AI in telecom research: self-driving cellular networks
CRNA believes in the idea of self-driving networks, and in this blog post we present our work on intelligent control for self-driving Radio Access Networks (RAN) based on Deep Reinforcement Learning.
In our last blog post, we presented our use of AI to classify network outages automatically, but the realisation of self-driving networks does not just require correct classification. In this post, we’ll go deeper into how we leverage the potential of AI to automate network operations.
The advances in deep reinforcement learning have led to outstanding success in various domains. In a recent study, we explored the potential of deep reinforcement learning to improve the performance of radio access networks, which is the most challenging part of cellular networks. We developed an AI agent capable of making autonomous decisions on three key cellular network control actions: adjustment of the base station antenna tilt, initiating performance-triggered handovers to less congested cells (fewer users or less traffic), and optimising the data rate tailored to best suit the specific type of traffic.
Centralized or Distributed Control Architectures for Cellular Networks?
We have presented a novel control framework, Intelligent Control Radio Access Network (ICRAN), for optimizing the utilization of resources while minimizing violations in service-level agreements (SLAs) in a multi-slice RAN. A multi-slice RAN is a single RAN infrastructure that can support multiple networks simultaneously, and each network can serve specific purposes or types of data traffic.
Inspired by the remarkable achievements that Deep Reinforcement Learning (DRL) has shown in solving complex control problems in highly dynamic environments, such as mobile networks, ICRAN comprises two DRL-based architectures: centralized ICRAN-C and distributed ICRAN-D. The first has a centralized controller coordinating all base stations in a network, and the latter has distributed control and coordination between a limited number of base stations. The distributed framework is more feasible as this method generates less telemetry traffic.
The setup of the two frameworks is illustrated below:
Outperforming state-of-the-art methods
In our findings, we see that the proposed ICRAN-C and ICRAN-D algorithms outperformed the state-of-the-art methods for optimizing network resources.
The substantial advantages granted by ICRAN have been confirmed through extensive simulations in the network simulator ns-3. It outperforms other slicing schemes and recent works in terms of resource utilization and Quality of Service assurance. ICRAN is, to the best of our knowledge, the only framework that simultaneously addresses multiple RAN problems. We believe that ICRAN is implementable and can serve as an important point on the way to realizing self-driving mobile networks.
We conclude that AI is essential in future telecom networks, enabling the vision of self-driving networks. This conclusion is based on the demonstration of how self-optimizing of performance in cellular networks based on DRL can be used to outperform heuristics optimization and other competing, traditional algorithms.
To read the full paper, see: A. H. Ahmed and A. Elmokashfi, SimulaMet, "ICRAN: Intelligent Control for Self-Driving RAN Based on Deep Reinforcement Learning," in IEEE Transactions on Network and Service Management, vol. 19, no. 3, pp. 2751-2766, Sept. 2022, doi: 10.1109/TNSM.2022.3191746.
(This blog post is written by Postdoctoral Fellow Azza Ahmed, Research Professor Haakon Bryhni, and Maria Normann.)