Graph Neural Network-based Continual Learning for Resource Allocation in Dynamic Wireless Environments

Recently, deep learning (DL) has gained significant attention for addressing optimization problems in the field of wireless communication. However, existing methods that train models on a single distribution struggle with adaptability to dynamic wireless environments, leading to the occurrence of the “catastrophic forgetting phenomenon”. These methods also suffer from poor scalability and generalization when the network size expands. To tackle these challenges, we propose a graph neural network-based continual learning (CL-GNN) framework that combines graph-based optimization with adaptive sample selection. Our method maps the wireless optimization problem into the graph neural network (GNN) structure to formulate the graph optimization problem. Then each GNN layer conducts the message passing process after graph modeling. Both homogeneous and heterogeneous GNNs are designed, capable of being implemented independently of training network size. Meanwhile, the framework incorporates a performance-based sample selection mechanism, enabling the GNN model to continuously augment the acquired knowledge in dynamic wireless environments. Furthermore, by integrating model-driven DL with domain expert knowledge, the framework additionally reduces computational complexity while maintaining robust performance across varying user densities and channel distributions. We evaluate the effectiveness of the CL-GNN through comprehensive evaluations of its adaptability, scalability, generalization in power control and beamforming optimization. Simulation results show that the CLGNN further reduces catastrophic forgetting by 10% compared to random sampling, retaining over 95% of performance ratio, which is close to the joint training baseline, in two applications. And compared to state-of-the-art methods such as the power control GNN (PCGNN) method, the heterogeneous GNN method and the beamforming neural network (BNN) method, the CLGNN exhibits 20% higher performance ratio in dynamic wireless environments. The CL-GNN also enables knowledge transfer from small-network training to large-network deployment while maintaining competitive performance. In summary, this work establishes a new benchmark for dynamic wireless network optimization.