AI and Reinforcement Learning Can Help Optimize Massive MIMO Performance

According to Omdia's 2021 ICT Enterprise Insights (ICTEI) Survey, 58% of service providers said they would increase their spend in artificial Intelligence (AI) tools in 2021, and 78% of service providers ranked the use of AI to automate network operations.

July 14, 2021

3 Min Read
AI and Reinforcement Learning Can Help Optimize Massive MIMO Performance

According to Omdia’s 2021 ICT Enterprise Insights (ICTEI) Survey, 58% of service providers said they would increase their spend in artificial Intelligence (AI) tools in 2021, and 78% of service providers ranked the use of AI to automate network operations as a top project for 2021. Plan and optimize network resource allocation, automate and optimize field operations, and predict, prevent and remediate network faults were top ranked use cases in this survey. The survey conducted between July and September 2020 had 419 respondents from the telecom sector.

First introduced for TDD LTE, massive MIMO became widespread with the launch of 5G. This technology brings great benefits in terms of capacity, coverage, and user experience but it is also complex to implement and operate. One of the key principles of massive MIMO is to increase the number of antenna elements and to create multiple narrow signal beams precisely directed at devices, instead of broadcasting the signal in an undifferentiated manner across a wide area, thereby increasing efficiency and reducing interferences.

While the configuration of a traditional passive antenna involves a single beam and is relatively simple, the configuration of a massive MIMO antenna on the contrary is sophisticated. It involves many parameters and variables such as the number of beams, horizontal and sometimes vertical beamforming, antenna tilt, azimuth, aperture, etc. This results in thousands of different possible beam patterns, and available options to consider when trying to optimize the performance.

In addition, this is not sufficient to just configure the antenna at the time of installation, it is an ongoing process as users and traffic distribution change. In short, the massive MIMO configuration needs to be dynamic, continuous, and to happen in real-time. A manual approach is therefore not practical, automation is required.

This is where AI is expected to come to the rescue. However, common AI learning methodologies like supervised and unsupervised learning are not well equipped for the task. Supervised learning takes time and cannot deliver the real-time insights required by massive MIMO. Unsupervised learning is more capable of real-time configuration but can typically be less accurate which is also a limitation for this kind of critical application.

With its CHIME platform, Cellwize suggests adopting a third approach called reinforcement learning. In layman’s terms, reinforcement learning is learning in real-time and through trial and error. The AI agent observes, selects, and takes an action, and then it analyzes the outcome of this action and updates the strategy accordingly. The process is then repeated until the optimal configuration is obtained, which in the case of massive MIMO is the configuration which maximizes performance for the highest number of users at a given time.

Figure 1: (Source: Cellwize) (Source: Cellwize)

An operator will rightly be cautious when hearing about a trial and error approach as any negative impact on the user experience should be avoided in the process. Therefore, the initial implementation must first happen off-network, in a lab environment. The lab simulates the network and how the implementation of a new configuration impacts key performance indicators such as coverage, throughput, spectral efficiency, etc. It also simulates the impact on neighboring cells. The solution only moves into production when a satisfactory level of performance has been achieved.

Massive MIMO, due to its complexity, is the ideal use case to demonstrate how AI, and reinforcement learning, can support network optimization and automation. But there are other mobile network use cases that can take advantage of reinforcement learning, for example the control of simpler traditional antennas electrical tilt to reduce interference.

To learn more about AI reinforcement learning, massive MIMO optimization and Cellwize’s solutions, you can download the whitepaper AI with reinforcement learning: the secret sauce for 5G success.

This blog is sponsored by Cellwize.

Subscribe and receive the latest news from the industry.
Join 62,000+ members. Yes it's completely free.

You May Also Like