This paper examines the network-centric approach for enhancing the efficiency of mobile agent control. The primary distinction of this approach is the use of a "wandering center." If communication with the active center is lost, a new center is appointed to continue control, ensuring the system's survivability and improving the overall efficiency of mobile agent management. An experimental study was conducted to evaluate the parameters and effectiveness of network-centric control of mobile agents. The study utilized a developed model of mobile agent interaction based on networkcentric control, along with a simulation model of agent behavior in a dynamic environment. Additionally, a neural network was created that accurately predicts the probability of hitting the target (%) under changing dynamics. The use of this neural network also helped identify input parameters with minimal impact on the outcome.
Today, most existing mobile agents are controlled manually using remote controls that operate on radio channels. However, this manual control poses several challenges, including the need for specialized operator training, limited operational range, and dependency on weather conditions [1,
In model aviation and professional applications such as the civil sector, agriculture, military, law
enforcement, and other fields, mobile agents are in high demand. Selecting the optimal models and
control systems is crucial for effectively monitoring ground-based targets [
Thus, pilotless mobile agents present a more efficient and economical alternative to manned
aircraft for many tasks [
The main challenge in controlling mobile agents lies in the reliance on centralized control systems,
where the control center serves as a vulnerable element; if communication with it is lost, further
control becomes impossible. In contrast, decentralized control faces issues with coordinating and
circulating large volumes of information, leading to slower response times. A promising alternative
is a network-centric approach, which allows control to be transferred to an alternative center when
needed [
An analysis of recent scientific literature reveals that, although modern information technologies have advanced considerably, the network-centric approach to control remains underdeveloped compared to centralized and decentralized methods.
Three main approaches to the control of mobile agents can be identified: centralized, decentralized, and network-centric. In centralized control, a single command center issues control signals to all mobile agents. If this center is disabled or compromised, all mobile agents lose connectivity and cannot be controlled.
Decentralized control, while reducing dependence on a single command point, has the drawback
of coordination challenges. For example, if mobile agents need to rapidly reconfigure, a large amount
of information must be transmitted to them, significantly reducing the overall speed of the control
system [
The network-centric approach is designed to address the limitations of the previous models. This approach integrates all forces and resources into a single information system, enabling control objectives to be met even in dynamic, complex environments that are subject to unpredictable interference. Such interference can be irregular, intermittent, and variable, yet network-centric control is still effective under these conditions [21, 22].
Consider a control network that a subset of mobile agents follows. Among the entire group of mobile agents, approximately 10% are selected as nodes that hold partial control information (Fig. 1). These nodes, equipped with control functions, form a network, and from them, one is designated as the primary control center [23, 24].
Only the node mobile agents (NMA) are coordinated directly by the control center; these nodes, in turn, control the remaining mobile agents (MA). This setup minimizes communication interference since individual control of each node could overwhelm communication channels. For example, if the command is to change direction quickly, adjusting the coordinates for each mobile agent individually would require significant time. Furthermore, if there is only one primary control agent and it suddenly stops responding, is damaged, or is compromised, all mobile agents would lose direction, become vulnerable to interception, or cease to operate [25–27].
A network-centric system consists of a primary control mobile agent (MMA) that continuously sends commands to other mobile agents, setting parameters such as movement direction [28, 29]. Alongside commands, it also relays information from the control center to the node mobile agents. If the primary control agent fails for any reason, copies of all information, including control data for all mobile agents, are preserved within the node mobile agents. In this way, an adversary would need to disable every node agent to compromise the entire network – a challenging task since the enemy cannot readily identify which agents’ function as nodes. This network organization enhances the resilience of the control system.
A key feature of a network-centric system is its “wandering center.” If the primary control mobile agent is identified and disabled by an adversary, the stability of the system remains unaffected, as control can quickly be reassigned. A new control agent is selected from the remaining active node agents, allowing for seamless continuity. This flexibility means that control can be handed over to another node at any moment, ensuring sustained operation. In this network-centric mobile agent control system, two types of information are managed: control: this includes global coordinates transmitted from the control center to the main control agent. command: this includes the coordinates transmitted by the control agent to node agents, which then relay local coordinates to other agents. Thus, mobile agents navigate based on local rather than global coordinates, as provided by the node agents.
In summary, a key strength of the network-centric approach is the use of a roving control center. If connections to the current center are lost, a new center is assigned to maintain control. This design ensures system survivability, thereby greatly enhancing the overall effectiveness of mobile agent control.
Based on the network-centric control algorithm model for mobile agents in dynamic environments and the simulation model of interacting mobile agents, it is essential to conduct experimental studies to assess parameters and evaluate the effectiveness of network-centric control.
To validate the scientific and methodological framework developed, we performed mathematical modeling to analyze the effectiveness of network-centric control for mobile agents in a dynamic environment.
The software tools used in this study include: Simulations were performed on a personal computer with an Intel® Celeron® processor (3.2 GHz) and 16 GB of RAM running on Microsoft Windows 10. Microsoft SQL Server 2022 was utilized as the database management system. Microsoft Visual Studio 2022 Community Edition served as the development environment. Development was conducted in C# using an object-oriented approach. Experiments in the simulation environment included setting initial parameters, performing intermediate calculations, and visualizing results with Windows Presentation Foundation.
The minimum required percentage of node mobile agents % (nodePercentage) needed to reach the target was calculated as follows. A target is considered "hit" if both of the following conditions are met: The percentage of mobile agents that reached the target (targetHitPercentage) must be at least equal to the required percentage of mobile agents needed to hit the target (targetHitPercentage). Among the mobile agents that reached the target, there must be at least one node mobile agent (targetReachedCount > 0).
To evaluate these conditions, 100 experimental trials were conducted, using the following parameters: Number of mobile agents (numberOfAgents): 100. Minimum percentage of mobile agents required to hit the target (targetHitPercentage): 5%. Percentage of node mobile agents (nodePercentage): ranging from 1% to 20% of the total number of mobile agents. Initial (startLatitude, startLongitude) and target (targetLatitude, targetLongitude) geographic coordinates for node mobile agents. Percentage of mobile agents that lost communication (lossPercentage), randomly generated between 50% and 90%. For each agent that lost communication, the geographic coordinates were set to NULL.
The experimental results were used to plot graphs illustrating the relationship between the percentage of node mobile agents % (nodePercentage) and the probability of hitting the target (targetHitPercentage) in each experiment (Fig. 2).
Based on the experimental results, the average effectiveness of target hits with varying numbers of node mobile agents was analyzed (see Table 1 and Fig. 3). 16% or more node mobile agents: 100% effectiveness. 12–15% node mobile agents: 95% effectiveness. 10–11% node mobile agents: 90% effectiveness. 7–9% node mobile agents: 80% effectiveness.
0–6% node mobile agents: less than 80% effectiveness.
From the above, it can be concluded that to achieve effective target damage (80% and above), the minimum required number of node mobile agents (nodePercentage) is between 7% and 9%. Therefore, using more than 9% of node mobile agents out of the total number of mobile agents is not advisable.
Evaluating the effectiveness of network-centric control for mobile agents in a dynamic environment, using neural networks, plays a crucial role in determining the suitability of this approach for various tasks. The main stages of this method can be described as follows: 1. Training Data Preparation: collect and prepare a large dataset representing different scenarios and dynamic environmental conditions. 2. Creating a Dataset for Network Training: split the collected data into training and test sets to evaluate the network's adaptability to new situations. 3. Designing and Configuring the Neural Network: develop a neural network architecture suited for controlling mobile agents in dynamic environments.
set the parameters and functions to optimize training. 4. Neural Network Training: use the training set to teach the network to recognize and solve agent control tasks in various environments. 5. Validation and Performance Evaluation: use the test set to assess the network’s accuracy and efficiency under real-world conditions.
analyze the results in the context of specific tasks and environmental constraints. 6. Improvement and Adaptation: use the results to refine the network architecture and its parameters.
adapt the network to changes in environmental dynamics and control tasks.
The process of training and creating the neural network was carried out using the Deductor Studio Academic analytical platform, which provides highly efficient tools for data analysis and processing. A critical step in this process was utilizing a specially created and carefully prepared dataset to train the network.
The dataset included information from solving 1,000 different problems, which played a key role in enhancing the adaptability and accuracy of the neural network. This approach to data preparation contributed significantly to the network’s ability to solve diverse tasks and ensured its high efficiency across various contexts.
For the parameters listed in Table 2, a mechanism for estimating the probability of target damage (%) based on the neural network can be applied. targetHitPercentage targetHitCount nodePercentage nodeCount startLatitude startLongitude targetLatitude targetLongitude lossPercentage lossCount targetHitProbability
The number of mobile agents that will be enough to hit the target, % The number of mobile agents that will be enough to hit the target, units.
Number of nodal mobile agents, % Number of nodal mobile agents, units Initial geographic coordinates of the mobile agent (latitude), degrees.
Initial geographic coordinates of the mobile agent (longitude), degrees.
Final geographical coordinates of the mobile agent (latitude), degrees.
Final geographic coordinates of the mobile agent (longitude), degrees.
Number of mobile agents with which communication was lost, % The number of mobile agents with which communication was lost, units.
Probability of hitting the target, %
The parameters from Table 2 were provided as input data for building a neural network (Fig. 4).
Based on the results of the correlation analysis of the raw data, it was determined that not all fields should be used for neural network training. Only the fields listed in Table 3 are relevant, as the remaining fields have an insignificant impact on the resulting value (Fig. 5).
The obtained dataset was used to train a multilayer perceptron neural network. The following network structure was established: five layers of neurons, with the first (input) layer containing 7 neurons, three hidden layers containing 7, 3, and 5 neurons respectively, and the fifth (output) layer containing 1 neuron. The activation function used is sigmoid. The architecture of the network – defined by the number of layers, their sizes, and the activation function – determines its ability to solve specific tasks. This framework demonstrates how data moves through the network, from the input to the output layer, with each layer performing calculations using weights and the activation function. The backpropagation method was employed for training.
The schematic representation of the neural network structure is shown in Figure 6, where the line colors indicate the values of the weighting factors. The output of the neural network is the probability of a target being hit (%). By applying the parameters of a different task to the input, the network generates a predictive output for the probability of hitting the target (%).
During the training of the neural network, the probability values of hitting the target (%) were obtained. On average, the deviation across 1,000 experiments was 3.8%. This suggests that the neural network trained using this methodology yields results that closely match the efficiency of the mobile agent control system (Table 4).
The structure of the neural network selected for training resulted in the smallest relative calculation error, while errors for other network configurations exceeded 5%. Therefore, the constructed neural network can be effectively used to solve similar problems, as it accurately reflects the results and allows for the estimation of the probability of hitting the target (%) in dynamically changing environments. Additionally, the use of the neural network helped identify input parameters with minimal impact on the outcome.
This study reviews the network-centric approach as a method for enhancing the efficiency of mobile agent control systems. The key distinction of this approach, compared to others, is the use of a "wandering center." If communication with the current center is lost, a new center is appointed, ensuring the continuity of the control process. This structure guarantees the system's survivability and, as demonstrated, offers significant potential for improvement, ultimately increasing the overall efficiency of mobile agent control.
The research highlights the effectiveness of network-centric control when compared to centralized control (in which no nodal mobile agents are used). It was found that as the number of nodal mobile agents increases, the effectiveness of hitting the target improves. Notably, effective performance (80% or higher) is achieved with just 9% of the total number of mobile agents functioning as nodal agents, suggesting that exceeding this threshold would not provide additional benefits.
Finally, the results from the mobile agent control system were compared with the output from the neural network training, which estimates the probability of hitting the target with a relative deviation of 3.8%. This confirms that the neural network, trained using the proposed methodology, produces results that closely align with the efficiency of the mobile agent control system.
[1] Moutinho A. Modeling and nonlinear control for airship autonomous flight. Instituto Superior
Tecnico, Technical University of Lisbon. 2017. P. 55–65. [2] Fukau T., Kanzawa T., Osuka K. Inverse optimal tracking control of an aerial blimp robot.
Proceedings of the 5th International Workshop. Robot Motion and Control. 2015. P. 193–198. [3] Ko J., Klein D., Fox D., Haehne D. Gaussian Processes and Reinforcement Learning for Identification and Control of an Autonomous Blimp Proc. of the IEEE Int. Conf. Robotics and Automation. 2017. P. 742–747. [21] Canfora G. Formal Tool for Identifying Mobile Malicious Behavior. IEEE Transactions on
Software Engineering. 2019. Vol. 45. P. 1230–1252. [22] How J., Frazolli E., Chowdhary G. Linear Flight Control Techniques. Handbook of Unmanned
Aerial Vehicles. Dordrecht; Heidelberg; New York; London: Springer. 2018. P. 529–576. [23] Lotfi A. Zadeh. 1994. Fuzzy logic, neural networks, and soft computing. Commun. ACM 37, 3 (March 1994), 77–84. DOI:https://doi.org/10.1145/175247.175255 [24] Varley TF, Sporns O, Schaffelhofer S, Scherberger H, Dann B. Information-processing dynamics in neural networks of macaque cerebral cortex reflect cognitive state and behavior. Proc Natl Acad Sci US A. 2023 Jan 10; 120(2): e2207677120. DOI:https://doi.org/10.1073/pnas.2207677120 [25] Paulo Vitor de Campos Souza. Fuzzy neural networks and neuro-fuzzy networks: A review of the main techniques and applications used in the literature. Applied Soft Computing. Vol. 92. 2020. ISSN 1568-4946. DOI:https://doi.org/10.1016/j.asoc.2020.106275 [26] AM Zador. A critique of pure learning and what artificial neural networks can learn from animal brains. Nat Commun 10, 3770 (2019).https://doi.org/10.1038/s41467-019-11786-6 [27] V. Mukhin, V. Zavgorodnii, V. Liskin, S. Syrota, V. Koval, L. Honchar. Classification of Information Objects with Fuzzy Parameters in E-Learning Systems. IEEE 12th International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS). Dortmund. Germany. 2023. P. 1189-1193. DOI: https://10.1109/IDAACS58523.2023.10348768 [28] Ghafor K. Multifunctional Models, Including an Artificial Neural Network, to Predict the Compressive Strength of Self-Compacting Concrete. Applied Sciences. 2022; 12(16):8161.
DOI:https://doi.org/10.3390/app12168161 [29] Vadym Mukhin; Yaroslav Kornaga; Yurii Bazaka; Ievgen Krylov; Andrii Barabash; Alla Yakovleva; Oleg Mukhin "The Testing Mechanism for Software and Services Based on Mike Cohn's Testing Pyramid Modification," 2021 11th IEEE International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS), Cracow, Poland, 2021, pp. 589-595, doi: 10.1109/IDAACS53288.2021.9660999.