THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE11
46
https://www.theamericanjournals.com/index.php/tajet
PUBLISHED DATE: - 23-11-2024
https://doi.org/10.37547/tajet/Volume06Issue11-06
PAGE NO.: - 46-53
METHODS OF TRAINING AND ADAPTATION
OF AI AGENTS IN COMPLEX PROCESS
CONTROL SYSTEMS
Oleksandr Khodorkovskyi
CEO, Quantum Core, Kyiv, Ukraine
INTRODUCTION
Modern control systems are becoming increasingly
complex and multifaceted, necessitating continual
adaptation to changing conditions and the ability
to operate effectively under high uncertainty. In
such environments, traditional automation
approaches prove insufficiently flexible, as they
struggle to respond appropriately to unpredictable
changes in processes and external factors. In
recent years, artificial intelligence (AI) and
machine learning have gained considerable
prominence in addressing these challenges,
offering adaptive solutions that can learn
autonomously, analyze data, and make decisions
based on the current situation. Among various AI
approaches, particular emphasis is placed on
methods for training and adapting AI agents, which
provide the flexibility and autonomy demanded by
complex control systems.
The relevance of this study is underscored by the
growing need for adaptive and intelligent systems
RESEARCH ARTICLE
Open Access
Abstract
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE11
47
https://www.theamericanjournals.com/index.php/tajet
capable of efficient operation in industries such as
manufacturing, energy, transportation, and other
sectors that require rapid responses to change.
Applying machine learning techniques, such as
neuroevolution and reinforcement learning,
allows AI agents to learn unique process
characteristics and adapt to new conditions in real
time, significantly enhancing their effectiveness.
Despite notable progress in AI, integrating
trainable AI agents into complex control systems
remains a challenging task. The need for
continuous learning and adaptation of agents to
dynamic conditions, along with the ability to act
based on large-scale data analysis, demands the
development and testing of new methods. Current
research
indicates
that
approaches
like
reinforcement learning and deep neural networks
enable the creation of systems that not only adapt
to changes but also continuously improve their
performance during operation. Consequently,
exploring methods for training and adapting AI
agents is an essential step in developing fully
autonomous control systems capable of efficient
functioning in complex environments.
The objective of this study is to investigate and
analyze contemporary methods for training and
adapting AI agents in control systems managing
complex processes, as well as to assess their
effectiveness in ensuring reliability, resilience, and
high performance.
METHODS
This study employed comparative analysis,
systematization, synthesis, and examination of
practical examples that illustrate the use of
training and subsequent adaptation methods for AI
agents in control systems for complex processes.
Thon C. et al. [1] point out that artificial intelligence
(AI) plays a crucial role in various fields,
particularly in process engineering, where it
optimizes complex industrial processes to enhance
efficiency and reduce costs. Azeem M. et al. [9] note
that in the manufacturing sector, the use of big data
addresses significant challenges and improves
production efficiency. The practical aspect of this
application was considered in this study through
the example of Volkswagen [10]. Additionally, the
practical side of AI applications was explored
based on Tesla's experience [12]. In aviation, AI
contributes to improving aircraft fuel efficiency
through advanced technologies, supporting the
industry's sustainable development [11].
Soori M., Arezoo B., and Dastres R. [7] observe that
in robotics, the integration of AI, machine learning,
and deep learning is transforming robot
capabilities, making them more complex and
intelligent. Galván E. and Mooney P. [5] believe that
these advancements are complemented by
research in the neuroevolution of deep neural
networks, which enhances training algorithms and
addresses future challenges in AI systems.
Heuillet A., Couthouis F., and Díaz-Rodríguez N. [6]
discuss the necessity for transparent and
interpretable
AI
models,
especially
in
reinforcement learning, where understanding AI
decision-making processes builds trust and
promotes broader adoption of the technology.
Practical applications are developing automated
machine learning approaches for real-time fault
detection and diagnosis, which improve system
reliability and maintenance, as noted by Leite D. et
al. [8].
Beyond technical achievements, social aspects of
AI are also subjects of research. De Togni G. et al.
[3] examine emotional and relational aspects of AI
systems to make them more "intelligent" and
"caring." Moradbakhti L., Schreibelmayr S., and
Mara M. [4] explore gender dynamics in AI through
the lens of voice assistants, investigating how these
technologies meet users' basic psychological
needs. The evolution and future of conversational
agents are analyzed with a proposed research
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE11
48
https://www.theamericanjournals.com/index.php/tajet
agenda for exploring new frontiers in AI
development, according to Schöbel S. et al. [2].
In healthcare, AI shows significant potential.
Platforms such as IBM Watson Health develop AI-
based solutions for personalized cancer treatment,
demonstrating AI's impact on providing
individualized medical care [13].
RESULTS AND DISCUSSION
Modern control systems for complex processes
require adaptive and autonomous solutions
capable of responding to changing conditions in
real time. Artificial intelligence (AI), particularly
learning-enabled agents, has become a key
component of such systems. AI agents can learn
specific process characteristics, predict behavioral
changes, and optimize outcomes based on
accumulated data. The scenarios presented below
in Table 1 demonstrate how these systems can
transform traditional methods of operation.
Table 1. Scenarios for the use of AI agents [4]
Scenario
Description
Roles and Tasks of AI
Agents
Advantages
Automation of
Credit
Underwriting
Credit risk assessment
requires thorough data
analysis of the borrower, type
of credit, and other factors,
which is time-consuming and
involves collaboration among
specialists. AI agents can
automate and expedite this
process by performing
functional roles.
- Intermediary Agent:
connects the borrower with
the financial institution.
- Data Processing Agent:
verifies documentation,
calculates financial metrics.
- Sub-agents: verify results
and analyze errors.
- Reduces
evaluation time by
20–60%.
- Increases speed
and quality of
analysis.
Updating Legacy
Software Systems
Updating old software
requires code analysis,
documentation, and updates,
demanding significant
resources. Agent systems can
distribute tasks among AI
agents specializing in
different aspects of the
process, optimizing and
accelerating modernization.
- Code Analysis Agent:
performs structural analysis.
- Quality Assurance Agent:
ensures compliance with
requirements.
- Development and Testing
Agent: implements and tests
updates.
- Saves resources
in development
and maintenance.
- Enhances
productivity and
operational
efficiency.
Creating
Marketing
Campaigns
Conducting online campaigns
requires coordination and
consideration of various
platforms' specifics. Agent
- Audience Analysis Agent:
collects and analyzes
information by segments.
- Content Creation Agent:
- Reduces time
costs.
- Improves
campaign
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE11
49
https://www.theamericanjournals.com/index.php/tajet
systems assist marketing
teams by auditing audiences,
creating, and testing content,
thus speeding up the process
and improving relevance.
adapts materials for different
segments.
relevance and
effectiveness.
For the effective implementation of agent-based
systems, organizations should consider several
key factors. First, it is essential to structure
knowledge about business processes to provide a
foundation for agent training. The technological
infrastructure must also be adapted to support
agent-based systems, ensuring their integration
into existing workflows. Finally, ongoing
supervision by specialists is crucial to assess the
accuracy and efficiency of agent operations,
creating conditions for their continued learning
[4].
Neuroevolutionary
algorithms,
combining
principles of evolutionary strategies and neural
networks, represent unique methods that allow
artificial intelligence systems to adapt to complex
tasks. This approach, inspired by natural
processes, employs artificial neural networks
optimized through mechanisms resembling
natural selection. The neuroevolutionary learning
process is based on principles similar to evolution:
a population of networks is generated randomly,
with each network evaluated for task performance.
Fitness scores determine which specimens will
reproduce, applying genetic operators such as
mutation and crossover to produce more adaptive
offspring. Gradually, with each generation,
algorithms enhance their performance. Table 2
below describes the main advantages and
disadvantages inherent in the neuroevolutionary
approach for training and adapting AI agents in the
management of complex process systems.
Table 2. The main advantages and disadvantages are inherent in the
neuroevolutionary approach in the training and adaptation of AI agents [5]
Advantages
Disadvantages
Ability to find global optima: Neuroevolutionary
algorithms avoid local minima, which is
particularly important for complex systems with
nonlinear dynamics.
High computational cost: The numerous iterations
and need for evolutionary modeling result in
significant computational expenses.
Adaptation to changing conditions: This approach
enables
agents
to
adapt
effectively
to
environmental changes or control parameters.
Parameter
tuning
complexity:
Selecting
hyperparameters (e.g., mutation rate, population
size) requires expert knowledge and can
significantly impact outcomes.
Resilience to unforeseen factors: Evolutionary
methods can find effective strategies even in the
Need for large training datasets: Successful
evolution
requires
numerous
simulations,
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE11
50
https://www.theamericanjournals.com/index.php/tajet
presence of noise and random factors affecting
learning.
increasing
training
time
and
resource
consumption.
Ability to generate diverse solutions: The
evolutionary process creates multiple solutions,
enhancing the chances of finding effective
strategies for complex tasks.
Difficulty in analyzing and interpreting solutions:
Neuroevolutionary outcomes can be challenging
to interpret, complicating validation and safety
assessments.
Reduced dependence on prior knowledge:
Neuroevolutionary algorithms do not require a
detailed task description and can learn with
limited information about the system.
Risk of undesired behavior: Evolutionary
algorithms may produce agent behaviors that do
not align with system goals or may even be
detrimental.
High flexibility and scalability: Neuroevolution
can be used to configure agents in systems of
varying complexity, in both real and simulated
environments.
Extended training duration: The need for
evolutionary iterations can make agent training
time-consuming,
reducing
the
method's
immediacy.
Thus, the data optimizes logistics management,
risk forecasting, data processing, and the creation
of adaptive models that learn despite a limited
amount of available information [5].
Reinforcement learning (RL) is a method in which
an agent, through interaction with the
environment, learns from the outcomes of its
actions. This method, based on a "trial and error"
model, allows AI to adapt and achieve objectives by
accumulating positive and negative rewards. Q-
learning, a popular approach that uses a Q-value
table, helps an agent discover optimal action
strategies by updating values based on observed
rewards. Policy gradients offer an alternative
approach, where the agent learns directly through
functions mapping states to actions, enabling it to
maximize expected rewards. These methods, when
combined with neural networks, make AI more
flexible and capable of making strategic decisions
in complex situations [6].
Table 3 below presents the advantages and
disadvantages of the reinforcement learning (RL)
method for enhancing the intelligence of AI agents.
Table 3. Advantages and disadvantages of the reinforcement learning (RL)
method for the development of intelligence of AI agents [7].
Advantages
Disadvantages
Adaptive learning through trial and error: RL
allows agents to learn from experience, adjusting
behavior
based
on
feedback
from
the
environment.
High computational complexity: Achieving
optimal strategies requires many iterations,
increasing computational costs.
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE11
51
https://www.theamericanjournals.com/index.php/tajet
Ability to find optimal strategies: RL can
optimize agent actions to achieve maximum
rewards, which is suitable for complex tasks.
Unstable
learning:
RL
often
encounters
inconsistent performance due to initial conditions
and environmental randomness.
Applicability to dynamic environments: RL is
well-suited for environments with changing
conditions, as the agent learns to adapt based on
current data.
Risk of overfitting to specific scenarios: The agent
may "memorize" actions for particular situations
and become ineffective in new conditions.
Minimal knowledge requirements about the
environment: RL enables learning even without
knowledge of the environment's mathematical
model, relying on interaction and receiving
rewards.
Extended
training
period:
In
real-world
applications, reinforcement learning can take a
long time due to the need for extensive interactions
with the environment.
Development of autonomy and independent
decision-making: Through learning, the agent
becomes capable of making decisions and
achieving goals without constant supervision.
Risk of negative behavior: If the reward function
is incorrectly configured, the agent may develop
undesirable or even harmful strategies that do not
align with system objectives.
Ability to learn in uncertain conditions: RL can
learn and make decisions even with noise and
incomplete information about the environment.
Knowledge transfer issues: Acquired strategies
may be of limited use in other tasks, requiring a
new training cycle for each new environment.
Next, practical examples of the use of training and
adaptation methods for AI agents will be
examined. For instance, Siemens applies deep
learning and neural networks in automation
systems for failure prediction and equipment fault
diagnosis. Its analytical platform, MindSphere,
uses neural network models to analyze real-time
data, optimizing production processes and
identifying potential issues before they arise. This
approach has enabled the company to reduce
emergency shutdowns, minimizing production
losses and enhancing equipment reliability [8].
General Electric has also implemented neural
network models in its Predix system, which
focuses on analyzing data from the industrial
Internet of Things (IIoT). This platform collects
and processes data from numerous sensors and
devices, allowing it to predict machine condition
changes, minimize downtime risks, and extend
equipment life. In critical sectors such as energy
and aviation, Predix, powered by neural network
methods,
has
significantly
improved
responsiveness and reliability in managing
complex processes [9].
Volkswagen uses evolutionary algorithms and
genetic programming methods to optimize supply
chains and production processes. The Digital
Production Platform (DPP) project, developed in
collaboration with AWS, utilizes algorithms to
predict production needs, manage logistics, and
minimize operational risks. The flexibility of
evolutionary methods allows the company to
quickly adapt to demand changes and avoid
production disruptions [10].
At Airbus, evolutionary algorithms are used to
optimize aircraft design, reducing weight and
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE11
52
https://www.theamericanjournals.com/index.php/tajet
improving fuel efficiency. These algorithms enable
the discovery of unique design solutions that
reduce material costs and improve operational
performance. This approach helps lower
operational costs and makes aircraft more
environmentally friendly [11].
Hybrid methods are used in companies such as
IBM and Tesla. For example, Tesla combines deep
learning and RL to develop autonomous vehicle
control systems. This approach allows systems to
analyze complex road situations in real time,
optimizing routes, avoiding obstacles, and
adapting to changing conditions [12].
In healthcare projects, IBM Watson uses hybrid
learning, combining neural networks and RL to
support medical decision-making. For instance, in
oncology clinics, Watson analyzes patient data and
recommends optimal treatment options based on
a comprehensive analysis of medical data and
scientific publications. This approach ensures
system flexibility, helping physicians make more
informed decisions and enhancing the quality of
medical care [13].
Thus, the integration of AI agents with training and
adaptation methods in complex processes, as
demonstrated by leading corporations, confirms
the importance of these technologies in improving
efficiency, reducing costs, and enhancing
operational performance.
CONCLUSION
The study confirms that integrating AI agents into
control
systems
for
complex
processes
significantly
enhances
their
adaptability,
resilience, and efficiency. The examined machine
learning methods, including reinforcement
learning and neuroevolutionary algorithms, have
demonstrated their capacity to provide high
predictive accuracy and rapid responsiveness to
environmental changes. The use of deep neural
networks and adaptive algorithms allows AI
agents not only to analyze current data but also to
improve behavior models based on experience,
which is critically important for complex
multitasking systems.
The research results showed that adaptive AI
agents can successfully address optimization tasks
and minimize disruptions, reducing costs and
improving operational performance. These
technologies are especially relevant to industrial
and infrastructure sectors, where system stability
and reliability are paramount. Agents with
continuous learning capabilities have proven
effective in automated management tasks, opening
up vast prospects for their use in resource
management, logistics, and forecasting.
Thus, the study underscores the significance of
using AI agents in modern control systems,
highlighting the promise of further research in this
field. The application of adaptive learning methods
for creating autonomous AI agents not only
improves management efficiency but also paves
the way for developing new operational models
that are resilient to external influences and capable
of self-improvement.
REFERENCES
1.
Thon C. et al. Artificial intelligence in process
engineering //Advanced Intelligent Systems.
–
2021.
–
Vol. 3.
–
No. 6.
–
p. 2000261.
2.
Schöbel S. et al. Charting the evolution and
future of conversational agents: A research
agenda along five waves and new frontiers
//Information Systems Frontiers.
–
2024.
–
vol.
26.
–
No. 2.
–
pp. 729-754.
3.
De Togni G. et al. What makes AI
‘intelligent’and ‘caring’? Exploring
affect and
relationality across three sites of intelligence
and care //Social Science & Medicine.
–
2021.
–
vol. 277.
–
p. 113874.
4.
Moradbakhti L., Schreibelmayr S., Mara M. Do
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE11
53
https://www.theamericanjournals.com/index.php/tajet
men not need “feminist” artificial intelligence?
Agentic and gendered voice assistants in the
light of basic psychological needs //Frontiers
in psychology.
–
2022.
–
vol. 13.
–
p. 855091.
5.
Galván E., Mooney P. Neuroevolution in deep
neural networks: Current trends and future
challenges //IEEE Transactions on Artificial
Intelligence.
–
2021.
–
Vol. 2.
–
No. 6.
–
pp. 476-
493.
6.
Heuillet A., Couthouis F., Díaz-Rodríguez N.
Explainability in deep reinforcement learning
//Knowledge-Based Systems.
–
2021.
–
Vol.
214.
–
p. 106685.
7.
Soori M., Arezoo B., Dastres R. Artificial
intelligence, machine learning and deep
learning in advanced robotics, a review
//Cognitive Robotics.
–
2023.
–
Vol. 3.
–
pp. 54-
70.
8.
Leite D. et al. An automated machine learning
approach for real-time fault detection and
diagnosis //Sensors.
–
2022.
–
vol. 22.
–
no. 16.
–
p. 6138.
9.
Azeem M. et al. Big data applications to take up
major challenges across manufacturing
industries: A brief review //Materials Today:
Proceedings.
–
2022.
–
Vol. 49.
–
pp. 339-348.
10.
The Volkswagen Group on AWS. [Electronic
resource]
Access
mode:
https://aws.amazon.com/ru/solutions/case-
studies/innovators/volkswagen-group
/
(accessed 10/23/2024).
11.
The Top 6 Technologies for Improving Aircraft
Fuel Efficiency. [Electronic resource] Access
mode:
https://www.prescouter.com/2018/01/techn
ologies-improving-aircraft-fuel-efficiency
/
(accessed 10/23/2024).
12.
How Artificial Intelligence is Revolutionizing
the World of Tesla. [Electronic resource]
Access
mode:
https://aiforsocialgood.ca/blog/how-
artificial-intelligence-is-revolutionizing-the-
world-of-tesla (accessed 10/23/2024).
13.
AI-Powered Platform from IBM Watson Health
to Personalize Cancer Treatment. [Electronic
resource]
Access
mode:
https://medium.com/@AIadvice/ai-powered-
platform-from-ibm-watson-health-to-
personalize-cancer-treatment-57116b973398
(accessed 10/23/2024).
