Application of Large Models in Communications Field

With the continuous development of 5G networks, the scale of mobile communication networks is expanding, and the co-exsitence of multiple frequencies and multiple systems is making networks more complicated. In the 5G era, alongside the Internet of everything (IoE), various new services are emerging, and there are differentiated ToC and ToB service requirements. A large number of terminals with varying types and capabilities are accessing the networks. Under the traditional network construction and operation mode, the level of network automation is low, the network lifecycle is disconnected at different stages, and the manpower cost associated with "planning, construction, maintenance and optimization" is high. To meet these challenges, automation and intelligence have become critical development directions for future network operation and maintenance.

In 2019, TM Forum introduced the concept of the "autonomous network", which is similar to auto-driving. The autonomous network is divided into six levels (L0–L5) as shown in Fig. 1. At L0-L3, users' requirements, objectives, and constraints can be achieved through policy-driven operations, with requirements transferred through existing interfaces. At L4-L5, the system can adaptively determine its own behaviors through intent-driven interaction, reducing the need for human adaptation. This capability will translate into service flexibility through the introduction of new, customized services without human intervention.

The ultimate goal of autonomous networks is to achieve "full autonomous networks", a consensus that has been widely acknowledged in the communications industry. At present, the industry’s autonomous level is generally at L3 (conditional autonomous networks).

In the L3 phase, network operations are automatically performed through pre-defined and pre-tested scripts and policies, greatly improving network O&M efficiency. However, key decision-making points still rely on human experience and require human participation. As the autonomous network evolves into the L4 and L5 phases, the obvious change is to minimize the dependence of network operation on human resources. This requires the network to independently formulate optimal solution and automatically execute tasks. According to the definition in the Autonomous Networks Whitepaper released by TM Forum, an autonomous network is a system managed in accordance with specified objectives or expectations. These objectives or expectations are called intentions, including requirements, objectives, and constraints abstracted in a simplified manner. In brief, the intent is "what" rather than "how". This means the user should tell the system what to do, instead of how to do it, greatly reducing the complexity of work.

Large models, represented by ChatGPT, have unprecedentedly promoted the evolution of intent-driven capabilities of autonomous networks and have led the research and practice within the current industry.

Building Large Model Capabilities

A large model is rooted in machine learning. Machine learning is a branch of artificial intelligence, which refers to the process of providing data to a program to train the program to identify data features without manual intervention. Machine learning covers multiple algorithms and technologies, such as linear regression, support vector machines, and deep learning based on neural networks. A large model is a machine learning model with a large number of parameters (typically 10 billion or more) and a complex network structure. Such a model is usually pre-trained using large-scale training data and requires a large amount of resources for both training and deployment. Key features of a large model include:

• Understanding: Large models well understand human intentions and demonstrate strong instruction-following capabilities.
• Memorization: Large models hold multiple rounds of dialogs without forgetting the contents of previous dialogs.
• Responsiveness: Large models derive comprehensive user concerns from a large number of user interactions, improve task scenarios, and provide responses that closely match user requirements.
• Smoothness and logic: The smoothness of language and the logical rules of large models have surpassed those of most humans.

Since OpenAI released the ChatGPT, large technology enterprises have successively released their own large language models, such as META's LLaMA, Google's PaLM-E, Baidu's ERNIE Bot, and Alibaba's Tongyi Qianwen.

Utilizing the intent recognition capability of large models, the network O&M mode shifts from "How to do" to "What to do", ensuring a fast service experience driven by natural language and free from process interference. For example, in a network fault handling scenario, the function call and API mapping capabilities of large models are used to achieve coordinated orchestration of the structured data models, enabling quick fault detection and alarm association convergence. Based on the output results,  large models perform secondary analysis and summarization to generate a fault summary.

In the closed-loop fault handling process comprising solution recommendation, solution implementation, and quality inspection, historical fault cases are retrieved using the retrieval augmented generation (RAG) capability of large models to generate a handling solution. Next, intent identification and transfer are completed through man-machine interaction, and fault self-healing instructions are executed. Then, interactive quality inspection is performed, utilizing natural language to SQL (NL2SQL) for self-querying alarm status. Finally, the associated knowledge is extracted through the large model to achieve knowledge recycling and model capability iteration.

Based on the Nebula large telecom model, ZTE has developed a wireless network guarantee assistant that empowers all three phases (beforehand, in-process, and afterward). This assistant quickly generates assurance solutions, implements various assurance tasks including perception evaluation, large-screen monitoring, data statistics and analysis, and cross-domain delimitation and positioning, generates review and summary reports, and recycles knowledge into the database, significantly improving network assurance efficiency in major scenarios and reducing O&M manpower. At the 2023 World Internet Conference Wuzhen Summit, ZTE's guarantee assistant, leveraging the large model, innovated practices for wireless network guarantee scenarios with an end-to-end closed-loop approach and provided generative AI dialog interaction to innovate traditional O&M modes, improving operation efficiency and reducing manpower investment by more than 30%.

Technical Challenges Faced by Large Models and Countermeasures

Compared with traditional methods, large models excel in intent identification capabilities, but still face some challenges in the automatic and intelligent network O&M scenario.

• Large model hallucinations: Hallucinations refer to information generated by large models that conflicts with the source or cannot be verified by the available source. These hallucinations may cause large models to generate unexpected output. However, the O&M field has low tolerance for errors, making it crucial to minimize these hallucinations.
• Weak interpretability: Large models are recognized as black-box models with complex neural network structures, resulting in generated content that is not easily interpretable. However, the O&M field requires high interpretability of results.
• Insufficient O&M corpus: The implementation of large models is hampered by a lack of a large O&M corpus. Especially, there is a shortage of both quality and quantity in private corpus.
• Combination with existing tools: It is a challenge to integrate large models with a large number of automatic network O&M tools in the existing network.

To address these challenges, several solutions can be implemented. For example, to mitigate large model hallucinations, we can increase the proportion of explicit knowledge through RAG, utilizing knowledge graphs. To enhance result interpretability, we can employ an “evidence-based” generation strategy. When integrating large models with automatic O&M tools, agents can be used to facilitate these processes.

As a significant technology in the field of artificial intelligence, large models are driving digital transformation across various industries. In the communications field, the natural language processing and intent recognition capabilities provided by large models offer more intelligent and convenient solutions for intent-based autonomous network construction. With the development of technologies and the expansion of application scenarios, large models will play an increasingly important role in the communications field.