AI4C1 – Neural Networks for Wireless Systems
Wednesday, 4 June 2025, 11:00-12:30, room 1.G
Session Chair: Dani Korpi (Nokia Bell Labs, FI)
Delta Optimization for Event-Based Sampling in Goal-Oriented Communication Systems
Mehmet Hakan Durak (Erzurum Technical University, Turkey); Hirley Alves (University of Oulu, Finland)
Goal-oriented communication (GOC) has emerged as a promising paradigm in modern wireless systems, emphasizing transmitting information relevant to the goal rather than the raw data itself. This study investigates event-based sampling methods, specifically the Send-on-Delta (SOD) and Send-on-Delta with Linear Prediction (SODwLP) algorithms, to explore the trade-off between normalized mean square error (NMSE) and the number of samples (NoS) in resource-constrained scenarios. While SOD achieves fewer transmitted samples compared to SODwLP, it exhibits higher NMSE, making it less favorable in terms of signal reconstruction accuracy. Through a grid search optimization strategy, we demonstrated that SODwLP achieves a lower overall cost at equivalent delta values and consistently selects larger optimal delta thresholds across various weight configurations, balancing communication efficiency and accuracy more effectively. These results highlight the adaptability of SODwLP for GOC systems, where specific priorities such as accuracy or energy efficiency can significantly influence the choice of sampling parameters. Future research could extend these findings by employing advanced optimization techniques and evaluating real-world applications in IoT and 6G networks to enhance the practical relevance of these algorithms.
Uplink OFDM Channel Prediction with Hybrid CNN-LSTM for 6G Non-Terrestrial Networks
Bruno De Filippo, Carla Amatetti and Alessandro Vanelli-Coralli (University of Bologna, Italy)
Wireless communications are typically subject to complex channel dynamics, requiring the transmission of pilot sequences to estimate and equalize such effects and correctly receive information bits. This is especially true in 6G non-terrestrial networks (NTNs) in low Earth orbit, where one end of the communication link orbits around the Earth at several kilometers per second, and a multi-carrier waveform, such as orthogonal frequency division multiplexing (OFDM), is employed. To minimize the pilot overhead, we remove pilot symbols every other OFDM slot and propose a channel predictor to obtain the channel frequency response (CFR) matrix in absence of pilots. The algorithm employs an encoder-decoder convolutional neural network and a long short-term memory layer, along with skip connections, to predict the CFR matrix on the upcoming slot based on the current one. We demonstrate the effectiveness of the proposed predictor through numerical simulations in tapped delay line channel models, highlighting the effective throughput improvement. We further assess the generalization capabilities of the model, showing minimal throughput degradation when testing under different Doppler spreads and in both line of sight (LoS) and non-LoS propagation conditions. Finally, we discuss computational-complexity-related aspects of the lightweight hybrid CNN-LSTM architecture.
Transformer-Aided CSI Prediction for Interference Alignment in MIMO Systems
Samitha Gunarathne, Nurul Huda Mahmood and Matti Latva-aho (University of Oulu, Finland)
Interference alignment (IA) is a degrees-of-freedom optimal interference management technique which approaches the capacity of an interference network at high signal-to-noise ratio (SNR) regime. It works as a cooperative precoding scheme in which different transmitters coordinate their transmissions so that all interference signals at an unintended receiver are confined to the same subspace. Each receiver can recover the desired signal by eliminating the aligned interferences using a suitable receiver design. However, IA requires full channel state information (CSI) of all involved links at all transmitters, which incurs a huge feedback overhead and is not feasible in practice. In this study, we propose a transformer-aided CSI prediction for signaling overhead reduction. Our proposed transformer-based CSI prediction demonstrates superior accuracy compared to the conventional methods. The predicted CSI is utilized to perform IA and compute the precoding matrices in a multi-antenna system, allowing them to efficiently implement IA scheme. Numerical results show that the achievable rate with the proposed transformer-based method is within 98% of the achievable rate with perfect CSI. In comparison, conventional deep learning-based CSI estimation approaches, namely long short-term memory networks and convolutional neural networks, is found to achieve only 90% and 88% of the ideal rate, respectively.
Causal Graph Generation and Validation for Cognitive 6G Networks
Mehmet Karaca (Ericsson Research, Turkey); Jishnu Sadasivan (Ericsson Research India, India); Alexandros Palaios (Ericsson Research, Germany); András Zahemszky (Ericsson Research, Sweden)
In this paper, we investigate causal learning and discovery in 6G networking context towards the ambition of achieving cognitive and autonomous network. This paper introduces a novel two-phase causal graph generation and verification framework tailored for autonomous networks in the 6G era. While existing causal discovery techniques provide valuable insights into network KPIs and actions, they often lack the robustness and adaptability required for real-world deployments. To address these challenges, we propose a structured offline-online approach that enhances both the accuracy and practical utility of causal graphs. In the offline phase, we generate multiple candidate causal graphs by leveraging a combination of state-of-the-art causal discovery methods. This process allows us to systematically explore different causal structures while incorporating domain-specific constraints. In the online phase, we validate and refine these candidate graphs in a working network environment. By observing real network behavior, applying controlled interventions, and analyzing the impact on key network KPIs, we iteratively evaluate the reliability of the generated causal graphs. We implement and test our approach in a realistic network emulator, demonstrating its effectiveness, and the experimental results show that our method improves causal inference accuracy and intervention predictability.
Port Selection for Fluid Antenna Systems via Conditional Generative Adversarial Networks
Mahdi Eskandari, Alister G. Burr and Kanapathippillai Cumanan (University of York, United Kingdom (Great Britain)); Kai Kit Wong (University College London, United Kingdom (Great Britain))
Fluid antenna systems (FAS) present an innovative solution to enhance diversity in compact mobile devices. They dynamically adjust the position of radiating elements (known as ports), optimizing their placement to mitigate signal fades effectively. Despite their potential to enhance the signal-to-noise ratio (SNR), practical implementation faces challenges in port selection due to the vast volume of required channel estimation for all of the ports. In this letter, we propose a novel approach employing conditional generative adversarial networks (cGANs) to streamline port selection processes. This is done by leveraging the correlation among the ports due to their close proximity to generate unobserved channels for unobserved ports. Through extensive simulations, our findings demonstrate significant re- ductions in outage probability with minimal observed ports, showcasing the efficacy of our proposed algorithm.
