The American Journal of Engineering and Technology
6
https://www.theamericanjournals.com/index.php/tajet
TYPE
Original Research
PAGE NO.
6-20
10.37547/tajet/Volume07Issue03-02
OPEN ACCESS
SUBMITED
01 January 2025
ACCEPTED
02 February 2025
PUBLISHED
03 March 2025
VOLUME
Vol.07 Issue03 2025
CITATION
Md Zahidur Rahman Farazi. (2025). Enhancing supply chain resilience with
multi-agent systems and machine learning: a framework for adaptive
decision-making. The American Journal of Engineering and Technology,
7(03), 6
–
https://doi.org/10.37547/tajet/Volume07Issue03-02
COPYRIGHT
© 2025 Original content from this work may be used under the terms
of the creative commons attributes 4.0 License.
Enhancing supply chain
resilience with multi-agent
systems and machine
learning: a framework for
adaptive decision-making
Md Zahidur Rahman Farazi
Department of Information Systems and Operations Management, The
University of Texas at Arlington
Abstract:
The research focuses on how Multi-Agent
Systems (MAS) coupled with Machine Learning (ML) can
help manage the challenges and risks associated with
new-generation supply chains networks. The proposed
MAS-ML framework improves flexibility, adaptability,
and predictiveness in essential roles in supply chain
management (SCM), including demand forecasting,
inventory management, production planning, and SCM
logistics. The framework is based on decentralised
decision-making where each agent is responsible for a
particular supply chain activity but employs real-time
data foresight from the ML model to streamline the
activities. This decentralisation enables resilience in
supply chains, which can experience events such as
demand variability and transportation disruptions.
MAS-ML is presented in this paper as the solution
capable of enhancing supply chain performance,
reliability, and cost optimisation in situations
characterised by risk and uncertainty, such as the
current global pandemic. In addition, this paper
presents potential research areas, such as the
integration of more enhanced deep learning algorithms,
the extension of proposing MAS-ML into other sectors,
and the addressing of ethical and transparency concerns
associated with AI-based decision-making systems. The
proposed
MAS-ML
framework
improves
the
adaptability and resiliency of supply chains, providing a
flexible solution for modern supply chain problems.
Keywords:
Multi-Agent Systems (MAS), Machine
Learning (ML), Supply Chain Management, Demand
Forecasting, Inventory Management, Production
The American Journal of Engineering and Technology
7
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
Planning, Logistics Optimisation, Supply Chain
Flexibility, Decentralised Decision-Making, Supply
Chain Resilience, Predictive Analytics, Random Forest,
Gradient Boosting.
Introduction:
Decision-making in today's global supply
chain
environment
involves
robust
strategic
framework and techniques than traditional supply
chain models (Katsaliaki et al., 2022). The proposed
MAS-ML framework is a holistic method that addresses
the
adaptability
incorporating
predictive
computational ability. Supply chains are the principles
of the global economy, which allow the generation of
materials, the distribution of products, and the
delivery of them to clients worldwide. More recently,
though, the supply chains have been threatened in
ways that have opened the supply chain experts to
new forms of management. COVID-19 has disrupted
operations worldwide, resulting in factory closures,
transportation constraints, and heightened demand
for specific goods, including personal protection
equipment (Camur et al., 2023). The pandemic caused
severe supply chain disruptions, with many companies
facing delays in production and shipping (Xu et al.,
2020). For instance, global automotive manufacturing
slowed significantly due to factory shutdowns in key
regions like China, the US, and Europe.
Many businesses were forced to adopt new strategies,
such as digitalisation to maintain supply chain
resilience during the crisis. Supply chains grow in size
and sophistication, and demand fluctuation, inventory
status
and
transportation
complications
are
challenging to forecast and control (Ghadge et al.,
2020). Traditional supply chain systems with standard
planning rely on accurate forecasts, and a centralised
control system cannot handle these challenges in real-
time (Epiphaniou et al., 2020). For instance, rigid
production
planning
and
arithmetic
demand
forecasting approaches regularly overlook consumer
demand fluctuations and transportation chain
disruptions.
The major problems of modern supply chains are
uncertainty. The current demand forecasting models
have some drawbacks in predicting the consumers’
changing behaviour, primarily influenced by digital
platforms and volatile economic environments
(Kalkanci., 2011). Supply chain managers mainly deal
with global transportation networks, where delays,
lockups and cost volatility are strangers. The older
supply chain management systems that used historical
parameters
and
forecasts
cannot
accurately
accommodate these variables (Katsaliaki et al., 2022).
In this regard, conventional decision-making models
can be ineffective in many settings because they do not
allow one to predict the appearance of a new factor in
the process. That is why, with increasing levels of supply
chain integration, the impact of such inefficiencies is
much more significant. Thus, new approaches, which
may effectively accommodate flexibility emerging due
to uncertainty, are deemed necessary.
This work proposes a novel approach known as MAS-
ML o address these challenges affecting supply chain
management in the current environment of fluctuating
uncertainty level. The integration of Multi-Agent
Systems with Machine Learning enables the
enhancement of reliability and flexibility of supply chain.
Integrating MAS and ML for the development of the
MAS-ML approach makes it possible to increase
flexibility and adaptability that include demand
forecasting,
inventory
management,
production
planning, and logistics functions. Therefore, logistics
agents can use of predictive models in order to find
optimal routes that can improve efficiency in delivery of
products. MAS linked with ML provides a strong
background useful for addressing more complicated and
supply chains (Pasupuleti et al., 2024).
MAS and ML add a new perspective in supply chain
management due to the flexibility of handling the
modern supply chain nature, which is full of
uncertainties (Stoychev, 2023). Implementation of the
MAS framework would enable each of the supply chain
agents to operate autonomously, therefore ensuring
the fast and effective decision-making regarding the
real-time data that constantly flows into the system
(Rajbala et al., 2023). This research presents a significant
contribution to literature as it seeks to understand the
factors responsible for the growing concern of supply
chain management in the modern world economy. The
overall purpose of the MAS-ML is to assist the
companies
in
improving
the
organisational
performance, reducing costs and enhancing the
customer satisfaction due to more efficient supply
chain. Furthermore, the adoption of ML models into the
MAS framework helps the supply chain managers to
reduce the risk that arises when their supply chain is
disrupted, hence improving operations resilience
(Farazi, 2024).
Literature Review
Supply Chain Resilience
Supply chain responsiveness can be defined as the
capability of the supply chain to protect against
potential disruptions, mitigate them, and continue with
business operations. Today, supply chains are highly
susceptible to disruptions caused by factors like the
COVID-19 outbreak (Raja Santhi and Muthuswamy,
2022). As the advent of the COVID-19 pandemic has
The American Journal of Engineering and Technology
8
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
shown, there is a need for robust and flexible supply
chains that are ready to face rapid fluctuations in
demand, production and delivery (Frederico, 2021).
Industry 4.0 technologies are a robust technique that
suggests digitising human activity systems through
intelligent technologies such as IoT, AI, and data
analytics for real-time monitoring and decision-
making. Predictive analytics, one of the main segments
of Industry 4.0, enables the supply chain manager to
anticipate disruptions that could occur and plan how the
operations will look through the use of past and present
data (Spieske and Birkel, 2021). For example, IoT devices
can offer Managers real-time insights into inventory
availability, while AI-driven systems can provide
different suppliers or transportation routes during
disruptions.
Figure 1: Industrial revolutions from Industry 1.0 to Industry 5.0 (Source: Folgado et al., 2024)
The other important factor that defines supply chain
vulnerability is that many supply chain managers need
help knowing where to begin when handling risk (Wu
and Chen, 2014). When it comes to decision-making in
the context of supply chain management, the reason
itself is bounded by the fact that no manager can make
a decision when encountering supply chain disruptions
or uncertainty in the market. Hence, there is a need for
systems that can use data to counter human cognitive
biases or shortcomings in real-time.
Figure 2: Supply Chain Resilience Source: https://www.altexsoft.com/blog/supply-chain-resilience/
Multi-Agent Systems (MAS)
Multi-agent systems (MAS) are distributed, self-
organised systems where numerous agents are
associated with various roles or activities in a supply
chain, like inventories, production, or transportation
(Gui et al., 2024). MA supports real-time decision-
making since the agents can work autonomously while
acquiring and exchanging data to meet goal-oriented
supply chain goals (Nitsche et al., 2023). However, the
decentralised functionality of MAS makes it useful in
supply chains because various functions may need to
respond to new circumstances or changes in demand
promptly. This means that the development of MAS in
the supply chain emphasises agents that adapt to
different strategies. This also allows agents to see what
has happened in their surroundings, gain knowledge of
the events, and plan their actions for the situation (Lee
et al., 2019). This flexibility makes MAS highly suitable
for solving uncertainties and dynamic environments
because agents can operate in real-time without
resorting to a higher authority.
Figure 3: Multi-Agent System (Source: Kishore et al., 2006)
The American Journal of Engineering and Technology
9
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
The current MAS design for the supply chain
application was decentralised decision making, where
every agent is capable of making its own decisions
based on the role it plays in the supply chain
application. This makes the supply chain adaptive to
market forces and conditions, thus leading to
increased
flexibility
of
the
supply
chain
(Gružauskas,2020). Artificial Intelligence (AI) has come
to be regarded as an empowering technology within
the supply chain field to enhance the forecast models
and support functions in the decision-making process
(Mahraz, 2022). Random forests and Neural networks
are some of the applications of the ML in the supply
chain management to analyse big data, and identity
hidden trends (Ni et al., 2020). Random Forests have
been implemented in the supply chain management
context for demand fluctuation forecasting and
inventory control (Makkar et al., 2020). For instance, a
supply chain manager uses Random Forests to forecast
the demand for a particular product during a holiday
and subsequently order the required stock (Kosasih
and Brintrup, 2022). This makes it easier to predict
which products are likely to be popular and which are
likely to be unpopular, and this leads to minimisation
of situations where stock out or overstocking occurs
frequently.
Machine Learning (ML) in Supply Chains
Neural Network techniques in ML are known to be
effective in identifying non-linear patterns in the supply
chain data. It illustrates how Neural Networks can
simulate the interaction between market demand,
production capabilities, and supplier lead time to
enhance production scheduling and procurement.
Neural Networks can update their knowledge from new
data, making it easier to improve on the forecast made
if conditions change frequently. The application of ML in
the supply chain results in improved decision-making
decreased operational executions, and significant
improvement in supply chain activity (Quayson et al.,
2023). Understanding the interplay between contracts
and behaviour supplies an essential perspective on
supply chain dynamics. Many supply chain members,
including the supplier, manufacturer, and distributors,
develop legal and business contracts that define risk,
responsibility and reward (Farazi, 2024). However, a
fixed traditional contract model could more effectively
address risks, uncertainty, and what people do in the
real world. That is where the role of Behavioral
Economics, which looks at how people depart from
rational choice, comes in.
Figure 4: ML system configuration (Sharma et al., 2020).
Chen and Rong (2020) look into factors such as
contract complexity and individual behaviour that
affect supply chain performance. For instance, the
complexity
of
contracts,
such
as
extensive
documentation,
exposes
the
parties
to
misunderstanding or misinterpretation, causing
inefficiency and conflicts among the supply chain
members (Li et al., 2020). Furthermore, bounded
rationality the finite capabilities of the decision-makers
may lead to robust decisions. Some of these
behavioural factors must be considered whenever
contracts are being developed to fit into the supply
chain and in a capability /constrained environment
(Chen, 2013).
According to behavioural contract theory, contracts
ought to be made more fundamental and should also
be made to bring out the concept of alignment of
incentives. For example, performance-based deals
with incentives that can be points such as on-time
delivery or expense reduction are likely to reduce the
impact of bounded rationality and enhance overall
supply chain effectiveness (Li et al., 2009). Thus, it
enables the companies to identify the behavioural
characteristics of supply chain partners and develop
contracts that can facilitate cooperation and minimise
uncertainties in the supply chain. The review of supply
chain resilience, Multi-Agent Systems (MAS), Machine
Learning (ML), and contract design reveals a clear trend
towards incorporating advanced technologies and
behavioural insights to optimise decision-making in
complex and dynamic supply chains.
MAS and ML is a promising approach for real-time and
decentralised
decision-making
for
achieving
cooperation throughout contracting networks (Farazi,
2024). In combination, these methods provide a full
solution to contemporary problems of managing supply
chains, thus increasin
g the organisations’ resistance to
unpredictable situations.
The American Journal of Engineering and Technology
10
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
Proposed MAS-ML Framework
Figure 5: Proposed MAS Framework
The decision making process in SCM is most often
agility and flexibility oriented in the present world due
to increased volatility in supply chain. The integration
of both Multi-Agent Systems (MAS) and Machine
Learning (ML) is proposed to address these problems
and provide a robust and efficient approach to supply
chain exposure. In MAS, supply chain functions such as
demand forecasting, inventory control, manufacturing
planning, and logistics are considered as objects that
can act independently and make decisions. The
Random Forest and Gradient Boosting Machines
scrutinise all the historical data along with the current
data and the agents get prompt and forecasted
decisions. MAS-ML models possess the capability for
near real-time changes and thus offer greater flexibility
and responsiveness in supply chain operations most
notably in the market context.
Machine Learning Integration for Predictive Analytics
Forecasting of demand is an essential feature of the
supply chain management. Random Forest and GBM in
the MAS framework can better estimate demand using
historical data and trends (Seyedan and Mafakheri,
2020). The demand forecasting models created actual
time prediction utilised by the MAS framework to
present demand requirement in the future (Zohra
Benhamida et al., 2021). These forecasts are useful
especially to agents such as inventory and production
managers for the purpose of ensuring that inventory
matches the supply chain management needs.
Another strategic function of the organisational
structure that the ML models in the framework include
is Inventory Optimisation (Chowdhury et al., 2024). It
allows the inventory agent to forecast for the right
quantities of stock. The inventory agent defines the
stock using the prediction from machine learning so
that the agent can reduce supplies depending on the
fluctuations in demand (Sakib, 2021). This approach
optimises costs while ensuring that products are
available to meet customer demand, thus improving
overall supply chain performance (Li and Chen, 2020).
Production Scheduling is related to demand forecasting
and inventory management since the former depends
on the latter to be accurate. Once the demand
prediction is produced, the production agent computes
the
"Production_needed"
and
the
"Production_schedule" to ensure the production
capacity is efficiently used. The simulation made it
easier to explain how to forecast demand in scheduling
the production process, the available material, and the
workforce within limited production abilities. Such
flexibility in product scheduling enables the supply chain
to quickly meet the higher or lower demand in the
market without undue lead time.
The MAS implement Logistics and Transportation
Optimisation as another primary function wherein the
use of machine learning models is done to reduce
transportation costs and routes (Adi et al., 2021). The
transportation agents used supervised machine learning
to forecast the optimal means of transporting the
merchandise depending on cost and time (Barua et al.,
2020). These predictions enabled the logistics agent to
optimise transportation plans by constantly adapting
costs and delivery time. In addition to cutting transport
costs, real-time optimisation capabilities improve
logistics' effectiveness, dependability, and adaptability
to demand or supply shock spikes.
Agent-Based Decision-Making
The primary stages in the MAS-ML framework are Agent
Interaction and Coordination. The decision-making is
decentralised, and each agent works independently,
although he may consult and provide information to
other agents (Antons and Arlinghaus, 2022). On the
other hand, the logistics agent works closely with the
inventory and production agents to ensure that all the
right products reach the appropriate point of sale at an
appropriate time (Aliawadi and SINGH, 2021). It allows
The American Journal of Engineering and Technology
11
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
real-time decision-making without any central
authority, as with traditional organisations. The above
coding simulation depicts how every agent operates to
ensure sufficient supply chain flow, not only in cases
where demand and logistics conditions shift.
The MAS framework helped exemplify the adaptive
response through agents who modify the production
timetables and logistical strategies based on data
collected at the time. It also maintains the flexibility of
the supply chain, which is essential for overcoming
disruptions (Xu et al., 2021). Another strategic
component of the MAS framework is logistics and
inventory collaboration. Logistics and inventory agents
play an essential role in product movement from the
point of production to the distribution centres or
customers (Kramarz and Kmiecik, 2022). It was found
that the logistics agent predicted the likely transport
pattern and the inventory agent checked whether the
stocks were adequate.
Resilience and Adaptability
Controlling uncertainty is one of the most significant
objectives of the MAS-ML framework. There are
various sources of uncertainty in supply chain such as
demand uncertainty, disruption in production and
distribution, and any other unpredictable situation
(Kumar and Sharma, 2021). The MAS-ML framework
addresses such uncertainties since the supply chain
agents are allowed to change the response
dynamically. Real-time decision making facilitates the
execution of machine learning models within the
context of MAS framework (Pereira and Frazzon,
2021). Other advantages of the MAS-ML framework
are Cost and Efficiency Improvements. Through the
coordination of production, inventory, and logistics, the
framework aims at reducing expenditures while
enhancing supply chain operations (Pasupuleti et al.,
2024). For instance, the linear programming model
sought to reduce holding costs while also striving to
ensure that stocks, inventories were adequate. This
synergy cost and efficiency not only increase the
profitability of the supply chain but also over supply
chain disruption (Buschiazzo et al., 2020).
Evaluation of the Framework
The primary KPIs suitable for assessing the effectiveness
of the MAS-ML framework are as follows: predictive
accuracy, production line quantity, decreased lead time,
and reduced logistics costs (Islam et al., 2024).
Additionally, linear programming optimised stock levels,
reducing holding costs and improving overall efficiency.
The above metrics enable an understanding of the level
of improvement in supply chain resiliency due to the
implementation of the MAS-ML framework.
This shows that through using the MAS-ML framework,
the supply chain gains high flexibility and
responsiveness. Thus, the applied framework optimizes
supply chain activities in terms of its production
schedule, logistics, and transportation plan to reduce
time and cost of the supply chain (Ikevuje et al., 2020).
METHODOLOGY
This section focuses on the approach used in realising
and evaluating the integration between MAS and ML in
enhancing supply chain resilience. The four elements
are data capture, machine learning model design, MAS
emulation, and performance assessment indicators.
Figure 6: Proposed Methodology Diagram
The American Journal of Engineering and Technology
12
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
Data Collection
MAS-ML can collect a considerable volume of data that
define the supply chain, including inventory, rate of
production, demand, cost of shipping, and time. These
datasets were collected from real-life supply chain
environments or well-articulated simulated scenarios.
In the coding simulation, variables like "Stock levels,"
"Production volumes," and "Shipping costs" were
crucial in modelling the decision of agents in the MAS
framework. It ensures that the agents have historical
and real-time information to make predictive and
adaptive decisions. Data Preprocessing was vital to
guarantee that the data collected was free from
blindness and biases. The missing values were either
imputed appropriately by measures like the mean or
omitted, and records with missing values were also
omitted. Numerical data was scaled correctly, which
made it easier for the machine learning models to
place all the values in a standard range. Variables like
"Product types" and "Transportation modes" were hot
encoded to make it possible for machine learning
algorithms to process them. This preprocessing
corresponds to the steps described in the coding file
part,
where
supply
chain
attributes
were
discretised/mapped to a more suitable format for ML
models.
Feature Selection was the most crucial step when
deciding which variables would be input to the machine
learning models. Some attributes, such as order
quantities, lead time, and shipping cost, were selected
because they are relevant inputs in generating demand
forecasts and are primarily used in optimising
production processes.
Machine Learning Model Development
Model Selection was based on the challenge of
algorithms integrated supply chain and dynamic data.
Random Forest and Gradient Boosting were chosen as
the primary models because these algorithms showed
high rates of predictive accuracy and are suitable for
working with large datasets with many attributes
(Callens et al., 2020). In the coding implementation of
these models, positive feedback was achieved in the
ability to predict demand and enhance real-time
reinforcement decisions for supply chain agents. Due to
their stability and ability to avoid overfitting, they are
ideal for unpredictable demand.
Figure 7: Machine Learning Methodologies Source:
https://www.forbes.com/sites/louiscolumbus/2018/06/11/10-ways-machine-learning-is-revolutionizing-
supply-chain-management/
Training and Testing entailed data division or the
division of a dataset into training and testing sets. The
hyperparameter optimisation process achieved the
best results for the machine learning models (Yang and
Shami, 2020). The Random Forest model was used in
the coding simulation process, and the estimators
were repeatedly searched and improved to get the
optimum values. This tuning of the model guarantees
that it is accurately tuned depending on the level of
demand to make the best decisions for the agents.
Regarding Model Evaluation, the measures involved
were Mean Square Error (MSE) and Mean Absolute
Error (MAE) to determine the accuracy of the models
generated.
Multi-Agent System (MAS) Simulation
The MAS Framework Setup had an agent design that
dealt with different supply chain functions, including
demand forecast, inventory, production schedule, and
logistics. During the coding simulation, agents
responded proactively to the demand requirements and
logistics conditions on the field shared in real-time to
arrive at decisions coherent with the overall goals of the
supply chain.
Regarding simulation facilities, AIFs were developed to
depict how agents in a particular supply chain interact
and self-organise in response to changing conditions.
The actual changes to the production schedule and
logistic operations to match the predicted demand were
The American Journal of Engineering and Technology
13
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
made while testing the MAS framework in the
Simulation Environment. The coding illustrated how
this environment emulated the real-world conditions
under which the agents operated regarding fluctuating
demands and supply chain challenges. A simulated
supply chain environment was created to validate the
feasibility and efficiency of the proposed MAS-ML
framework to handle the supply chain operations.
Every agent integrates decision logic to control the
decision-making process. For instance, the Production
Agent used a simple rule that answered questions
regarding production timing and quantity based on
estimated demand and available inventory. This logic
helped direct the accurate production of the required
products without using many raw materials when they
were not needed. Such heuristics were used in the
coding example to improve the production schedules
and make the supply chain as efficient as possible.
Optimisation Techniques
Linear programming was use to optimise the Stock
Holding Cost. The scipy' linprog' function was used to
solve the problem of minimising the cost of holding
stock. This optimisation made it easy for the agents to
provide adequate stock to clients after meeting the
costs of acquiring these stocks without spending much
money. Logistics Optimisation aims to find the best
transportation routes by employing a predictive
model, reducing shipping expenses. In the coding
simulation, resource allocation was seen in calculating
shipping costs and choosing transport infrastructure to
ensure the optimum use of resources and timely
delivery.
Production planning was employed where demand was
identified to be higher than demand in stores. It
managed to avoid overproduction while at the same
time focusing on areas that required more attention.
The performance measures involve the Mean Square
Error and Mean Absolute Error of the demand forecast,
and the usage of the logistic resources, for flexibility and
cost analysis of the framework.
Implementation: Integration of MAS and ML in the
Case Study
To demonstrate the applicability of the proposed Multi-
Agent System (MAS) and Machine Learning (ML)
framework, this paper presents a manufacturing
company’s supply chain example in which the firm
specialises in consumer electronics. This is because the
demand in this area is very unpredictable due to
changes in the customers’ preference in the
advancement in technologies. Therefore, the MAS-ML
framework is anticipated to improve the flexibility as
well as increase efficiency and capability of responding
to actual changes at a comparatively cheaper cost.
The MAS-ML framework was implemented in four key
areas: Demand forecasting, inventory management,
production planning, logistics optimisation, and selling
prices. The detailed diagram with steps is depicted
below:
Figure 8: MAS-ML Framework
Demand Forecasting Agent
The demand forecasting agent was designed to
calculate the future demand for a product and in this
context, the clients’ sales history, market trends and
behaviour were applied (Zohdi et al., 2022). The agent
used advanced machine learning techniques like
Random Forests and Gradient Boosting Machines to
develop
precise
demand
forecasting.
When
incorporating the use of ML into this agent, the
company was able to move from static to dynamic in
terms of the forecast and this means that the
predictions made by the agent could be refreshed from
time to time
Inventory Management Agent
The inventory management agent adapted the number
of stocks to the demand forecast that the demand
forecasting agent obtained. This agent ensured no
stockouts or overstocking of products and components.
It retained the ML predictions to proactively estimate
the demand shortly and determine the right quantity of
stocks to help reduce holding costs while making the
products easily available. The inventory management
agent also liaised with the logistics agent to ensure that
the frequency of restocking matched the company's
manufacturing and delivery timetable.
The American Journal of Engineering and Technology
14
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
Production Planning Agent
The production planning agent was supposed to
respond to the predicted demand and change the
production schedule accordingly. This agent applied
the
demand
forecasts
to
determine
the
“Production_needed” and “Production_schedule"
variables to maintain total capacity without excess
inventory. Where there were changes in demand
forecast, the production planning agent had to make
recommendations to increase or reduce production
depending on available labour, raw materials, and lead
time to produce the products. This performance was
achieved by linking the demand forecasts derived from
the ML algorithm to real-time company production
data through this agent to increase production
efficiency while reducing waste.
Logistics Optimisation Agent
The logistics agent ensured that the movement of
goods was timely and effective, all in the most
affordable manner. This agent estimated the shipping
time, cost of transportation, and lead time using the
predictive capabilities of the machine learning models.
By including this information, the logistics agent could
redirect deliveries where needed to minimise delays
and expenses related to transportation. They also
collaborated with the inventory and production agents
to efficiently utilise transportation resources,
especially when responding to shifts in demand or
supply shocks.
RESULTS
The simulation of the MAS-ML framework generated
the following results that enhanced the operation of
the supply chain:
Multi-Agent System
Figure 9: Multi-Agent System
The provided code depicts the Multi-Agent System
where the Adjustment Agent freezes the new
production level if it exceeds the predicted number.
The Production Agent reacts to the forecasted demand
and changes its production level similarly. Each agent
begins with given inventory and production rates.
When the forecasted demand exceeds the current
inventory, the agents raise production by the
discrepancy and set the current inventory to zero. In
this case, both agents had very little inventory left, and
the new productions were set to be 322.24 units to
meet the demand. This simulation mimics the ability of
agents to change the production rate in response to
forecasted demand to achieve supply chain
equilibrium.
Figure 10: Mean Absolute Error
The code determines Mean Absolute Error (MAE) for
demand forecasting using the predicted values (y_pred)
and actual test values (y_test). MAE calculates the mean
of the absolute differences between the actual and
predicted values without referencing signs. In this case,
the MAE is about 361.72. Therefore, the predicted
demand is 361.72, meaning the actual demand is
increasing. A lower MAE mean that the model is more
accurate, while a higher value shows that the error rate
of the forecasting model needs to be improved. It helps
evaluate the forecast accuracy of demand by the model.
Figure 11: MAE for Inventory, Production, Logistics
In the simulated model, three agents, Inventory,
Production, and Logistics, adjust to the demand based
on the machine learning prediction. Firstly, each agent
has initial conditions of inventory, production and
logistics levels entered by the user. Upon receiving the
following demand forecast of 331.4 units, the Inventory
and Production Agents reduce the inventory level to
zero and then scale the production according to the
needs. Due to the increase in production to 101 units,
The Logistics Agent adapts the company's logistics
resources. Self-organised behaviour across agents
shows how agents forecast demand, make operational
decisions, and achieve supply chain equilibrium. This
result is aligned with the proposed goal of improving
supply chain robustness through utilising multi-agent
systems (MAS) with machine learning to facilitate timely
adaption in response to market conditions and increase
flexibility and responsivity of the supply chain.
Stock Levels vs Predicted Demand
The American Journal of Engineering and Technology
15
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
Figure 12: Stock Levels vs Predicted Demand
This graph represents the actual stock level of a
particular product against the expected demand rate
when the given test data is applied. The blue line refers
to the stock level, which is generally low, adding below
100 units often. The orange colour is used for the
forecasted demand and is much higher and varies
between 300 and 700 units. The difference between
stocks and forecasted demand also shows that current
inventory is insufficient for the forecasted market
needs, and always having a higher inventory is only
sometimes useful due to its variation, which calls for
flexibility in inventory management.
Scheduled Production Volumes
Figure 13: Scheduled Production Volumes
This bar chart illustrates the amount of production
planned in numerous test instances. Each bar's length
is proportional to the production capacity needed
based on the forecasted demand. The planned
production quantities are different and range from
about 100 to more than 600 units. These oscillations in
the amount of production volumes indicate that
demand predictions can be quite volatile, and it is
essential to accurately align production capabilities to
the real-time market.
Inventory Levels and Production Needed
Figure 14: Inventory Levels and Production Needed
The above chart shows the initial stock level in yellow.
The line in red represents the production that needs to
be undertaken to meet the anticipated demand. Except
that the actual production demand substantially
exceeds the initial stock, more inventory must be
needed to cover the need. Therefore, new production is
required, and there is a deficiency in current resources
and potential production, stressing the significance of
proper supply estimation and flexible manufacturing
methods.
Optimal Stock Levels Based on Cost Optimisation
Figure 15: Optimal Stock Levels Based on Cost
Optimisation
This graph illustrates the changes to the stock levels
after a cost-optimisation algorithm has been run on the
company inventory. It is observed from the chart that
the optimisation leads to a highly stochastic and
extremely high stock level of more than 600 units for an
item single index. In contrast, the rest of the indices
reveal zero or almost zero stock levels. This suggests
that the cost-optimisation model ensures that stock is
maximised for a particular item driven by high expected
demand or a constrained supply chain. This indicates a
need for additional optimisation regarding the stock of
different products in the store.
Inventory Levels vs Predicted Demand
The American Journal of Engineering and Technology
16
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
Figure 16: Inventory Levels vs Predicted Demand
This plot shows how the inventory (orange) and the
forecasted demand (blue) correlate with augmented
test data. The demand forecast also demonstrates high
volatility and reaches a level ranging from 300 to 700
units. As for inventory, it is very low, approximately
100 units for a while. This is to argue that raw material
and other inventory flows have a huge variability in
customer demand. This calls for improved inventory
stock management methods to fit the volatile and
unpredictable market. The agents may have to
customise their production scheduling and inventory
restocking to maintain the equilibrium supply chain.
Production Scheduling Based on Predicted Demand
Figure 17: Production Scheduling Based on Predicted
Demand
This bar chart represents the planned production
quantities in light of the forecasted consumer
requirements. The production volumes vary from 100
to 600 units in different test iterations. The variability
in the production schedule visually demonstrates how
this system adapts to changes in demand by the hour.
This is a sign of the real-time adaptation of production
levels to changes in demand. It shows how crucial it is
to establish workable solutions, such as an adaptive
workable supply chain.
Logistics Resources Utilization (Shipping Costs) Over
Time
Figure 18: Logistics Resources Utilisation (Shipping
Costs) Over Time
This graph measures shipping cost (logistics resources)
in time. The shipping costs are volatile, ranging from 2
to 8 units. This variability implies that the logistics
operations are flexible enough to adapt to production
rate changes and demand. Seasonal changes could be
another reason for the fluctuations, as the flow of
production lines dictates various transportation
requirements. Congestion costs may reflect the volume
of production or sales, while low congestion costs may
indicate little pressure on demand. It becomes critical,
therefore, to manage these costs well to minimise
fluctuations and maximise efficiency in supply chain
costs.
Inventory, Production, and Logistics Interaction
Figure 19: Inventory, Production, and Logistics
Interaction
This graph portrays the inventory levels, planned
production, logistics, and shipping costs. The green line
corresponds to the planned production, which varies
widely from 200 to 600 units, indicating how the system
adjusts to expected customer demand. The orange
colour represents the degree of inventory position less
than 100 units throughout the test data points. The
purple corresponds to the logistics costs and that there
is minimal stress on transportation factors. The
integration of these elements also reveal that the
process of synchronising inventory management,
manufacturing planning and transportation is critical in
improving supply chain performance.
The American Journal of Engineering and Technology
17
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
Table 1: Comparison Metric
Metric
Test Data Index 1 Test Data Index 2
Test Data Index 3 Average Value
Mean Absolute Error (MAE)
361.72
354.12
369.82
361.89
Scheduled Production (Units)
322.24
500.45
410.34
411.68
Logistics Costs (Shipping)
21.00
51.00
101.00
57.67
Inventory Levels (Units)
100
90
80
90.00
DISCUSSIONS
This paper explores integrating multi-agent systems
(MAS) and machine learning to enhance the reliability
and flexibility of supply chain systems. Balancing
supply and demand is critical, and with historical data,
Random Forest Regression produces reasonably good
forecasts. In this regard, the research employs the
Random Forest model to forecast future demand,
including product type, transportation mode, and
inventory level. The Mean Absolute Error (MAE)
obtained in the predictions enable the evaluation of
the model and its response to changes in demand. The
machine learning component is used to provide a
mechanism for an agent to respond to variations in
demand forecasts and alter their actions accordingly.
Machine learning guarantees that the supply chain can
be more sensitive and active in detecting and
minimising supply and demand variability situations in
which there is an excess of actual or lack of inventory.
The MAS framework describes the central part of the
adaptation decision-making process within the supply
chain. Every agent, such as Inventory Agent,
Production Agent, and Logistics Agent, is a partially
autonomous entity that adapts its actions with the
help of forecasted demand and actual data. The
Inventory Agent oversees stock status, the Production
Agent coordinates production based on estimated
demand. The role of the Logistics Agent is to minimise
shipping and transportation expenses. Such agents
interact and coordinate to allow all the processes in the
logistics network, including inventory, production, and
distribution, to run concurrently. The autonomous
adaptive decision-making capability present in this
system is a significant plus point in terms of the
flexibility of the supply chain.
Linear programming is used for cost optimisation
within the framework to minimise the stock holding
cost while ensuring that the production and logistics
capacities are adequate to meet the entire demand.
This approach allows the system to determine the least
costly combination of resource distribution and output
across the supply chain so that no agent produces too
much, too little, or at a rate or volume that can be
considered inefficient.
Maintaining low operational costs to meet the
fluctuating demand is always a challenge for any
supply chain. It dictates that optimisation algorithms
guarantee that production and logistics are being run
under the right parameters, given the estimates on
demand and stocks. The result of costs is minimised, but
the system remains flexible enough to respond to shifts
or changes in demand and operational symmetry in that
the system can respond promptly and effectively to
changes at any level. However, this approach's highly
attractive characteristic is the agents' cooperation.
Since information is shared and decisions are made
collectively within the MAS framework, agents can
address any disruptions or fluctuations in demand as
soon as they arise.
CONCLUSION
This paper evaluates the Multi-Agent System (MAS) and
Machine Learning (ML) framework to support robust
supply chains in the emerging digital economy. MAS and
ML are the best approach to most problems associated
with conventional supply chain management systems.
The decision-making is decentralised through the
introduction of autonomous agents with the machine’s
learning predictive knowledge. Real-time solutions for
demand forecasting, inventory planning, production
planning, and logistics can be provided with the help of
the framework. One of the most crucial advantages of
this adaptability in decision-making is that supply chains
can bounce back from disruptions, thus achieving higher
total efficiency and reduced expenditures. An important
observation that has been made in this study is that
Random Forests can be used to make good predictions
of metrics like demand, lead times and shipping costs in
a supply chain. For example, the demand forecasting
agent can observe a rapid rise or fall in demand which
can be useful to the inventory management agent when
determining stocking levels and avoiding costly mistakes
such as overstocking or stock out.
The decentralised MAS-ML framework minimises agent
central authority decision-making, which is associated
with high rigidity, thereby enhancing flexibility and
responsibility in the supply chain. In most conventional
organisational systems, the top functions and makes
decisions, which take time to filter down to lower levels.
MAS allows every agent to function independently and
in a real-time environment. For instance, whenever the
logistics agent faces a transportation delay, it can
promptly reschedule shipments and alert the
production and inventory agents, thereby minimising
disruptions throughout the supply chain.
The American Journal of Engineering and Technology
18
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
The proposed MAS-ML framework presents a notable
improvement in increasing the resilience of supply
chains, but the following directions for subsequent
studies can strengthen the framework and raise the
possibility of its expansion to other sectors. The
possible development is deepening the framework by
incorporating more sophisticated deep learning
methods. Although algorithms like random forest are
good predictors, there may be more accurate deep
learning structures like LSTM networks or CNNs for
identifying intricate patterns in the supply chain data.
Deep learning models have the flexibility to use large
amounts of data, which have temporal dependencies,
and this has long been seen in global supply chain
systems. Additionally, RL could train an agent to arrive
at further well-grounded, long-term decisions in light
of the state transitions in the supply chain.
A closer look at applying MAS-ML methodology to
industries other than manufacturing and consuming
goods is a promising area for further study. The ideas
of MAS and ML can be extended to several industries,
including healthcare, pharmaceuticals, and energy,
where the supply chains are equally as intricate and
unpredictable as those in manufacturing and require
tremendous flexibility. In addition, future work may
also seek to refine the structure and feasibility of the
framework. This involves enhancing the framework's
scalability to process large volumes of real-time data
from IoT devices, sensors, and blockchain networks.
Such an approach could guarantee that humans' and
machines' strengths are optimally utilised to deliver
the best results.
REFERENCES
Adi, T.N., Bae, H. and Iskandar, Y.A., 2021.
Interterminal truck routing optimisation using
cooperative multiagent deep reinforcement learning.
Processes, 9(10), p.1728.
Ailawadi, S.C. and SINGH, P.R., 2021. Logistics and
Supply Chain Management. PHI Learning Pvt. Ltd.
Antons, O. and Arlinghaus, J.C., 2022. Distributing
decision-making authority in manufacturing
–
review
and roadmap for the factory of the future.
International Journal of Production Research, 60(13),
pp.4342-4360.
Barua, L., Zou, B. and Zhou, Y., 2020. Machine learning
for international freight transportation management:
A comprehensive review. Research in Transportation
Business & Management, 34, p.100453.
Buschiazzo, M., Mula, J. and Campuzano-Bolarin, F.,
2020. Simulation optimisation for the inventory
management of healthcare supplies. International
Journal of Simulation Modelling, 19(2), pp.255-266.
Callens, A., Morichon, D., Abadie, S., Delpey, M. and
Liquet, B., 2020. Using Random forest and Gradient
boosting trees to improve wave forecast at a specific
location. Applied Ocean Research, 104, p.102339.
Chowdhury, N. R. H., Masum, N. a. A., Farazi, M. Z. R., &
Jahan, N. I. (2024). “The impact of predictive analytics
on financial risk management in businesses.” World
Journal of Advanced Research and Reviews, 23(3),
1378
–
1386.
Epiphaniou, G., Bottarelli, M., Al-Khateeb, H., Ersotelos,
N.T., Kanyaru, J. and Nahar, V., 2020. Smart distributed
ledger technologies in Industry 4.0: Challenges and
opportunities in supply chain management. Cyber
Defence in the Age of AI, Smart Societies and
Augmented Humanity, pp.319-345.
Farazi, M. Z. R. (2024). “Designing circular supply chains
with digital technologies for competitive sustainability:
An operation management perspective.” International
Journal of Science and Research Archive, 13(1), 2346
–
2359.
Farazi, M. Z. R. (2024). “Evaluating the impact of AI and
blockchain on credit risk mitigation: A predictive
analytic
approach
using
machine
learning.”
International Journal of Science and Research Archive,
13(1), 575
–
582.
Farazi,
M. Z. R. (2024). “Exploring the Role of Artificial
Intelligence in Managing Emerging Risks: An In-Depth
Study of AI Applications in Financial Institutions’ Risk
Frameworks.” The American Journal of Management
and Economics Innovations, 6(10), 20
–
40.
Folgado, F.J., Calderón, D., González, I. and Calderón,
A.J., 2024. Review of Industry 4.0 from the perspective
of automation and supervision systems: Definitions,
architectures and recent trends. Electronics, 13(4),
p.782.
Frederico, G.F., 2021. Towards a supply chain 4.0 on the
post-COVID-19 pandemic: a conceptual and strategic
discussion for more resilient supply chains. Rajagiri
Management Journal, 15(2), pp.94-104.
Ghadge, A., Er Kara, M., Moradlou, H. and Goswami, M.,
2020. The impact of Industry 4.0 implementation on
supply chains. Journal of Manufacturing Technology
Management, 31(4), pp.669-686.
Gonçalves, J.N., Carvalho, M.S. and Cortez, P., 2020.
Operations research models and methods for safety
stock determination: A review. Operations Research
Perspectives, 7, p.100164.
Gružauskas, V., 2020. Supply chain resilience in the
context of sustainable food industry (Doctoral
dissertation, Kauno technologijos universitetas).
Gui, Y., Zhang, Z., Tang, D., Zhu, H. and Zhang, Y., 2024.
The American Journal of Engineering and Technology
19
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
Collaborative dynamic scheduling in a self-organising
manufacturing
system
using
multi-agent
reinforcement
learning.
Advanced
Engineering
Informatics, 62, p.102646.
Ikevuje, A.H., Anaba, D.C. and Iheanyichukwu, U.T.,
2024. Optimising supply chain operations using IoT
devices and data analytics for improved efficiency.
Magna Scientia Advanced Research and Reviews,
11(2), pp.070-079.
Islam, M.T., Ayon, E.H., Ghosh, B.P., MD, S.C., Shahid,
R., Rahman, S., Bhuiyan, M.S. and Nguyen, T.N., 2024.
Revolutionising Retail: A Hybrid Machine Learning
Approach for Precision Demand Forecasting and
Strategic Decision-Making in Global Commerce.
Journal of Computer Science and Technology Studies,
6(1), pp.33-39.
Kalkanci, B., Chen, K.Y. and Erhun, F., 2011. Contract
complexity and performance under asymmetric
demand information: An experimental evaluation.
Management science, 57(4), pp.689-704.
Katsaliaki, K., Galetsi, P. and Kumar, S., 2022. Supply
chain disruptions and resilience: A major review and
future research agenda. Annals of Operations
Research, pp.1-38.
Kishore, R., Zhang, H. and Ramesh, R., 2006. Enterprise
integration using the agent paradigm: foundations of
multi-agent-based integrative business information
systems. Decision support systems, 42(1), pp.48-78.
Kramarz, M. and Kmiecik, M., 2022. Quality of
Forecasts as the Factor Determining the Coordination
of Logistics Processes by Logistic Operator.
Sustainability, 14(2), p.1013.
Lee, Y.S. and Sikora, R., 2019. Application of adaptive
strategy for supply chain agent. Information Systems
and e-Business Management, 17, pp.117-157.
Li, J., Sikora, R.T. and Shaw, M.J., 2009. Supply Chain
Management: A Multi-Agent System Framework. In
Supply
Chain
Management
and
Knowledge
Management: Integrating Critical Perspectives in
Theory and Practice (pp. 151-169). London: Palgrave
Macmillan UK.
Li, S. and Chen, K.Y., 2020. The commitment
conundrum of inventory sharing. Production and
Operations Management, 29(2), pp.353-370.
Li, S., Chen, K.Y. and Rong, Y., 2020. The behavioral
promise and pitfalls in compensating store managers.
Management Science, 66(10), pp.4899-4919.
Mahraz, M.I., Benabbou, L. and Berrado, A., 2022.
Machine learning in supply chain management: A
systematic literature review. International Journal of
Supply and Operations Management, 9(4), pp.398-
416.
Makkar, S., Devi, G.N.R. and Solanki, V.K., 2020.
Applications of machine learning techniques in supply
chain optimisation. In ICICCT 2019
–
System Reliability,
Quality Control, Safety, Maintenance and Management:
Applications to Electrical, Electronics and Computer
Science and Engineering (pp. 861-869). Springer
Singapore.
Ni, D., Xiao, Z. and Lim, M.K., 2020. A systematic review
of the research trends of machine learning in supply
chain management. International Journal of Machine
Learning and Cybernetics, 11, pp.1463-1482.
Nitsche, B., Brands, J., Treiblmaier, H. and Gebhardt, J.,
2023. The impact of multiagent systems on autonomous
production and supply chain networks: use cases,
barriers and contributions to logistics network
resilience. Supply Chain Management: An International
Journal, 28(5), pp.894-908.
Pasupuleti, V., Thuraka, B., Kodete, C.S. and Malisetty,
S., 2024. Enhancing supply chain agility and
sustainability through machine learning: Optimisation
techniques for logistics and inventory management.
Logistics, 8(3), p.73.
Pereira, M.M. and Frazzon, E.M., 2021. A data-driven
approach to adaptive synchronisation of demand and
supply in omni-channel retail supply chains.
International Journal of Information Management, 57,
p.102165.
Quayson, M., Bai, C., Effah, D. and Ofori, K.S., 2023.
Machine Learning and Supply Chain Management. The
Palgrave Handbook of Supply Chain Management, pp.1-
29.
Raja Santhi, A. and Muthuswamy, P., 2022. Pandemic,
war, natural calamities, and sustainability: Industry 4.0
technologies to overcome traditional and contemporary
supply chain challenges. Logistics, 6(4), p.81.
Rajbala, R., Nain, P.K.S. and Kumar, A., 2023. Intelligent
Agent-Based Supply Chain Management Using Service-
Oriented Architecture. In Contemporary Studies of Risks
in Emerging Technology, Part A (pp. 111-126). Emerald
Publishing Limited.
Sakib, S.N., 2021. The application of the inventory
models to manage and control overstocking in the
production system.
Seyedan, M. and Mafakheri, F., 2020. Predictive big data
analytics for supply chain demand forecasting: methods,
applications, and research opportunities. Journal of Big
Data, 7(1), p.53.
Sharma, R., Kamble, S. S., Gunasekaran, A., Kumar, V.,
and Kumar, A. (2020). A systematic literature review on
machine
learning
applications
for
sustainable
The American Journal of Engineering and Technology
20
https://www.theamericanjournals.com/index.php/tajet
The American Journal of Engineering and Technology
agriculture supply chain performance. Comput. Oper.
Res. 119:104926. doi: 10.1016/j.cor.2020.104926
Stoychev, V., 2023. The potential benefits of
implementing machine learning in supply chain
management (Doctoral dissertation, Technische
Hochschule Ingolstadt).
Wu, D.Y. and Chen, K.Y., 2014. Supply chain contract
design: Impact of bounded rationality and individual
heterogeneity.
Production
and
Operations
Management, 23(2), pp.253-268.
Xu, L., Mak, S. and Brintrup, A., 2021. Will bots take
over the supply chain? Revisiting agent-based supply
chain automation. International Journal of Production
Economics, 241, p.108279.
Xu, Z., Elomri, A., Kerbache, L. and El Omri, A., 2020.
Impacts of COVID-19 on global supply chains: Facts and
perspectives. IEEE engineering management review,
48(3), pp.153-166.
Yang, L. and Shami, A., 2020. On hyperparameter
optimisation of machine learning algorithms: Theory
and practice. Neurocomputing, 415, pp.295-316.
Zohdi, M., Rafiee, M., Kayvanfar, V. and Salamiraad, A.,
2022. Demand forecasting based machine learning
algorithms on customer information: an applied
approach. International Journal of Information
Technology, 14(4), pp.1937-1947.
Zohra Benhamida, F., Kaddouri, O., Ouhrouche, T.,
Benaichouche, M., Casado-Mansilla, D. and López-de-
Ipina, D., 2021. Demand forecasting tool for inventory
control smart systems. Journal of Communications
Software and Systems, 17(2), pp.185-196.
