14th - 15th November 2024 / Venue : Saadiyat Rotana Resorts, Abu Dhabi, UAE.
Title: 5G Network Automation Using Local Large Language Models and Retrieval- Augmented Generation
Institution: Sharif University of Technology, Tehran, Iran
Authors: Ahmadreza Mejlesara, Ali Mejlesi, Ali Mamaghani, Alireza Shokrani, Babak Khalaj
With the global transition to 5G technology, private networks have become increasingly important but often require extensive networking and software expertise to set up. At the same time, advancements in artificial intelligence, especially large language models (LLMs), are revolutionizing communication infrastructures by simplifying complex technical tasks. Our demo, titled "5G Network Automation Using Local Large Language Models and Retrieval-Augmented Generation (RAG)," presents a novel approach to automating 5G network management using locally deployed LLMs. By leveraging RAG, our system enhances the accuracy and efficiency of network configurations, making private 5G networks accessible to users without deep technical knowledge. Running LLMs on local devices emphasizes privacy and cost-efficiency by eliminating the need to transmit sensitive data over external cloud-based APIs. The demonstration features a locally deployed, lightweight LLaMA-3 8b Q-4b (LLaMA-3 8 billion instruct 4 bit quantized) LLM enhanced with Retrieval-Augmented Generation (RAG) to automate the configuration and management of private 5G networks. Unlike traditional cloud-based LLMs that require continuous API access raising concerns about privacy, cost, and scalability this solution processes all data locally, enhancing data privacy and reducing operational costs. Model quantization reduces the resource requirements, allowing the LLM to run efficiently on edge devices like laptops equipped with NVIDIA 3060 GPUs.
Title: Leveraging Fine-Tuned Retrieval-Augmented Generation with Long-Context Support: For 3GPP Standards
Institution: Khalifa University, 6G Research Centre, Department of Computer Science, Khalifa University, Abu Dhabi, UAE
Authors: Nouf Alabbasi , Omar Erak ,Omar Alhussein and Ismail Lotfi, Amr Hussein, Sami Muhaidat, Merouane Debbah
Retrieval-Augmented Generation (RAG) augments text generation with information retrieval, enabling models to produce more accurate and contextually aware responses. Furthermore, integrating RAG into telecommunication systems involves deploying LLM frameworks on user equipment and edge devices, a process which presents a significant challenge due to the involved computational intensity of LLMs, both in terms of training and inference costs. This challenge highlights the appeal of small language models (SLMs). These language models (LMs) offer computational and storage efficiency while maintaining adequate performance, suggesting their suitability for deployment on edge devices and possible enablement of on-edge artificial intelligence (AI).
In this demo we would demonstrate a carefully developed Phi-2 based fine-tuned RAG system to serve as an oracle for communication networks. Our RAG system leverages a forward-looking semantic chunking (or parsing) strategy that adaptively determines breakpoints between sentences based on embedding similarity. This approach enables the system to effectively process documents with diverse formatting. In the 3GPP documents, a query can often relate to multiple similar contexts, as discussions and paragraphs on related topics may appear in various sections or be phrased similarly. Therefore, we utilize a re-ranking algorithm to further rank the retrieved chunks based on their relevance to the input query.
Title: Intelligent Digital Twin for 6G Networks
Institution: BubbleRAN and EURECOM, France
Authors: CHATZISTEFANIDIS Ilias, Andrea Leone, Navid Nikaein, Alireza Mohammadi, Mikel Irazabal.
This demonstration presents an intelligent digital twin for 6G networks, utilizing cutting- edge technologies such as cloud-native 5G, intent-based networking, and Generative AI. We showcase the pathway to building network intelligent based on digital twin networks that will drive the evolution towards 6G.
Our digital twin is designed atop an intent-based 5G platform featuring cloud-native controllers. It integrates key layers and interfaces between the physical network, the digital twin, and the applications/business stakeholders. Additionally, we implement Large Language Models (LLMs) to create autonomous agents capable of utilizing data from both the physical and digital networks to enable distributed, data-driven network operations.
The demo emphasizes how the digital twin evolves in real time in response to changes in the physical network and explores what-if analysis techniques to identify optimal actions within the twin network, which can then be enforced on the physical network.
Overall, this demo illustrates how operators can develop detailed digital twins to enhance strategic decision-making and efficiency in their operational networks. The design and prototype offer a foundation for future work and can extend beyond connectivity, supporting emerging applications like Collaborative Communication, the Metaverse, Extended Reality (XR), and more.
Title: Next generation field test system for 6G related KPI’s, and innovation in Channel Sounding for 6G.
Institution: Anritsu EMEA GmBH
Authors: Amish Lad, Borrill, Jonathan and Bordin, Marco
In 6G there is a new approach to KPI's where the emphasis is more to higher level 'user perceived' performance, as well as expecting more demanding targets for latency, jitter, per- user throughput, etc. So this new generation of test system is aimed on measuring these new types of KPI. The demo system easily generates 'heat maps' of all of the KPI's, and links the actual user KPI's to network events and L3 messages. It includes a fully virtualised network side component, so that it can easily and flexibly deployed into any cloud/virtual environment where the network, MEC, or service is being hosted. In this demo, the network side component will be an AWS EC2 instance, that will be deployed simultaneously into AWS Wavelength edge locations and into AWS Central Zone, to enable benchmarking of MEC capabilities. For next generation networking applications, there is a need for very high accuracy latency measurements, using real data traffic flows, that can only be supported with this type of solution. For standards bodies and industry forums the focus is on making and visualising real field measurements of next generation KPI's, this demo system is supporting on-going efforts to define and set suitable targets for these KPI values. For the channel sounding, we will present an innovative VNA based approach using the ME7869A PhaseLync system. This enables a unique and easy testing simultaneously in FR1/FR2/FR3 bands, with compact form factor and easily automated control. We will show results from indoor factory measurements, outdoor multipath conditions, and from evaluating FR3 Extreme Large Antenna Arrays (ELAA) propagation effects especally in near field conditions.