Decentralized Routing and Radio Resource Allocation in Wireless Ad Hoc Networks via Graph Reinforcement Learning

In wireless Ad Hoc networks, optimal routing is closely coupled with radio resource allocation. This paper adopts a cross-layer approach and improves the performance of multi-hop transmission by joint routing and power control. A hop-by-hop optimization framework based on deep reinforcement learning (DRL) is proposed for efficient and decentralized decisions on relay nodes and transmit power. For decision at each hop, graphs are used by the agent to flexibly characterize irregular topological and interference relations among neighboring nodes. Then, a novel graph neural network (GNN)-based model architecture is designed as the agent’s policy to exploit these graph relations and handle input/output of fluctuating dimensions. Particularly, information exchange among nodes during optimization strictly adheres to practical limitations on their sensing ability, fitting the distributed configuration of Ad Hoc networks. The learnt policy can quickly establish high-quality paths between sources and destinations, exhibiting strong generalization ability to random network topologies. Numerical results show that our proposed method outweighs both heuristic benchmarks and state-of-the-art DRL counterparts in terms of both throughput and path connection probability.