Resilient AI-Based for Post-5G Radio Access Networks

1. General context

The upcoming post-5G networks are expected to deliver ubiquitous broadband access for an enormous number of connected devices. It constitutes highly complicated environment including cutting-edge technologies, such as massive IoT, edge computing, and network slicing. 6G networks are supposed to accommodate extremely diverse use cases ranging from ULLR services that need ultra-low latency and reliability, such as autonomous driving and remote surgery, to high-throughput services that prioritize large data transfer, and broadcast service systems focusing on distributing media to a wide audience. This entails a tangled and confusing ecosystem requiring considerable and costly management processes where conventional solutions have reached their limits. To solve that major issue, AI has been suggested as alternative [1].

2. Goals and Challenges

In earlier network generations, humans can still maintain control over the management process. Actually, leading network operators are still performing human-operated management in order to manage their transport infrastructure. This strategy is approaching its limits and is unlikely to effectively handle the growing dynamic, scale, and complexity of future networks. In such a case, adopting AI-based management processes offers an appealing and, to some extent, inevitable alternative. That is why, we aim in this work to develop an AI dedicated to perform customized cells management in order to improve radio cell resiliency and efficiency. This task includes the following identified challenges:

Challenge 1: Data quality and availability

Building robust and unbiased AI-based solutions rely heavily on large volumes of high-quality data. This requires understanding the nuts and bolts of real post-5G cellular network in order to assess the available data and explore its underlying features and relationships. A compromise should be found that balances the data collection cost with its relevance.

Challenge 2: AI model selection and training

Since AI is not a one-size-fits-all solution, a considerable work should be made in order to find the suitable AI for each management operation, e.g., resource allocation, fault recovery and load balancing. This modeling phase should account for data characteristics and operational constraints. In our case, we will mainly focus on radio cell failure management using Multi-Agent Deep Reinforcement Learning (MARL), which is presumed as one of the most promising techniques for such a context.

Challenge 3: AI deployment and post-deployment

The AI deployment over radio cell is a crucial aspect that should be addressed. The deployment strategy should account for AI prerequisite in addition to several operational criteria such as latency, resource availability, and scalability. Moreover, it is required to foresee in which way the AI should evolve in order to follow varying network and traffic conditions.