Solving Integrated Access Backhaul Frequency Allocation Problem with Graph-Based Deep Reinforcement Learning

We present a novel deep reinforcement learning (DRL) framework for ultra-dense integrated access and backhaul (IAB) frequency allocation, formulated as a combinatorial optimization problem. Our approach leverages graph neural networks (GNNs) to model the network as a graph optimization task, capturing intricate spatial dependencies. To further enhance learning, we integrate a transformer-based attention mechanism that dynamically prioritizes critical factors such as channel states, interference patterns, and resource allocation states. Unlike conventional time-based DRL optimization methods, our framework employs an iterative optimization approach that allocates one resource per step, thereby effectively addressing the curse of dimensionality in DRL.