A practical distributed learning approach for NORA in relay-assisted ALOHA networks

• A fully distributed reinforcement learning framework is proposed for relay-assisted NOMA in slotted ALOHA networks under practical IoT constraints, without information exchange between devices or centralized coordination. • Theoretical analysis derives the achievable transmission success rate for the proposed learning-based random access scheme. • Experimental evaluation of greedy and ϵ-greedy based access strategies illustrates the trade-off between exploration and exploitation and the impact on learning convergence. • A double-estimation based access schemes are introduced to accelerate convergence and enhance learning stability. This paper investigates a multi-hop random access ALOHA network that supports multiple devices communicating with a destination through a relay over multiple channels. By integrating non-orthogonal random access (NORA) into the ALOHA framework, each device can randomly select different channel-slot pairs and transmit data packets with distinct power levels, thereby reducing collisions across two-hop links. To enable distributed adaptation of channel-power selection, we propose a multi-agent learning framework based on multi-armed bandit (MAB) algorithms, allowing devices to independently learn optimal access strategies that maximize successful transmission rates over relay-assisted channels. We first analyze the average rewards of random access actions as the baseline performance when the channel selection probabilities of devices are known to each other. In the practical distributed setting, where users act independently without coordination, we develop four MAB-based NORA algorithms, including double-estimation integrated greedy and non-greedy methods. Simulation results demonstrate that the proposed approaches outperform benchmark schemes in success rate and convergence speed. Furthermore, the use of double estimation effectively mitigates reward fluctuations, leading to more stable learning. Notably, when the two-hop links exhibit symmetric average channel conditions, the proposed RL-aided NORA schemes with greedy selection improve access success rates compared with asymmetric configurations.