Research problem:

    Domain-Aware AI for Wireless Communications: Beyond the Black-Box Paradigm


Brief Description:

Artificial intelligence (AI) is transforming all fields in this world. Without doubt, communication researchers should actively embrace these advances. However, many existing works are blindly applying generic AI tools, particularly deep learning, as black-box solutions without understanding their connection to communication principles. Superficial parameter tuning or trial-and-error combinations of AI techniques do not constitute meaningful progress.

Communication systems are built on well-established models, theories, and principles developed over decades. To truly advance the field, we should leverage this rich domain knowledge within the AI framework. Our goal is to integrate, rather than discard, our foundational understanding, creating solutions that respect the structure of communication systems. This approach will lead to more interpretable, efficient, and robust innovations.

In this research thrust, we explored several directions:

• The first one is deep unfolding, which takes an iterative optimization algorithm to guide the design of the neural network architecture. As many wireless resource allocation problems would result in iterative algorithms, this makes the neural network architecture tailor to the problem at hand, thus resulting in easier and faster convergence using less training data.

• The second one is complex-valued neural networks. The signals and channels of wireless communications by nature are complex-valued. But most implementation of AI in wireless systems separate the real and imaginary parts and treat them as independent components so as to fit into the mainstream AI tools. We advocate that complex-valued neural network should be used such that the intricate relationship between the real and imaginary parts of the signals is respected and maintained during the learning and inference processes.

• Wireless networks can be interpreted as graphs. Therefore, graph neural networks (GNNs) are natural representation of wireless networks. However, standard GNN only provides node update mechanism, which is not sufficient in modeling complex wireless resource allocation problems. For example, in cooperative beamforming, the to-be-designed beamformers must be modeled on the edges while the transmit powers are modeled on the nodes. This makes it necessary to extend the GNNs from node-only update to simultaneous edge-node update.

• A prerequisite in many deep learning tools is the consistency of distributions of training and testing data. Even a minor shift in the testing data distribution would result in significant performance degradation. While we can keep training (or fine-tuning) the model if we have new data from the shifted distribution, another challenge arises: the deep learning model tends to forget what it learned long-time ago. This is known as catastrophic forgetting, and is especially relevant in wireless communications, as the wireless environment does not stay the same over time. To mitigate the catastrophic forgetting, we propose to embed update due to new information into the null space of the weighting matrices and delete information corresponding to the smallest singular value of the weighting matrices. It is surprising that while these concepts are basic in linear algebra, they are very effective in keeping the learned information in the neural network.

Related Publications:

Deep unfolding in wireless communications:

1. Qingfeng Lin, Yang Li, Yik-Chung Wu, and Rui Zhang, ``Intelligent Reflecting Surface Aided Activity Detection for Massive Access: Performance Analysis and Learning Approach," IEEE Trans. on Wireless Communications, vol. 23. no. 11, pp. 16935-16949, Nov. 2024.

2. Qingfeng Lin, Yang Li, Wei-Bin Kou, Tsung-Hui Chang, and Yik-Chung Wu, ``Communication-Efficient Activity Detection for Cell-Free Massive MIMO: An Augmented Model-Driven End-to-End Learning Framework," IEEE Trans. on Wireless Communications, vol. 23. no. 10, pp. 12888-12903, Oct. 2024

3. Bokai Xu, Jiayi Zhang, Qingfeng Lin, Huahua Xiao, Yik-Chung Wu, and Bo Ai, ``Deep Unfolding Beamforming and Power Control Designs for Multi-Port Matching Networks," IEEE Trans. on Wireless Communications, vol. 24. no. 2, pp. 1401-1414, Feb. 2025.

4. Zeyi Ren, Qingfeng Lin, Jingreng Lei, Yang Li, and Yik-Chung Wu, ``Mixture of Experts-augmented Deep Unfolding for Activity Detection in IRS-aided Systems," IEEE Wireless Communications Letters, vol. 14, no. 9, pp. 2912-2916, Sept. 2025.

Complex-valued neural networks for wireless communications:

1.

2.

3.

4.

5. Yang Leng, Qingfeng Lin, Long-Yin Yung, Jingreng Lei, Yang Li, and Yik-Chung Wu, ``Unveiling the Power of Complex-Valued Transformers in Wireless Communications," IEEE Transactions on Communications, vol. 74, pp. 612-627, 2026.

6. Jingreng Lei, Yang Li, Long-Yin Yung, Yang Leng, Qingfeng Lin, and Yik-Chung Wu, ``Understanding Complex-Valued Transformer for Modulation Recognition," IEEE Wireless Communications Letters, vol. 13, no. 12, pp. 3523 - 3527, Dec 2024.

Simultaneous edge-node update Graph Neural Network:

7. Yunqi Wang, Yang Li, QingJiang Shi, and Yik-Chung Wu, ``ENGNN: A General Edge-Update Empowered GNN Architecture for Radio Resource Management in Wireless Networks," IEEE Trans. on Wireless Communications, vol. 23, no. 6, pp. 5330-5344, Jun 2024.

Avoiding catastrophic forgetting in changing environment:

8. Zhenrong Liu, Yang Li, Yik-Chung Wu, and Yi Gong, ``Learning to Optimize Resource Allocation in Dynamic Wireless Environments: Embracing the New While Engaging the Old," IEEE Trans. on Wireless Communications, vol. 24. no. 9, pp. 7346-7359, Sep. 2025.