Multi-agent access control in NB-IoT networks based on deep reinforcement learning

This paper studies mechanisms for managing the random access procedure (Random Access Channel, RACH) in networks of narrowband Internet of Things (NB-IoT), where a high number of devices leads to channel congestion and a significant increase in collisions. The subject of the research is the dynamic allocation of RACH time slots based on deep reinforcement learning algorithms. Particular attention is paid to multi-agent architecture, where each agent specializes in a certain aspect of management (proactive detection of load peaks, coordination of RACH slots, traffic differentiation, spatial balancing and energy saving). The authors analyze in detail the causes of RACH congestion and compare traditional static schemes and single-agent RL approaches. The proposed system is implemented in the NS-3 environment using modern Deep Reinforcement Learning methods (DQN, A2C, DDPG). The scientific novelty lies in the decomposition of mass access tasks between several agents, which provides deeper adaptation to the load and reduces the likelihood of collisions. As a result of the experiments, a significant reduction in collisions was achieved compared to the single-agent method, a reduction in connection delays and an increase in the energy efficiency of devices. Prospects for further research include experimental validation of the developed concept on physical hardware.