Exploring the Potential of AI in Network Slicing for 5G Networks: An Optimisation Framework

Abstract

ABSTRACT

The integration of artificial intelligence (AI) into 5G network slicing is essential for overcoming limitations in autonomous management and precise resource allocation in complex network environments. Traditional methods struggle with dynamic adaptability, often requiring manual intervention and lacking scalability. This research leverages AI models, specifically logistic regression and long short‐term memory (LSTM) to automate and optimise real time slice allocation. During testing across various values of the regularisation parameter (alpha), the models achieved classification accuracy up to 95% at alpha = 0.1 and maintained over 65% at higher values, demonstrating robustness. We also implement dynamic programming of segment routing over IPv6 (SRv6) Identifiers, enabling accurate differentiation of up to 40,000 enhanced mobile broadband (eMBB) slices, as well as ultra‐reliable low‐latency communication (URLLC) and massive machine type communications (mMTC) types. An adaptive application programming interface (API) based framework further adjusts SRv6 traffic engineering (SRv6 TE) policies in real time, ensuring uninterrupted service. High receiver operating characteristic‐area under the curve (ROC AUC) scores, reaching 0.99, validate the model's strong classification performance. This approach advances automated 5G slicing by enhancing responsiveness, scalability and service quality.

Introduction

The roll-out of 5G technology has revolutionised telecommunications by enabling high data rates, low latency and scalable connectivity. At the heart of this transformation lies network slicing, which allows operators to create multiple virtual networks within a single physical infrastructure, each tailored to specific application requirements such as high-speed video streaming or industrial internet of things (IoT). However, the increasing complexity and dynamism of modern network environments present significant challenges in real-time performance optimisation and resource utilisation. To address these challenges, this research integrates machine learning (ML) and artificial intelligence (AI) techniques, specifically logistic regression and LSTM networks, into the 5G network slicing architecture. These models automate the assignment of SRv6 identifiers and dynamically adjust SRv6 TE policies in response to network incidents like link failures. This AI-driven approach aims to enhance resilience, reduce manual intervention and consistently meet quality of service (QoS) demands [1]. The study investigates the following research questions:

• How can AI-driven models improve automation and resource allocation in 5G network slicing?

• What is the impact of adaptive AI control mechanisms on QoS during network disruptions?

• How does AI integration affect the overall resilience and efficiency of 5G networks?

Key contributions of this research include the following:

• A novel integration of logistic regression and LSTM models to support real-time network management.
• Automated SRv6 segment identifier (SID) assignment to streamline operations and reduce manual configuration.
• Adaptive SRv6 TE policy updates that maintain service quality during failures.
• A comprehensive evaluation of AI's impact on network resilience and operational efficiency.
• A forward-looking framework to support future research in AI-driven 5G network management.

The remainder of this article is structured as follows: Section 2 reviews related work, Section 3 outlines the methodology, Section 4 presents and analyses results and Section 5 concludes with key findings and future directions.

Literature Review

Building on the challenges and research questions outlined in the introduction, this section surveys recent advancements in machine learning–enabled network slicing and SRv6-based 5G architectures. As 5G networks become more complex with the increasing number of connected devices and diverse service demands, the need for intelligent, efficient network slicing has grown significantly. Artificial intelligence and machine learning have emerged as key technologies to support automation in slice design, deployment, fault detection and resource management. Gupta and Misra [2] used the unicauca version 2 public dataset, containing seven days of traffic and 87 attributes, to compare several ML algorithms for slice selection. random forest and K-nearest neighbours (KNN) outperformed support vector machine (SVM) and decision tree in terms of accuracy, using simulations run on an NVIDIA graphics processing unit (GPU). To further address challenges in slicing performance, Abidi et al. [3] developed a metaheuristic approach using Glowworm Swarm Optimisation and Deer Hunting Optimisation. These algorithms improved feature weighting based on parameters such as device type, bandwidth and packet delay. The system then used a hybrid deep belief network and neural network to classify slices into eMBB, mMTC and URLLC categories. Evaluation was conducted using simulated data. Additionally, DeepSlice [4] applied deep learning to slice management by predicting device types based on key performance indicators such as resource usage and traffic load. The model enabled dynamic slice allocation and reallocation under congestion, although it was tested in a simulated environment. Other researchers have focused on integrating AI with network softwarisation technologies. Another prominent approach to enhancing 5G network slicing involves the integration of network function virtualisation (NFV), software-defined networking (SDN), ML, AI and big data technologies. This convergence is vital for building a comprehensive architectural and experimental framework that enables intelligent, automated and optimised network management for next-generation networks. By leveraging ML, AI, and big data analytics, this approach aims to fulfil the growing demands for intelligence, faster management and end-to-end automation in 5G infrastructures [5]. Costa-Perez [6] proposed a reinforcement learning-based slicing framework with forecasting, admission control and scheduling.

Simulations showed better service level agreement (SLA) compliance, resource use and mobility handling than static methods. Furthermore, Messaoud et al. [7] proposed a dynamic slicing model using online Gaussian Mixture Models to cluster devices and assign them to slices in real time. Their method aimed to maintain QoS by mitigating resource starvation and prioritising time-sensitive traffic. Tests showed it adapted effectively to varying traffic demands. Similarly, Singh et al. [8] proposed a load balancing approach using an ML-based sub-slicing framework. Each logical slice was divided into sub-slices, each optimised for specific performance goals such as spectral efficiency, low latency, or energy savings. They applied SVM and K-means clustering to analyse application requirements and group services with similar characteristics. As 5G networks scale with increasing devices and service diversity, ML/AI have become key to automating slice management. Techniques like sub-slicing and clustering enhance resource efficiency and service quality. Recent efforts have also combined ML with SRv6 to improve flexibility and reliability in slicing. In the study titled ‘Dynamic Reliability Scheme for SRv6-Based SFC Orchestration’, [9] the authors proposed a dynamic approach to enhance service function chaining by minimising reliance on static topologies. They used extended sub-type-length-value (TLV) fields in the SRv6 header to embed service function attributes, along with an SRv6 ping mechanism for robust orchestration. Tested under SD-WAN failure scenarios, the system redirected traffic within 50 milliseconds, significantly reducing service activation time and enabling fast, dynamic 5G slice provisioning. In ‘SRv6 Strategy Based on Network State and Behavior Prediction’, [10] a source routing approach was proposed to deliver routing tables directly to source nodes, bypassing intermediates. Built on an SDN-SRv6 information-centric networking (ICN) architecture with dual-path protection, it improved reliability, energy efficiency and congestion control under heavy load, especially against malicious nodes. ‘A Solution for Establishing Traffic Routes per Application’ [11] proposed a route control function for per-application routing, generating route identifiers and integrating with core functions like policy control function (PCF) and session management function (SMF). While it enhanced QoS through fine-grained differentiation, it faced issues with scalability, implementation complexity and standardisation. In ‘Deep Learning-Based Network Slice Recognition’, [12] fully connected neural networks achieved 97.1% accuracy across eMBB, URLLC and mMTC categories. Performance was especially strong for URLLC, though slightly lower for eMBB, and the model needs scaling for real-world deployment.

In conclusion, existing research has made significant progress in applying ML models and advancing SRv6 for 5G network slicing. Studies have utilised algorithms such as Random Forest, SVM, KNN and decision trees to enhance slice selection and network efficiency. Considerable work has also focused on improving SRv6 to support resource allocation, service differentiation and slice management. However, a key gap remains in achieving dynamic, automated SRv6 implementation at the core IP level. Real-time configuration of slice specific SIDs and continuous traffic flow monitoring are still underexplored. Traditional solutions like multiprotocol label switching (MPLS) and software defined wide area network (SD-WAN) offer basic traffic engineering but lack the adaptability and responsiveness required for modern 5G networks. MPLS relies on manual QoS configuration and lacks dynamic rerouting, while SD-WAN operates at the application layer and is unsuitable for fine-grained, per-slice control.

In contrast, our proposed methodology advances the field by introducing a real-time, AI-driven SRv6 framework. By leveraging logistic regression and LSTM models, our approach predicts slice types such as eMBB, uRLLC and mMTC and dynamically assigns and programs corresponding SIDs. This system introduces an API-based orchestration layer that continuously monitors network conditions and automatically reconfigures SRv6 TE policies to adapt to congestion, high traffic loads or failures. By automating the SID assignment process and enabling resilient, slice-aware routing at the IP layer, our approach ensures uninterrupted traffic flow and optimises resource utilisation. In doing so, our study addresses key limitations in existing SRv6 and TE solutions, providing a novel, adaptive and fully automated framework for intelligent 5G network slice management.

Network Slicing

Having examined prior research efforts and their limitations, this section introduces the fundamental concept of network slicing and outlines the architectural foundations upon which our AI-driven framework is built. Network slicing is a transformative telecommunications technology that enables the creation of multiple, isolated virtual network slices on a shared physical infrastructure. Each slice is tailored to meet the specific requirements of different services, applications or user groups, allowing for efficient allocation of network resources while ensuring optimal performance, security and QoS in line with defined SLAs [13].

These slices illustrated in Figure 1 provide dedicated environments for various service types or industry users, each with its own logical topology, reliability parameters, security levels and SLA specifications to meet distinct demands. Network slicing empowers carriers to efficiently cater to the varying needs of diverse customer segments by providing customised network functionalities for specific applications, such as the internet of vehicles (IoV) and smart agriculture. The integration of AI and ML technologies into this framework can significantly enhance the capabilities of network slicing by automating configuration, resource allocation and performance optimisation, ultimately addressing challenges related to dynamic network environments. There are different types of network slicing, including IP transport slicing, core network slicing and radio access network (RAN) slicing. Let's explore each type in detail:

• IP transport slicing: IP transport slicing focuses on segmenting and managing the transport network portion of the overall network infrastructure. It involves dividing the IP-based transport network into multiple virtual slices, where each slice operates as an independent network with its own dedicated resources and customised network functions. IP transport slicing enables efficient resource allocation, traffic management and quality assurance for different services and applications [14].

[IMAGE OMITTED. SEE PDF]

Slice IDs play a crucial role in the implementation of network slicing. To begin the process of network slicing, it is necessary to generate network slice instances and assign unique slice IDs. The physical network is then segmented into multiple network slices by allocating interface forwarding resources accordingly.

In the illustrated network (Figure 2), specific network slice instances are established on nodes P1 and P3. Furthermore, network slice interfaces are meticulously planned and linked to the respective network slice instances while ensuring that adequate forwarding resources, such as bandwidth, are reserved to accommodate the specific requirements of each slice.

• 2.

  Core network slicing: This involves virtualising and customising core network components (routers, switches and network functions) to create isolated logical networks. Each slice operates independently with its own resources and functions tailored to specific service requirements. This enables efficient service provisioning, routing, and QoS management within the core network [15].

• 3.

  RAN slicing: This involves virtualising and partitioning the RAN, including base stations and antennas. This creates multiple logical RANs, each serving specific users or applications. Dedicated radio resources and customised RAN configurations ensure optimised performance, coverage and capacity [16].

[IMAGE OMITTED. SEE PDF]

Combining these three types of slicing creates holistic, end-to-end network slices spanning transport, core and RAN networks. Figure 3 provides an overview of this comprehensive approach. In summary, network slicing, empowered by AI/ML, enables network operators and service providers to efficiently manage and deliver a wide range of services with diverse requirements. This integration maximises resource utilisation and ensures optimal performance and QoS for each service, while also addressing the inherent challenges of real-time network management in a rapidly evolving landscape.

[IMAGE OMITTED. SEE PDF]

Transport IP Network Slicing

While the previous section outlined the broader architecture and types of network slicing, this section narrows the focus to transport IP slicing, an essential enabler for scalable, service-differentiated delivery in 5G core infrastructure. In the past, transport networks supporting 2G/3G/4G services commonly used multi-protocol label switching (MPLS). However, deploying MPLS in large-scale networks introduced complexities, including managing multiple control plane protocols and configuring numerous devices across diverse domains. This research introduces end-to-end (E2E) slicing technology for 5G networks, focusing on transport layer slicing and its capabilities [17]. Transport layer slicing, specifically targeting the transport layer of the internet protocol (IP) network, creates isolated and dedicated virtual networks within the same physical infrastructure. This allows for efficient resource utilisation and customised services for different traffic types. Several technologies can achieve this, including IP/MPLS, virtual private networks (VPNs) and SR. This paper focuses on SR in conjunction with SDN due to its significant impact on network automation, as shown in Figure 4.

[IMAGE OMITTED. SEE PDF]

SR, particularly SRv6, simplifies network architecture by requiring only interior gateway protocol (IGP) and border gateway protocol (BGP) for operation. SRv6 is deployed only at ingress and egress nodes, requiring only IPv6 support from transit nodes. Without altering the original IPv6 packet format, SRv6 adds an SR header (SRH), guiding data forwarding via an explicit IPv6 path defined by a list of SRv6 SIDs [18].

SRV6 Policy

SRv6 TE Policy leverages the source routing mechanism of SR to precisely control packet forwarding within a network. This is achieved by including an ordered list of SIDs in the packet header at the source node. These SIDs explicitly define the end-to-end path from source to destination, overriding the standard IGP computed shortest path [18]. This deterministic approach offers precise control over traffic flow, enabling efficient TE. In our proposed AI-driven framework, the SRv6 TE policy plays a crucial role in ensuring optimal traffic routing for different network slices. By programmatically assigning specific SIDs to packets based on predictions from our logistic regression and LSTM models, we can dynamically steer traffic through the network, optimising resource allocation and adapting to network conditions and demands in real time. For example, during periods of congestion, our system can dynamically adjust the assigned SIDs to guide traffic away from congested paths, ensuring uninterrupted traffic flow even during network failures. This precise control offered by SRv6 TE Policy translates to significant advantages in network management. By explicitly defining paths, network operators can ensure that end-to-end service requirements are consistently met, enhancing overall network quality. This capability becomes particularly valuable when combined with SDN, creating a highly flexible and responsive infrastructure ideally suited for service-driven network environments. Therefore, SRv6 TE Policy, with its deterministic path control and adaptability, is highly recommended as the optimal operational mode for SRv6 in such scenarios. The SRv6 TE Policy working mechanism is explained in detail in Figure 5 below for better understanding.

[IMAGE OMITTED. SEE PDF]

As illustrated in Figure 6, we present a comprehensive design for an AI-driven approach to 5G network slicing, leveraging the capabilities of logistic regression and LSTM models. The objective of this design is to provide efficient, real-time classification and predictive capabilities that facilitate network slice management through automated SRv6 SID assignment, proactive TE and congestion monitoring.

[IMAGE OMITTED. SEE PDF]

• PE3 as a forwarder: PE3, acting as a forwarder, communicates the network's topology information to the controller using BGP Link-State (BGP-LS). This information encompasses details about nodes, links and TE attributes, including link cost, bandwidth and delay.

• BGP peer relationship: The controller and the head-end forwarder establish a BGP peer relationship specifically for the BGP IPv6 SR-policy address family.

• SRv6 TE policy computation: The controller performs computations to create an SRv6 TE Policy, considering various factors like link cost, latency and packet loss rate. Subsequently, the controller transmits the SRv6 TE policy to the head-end forwarder via the established BGP peer relationship. The head-end forwarder then utilises this information to generate SRv6 TE policy entries.

• Path information delivery: The controller delivers path information to the ingress (PE1) of the network. The ingress node then generates SRv6 TE policies. SRv6 TE policies can be used to steer traffic through the network in a variety of ways. For example, traffic can be steered based on the colour values of routes, the differentiated services code point (DSCP) values of packets or the service class values of packets.

Proposed Design

Building on the transport slicing architecture and SRv6 TE mechanisms discussed in the previous section, we now introduce our proposed AI-enhanced framework that enables intelligent slice classification, automated SID assignment and dynamic traffic engineering in real time.

The architecture comprises three main layers: the network orchestration layer, slice layers and user access and external connections.

• Network orchestration layer

At the core of this architecture is the network orchestration layer, which acts as the primary control and monitoring hub. This layer includes:

API: The interface enabling dynamic communication between the AI models, controllers, and underlying network infrastructure.

AI models (logistic regression and LSTM): The AI component performs two crucial functions:

• Logistic regression: Predicts the appropriate slice type for incoming traffic requests, enabling efficient, real-time SID assignment for each slice.
• LSTM model: Continuously monitors network health to detect congestion, high traffic volumes, or failures (such as a link going down). It can proactively reconfigure SRv6 TE policies to ensure uninterrupted service.

Controller: Serves as the intermediary between the orchestration layer and the slice layers, executing configuration commands and handling adjustments in SID assignment or TE policy.

This orchestration layer automates the SRv6 SID assignment through the API and dynamically modifies network paths based on real-time analysis from the AI models.

• 2.

  Slice layers

Below the orchestration layer are three distinct slice layers (eMBB, uRLLC and mMTC), each designed with specific traffic demands and QoS requirements in mind:

• eMBB slice (100::1): Designed to handle high-bandwidth, low- latency applications such as streaming or gaming. Represented in blue, this slice is assigned the SRv6 SID of 100::1.
• uRLLC slice (200::1): Designed for applications with strict latency and reliability requirements, such as healthcare or autonomous driving. This slice, in red, has an SRv6 SID of 200::1.
• mMTC slice (300::1): Allocated for massive IoT communications where high latency is acceptable. This orange slice is designated with the SRv6 SID 300::1.

Each slice layer includes dedicated router nodes and paths, isolated from the others. These paths are tailored based on the QoS requirements of each slice type, ensuring that data flows in a streamlined manner for specific applications.

• 3.

  User access and external connections

On the user side, this architecture supports various devices and applications, including hospitals, IoV, and robotic devices. The network orchestration layer automatically assigns these users to the correct slice type, guided by their application needs.

On the network side, this system connects seamlessly to 5G services, represented by the 5G and cloud icons, ensuring efficient interaction with broader telecommunications infrastructure. To validate this architecture in a real-world scenario, the next section presents a Cisco-based network design that interfaces directly with our AI models and orchestration components.

Network Design

The proposed topology in Figure 7 showcases a Cisco-based network topology that has been configured to meet our requirements in this article and to connect it to our model and API framework. The network topology is designed to provide a holistic simulation environment. It consists of 11 routers that manage data routing, and one router, which is R9, will act as the main controller of the whole topology.

[IMAGE OMITTED. SEE PDF]

IPv6 loopback interfaces are configured on each router, and the IPv6 is enabled on the data interfaces without the need of any IP configuration.

Intermediate system to intermediate system (IS-IS) is used as the IGP by enabling IS-IS on all the core interfaces, which are from R1 through R9, since R10 and R11 are considered as customer equipment (CE) routers. SRv6 is enabled on all routers and inside the IS-IS as well to advertise the locator and functions. Each locator will have the following structure: fcbb:bb00:DD0X::/48 (DD is the domain ID and X is the router ID).

To proceed with the implementation, L3VPN and BGP are configured on the PE routers R1 and R8. An internal BGP (IBGP) connection will be established between the PE and the controller R9, which will act as a route reflector as shown below in Figure 8.

[IMAGE OMITTED. SEE PDF]

To accomplish the desired SRv6 TE policy configuration as needed, TE should be enabled, which is always in a single direction. In this study TE is configured from router R1 to R8. Router R9 is the main controller that will run the proposed AI model with the API framework that we will talk about later in this article. To achieve all the targeted automation, SR-PCE must be enabled on R9 as shown below in Figure 9. Note that SR-PCE refers to a desired outcome or path through the network that meets certain criteria (e.g., bandwidth requirements and path redundancy), and the PCE is mainly software that adheres to the specified SR policy. Furthermore, SR-PCE must see the full topology to collect all the in- formation needed and monitor the topology through BGP-LS messages, BGP-LS has been configured between R9 and R4, no need to configure it with more than one router since there is one single IS-IS process running as IGP. The last step in configuring and finalising the desired network is to configure an SRv6 TE policy with latency according to the diagram in Figure 10. The link has been assigned some intentionally defined latency just for testing, and later, the AI model will be trained for further enhancement.

[IMAGE OMITTED. SEE PDF]

[IMAGE OMITTED. SEE PDF]

Then using the API any link failure in the topology will be checked and detected. SRv6 will be altered in this case, so we need to change it manually to maintain the reachability of any client. However, using this TE policy, it will be changed automatically.

Network partitioning into these three slices allows for optimised QoS parameters, resource allocation and traffic management policies, ensuring efficient service delivery across diverse use cases without compromising performance. This tailored approach maximises infrastructure utility, offering service providers significant flexibility and customisation capabilities in meeting the specific needs of various industries and applications. The adaptability of this multi-slice approach highlights the versatility of modern networking technologies and the performance level of 5G as shown in Figure 11.

[IMAGE OMITTED. SEE PDF]

The primary objective of this network setup is to partition it into three distinct slices (eMBB, uRLLC and mMTC) each governed by a dedicated SRv6 TE policy. These slices, described in detail in Section 5, enable optimized service delivery based on their specific QoS profiles, as shown in Figure 12.

[IMAGE OMITTED. SEE PDF]

As previously discussed, SRv6 extends SR, using IPv6 headers to encode segments. SRv6 simplifies the header format, supporting end- to-end paths by encoding complete path information in the packet header. This simplifies service chaining and avoids complex configurations. SRv6 is particularly valuable in modern networking contexts, including SDN, 5G networks and cloud native architectures, enabling dynamic path selection, TE and network slicing. In this work, SRv6 SIDs are used to control packet paths based on predictions from our ML model. For testing purposes, three loopback interfaces representing eMBB, URLLC and mMTC services were configured on Customer Edge (CE) routers R10 and R11. Each interface uses a distinct SRv6 TE policy tailored to its service requirements (e.g., eMBB uses SID 100::1, URLLC uses 200::1 and mMTC uses 300::1). This setup facilitates robust and seamless host-to-host communication. The AI model predicts the appropriate slice type for each packet traversing the SR domain, enabling automatic SID assignment and path selection. Furthermore, the system dynamically adjusts the SRv6 TE policy to adapt to link failures. Building on this network foundation, the next section details the design, configuration and integration of our AI models logistic regression and LSTM used for intelligent slice classification and dynamic SRv6 TE policy control.

Simulation

With AI Model

This section outlines the development, configuration and extensive evaluation of an AI-driven approach for 5G network slice management. The methodology utilises two complementary ML models: logistic regression for swift, high-accuracy classification of network slice types and LSTM networks for predicting future network demands based on historical traffic. This combined approach leverages the unique strengths of traditional ML and deep learning to address the complex, dynamic requirements of 5G network slicing. Figure 13 is a summarised flow chart that describes the whole functionality of the model.

[IMAGE OMITTED. SEE PDF]

Logistic Regression for Immediate Slice Type Classification

• Model selection and justification: Logistic regression was selected for its efficiency and interpretability in real-time multi-class classification tasks. It enables fast and accurate identification of slice types: the Logistic Regression class from the Scikit-learn library, which manages this optimisation internally. (eMBB, URLLC, and mMTC), which collectively play a vital role in maintaining stringent QoS requirements within dynamic 5G environments.

• Mathematical formulation: The logistic regression model estimates the probability that a given input feature vector x ∈ R³, consisting of:

  • x₁: QoS class identifier (QCI)
  • x₂: Packet loss rate
  • x₃: Packet delay budget
  belongs to a specific slice type k ∈ {1, 2, 3}, corresponding to the eMBB, uRLLC and mMTC categories. This probability is computed using the softmax function: 1 $$$\begin{equation}P\left( {y = k{\mathrm{|}}x} \right) = \frac{{{{e}^{w_k^Tx + {{b}_k}}}}}{{\sum_{j = 1}^3 {{e}^{w_j^Tx + {{b}_j}}}}}\end{equation}$$$where:
  • w_k ∈ R³ is the weight vector for class k,
  • b_k is the bias term for class k,
  • K = 3 represents the number of classes (slice types),
  • P (y = K | x) denotes the conditional probability that the input vector x belongs to class k.

  The numerator $${{ \in }^{W_k^T + {{b}_k}}}$$ represents the normalised score for class k, while the denominator normalises the scores across all classes to ensure the probabilities sum to 1.

  To elaborate further, the model first computes a linear combination (score) of the input features for each class k as follows: 2 $$$\begin{equation}{\mathrm{Scor}}{{{\mathrm{e}}}_k} = {{w}_{k1}}{{x}_1} + {{w}_{k2}}{{x}_2} + {{w}_{k3}}{{x}_3} + {{b}_k}\ \end{equation}$$$

  This dot product between the weight vector w_k and input vector x, plus the bias b_k, gives the raw classification score. These scores are passed through the softmax function to yield a probability distribution over the three slice types. The model then selects the class with the highest predicted probability as the output.

  The parameters w_k and b_x are learned during training using multinominal logistic regression. In our implementation we employed.

• Architecture and configuration: Using a one-vs-rest (OvR) strategy, the model handles multi-class classification by evaluating each class independently. This enables the system to rapidly classify traffic based on real-time network conditions and assign the appropriate slice accordingly.
• Hyperparameter tuning: To ensure generalisation across variable network conditions, we tuned the regularisation parameter α within the range [0.01, 1.0]. In logistic regression, the objective is to minimise a regularised loss function of the form: 3 $$$\begin{equation}\mathcal{L}\left( {w,b} \right) = - \frac{1}{N}\sum_{i = 1}^N \sum_{k = 1}^K {{y}_{i,\ k}}\log \left( {\widehat {{{y}_{i,k}}}} \right) + {{\alpha}}|w|_2^2\ \end{equation}$$$

where:

• N is the number of training samples,
• K = 3 is the number of classes (eMBB, uRLLC and mMTC),
• y_{i, k} ∈ {0, 1} is the true label of sample i for class k (one-hot encoded),
• ˆy_{i, k} ∈ (0, 1) is the predicted probability that sample i belongs to class k,
• w is the weight vector of the model,

• $${\bm\alpha} | {\bm W} |_{\bm 2}^{\bm 2}$$ is the L2 regularisation term.

  The summation iterates over all classes, and y_i,k acts as an indicator function that ensures only the term corresponding to the true class contributes to the loss. This is a standard formulation in multinomial logistic regression. The use of the (1/N) factor assumes uniform weighting of all training samples, which aligns with standard empirical risk minimisation principles. The formulation is based on cross-entropy and should not be confused with minimum error entropy (MEE).

  The L2 regularisation penalises large weight values, helping to avoid overfitting. Our experiments showed that even at higher values of α, the model-maintained classification accuracy above 89%, confirming strong generalisation and robustness across diverse traffic conditions.

• Solver and computational efficiency: The ‘lbfgs’ solver (limited-memory Broyden–Fletcher–Goldfarb–Shanno) was selected for its fast convergence, low memory footprint and effectiveness on large-scale, multi-class datasets. It is a quasi-Newton method that approximates the inverse Hessian matrix to minimise the logistic loss function: 4 $$$\begin{equation}\ \mathop {\min }\limits_{{\theta}} \ \mathcal{L}\left( {{\theta}} \right) = \sum_{i = 1}^n \ \log \left( {1 + \exp \left( { - {{y}_i} \cdot {{{{\theta}}}^ \top }{{x}_i}} \right)} \right) + \frac{{{\alpha}}}{2}|{{\theta}}{{|}^2}\end{equation}$$$where θ represents the model parameters, x_i is the input feature vector, y_i ∈ {–1, +1} is the class label and α is the regularisation parameter. This expression corresponds to the negative log-likelihood of the logistic model rather than an entropy computation. L-BFGS minimises this differentiable objective efficiently by approximating the inverse Hessian without explicitly storing it, which makes it well suited for large-scale optimisation.
• Evaluation and Feature Insights: The model was evaluated using accuracy and ROC-AUC metrics for each class, verifying its strong classification capabilities. Feature importance analysis highlighted packet loss rate and packet delay budget as the most significant contributors, offering interpretable insights for network optimisation.
• Practical Implications: The probabilistic output of logistic regression allows network operators to make fast, data-driven decisions with interpretable confidence levels. Its simplicity and efficiency make it a strong candidate for real-time deployment in 5G slicing systems, ensuring dynamic and accurate slice management under varying traffic conditions.

LSTM for Future Network Behaviour Prediction

The long short-term memory (LSTM) network was selected for its advanced capability in sequence modelling and capturing temporal dependencies, which are crucial for forecasting network behaviours in 5G slicing. Unlike standard models, LSTM can retain information across sequences, making it highly suited for understanding and predicting trends in network demand and adapting slice configurations proactively. This predictive ability supports SRv6 TE policies, enabling pre-emptive management of resources in response to forecasted congestion or service degradation.

• Architecture: Optimised for Temporal Prediction in Network Slicing

  The model architecture comprises a single LSTM layer with 50 hidden units to balance complexity and computation, followed by a fully connected output layer with a softmax activation function for multi class classification. This structure enables the LSTM to capture and process sequential patterns within network traffic data, learning dependencies over time to predict shifts in network demand accurately. By leveraging sequence-based learning, the model goes beyond static prediction, identifying demand patterns that vary with network activity, enabling dynamic and adaptive SRv6 TE policy adjustments.

• Hyperparameter Configuration for Model Convergence and Stability

  To ensure both accuracy and stability in predictions, the LSTM was trained with the Adam optimiser and a categorical cross-entropy loss function, minimising prediction errors across multiple classes.

  The model was trained on 80% of the dataset and tested on the remaining 20%, achieving strong generalisation performance suitable for real-time 5G network slicing systems.

  Training spanned 20 epochs with a batch size of 32, an optimal balance between convergence and overfitting prevention. This configuration allows the LSTM to converge efficiently while retaining generalisability across diverse network conditions, ensuring consistent predictive accuracy in new scenarios without excessive fine-tuning.

• Data Reshaping and Gaussian Noise for Realistic Modelling

  Network data was reshaped into a 3D input format (samples, timesteps and features), matching the LSTM's input structure to maintain the temporal relationship between data points. To enhance model robustness, Gaussian noise was added to critical features like packet loss and latency, simulating variability found in real-world network environments. This addition strengthens the model's resilience to fluctuations, ensuring it can reliably anticipate slice demand under dynamic network conditions.

• Metrics and Model Performance Evaluation

  • Epoch-based accuracy and loss tracking: Accuracy and loss were tracked across epochs, demonstrating steady model convergence without signs of overfitting.
  • Confusion matrix and slice demand analysis: The confusion matrix highlights the accuracy of each slice prediction, allowing fine-tuning and validating the LSTM's proficiency in forecasting slice demand with minimal misclassifications.
  • Predicted slice type distributions: Analysed to confirm the model's ability to identify trends, further supporting proactive congestion management.

• Mathematical Formulation of the LSTM Model

  The LSTM model processes input sequences using gated memory units to capture temporal dependencies. Although the internal operations of LSTM involve multiple non-linear transformations and gating mechanisms (input, forget and output gates), the output computation in our classification setup can be summarised as follows: 5 $$$\begin{equation}{\mathrm{\ Output}} = {\mathrm{softmax}}\left( {W \cdot h + b} \right)\end{equation}$$$

  where:

  • h is the hidden state vector produced by the LSTM layer (dimension 50),
  • W is the weight matrix of the dense output layer (dimension 3 × 50),
  • b is the bias vector of the dense output layer (dimension 3),
  • softmax maps the output to a probability distribution across the 3 slice types: eMBB, URLLC and mMTC.

  The input to the model is a feature vector x ∈ R³ representing QoS parameters: QCI, packet loss rate and packet delay budget. The LSTM layer processes this input (reshaped as [1, 3]) through its recurrent structure, applying a ReLU activation function rather than the standard tanh, yielding the hidden representation h. This is then passed to the dense layer for classification.

• LSTM Operations for Sequence Processing

  The full LSTM operations are defined by the following equations: 6 $$$\begin{equation}{{f}_t} = {{\sigma}}\left( {{{W}_f}{{x}_t} + {{U}_f}{{h}_{t - 1}} + {{b}_f}} \right)\end{equation}$$$7 $$$\begin{equation}{{i}_t} = {{\sigma}}\left( {{{W}_i}{{x}_t} + {{U}_i}{{h}_{t - 1}} + {{b}_i}} \right)\end{equation}$$$8 $$$\begin{equation}\widetilde {{{C}_t}} = \tanh \left( {{{W}_c}{{x}_t} + {{U}_c}{{h}_{t - 1}} + {{b}_c}} \right)\end{equation}$$$9 $$$\begin{equation}{{C}_t} = {{f}_t} \odot {{C}_{t - 1}} + {{i}_t} \odot \widetilde {{{C}_t}}\end{equation}$$$10 $$$\begin{equation}{{o}_t} = {{\sigma}}\left( {{{W}_o}{{x}_t} + {{U}_o}{{h}_{t - 1}} + {{b}_o}} \right)\end{equation}$$$11 $$$\begin{equation}{{h}_t} = {{o}_t} \odot \tanh \left( {{{C}_t}} \right)\end{equation}$$$

These represent the forget gate f_t, input gate i_t, candidate cell state C˜t, updated cell state C_t, output gate o_t and hidden state h_t at time step t. The symbol ⊙ denotes element-wise multiplication and σ is the sigmoid function.

In our case, since the input sequence has only one time step, these operations are executed once per sample, and the final hidden state h_t is passed to the dense layer for classification.

Adaptive API Framework for Real-Time Network Monitoring and Reconfiguration

To fully operationalise the predictive capabilities of both models, we developed an adaptive API framework integrating the trained LSTM and logistic regression models into the network topology. This framework serves as a dynamic decision-making system, automatically assigning SRv6 SIDs based on predicted slice types.

API Framework Key Components

• program_srv6_sid function: Simulates real-time SRv6 SID assignment based on model predictions, incorporating real-world variability. This function allows the network to adapt to changing conditions without manual intervention.

• check_for_congestion function: Monitors the distribution of predicted slice types in real time, triggering alerts when traffic volumes exceed predefined thresholds. This enables swift responses to congestion, optimising QoS across slice types.

• add_gaussian_noise function: Injects variability into core features such as packet loss and latency, mirroring unpredictable network conditions. This addition fortifies the model's resilience, allowing it to handle real-time fluctuations in a dynamic telecom environment.

Model Evaluation Insights and Practical Applications

Both models were evaluated with detailed metrics to confirm reliability across classification and sequential prediction tasks:

• Accuracy and ROC-AUC: The logistic regression model exhibited robust classification, achieving AUC values near 0.99, signifying exceptional discriminatory power across slice types.
• Precision, recall and F1-score: Precision, recall, and F1-scores were calculated for each slice type, showing high accuracy in distinguishing between eMBB, URLLC and mMTC classes.
• Confusion matrix analysis: Misclassification patterns in the confusion matrix were examined, highlighting areas for fine-tuning model predictions and supporting enhanced accuracy in slice assignment.
• LSTM epoch performance tracking: Tracking LSTM accuracy and loss over epochs validated the model's temporal prediction accuracy and revealed minimal overfitting, essential for maintaining reliability in live networks.

Final Observations and Methodology Contributions

The choice of Logistic Regression and LSTM is rooted in the balance they offer between simplicity, interpretability and the ability to capture temporal dependencies essential for proactive 5G network slicing:

• Logistic regression: Provides an efficient, interpretable solution for real-time slice classification.
• LSTM: Captures long-term network traffic patterns, enabling proactive resource management.
• Adaptive API framework: Ensures the dynamic assignment of SIDs based on slice predictions, with real-time reconfiguration for continuous network optimisation.

This hybrid methodology underscores a significant advancement in AI-driven 5G network slicing, achieving real-time adaptability with minimal human intervention.

AI Models Contribution to Dynamic Network Slice Management

The integration of AI models enables dynamic and adaptive SRv6 TE, resource allocation and QoS optimisation within 5G network slicing.

• Dynamic SRv6 TE Policy Adjustment

Predictive insights:

• By employing LSTM networks for future network demand prediction, the system can anticipate changes in traffic patterns and adjust SRv6 TE policies accordingly.
• For instance, if a surge in demand for a specific slice type (e.g., eMBB) is predicted, the system can pre-emptively allocate more resources to that slice, optimising traffic routing.

Adaptive policy adjustments:

• The adaptive API framework facilitates the dynamic assignment of SRv6 SIDs based on real-time slice predictions.
• This responsiveness allows for quick adjustments in routing paths, enabling the network to adapt to varying traffic loads and conditions.
• During peak usage times, the model can shift traffic from congested paths to less utilised ones, minimising delays and improving overall network performance.

• b.Resource Allocation Optimisation

Automated resource management:

• The logistic regression model, with its high accuracy in classifying slice types, ensures that resources are efficiently allocated according to the specific requirements of each slice.
• This automation reduces the need for manual intervention, allowing for quicker response times to changing network demands.

Real-time monitoring and adjustments:

• The check_for_congestion function continuously monitors the distribution of slice types and traffic levels.
• When thresholds are exceeded, the system can trigger alerts and initiate automated resource reallocation.
• This capability helps to prevent bottlenecks and ensures that each slice receives the necessary resources to meet its QoS requirements.

• c.QoS Improvement

Predictive QoS management:

• By accurately forecasting network demands with LSTM models, the system ensures that sufficient resources are allocated to slices requiring low latency and high reliability. For example, in critical applications like autonomous driving (URLLC), the LSTM can predict spikes in demand and facilitate proactive measures to meet stringent QoS metrics.

Continuous improvement through feedback loops:

• The integration of feedback mechanisms into the API framework allows for the collection of real-time performance data, which can be fed back into the AI models. This continuous learning process enables the models to refine their predictions and resource allocation strategies over time. As a result, enhanced QoS and user satisfaction are achieved through adaptive improvements in model performance.

Enhanced fault tolerance:

• The incorporation of Gaussian noise to simulate real-world variability helps the models develop robustness against unpredictable network conditions.
• By training on this noisy data, the models can better adapt to sudden changes in traffic patterns.
• This enables the system to maintain QoS even during adverse conditions, such as network failures or unexpected surges in demand.

API Architecture

The AI model has been built on Google Colab and the network setup has been simulated on a local network. This was done to fully integrate all simulations in one model, such as network topology, Google Colab and the API code, which has been built using Flask. In addition, ngrok is used to create a tunnel between the local network and Google Colab to run the API accordingly. Below in Figure 14, a detailed flowchart will show how the API architecture has been built.

[IMAGE OMITTED. SEE PDF]

• Workflow Explanation

The workflow integrates multiple components, from initiating API requests in Google Colab to configuring SRv6 policies based on AI model predictions. The steps are as follows:

• Google Colab: The process begins here, where a user or an automated script initiates an API request.
• Ngrok: This service is used to create a secure tunnel to the Flask API, which is crucial when the API is running locally or in a non-public environment.
• Flask API: The central component where the logic of the API is defined. This handles incoming requests routed through Ngrok.
• Load LSTM model: This branch of the Flask API indicates that there is an endpoint or function dedicated to loading an LSTM model.
• Check interface status: After the LSTM model is loaded, the API performs a status check on network interfaces, which could involve verifying the health or connectivity status of router interfaces.
• Load logistic regression model: Parallel to the LSTM model, a logistic regression model is also loaded, providing another predictive functionality within the API.
• Check predicted slice: The API checks which network slice has been predicted by the logistic regression model. This step determines the type of network traffic (e.g., eMBB, URLLC, and mMTC) expected to flow through the network.
• /configure-and-adjust srv6 TE policy: This endpoint manages the configuration and adjustment of SRv6 TE policies, potentially responding to interface status checks.
• /configure-srv6 SID endpoint: Configures SRv6 SIDs based on the predicted slice types.

With the AI models, adaptive API framework, and network simulation fully integrated, we now proceed to evaluate the performance of our system. The next section presents the experimental results and assesses the effectiveness of the proposed AI-driven approach in real-time 5G network slicing scenarios.

Results

This section discusses the experimental outcomes, emphasising the performance of the AI-driven approach in network slicing within simulated 5G environments and the interaction between the model and the API framework.

As illustrated in Figure 15, traffic between R10 and endpoints

[IMAGE OMITTED. SEE PDF]

11.11.11.2 and 11.11.11.3, hosted on R11, demonstrates the model's ability to manage slice-based traffic. Each endpoint corresponds to services classified by the model into different slice types, achieving effective traffic segregation and allocation based on application demands.

Figure 16 provides an overview of the PCE topology, displaying details gathered by the BGP-LS protocol, including the number of active nodes, links, additions and removals within the network. This real-time topology data is not only essential for immediate adjustments but is also integrated as feedback into the AI model to continuously improve its prediction capabilities. By updating the model with this information, the system enhances its understanding of network dynamics, enabling more accurate predictions and adaptability to changing network states.

[IMAGE OMITTED. SEE PDF]

In Figure 17, part of the TE policy output is displayed, highlighting how the AI model's predictions are utilised to optimise SRv6 paths dynamically. The policy output indicates that the metric type is latency, utilising colour code 300 to denote specific latency-sensitive paths. Here, the AI model dynamically selects a binding SID for each network slice type, based on real-time analysis of network conditions.

[IMAGE OMITTED. SEE PDF]

Candidate Path Selection in SRv6 Routing

Within this scenario, two candidate paths are available for traffic routing:

• Path with preference 200: This path is the default option initially calculated by the head-end, representing a typical route for traffic under normal conditions.

• Path with preference 100: This secondary path is automatically activated when a network issue is detected, ensuring seamless fail-over and network resilience.

In this instance, the SRv6 policy automatically selected the path with preference 100. This decision was informed by the model, which detected that the default path with preference 200 was down. Consequently, the system reconfigured the SRv6 TE policy to use the more reliable path with preference 100, ensuring uninterrupted traffic flow and maintaining optimal QoS. This dynamic path selection exemplifies the model's capability to adapt TE strategies autonomously, reinforcing the robustness and responsiveness of AI-integrated 5G network slicing. AI model outputs and accuracy: In the model's evaluation, Figure 18 shows the AI-driven model's programming of SIDs across different packet parameters, validating the ability of the trained model to program SIDs in alignment with predicted slice types (eMBB, uRLLC and mMTC). The logistic regression model achieved an accuracy of 0.89 (89%) under conditions of additional noise, underscoring its robustness and adaptability.

• Predicted Slice Type Distribution The AI model automatically classifies incoming traffic into one of the predefined slices (eMBB, uRLLC or mMTC, as introduced in Section 5), using a probability function derived from the multinomial logistic regression formulation described in Equation (1). Each traffic sample is represented as a feature vector, $${\mathrm{X = }}[ {{{{\mathrm{X}}}_1},\,{{{\mathrm{X}}}_2},\,{{{\mathrm{X}}}_3} \in {{{\mathrm{R}}}_3}} ]$$

• comprising the QCI value, packet delay budget and packet loss rate.

[IMAGE OMITTED. SEE PDF]

The model computes a linear score for each class based on the learned weights and biases, as shown in Equation (2). These scores are transformed into a probability distribution via the softmax function. The slice type with the highest predicted probability is then selected.

• For eMBB, which typically tolerates higher delays but requires higher throughput, the model tends to favour samples with moderate QCI values and larger delay budgets, resulting in a higher softmax probability for class 1.
• For URLLC, which requires ultra-low latency and reliability, the model identifies traffic with low QCI values and strict delay/loss constraints, resulting in a higher score for class 2.
• For mMTC, which supports massive device connectivity with high tolerance for delay and packet loss, the classifier assigns higher probabilities to samples with high QCI values and relaxed performance needs, corresponding to class 3.

The regularised loss function in Equation (3) helps prevent overfitting by penalising large weights, allowing the model to generalise well across diverse traffic patterns. As shown in Figure 19, the resulting slice type distribution reflects the model's capacity to capture the unique characteristics of each 5G slice and classify traffic accordingly with high accuracy and consistency.

• Performance in predicted slice types: During validation, the AI model demonstrated its proficiency across all slices, accurately predicting approximately:
• 40,000 instances for eMBB,
• 35,000 instances for uRLLC and
• 25,000 instances for mMTC.

[IMAGE OMITTED. SEE PDF]

These results are not only reflective of the model's statistical performance but also underline its ability to mathematically discriminate between classes based on feature-driven score differentials and learned softmax thresholds. The predicted distribution showcases the model's capacity to extract and respond to patterns inherent in the distinct traffic profiles of each slice type. This underscores the robustness and scalability of the classifier, as it adapts to dynamic traffic inputs while maintaining high decision confidence, akin to an expert system capable of self-adjusting in real time.

Figures 20 and 21 illustrate the congestion alerts generated by the model when the traffic exceeds predefined thresholds, such as a potential congestion alert in uRLLC with 25,251 occurrences. Figure 21 also shows the detection of a link-down event, with the model responding by automatically updating SRv6 TE policies to maintain connectivity.

[IMAGE OMITTED. SEE PDF]

[IMAGE OMITTED. SEE PDF]

In Figures 22 and 23, the accuracy and loss metrics of the LSTM model are displayed, revealing an accuracy of 0.98 (98%), with consistent performance across the training epoch. The model's minimal loss value across epochs illustrates the stability and learning efficiency of the LSTM model.

[IMAGE OMITTED. SEE PDF]

[IMAGE OMITTED. SEE PDF]

Figure 23 presents ROC curve, showcasing the model's ability to distinguish between slice types with AUC scores of:

• 0.97 for Slice 0,
• 0.99 for Slice 1, and
• 0.99 for Slice 2.

The ROC curve and AUC ROC curve shown in Figure 23 provides a visual assessment of the classification model's performance across multiple slice types. Each curve on the ROC plot corresponds to one slice type (eMBB, uRLLC or mMTC), with the true positive rate (TPR) plotted against the false positive rate (FPR). A higher TPR indicates that the model accurately identifies the correct slice of types, while a lower FPR shows it avoids incorrect classifications. AUC quantifies the model's discrimination ability. High AUC values, especially those close to 1.0 (as achieved by the model), indicate a strong performance in distinguishing between slice types. For instance, an AUC of 0.99 for Slice 2 underscores the model's exceptional capability in correctly identifying this slice type, minimising both false positives and false negatives. The ROC and AUC provide critical metrics to validate the model's reliability in real-world applications, where precise slice type classification is essential for maintaining QoS across varying service requirements.

Dynamic accuracy with alpha variation (Figures 24 and 25) illustrates the effect of altering the regularisation parameter, alpha, on the accuracy of both logistic regression and LSTM models. Starting with an alpha of 0, accuracy for both models consistently exceeded 0.9, indicating a high level of resilience and strong initial performance. As alpha increases, accuracy remains stable, staying above 0.6 even at higher alpha values like 5000. This stability highlights the models’ robustness under diverse conditions, emphasising their adaptability to real-world fluctuations.

[IMAGE OMITTED. SEE PDF]

Figure 24 focused on logistic regression, shows the accuracy trend as alpha values rise, directly influencing regularisation. Lower alpha values (e.g., 0.1) help the model avoid overfitting, striking a balance between complexity and generalisation, resulting in a more versatile model. Interestingly, the logistic regression model maintains a high accuracy (above 65%) across a broad alpha range, underscoring its resilience to different levels of regularisation and network data variability. The LSTM model, represented in Figure 25 demonstrates robustness, adapting well across a spectrum of alpha values. Even at higher alpha values, its accuracy shows minimal decline, reinforcing its capability to handle dynamic network conditions effectively.

[IMAGE OMITTED. SEE PDF]

This adaptability ensures that, despite fluctuating alpha values and network conditions, both models maintain reliable accuracy, allowing them to classify slice types consistently and accurately. The observed stability across alpha variations highlights the models’ readiness for deployment in a live 5G environment. Their ability to sustain high-quality predictions under dynamic conditions is essential for ensuring stable, uninterrupted network performance. This reliability, even as network traffic, latency and user demand shift, underpins the robustness of the AI models, supporting continuous, adaptive optimisation of network resources.

Deployment and Real-Time Adaptation With API Framework

After training, the models were deployed into an API framework interfaced with the network router. For instance, Figure 26a illustrates the automatic configuration of SRv6 SIDs by the API for a uRLLC slice, while Figure 26b shows real-time detection and response to link-down events. Upon identifying a link failure, the API promptly reconfigures the SRv6 TE policy to ensure service continuity by redirecting traffic from R10 to R11 through an alternative path comprising R1, R2, R5, R6 and R8.

[IMAGE OMITTED. SEE PDF]

Feature Importance Analysis

The feature importance analysis illustrated in Figure 27 highlights the relative influence of specific network parameters on the AI-driven model's classification of slice types. Among the evaluated features:

• Packet delay budget (latency): This parameter has a substantial negative impact, indicating its strong influence on classification outcomes.

  • High latency values likely degrade the classification score, emphasising the critical role of low latency in applications requiring fast response times, such as uRLLC.

• Packet loss rate (reliability): Demonstrates a significant positive influence, underscoring its importance in slice assignments.

  • High reliability is essential, particularly for eMBB applications that rely on sustained data flow with minimal packet loss.

• QCI (quality of service class identifier) input 6: Although less influential than the other two features, QCI still contributes positively to the model's predictive capabilities.

  • It aids in distinguishing among slice types based on QoS levels.

[IMAGE OMITTED. SEE PDF]

This analysis confirms that latency and reliability are key indicators in slice classification, aligning with the performance demands of diverse 5G applications. The model effectively prioritises these metrics, enabling precise and adaptive network management.

Comparative Analysis: AI Model Versus Traditional Techniques

The AI-driven model, particularly logistic regression and LSTM, achieves an accuracy above 95% at alpha 0.1 and maintains an accuracy above 65% while increasing alpha to the maximum, with the ROC analysis showing AUC values up to 0.99, indicating exceptional precision in distinguishing between slice types (eMBB, uRLLC and mMTC). By comparison, traditional techniques, such as static MPLS configurations or manual SRv6 TE policies, lack the predictive metrics and adaptability to real-time network changes. Additionally, KNN, SVM and random forest models, while evaluated, achieved significantly lower accuracy levels (around 60%), further affirming the effectiveness of our selected AI models. Figure 28 presents a comparison of all models evaluated, illustrating that logistic regression and LSTM models consistently outperform KNN, SVM and random forest in terms of accuracy and adaptability to dynamic network demands.

[IMAGE OMITTED. SEE PDF]

Traditional techniques, which generally rely on pre-set routing tables, lack such predictive metrics, offering limited insight into classification performance and a less effective approach to handling real-time demands on QoS and resource prioritisation. The AI model's high ROC and AUC scores affirm its reliability in maintaining QoS, even as user demand and traffic complexity increase.

• Scalability and Responsiveness

In traditional network models, configurations are often static, with predefined routing and TE policies that limit the network's ability to adapt dynamically. The AI framework, however, provides real-time reconfiguration of SRv6 SIDs and dynamic SRv6 TE policy updates in response to real-time traffic changes, network failures and user demands. This capability enables the model to proactively route traffic through optimised paths, minimising latency and maximising throughput without manual intervention. For example, the AI model's automatic reprogramming of SIDs and rerouting through TE policies during network failures keeps the network operational without interruption. In contrast, traditional MPLS and manual SRv6 approaches may struggle to maintain service continuity, often requiring manual reconfiguration and causing delays in recovery from disruptions. The LSTM model within the AI framework forecasts network behaviour, enabling pre-emptive resource adjustments based on anticipated demand. This contrasts with traditional TE methods that often rely on reactive measures, leading to inefficiencies in load management and congestion control.

• Efficiency in Resource Allocation and QoS Maintenance Traditional techniques, such as MPLS-based TE or standard SD-WAN routing, allocate resources based on static policies, which often fail to account for real-time QoS variations and specific application needs. By contrast, the AI model utilises logistic regression for immediate slice classification and LSTM for predictive resource management, allowing the network to optimise resource allocation in line with current and forecasted demands. This approach ensures that each application type (eMBB, uRLLC and mMTC) receives the appropriate QoS, which is particularly beneficial for applications with stringent requirements, like ultrareliable low latency (uRLLC) or high bandwidth (eMBB) services. The adaptive API-based framework introduced in this research further reinforces this capability, as it monitors for congestion and dynamically adjusts the network's TE policies. Traditional models typically lack this level of automation and adaptability, which can lead to suboptimal resource distribution, increased packet loss and reduced overall network efficiency. By dynamically assigning SIDs and automatically updating SRv6 policies, the AI model reduces packet rerouting delays, enhances service reliability and improves the end-user experience, particularly during peak network loads.
• Flexibility and Adaptation to Real-World Variability

The AI model's dynamic accuracy across different alpha values signifies its flexibility and resilience in varying network conditions, a quality that traditional methods lack. While MPLS and basic SRv6 techniques are often rigid and rely on manual adjustments to maintain network stability, the AI model maintains high accuracy (above 89%) even as the alpha parameter, representing regularisation, increases. This allows it to adapt to diverse traffic patterns, workload fluctuations and potential network disruptions, ensuring stable performance across different scenarios. In summary, the AI-driven approach introduces a scalable, adaptive and intelligent network management solution that surpasses traditional models in predictive power, resource efficiency and real-time adaptability. The ability to autonomously configure, monitor and optimise SRv6 TE policies based on AI predictions ensures a resilient, high-performance network capable of meeting the dynamic demands of modern 5G applications. This positions the AI-driven model as a significant advancement over conventional techniques, establishing a new standard in efficient, QoS-focused 5G network slicing. The next section highlights the significant contributions of our research, comparing it with existing literature and emphasising the technical innovations introduced at the core routing and orchestration layers.

Significant Contribution of This Research

Compared to previous literature reviews, this enhanced model design addresses several issues and differs in the simulation and project design approach. This research focuses on deploying a modified slicing model, particularly at the core network layer, which encompasses routing and switching technologies. SR and SRv6 were implemented on routers, alongside routing protocols (IS-IS, BGP), to facilitate parallel traffic distribution. Additionally, SIDs were automatically assigned to each interface by integrating with an SDN controller and predicting slice types using the multinomial logistic regression model. The LSTM model's prediction capabilities, including anticipating link failures, traffic volume and congestion events, as well as automatically adjusting SRv6 TE policies upon link failure, were successfully tested. This feature was achieved by implementing an API application capable of continuously communicating with the router to check link statuses. When a failure is detected, the system can dynamically push a new SRv6 TE policy to reroute traffic, ensuring uninterrupted connectivity and maintaining QoS. However, while the framework's performance in controlled conditions shows promise, there are limitations to consider in deployment. The approach relies on real-time prediction and adjustment of network configurations, which could encounter scalability challenges in large-scale deployments or environments with diverse network topologies. Furthermore, while automated SRv6 TE policy reconfiguration is highly beneficial for dynamic network conditions, issues related to compatibility with legacy infrastructure, network device diversity and the computational demands of real-time monitoring and adjustments should be addressed in future work. Additionally, training the models on real-world traffic data could introduce variability, as the model's performance may depend on the quality and volume of data used. Unlike previous studies that employed ML models like SVM and KNN for network slicing, our research proposes an integrated approach using multinomial logistic regression and LSTM models. This combination enhances both the predictive accuracy and adaptability of network slice classification and management within a 5G framework. By automating SRv6 SID assignment and traffic routing, this study builds upon traditional techniques and paves the way for future advancements in the management and optimisation of 5G networks. While these results mark significant progress, further research is needed to optimise real-world deployment, address scalability concerns and integrate the solution with diverse network infrastructures. The practical application of this framework in large-scale 5G networks requires overcoming challenges related to deployment, standardisation and continuous monitoring to ensure its effectiveness and adaptability. Building upon these contributions and technical innovations, the final section offers a comprehensive summary of our findings, outlines the practical implications of our AI-driven approach and discusses avenues for future enhancement and large-scale deployment.

Conclusion and Future Work

This research highlights the transformative potential of AI-driven models in advancing 5G network slicing by enabling automated, precise SRv6 SID assignment and real-time adaptation of network configurations. By integrating logistic regression and LSTM networks within a dynamically adaptive API framework, our approach replaces static configurations with intelligent, responsive control, significantly improving QoS and traffic continuity. The system achieved classification accuracy up to 99% at an alpha value of 0.1 and remained robust above 65% at higher values, while effectively distinguishing up to 40,000 eMBB slices and separating uRLLC and mMTC types. High ROC AUC scores (up to 0.99) further validate the model's strong performance. However, the AI-driven solution presented in this study introduces higher development and operational costs compared to traditional, rule-based approaches. The increased complexity of implementing advanced AI models, specialised expertise and computational requirements contribute to the higher initial development costs. Furthermore, the ongoing operational costs, including continuous model maintenance, API upkeep and real-time network performance monitoring, must be considered. Despite these higher costs, the long-term benefits of enhanced network adaptability, reduced manual intervention and more efficient resource utilisation justify the investment. This framework paves the way for intelligent traffic routing, load balancing, and self-optimisation in core IP networks. Future research directions include:

• Real-world deployment: Testing AI-driven slicing in live 5G networks to validate scalability and practical performance.
• Model extension: Exploring advanced models like deep reinforcement learning to enhance adaptability and decision-making.
• Interoperability and standardisation: Ensuring seamless integration across heterogeneous networks to foster industry adoption.
• Edge integration and energy efficiency: Leveraging edge computing for low-latency decisions and supporting sustainable, green networking.
• User experience and security: Enhancing user satisfaction and safeguarding network data with robust AI-based security mechanisms.
• Network device diversity: Addressing the challenges posed by diverse network devices in terms of compatibility, management and integration.
• Computational demands: Tackling the computational overhead required for real-time monitoring and continuous adjustments in dynamic network conditions.

In conclusion, while the AI-driven framework involves higher costs, it offers a significant leap forward in network management capabilities, laying the foundation for more adaptive, efficient and secure 5G infrastructures. This study encourages continued innovation in AI-driven network management, offering a roadmap for next-generation applications and networks that can dynamically meet evolving demands.

Author Contributions

Zeina Boufakhreddine: conceptualisation, methodology, AI model development, simulation and writing – original draft. Alain Nohra: network implementation, support in AI model integration, visualisation and writing – review and editing. Gaby Abou Haidar: supervision, project administration and writing – review and editing. Roger Achkar: technical validation, resource coordination and technical review. Michel Owayjan: advisory on AI integration strategy, review oversight and final approval of the manuscript.

Acknowledgements

This work was supported by the Faculty of Engineering and Computer Science at the American University of Science and Technology (AUST), Beirut. The authors would like to express their sincere gratitude to Dr. Gaby Abou Haidar for his valuable supervision throughout the course of this research. Special thanks are also extended to Dr. Roger Achkar and Mr. Michel Owayjan for their insightful reviews and for providing the technical resources and guidance necessary to complete this study.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting Information

Additional supporting information is not included in the online version of this article. However, access to the simulation environment, trained AI models or source code used in this study can be made available upon reasonable request to the corresponding author.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

References

M. Gramaglia, V. Sciancalepore, F. J. Fernandez‐Maestro, R. Perez, P. Serrano, and A. Banchs, “Experimenting With SRv6: A Tunneling Protocol Supporting Network Slicing in 5G and Beyond,” in IEEE 25th International Workshop on Computer Aided Modeling and Design of Communication Links and Networks (IEEE, 2020), 1–6.

R. Gupta and R. Misra, “Machine Learning‐Based Slice Allocation Algorithms in 5G Networks,” in 2019 International Conference on Advances in Computing, Communication and Control (IEEE, 2019), 1–4.

M. Abidi, H. Alkhalefah, K. Moiduddin, et al., “Optimal 5G Network Slicing Using Machine Learning and Deep Learning Concepts,” Computer Standards and Interfaces 76 (2021): 103518, https://doi.org/10.1016/j.csi.2021.103518.

A. Thantharate and B. Reddy, “DeepSlice: A Deep Learning Approach Towards an Efficient and Reliable Network Slicing in 5G Networks,” in 2019 IEEE 10th Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (IEEE, 2019), 762–767.

L. V. Le, B. S. P. Lin, L. P. Tung, and D. Sinh, “SDN/NFV, Machine Learning, and Big Data Driven Network Slicing for 5G,” in IEEE 5G World Forum (IEEE, 2018), 20–25.

V. Sciancalepore, C. P. Xie, A. Banchs, and M. Gramaglia, “RL‐NSB: Reinforcement Learning‐Based 5G Network Slice Broker,” IEEE/ACM Transactions on Networking 27 (2019): 1543–1557.

S. Messaoud, B. Dammak, A. Sassi, and M. Messaoud, “Slicing Optimization Based on Machine Learning Tool for Industrial IoT 4.0,” in Proceedings of the 2021 IEEE International Conference on Design & Test of Integrated Micro & Nano‐Systems (IEEE, 2021), 1–5.

S. Singh, M. Salim, J. Cha, Y. Pan, and J. Park, “Machine Learning‐Based Network Sub‐Slicing Framework in a Sustainable 5G Environment,” Sustainability 12, no. 15 (2020): 6250, https://doi.org/10.3390/su12156250.

Q. Miao, “Dynamic Reliability Scheme for SRv6‐Based SFC Orchestration,” in Proceedings of the 2023 5th International Conference on Frontiers Technology of Information and Computer (IEEE, 2023), 1245–1248.

P. Hao and F. Yang, “SRv6 Strategy Based on Network State and Behavior Prediction in 5G Network,” in 2023 9th International Conference on Computer and Communications (IEEE, 2023), 1044–1048.

L. J. Kim and S. Kim, “A Solution for Establishing Traffic Routes Per Application in 5G‐Advanced and 6G Mobile Networks,” in 2023 14th International Conference on Information and Communication Technology Convergence (IEEE, 2023), 1384–1386.

H. Rashid and S. Jeong, “Deep Learning‐Based Network Slice Recognition,” IEEE Communications Magazine 59 (2023): 297–299, https://doi.org/10.1109/ICUFN57995.2023.10199606.

P. Subedi, A. Alsadoon, and P. Prasad, “Network Slicing: A Next Generation 5G Perspective,” EURASIP Journal on Wireless Communications and Networking 2021, no. 102 (2021): 1–26.

R. Rokui, D. Dufresne, J. Hesselbach, and C. Choi, “A Standards‐Based, Model‐Driven Solution for 5G Transport Slice Automation and Assurance,” in Proceedings of the 2020 6th IEEE Conference on Network Softwarization (IEEE, 2020), 106–113.

D. Sattar and A. Matrawy, “Optimal Slice Allocation in 5G Core Networks,” IEEE Networking Letters 1, no. 2 (2019): 48–51, https://doi.org/10.1109/LNET.2019.2908351.

N. M. Vo, T. Le, T. Nguyen, and P. Pham, “Slicing the Edge: Resource Allocation for RAN Network Slicing,” IEEE Wireless Communications Letters 7, no. 6 (2018): 970–973, https://doi.org/10.1109/LWC.2018.2849161.

U. Kaippallimalil, S. Salim, M. Moustafa, and J. Jha, “Traffic Engineered Transport for 5G Networks,” in Traffic Engineered Transport for 5G Networks (IEEE, 2019), 1–6.

Y. Tian, Z. Wang, X. Yin, et al., “Traffic Engineering in Partially Deployed Segment Routing over IPv6 Network With Deep Reinforcement Learning,” IEEE/ACM Transactions on Networking 28, no. 4 (2020): 1573–1586, https://doi.org/10.1109/TNET.2020.2987866.