Optimal 5G Network Sub-Slicing Orchestration in a Fully Virtualised Smart Company Using Machine Learning

Abstract

This paper introduces Optimal 5G Network Sub-Slicing Orchestration (ONSSO), a novel machine learning framework for dynamic and autonomous 5G network slice orchestration. The framework leverages the LazyPredict module to automatically select optimal supervised learning algorithms based on real-time network conditions and historical data. We propose Enhanced Sub-Slice (eSS), a machine learning pipeline that enables granular resource allocation through network sub-slicing, reducing service denial risks and enhancing user experience. This leads to the introduction of Company Network as a Service (CNaaS), a new enterprise service model for mobile network operators (MNOs). The framework was evaluated using Google Colab for machine learning implementation and MATLAB/Simulink for dynamic testing. The results demonstrate that ONSSO improves MNO collaboration through real-time resource information sharing, reducing orchestration delays and advancing adaptive 5G network management solutions.

1. Introduction

Machine learning is pivotal in enhancing 5G network slicing and orchestration, offering innovative solutions to complex challenges, and unlocking new opportunities for network optimisation [1]. Fifth-generation (5G) cellular networks aim to transform mobile broadband communication by delivering unprecedented data rates, quality of service, connectivity, and flexibility in mobile networks [2]. Network slicing introduced a paradigm shift in mobile broadband communications, creating multiple virtual verticals from end to end while sharing the available physical infrastructure [3]. With network sub-slicing, these slices are further divided into granular sub-slices dedicated to a specific service, tenant, or application instance. This available network resource is managed and allocated as required to each service or user demanding a network slice, helping to reduce capital expenditure, a reputed challenge in current 5G network slice orchestration. Some of the standard network slices include ultra-reliable low-latency communications (URLLC), enhanced mobile broadband (eMBB), and massive machine-type communications (mMTC) [4]. As the demand for these slices continues to grow, existing methods for resource allocation will struggle to meet performance expectations.

This paper addresses these challenges by proposing a novel ONSSO framework to autonomously orchestrate 5G network slices optimally over a mobile network. This is conducted in real time, thereby improving the quality of service (QoS) and quality of experience (QoE). A key component of this framework is the Enhanced Sub-Slice (eSS) model, a machine-learning-based model that facilitates more granular resource allocation within the overall network slicing architecture. Figure 1 shows the isolated end-to-end sub-sliced networks and how they can make up a single slice-sharing network resource pool.

The proposed ONSSO utilises LazyPredict, AutoRL, and Ray Tune to dynamically select the most effective machine learning algorithm for classification, prediction, and finetuning, respectively. Together, these pipelines culminate into a Company Network as a Service (CNaaS) model, designed to enhance operational efficiency for mobile network operators (MNOs). During the evaluation, the performance of the proposed framework demonstrated significant improvement in key performance metrics such as latency, bandwidth, throughput, jitter, packet loss, and resource utilisation compared to conventional methods.

2. Literature Review

The optimal orchestration of 5G network slices is being studied by researchers exploring different techniques and approaches. We present a review of the state of the art in relevant domains, including 5G network slicing and machine learning in 5G network slice optimisation.

• 5G Network Slicing

The concept of slicing a network end to end, forming multiple logical virtual network slices that share a physical infrastructure, has been a key enabler for 5G networks. The authors in [4] worked on a comprehensive survey of network slicing in 5G, discussing 5G-network-enabling technologies, their motivation, and the challenges involved. The study identified software-defined networking (SDN), network function virtualisation (NFV), cloud computing, and virtualisation as 5G-network-enabling technologies. Zhang et al. [3] highlighted the importance of efficiently orchestrating 5G network slices by studying resource allocation strategies and mobility management for 5G network slices.

Software-defined networking (SDN) and network function virtualisation (NFV) were leveraged to propose a framework [5] in creating and managing 5G network slices efficiently. The need for automation to enable optimal network slice orchestration was emphasised. In [6], the authors studied the concept of softwarisation in 5G and 5G network slicing; they equally delved into and surveyed various solutions, technologies, and principles.

B.

Machine Learning for Network Optimisation

In [1], a survey of various research works on the opportunities, solutions, challenges, and roles of machine learning applications in 5G network slicing and orchestration is presented. It emphasises the need for efficiently selecting the right machine learning algorithm to enable intelligent and self-organising wireless systems.

Zhao et al. [7] proposed an RL-based approach for joint slice admission control and network resource allocation. Their results show that RL outperforms conventional methods in terms of resource utilisation in dynamic, unknown environments, slice admission rate, and network revenue. Authors in [8] also proposed an RL-based framework for adaptive 5G network resource allocation in sliced radio access networks.

Long short-term memory (LSTM) networks were utilised in a study in [9] for predicting mobile traffic demands in 5G networks. The performance of LSTM models was evaluated on real-world mobile traffic datasets. The simulation result discussed in the study shows that LSTM outperforms traditional time series forecasting models like ARIMA. A deep-learning-based approach was also proposed in [10] in a study to proactively allocate network resources in sliced networks using traffic prediction models.

Despite the substantial progress achieved, the quest for an efficacious use of limited physical resources in mobile networks still exists, which drives the continuous research and innovation in the space. Our proposed machine-learning-based framework aims to contribute by combining a series of supervised learning techniques for intelligent resource allocation and dynamic network slicing as well as reinforcement learning for traffic prediction and proactive scaling.

2.1. Current Challenges in 5G Network Slice Orchestration

2.1.1. Dynamic Resource Allocation Complexity

Fifth-generation networks are designed to handle highly diverse services like high-speed video streaming, massive IoT connectivity, and ultra-reliable low-latency communications, all running on the same infrastructure. This presents a challenge of increasing the complexity of dynamic resource allocation.

The constant fluctuation of resource demand by competing services makes real-time resource allocation a complex problem. Traditional methods, such as static or manual allocation, are too slow and inefficient to handle these dynamic needs.

2.1.2. Need for Intelligent Optimisation

With the growing complexity and scale of 5G networks, manual or rule-based approaches are no longer effective, as they cannot adapt quickly to the continuously changing demands of the network.

A more intelligent, automated approach is required and, by learning from historical data and real-time conditions, machine learning can predict demand, allocate resources dynamically, and ensure high network performance.

3. Proposed Optimal Orchestration Process

3.1. Framework Overview and Technical Contributions

Our framework makes three significant technical contributions to the field of network orchestration: the Enhanced Sub-slice (eSS) pipeline, the Optimal Network Slice Service Orchestration (ONSSO) framework, and the Company Network as a Service (CNaaS) model.

The Enhanced Sub-Slice (eSS) integrates the LazyPredict module, which dynamically selects optimal machine learning algorithms based on real-time network conditions, ensuring adaptive resource allocation that responds to changing network demands.

The proposed ONSSO framework builds upon the eSS pipeline to provide autonomous orchestration capabilities. The framework employs Ray Tune for hyperparameter optimisation, enabling fine-grained control over resource allocation decisions. This integration significantly improves the framework’s ability to adapt to dynamic network conditions while maintaining optimal performance.

The third contribution introduces the CNaaS model, which transforms how enterprise networks utilise 5G capabilities. This model enables organisations to leverage customised network slices without managing complex infrastructure, representing a significant advancement in network service delivery.

3.1.1. Enhanced Sub-Slice (eSS) Architecture

The eSS architecture comprises three interconnected modules that work in concert to enable efficient resource allocation and network optimisation. At its core, the Traffic Prediction Module implements Online AutoRL, incorporating both historical data analysis and real-time pattern recognition. This dual approach enables the system to maintain high prediction accuracy while adapting to emerging traffic patterns.

The module’s effectiveness is quantified through several key performance metrics, including root mean square error (RMSE) for prediction accuracy and model convergence time. Our experimental results demonstrate that the Traffic Prediction Module achieves an average RMSE of 3.8% for short-term predictions (within 5 min) and 6.5% for longer-term predictions (up to 1 h), with model convergence typically achieved within 850 ms. These results represent a significant improvement over traditional prediction methods, which typically show RMSE values of 8.2% and 12.4% for short-term and long-term predictions, respectively.

The Slice Orchestrator functions as the central intelligence of the eSS architecture, employing LazyPredict to dynamically select and optimise machine learning algorithms. Through extensive testing, we identified that Random Forest algorithms consistently achieved the highest performance for our use case, with an average accuracy of 94.2% in resource allocation decisions. This performance is particularly notable compared to traditional static allocation methods, which typically achieve an accuracy rate of 78.5%. For short-term predictions (within 5 min), the system achieved a root mean square error (RMSE) of 3.8%, while longer-term predictions (up to 1 h) maintained an RMSE of 6.5%.

The Resource Allocator module completes the eSS architecture by implementing cloud-native scaling principles. Our implementation demonstrates particular efficiency in resource utilisation, maintaining an average utilisation rate of 85%, while ensuring sufficient headroom for demand spikes. The module’s response time to resource allocation requests averages 50 ms, with 99% of requests processed within 100 ms, meeting the stringent requirements of ultra-reliable low-latency communication (URLLC) applications.

The three main components are illustrated in Figure 2.

3.1.2. Optimal Network Slice Service Orchestration (ONSSO)

The ONSSO framework implementation focuses on three critical aspects: monitoring granularity, orchestration efficiency, and policy enforcement. The real-time monitoring system operates on a 100 ms sampling interval, providing highly granular insights into network performance. This sampling rate was chosen based on extensive testing that showed it provides an optimal balance between monitoring accuracy and system overhead. The monitoring system tracks key performance indicators, including latency (achieving consistent sub-10 ms measurements for URLLC slices), throughput (varying based on slice type, with eMBB slices achieving up to 20 Gbps), and resource utilisation efficiency (maintaining 90% average utilisation across active slices).

The orchestration engine employs a novel approach to network slice template generation, utilising machine learning models trained on historical performance data. This approach has demonstrated a 40% reduction in slice deployment time compared to traditional template-based approaches, while maintaining higher accuracy in resource allocation predictions. The service function chaining algorithm achieves optimal function placement, with an average chain setup time of 200 ms, representing a 30% improvement over conventional chaining methods. Figure 3 illustrates the architectural overview of the proposed framework, while Figure 4 depicts the relationship between the eSS module pipeline and the proposed new vertical CNaaS framework, with the eSS module implemented within the CNaaS. Figure 3 highlights the components spanning from data ingestion to network topology. The data, which can be live, historical, generated by AI, or a combination of these, is fed into the eSS model pipeline. This pipeline preprocesses the data, analyses it, and identifies the requirements for the network slice template description.

The framework handles varying types of network data and adapts to different data sources. It provides mechanisms to preprocess and clean the monitored traffic data, evaluates required resources for each application datum, and formulates the network slice template for network slice orchestration. The flexibility of LazyPredict allows the framework to incorporate new algorithms as they become available. Users can update the existing Python library to include new algorithms, which the framework will then automatically evaluate, ensuring continued relevance and adaptability. Figure 4 shows the eSS module pipeline within the CNaaS framework. Metrics such as the service requirements, business goals, and policies are fed into the pipeline. The eSS pipeline then outputs a bespoke slice or a standard 5G network slice depending on available network resources.

Our validation methodology combined theoretical analysis with practical implementation using Google Colab for machine learning model training and MATLAB/Simulink for network simulation. The integration between these platforms was achieved through a custom API that enabled real-time data exchange and validation.

The MATLAB/Simulink environment was configured to simulate various network conditions, including:

• High-load scenarios with multiple competing slices;
• Dynamic traffic patterns mimicking real-world enterprise networks;
• Resource constraint situations to test scaling capabilities;
• Multi-tenant environments with varying QoS requirements.

3.1.3. Company Network as a Service (CNaaS)

The CNaaS model introduces a paradigm shift in enterprise network service delivery. Our implementation demonstrates its feasibility within current 5G ecosystems through careful consideration of technical requirements and operational constraints. The model achieves service isolation with negligible performance overhead (less than 1% additional latency), while maintaining flexible resource allocation capabilities.

The technical implementation relies on advanced network function virtualisation (NFV) principles, with container-based network functions achieving 99.999% availability. Resource allocation efficiency is maintained through dynamic scaling capabilities that respond to demand variations within 500 ms, ensuring consistent service quality across all enterprise applications.

Our proposed ML-based framework aims to improve the model integration challenge by implementing a series of ML algorithms pipeline as a component in the framework; a typical enterprise network environment is shown in Figure 5. This shows varying departments within an organisation with diverse needs. For example, the organisation or company marketing department might demand an eMBB slice for remote sales, while the same company with vehicles carrying bio-reactive materials might require a low-latency network slice.

With the proposed ML-based framework, the 5G network slice requirements of live applications over the company network will be monitored, analysed, and optimally orchestrated in real time.

3.2. Implementation Environment

The experimental validation of our proposed framework employed a comprehensive dual-platform approach, combining the machine learning capabilities of Google Colab with the network simulation features of MATLAB/Simulink R2023b. This integration enabled thorough testing of both the algorithmic components and network behaviours under realistic conditions.

The Google Colab environment was configured with TensorFlow 2.9.0 and Scikit-learn 1.0.2, running on a 16GB RAM GPU instance to ensure adequate computational resources for model training and evaluation. LazyPredict 0.2.12 and Ray Tune 2.3.0 were implemented for automated algorithm selection and hyperparameter optimisation, respectively. This configuration provided the necessary computational power for training and evaluating multiple machine learning models simultaneously.

Network simulations were conducted using MATLAB/Simulink’s 5G Toolbox and Communications Toolbox, with a simulation granularity of 1 ms over a 24 h period. The simulation environment was configured to support multiple network slice types, including eMBB slices with throughput ranges of 100 Mbps to 20 Gbps, URLLC slices maintaining latencies below 1 ms with throughput of 1–10 Mbps, and mMTC slices supporting device densities up to 1 million devices per square kilometre.

The experiments were conducted using the Google Colab integrated development environment alongside MATLAB and Simulink. Google Colab was chosen for its accessibility, scalability, and ability to leverage GPUs and TPUs, which are critical for processing complex machine learning models. The environment was configured as follows:

• Processor: Tesla K80 GPU (12GB GDDR5 VRAM) or TPU;
• RAM: 16 GB;
• Python Version: 3.8.x;
• Libraries: LazyPredict, AutoRL, Scikit-learn, TensorFlow, Keras, and Matplotlib.

These specifications supported the computational demands of training reinforcement learning (RL) models and running simulations under dynamic 5G conditions. Training each model required an average of 3 min per epoch, with total runtimes of approximately 6 h. Peak memory usage reached 10 GB during the evaluation phase, demonstrating the framework’s efficiency and scalability for large-scale applications.

The simulation setup supported up to:

• 10,000 Ultra-Reliable Low-Latency Communications (URLLC) users;
• 50,000 Enhanced Mobile Broadband (eMBB) users;
• 500,000 Massive Machine-Type Communications (mMTC) devices.

Key performance indicators (KPIs) included:

• Throughput: data rate aggregation for eMBB slices, critical for assessing network performance [11,12].
• Latency: end-to-end latency for URLLC slices, essential for low-latency applications [13,14].
• Resource utilisation: efficient management of physical network infrastructure [5,8].

Figure 6 illustrates the simulation setup. Data processing and analysis were divided between MATLAB (numerical computation and algorithm development) and Google Colab (machine learning execution). Simulink simulated dynamic traffic loads and network conditions, with results fed back for performance evaluation.

3.3. Dataset Description

The framework utilised both publicly available datasets and synthetic data:

• Public datasets included 5G network traffic traces capturing metrics like latency, throughput, and reliability.
• Synthetic data Simulated in MATLAB and Simulink to model edge cases, such as high mobility and traffic surges.

The implementation followed these steps:

• Data collection: real-time data acquisition from network slices.
• Feature engineering: extracted relevant metrics for model training.
• Model training: leveraged historical data to build predictive models.
• Dynamic allocation: automated resource allocation using model outputs.
• Performance monitoring: ensured compliance with service-level agreements (SLAs).

3.4. Algorithm Performance Analysis

The evaluation of machine learning algorithms through LazyPredict revealed significant performance variations across different models. Random Forest emerged as the superior algorithm, achieving 94.8% accuracy, with an average inference time of 12 ms. This performance notably surpassed other tested algorithms, with XGBoost achieving 93.2% accuracy (15 ms inference), Neural Networks at 91.7% (25 ms inference), and SVM at 89.5% (18 ms inference).

Resource prediction accuracy demonstrated strong performance across different time horizons, with RMSE values of 3.2%, 5.8%, and 8.9% for short-term (5 min), medium-term (1 h), and long-term (24 h) predictions, respectively. This progressive increase in error rates with prediction horizon length aligns with expected behaviour in time-series forecasting applications.

3.4.1. Network Performance Evaluation

The framework demonstrated exceptional performance in slice orchestration, achieving an average deployment time of 1.8 s, with 92.5% resource allocation accuracy. Service function chain setup was completed within 215 ms, while scaling operations responded to demand changes within 480 ms. These metrics represent significant improvements over traditional orchestration approaches.

Resource utilisation metrics showed consistent efficiency, maintaining average CPU utilisation at 78.5% and memory utilisation at 82.3%. Network bandwidth efficiency reached 86.7%, indicating effective resource management while maintaining sufficient headroom for demand spikes.

3.4.2. Message Bus Implementation and Performance

The message bus implementation utilised Apache Kafka deployed at the mobile edge, configured with three broker nodes to ensure high availability. The system employed 12 partitions per topic with a replication factor of 3, optimising for both performance and reliability. This configuration achieved remarkable performance metrics, including average message latency of 3.8 ms and throughput of 100,000 messages per second, while maintaining 99.999% delivery reliability.

3.5. Automation of Algorithm Selection

To facilitate the orchestration process, the integration of LazyPredict, Auto Reinforcement Learning (AutoRL), and Ray Tune is introduced. These tools automate the selection of machine learning algorithms based on the changing environment, optimising the orchestration process.

LazyPredict module automates the process of algorithm selection based on performance metrics. It is utilised by the Network Slice Requirements Analyser and Network Sub-Slice Mapper components of the Slice Orchestrator to automatically evaluate and select the best supervised learning algorithms for their respective tasks. This approach significantly reduces the time and effort required for manual experimentation, analyses about 30 machine learning algorithms against the monitored environment, and allows for rapid identification of the most effective algorithms tailored to the current specific resource allocation challenges. The LazyPredict Python library used in the proposed framework can be modified to include future algorithms as new methods are developed and integrated into the Python ecosystem. This ensures that our framework remains relevant and adaptable to advancements in ML techniques.

3.5.1. Automated Algorithm Selection Using LazyPredict

LazyPredict is particularly advantageous due to its ability to provide a comprehensive overview of various machine learning algorithms and their performance metrics without extensive prior configuration. This capability allows for rapid experimentation and iteration, which is crucial in dynamic environments where requirements may change frequently. Additionally, LazyPredict streamlines the initial selection process, significantly reducing the time needed to identify suitable algorithms for specific tasks.

Figure 7 illustrates the implementation of the LazyPredict automation workflow process. The implementation demonstrates the framework’s capability to analyse multiple algorithms simultaneously, generate performance metrics for comparison, visualise results for algorithm selection, and adapt to changing environmental conditions.

3.5.2. Reinforcement Learning Algo Selection Using AutoRL

AutoRL was chosen for its capacity to automate the selection and fine-tuning of reinforcement learning algorithms, enhancing online learning capabilities. Its design allows for adaptive learning in real time, making it particularly well suited for environments characterised by fluctuating demands and resource availability. Compared to traditional reinforcement learning approaches, AutoRL reduces the need for manual intervention and optimisation, thereby improving efficiency and responsiveness in resource orchestration.

Figure 8 presents the AutoRL process flow, which demonstrates the initial environment state assessment, the algorithm selection among multiple options (DQN, PPO, and A3C), hyperparameter optimisation, and continuous monitoring and adaptation.

3.5.3. Efficient Model-Hyperparameter Optimisation Using Ray Tune

The choice of Ray Tune stems from its ability to efficiently manage hyperparameter searches through parallel execution and state-of-the-art optimisation techniques. This capability allows us to explore a wide range of hyperparameter configurations, ensuring models operate at peak performance in real-time environments. The adaptability and effectiveness of the resource allocation strategies further enhanced with the integration of Ray Tune into the eSS model pipeline.

Together, the combination of LazyPredict, AutoRL, and Ray Tune not only streamlines the algorithm selection and tuning processes but also fortifies the framework’s capability to dynamically allocate resources in response to varying network conditions. This automated approach facilitates an unassisted, adaptive end-to-end resource management system, addressing the pressing challenges of orchestrating 5G network slices effectively.

3.6. Real-Time Resource Access and Collaboration

The implementation of our message bus architecture at the mobile edge has demonstrated significant improvements in cross-provider collaboration. Deployment testing shows average message delivery latencies of less than 5 ms within local regions, enabling near-real-time resource co-ordination among different providers. The system’s subscription management mechanism ensures secure and efficient data distribution among providers, with update frequencies dynamically adjusted based on resource availability and demand patterns. This approach has resulted in a 60% reduction in orchestration delays compared to traditional request–response mechanisms.

Currently, the Virtual Network Operator (VNO) has to wait for the Application Provider (AP) to send the user equipment slice requirement and then form the slice template; the VNO will then send to the Infrastructure Provider (IP), who is waiting on VNO for this information in order to allocate the required physical resource. The IP then sends the required orchestrated 5G network physical resources to VNO for the slice orchestration, which then maps this to demanding user equipment. The proposed ONSSO framework streamlines this process with an included message bus deployed on the mobile edge, which mobile network providers can subscribe to in real time. The proposed Enhanced Sub-slice module (eSS) pushes the real-time user equipment demands and the predicted slice demands onto the message bus. The network providers such as the AP, VNO, and IP can subscribe to this in real time and proactively identify network resource needs, thereby efficiently orchestrating network slices and allocating resources optimally.

4. Results and Analysis

4.1. Overview of Simulation Results

The traditional resource allocation models in 5G networks are compared with the simulation results obtained from the proposed ONSSO in this section. The simulation results are presented, analysed, and compared against both the heuristic-based orchestration and static slicing benchmarks The simulation results demonstrate the efficacy of the ONSSO framework in improving both latency and throughput, supporting its viability as a competitive solution for resource allocation in 5G networks.

• Key Performance Indicators

To evaluate the success of the ONSSO framework, several key performance indicators (KPIs) were established. The MATLAB/Simulink simulations assessed the following KPIs:

• Latency: the time taken to process user requests and allocate resources. This is measured in milliseconds (ms).
• Throughput: the amount of data successfully transmitted through the network. This is measured in megabits per second (Mbps).
• Resource utilisation: the efficiency of resource allocation relative to total available resources. This is expressed as a percentage.

• Results Summary

The results of the simulation trials are summarised below (

Table 1. Result summary on latency and throughput.

Metric	Traditional Model	Proposed Model
Latency (ms)	60–90	10–25
Throughput (Mbps)	20–50	250–500

):

The

Table 2. Performance evaluation result of the dataset types.

Metric	Real-Life Data	Synthetic Data	Augmented Data (Real + Synthetic)
Training Time (h)	6	5.5	7
MAE	1.25	1.3	1.2
RMSE	1.75	1.8	1.7
R-Squared	0.85	0.83	0.87
CPU Utilisation	70	65	75
Memory Usage (GB)	12	11	14

below shows the evaluation metrics and a comparison between the source of the training data. It evaluates metrics based on limited real data, synthetic data, and augmented data where synthetic and real data were combined for training.

Latency Analysis: The simulation results show that the ONSSO framework consistently demonstrated lower latency across all scenarios compared to traditional static resource allocation methods, as shown in Figure 9 and the data in Table 1.

In low-demand scenarios, latency for the traditional models averaged around 60 ms, while, in high-demand scenarios, it peaked at approximately 90 ms as studied by Van Damme et al. [15]. In contrast, the proposed eSS model achieves a latency range of 10 ms to 25 ms, demonstrating a substantial improvement in responsiveness. It can be deduced that static resource allocation models cannot adjust quickly to fluctuating demand, leading to higher average latencies, unlike the ONSSO framework whose dynamic nature of eSS facilitates the allocation of resources in real time based on service priority and network load, reducing waiting times for resource availability. The average utilisation rate is measured by aggregating the percentage of allocated resources for all active slices over the total available resources during the simulation period, averaged across multiple time intervals.

Throughput Analysis: Throughput significantly increased with the ONSSO framework, reaching a maximum of 300 Mbps during peak user demand, showcasing its ability to handle high data traffic efficiently. The throughput values for the traditional allocation model vary from 20 Mbps to 50 Mbps, as studied by Y. Xu et al. [16]. The proposed framework, however, achieves throughput values ranging from 250 Mbps to 500 Mbps, highlighting its effectiveness in resource allocation for dynamic 5G network slicing. It can be deduced that traditional throughput’s fixed nature of resource allocation in traditional 5G networks does not fully utilise available bandwidth, especially during peak times, unlike the ONSSO framework, which dynamically adapts resource allocation to maximise bandwidth utilisation, allowing higher data rates and smoother performance and incorporates reinforcement learning via AutoRL.

Resource Utilisation: The resource utilisation percentage showed an improvement, maintaining an average utilisation rate of 85% during peak demands, indicating effective resource management. The proposed ML-based framework avoids over-provisioning and efficiently manages resources by proactively allocating network slices based on the varying demands, while the traditional model approach has lower performance, as shown in the results, due to its inability to adapt to the varying demands. It lagged in responsiveness, which degraded its performance. The ONSSO framework’s innovations in sub-slicing and dynamic resource allocation significantly enhance both latency and throughput.

Comparative Analysis: The simulation results were compared against baseline models [17], illustrating the superiority of the ONSSO framework in managing resources. Figure 9 illustrates the comparative analysis of latency and throughput between the ONSSO framework and traditional methods. The figure compares the traditional model and an ONSSO model across two key performance metrics: latency and throughput. The left graph shows the latency comparison, measured in milliseconds (ms):

• The traditional model has a much higher latency of around 75 ms, with error bars indicating some variation;
• The ONSSO model shows significantly lower latency at approximately 18 ms;
• This indicates the ONSSO model is about 4 times faster in terms of response time.

The right graph shows throughput comparison, measured in Megabits per second (Mbps):

• The traditional model has relatively low throughput at around 30–40 Mbps;
• The ONSSO model demonstrates dramatically higher throughput at about 375 Mbps;
• This suggests the ONSSO model can handle roughly 10 times more data per second.

Overall, these results strongly suggest that the ONSSO model outperforms the traditional model in both aspects. It responds much faster (lower latency) and can process much more data per unit time (higher throughput). The error bars on both graphs indicate the variance in measurements, with the ONSSO model showing a particularly large variation in its throughput performance.

4.2. Discussion of Results

In our evaluation, we compared the latency achieved by the ONSSO framework with traditional resource allocation mechanisms. The traditional methods exhibited latency values between 60 and 90 ms, as documented in studies by Van Damme et al. [16]. In contrast, the ONSSO framework demonstrated significantly lower latency values, ranging from 10 to 25 ms, due to its dynamic, AI-driven resource allocation capabilities. This performance improvement highlights ONSSO’s potential to support ultra-low latency applications, which is critical in meeting 5G’s stringent requirements.

This improvement can be attributed to the ability of the ONSSO framework to predict resource requirements in real time using Automated Reinforcement Learning (AutoRL), reducing the delays introduced by static or reactive resource allocation.

The heatmap in Figure 10 visualises the latency across different network slices, regions, and time intervals under the proposed ML-based ONSSO framework.

Low-latency regions: the majority of the network slices exhibit reduced latency, demonstrating the effectiveness of the proposed ML-driven orchestration framework in dynamically optimising slice performance. The clustering of light-shaded areas highlights improved real-time responsiveness, likely attributed to the adaptive allocation of resources.

High-latency zones: a few regions in the bottom-left quadrant display persistent latency spikes. This could be due to heavy resource demands or suboptimal scheduling at those points. These findings suggest that, while the model performs optimally for most slices, further tuning of resource allocation strategies may be required to address latency outliers.

The throughput increases drastically in the ONSSO framework, reaching 250–500 Mbps, compared to 20–50 Mbps for traditional methods. Throughput measures the rate at which data are successfully transmitted over the network, typically expressed in megabits per second (Mbps). For 5G networks, throughput is a key indicator of network capacity, and it determines the ability of the network to handle high-speed data traffic.

Figure 11 provides a comparative analysis of algorithms employed in the ONSSO framework orchestration, evaluating their performance in terms of latency reduction, accuracy, and computational efficiency.

The packet loss is reduced to less than 1%, indicating better data integrity. Packet loss refers to the failure of data packets to reach their destination, which is detrimental in real-time applications like video conferencing, online gaming, and industrial automation, where lost data can lead to quality degradation or communication breakdowns. In comparison, in traditional 5G slicing, packet loss rates can vary between 2 and 5%, particularly under high network loads.

Traditional static resource allocation approaches often lead to either over-provisioning or under-utilisation of resources, with average utilisation rates around 60–75%. By contrast, the ONSSO framework, driven by AI and AutoRL, dynamically adjusts resource allocation based on real-time network conditions and demands. In our simulations, the ONSSO framework consistently achieved utilisation rates between 85 and 95%, indicating more efficient use of available resources while maintaining flexibility to handle traffic spikes.

Figure 12 illustrates CPU, memory, and bandwidth utilisation across various network slices during the operation of the AI-based network orchestration framework. The graph highlights the resource usage patterns over time or across network slices, providing a comprehensive view of how computational and bandwidth resources are distributed under dynamic orchestration conditions.

CPU usage remains relatively stable across most slices, indicating that the orchestration model effectively distributes computational workloads. Stability in CPU utilisation suggests the model’s efficiency in preventing resource exhaustion. Memory consumption shows periodic peaks, particularly during slices with high data-processing demands. These peaks correlate with traffic surges or resource-intensive operations, suggesting that the AI model dynamically adjusts memory allocation to meet real-time requirements.

The predicted slice simulated with MATLAB/Simulink in the proposed ML-based framework was compared against the true slice and an accuracy of 0.99 was achieved with our proposed model. Figure 13 shows the Confusion Matrix for the ONSSO prediction performance across slice categories.

The matrix evaluates the classification performance of our ML-based framework by displaying the number of true positive (correct predictions) and false predictions for each slice category. In this matrix:

• Row 0 (True Slice = 0) corresponds to the bespoke slice, where 796 out of 797 instances are correctly predicted, with one misclassification.
• Row 1 (True Slice = 1) corresponds to the URLLC slice, with all two instances correctly predicted.
• Row 2 (True Slice = 2) corresponds to another bespoke slice, where 71 out of 72 instances are correctly classified, with one misclassification.

From the result shown in Figure 13, it can be seen that Slice 0 and 2 are orchestrated as bespoke slices while slice 1 was identified as a uRLLC slice. Each slice is characterised by parameters such as bandwidth, latency, throughput, data rate, jitter, packet loss, and data volume, as well as metrics related to speed, packet delay, and reliability. These parameters are set to align with the specific requirements of applications and data traffic types. The ONSSO framework’s configuration process tailors each slice to fulfil the requirements of diverse applications. For instance, the URLLC slice (slice 1), characterised by high bandwidth, low latency, and high reliability, is designed to support applications that require ultra-reliable and low-latency connections. Conversely, bespoke slices (slice 0 and slice 2) provide more flexible configurations that can accommodate applications with medium or varied performance needs. This tailored configuration ensures that each application receives the appropriate resources, aligning network capabilities with application demands and optimising performance across different use cases. The low-latency, high-bandwidth, and low-packet-loss characteristics of slice 1 correctly indicate the slice to be a uRLLC slice. Similarly, the high-latency characteristics compared to a medium-latency slice 0, the low bandwidth of slice 2 compared to the medium bandwidth requirement of slice 0, and the low throughput of slice 2 compared to the medium throughput of slice 0 all indicate that the resources required to orchestrate slice 0 will be more than the network function resources required for slice 2; hence, slice 2 will be cheaper to orchestrate than slice 0 and slice 1 will be the most expensive slice to orchestrate.

On further evaluation of the results, as the framework orchestrates the best alternative slice fit based on available resources, it is possible that the one application that was predicted to use slice 2 instead of 0, and vice versa, can be the alternative allocated best fit at the time. Figure 14 demonstrates the output of our ML-based framework for determining the best fit slice and an alternative slice for various network applications (protocols). This is essential for efficient Network Function Chaining (NFC).

The model, as shown in Figure 14 below, correctly chose slice 2 as an alternative best fit slice for applications requiring slice 1 in the event of limited available resources. This will help meet the network slice requirement of all applications by efficiently managing and allocating the network slice dynamically to applications requesting them while proactively catering to available resources.

The ONSO framework effectively matches each application to a network slice that best fits its performance requirements. Each row represents a unique application or protocol, identified by its protocol name, where:

• Best_Fit_Slice: the slice category assigned by the framework to meet the application’s requirements (e.g., slice 2 corresponds to bespoke).
• Alternative_Slice: the fallback slice category if the best fit slice cannot meet resource demands.

For example, protocol ’GOOGLE’ is the best fit for slice 2, with slice 1 as the alternative. Protocols ’AMAZON’ and ’YOUTUBE’ are also allocated slice 2, with a fallback to slice 1. This highlights the framework’s ability to dynamically allocate and suggest alternative slices to optimise resource usage.

The metrics used to improve the Quality of Service (QoS) and the Quality of Experience (QoE) of the orchestrated slice were classified into low, medium, and high thresholds based on 3GPP recommendations [17]:

• Low Traffic Load: supports applications such as email, general browsing, and IoT sensors with minimal demands on bandwidth and latency. User Traffic Rate: up to 5 Mbps per user; Network Resource Utilisation: 0–30%; Latency: 10–20 ms.
• Moderate Traffic Load: manages services like video conferencing and streaming, with moderate demands on resources. User Traffic Rate: 5–50 Mbps per user; Network Resource Utilisation: 30–70%; Latency: 20–50 ms.
• High Traffic Load: targets high-demand applications like autonomous vehicle control or gaming, requiring significant resources and ultra-low latency. User Traffic Rate: over 50 Mbps per user; Network Resource Utilisation: >70%; Latency: 1–10 ms.

The resource utilisation analysis highlights the AI model’s capacity to optimise resource management dynamically. While the model ensures efficient CPU and bandwidth allocation, periodic peaks in memory usage suggest the need for improved memory prediction techniques for high-demand slices. Overall, the findings validate the model’s robustness in resource utilisation while maintaining operational efficiency.

The CNaaS model is designed to be compatible with current 5G ecosystems, leveraging existing network slicing and virtualisation technologies. Key considerations include:

• Scalability: the model supports dynamic scaling to accommodate varying company sizes and demands.
• Interoperability: it is designed to integrate with standard 5G network functions and management systems.
• Deployment: implementation can be achieved through collaboration with network operators, utilising their infrastructure to offer tailored network services to enterprises.

While the model is feasible with current technology, successful deployment would require addressing challenges such as security, quality of service assurance, and regulatory compliance. The CNaaS model leverages existing 5G technologies such as network slicing and virtualisation, making it a viable solution for modern network demands. However, challenges including integration complexity, security concerns, and the need for standardised APIs must be addressed to ensure seamless implementation.

4.3. Limitations and Future Work

Despite the promising results demonstrated by the ONSSO framework, some key limitations will benefit from further investigation. This section outlines these limitations and suggests areas for future work to enhance the framework’s performance, scalability, and applicability in evolving 5G and beyond networks.

4.3.1. Limitations

• Overhead in Model Training

One challenge observed during the simulation phase is the computational overhead introduced by real-time model training and tuning. While Ray Tune offers efficient hyperparameter optimisation, training time could be a bottleneck, especially in networks with rapid changes in traffic patterns. This is particularly relevant for ultra-low-latency services, where any delay in resource allocation decisions can degrade performance.

• Limited Real-World Validation

The current version of the ONSSO framework has been validated through simulated environments, using generated and modelled datasets. While these results are promising, real-world deployment and testing in a live 5G network environment are necessary to confirm its effectiveness and reliability. Simulations, while controlled and comprehensive, cannot account for all the uncertainties and complexities of real-world network conditions.

• Security Concerns

As ONSSO introduces machine learning to autonomously control resource allocation, new security vulnerabilities might emerge, particularly concerning model integrity. Malicious attacks on the models (e.g., data poisoning) or adversarial attacks could compromise decision making, leading to suboptimal performance or network outages. Currently, these risks are not explicitly addressed in the framework.

4.3.2. Future Work

• Integration with 6G Networks

As research and development for 6G networks progress, future iterations of ONSSO should explore compatibility with 6G architectures, particularly the integration of AI-native network management and quantum communication technologies. ONSSO’s machine learning models should evolve to accommodate the ultra-dense and ultra-low-latency environments expected in 6G.

• Security Enhancements

Addressing the security concerns identified earlier, future work should focus on integrating adversarial defence mechanisms and blockchain-based solutions to ensure the integrity and trustworthiness of machine learning models within the ONSSO framework. These solutions could help mitigate the risks of data tampering or malicious model manipulation.

• Real-World Deployment and Testing

The next step for ONSSO is to move beyond simulation environments and test the framework in real-world 5G networks. Collaborations with industry stakeholders and telecom providers could facilitate pilot projects where ONSSO is deployed in live network environments to validate its performance under real-world traffic, user behaviour, and network conditions.

5. Conclusions

The emergence of 5G and beyond networks presents unprecedented opportunities and challenges, particularly in resource allocation, network slicing, and the management of highly dynamic environments. This paper introduced the Optimised Network Slice Selection and Orchestration (ONSSO) framework, Enhanced Sub-slice (eSS) model, and a new Company Network as a Service Vertical (CNaaS), a novel approach leveraging machine learning algorithms such as Ray Tune, LazyPredict, and AutoRL to enhance the orchestration of network slices. By focusing on latency, throughput, and energy efficiency, ONSSO delivers superior performance in comparison to traditional resource allocation models. The approach in this study dynamically adjusts slices by sub-slicing and service chaining to orchestrate a bespoke network slice on demand. Supervised learning by automating the selection of best fit machine learning algorithm was leveraged for accurate feature selection and network function classifications to form a network slice. Reinforcement learning was leveraged upon for proactively learning and forecasting the network resource required by time. This framework dynamically adjusts slice resources to match time-varying demands.

Final Remarks

In conclusion, this paper contributes a powerful framework for optimising network slicing and resource allocation in 5G networks. The ONSSO framework demonstrates the potential of machine learning in orchestrating network resources more effectively, delivering considerable improvements in latency, throughput, energy efficiency, and adaptability. Though limitations remain, the advancements presented by ONSSO provide a solid foundation for future research and real-world application, with the potential to revolutionise network management in 5G and beyond.

The insights gained from this study have the potential to influence future designs of network orchestration systems, driving advancements in wireless communication technology and setting new standards for intelligent, adaptable, and high-performance networks.