Knowledge-Driven Deep Learning Paradigms for Wireless Network Optimization in 6G

TL;DR

The paper tackles the challenge of optimizing large-scale 6G wireless networks where purely model-based methods struggle with complexity and data-driven DL suffers from data hunger and limited interpretability. It proposes knowledge-driven DL as a principled paradigm that injects communication-domain knowledge into neural networks, delivering interpretable, data-efficient, and fast-inference solutions. A holistic closed-loop framework (knowledge source, representation, integration, application) and a taxonomy of knowledge integration approaches—knowledge-assisted, knowledge-fused, and knowledge-embedded DL—are presented, alongside a case study of a WMMSE-unrolled GNN for D2D resource management that achieves millisecond-level inference and superior scalability. The work highlights practical significance for 6G by enabling intelligent, constraint-aware optimization and identifies open issues in nonlinear constraints, theoretical guarantees, and knowledge selection and aggregation.

Abstract

In the sixth-generation (6G) networks, newly emerging diversified services of massive users in dynamic network environments are required to be satisfied by multi-dimensional heterogeneous resources. The resulting large-scale complicated network optimization problems are beyond the capability of model-based theoretical methods due to the overwhelming computational complexity and the long processing time. Although with fast online inference and universal approximation ability, data-driven deep learning (DL) heavily relies on abundant training data and lacks interpretability. To address these issues, a new paradigm called knowledge-driven DL has emerged, aiming to integrate proven domain knowledge into the construction of neural networks, thereby exploiting the strengths of both methods. This article provides a systematic review of knowledge-driven DL in wireless networks. Specifically, a holistic framework of knowledge-driven DL in wireless networks is proposed, where knowledge sources, knowledge representation, knowledge integration and knowledge application are forming as a closed loop. Then, a detailed taxonomy of knowledge integration approaches, including knowledge-assisted, knowledge-fused, and knowledge-embedded DL, is presented. Several open issues for future research are also discussed. The insights offered in this article provide a basic principle for the design of network optimization that incorporates communication-specific domain knowledge and DL, facilitating the realization of intelligent 6G networks.

Figures (5)

• Figure 1: The pipeline of knowledge-driven DL. Compared with data-driven DL, in knowledge-driven DL, prior knowledge with reasonable representations is integrated into neural networks to select the training data, devise the neural network model and customize the learning algorithms.
• Figure 2: The holistic framework of knowledge-driven DL for wireless network optimization. Within this framework, knowledge source in wireless networks, communication-specific domain knowledge with various representations, knowledge integration approaches in neural networks and knowledge application in wireless networks are formed as a closed loop to continuously refine the latest knowledge and drive future wireless network optimization.
• Figure 3: Three Main Knowledge Integration Approaches and their Corresponding Integrated Knowledge, Knowledge-driven Components in Neural Networks, Knowledge Representations, Potential Applications and Typical Examples in Wireless Networks.
• Figure 4: The proposed knowledge-driven DL approach for resource management in D2D networks. Both the graph-structured communication topology in D2D networks and the classical WMMSE algorithm for the sum rate maximization problem are integrated into neural networks, inspiring the proposed WMMSE algorithm unrolled GNN. In this figure, $\textbf{u}$, $\textbf{w}$ and $\textbf{v}$ are three variable blocks in the classical WMMSE algorithm and their scalar forms are variables of the $n$-th transceiver pair in D2D networks.
• Figure 5: The scalability comparison of the proposed WUGNN and other two approaches.