Introduction
Distributed Antenna Systems (DAS) have emerged as a pivotal technology for enhancing the capacity and coverage of wireless networks. By distributing Remote Access Units (RAUs) throughout a cell, DAS reduces the average distance between antennas and User Equipments (UEs), thereby mitigating path loss and improving signal quality. However, the performance of DAS is heavily dependent on effective power allocation strategies. Traditionally, maximizing Spectral Efficiency (SE) or Energy Efficiency (EE) requires solving complex non-convex optimization problems. While iterative algorithms like sub-gradient descent provide accurate solutions, their high computational complexity makes them unsuitable for real-time applications in dynamic mobile environments. This paper explores a machine learning-based alternative using the $k$-Nearest Neighbor ($k$-NN) algorithm to achieve near-optimal performance with significantly reduced latency.
System Model and the Power Allocation Problem
The study considers a single-cell DAS environment where $N$ RAUs are strategically placed to serve $K$ UEs. The goal is to determine the optimal transmit power for each RAU-UE link to maximize either the total system SE or the overall EE, subject to individual RAU power constraints.
In traditional frameworks, this involves iterative mathematical optimization. For SE maximization, the problem is often formulated as a sum-rate maximization under power constraints. For EE, the objective is to maximize the ratio of the total data rate to the total power consumption. These methods typically exhibit a computational complexity of $O(n^3)$, which becomes a bottleneck as the number of users and antennas grows.
The Machine Learning Approach: $k$-Nearest Neighbors
To overcome the limitations of iterative solvers, the authors propose a "lookup" approach based on the $k$-NN algorithm. The fundamental premise is spatial similarity: if a new set of user locations is similar to a previously solved scenario, the optimal power allocation schemes should also be similar.
Data Generation and Training
The process begins with the creation of a comprehensive database. Using traditional iterative methods, the researchers generated $10^5$ training samples. Each sample consists of a "feature vector" representing the coordinates of 10 UEs and a "label" representing the corresponding optimal power allocation scheme.
Inference via $k$-NN
When the system encounters a new group of users, it performs the following steps:
- Search: It calculates the Euclidean distance between the current user coordinates and those in the database.
- Selection: It identifies the $k$ most similar historical scenarios (the nearest neighbors).
- Weighting: Instead of a simple average, the system applies a weight function that accounts for the mean and standard deviation of the distances, ensuring that the most relevant historical data has the greatest influence on the predicted power allocation.
This shift reduces the operational complexity from $O(n^3)$ to $O(n)$, enabling near-instantaneous power allocation updates.
Performance Evaluation
The proposed $k$-NN scheme was evaluated against the traditional sub-gradient method across various transmit power levels. The results demonstrate that the machine learning approach serves as a highly accurate approximation.
As shown in the performance curves, the SE achieved by the $k$-NN method closely tracks the "ground truth" provided by the iterative solver. For instance, at a transmit power of 19 dBm, the SE gap was a negligible 0.11 bit/s/Hz. Similar results were observed for Energy Efficiency, where the average deviation was only 0.282 bit/Joule/Hz. These findings suggest that the $k$-NN model captures the underlying physics of the power allocation problem with high fidelity.
Conclusion and Future Directions
This research highlights the potential of machine learning to replace computationally intensive optimization in wireless communications. By leveraging historical data, the $k$-NN approach provides a practical path toward real-time power management in DAS.
Future work could extend this model by incorporating more realistic channel conditions, such as small-scale fading and inter-channel interference. Additionally, exploring more sophisticated architectures like Deep Neural Networks (DNNs) or Support Vector Machines (SVMs) might further enhance accuracy and scalability for larger, more complex network deployments.
Summary updated