Multi-Agent Deep Reinforcement Learning-Based Fairness-Aware Power Allocation in NOMA Systems under Imperfect SIC for 5G Communication Networks

In the invisible rivers of data that flow through our modern world, wireless networks face a constant struggle: how to serve more people with the same limited amount of spectrum. For decades, the standard solution was to divide the available radio frequencies into separate lanes, assigning each user their own exclusive path. This method, known as orthogonal multiple access, works reliably but is inefficient, like a highway where cars are forced to take turns on a single-lane road even when the other lanes are empty. A newer approach, called Non-Orthogonal Multiple Access, or NOMA, changes the rules entirely. Instead of separating users by time or frequency, it allows multiple users to share the exact same frequency at the same time. To make this work, the transmitter sends a single, combined signal where different users are assigned different levels of power. The user with the strongest connection receives a weaker signal, while the user with the weakest connection gets a stronger one. At the receiving end, a process called successive interference cancellation acts like a filter, stripping away the stronger signals to reveal the weaker ones underneath. This technique promises to pack far more data into the same space, a necessity for the fifth-generation, or 5G, networks that are currently connecting billions of devices.

However, the promise of NOMA hinges on a delicate balancing act: deciding exactly how much power to give to each user. If the power is distributed poorly, the users with weak connections might fail to receive their data, while the network as a whole wastes capacity. This problem becomes even harder in the real world, where the equipment is not perfect. The process of stripping away signals is prone to small errors, leaving behind a faint residue of interference that can degrade the connection. Furthermore, the wireless environment is chaotic; signals bounce off buildings and fade unpredictably. Traditional methods of managing this power often rely on fixed rules or complex, centralized calculations that struggle to keep up with these rapid changes. They tend to favor the users with the best connections, leaving those with poor signals behind, or they become too slow to adapt when conditions shift.

To address these challenges, a team of researchers at Kampala International University and other institutions in Nigeria proposed a new way to manage this power distribution using a form of artificial intelligence. Instead of relying on a single central brain to make decisions for the entire network, they designed a system where every user acts as an independent, intelligent agent. This approach, known as Multi-Agent Deep Reinforcement Learning, allows each user to learn from its own local experience. Imagine a group of drivers who, instead of waiting for a traffic control tower to tell them when to move, learn to navigate a busy intersection by watching the cars around them and adjusting their speed based on what works best for the group. In this system, each user observes its own signal quality and the interference it faces, then makes its own decision on how much power to request. These agents are not acting selfishly; they are trained with a shared goal that rewards the entire group for being fair and efficient.

The researchers tested this decentralized system in a simulated environment that mimicked the unpredictable conditions of a real 5G network. They compared their new method against several established techniques, including simple fixed rules, equal power distribution, and more advanced algorithms like genetic algorithms and centralized learning systems. The simulations ran thousands of times to ensure the results were reliable. The findings revealed that while the centralized system could achieve the highest total data speed under perfect conditions, the new multi-agent approach offered a superior balance of performance and reliability. Specifically, the new method significantly reduced the number of times users failed to connect, a metric known as outage probability, keeping it below ten percent even when signal strength was relatively low. It also maintained a high level of fairness, ensuring that users with poor connections were not left behind, achieving a fairness score of 0.96 at high signal strengths.

Perhaps most importantly, the study showed that this new approach is robust against the imperfections found in real-world hardware. When the researchers introduced errors into the signal cancellation process and the measurement of channel conditions, the centralized system's performance dropped more sharply than the decentralized one. Because the multi-agent system relies on local observations rather than a perfect global view, it was less shaken by these inaccuracies. In terms of data accuracy, the new method reduced the rate of errors to less than one in a thousand at high signal strengths, outperforming the other tested methods. The researchers concluded that while centralized intelligence has its place, a network where users learn to cooperate and adapt locally offers a more scalable and resilient solution for the future of wireless communication. This work suggests that the next generation of networks may not need a single, all-knowing controller, but rather a collection of smart, cooperative agents that can navigate the chaos of the wireless world together.