Subject: Wireless Innovations: Optimization, Deep Learning, and the Future of Connected Systems

Hi Elman,

This newsletter explores the latest advancements in wireless communication, sensing, and signal processing, with a focus on optimization and deep learning techniques. We'll delve into cutting-edge research on mmWave, cell-free massive MIMO, and integrated sensing and communication (ISAC), highlighting several impactful papers that are shaping the future of connected systems.

General Overview

This collection of papers explores diverse challenges and opportunities in wireless communication, sensing, and signal processing, with a notable emphasis on optimization and deep learning techniques. Several works address the complexities of mmWave and cell-free massive MIMO systems. Yan et al. (2024) (Yan et al., 2024) introduce a dynamic AP mode selection scheme for mmWave cell-free massive MIMO-ISAC, optimizing energy consumption and performance by strategically assigning APs as transmitters, receivers, or shutting them down. Similarly, Topal et al. (2024) (Topal et al., 2024) tackle energy efficiency in cell-free massive MIMO with wireless fronthaul, proposing joint antenna activation and power allocation to minimize power consumption while meeting QoS requirements. Liu et al. (2024) (Liu et al., 2024) investigate mobile cell-free mMIMO with mobile APs, leveraging multi-agent reinforcement learning (MARL) for joint mobility and power control, incorporating a graph neural network (GNN) for enhanced agent collaboration and scalability. These contributions collectively advance the understanding and practicality of deploying efficient and robust mmWave and cell-free systems.

Another prominent theme is the integration of sensing and communication (ISAC). Mason and Pegoraro (2024) (Mason & Pegoraro, 2024) utilize deep reinforcement learning to optimize channel sampling patterns in ISAC for improved micro-Doppler spectrogram estimation. Kurt and Guvensen (2024) (Kurt & Guvensen, 2024) address self-interference mitigation in full-duplex MIMO communications, proposing a joint approach of digital cancellation and spatial suppression based on a scatterer map. Benício et al. (2024) (Benício et al., 2024) explore RIS-assisted sensing, employing nested tensor decomposition for joint estimation of target parameters. These works showcase the potential of ISAC for enhanced sensing capabilities within communication systems.

Beyond ISAC, several papers delve into specific communication challenges. Rubio and Pascual-Iserte (2024) (Rubio & Pascual-Iserte, 2024) investigate user grouping and resource allocation in multiuser MIMO systems under SWIPT, optimizing the weighted sum rate while ensuring minimum harvested power. Aravanis et al. (2024) (Aravanis et al., 2024) derive a tractable closed-form approximation of the ergodic rate in Poisson cellular networks, providing insights into network densification strategies. Tang et al. (2024) (Tang et al., 2024) present BiCSI, a binary encoding and fingerprint-based matching algorithm for Wi-Fi indoor positioning, achieving high accuracy with reduced storage requirements. Muñoz et al. (2024) (Muñoz et al., 2024) focus on robust precoding for multi-user visible light communications with quantized channel information, formulating the problem as a second-order cone program.

Finally, applications of signal processing and deep learning extend beyond communication systems. Jang and Meyer (2024) (Jang & Meyer, 2024) propose a new statistical model for waveguide invariant-based range estimation in shallow water. Oshima et al. (2024) (Oshima et al., 2024) demonstrate radar-based measurement of body movements in classroom environments. San-José-Revuelta and Casaseca-de-la-Higuera (2024) (San-José-Revuelta & Casaseca-de-la-Higuera, 2024) introduce a modified flower pollination algorithm for equalization in DS/CDMA systems. Several papers utilize deep learning for diverse applications: Pappula and Anwar (2024) (Pappula & Anwar, 2024) develop an ADHD diagnostic interface based on EEG spectrograms, Kaseb et al. (2024) (Kaseb et al., 2024) propose PINN4PF, a deep learning architecture for power flow analysis, and Monge-Alvarez et al. (2024) (Monge-Alvarez et al., 2024) present a machine hearing system for robust cough detection. These diverse applications highlight the growing influence of deep learning in signal processing and related fields.

Paper Highlights

Adaptive Personalized Over-the-Air Federated Learning with Reflecting Intelligent Surfaces

Adaptive Personalized Over-the-Air Federated Learning with Reflecting Intelligent Surfaces by Jiayu Mao, Aylin Yener https://arxiv.org/abs/2412.03514

Caption: This graph showcases the superior test accuracy of ROAR-Fed with a noiseless downlink compared to a noisy downlink and other baseline algorithms ([28] and [31]) over several iterations. The results demonstrate the effectiveness of the proposed algorithm in mitigating the impact of noisy channels and imperfect CSI in over-the-air federated learning.

Over-the-air federated learning (OTA-FL) offers a compelling approach to distributed learning by exploiting the superposition property of wireless channels, allowing simultaneous model updates. However, noisy channels and imperfect channel state information (CSI) present significant hurdles. This paper introduces ROAR-Fed, a novel algorithm utilizing reconfigurable intelligent surfaces (RIS) to enhance OTA-FL performance by jointly optimizing communication, computation, and learning resources.

ROAR-Fed's adaptive resource allocation strategy is its core strength. The algorithm employs channel inversion for uplink power control and dynamically adjusts the number of local update steps to counteract fading and optimize both learning and communication resources. The non-convex problem of RIS phase shift design is tackled using successive convex approximation (SCA), iteratively solving simplified convex problems to determine optimal RIS configurations, thereby maximizing their effectiveness in improving the wireless environment. Downlink power control is also dynamically managed to mitigate the impact of noisy channels on the received global model's accuracy at edge devices.

A rigorous convergence analysis is provided, considering a general non-convex learning objective and quantifying the impact of imperfect CSI and communication noise on the convergence upper bound. The formula for the convergence upper bound highlights the interconnectedness of optimization error, uplink noise error, local update error, and other error terms related to statistical variations, channel estimation errors, and downlink noise, emphasizing the importance of the cross-layer design and the impact of imperfect CSI.

Experimental results on MNIST and Fashion-MNIST datasets showcase ROAR-Fed's superiority over state-of-the-art baselines, even under challenging conditions. Increasing the number of RIS elements demonstrably improves accuracy and convergence speed. The framework's extension to personalized federated learning (PFL) with personal RISs (PROAR-PFed) further demonstrates significant improvements in personalized task accuracy, highlighting the potential of RIS-assisted OTA-FL for enhancing distributed learning in realistic wireless environments.

Mobile Cell-Free Massive MIMO with Multi-Agent Reinforcement Learning: A Scalable Framework

Mobile Cell-Free Massive MIMO with Multi-Agent Reinforcement Learning: A Scalable Framework by Ziheng Liu, Jiayi Zhang, Yiyang Zhu, Enyu Shi, Bo Ai https://arxiv.org/abs/2412.02581

Caption: This diagram illustrates the architecture of a UAV-assisted cell-free massive MIMO system using multi-agent reinforcement learning (MARL) for joint mobility and power control optimization. The system employs a graph neural network (GNN) for inter-agent communication and a dynamic permutation network for efficient state-action space compression, enabling coordinated resource allocation between mobile access points (APs) and user equipment (UEs). The dashed red lines indicate downlink power control and mobility management for a specific UE, while the dashed black lines represent wireless fronthaul links between the APs and the central processing unit (CPU).

Cell-free massive MIMO, a cornerstone of 6G, leverages distributed access points (APs) to serve all user equipment (UEs) simultaneously, enhancing spectral efficiency and mitigating interference. However, coverage limitations and the need for dynamic power control, particularly with mobile APs like UAVs, present significant challenges. Traditional optimization methods struggle with computational complexity and scalability in these dynamic scenarios. This paper introduces SF-MADDPG, a scalable framework using multi-agent reinforcement learning (MARL) to optimize both mobility and power control in UAV-assisted cell-free massive MIMO.

SF-MADDPG addresses key challenges through several innovative components. A graph neural network (GNN) facilitates inter-agent communication, enabling efficient resource allocation through collaboration. A dynamic permutation network, leveraging permutation invariance and equivariance, compresses the state-action space, mitigating the dimensionality curse in MARL. A directional decoupling architecture with an attention-based intrinsic reward network ensures accurate reward partitioning, reflecting each mobile-AP's contribution to the overall system performance. The downlink achievable spectral efficiency (SE) is a key metric, capturing the effectiveness of the joint optimization.

Numerical results demonstrate SF-MADDPG's significant performance advantage over existing methods, including conventional MARL algorithms and heuristic power control schemes, with substantial improvements in sum SE. It also outperforms centralized MDGNN schemes, highlighting the importance of dynamic mobility optimization. The joint optimization of mobility and power control leads to more efficient agent trajectories and improved interference management. The framework's adaptability is validated across various system configurations, demonstrating consistent performance gains. The integration of GNNs, permutation networks, and directional decoupling architectures proves crucial for enhancing MARL performance in complex wireless systems. The emphasis on distinguishing between entity-uncorrelated and entity-correlated actions through permutation properties is key for reducing complexity and improving convergence.

Learn More by Using Less: Distributed Learning with Energy-Constrained Devices

Learn More by Using Less: Distributed Learning with Energy-Constrained Devices by Roberto Pereira, Cristian J. Vaca-Rubio, Luis Blanco https://arxiv.org/abs/2412.02289

Caption: This boxplot compares the test accuracy of LeanFed and FedAvg with varying client participation rates (100%, 80%, 50%, 20%, 10%) on image classification tasks. LeanFed consistently outperforms FedAvg, especially with lower participation rates, demonstrating its effectiveness in energy-constrained federated learning environments. The wider spread in FedAvg's accuracy reflects the instability caused by device dropout due to energy depletion.

Federated learning (FL) offers a decentralized and privacy-preserving approach to model training, but the varying energy capacities of participating devices present a significant challenge. Energy limitations not only affect model accuracy but also lead to device dropout, impacting convergence. This paper introduces LeanFed, an energy-aware FL framework that optimizes client selection and training workloads on battery-constrained devices.

LeanFed's core innovation lies in its adaptive data usage strategy. By dynamically adjusting the fraction of local data used by each device during training, LeanFed maximizes device participation across communication rounds while preventing premature battery depletion. This approach contrasts with traditional FedAvg, which typically uses the entire local dataset, potentially leading to early dropouts for energy-constrained devices. The fraction of data used is carefully determined to balance training contribution with energy preservation.

The evaluation of LeanFed against FedAvg on CIFAR-10 and CIFAR-100 datasets under various levels of data heterogeneity and device participation rates demonstrates its effectiveness. LeanFed consistently improves model accuracy and stability, particularly in scenarios with high data heterogeneity and limited battery life. The results highlight the detrimental effect of device dropout on FedAvg's performance, with accuracy degrading as devices become inactive due to energy constraints. LeanFed's ability to mitigate dropout and ensure balanced contributions throughout training leads to more robust convergence and higher final accuracy.

Conclusion

This newsletter highlights a convergence of critical themes in wireless systems research. The exploration of mmWave and cell-free massive MIMO architectures, coupled with the innovative application of optimization and deep learning techniques, is paving the way for more efficient and robust networks. The increasing focus on integrated sensing and communication (ISAC) further expands the possibilities of these systems, enabling enhanced sensing capabilities within the communication framework. The presented papers showcase not only advancements in specific areas like resource allocation and interference mitigation but also a broader trend towards intelligent and adaptive systems. The development of algorithms like ROAR-Fed and SF-MADDPG exemplifies the power of cross-layer optimization and multi-agent learning in tackling complex challenges in dynamic wireless environments. Furthermore, the emphasis on energy efficiency in distributed learning, as demonstrated by LeanFed, underscores the growing importance of considering real-world constraints in algorithm design. These advancements collectively contribute to a vision of future wireless systems that are not only more powerful and versatile but also more sustainable and adaptable to the diverse needs of connected devices.