See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/368345493

Applications of Artificial Intelligence in Inventory Management: A Systematic

Review of the Literature

Article in Archives of Computational Methods in Engineering · February 2023

DOI: 10.1007/s11831-022-09879-5

CITATIONS READS

17 4,852

3 authors:

Özge Albayrak Ünal Burak Erkayman

Ataturk University Ataturk University
10 PUBLICATIONS 60 CITATIONS 34 PUBLICATIONS 257 CITATIONS

SEE PROFILE SEE PROFILE

Bilal Usanmaz
Ataturk University
9 PUBLICATIONS 37 CITATIONS

SEE PROFILE

All content following this page was uploaded by Özge Albayrak Ünal on 25 October 2023.

The user has requested enhancement of the downloaded file.

Archives of Computational Methods in Engineering
https://doi.org/10.1007/s11831-022-09879-5

REVIEW ARTICLE

Applications of Artificial Intelligence in Inventory Management:

A Systematic Review of the Literature
Özge Albayrak Ünal1 · Burak Erkayman1 · Bilal Usanmaz2

Received: 13 August 2022 / Accepted: 23 December 2022

© The Author(s) under exclusive licence to International Center for Numerical Methods in Engineering (CIMNE) 2023

Abstract
Today, companies that want to keep up with technological development and globalization must be able to effectively manage
their supply chains to achieve high quality, increased efficiency, and low costs. Diversified customer needs, global competi-
tors, and market competition have led companies to pay more attention to inventory management. This article provides a
comprehensive and up-to-date review of Artificial Intelligence (AI) applications used in inventory management through a
systematic literature review. As a result of this analysis, which focused on research articles in two scientific databases pub-
lished between 2012 and 2022 for detailed study, 59 articles were identified. Furthermore, the current situation is summarized
and possible future aspects of inventory management are identified. The results show that the interest in AI methods has
increased in recent years and machine learning algorithms are the most commonly used methods. This study is meticulously
and comprehensively conducted so it will probably make significant contributions to the further studies in this field.

keywords Inventory Management · Artificial Intelligence · Machine Learning · Deep Learning · Systematic Literature
Review

1 Introduction a continuous process of planning, organizing and controlling

which minimizes inventory investment while balancing sup-
As a result of increasing competition, customers today are ply and demand [2].
looking for the products they buy in global markets, at the Inventory management is primarily concerned with the
right time, in the right place, with good quality and at lower planning and controlling an industry’s inventory and is
prices. Supply chains are complex systems that connect the an important component of supply chain management. It
world. Inventory management is essential for achieving the includes issues such as estimating material requirements at
goals of efficient supply chains, controlling costs, and deliv- various points in the supply chain, determining necessary
ering to customers with minimal delays [1] and goes hand material’s amount, ordering frequency, and safety stock lev-
in hand with the supply chain. The word inventory refers els. It also includes inventory visibility, inventory forecast-
to investment in materials and end-products throughout the ing, inventory management, lead time, inventory shipping
supply chain for use in production or distribution to the end- costs, inventory forecasting, inventory valuation, forecast-
customer. Singh and Verma define inventory management as ing future inventory prices, available physical space, qual-
ity management, returns and defective goods, and demand
* Burak Erkayman forecasting [2]. It plays a very important role in reducing
[email protected] overall costs and rapid response objectives. Effective inven-
Özge Albayrak Ünal tory management requires the right inventory in the right
[email protected] place at the right time to minimize system costs and meet
Bilal Usanmaz customer needs.
[email protected] In inventory management, it is very important to avoid
uncertainties, which usually occur in demand forecasts
1
Department of Industrial Engineering, Engineering Faculty, during the lead time. Demand forecasts form the basis
Ataturk University, 25240 Erzurum, Turkey
of all planning activities as they are the input for many
2
Department of Computer Engineering, Engineering Faculty, operational decisions [3, 4]. Manufacturing companies
Ataturk University, 25240 Erzurum, Turkey

13
Vol.:(0123456789)
 Ö. Albayrak Ünal et al.

significantly rely on demand forecasting since it deter- 3 By analyzing the status of AI methods, the most com-
mines production planning. When the demand process is monly used methods in inventory management are iden-
unknown, we generally depend on forecasting expected tified.
demand during the lead time. Accurate and reliable 4 Future research directions in the field of inventory man-
demand forecasts are critical as they thoroughly guide agement are identified.
supply chain managers’ planning, all major operational
decisions. Inaccurate estimations mislead service level The article is organized as follows: After the Sect. 1,
goals and create additional costs like shortages, lost rev- information regarding the research methodology is given
enue, or excess inventory. Over demand may lead to inven- in Sect. 2. The literature review and related work evalua-
tory depletion, while additional inventory costs may be tion strategy are described in Sect. 3. Section 4 contains the
incurred if demand is below expectations. With sufficient practical implications, which explores the research questions
inventory, bottlenecks can be avoided or unnecessarily in detail. Conclusion and future work directions are given
high inventory costs can be avoided by keeping a corre- in Sect. 5.
spondingly lower inventory level. Inventory management
is used in various domains such as retail [5, 6], logistics
[7, 8], supply chain management [9, 10], and availability 2 Research Methodology
of multiple sources of supply [11].
Development in artificial intelligence technologies, one of Systematic literature reviews, which are becoming increas-
the innovations that technological development has brought ingly popular in academic research, are based on a rigorous,
with the improving capabilities of computer hardware, robust, well-defined, and reliable methodology for literature
allows the inventory management applications turn into an search, and they allows the reader to rapidly sample and
intelligent process. Machine learning (ML) and deep learn- evaluate the related field [12]. They are used to gain a new
ing (DL) methods, which are sub-branches of AI, play an and comprehensive understanding of the relevant field and
important role in this sense. Number of studies conducted in identify other useful areas of research. This approach aims
recent years shows that the interest in ML and DL techniques to identify, interpret, evaluate, and categorize all articles
is gradually increasing. These methods can quickly analyze related to the identified research question(s). Compared to
large and diverse data sets, and they improve the accuracy a literature review, which focuses primarily on the descrip-
of demand forecasting. In addition, combination of inven- tive results of a particular field of knowledge, a systematic
tory management with AI techniques makes it an efficient literature review provides a more useful and comprehensive
and flexible process with lower operational costs, supplies overview of research fields [13].
faster response times for customers as well as more contex- The current study followed a five-step process to better
tual information. Researchers were allowed to focus on ques- understand the existing literature, identify research questions
tions aimed at clarifying the true potentials and potential and keywords, and develop further steps. Figure 1 shows the
weaknesses of such algorithms, making it easier for them to stages of the systematic literature review used in the study,
focus on these techniques. and Fig. 2 shows the number of articles examined through
The aim of this article is to provide a comprehensive over- the review process.
view of the current and future research potential of inventory
management through a systematic literature review of arti- 1. Determination of research questions Probably the most
cles from 2012 to 2022. Inventory management and related important and difficult part of the research design is the
AI techniques are categorized and presented in a way that determination of the research questions. The formulation
facilitates orientation for researchers in the field. A biblio- of a research question leads to the selection of research
metric analysis was used to ensure that studies were inves- strategies and methods, i.e. the research is conducted
tigated and explored in depth with a quantitative analysis. based on the research questions. Formulating a research
This study comprehensively explores and investigates question plays an important role in the research process
AI methods used in inventory management. It also aims to as it helps to combat the collection and analysis of unre-
summarize the current applications and points out possible lated data [14].
future directions for inventory management. The contribu-
tions of this study are as follows: Research questions were formulated to accurately
determine the purpose considering the scope of the study.
1 The relevant literature by developing a classification Research questions of this study are as follows:
scheme based on previous studies are reviewed.
2 A step-by-step approach to conduct a systematic litera- • RQ1. What is the current state of research in inventory
ture review is provided. management?

13
Applications of Artificial Intelligence in Inventory Management: A Systematic Review of the…

Fig. 1  Stages of systematic literature review

Fig. 2  Number of articles

reviewed according to the
results of the search

• RQ2. What are the areas and sub-areas in inventory 2. Databases examined in the study Web of Science and
management research where AI methods are being Scopus, two established academic databases that provide
applied? broad access to a wide range of peer-reviewed literature,
• RQ3. Which AI methods are being used in inventory were utilize to identify the related studies on research
management studies? topic.

13
 Ö. Albayrak Ünal et al.

  Selected studies distribution in terms of three main

topics, namely inventory problems, demand forecast and
inventory classification, are shown in Table 1. It is clear
from this table that majority of the studies have been
conducted in the field of demand forecasting.

5. Analysis of results Findings, limitations and recommen-

dation were examined in this step that will be presented
in Sect. 5 of the study.
Fig. 3  The identifying keywords used in the study

2.1 Bibliometric Analysis
Table 1  Search results
Web of sci- SCOPUS Total Bibliometrics is the census-based field of study to survey
ence
published articles, journals, and book editions using math-
Inventory problems 10 9 19 ematical and statistical techniques. Bibliometric analysis,
Demand forecast 13 14 27 on the other hand, is a research method that examines
Inventory classification 10 3 13 the characteristics of studies and research in a particular
Total 33 26 59 field with a quantitative analysis, i.e. it analyzes certain
characteristics of documents/publications such as journal,
subject, number of authors, publication information. The
bibliometric analysis applied in this study corresponds to
3. Identifying keywords Web of Science and Scopus search
the answer to the research question RQ1. In this sense, a
were performed based on title, abstract, and manuscript
quantitative analysis and data visualization of the stud-
keywords (TITLE-ABS-KEY) in May 2022. The cor-
ies is presented according to the criteria of publication
responding literature was searched using the keywords
trends, source distribution, distribution by country, most
given in Fig. 3.
cited articles, and keyword analysis.
Here, the “OR” operator is used to combine keywords
within the same group while the “AND” operator is used to
2.1.1 Publication Trend
combine the two main keyword groups. As a result of the
initial search, 338 results were obtained in Web of Science
The current study examines 59 AI studies for inventory
and 477 results in Scopus.
management and covers the years 2012–2022. Distribu-
tion of studies in terms of the publication year is shown
4. Inclusion/Exclusion criteria. Inclusion/Exclusion crite-
in Fig. 4. It is clear from this figure that there has been
ria Inclusion/exclusion criteria were established to select
a rapid growth recently, and about 40% of these studies
the most relevant articles. Articles whose language of
were published in 2021. The reason for the low number of
publication was English in the period from 2012 to 2022
articles published in 2022 is that the literature review was
were included in the study. After applying these filters,
conducted in May 2022.
175 results were obtained in Web of Science and 186
results in Scopus. Following biochemistry and telecom-
2.1.2 Resource Distributions
munications-oriented studies exclusion, 130 articles in
Web of Science and 163 articles in Scopus remained in
The distribution of 59 articles among 49 journals are shown
the search list. Research results from the two databases
in Fig. 5. About 70% of the articles were published as a sin-
were then merged using Endnote software, removing
gle study in a particular journal. European Journal of Opera-
repetitive publications yielding 195 articles in the list.
tional Research, on the other hand, hosts the majority of the
This process is illustrated in Fig. 2. articles that were
publications with 10% followed by Applied Science with 7%
insufficient to answer the research questions and not
and International Journal of Production Economics with 4%.
directly related to the topic were excluded. Through the
Expert Systems with Application, Computers & Industrial
filtering process, the number of articles was reduced to
Engineering, and Mathematical Problems in Engineering
59 that are included in the study.
journals were contributed to publications with 3% each.

13
Applications of Artificial Intelligence in Inventory Management: A Systematic Review of the…

2.1.5 Keyword Analysis

Identifying commonly used keywords related to inventory

management and artificial intelligence techniques is very
important to determine the focus of a study. Therefore, a co-
occurrence analysis was performed. With this analysis, 8 dif-
ferent clusters containing the largest number of items were
identified utilizing minimum twice co-occurrence threshold.
The mapping of 63 qualified keywords out of a total of 603
keywords and their interactions with each other are shown
Fig. 4  Publication trends of articles from 2012 to 2022 in Fig. 7. Here, each cluster is represented by a different
color to indicate that objects in the same cluster have more
similarity than objects in a different cluster. Different objects
are connected by lines representing the links between them.
The most frequently used keywords in the selected lit-
erature are “inventory management” (occurrence = 30, total
link strength = 47), “machine learning” (occurrence = 30,
total link strength = 37) and “demand forecasting” (occur-
rence = 25, total link strength = 36). The keywords and their
numbers are detailed in Table 3.

3 Review of Literature

This section addresses research questions RQ2 and RQ3.

The articles studied were analyzed in detail and discussed
under three main headings such as inventory problems,
inventory classification, and demand forecasting. Each sec-
Fig. 5  Resource distributions
tion is also divided into subsections, and the articles used in
the study are presented in the form of a summary table after
being examined against the characteristics given in Fig. 8.
2.1.3 Distribution of Articles by Country of Origin In addition, studies that did not have similar characteris-
tics in each category were examined in more detail. The list
Geographic distribution of countries, defined based on the with the full names of the algorithms whose abbreviations
corresponding authors’ institution origin, was also stud- are given in the summary table can be found in Appendix.
ied. In this sense, there were 21 countries contributed. Top Although this study is limited according to the character-
eight countries with a major contribution along with others, istics given in Fig. 8, important studies on this topic are as
combined into a single entry, are given as a bar graph in follows: Aggarwal provided a comprehensive overview of
Fig. 6. It can be concluded, from this figure, that the topic traditional inventory systems [20]. Giannoccaro et al. used
is of global interest. Although there are many China based the reinforcement learning algorithm to determine a near-
articles (about 26%), the number of articles from Canada, optimal inventory policy [21]. Cachon and Fisher demon-
Belgium, USA and Turkey is also significant. strated the various logistical advantages of sharing inventory
management related information with supply chain partners
2.1.4 Most Cited Articles [22]. Partovi and Anandarajan were used artificial neural
networks to classify stock keeping units in a pharmaceutical
The most cited five studies along with their application company [23]. Giannoccaro et al. presented a method based
focuses are tabulated in Table 2. Here, Nguyen and Med- on fuzzy set theory and cascade inventory theory to define
jaher is the most cited article dealing with the development inventory management policy in the supply chain [24].
of a predictive maintenance framework based on sensor
measurements [15]. Tabernik and Skocaj studied object
3.1 Inventory Problems
detection and recognition [16]. Mohammaditabar et al., Liu
et al. and Kartal et al. focused on classification of inventories
This section discusses issues related to general inventory
using different methods [17–19].
problems associated with optimization, inventory control,

13
 Ö. Albayrak Ünal et al.

Fig. 6  Distribution of published articles by country of origin

Table 2  Most cited articles based on learning using existing data, and enable dynamic
learning in a changing environment without the need for a
Authors Year Citations Application Focuses
predetermined environment model [29].
Nguyen and Medjaher 2019 150 Data driven approach The data-driven approach requires manual processing and
Tabernik and Skocaj 2020 126 Object detection analysis of data by experts [46]. Inventory problems using ML
Mohammaditabar et al. 2012 111 Inventory classification methods were explored in this subsection. Pirayesh Neghab
Liu et al. 2016 102 MCABC inventory et al. studied the ordering problem for a single news vendor
classification problem to minimize the expected cost. To solve the model,
Kartal et al. 2016 86 Inventory classification they proposed a new algorithm called HMMNV based on the
integration of neural networks and hidden Markov models,
and policy setting. It is aimed to determine optimal inven- and demonstrated the performance of the algorithm using
tory level, minimize operating costs, and maintain that level crude oil demand data. Their model performed 27% better
on a regular basis. In order for production processes to con- compared to other methods in terms of system cost [30].
tinue without an interruption, requirements must be supplied The optimization subsection examines how products are
in a timely manner. Order fulfillment [25], situations where procured, managed, used, and product policies are deter-
the distribution of demand suddenly changes [26], dynamic mined. Better results were obtained based on the over-pro-
inventory management and control [27] and inventory con- visioning, ski-rental, and max-min approaches in a problem
trol [28] are studied in the literature. of replenishing the drug volume to manage a hospital’s drug
As a result of the research in the field of inventory man- inventory by Zwaida et al. With this study they have proven
agement problems, 18 articles were identified, which are that DRL is a promising method [33]. Kara and Doğan eval-
presented in Table 4. ML algorithms have been mainly used uated inventory management performance under stochastic
in this field, and the results have shown that they can be well customer demand and lead time to minimize a retailer’s total
adapted to this type of problems. Inventory control aims to cost. They observed that short-lived products showed better
continuously monitor inventory movements, customer pur- results under higher variance in demand [32].
chasing tendencies, and timing analysis using cumulative One of the widely studied areas in inventory problems
sales data. Most articles on inventory control have used rein- is the inventory control. Meisheri et al. suggested a new
forcement learning algorithms to track and control inventory approach of reinforcement learning to solve practical scenar-
movements, solve complex sequential decision problems ios with unit weights and quantities, shelf life and capacity,

13
Applications of Artificial Intelligence in Inventory Management: A Systematic Review of the…

Fig. 7  Keyword co-occurrence map

Table 3  Keywords contained in the eight clusters

Cluster Keywords Count

Inventory management, ABC classification, mixed integer programming, classification, inventory control, inventory 9
Cluster 1 classification, inventory, multi-criteria decision making, multi- criteria inventory classification
Inventory optimization, LSTM, genetic algorithms, smart manufacturing, random forest, stochastic optimization, 9
Cluster 2 data-driven decision making, prescriptive analytics, XGBoost
Demand forecasting, artificial intelligence, blockchain, deep learning, Internet of Things, supply chain, inventory 9
Cluster 3 management system, service level, logistic
ABC analysis, optimization, management, location awareness, feature extraction, radio frequency identification, 8
Cluster 4 costs
Supply chain management, production, e-commerce, inventory theory and control, retailing, hidden markov model, 8
Cluster 5 operations management, inventory
Spare parts, classification, artificial neural network, deep reinforcement learning, bullwhip effect, inventory control, 7
Cluster 6 production
Machine learning, data mining, neural networks, big data, blood demand, perishable inventory management, CART​ 7
Cluster 7
Forecasting, intermittent demand, inventory, newsvendor, demand, inventory management 6
Cluster 8

cross-product constraints, stochastic demand, and multiple and aimed to highlight potential research avenues that could
products for an inventory control problem of multiple items expand their scope by improving them [36].
with different lead times. The proposed solution method Establishing appropriate inventory policies is very
using DQN and PPO algorithms were compared, and the important in order to increase inventory management per-
results showed that the method is robust and very close to formance. Decision makers must consider the impact of
the optimum [39]. De Moor et al. demonstrated that rein- complex interactions by incorporating both internal and
forcement learning in perishable inventory management can external factors. Priore et al. have developed a dynamic
stabilize the training process and improve DRL performance framework for managing inventory in a supply chain.
in inventory control. Results showed that reward shaping Inductive learning algorithms were used to understand the
had an overall positive effect [35]. Boute et al. used DRL complex interaction between controllable and uncontrol-
algorithms to facilitate applications in inventory control lable factors that affect labor performance. They obtained

13
 Ö. Albayrak Ünal et al.

Fig. 8  Features used to analyze

the articles used in the study

successful results for the inventory management problem increase profits, market their products, and improve cus-
by selecting the best inventory policy for a wholesaler with tomer satisfaction [55] by preventing the inventory from
an average accuracy of 88% [40]. being depleted. It improves the developing of an adaptive
Inventory visibility provides an overview of inven- pricing strategy for better revenue management [56]. Simple
tory that allows real time tracking. Demey and Wolff pre- moving average, exponential smoothing, Croston method,
sented an inventory management system called SIMISS Sytetos-Boylan approach, etc. are used in demand forecast-
to improve the visibility of lost items in the International ing as the traditional methods. In recent years, on the other
Space Station, and they used a classical decision tree to hand, applications of AI methods such as clustering, k-near-
show its performance. With this system, they were able to est neighbor, neural networks, regression analysis, decision
reduce supply costs for long-term missions. [41]. tree, support vector machines, Gaussian processes, regres-
One of the methods used in inventory management is sion and long-short term memory, etc. have employed.
object detection and recognition. Tabernik and Skocaj Unlike traditional demand forecasting methods, Machine
proposed a Deep Learning based system for traffic sign Learning does not focus on priori assumptions, but learns
inventory management using convolutional neural net- from the available data to produce the most accurate result.
works. They concluded that the DL based approach has This, in turn, helps ML-based forecasters increase customer
high accuracy and speed for many traffic sign categories engagement and create more accurate demand forecasts as
with an average error rate of 2–3% [16]. Merrad et al. have they expand into new markets or channels [57].
proposed a reliable and efficient object detection solution Demand forecasting is mainly concerned with the deter-
to overcome the inventory management problem of detect- mining the level of demand for a future period. It is the
ing out-of-stocks in warehouses [42]. most studied field as shown in Table 5. Demand forecast-
The digital twin, which connects the physical and vir- ing field consists of the sub-fields of customer forecasting,
tual worlds [47], can address the challenge of seamless product forecasting, sales forecasting, order forecasting,
integration between advanced data analytics and IoT [48]. and information finding/information exchange. Based on
Kegenbekov and Jackson offered a solution to outperform the reviews, included 28 demand forecasting articles were
the base stock policy. They also introduced digital supply analyzed according to their sub-fields.
chain twins for real-world applications [38]. This section shows that stable predictive performance
Computer vision enhances human-machine interaction is achieved by tackling complex problems without requir-
by providing improved visualization of the natural world ing long demand histories to properly adjust parameters.
on a digital platform similar to the human brain [49, 50]. The benefits of ML in inventory management are reducing
Kalinov et al. found that flight path correction using the costs, determining the amount of material to procure, esti-
CNN approach provided a higher precision score compared mating material demand, and classifying inventory.
to the standard snake-based grid flight trajectory method. Companies try to predict markets’ future trends in order
They reduced the time of an inventory process without to understand customer demand, and they plan the decisions
decreasing the percentage of barcode recognition [43]. to be made and the activities to be executed. However, it
is very difficult to make an effective forecasting since the
3.2 Demand Forecasting customer demand often fluctuates due to various factors.
Deng and Liu have shown that their proposed deep inven-
This section presents studies on inventory management tory management method can effectively predict customer
and demand forecasting. It relates to the processes such as demand trends with a predictive accuracy of more than
cost-effective inventory management [51], inventory con- 80% and reduce overall costs by about 25% [58]. K-nearest
trol [52], stock forecasting [53] and sales forecasting [54]. neighbor models were used for monthly customer demand
Accurate demand forecasting is considered a necessity for by Kack and Freitag. They obtained high average estimation
proper inventory management. It supports companies to accuracy with short computation times [59].

13
Applications of Artificial Intelligence in Inventory Management: A Systematic Review of the…

Table 4  Analysis of studies for inventory problems

Sub-Field Study Cita- Algorithm Type Algorithms Goals/ Approach
tions

Data-Driven [15] 150 DL Long-Short Term Memory A new dynamic predictive maintenance framework based
Approach (LSTM) on sensor measurements is presented.
[30] 2 DL Deep Neural Network A new integrated estimation and optimization approach is
(DNN) implemented to solve an inventory problem that consid-
ers both observable and unobservable characteristics that
affect the randomness of demand.
Optimization [31] 19 ML Automated Innovization The inventory management problem has been applied to an
AI Genetic Algorithm (GA) automated innovation framework using genetic program-
ming to achieve higher levels of innovation.
[32] 83 ML Q-learning and Sarsa A reinforcement learning-based approach was used to
AI algorithms, GA understand the importance of stock age policy in perish-
able stock systems.
[33] 12 ML Deep Reinforcement An online Deep Reinforced Learning-based solution for the
Learning (DRL) drug refill optimization problem in a hospital is proposed.
Inventory Con- [34] 1 ML State-Action-Reward- A reinforcement learning approach was proposed to use
trol State-Action (SARSA) knowledge of the structural components of inventory
Algorithm management problems, and the results showed the appli-
cability of the approach.
[35] 3 ML Deep-Q-Network (DQN) A reward design method was used to accelerate training
and learning in perishable inventory management.
[36] 21 ML DRL A roadmap for the application of DRL in inventory control
was presented, along with a list of issues that need to be
addressed for the solution to work.
[37] 44 ML DRL The problems of lost sales, dual-sourcing, and multi-eche-
lon inventory management were evaluated.
[38] 8 ML DRL Demonstrated how inbound and outbound data flows can
Digital Twin be synchronized and business continuity supported when
end-to-end visibility is provided based on the Proximal
Policy Optimization algorithm.
[39] 5 ML DQN A general framework using RL algorithms is presented,
Proximal Policy Optimiza- explaining why the general inventory problem cannot be
tion (PPO) solved using classical optimization techniques.
Inventory Policy [40] 70 ML Inductive Learning Algo- It provides a dynamic framework for periodically determin-
rithm ing the best renewal rule for a given node in the supply
chain and the determination of the appropriate inventory
policy.
Stock Visibility [41] 6 ML Decision Tree (DT) An inventory management system called SIMISS was
offered to increase the visibility of lost items.
Object Recogni- [16] 126 DL Mask R-CNN A Deep Learning based system is proposed to detecting
tion and recognizing multiple traffic sign categories.
[42] 1 ML Random Sample Con- A real-time machine learning-based notification system
sensus (RANSAC), for the inventory shortage problem in warehouses was
Scale Invariant Feature presented.
Transform (SIFT)
[43] 22 Computer Vision Convolutional Neural Presented a hybrid robotic system based on unmanned
Network (CNN) aerial vehicles for detecting and scanning of barcodes
scanned as landmarks in a real warehouse in low light
conditions.
Customer Senti- [44] – Computer Vision Mask-R-CNN An innovative pipeline is proposed that integrates
ment Analysis DL YOLOv5 advanced Deep Learning technologies and includes a
visual AI-based customer sentiment assessment engine.
Maintenance [45] – ML DRL A hybrid solution algorithm based on a Double Deep
Planning AI Simulated Annealing (SA) Q-Network for maintenance planning was developed.

13
 Table 5  Analysis of studies for demand forecasting field
Sub-Field Study Citations Algorithm Type Algorithms Goals/Approach

13
Customer Estimate [58] 1 DL LSTM A deep inventory management method is proposed to forecast future
customer demand.
[59] 38 ML K-Nearest Neighbor (KNN) The local forecasting performance of K-nearest neighbor models is
examined in estimating the monthly customer demand of a manufac-
turing company.
Product Estimate [60] 52 ML K-means Clustering Algorithm, ARIMA A forecasting model for retailers based on customer segmentation was
Statistical Methods developed to improve inventory performance.
[61] 13 Time Series Models Moving Average, Neural Network (NN), By using different methods to estimate the print volume of Taiwan’s
DL ARIMA leading educational publisher, they were able to increase the accuracy
of demand forecasting by 3.7% and reduce capacity planning costs by
8.3%.
[62] 47 ML CART, Decisions about the number of transfusions and blood orders in a
DL KNN, hospital network were made using forecasting techniques. Comparing
Random Forest (RF), MLP, Artificial Neural Network the Multilayer Perceptron (MLP) model with other approaches, it was
(ANN) found that it can produce efficient decisions.
[63] 20 ML K-Means Clustering Algorithm, RF, Quantile Regression The proposed machine learning-based approach aims to provide pre-
Forest (QRF) launch forecasts and support inventory management decisions by using
historical sales data of existing and new products as well as previously
launched products.
[64] 6 DL Recurrent Neural Network (RNN) A forecasting framework is proposed that uses information on non-
demand characteristics such as past demand and downstream inventory
data to forecast the demand for drugs in a group.
[65] 5 Statistical Methods Comb-TSB, After a literature review on demand forecasting methods for time series
ML ClustAvg data, a new demand forecasting tool is proposed based on analysis
results and findings.
[66] 3 DL CNN, A deep learning-based demand forecasting strategy for cold chain prod-
LSTM, uct demand prediction is presented.
Graph Embedding
[67] 20 Statistical Methods GAMLSS An application of Generalized Additive Models for Position, Scale, and
ML Shape (GAMLSS) has been proposed to build distribution regression
models to forecast demand for a large number of perishable products.
[68] 9 Time Series Models ML Loess, A hybrid model is presented that combines seasonal and trend decom-
Extreme Gradient Boosting (XGBoost) position, considering inventory and replenishment constraints, to
estimate future red blood cell demand.
[69] 1 Time Series Models Simple Exponential Smoothing, Three time series forecasting methods were selected for demand
AI Quadratic Exponential Smoothing, forecasting according to periodic demand, static demand, and trend
Feature Synthesis, demand for sensitive parts based on spare parts, and a genetic algo-
GA rithm was used to check the performance of the inventory management
system.
Ö. Albayrak Ünal et al.
 Table 5  (continued)
Sub-Field Study Citations Algorithm Type Algorithms Goals/Approach

[70] 4 ML XGBoost The future need for red blood cells was predicted.
[71] 4 ML Linear Regression, Nonlinear Regression, ANN, SVR The number of spare parts for construction machinery that the customer
DL will demand in the future was estimated.
[72] 8 ML RF, XGBoost, Stochastic optimization and ML methods were used to examine the ten
KNN, DT most commonly used drugs to minimize the need for emergency sup-
plies and inventory in a hospital ward.
[73] 1 AI Multiple Linear Regression, An IoT-based inventory management system is presented that combines
GA, the causal method of multiple linear regressions with genetic algo-
Internet of Things rithms for product demand forecasting of a semiconductor manufac-
turer.
[74] – ML Generalized Linear Model, Generalized Additive Model, A two-stage hybrid model is proposed to study the demand uncertainty
Multivariate Adaptive Regression Splines, RF, Bayes- of a food bank.
ian Additive Regression Trees
[75] 2 ML RF, Gaussian Process (GP), NN, XGBoost, DT A new ML method for estimating red blood cell requirements was
developed and compared with the results of four widely used ML
algorithms.
[76] 6 ML Linear regression, GAMLSS, Classification-based model selection is presented as a new approach to
Statistical Methods QuantReg, QRF, ARIMAX demand forecasting for perishable retail products.
Sales [77] 35 ML XGBoost A three-stage XGBoost-based forecasting model was created to predict
Forecast the sales characteristics and trend of a commodity data series. The
CA -XGBoost model was found to outperform the Autoregressive
Integrated Moving Average (ARIMA), XGBoost, C-XGBoost, and
A-XGBoost models.
[78] 2 ML LSTM, A sales forecasting model called M- GNA- XGBoost has been proposed
DL Generative Adversarial Networks, to effectively forecast sales for each product in online stores and imple-
XGBoost ment digital marketing strategies.
Applications of Artificial Intelligence in Inventory Management: A Systematic Review of the…

[79] – ML LSTM, A framework has been proposed to reliably predict a company’s inven-
Core Density Estimation tory data.
Order [80] 1 ML RF, KNN, NN, Logistic Regression, Balanced Blagging ML methods are used to solve the problem of backorders prediction
Forecast DL (BB), SVM, XGBoost, problem and maximizing profit on decisions about outstanding orders.
LightGBM A post-hoc explanatory model was applied to the best performing
model by comparing it to other commonly used ML methods.
[81] 22 ML RF, Product preorders were estimated using a tree-based predictive model.
Gradient Boosting Machine
[82] 23 ML A new ML method called Weighted Majority Newsvendor Shifting is
proposed to solve a newsvendor problem in estimating an unknown
demand without knowing the type of distribution, variance, and mean.

13
 Ö. Albayrak Ünal et al.

One of the most studied topics in the field of demand fore-

A case study of a U.S. pharmaceutical manufacturer examined the use of

A knowledge discovery system has been researched and developed with
casting is product demand forecasting. There exist 14 studies

roulette genetic algorithm for demand forecasting in a small textile

conducted on this topic. Benhamida et al. have proposed the

downward information to improve short-term demand forecasts.

A digital twin integrating smart warehouse and production with a

Comb-TSB hybrid method for intermittent and lumpy demand
models. In this method, TSB relies on an automatic separa-
tion decomposition between a combination method consisting
mainly of statistical models, while the Comb method uses two
statistical estimation methods (ARIMA and Theta) and ML-
based MLP method. Their cluster-based approach was validated
in a case study [65]. In another study, Abbasi et al. used four
common predictive machine learning models to make blood
ANN for inventory forecasting.

transfusion decisions in a hospital network. They concluded that

use of a trained neural network model reduced average daily
company was proposed.

costs by about 29% compared with the current policy [62].

Zhang et al. compared the ARIMA, PSO-ELM, and XGBoost
Goals/Approach

methods with an empirical evaluation based on real data to vali-

date the performance of the cold chain storage demand forecast-
ing scheme. While their proposed method had the best per-
formance, the ARIMA method yielded the worst [66]. Ulrich
et al. applied GAMLSS to build distribution regression models
to predict the demand for a large number of perishable items.
They observed the results of linear regression, log-linear regres-
sion, log-log regression, quantile regression, random forests,
and quantile regression forest methods as performance criteria
for the GAMLSS approach [67]. To forecast demand for eight
Naive, Exponential Smoothing (ETS),

products in a supermarket in India, Bala segmented customers

based on various characteristics through cluster analysis and
implemented various forecasting models based on ARIMA and
neural network based ARIMA. The given forecasting model
TREPAN Algorithm

ensured to improve inventory performance by increasing the

level of service to customers and decreasing the level of inven-
Digital Twin
Lasso, MLP,

tory [60]. Li et al. suggested an integrated ordering strategy to

Algorithms

ARIMA-W
ARIMAX,

SVR, RF,

forecast future red cell demand. They reduced the inventory by

ETS-W,
ANN,

40% and ordering frequency by 60%, resulting in significant

GA

cost savings for blood suppliers [68].

Sales forecasting is an analytical technique that seeks to
Time Series Models ML

predict and understand consumer demand in businesses and

support managerial decision-making. Effective demand fore-
Algorithm Type

casting not only supply companies a competitive advantage,

but also plays a very important role in the strategic decisions
they make to satisfy consumer demands and satisfaction,
DL

AI

maintain market share, and manage costs [86]. In this sense,

Wang and Yang used an effective sales forecasting model
Study Citations

called M-GNA-XGBoost, which provides effective forecast-

ing of sales in a short time to perform digital marketing strat-
4

egies with machine learning. The root mean and absolute

Knowledge Discovery, [83]

[84]

[85]

error mean of the error squares of the model were obtained

as 11.9 and 8.23, respectively. Based on these results, it was
Knowledge Sharing
Table 5  (continued)

proven that the efficiency of the method was increased [78].

Ji et al. created a C-XGBoost model based on the clustering
Optimization

algorithm according to their three-stage XGBoost model,

Sub-Field

which also includes sales features. To achieve higher pre-

diction accuracy, an A-XGBoost model was created using

13
Applications of Artificial Intelligence in Inventory Management: A Systematic Review of the…

ARIMA for the linear part and an XGBoost model for the Material management is the optimization of inventory
nonlinear part. C-XGBoost and A-XGBoost models were through classification of materials to ensure continuity of
weighted to form the final model, CA-XGBoost model [77]. quality performance and to control the material distribu-
In order forecasting problem, a customer places an order tion cycle. In this context, using an earthquake disaster
for future production and shipment if a product is out of as a case study, Huang et al. divided emergency sup-
stock or temporarily unavailable. Product order short- plies (medical gauze, leech, bandage, alcohol, saline, and
ages are very important in inventory management as they blood pressure monitor) into three main categories such
affect the entire supply chain. According to Ntakolia et al. as importance, scarcity, and time to analyze them effec-
found that the best models for the back-order estimation tively. Their model provided a prediction accuracy of up
problem were the isotonic regression method and the post- to 92.45% compared to other models [88].
calibrated LGBM model. The explainability analysis of Inventory classification is used in many sectors such as
the best model was performed, and the results showed a healthcare, defense, and automotive production. Various meth-
similar performance depending on the area under the curve ods were applied to find a better classification model for obtain-
(AUC) value 0.95 of the RF, XGB, LGBM, and BB models. ing, organizing, and analyzing useful information by collecting
This result also introduced a significant contribution to the large amounts of data. Maathavan and Venkatraman imple-
accurate forecasting of future demand and backorders [80]. mented various machine learning methods to find a better clas-
Demand forecasting is used in knowledge discovery or sification model for electronic health records [91]. The results
knowledge sharing to obtain information between variables. showed that the KNN method achieved the highest performance
Lee et al. designed an inventory information discovery sys- in terms of precision, recall, and F1 score. According to the
tem to capture and predict information between variables average execution time of the encryption algorithms, KNN still
and extract the information learned from a ANN in the form has the best performance. The genetic algorithm was used to
of decision trees and explain the demand forecast result with estimate the weights of the criteria to perform the ABC inven-
the TREPAN algorithm [83]. tory classification and MCDM methods were preferred to cal-
culate the weighted score of the inventory items by Kaabi et al.
3.3 Inventory Classification [90]. The results show that the genetic algorithm-based models
outperform the existing classification models in terms of total
This section presents studies on the classification of cost and inventory turnover functions. In the model by Aktepe
inventories. The basis of efficient inventory management et al. for the classification of inventory items in terms of mul-
is based on accurate inventory classification. It provides tiple criteria, the inventory units are first classified according
benefits such as protecting inventories of critical raw to the decision rules of an expert system and then the k-means
materials and finished products, controlling inventories of clustering algorithm is used to make the second assignment
semi-finished products, and minimizing inventory costs [87]. When expert systems and k-means clustering algorithms
[23]. This way, companies have a competitive advantage are not sufficient for grouping some items, they have developed
as they can provide the best service to their customers. The a fuzzy rule-based method to make the final assignment.
traditional ABC method of classifying inventory consid- Product management includes all the strategic directions
ers only total amount of annual usage. It ignores other and practices that enable the development, marketing and
criteria and results in an inaccurate classification. Multi improvement of products. In this context, García-Barrios
Criteria Decision Making (MCDM) methods incorporating et al. studied the problem of sourcing impulse purchase
order size, delivery time, inventory costs, suppliers, etc., products and showed that this problem can be solved with
have been developed as a solution for this. Development of low-cost groupings based on the clustering process [92].
technology has led the companies to more efficient meth- They have shown that the proposed method can be used to
ods of classifying their inventories. The use of DL and cluster impulse purchase products more effectively.
ML methods provides promising results for classification Spare parts management plays an important role in main-
by automating complex decision-making processes and tenance planning and logistics activities. Decision makers
processing large amounts of data easily. can determine the optimal strategy for inventory manage-
For the sect. 3.3, 13 articles were identified and are ment by overcoming the problems such as explanation and
listed in Table 6. It shows the studies using AI and ML learning ability and criteria selection based on right clas-
methods. The results show that ML methods classify sification methods. In this manner, Zhang et al. concluded
inventories in more detail and provide efficient strategies that the order of importance of spare parts is same, and the
[87]. The k-means clustering algorithm, which splits the gravity values are very close when they are compared with
data into multiple groups with similar characteristics, is the actual gravity values to check the validity of the model
widely employed in the articles. DL methods have also [94]. Moreover, they ensured the correct evaluation of the
been applied in many studies. importance of these items by adding new spare parts.

13
 Ö. Albayrak Ünal et al.

3.4 Methods Used in Studies authors, they are actively working on this topic, and they are
leading the literature by developing new methods.
Table 7 lists the artificial intelligence methods used in this Keyword analysis is very important for a study. The
study and the number of uses. VOS (visualization of similarities) Viewer software, was
In the literature, many AI based methods have been used preferred to visualize the relationships between the articles.
for inventory management. Among these, the random forest It performs clustering techniques and provides a graphical
method is the most commonly used one with 9 occurrences. representation of bibliometric maps using custom labeling
The K-Means clustering algorithm, encountered in 8 stud- algorithms or density metaphors [97]. Based on this map,
ies, is the second most widely applied method. XGBoost one can see the most common variables of the inventory
and ANN, employed 7 times, are the third most preferred management concept and the relationship pattern between
methods followed by SVM, GA and RA with 6 times. On the these variables. It can be concluded that inventory manage-
other hand, RNN, CART, and DNN were used once. Note ment, machine learning, and demand forecasting have the
that the total number of artificial intelligence techniques is most similarities and connections among all items, as they
greater than the number of articles since some articles uti- have the largest nodes. It is clear from Fig. 7 that the concept
lizes more than one method. of inventory management is often associated with concepts
such as demand forecasting, classification, control, optimi-
zation, and artificial intelligence methods (deep learning,
4 Practical Implications machine learning, and reinforcement learning). According to
the distribution of keywords, location awareness, smart man-
This section examines the RQ answers in a more compre- ufacturing, big data, e-commerce, Internet of Things, and
hensive and detailed manner. deep reinforcement learning are the current areas of work.
RQ1: What is the current state of research in inventory RQ2. What are the areas and sub-areas in inventory man-
management? agement research where AI methods are being applied?
The purpose of this question is to obtain a quantitative AI methods, which are widely used in many industries
overview of the current work on this topic. A bibliometric and fields, affect almost all areas of inventory management.
analysis was conducted to answer the question. This analy- Considering the relevant literature, studies have been carried
sis provided a detailed data visualization that allowed the out in many areas such as operations management, inventory
perception of key characteristics of the literature. It also control, inventory visibility, supply chain management, opti-
defined the publication trends, source distributions, influ- mization, knowledge discovery, object recognition, estima-
ential authors, regions, and keywords of the articles included tion, classification, and etc. The contributions of AI methods
in the current study. In addition, the use of network analysis to the field and sub-fields of inventory management and their
revealed important research and relationships issues. impact on the literature are examined in more detail. As a
For publication trends, articles covering the years of result of the keyword analysis, it is observed that inventory
2012–2022 period and containing AI techniques in inventory management is mainly focused on three categories: inven-
management were examined. Since 2010–2012, considered tory problems, demand forecasting, and classification. For
the beginning of Deep Learning practical implementations, this reason, this study conducted a comprehensive investiga-
the number of published articles has increased. The growth tion by dividing these areas into subcategories.
was observed especially in 2021. This is because researchers The inventory problem articles are divided into data-
are getting familiar with the potential of these methods in driven approach, optimization, inventory control, inventory
inventory management studies and discovering that employ- policy, inventory visibility, and object recognition subsec-
ing such methods yield in a performance increase. In addi- tions. These fields are interrelated if meeting service quality
tion, the rapid development of artificial intelligence technol- requirements, managing product procurement and admin-
ogy has made inventory management processes intelligent. istration, and minimizing costs are being investigated. The
Although there exist many studies on inventory management main methods performed in this field are ML and DL. These
in the literature, the number of studies dealing with artificial methods are able to process huge amounts of data quickly
intelligence applications is limited. to consistently identify patterns and gain insights that may
Evaluation of the distribution of studies shows that most be too complex for the human mind to manage. The genetic
of the journals are production-related, indicating the active algorithm method, on the other hand, is an optimization
role of inventory management research. It appears that publi- method and was used along with the reinforcement learn-
cations from China account for more than a quarter of the 59 ing algorithm. The digital twin was also utilized with the
examined articles. It can be concluded that AI applications reinforcement learning algorithm. A detailed overview of
in inventory management are intensively pursued by Chinese these methods is given in Table 4.

13
 Table 6  Analysis of studies for the inventory classification field
Sub-Field Study Citations Algorithm Type Algorithms Goals/ Appoach

Material Management [88] 2 DL Back-Propagation Neural Network A neural network-based medical equipment management model for
(BPNN) emergency classification is proposed.
[89] 1 ML K-means Clustering Algorithm A material recognition method was developed for dynamic inven-
tory management of aerospace companies, including versatile and
general materials. Inventory classification based on the numerical
concept area was performed, which can automatically identify
clusters of materials with the net scale function.
Health Services/ Defense [90] 10 AI GA, WS, TOPSIS A benchmark dataset consisting of 47 items was used to test the
Industry/ Automotive MCDM model’s performance by proposing a hybrid method for inventory
Sector classification.
[91] – ML Support Vector Machine (SVM), MNB, DT, RF, GB, KNN An efficient encryption technique has been developed to secure user
data by classifying electronic health records with machine learning
techniques.
[87] 12 AI Expert Systems, A new classification algorithm was developed by integrating the clas-
ML Fuzzy Logic, sical ABC classification with various methods in a large defense
K-means Clustering Algorithm industry company.
[19] 79 MCDM, SAW, AHP In a case study conducted in an automotive company by integrat-
ML VIKOR, Naïve Bayes, Bayesian Network, ANN, SVM ing machine learning algorithms and MCDM methods, the SVM
method was found to have the best classification accuracy.
Inventory Control [17] 105 AI SA An inventory control system is proposed that simultaneous classifies
inventory and groups items by inventory cost and similarity, with
appropriate guidelines established for each product group.
MCABC Problem [18] 98 ML K-means Clustering Algorithm, SA Cluster analysis was used to progress at different levels of detail
AI in forming the preference order of the clusters defined for the
MCABC problem, and the simulated annealing method was used
to search for the optimal classification according to the hierarchy of
Applications of Artificial Intelligence in Inventory Management: A Systematic Review of the…

clusters from top to bottom.

Product Management [92] – ML K-means Clustering Algorithm, Products with similar demand, order, or cost characteristics were
grouped to find a near-optimal inventory grouping solution for
managing multiple impulse purchase SKUs.
Spare Part [93] 1 DL CNN A new approach to the classification of spare parts is presented to
perform multi-criteria classification based on a hierarchical struc-
ture through image recognition.
[94] 1 ML K-means Clustering Algorithm, The importance of maintenance spare parts was objectively classified
DL BPNN and evaluated.
Object Detection [95] 67 Computer Vision A new system has been introduced to recognize traffic signs from
Google Street View API images and create an inventory of traffic
signs.
[96] 63 Computer Vision Adaboost, The performance of three computer vision algorithms in detecting
Linear SVM, Nonlinear SVM and classification of traffic signs for US highways is presented and
validated.

13
 Ö. Albayrak Ünal et al.

Table 7  Number of methods used in the study

Methods Number of Methods Number of Methods Number
Usage Usage of Usage

RF 9 QRF 3 Internet of things 1

K-Means Clustering Algorithm 8 Digital Twin 3 LASSO 1
XGBoost 7 GAMLSS 2 CART​ 1
ANN 7 MLP 2 Automated innovization 1
SVM 6 BPNN 2 Fuzzy logic 1
GA 6 ETS 2 MNB 1
RA 6 Naïve/Naive Bayes 2 GB 1
KNN 5 Q-Learning 2 Expert systems 1
Time Series Models 5 Sarsa Algorithms 2 QuantReg 1
LSTM 5 DQN 2 PPO 1
CNN 5 DNN 2 GP 1
ARIMA 5 RNN 1 BB 1
DT 4 Graph Embedding 1 Inductive Learning Algorithm 1
DRL 4 Generative Adversarial Networks 1 SIFT algorithm 1
SA 3 Core Density Estimation 1 Yolov5 1

In the field of demand forecasting, the articles reviewed From the review of literature, it is clear that many AI
consist of the sub-areas of customer forecasting, product techniques have been applied to inventory management. In
forecasting, sales forecasting, order forecasting, and infor- this sense, the most common and effective method is the
mation exchange. In this sense, demand forecasts play an random forest method, which was used in 9 studies. It is a
important role in managing key operations. They aim to popular machine learning method introduced by Breiman
meet incoming demand in the shortest possible time and at in 2001 and used to develop predictive models [98]. RF, an
the lowest possible cost. Here, the methods of ML provide ensemble learning method, is widely used in many fields
better results as they can handle complex interdependencies such as computer vision, product demand forecasting, scrap
among many causal factors affecting the demand [57]. While feature filtering, and data mining [99] due to its excellent
CNN, LSTM, ANN and RNN were preferred more among performance and efficient training process. This unified
DL methods, KNN, QuantReg, QRF, RF, SVM, XGBoost, machine learning algorithm creates a group of trees where
CART and GAMLSS were preferred from ML methods. The each tree, combined with a set of tree classifiers, votes a unit
simple exponential smoothing, the quadratic exponential for the most popular class, and, the final ranking is obtained
smoothing, the feature synthesis, and Loess methods were by combining such results [100]. Cluster analysis is the sec-
widely implemented among all time series methods. Also, ond most commonly used method. Cluster analysis, in short
ML and DL methods were jointly applied in many studies, clustering, is referred as an unsupervised learning approach
while others used them separately. There are many stud- since it does not use label information. It provides infor-
ies that combine ML methods with time series methods. A mation about the data by dividing the objects into clusters,
detailed overview of these methods is presented in Table 5. ensuring that the objects in one cluster are more related than
In the field of inventory classification, material manage- objects in other clusters [101].
ment consists of sub-fields on a sectoral basis (healthcare, Following cluster analysis, XGBoost and ANN methods
automotive, defense), inventory control, object detection, and were used in 7 studies. Here, XGBoost is another ensemble
product management. While effective use of ML applications learning method that has achieved remarkable results in many
can be observed in this area, GA, fuzzy logic, expert systems research areas. It automatically implements multi-threading of
and simulated annealing from artificial intelligence methods CPU to perform parallel computation and optimize the algo-
are also adopted. These methods have been implemented rithm. This way, the training speed and prediction accuracy
along with ML and MCDM algorithms. In addition, SAW, of the models significantly improve [102]. ANN, on the other
AHP, TOPSIS and VIKOR are commonly preferred MCDM hand, is one of the most popular machine learning methods.
methods. Another topic in inventory classification that was It is an information processing technique that is used to obtain
applied with machine learning algorithms is Computer vision. patterns, information or connections in large amount of data.
A detailed overview of these methods is given in Table 6. It belongs to the class of black-box models, since ANNs do
RQ3. Which AI methods are being used in inventory not require knowledge of the physical parameters of the pro-
management studies? cess [103]. ANN-based methods are useful tools in modeling

13
Applications of Artificial Intelligence in Inventory Management: A Systematic Review of the…

various engineering systems for real-world conditions without a bibliometric analysis was initially performed to determine
having to solve complex mathematical models. the publication trend, source distribution, publications’
The results revealed that machine learning methods are country of origin, the most cited articles, and the most fre-
generally used in demand forecasting and classification, while quently used keywords. Finally, the techniques employed
reinforcement learning algorithms are preferred in inventory in the reviewed articles are discussed and current research
problems. In terms of methods’ variety, the field of demand developments are summarized. A detailed analysis of the
forecasting is quite rich, with 39 techniques. This is because research questions is also provided in the sect. 4.
some studies include various AI techniques to compare the It has been observed that lower accuracy rates, increased
results obtained from these methods that have a powerful effect, wasted labor and time, and higher inventory costs occurred
when applied separately or in combination with others. In terms in studies using statistical or time series methods compared
of AI techniques usage, inventory classification is the second to the studies using AI methods. While some other stud-
widely-used area with 23 algorithms. Here, the most common ies employed both methods in their studies and concluded
method is K-means clustering algorithm, while SVM, BPNN that AI methods offer improvements in predictive accuracy
and simulated annealing are other frequently implemented and achieve higher accuracy rates. Demand forecasting
methods. Hybrid techniques have also been employed in many errors, on the other hand, result in high inventory, storage,
studies. For example, Liu et al. classified products based on and transportation costs. For this type of problems, AI has
clustering analysis and then searched for the most appropriate been shown to quickly analyze large and diverse data sets,
solution with a simulated annealing algorithm [18]. increase the accuracy of demand forecasting, and enable effi-
Inventory problems ranked third with 17 techniques, with cient and flexible inventory management with lower inven-
Q-learning and Sarsa algorithms. For this, single methods tory and operational costs as well as providing customers
were utilized instead of multiple techniques with limited with faster response time and more contextual knowledge.
number of studies. This is because AI methods are recently Examining the distribution of keywords, we conclude
being replaced by ML and DL applications as they mostly that location awareness, smart manufacturing, Big Data,
yield better and reliable results. e-commerce, Internet of Things, digital twin, and DRL are
current areas of work. Among these, reinforcement learn-
ing algorithms are becoming increasingly popular in inven-
5 Conclusion tory management. Its use enables dynamic learning that
accounts problems with higher complexity. There is also
Given the developments in ML and DL algorithms, inter- a trend towards the Internet of Things and the digital twin.
est in these methods has greatly increased in many fields. Despite the limited use of digital twin technology in inven-
Inventory management and artificial intelligence concepts tory management, it has recently attracted a great interest
now occupy an important place and are being studied by and is growing rapidly worldwide.
many researchers parallel to the growing need with the In the field-based analysis, ML methods are typically
developing technology. In view of the breadth of the field, used in the areas of demand forecasting and classification,
this study is mainly descriptive and provides an assessment while reinforcement learning algorithms are preferred for
of the potential that AI methods can offer to solve problems inventory problems especially in inventory control, optimi-
in inventory management. This research study has exam- zation and stock visibility. Computer vision, on the other
ined the areas in which these methods are used without hand, is mainly considered technique in object recognition.
considering the best AI method. Not only studies employ- This study is intended to provide researchers and practi-
ing AI methods in inventory management are examined, tioners with a starting point and roadmap for AI techniques
but also the latest research trends are highlighted. in inventory management by highlighting the most popular
The aim of this study is to provide a systematic review research topics. It revealed that there is a gap in the applica-
of the literature focusing on the recent development and tion of AI methods in inventory management, and all appli-
application of AI methods in inventory management. For cations have not been comprehensively discovered yet. It
this purpose, 59 research studies were evaluated based on also shows that collaboration among researchers needs to be
pre-defined criteria as a result of the examination of 815 improved for increased performance. Nonetheless, current
articles from two major databases. These articles are divided study has some limitations. Use of two different databases
into three categories such as inventory problems, demand during the article filtering procedure causes a confusion in
forecasting, and inventory classification. Considering the keyword, publication period and language selection. Also,
categories of reviewed articles, answers were obtained to including only articles in the study may cause some other
the three research questions of the current situation in inven- valuable sources such as books, conference papers and busi-
tory management (RQ1), the field and sub-fields used in the ness reports, and etc. to be overlooked. Therefore, this study
study (RQ2), and the AI methods used (RQ3). In this sense, may be expanded including additional sources.

13
 Ö. Albayrak Ünal et al.

Appendix Funding This research did not receive any specific grant from funding
agencies in the public, commercial, or not-for-profit sectors.

Full names of the algorithms used in the study.

Table 8
References
1. Perez HD et al (2021) Algorithmic approaches to inventory man-
Table 8  Full names of the algorithms used in the study agement optimization. Processes 9(1):102
2. Singh D, Verma A (2018) Inventory management in supply
Acronym Full Name
chain. Mater Today 5(2):3867–3872
AHP Analytic Hierarchy Process 3. Haberleitner H, Meyr H, Taudes A (2010) Implementation of a
demand planning system using advance order information. Int J
AI Artificial Intelligence Prod Econ 128(2):518–526
ANN Artificial Neural Network 4. Abbasimehr H, Shabani M, Yousefi M (2020) An optimized
ARIMA Autoregressive Integrated Moving Average model using LSTM network for demand forecasting. Comput
ARIMAX Multivariate ARIMA Ind Eng 143:106435
5. Beutel A-L, Minner S (2012) Safety stock planning under causal
BPNN Back-Propagation Neural Network demand forecasting. Int J Prod Econ 140(2):637–645
BB Balanced Blagging 6. Sridhar P, Vishnu C, Sridharan R (2021) Simulation of inventory
CART​ Classification and Regression Tree management systems in retail stores: a case study. Mater Today
CNN Convolutional Neural Network 47:5130–5134
7. Granillo-Macías R (2020) Inventory management and logis-
DL Deep Learning tics optimization: a data mining practical approach. LogForum
DNN Deep Neural Network 16(4):535–547
DQN Deep-Q-Network 8. Nallusamy S (2021) Performance measurement on inventory
DRL Deep Reinforcement Learning management and logistics through various forecasting tech-
niques. Int J Perform Eng 17(2):216–228
DT Decision Tree 9. Acosta ICG et al (2018) Design of an inventory management sys-
ETS Exponential Smoothing tem in an agricultural supply chain considering the deterioration
GP Gaussian Process of the product: the case of small citrus producers in a developing
GAMLSS Generalized Additive Models for Position, Scale, and country. J Appl Eng Sci 16(4):523–537
Shape 10. Varghese V et al (2012) Applying actual usage inventory man-
agement best practice in a health care supply chain. Int J Supply
GA Genetic Algorithm Chain Manage 1(2):1–10
GB Gradient Boosting 11. Xie C, Wang L, Yang C (2021) Robust inventory management
KNN K-Nearest Neighbor with multiple supply sources. Eur J Oper Res 295(2):463–474
LSTM Long-Short Term Memory 12. Tranfield D, Denyer D, Smart P (2003) Towards a methodology
for developing evidence-informed management knowledge by
ML Machine Learning means of systematic review. Br J Manag 14(3):207–222
MLP Multilayer Perceptron 13. Evangelista P, Durst S (2015) Knowledge management in envi-
MNB Multinomial Naive Bayes ronmental sustainability practices of third-party logistics service
MCABC Multi-Criteria ABC Classification providers. Vine 45(4):509–529
14. Bryman A (2007) The research question in social research: what
MCDM Multi Criteria Decision Making is its role? Int J Soc Res Methodol 10(1):5–20
PPO Proximal Policy Optimization 15. Nguyen KTP, Medjaher K (2019) A new dynamic predictive
QuantReg Quantile Regression maintenance framework using deep learning for failure prognos-
QRF Quantile Regression Forest tics. Reliab Eng Syst Saf 188:251–262
16. Tabernik D, Skocaj D (2020) Deep learning for large-scale traf-
RF Random Forest fic-sign detection and recognition. IEEE Trans Intell Transp Syst
RANSAC Random Sample Consensus 21(4):1427–1440
RNN Recurrent Neural Network 17. Mohammaditabar D, Ghodsypour SH, O’Brien C (2012) Inven-
SAW Simple Additive Weighting tory control system design by integrating inventory classification
and policy selection. Int J Prod Econ 140(2):655–659
SIFT Scale Invariant Feature Transform 18. Liu JP et al (2016) A classification approach based on the out-
SARSA State–action–reward–state–action ranking model for multiple criteria ABC analysis. Omega-Inter-
SA Simulated Annealing national J Manage Sci 61:19–34
SVM Support Vector Machine 19. Kartal H et al (2016) An integrated decision analytic frame-
work of machine learning with multi-criteria decision making
TOPSIS Technique for Order Preference by Similarity to Ideal for multi-attribute inventory classification. Comput Ind Eng
Solution 101:599–613
WS Weighted Sum 20. Aggarwal SC (1974) A review of current inventory theory and
XGBoost Extreme Gradient Boosting its applications. Int J Prod Res 12(4):443–482

13
Applications of Artificial Intelligence in Inventory Management: A Systematic Review of the…

21. Giannoccaro I, Pontrandolfo P (2002) Inventory management 40. Priore P et al (2019) Applying machine learning to the dynamic
in supply chains: a reinforcement learning approach. Int J Prod selection of replenishment policies in fast-changing supply chain
Econ 78(2):153–161 environments. Int J Prod Res 57(11):3663–3677
22. Cachon GP, Fisher M (2000) Supply chain inventory man- 41. Demey YT, Wolff M (2017) SIMISS: a model-based search-
agement and the value of shared information. Manage Sci ing strategy for inventory management systems. IEEE Internet
46(8):1032–1048 Things J 4(1):172–182
23. Partovi FY, Anandarajan M (2002) Classifying inventory 42. Merrad Y et al (2020) A real-time mobile notification system for
using an artificial neural network approach. Comput Ind Eng inventory stock out detection using SIFT and RANSAC. Int J
41(4):389–404 Interactive Mobile Technol. https://​doi.​org/​10.​3991/​ijim.​v14i05.​
24. Giannoccaro I, Pontrandolfo P, Scozzi B (2003) A fuzzy echelon 13315
approach for inventory management in supply chains. Eur J Oper 43. Kalinov I et al (2020) WareVision: CNN barcode detection-based
Res 149(1):185–196 uav trajectory optimization for autonomous warehouse stocktak-
25. Yang L, Li H, Campbell JF (2020) Improving order fulfillment ing. IEEE Rob Autom Lett 5(4):6647–6653
performance through integrated inventory management in a 44. Giaconia C, Chamas A (2022) GAIA: great-distribution artificial
multi-item finished goods system. J Bus Logistics 41(1):54–66 intelligence-based algorithm for advanced large-scale commer-
26. Wang Z, Mersereau AJ (2017) Bayesian inventory management cial store management. Appl Sci-Basel 12(9):4798
with potential change-points in demand. Prod Oper Manage 45. Kosanoglu F, Atmis M, Turan HH (2022) A deep reinforcement
26(2):341–359 learning assisted simulated annealing algorithm for a mainte-
27. Calle M et al (2016) Integrated management of inventory and nance planning problem. Annals of Operations Research 15:1–32
production systems based on floating decoupling point and real- 46. Deutsch J, He D (2017) Using deep learning-based approach to
time information: a simulation based analysis. Int J Prod Econ predict remaining useful life of rotating components. IEEE Trans
181:48–57 Syst Man Cybernetics: Syst 48(1):11–20
28. KP ASR, Nayak N (2017) A study on the effectiveness of inven- 47. Tao F et al (2019) Digital twin-driven product design framework.
tory management and control system in a milk producer organi- Int J Prod Res 57(12):3935–3953
sation. Int J Logistics Syst Manage 28(2):253–266 48. Fuller A et al (2020) Digital twin: enabling technologies, chal-
29. Rana R, Oliveira FS (2015) Dynamic pricing policies for inter- lenges and open research. IEEE Access 8:108952–108971
dependent perishable products or services using reinforcement 49. Abosuliman SS, Almagrabi AO (2021) Computer vision assisted
learning. Expert Syst Appl 42(1):426–436 human computer interaction for logistics management using deep
30. Pirayesh Neghab D, Khayyati S, Karaesmen F (2022) An inte- learning. Comput Electr Eng 96:107555
grated data-driven method using deep learning for a newsven- 50. Mao J et al (2020) The importance of public support in the imple-
dor problem with unobservable features. Eur J Operational Res mentation of green transportation in the smart cities. Computa-
302(2):482–496 tional Intell. https://​doi.​org/​10.​1111/​coin.​12326
31. Bandaru S et al (2015) Generalized higher-level automated 51. Tian X, Wang H, Erjiang E (2021) Forecasting intermittent
innovization with application to inventory management. Eur J demand for inventory management by retailers: a new approach.
Oper Res 243(2):480–496 J Retailing Consumer Serv 62:102662
32. Kara A, Dogan I (2018) Reinforcement learning approaches for 52. do Rego JR, De Mesquita MA (2015) Demand forecasting and
specifying ordering policies of perishable inventory systems. inventory control: a simulation study on automotive spare parts.
Expert Syst Appl 91:150–158 Int J Prod Econ 161:1–16
33. Zwaida TA, Pham C, Beauregard Y (2021) Optimization of 53. Tangtisanon P (2018) Web service based food additive inven-
inventory management to prevent drug shortages in the hospi- tory management with forecasting system. in 2018 3rd Interna-
tal supply chain. Appl Sci 11(6):2726 tional Conference on Computer and Communication Systems
34. Katanyukul T (2014) Ruminative reinforcement learning: (ICCCS). IEEE
improve intelligent inventory control by ruminating on the 54. Yu Q et al (2017) Application of long short-term memory
past. J Comput 9(7):1530–1535 neural network to sales forecasting in retail—a case study. in
35. De Moor BJ, Gijsbrechts J, Boute RN (2021) Reward shaping International Workshop of Advanced Manufacturing and Auto-
to improve the performance of deep reinforcement learning mation. Springer
in perishable inventory management. Eur J Operational Res 55. Kumar A, Shankar R, Aljohani NR (2020) A big data driven
301(2):535–545 framework for demand-driven forecasting with effects of mar-
36. Boute RN et al (2021) Deep reinforcement learning for inven- keting-mix variables. Ind Mark Manage 90:493–507
tory control: a roadmap. Eur J Operational Res 298(2):401–412 56. Seyedan M, Mafakheri F, Wang C (2022) Cluster-based
37. Gijsbrechts J et al (2021) Can deep reinforcement learning demand forecasting using bayesian model averaging: an ensem-
improve inventory management? Performance on lost sales, ble learning approach. Decis Anal J 3:100033
dual-sourcing, and multi-echelon problems. M SOM-Manuf 57. Feizabadi J (2022) Machine learning demand forecast-
Serv Operations Manag. https://​doi.​org/​10.​1287/​msom.​2021.​ ing and supply chain performance. Int J Logistics Res Appl
1064 25(2):119–142
38. Kegenbekov Z, Jackson I (2021) Adaptive supply chain: demand– 58. Deng CN, Liu YJ (2021) A deep learning-based inventory
supply synchronization using deep reinforcement learning. Algo- management and demand prediction optimization method for
rithms 14(8):240 anomaly detection. Wirel Commun Mobile Comput. https://​
39. Meisheri H et al (2022) Scalable multi-product inventory control doi.​org/​10.​1155/​2021/​99693​57
with lead time constraints using reinforcement learning. Neural 59. Kack M, Freitag M (2021) Forecasting of customer demands
Comput Appl 34(3):1735–1757 for production planning by local k-nearest neighbor models.
Int J Prod Econ 231:107837

13
 Ö. Albayrak Ünal et al.

60. Bala PK (2012) Improving inventory performance with cluster- and gradient boosting machine learning techniques. J Big Data
ing based demand forecasts. J Modelling Manage 7(1):23–37 7(1):1–22
61. Lee CY, Liang CL (2018) Manufacturer’s printing forecast, 82. O’Neil S et al (2016) Newsvendor problems with demand shocks
reprinting decision, and contract design in the educational pub- and unknown demand distributions. Decis Sci 47(1):125–156
lishing industry. Comput Ind Eng 125:678–687 83. Lee CKM et al (2017) Design and development of inven-
62. Abbasi B et al (2020) Predicting solutions of large-scale tory knowledge discovery system. Enterp Inform Syst
optimization problems via machine learning: a case study 11(8):1262–1282
in blood supply chain management. Comput Operations Res 84. Van Belle J, Guns T, Verbeke W (2021) Using shared sell-
119:104941 through data to forecast wholesaler demand in multi-echelon
63. van Steenbergen RM, Mes MRK (2020) Forecasting demand supply chains. Eur J Oper Res 288(2):466–479
profiles of new products. Decis Support Syst 139:113401 85. Tang Y-M et al (2022) Integrated smart warehouse and manu-
64. Zhu XD et al (2021) Demand forecasting with supply-chain infor- facturing management with demand forecasting in small-scale
mation and machine learning: evidence in the Pharmaceutical cyclical industries. Machines 10(6):472
Industry. Prod Oper Manage 30(9):3231–3252 86. Ecemiş O, Irmak S (2018) Paslanmaz çelik sektörü satış tah-
65. Benhamida FZ et al (2021) Demand forecasting tool for inventory mininde veri madenciliği yöntemlerinin karşılaştırılması. Kilis
control smart systemsy. J Commun Softw Syst 17(2):185–196 7 Aralık Üniversitesi Sosyal Bilimler Dergisi. 8:148–16915
66. Zhang P et al (2021) Pharmaceutical cold chain management 87. Aktepe A et al (2018) An inventory classification approach com-
based on blockchain and deep learning. J Internet Technol bining expert systems, clustering, and fuzzy logic with the abc
22(7):1531–1542 method, and an application. S Afr J Ind Eng 29(1):49–62
67. Ulrich M et al (2021) Distributional regression for demand fore- 88. Huang B, Gan W, Li Z (2021) Application of medical material
casting in e-grocery. Eur J Oper Res 294(3):831–842 inventory model under deep learning in supply planning of public
68. Li N et al (2021) A decision integration strategy for short-term emergency. IEEE Access 9:44128–44138
demand forecasting and ordering for red blood cell components. 89. Wang A, Gao XD (2021) A variable-scale dynamic clustering
Oper Res Health Care 29:100290 method. Comput Commun 171:163–172
69. Ran H (2021) Construction and optimization of inventory man- 90. Kaabi H, Jabeur K, Ladhari T (2018) A genetic algorithm-based
agement system via cloud-edge collaborative computing in sup- classification approach for multicriteria ABC analysis. Int J
ply chain environment in the internet of things era. PLoS ONE Inform Technol Decis Mak 17(6):1805–1837
16:e0259284 91. Maathavan KSK, Venkatraman S (2022) A secure encrypted clas-
70. Sun X et al (2021) RBC inventory-management system sified electronic healthcare data for public cloud environment.
based on XGBoost model. Indian J Hematol Blood Transfus Intell Autom Soft Comput 32(2):765–779
37(1):126–133 92. García-Barrios D et al (2021) A machine learning based method
71. Aktepe A, Yanik E, Ersoz S (2021) Demand forecasting applica- for managing multiple impulse purchase products: an inventory
tion with regression and artificial intelligence methods in a con- management approach. J Eng Sci Technol Rev 14(1):25–37
struction machinery company. J Intell Manuf 32(6):1587–1604 93. Yang K et al (2021) Multi-criteria spare parts classification
72. Galli L et al (2021) Prescriptive analytics for inventory manage- using the deep convolutional neural network method. Appl Sci
ment in health. J Oper Res Soc 72(10):2211–2224 11(15):7088
73. Eljaouhari A et al (2022) Demand forecasting application with 94. Zhang S et al (2020) Importance degree evaluation of spare parts
regression and iot based inventory management system: a case based on clustering algorithm and back-propagation neural net-
study of a semiconductor manufacturing company. Int J Eng Res work. Math Problems Eng. https://​doi.​org/​10.​1155/​2020/​61618​
Afr 60:189–210 25
74. Sucharitha RS, Lee S (2022) GMM clustering for in-depth food 95. Balali V, Ashouri Rad A, Golparvar-Fard M (2015) Detection,
accessibility pattern exploration and prediction model of food classification, and mapping of U.S. traffic signs using google
demand behavior. Socio-Economic Plan Sci. https://​doi.​org/​10.​ street view images for roadway inventory management. Vis Eng
48550/​arXiv.​2202.​01347 3(1):1–18
75. Wang Z et al (2022) Development and validation of a machine 96. Balali V, Golparvar-Fard M (2016) Evaluation of multiclass traf-
learning method to predict intraoperative red blood cell transfu- fic sign detection and classification methods for us roadway asset
sions in cardiothoracic surgery. Sci Rep 12(1):1–9 inventory management. J Comput Civil Eng 30(2):04015022
76. Ulrich M et al (2022) Classification-based model selection in 97. Van Eck NJ et al (2010) A comparison of two techniques for
retail demand forecasting. Int J Forecast 38(1):209–223 bibliometric mapping: multidimensional scaling and VOS. J Am
77. Ji S et al (2019) An application of a three-stage XGboost-based Soc Inform Sci Technol 61(12):2405–2416
model to sales forecasting of a cross-border e-commerce enter- 98. Breiman L (2001) Random forests. Mach Learn 45(1):5–32
prise. Math Probl Eng. https://​doi.​org/​10.​1155/​2019/​85032​52 99. Bai J et al (2022) Multinomial random forest. Pattern Recogn
78. Wang S, Yang Y (2021) M-GAN-XGBOOST model for sales 122:108331
prediction and precision marketing strategy making of each prod- 100. Liu Y, Wang Y, Zhang J (2012) New machine learning algorithm:
uct in online stores. Data Technol Appl 55(5):749–770 Random forest. International Conference on Information Com-
79. Kim M et al (2022) Framework of 2D KDE and LSTM-based puting and Applications. Springer, Berlin
forecasting for cost-effective inventory management in smart 101. Hong JS, Lie CH (1993) Joint reliability-importance of two edges
manufacturing. Appl Sci-Basel 12(5):2380 in an undirected network. IEEE Trans Reliab 42(1):17–23
80. Ntakolia C et al (2021) An explainable machine learning model 102. Ma M et al (2021) XGBoost-based method for flash flood risk
for material backorder prediction in inventory management. Sen- assessment. J Hydrol 598:126382
sors 21(23):7926 103. Jawad J, Hawari AH, Zaidi SJ (2021) Artificial neural network
81. Islam S, Amin SH (2020) Prediction of probable backorder modeling of wastewater treatment and desalination using mem-
scenarios in the supply chain using distributed random forest brane processes: a review. Chem Eng J 419:129540

13
Applications of Artificial Intelligence in Inventory Management: A Systematic Review of the…

Publisher’s Note Springer Nature remains neutral with regard to author(s) or other rightsholder(s); author self-archiving of the accepted
jurisdictional claims in published maps and institutional affiliations. manuscript version of this article is solely governed by the terms of
such publishing agreement and applicable law.
Springer Nature or its licensor (e.g. a society or other partner) holds
exclusive rights to this article under a publishing agreement with the

13

View publication stats