed. In
fact, the OR module relies on a mathematical model that can manage
only 1-layer BOM. Moreover, at present, the optimization does not take
into account the material flows that are used for production in the
main market. Given the limited production and movement of materials
for the secondary market, there have been no operational problems in
deploying the proposed solutions. However, at a later stage it is possible
to consider the production and supply volume of both markets and to
optimize the production (either of the second market or of both) having
this wider vision leading to further savings for the company.
In considering alternative solutions, let us point out how machine
learning or deep learning approaches often does not reach an optimal
solution and, above all, they are based on training data that are totally
not applicable on this scenario.
Our analysis on real data from a worldwide scenario has shown
variations of the trends with changes in the opportunity parameter 𝜇.
In each configuration, transfers of components mainly involve stocked
parts, due to low transfer rates and negligible fixed transportation
costs. Because there are possible different trends for the various (but
equivalent in terms of objective function values) solutions found, future
research can be moved in the direction of driving the solution obtained
to a better usage of the warehouse.
Future steps will be also focused on the analysis of the scalability of
the system to cover other areas of logistic and procurement, as well as
on a deeper parameter evaluation, both over the proposed dashboard
and the DSS, to further improve action planning.
CRediT authorship contribution statement
Emanuele Frontoni: Funding acquisition, Investigation, Software.
Fabrizio Marinelli: Conceptualization, Validation. Roberto Rosetti:
Methodology, Software, Visualization, Writing - original draft. Primo
Zingaretti: Supervision, Writing - review & editing.
Acknowledgements
We are grateful and wish to thank the anonymous Reviewers whose
sharp remarks helped us to improve the presentation of our results.
E. Frontoni et al.
References
Angelus, A. (2011). A multiechelon inventory problem with secondary market sales.
Management Science, 57(12), 2145–2162.
Antosz, K., & Ratnayake, R. C. (2019). Spare parts’ criticality assessment and prioriti-
zation for enhancing manufacturing systems availability and reliability. Journal of
Manufacturing Systems, 50, 212–225.
Arbib, C., Di Iorio, F., Marinelli, F., & Rossi, F. (2002). Cutting and reuse: An application
from automobile component manufacturing. Operations Research, 50(6), 923–934.
Calmon, A. P., & Graves, S. C. (2017). Inventory management in a consumer electronics
closed-loop supply chain. Manufacturing & Service Operations Management, 19(4),
568–585.
Cherri, A. C., Arenales, M. N., Yanasse, H. H., Poldi, K. C., & Vianna, A. C. G.
(2014). The one-dimensional cutting stock problem with usable leftovers - A survey.
European Journal of Operational Research, 236(2), 395–402.
Chung, W., Talluri, S., & Kovács, G. (2018). Investigating the effects of lead-time
uncertainties and safety stocks on logistical performance in a border-crossing JIT
supply chain. Computers & Industrial Engineering, 118, 440–450.
Cortellessa, V., Marinelli, F., & Potena, P. (2008). An optimization framework for "build-
or-buy" decisions in software architecture. Computers & Operations Research, 35(10),
3090–3106.
CPLEX Optimizer. (2020). https://www.ibm.com/it-it/analytics/cplex-optimizer. (Ac-
cessed 25 April 2020).
Elasticsearch. (2020). https://www.elastic.co/elasticsearch/. (Accessed 25 April 2020).
Fang, C.-C., Lai, M.-H., & Huang, Y.-S. (2017). Production planning of new and
remanufacturing products in hybrid production systems. Computers & Industrial
Engineering, 108, 88–99.
Fu, K., Gong, X., & Liang, G. (2019). Managing perishable inventory systems with
product returns and remanufacturing. Production and Operations Management, 28(6),
1366–1386.
Galizia, F., ElMaraghy, H., Bortolini, M., & Mora, C. (2019). Product platforms design,
selection and customisation in high-variety manufacturing. International Journal of
Productions Research.
Han, J.-H., Lee, J.-Y., & Kim, Y.-D. (2019). Production planning in a two-level supply
chain for production-time-dependent products with dynamic demands. Computers
& Industrial Engineering, 135, 1–9.
Jena, S. K., & Sarmah, S. (2016). Price and service co-opetiton under uncertain demand
and condition of used items in a remanufacturing system. International Journal of
Production Economics, 173, 1–21.
Kibana. (2020). https://www.elastic.co/kibana. (Accessed 25 April 2020).
Kumar, K., & Aouam, T. (2019). Extending the strategic safety stock placement model
to consider tactical production smoothing. European Journal of Operational Research,
279(2), 429–448.
12
Computers & Industrial Engineering 148 (2020) 106697
Lee, H., & Whang, S. (2002). The impact of the secondary market on the supply chain.
Management Science, 48(6), 719–731.
Lei, L., Zhong, H., & Chaovalitwongse, W. A. (2009). On the integrated production
and distribution problem with bidirectional flows. INFORMS Journal on Computing,
21(4), 585–598.
Liotta, G., Kaihara, T., & Stecca, G. (2016). Optimization and simulation of collaborative
networks for sustainable production and transportation. IEEE Transactions on
Industrial Informatics, 12(1), 417–424.
Montalto, A., Graziosi, S., Bordegoni, M., & Landro, L. D. (2018). Combining aesthetics
and engineering specifications for fashion-driven product design: A case study on
spectacle frames. Computers in Industry, 95, 102–112.
(0000). Obsolescences dataset, anonymous url, Anonimous url.
Phuc, P. N. K., Yu, V. F., & Tsao, Y.-C. (2016). Optimizing fuzzy reverse supply chain
for end-of-life vehicles. Computers & Industrial Engineering, 757–765.
Pinedo, M. L. (2009). Planning and scheduling in manufacturing and services.
Springer-Verlag, New York.
Pochet, Y., & Wolsey, L. A. (2006). Production planning by mixed integer programming
(Springer series in operations research and financial engineering). Secaucus, NJ, USA:
Springer-Verlag New York, Inc..
Ramezanian, R., Rahmani, D., & Barzinpour, F. (2012). An aggregate production
planning model for two phase production systems: Solving with genetic algorithm
and tabu search. Expert Systems with Applications, 39(1), 1256–1263.
Rasmi, S. A. B., Kazan, C., & Türkay, M. (2019). A multi-criteria decision analysis
to include environmental, social, and cultural issues in the sustainable aggregate
production plans. Computers & Industrial Engineering, 132, 348–360.
Saha, C., Aqlan, F., Lam, S. S., & Boldrin, W. (2016). A decision support system for
real-time order management in a heterogeneous production environment. Expert
Systems with Applications, 60, 16–26.
Takai, S., & Sengupta, S. (2017). An approach to evaluate the profitability of component
commonality. Journal of Mechanical Design, 139(7).
Takriti, S., Fourer, R., Gay, M., & Kernighan, B. (1994). AMPL: A modeling language for
mathematical programming. JSTOR.
Vié, M.-S., Zufferey, N., & Cordeau, J.-F. (2019). Solving the wire-harness design
problem at a European car manufacturer. European Journal of Operational Research,
272(2), 712–724.
Wang, S.-C., & Yeh, M.-F. (2014). A modified particle swarm optimization for aggregate
production planning. Expert Systems with Applications, 41(6), 3069–3077.
Yaghin, R. G. (2020). Enhancing supply chain production-marketing planning with
geometric multivariate demand function (a case study of textile industry). Computers
& Industrial Engineering, 140, 206–220.