International audience

With the rapid advancement of telecommunications, 6G networks are expected to become more intelligent and capable of making autonomous decisions. Artificial Intelligence (AI) will play a crucial role in achieving this, particularly through the use of Large Language Models (LLMs). These models are increasingly being adopted for networking tasks due to their advanced capabilities in coding, reasoning, and language processing. LLMs have significant potential to support the development of autonomous networks by reducing or even eliminating the need for human intervention. However, LLMs are computationally expensive, which necessitates their shared use across different 6 G applications, i.e., a single LLM might be required to perform multiple tasks within a 6 G network. To this end, routing tasks to the appropriate LLMs presents several challenges: (i) the arrival time of tasks is unpredictable, (ii) tasks must meet specific deadlines, which are part of the ServiceLevel Objectives (SLOs), and (iii) each LLM may perform better on different types of tasks, leading to varying task scores. In this paper, we propose a Deep Reinforcement Learning (DRL) approach for routing tasks to a set of LLMs (task scheduling). Our goal is to maximize task scores while ensuring their deadlines are met. Evaluations conducted under real-world conditions show that our DRL-based approach outperforms traditional methods like Round-Robin (RR) and random scheduling.

With the high demands of 6 G networking services in terms of Quality of Service (QoS), managing these networks requires intelligent next-generation Operations Support Systems (OSSs). According to standardization bodies such as ETSI and 3GPP, OSS must support end-to-end, cross-domain management across all 6 G domains. They are making significant efforts to standardize Application Programming Interfaces (APIs) to enable Intent-Based Networking (IBN), which simplifies network management by allowing users to express their intentions in a declarative manner. However, these systems remain complex for users with limited domain knowledge who need to interact with these standardized APIs. Moreover, adding new functionalities to OSS often requires users to learn new API endpoints and structures, which can be time-consuming. To address these challenges, we propose enabling natural language interaction with OSS by leveraging Large Language Models (LLMs). Our approach offers two key advantages: simplifying user interaction with the system using natural language, and enabling the system to autonomously adapt to new API features. Since fulfilling a user's intent may involve multiple low-level API calls, our solution is designed to plan and execute them in a coordinated manner. We employ multi-agent LLMs with a hierarchical planning mechanism, creating a chatbot-like system that processes natural language inputs effectively. Real-world experiments conducted at EURECOM's OSS demonstrated that the proposed approach can efficiently manage all 6 G domains using natural language.

Intent-Based Networking (IBN) is a promising paradigm for next generation networks, enabling automated network management based on user-defined business network requirements (Intents). However, current IBN approaches consider that users require expertise in some formal and technical models (e.g., Network Service Descriptors - NSDs) to define these Intents, necessitating substantial effort. A natural progression of IBN systems is to define Intents using natural language instead of structured models. However, dealing with this becomes challenging due to the unstructured and ambiguous nature of natural language. Fortunately, Large Language Models (LLMs) are becoming very powerful in understanding human language, making them well-suited for this task. This paper proposes an LLM-based Intent translation system that allows users to express Intents in natural language, which the system subsequently converts into NSDs. Moreover, we employ a Human Feedback (HF) loop that enables the system to learn from past experiences. Evaluations conducted at the EURECOM 5G facility [1] confirm the effectiveness of our approach in generating accurate NSDs suitable for deployment on an edge computing cluster.

The Network Data Analytics Function (NWDAF) is a key component of the 5G Core Network (CN) architecture whose role is to generate analytics and insights from the network data to accommodate end users and improve the network performance. NWDAF allows the collection, processing, and analysis of network data to enable a variety of applications, such as User Equipment (UE) mobility analytics and UE abnormal behaviour. Although defined by 3GPP, realizing these applications is still an open problem. To fill this gap: (i) we propose a microservices architecture of NWDAF to plug the 3GPP applications as mi-croservices enabling greater flexibility and scalability of NWDAF; (ii) devise a Machine Learning (ML) algorithm, specifically an LSTM Auto-encoder whose role is to detect abnormal traffic events using real network data extracted from the Milano dataset [1]; (iii) we integrate and test the abnormal traffic detection algorithm in the NWDAF based on OpenAirInterface (OAI) 5G CN and RAN [2]. The experimental results show the ability of NWDAF to collect data from a real 5G CN using 3GPP-compliant interfaces and detect abnormal traffic generated by a real UE using ML.