The rapid advancement in large foundation models is propelling the paradigm shifts across various industries. One significant change is that agents, instead of traditional machines or humans, will be the primary participants in the future production process, which consequently requires a novel AI-native communication system tailored for agent communications. Integrating the ability of large language models (LLMs) with task-oriented semantic communication is a potential approach. However, the output of existing LLM is human language, which is highly constrained and sub-optimal for agent-type communication. In this paper, we innovatively propose a task-oriented agent communication system. Specifically, we leverage the original LLM to learn a specialized machine language represented by token embeddings. Simultaneously, a multi-modal LLM is trained to comprehend the application task and to extract essential implicit information from multi-modal inputs, subsequently expressing it using machine language tokens. This representation is significantly more efficient for transmission over the air interface. Furthermore, to reduce transmission overhead, we introduce a joint token and channel coding (JTCC) scheme that compresses the token sequence by exploiting its sparsity while enhancing robustness against channel noise. Extensive experiments demonstrate that our approach reduces transmission overhead for downstream tasks while enhancing accuracy relative to the SOTA methods.

Existing works on machine learning (ML)-empowered wireless communication primarily focus on monolithic scenarios and single tasks. However, with the blooming growth of communication task classes coupled with various task requirements in future 6G systems, this working pattern is obviously unsustainable. Therefore, identifying a groundbreaking paradigm that enables a universal model to solve multiple tasks in the physical layer within diverse scenarios is crucial for future system evolution. This paper aims to fundamentally address the curse of ML model generalization across diverse scenarios and tasks by unleashing multi-modal feature integration capabilities in future systems. Given the universality of electromagnetic propagation theory, the communication process is determined by the scattering environment, which can be more comprehensively characterized by cross-modal perception, thus providing sufficient information for all communication tasks across varied environments. This fact motivates us to propose a transformative two-stage multi-modal pre-training and downstream task adaptation paradigm. In the pre-training stage, we introduce a multi-modal two-tower model and a corresponding contrastive learning method to integrate the explicit description of the scattering environment and implicit channel state information (CSI) into a universal representation, which encapsulates rich high-level knowledge and can be leveraged for all downstream tasks in different scenarios. Additionally, we present two specially designed model structures to enhance the interaction of communication modalities. In the second stage, based on the frozen pre-trained model, we propose a direct method and a pluggable method for flexible and low-cost task adaptation. Experimental results demonstrate that our proposed approach significantly outperforms benchmarks in both task performance and tuning parameter size for exemplary sub-tasks in unseen scenarios.

CSI compression, which serves as the first-tire use case for AI/ML applications in 3GPP, has gained widespread attention. The 3GPP-defined Type-3 training with CSI encoder and decoder sequentially trained in different sessions has been accorded a prioritized option due to its exceptional capacity for preserving intellectual property. However, current framework incorporating strong constraints on the latent feature, compromises the principle of separate training. In addition, in the context of a single UE served by multiple NW vendors, the adoption of a unified UE encoder accommodating multiple NW vendors is compelling from complexity perspective, yet remains unexplored. This paper proposes a novel separate training framework followed by a UE-side model training scheme, which leverages the vector quantizer (VQ) with a vendor-proprietary codebook to conceal the interpretation between disclosed codewords and latent representations. We further extend this scheme to encompass multi-NW-vendor scenarios, which enables a unified encoder compatible with multiple NW decoders. Simulation results on the system level simulation dataset and real-world over-the-air dataset demonstrate that sequentially trained encoder-decoder can cooperate seamlessly with negligible performance degradation compared to the Type-1 encoder-decoder joint training. Furthermore, the proposed scheme remains insensitive to changes in the number of NW vendors and codebook variations.

The deep learning (DL)-based channel state information (CSI) feedback methods have been intensively explored in recent years. Most existing works are trained offline based on the prestored datasets. However, in real-world deployment, the pretrained model may not fit the field environment due to the wireless environment changes. In a previous study, a supervised learning approach has been introduced to deal with this CSI feedback transferability problem by finetuning the model. Nevertheless, the transmission for the original uncompressed CSI data will cause intense traffic in air interface. In this paper, a novel unsupervised transfer learning framework named Discrepancy-based Domain Adaptation Network (DDA-Net) is proposed to solve this problem. By minimizing the discrepancy between the CSI codeword datasets from the pretrained environment and field environment, the encoder and decoder are finetuned to extract the common features in both environments, so that the DL-based CSI feedback model can also work properly in a drifted environment without transmitting any original uncompressed CSI data. Simulation results show that the DDA approach can augment CSI feedback reconstruction accuracy and combat overfitting problems in the deployment environment. Compared to the existing supervised learning approach, the DDA approach can achieve similar CSI recovery performance without transmitting the original uncompressed CSI data, reducing considerable transmission traffic.

Deep learning (DL) based receiver (DeepRx) has been proven to be able to greatly improve the data transmission performance compared to conventional OFDM (orthogonal frequency-division multiplexing) receivers. In order to accommodate DFT-s-OFDM (discrete Fourier transform-spread OFDM), which is a widely used single carrier waveform for uplink transmission characterized with low peak to average power ratio (PAPR), the internal structure of DeepRx needs to be redesigned. In this paper, we propose a novel deep neural network based receiver, referred to as IDFT-deprecoding embedded deep receiver (iDeepRx) customized for DFT-s-OFDM. By embedding an untrainable functionality of IDFT-deprecoding between two trainable neural network structures, the proposed iDeepRx can support DFT-s-OFDM transmission and mitigate implicit link and hardware-induced channel impairments. We demonstrate a single-layer 100 Gb/s error-free data transmission using iDeepRx over 20 GHz bandwidth at 220 GHz carrier frequency and observe 2 dB+ relative performance gain over the conventional receiver. Moreover, intermediate output of iDeepRx has been extracted with constellation-like patterns for visualization and performance monitoring during training and inference, which helps in making the internal working mechanism of iDeepRx more explainable.

The 6G network enables a subnetwork-wide evolution, resulting in a “network of subnetworks”. However, due to the dynamic mobility of wireless subnetworks, the data transmission of intra-subnetwork and inter-subnetwork will inevitably interfere with each other, which poses a great challenge to radio resource management. Moreover, most existing approaches require the instantaneous channel gain between subnetworks, which are usually difficult to be collected. To tackle these issues, in this paper we propose a novel effective intelligent radio resource management method using multi-agent deep reinforcement learning (MARL), which only needs the sum of received power, named received signal strength indicator (RSSI), on each channel instead of channel gains. However, to directly separate individual interference from RSSI is an almost impossible thing. To this end, we further propose a novel MARL architecture, named GA-Net, which integrates a hard attention layer to model the importance distribution of inter-subnetwork relationships based on RSSI and excludes the impact of unrelated subnetworks, and employs a graph attention network with a multi-head attention layer to exact the features and calculate their weights that will impact individual throughput. Experimental results prove that our proposed framework significantly outperforms both traditional and MARL-based methods in various aspects.

In this paper, we focus on the network deployment of NR-U system coexisting with WiFi system in the same unlicensed spectrum. In NR-U study, detecting the presence of WiFi was proposed as one candidate solution for coexistence fairness. However, it is very challenge to do such inter-RAT signaling detection by current NR-U receiver due to lack of time and frequency synchronization. We propose a dual-functional deep receiver for NR-U system, which is able to conduct both NR-U data receiving and WiFi preamble detection by the same radio frequency (RF) units. Attention mechanism assisted deep learning is used to recognize the signaling pattern of WiFi preamble to better overcome unknown frequency offset and misaligned receive timing. From simulation evaluation, it can be observed that the ML based approach outperforms the legacy correlation and threshold based signal detection method in all SNR regions. Even with the assumption of frequency error and misaligned timing, the ML based algorithm can still achieve larger than 90% detection probability in case of -5dB SNR. Hence, with proposed solution, inter-RAT signaling detection could be a practicable coexistence manner to be utilized in existing unlicensed bands or greenfield 7GHz unlicensed bands.

The time-varying and non-stationary channel characteristics caused by high Doppler effects is a significant challenge for modern wireless communication systems. Aiming at the high-speed scenarios, a constellation image analysis aided channel estimation using a hybrid neural network (NN) structure, named PR-SRNN, is proposed to combat Doppler effects. Only 1 demodulation reference signal (DMRS) pilot, rather than multiple pilots as in previous reports, is needed by the proposed PR-SRNN in high-speed scenarios. The impacts from Doppler effects are analyzed and compensated by an intelligent pattern recognition neural network (PRNN) based on the contaminated constellation images. The output layer of PRNN is then merged with a super resolution NN (SRNN) which preliminarily reconstructs the channel estimation in frequency domain. Adaptive features learning against Doppler effects and synergic compensation against multipath fading are learnt jointly. It is the first time to our knowledge that a pattern recognition NN on constellation images is introduced to the physical layer, functioning as expert-like analysis system. The simulation results based on both 3GPP statistical channel models and ray tracing show that PR-SRNN exhibits robustness against diverse degrees of Doppler effects beyond the pretrained scope. Amongst different framework candidates of super resolution (SR) NNs, residual convolution SRNN with channel attention has been selected regarding its performance superiority in terms of loss and convergence speed. Furthermore, the cross validation between PR-SRNN and our previously proposed SubSRNN which takes in extra semantic information as a 2ndinput proves that PR-SRNN is effective for Dopplers effects without extra side information to be reported.

The advance of deep learning in computer vision has been leveraged for channel estimation in radio receiver since high-resolution full channel response can be reconstructed from raw channel estimate achieved based on sparse pilots. Despite that significant performance gains have been derived by using powerful neural networks (NN) especially in given channel conditions, it is still challenging and crucial to augment NN's robustness in untrained scenarios in terms of user equipments' (UE) locations and velocities. In this paper, we proposed a super-resolution (SR) NN as the backbone structure for high-resolution channel response reconstruction from low-resolution raw estimate based on sparse pilots. On top of that, a specialized sub NN structure is embedded to adaptively combat against vari-ant Doppler-induced diverse phase rotations between orthogonal frequency division multiplexing (OFDM) symbols in time domain. The Sub-NN with semantic information like UEs' velocities as the input can learn multi-path-propagation dependent Doppler decomposition, and compensate the phase rotation accordingly. For the first time, we show how such specially tailored NN with extra information can be used for high-accuracy channel estimation using only 1 demodulation reference signal (DMRS), which is robust to locations and Dopplers that have not even been trained.

Massive Multiple Input Multiple Output (MIMO) at millimeter wave bands is able to boost the system throughput. A key challenge for the hybrid beamforming design in massive MIMO systems is the acquisition of the full channel state information, since the number of radio frequency chains is much smaller than that of the antennas. Conventional methods require a longer measurement time, a large overhead, or costly signal processing efforts. Therefore, we propose an efficient and adaptable deep neural network based low-rank channel recovery scheme for a hybrid array based massive MIMO system. The proposed neural network architecture includes a common feature extraction module and the adaptable recovery module. The feature extraction, built on the convolutional neural network with residual learning functionality, can efficiently learn the essential features from the low-rank measurements. The adaptable key recovery module maps the essential features to the full channel information. The proposed architecture enables an efficient learning procedure and can be easily adapted to different cases. Simulation results are carried out and compared with existing solutions, showing the potential of applying deep learning concepts in millimeter wave massive MIMO systems.