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

1. Introduction

The fifth generation (5G) cellular networks aims to transform mobile broadband communication delivering unprecedented quality of service, connectivity and flexibility in mobile networks [2]. Network slicing enables the paradigm shift in mobile broadband communications with its ability to create multiple virtual verticals from end to end while sharing the available physical infrastructure [3]. Network sub-slicing further divides each network slice into granular sub-slices which are dedicated to specific service, tenant or application instance. This manages the available network resource and allocates only what is required to each service or user demanding a network slice. This helps reduce the capital expenditure cost which was identified as a current 5G network slice orchestration challenge. Each orchestrated network sub-slice is configured to meet specific user equipment’s network requirement. There are standard network slices, such as ultra-reliable low-latency communications (URLLC), enhanced mobile broadband (eMBB), and massive machine-type communications (mMTC) [4]. Orchestrating a whole standard slice with all the required resource to configure that slice might result into over provisioning and allocating more resource than is required by the user equipment. This could increase the risk of denial of slice for other user equipment due to the limited available physical resource. The network resource available is limited and poses a challenge to mobile network availability on user equipment. This is currently a significant challenge and there is high demand as well as need to manage the limited available physical infrastructure efficiently to meet the needs of diverse equipment requesting network slice [5]. Efficiently provisioning a network slice to meet the varying network traffic patterns and fluctuating service demands requires smart and intricate decisions regarding the allocation of the limited available network resource including the physical infrastructure, scaling and isolation of the network functions [5]. The network sub-slice is made up of network functions which are chained together to form end-to-end network slice. The ability to define, reorganise, share and chain these specific required network functions enables optimal allocation of resources and allows for high flexibility and customisation. Figure 1. Shows the isolated end-to-end sub-sliced networks and how they can make up a single network slice sharing network resource pool.

5G networks are complex and dynamic so the tradition rule based of network slice management and orchestration may struggle to adapt to the needs of 5G networks [6]. Some of the tradition rule based network slice are:

• Static Slicing which has preconfigured and defined rules or policies. It does not offer flexibility and does not change with carrying user demands or network conditions.
• Threshold-based Scaling which is a reactive approach that scales up and down based on the predefined threshold to scale.
• Priority-based Resource Allocation which is another rule based network slice management whereby network slices are assigned priority levels. This ensures higher priority slices get the required resource before considering allocating resources to any other lower priority network slice.
• Heuristic-based Orchestration is a reactive approach, which uses Heuristic algorithms or rule of thumb to make decisions on network slice resource allocation. This decision could be based on various rule of thumb such as historical data, expert knowledge or simplified model of the slice behaviour
• Policy-based Management which is based on predefined policies to define the rules for allocating, orchestrating and scaling network resources

In this context, recent studies has shown that machine learning (ML) techniques have emerged as promising solutions with its ability to make smart and data driven decisions as well as enable a more robust, agile and efficient network slice orchestration [1].

In this context, recent studies has shown that machine learning (ML) techniques have emerged as promising solutions with its ability to make smart and data driven decisions as well as enable a more robust, agile and efficient network slice orchestration [1].

In this paper, we propose an end-to-end novel vertical CNaaS framework to optimally orchestrate 5G network slice over a whole network for collocated departments or distributed company departments networks in real time there by improving the Quality of Service (QoS) and Quality of Experience (QoE) across the entire company networks. As part of this proposed new vertical, a novel ML-based framework was also proposed for effectively sub-slicing network slice in real time to use only the required network function or suggest network functions that can still meet the user needs but with much reduced resource allocation depending on network resource availability and dynamic user requirements. This chains the defined network functions and optimally orchestrate a 5G network slice. Our framework leverages lazy predict which is a python module of series of supervised machine learning algorithms. The supervised algorithm contain both regression and classified machine learning algorithms which dynamically chooses the best-fit algorithm for a new unknown environment. Additionally, reinforcement learning (RL) techniques is employed to make intelligent predictions accurately on future network traffic pattern and make decisions on slice resource allocation on demand. This forms part of the proposed Enhanced Sub-Slice (eSS) module. The proposed eSS module also finds the best alternative fit network slice thereby ensuring all user devices get a network slice that meet their needs. The module also incorporates the concept of cloud computing for scaling to either pay for more resources or configure to use the default alternative best fit only. This caters for proactive scaling of slice resources to meet anticipated demands. We further discussed the following in this paper as part of the proposed novel -ML based 5G slice orchestration framework:

• Propose an ML-based framework for optimal orchestration a New Vertical 5G Enterprise Network As A Service – Company Network As A Service (cNAAS).
• Propose and validate an Enhanced Sub Slice Machine (eSS) Learning Pipeline Model
• Propose and validate an alternative Network Function Chaining
• Propose and validate an auto Machine Learning Model Comparison Fit In Real Time
• Propose an enabler for seamless collaboration and integration between network providers and users

The performance of our proposed ML based framework was extensively evaluated and the result demonstrated significant improvement in key performance metrics such as latency, bandwidth, throughput, jitter, packet loss and resource utilisation when compared to conventional methods. Explainable AI is showcased in all aspect of the validation.

2. Literature Review

Optimally orchestrating and managing 5G network slice is being studied by different researchers exploring different techniques and approaches. In this section, we review the state-of-the-art in relevant domains, including 5G network slicing, 5G network resource management, and the role machine learning has played in optimisation of 5G network slice.

A. 5G Network Slicing

The concept of slicing a virtual network end-to-end forming multiple logical network slice which shares physical infrastructure has been a key enabler for 5G networks. The study in [4] worked on a comprehensive survey of network slicing in 5G, which discussed the 5G networks enabling technologies, its motivation, and challenges involved. The study identified Software-Defined Networking (SDN), Network Functions Virtualisation (NFV), Cloud Computing and Virtualisation as the 5G networks enabling technologies. 5G network slice ability to support various use cases and diverse services, its ability to efficiently utilise resource by creating dedicated slices for services, multi-tenancy with the potential new business models and improved security and isolation were motivation highlighted by the study. Ability to optimally orchestrate a network slice end-to-end and reducing its life cycle complexity were some of the challenges they identified. Zhang et al. [3] highlighted the importance of efficiently orchestrating 5G network slice by studying resource allocation strategies and mobility management for 5G networks slice.

There has been several studies on frameworks and architectures to optimally orchestrate 5G network slice. Software-defined networking (SDN) and network function virtualisation (NFV) were leveraged upon to propose a framework in [5] in creating and managing 5G network slice efficiently. Authors identified the need for automation to enable optimal network slice orchestration. An SDN/NFV-based architecture for network slicing was proposed which comprises of a hierarchical control plane, a virtualised data plane, and slicing management and orchestration functions. It ultimately identified network slicing as a key enabler for supporting diverse user network slice requests. In [11], authors studies into softwarisation in 5G and 5G network slicing, they equally delved into and surveyed various solutions, technologies, and principles.

B. Resource Management in Network Slicing

Efficiently allocating the limited available network resource is crucial in optimally orchestrating a 5G network slice to meet specific user device service requirements. Researchers in [7] proposed a mobile transport platform architecture for flexible resource sharing in their study to solve the challenge of managing network resource allocation and efficiently sharing it. Their study also emphasised the need for automating network resource allocation for improved user experience and the need for efficient algorithm. In [13], a mixed-integer linear programming model was developed and for optimal end-to-end slice resource allocation. They presented a heuristic algorithm to solve the NP-hard problem efficiently. The simulation result from their study shows the overall slice satisfaction and network resource utilisation in comparison to the static allocation methods.

The demand for 5G network slice is ever changing and fluctuating. Researchers have been studying how to efficiently and dynamically scale network slice resources to meet the dynamic needs. Deep reinforcement learning-based approach was proposed in [14] for adaptive resource scaling in network slices. They formulated the problem as a Markov decision process and used deep reinforcement learning (DRL) to learn an optimal policy to optimally allocate network resource in a dynamic time-varied network environment. In another study to enable 5G network slice scaling and dynamic resource sharing, Yan et al. [15] proposed a hierarchical resource orchestration framework.

C. Machine Learning for Network Optimisation

Machine learning techniques have been studied by various researchers to optimise various aspects of communication networks with significant positive results. Research in [1] surveyed various works of researchers on the opportunities, solutions, challenges and roles that machine learning applications has played in 5G network slicing and orchestration. The application of supervised, unsupervised machine learning as well as reinforcement learning paradigms in 5G networks was discussed in the study. They emphasised the need for efficient and selecting the right machine learning algorithm so as to enable intelligent and a self-organising wireless systems.

Reinforcement learning (RL) is a machine-learning algorithm that has been widely used by many researchers and explored for 5G network slice resource management and allocation. In a study to optimise resource allocation and network slice admission by Zhao et al. [15], they proposed an RL-based approach for joint slice admission control and network resource allocation. The result shown in the study from the simulation result shows that RL outperforms conventional methods in terms of resource utilisation in dynamic unknown environment, slice admission rate and network revenue. Authors in [18] also proposed an RL-based framework for adaptive 5G network resource allocation in sliced radio access networks.

Another machine learning technique that has been studied for proactive network resource scaling and traffic prediction is supervised learning techniques. Long short-term memory (LSTM) networks was utilise in a study in [19] for predicting mobile traffic demands in 5G networks. The performance of LSTM models were evaluated on real world mobile traffic dataset. The simulation result as discussed in the study shows that LSTM outperforms traditional time series forecasting models like ARIMA. A deep learning-based approach was also proposed in [20] in a study to proactively allocate network resources in sliced networks using traffic prediction models.

Despite the substantial progress achieved, there remains a requirement for a comprehensive and intelligent frameworks that addresses the challenge of optimally allocating the limited available network resource and efficiently orchestrate a 5G network slice by integrating multiple machine learning techniques. Our proposed machine learning-based framework aims to fill this gap by combining series of supervised learning techniques for intelligent resource allocation and dynamic network slicing as well as reinforcement learning for traffic prediction and proactive scaling.

3. The Proposed CNaaS Framework

We propose a novel machine learning-based framework to address the challenge of 5G network optimal network slice orchestration and a novel ML-based model pipeline to address resource allocation. This approach utilises a synergistic combination of supervised learning techniques and reinforcement learning (RL) to dynamically sub-slice and optimally orchestrate 5G network slices. The approach we propose optimises network resource allocation in real time as demands dynamically changes and while proactively scale network resources to meet predicted future network slice demands. This proposed framework enables a new proposed vertical 5G Enterprise Network As A Service which is tagged Company Network As A Service (CNaaS). The CNaaS does not only automates the optimal orchestration of diverse 5G network slices end-to-end to meet the varying needs of diverse application across an entire company’s network but also transforms the way 5G mobile network slice providers collaborate.

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 networks physical resources to VNO for the slice orchestration who then maps this to demanding user equipment. The proposed CNaaS 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 unto 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 slice and allocating resources optimally. Recent works which highlights the potential of such hybrid approaches of using multiple machine learning algorithms for network optimisation has been the motivation behind the leveraging of multiple ML paradigms [15,16]. Their studies emphasises the advantage of using hybrid machine learning algorithm for network slice orchestration for prediction, monitoring and allocation of resources. The authors highlight the potential of zero touch network management with opportunities of hybrid ML for joint optimisation of network resource allocation. Key challenges such as data management, model integration and online learning were highlighted in their study. 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 2. This shows varying departments within an organisation with diverse needs. For example, the organisation or company marketing department might demand an eMBB slice for to 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, analyses and optimally orchestrated in real time.

A. System Architecture

To address the challenges of optimal 5G network slice orchestration and resource management, we propose a novel machine learning-driven framework that leverages the strengths of different ML techniques. Our approach utilises a synergistic combination of Lazypredict or AutoML, supervised learning, and time-series forecasting techniques to dynamically sub-slice and orchestrate 5G network slices. The framework optimises 5G network resource allocation in real time and proactively scales 5G network resources to meet anticipated future demands. Figure 3, shows the architecture overview of the proposed framework. While Figure 4. Shows the relationship between the eSS module pipeline and the proposed new vertical CNaaS framework with eSS being implemented within the CNaaS. Figure 3. highlights the components from data ingestion to network topology. The data which is live, historical, generative AI or an combination of any of these is fed into eSS model pipeline which pre-processes the data, analyses it, and identifies the requirements for the network slice template description. The allocation of resources by the eSS and orchestration of the 5G slice goes through sub-slice creation and modification to determine the best network functions and service functions to be allocated thereby efficiently managing the available resources. The performance of the sub-slice is monitored and optimised. Within the framework, the security and compliance function is applied based on the enterprise or company’s policies. The proposed framework is fully customisable and adaptable. Access control policies and other company’s specific policies are applied. This then goes through the service function chain policy engine which describes the service function templates and them configures the service function configuration. The best fit or alternative best fit is the allocated and the ability to auto scale is also implemented in the framework. The resource allocator uses the cloud-computing concept by auto scaling the virtual network function as required based on defined policy and budget while ensuring all applications diverse requirements are met in the dynamic environment.

One of the major challenges of the use of ML approach is the availability of 5G dataset. The more data is available, the better the model is trained. This research work uses augmented data of real world historical 5G network dataset and generative AI. This makes it possible for any new network to start using this framework and start learning with the available dataset with the limited available dataset from the start. As the historical dataset grows over the network, the proposed model keeps training and adapting to the network needs. We evaluated and compared the result of using just the historical dataset against the proposed augmented data and the result shows better accuracy with throughput, jitter, bandwidth and packet loss in augmented data compared to the historical dataset only. This shows the more available data helped with the model training and its accuracy.

The proposed machine learning-based framework consists of three main components, as illustrated in Figure 5:

1) a Traffic Prediction Module which primarily leverages reinforcement learning technique for anticipated 5G network slice traffic demand prediction and sending the traffic predictions to a slice orchestrator;

2) a Slice Orchestrator which utilises supervised learning models like Random Forest (RF) and SVM for resource allocation and then sends the resource allocation decisions to resource allocator;

3) a Resource Allocator which receives the resource allocation decision from the slice orchestrator and implements it.

In the Traffic Prediction Module, Online learning Automated Machine Learning (AutoML) is further leveraged on to enhance the RL’s agent performance and adaptability. Similar designs which were proposed for hierarchical network resource orchestration in sliced 5G networks inspired this modular architecture. [15,17]. From their study, the hierarchical orchestration framework enabled dynamic end-to-end 5G network slicing by supporting inter-domain network resource optimal allocation. The hierarchical orchestration framework has three layers: the service management, which handles the end-to-end network slice lifecycle management, the network slice orchestration which handles the orchestration of 5G network slice across multiple domains, and the resource orchestration which handles and manages the allocation of physical/virtual 5G network resource within a domain in order to meet the slice requirements. This supports flexible sharing of network resource across layers, within and across domains.

With the proposed Enhanced Sub-Slice (eSS) model pipeline, network functions can be optimally shared among isolated network sub-slices thereby efficiently managing, orchestrating and allocating the available resource. The eSS allocates only the required network functions to be chained to make a 5G network slice thereby reducing the risk of over provisioning and reducing cost of allocating standard slices. The eSS model also orchestrates the best alternative fit network slice to reduce the risk of denial of slice and improve Quality of Experience (QoE). Based on the available physical resources, the eSS will analyse the best-fit slice for all live applications and predicted live applications to come online and ensure the available resources are optimally allocated to meet the needs of live applications. It proposes the alternative network slice with lesser network resource requirements but yet still meeting the needs of the application to enhance QoS and QoE.

The Traffic Prediction Module:

The Traffic Prediction Module utilises a two stage learning process:

• Reinforcement learning which is an offline learning that learn historical data and predict traffic trends and anticipated demands of 5G network slice. This component is responsible for forecasting future network traffic patterns for each network slice.
• AutoML, which is an online learning that continually adapts to the dynamic environment conditions and evolving user requirements. The RL agent adaptively updates its policy based on the live feedback, refine its predictions and enables its ability to dynamically respond to the evolving traffic pattern.

The RL agent of the Traffic Prediction Module learns to predict the future 5G network slice requirements through trial-and-error interactions with its environment by observing the network's state and receiving rewards or penalties based on the accuracy of its predictions. The RL agent's state space is made up of various network parameters while the action space is made up of the possible slice configurations. Current slice utilisation, application-specific Quality of Service (QoS) requirements and user mobility patterns are some of the network parameters. Factors like bandwidth allocation, resource prioritisation, and latency targets, are some possible slice configurations. A crucial aspect of the Traffic Prediction Module is the exploration-exploitation trade-offs. The module exploits its current knowledge base to make informed and accurate predictions. The module also explores new slice configurations to uncover potential improvements.

The AuoML helps automate and optimise RL agent’s training by automating hyper-parameter tuning as the performance of RL algorithm often depends on the right selection of hyper-parameters, enhanced feature engineering to automatically identify the relevant features from the raw network data and traffic patterns, model selection by automatically selecting the right model or using ensembled techniques to automatically combine multiple models, transfer learning to automate the transfer of knowledge from pre-trained models, and continuous adaptation which continuously monitors the RL agents performance and automatically adjusts its parameters or the underlying RL algorithm. This Traffic Prediction Module enables a more robust 5G network slice demand predictions in dynamic 5G network environments.

The Slice Orchestrator:

The core component is the Slice Orchestrator and forms the core of the eSS module pipeline. It manages the overall 5G network slice orchestration process. The Traffic Prediction Module sends traffic predictions to the slice orchestrator and the slice orchestrator interacts with the Resource Allocator to optimise the allocation of network slice resource. The slice orchestrator consist of two subcomponents, which are:

• The Network Slice Requirements Analyser, eSS with an inbuilt metrics KPI estimator and which uses lazypredict, a python module to analyse the dataset, automatically select best fit machine learning algorithm for the given network dataset and then train the best supervised machine learning algorithm (e.g., Random Forests, SVM, Gradient Boosting, Neural Networks). This analyses the network traffic data to identify the 5G network slice requirement for the given network pattern. It identifies the best fit supervised learning classifier (e.g., Decision Trees, Support Vector Machines) for training, feature engineering, hyperparameter tuning and selecting the model with optimal performance considering the key metrics such as latency, throughput, bandwidth, jitter and packet loss rate. Although lazypredict is configured by default due to its lightweight capabilities, AutoML can also be configured where computational storage is not a challenge to take advantage of its online learning and automating features.
• Network Sub-Slice Mapper, part of the eSS which uses bespoke inbuilt python function to select the best-fit or alternative best fir network functions configurations to be chained based on available network resources and predicted network resource requirements for future slices. The most suitable network sub-slice is mapped to the applications for optimal performance considering the key metrics such as latency, throughput, bandwidth, jitter and packet loss rate.

A similar approach which was employed in recent studies on RL-based resource allocation was a key motivator to employing a supervised learning technique to learn an optimal policy for resource allocation through interactions with the environment (i.e., the 5G network slices) [12,15]. The problem was formulated as a Markov Decision Process in their study and Deep Q-network (DQN) algorithm was used to handle the high-dimensional state and action spaces.

Physical resources such as compute, network, and storage are allocated and managed by the Resource Allocator across the network slices. Based on the learned policy, the Resource Allocator exposes an interface for the Slice Orchestrator to request resource deallocations or allocations. This Resource Allocator component enhances and utilises existing frameworks for resource management in virtualised network environments [21,22]. Their study delved into the Network Function Virtualisation (NFV), its infrastructure and components. The VNF orchestration and deployment, service function chaining, performance optimisation and resource utilisation plays a critical role in optimally allocating shared network resources.

B. LazyPredict for Algorithm Selection

Lazypredict or AutoML depending on available computational resources 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. Lazypredict is a lightweight python module, which automatically selects and evaluates the best fit supervised learning technique. Where computational resource is available, AutoML, which is computational intensive, can be used to automate the process of feature engineering, model selection, and hyper-parameter tuning. Tools, such as AutoGluon, Google Cloud, or Auto-Sklearn enables efficient selection of the most appropriate ML algorithms for the given data and problem. Figure 6 shows the eSS module pipeline within the CNaaS framework. The network slice requirement is analysed within the pipeline using machine learning algorithm, the best fit machine learning algorithm based on the network dataset is automatically selected, and appropriate network functions required to meet the slice needs are optimally chained and mapped to the user equipment. Metrics such as the service requirements, business goals and policies, are fed unto the pipeline. The eSS pipeline then outputs a bespoke slice or a standard 5G network slice depending on available network resource.

C. Time-Series Forecasting for Traffic Prediction

Future traffic patterns for each network slice is predicted by the Traffic Prediction Module by utilising specialised time-series forecasting techniques. Ray Tune is used to automatically select the best RL technique to fit the analysed 5G network environment. Ray Tune is lightweight compared to AutoRL and strikes a good balance between functionalities and lightweight. It automates RL algorithm selection and hyperparameter tuning. These models are trained with the selected RL algorithm by Ray tune on historical traffic data, which may include features such as day of the week, frequency in a day, time intervals, time of day, and various contextual information, as suggested in previous work on mobile traffic prediction [8,9]. The traffic predictions for future time steps generated from the trained models are then fed into the Slice Orchestrator for proactively allocating and scaling resources.

The combination of the machine learning algorithms in the eSS module pipeline aims to optimally orchestrate 5G network slices and efficiently allocates network resources to ensure quality of service and improved quality of experience in an unknown networks with changing and diverse service requirements. The use of lazypredict techniques for selecting best fit machine learning algorithm with the use of supervised learning technique for analysing the network slice requirement and the sub-slice mapping , coupled with the use of specialised time-series forecasting methods for traffic prediction helps formulate our proposed framework. The limitations of single-paradigm is addressed by this integration of multiple ML techniques and leverages the strengths of each technique, as suggested by previous studies on hybrid ML approaches for network optimisation [10,11].

4. Case study

To evaluate the performance of our proposed ML-based framework for optimal 5G network slicing and orchestration, we conducted extensive simulations modeling a metropolitan 5G network deployment, following an approach similar to prior studies on network slicing and resource management [9,14]. The study introduced a slice-aware embedding technique to map the complex relationship between the 5G network slice and their required network functions. A proactive adjustment then periodically adjusts the network slice configuration based on predicted future slice demand requirements.

A. Simulation Setup

The simulated 5G network covers a metropolitan area of 10 km x 10 km, with a total of 25 macro cell sites. This network configuration is based on typical urban deployment scenarios considered in related works [3,17]. Three different standard network slice types were considered in the studies:

• eMBB (Enhanced Mobile Broadband): Caters for high speed, capacity and bandwidth mobile broadband services like immersive augmented reality (AR), video streaming, virtual reality (VR) and web browsing.
• URLLC (Ultra-Reliable Low-Latency Communications): Enables reliable data transmission and Caters to mission-critical applications with stringent latency such as autonomous cars, remote surgery and industrial automation.
• mMTC (Massive Machine-Type Communications): Enabling massive connectivity for IoT devices that requires low data rate and low energy consumptions requirements in smart city and sensor network applications.

For evaluation purposes, a maximum of 10,000 URLLC users, 50,000 eMBB users and 500,000 mMTC devices simultaneously was configured for the network infrastructure. This aligns with capacity estimations and requirements outlined in related 5G network slicing studies [4,6].

The key performance indicators (KPIs) considered are:

• Throughput: Recent studies [16,23] refers to this commonly used evaluating metric in aggregating data rate for eMBB slices during a network slice performance evaluaton.
• Latency: A critical metric for slices such as URLLC slices for End-to-end latency for URLLC slices. It’s a critical requirement for low-latency applications [24,25].
• Resource Utilisation: This is an essential goal for efficiently managing the underlying shared physical network infrastructure resources [5,8].

Based on recent studies of in real-world mobile networks [1,5] which studies the pattern of the 5G network, our simulation models accounts for the periodic bursts of mMTC slices, the dyanamic time-varying loads with diurnal patterns for eMBB and URLLC slices, and environmental factors like channel conditions and user mobility. Assumptions made were consistent with related simulation studies [19,26].The simulation studies developed an enhanced 5G network slice simulation framework on top of the ns-3 network simulator. Our simulation setup leveraged on the parameters, configurations and assumptions from their studies. The Slicesim supports the simulation of multiple end-to-end slice with their performance comparisons. While Slicesim handled the specific network slicing, Matlab and Simulink were used for the overall network simulation of the proposed framework and Google Colab was used for running the machine learning python code. The simulation experimental setup in this study is shown in Figure 7.

The data source is fed into the MATLAB and Google Colab. MATLAB handles the numerical computation, the algorithm development and the data analysis while the Google Colab runs the python code thereby leveraging available cloud resources and integrating machine learning models. MATLAB sends the processed data to Simulink which performs simulation and sends the performance evaluation results. The data input was varied and evaluated by varying different traffic loads and network conditions to mimic a real live dynamic 5G network environment.

B. Benchmarks

We compare the performance of our proposed framework against two benchmarks:

Static Slicing: The traditional static network resource allocation are allocated based on pre-defined configuration without any adjustment. Standard network slices are predefined, configured and allocated as specified. This approach is commonly used in existing networks [6,27].

Heuristic-based Orchestration: A reactive scaling and allocation of network resources based on observed demands, without any learning components or prediction. Recent studies has shown that some network slicing frameworks employs this rule-based orchestration techniques benchmark [18,19].

C. Implementation Details

The reinforcement learning component uses a Deep Q-Network (DQN) with a dense neural network architecture, trained using experience replay and a target network for improved stability, following best practices in DQN implementation [19,20]. The state space includes metrics like traffic loads, slice requirement, resource utilisation, and QoS indicators for each slice, as commonly used in RL-based network optimisation tasks [13,15]. The aim is to maximise a reward function that considers 5G network slice resource utilisation, admission rates, and fairness among slices. Dueling DQN and prioritising experience replay technique were used to improve performance and training stability while considering resource constraints.

The supervised learning component employs the light weight lazypredict module which evaluates the current network traffic data against series of supervised machine learning algorithms and choose the best fit for the data.

Similar simulation in recent works [28,29] were implemented by using a custom 5G network simulator built on top of Python libraries such as TensorFlow and NumPy. Before the models were deployed in the similar, the RL and supervised learning models were trained first. This adopts similar methodology in recent RL-based network optimisation studies [2,12]. Recent studies have used open source simulators leveraging on ns-3 network simulator. 5G-LENA-Slicer is used for simulating end-to-end 5G network slicing on commodity hardware. The results were evaluated by comparing the results with a real life test bed results. Slicer was another network slice simulator in their studies. It also supports and manages the creation of multiple 5G network slices with varied configurations. It also provides the ability to simulate 5G radio access technologies.

Following the network slicing performance evaluation guidelines suggestions [30,31] and varying the network traffic patterns and intensity, we evaluated both the proposed framework and benchmark methods. This was done across a range of scenarios covering diverse operating conditions. Standard techniques for performance evaluation and result analysis were used to statistically analyse the collected performance metrics to quantify improvements [32,33].

In the next section, the advantages of our proposed ML-based approach over the conventional methods was highlighter as we discussed the simulation results.

5. Performance Evaluation

In this section, the simulations results are presented, analysed and compared against both the heuristic-based orchestration and static slicing benchmarks.

Proposed ML-Based Framework vs Static-Based vs Heuristic-Based

A. Throughput Analysis

Our simulation results shows that the proposed ML-based framework consistently outperforms other benchmarks as shown in Figure 7. Higher throughput is achieved across all load levels for the eMBB slices under varying traffic loads.

The throughput gains at low to moderate loads are evident but not as significant as the throughput gain from moderate to high as the loas increase. This is because at low to moderate load, the network resources are readily available compared to when the available resources need to be efficiently managed at moderate to high loads. Our proposed ML- based framework's ability to dynamically and intelligently manage slices, scale network slices and allocate resources as demand changes becomes more pronounced. From the simulation, result shown In Figure 6, there is a significant 28% improvement in throughput for our proposed ML-based approach compared to static slicing and 19% improvement over heuristic orchestration at the highest load scenario.

B. Latency Performance

Our proposed ML-based framework shows significant advantage for URLLC slices with stringent latency requirements according to the analysed simulation results as shown in Figure 8.

The result shows that our ML-based approach consistently maintains low latency from low load to high load compared to the static and heuristic benchmarks which suffered degradation in performance due to the delayed reaction to scaling. Proactively scaling network slice and efficiently optimising slice allocations with our ML-based framework, effectively and consistently meets the 1ms latency target for URLLC slices unlike the benchmarks which violate this requirement under moderate to high loads.

C. Resource Utilisation

The simulation result shows that our proposed ML-based framework not only achieve better throughput and latency when compared to benchmarks, but it also achieves better resource unitisation as illustrated in Figure 9.

Our ML-based framework avoids over-provisioning and efficiently manages resources by proactively allocating network slice based on the varying demands while the static slicing approach has a lower performance as shown in the result due to its inability to adapt to the varying demands. The heuristic method lags behind in responsiveness which degrades it’s performance.

From the simulation result, an improvement of 15-29% was achieved by our proposed ML-based orchestration framework in resource utilisation compared to static slicing and an improvement of 8-18% was achieved compared to heuristic-based orchestration.

Proposed ML-Based Framework Performance Evaluation

The proposed ML-based framework which we have tagged as Optimal Network Sub-Slice Orchestrator(ONSSO) performance and accuracy was also evaluated. The predicted slice was compared against the true slice and accuracy of 0.99 was achieved with our proposed model. Figure 10.

The confusion matrix shows the high accuracy in our proposed model where out of 797 applications that required Slice 0, 796 got correctly predicted as slice 0 while only 1 got predicted as slice 2. Out of the 2 slices that required slice 1, all the 2 slices correctly got predicted. Out of 72 slices that required slice 2, 71 got correctly predicted and 1 was predicted as slice 0. On further evaluation of the results, as the framework orchestrates best alternative slice fit based on available resource, it is possible that the one application that was predicted slice 2 instead of 0 and vice versa can be the alternative allocated best fit at the time. Figure 11 shows the result of the predicted alternative best-fit slice.

Alterative Network Function Chaining

From the result shown in Figure 11, It can be seen that Slice 0 and 2 are orchestrated as bespoke slices while slice 1 was identified as a uRLLC slice. The low latency, high bandwidth and low packet loss characteristics of slice 1 correctly indicates the slice to be an 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 indicates that the resources required to orchestrate slice 0 will be more that the network functions 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. The model as shown in Figure 11 below correctly chose slice 2 as an alternative best fir 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 for available resources.

D. Conclusion and Discussion

The simulation results clearly demonstrate the effectiveness of the proposed ML-based framework for optimally orchestrating 5G network slice. Our approach dynamically adjust 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 upon for accurate feature selection and network functions 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 adjust slice resources to match time-varying demands.

The proposed ML-based framework outperforms the conventional static and heuristic-based methods across key metrics lie latency, throughput and resource utilisation. This is shown by its ability to manage high load and dynamically adjust proactively to varying complex demands effectively.

While the simulation results are promising, there are still potential areas for further improvement and exploration. Investigating other lighter libraries that could be used to evaluate multiple types of supervised and unsupervised machine learning algorithm to effectively choose the right algorithm for the model at a faster speed. Exploring transfer learning techniques to help with model training and smoother adaptation in different network environments. Additional contextual information such as channel conditions and mobility patterns can be added into the ML models in future and further studied.