Abstract
Transmit power control (PC) will become increasingly crucial in alleviating interference as the densification of the wireless networks continues towards 6G. However, the practicality of most PC methods suffers from high complexity, including the sensing and signalling overhead needed to obtain channel state information. In a highly dense scenario such as the deployment of short-range cells installed within production entities, termed in-factory subnetworks (InF-S), sensing and signalling overhead become a major limitation. In this paper, we represent the InF-S as a graph and resort to graph neural networks for solving the power control problem. We present four graph-attribution methods requiring different degrees of channel information corresponding to different levels of sensing and signalling overhead and study the complexity and performance tradeoffs of the resulting power control graph neural network (PCGNN)
algorithms. We then propose a PCGNN method with scalable sensing and signalling graph attribution
which can meet the stringent outage target while maximizing the global performance by 10% relative to
fixed power control. We further verified the size generalizability and robustness of the PCGNN methods
to different network settings.
algorithms. We then propose a PCGNN method with scalable sensing and signalling graph attribution
which can meet the stringent outage target while maximizing the global performance by 10% relative to
fixed power control. We further verified the size generalizability and robustness of the PCGNN methods
to different network settings.
Original language | English |
---|---|
Journal | IEEE Open Journal of the Communications Society |
Volume | 5 |
Pages (from-to) | 3120-3135 |
Number of pages | 16 |
ISSN | 2644-125X |
DOIs | |
Publication status | Published - 1 May 2024 |
Keywords
- Channel State Information
- Complexity theory
- Graph neural networks
- Interference
- Optimization
- Partial Channel Information
- Power control
- Robot sensing systems
- Sensors
- Subnetworks
- Unsupervised Learning