Constraint-Aware Sector Carrier Assignment in Cloud-RAN Using Heterogeneous Graph Neural Networks

Sector-Carrier (SC) assignment is a critical optimization problem in Cloud Radio Access Networks (Cloud-RAN), where achieving optimal SC-to-server mapping is essential for building scalable and high-performance next-generation wireless systems. Prior studies have shown the promise of deep reinforcement learning (DRL) for this task; however, two key limitations remain. First, existing approaches fail to exploit structural relationships among SCs; relationships often known a priori and expressible as soft or hard design constraints. Second, single-agent DRL methods face scalability challenges due to the exponential growth of the action space, leading to slow convergence. To address these limitations, we propose a multi-agent reinforcement learning (MARL) framework augmented with a heterogeneous graph neural network (GNN). The GNN models complex, predefined inter-SC relationships, enabling agents to reason jointly over two types of pairwise requirements: resiliency and mobility (modeled as soft constraints), while ensuring server utilization remains within hard capacity limits. By distributing decision-making across multiple agents and leveraging the GNN to encode structural priors, the proposed framework effectively decomposes the large action space and scales efficiently to larger deployments. Experimental results show that the proposed method reduces convergence steps by 50% and achieves up to 66% faster execution time compared to single-agent baselines using homogeneous graph models, with only a 65% increase in GPU memory usage.