Revisiting Topological Interference Management: A Learning-to-Code on Graphs Perspective
Zhiwei Shan, Xinping Yi, Han Yu, Chung-Shou Liao, Shi Jin
TL;DR
This paper tackles topological interference management (TIM) by reframing IA as a vector-assignment problem and introducing learning-to-code on graphs (LCG). The approach uses graph neural networks to model topology-aware policies and reinforcement learning to iteratively assign IA beamforming vectors, yielding both known one-to-one IA schemes and novel subspace IA solutions, especially in SIMO settings. Empirical results show that LCG recovers optimal solutions for the majority of network topologies (often aligning with MAIS bounds) and provides fast online inference, with strong generalization and transferability across graph sizes. The work demonstrates the potential of learning-based, graph-structured coding design to automate TIM solutions and motivate further exploration of scalable IA schemes in complex networks.
Abstract
The advance of topological interference management (TIM) has been one of the driving forces of recent developments in network information theory. However, state-of-the-art coding schemes for TIM are usually handcrafted for specific families of network topologies, relying critically on experts' domain knowledge and sophisticated treatments. The lack of systematic and automatic generation of solutions inevitably restricts their potential wider applications to wireless communication systems, due to the limited generalizability of coding schemes to wider network configurations. To address such an issue, this work makes the first attempt to advocate revisiting topological interference alignment (IA) from a novel learning-to-code perspective. Specifically, we recast the one-to-one and subspace IA conditions as vector assignment policies and propose a unifying learning-to-code on graphs (LCG) framework by leveraging graph neural networks (GNNs) for capturing topological structures and reinforcement learning (RL) for decision-making of IA beamforming vector assignment. Interestingly, the proposed LCG framework is capable of recovering known one-to-one scalar/vector IA solutions for a significantly wider range of network topologies, and more remarkably of discovering new subspace IA coding schemes for multiple-antenna cases that are challenging to be handcrafted. The extensive experiments demonstrate that the LCG framework is an effective way to automatically produce systematic coding solutions to the TIM instances with arbitrary network topologies, and at the same time, the underlying learning algorithm is efficient with respect to online inference time and possesses excellent generalizability and transferability for practical deployment.