Multiobjective Constraints‐Aware Adaptive Deep Actor–Critic Reinforcement Learning for Spectrum Management and Congestion Control for Traffic Reduction in D2D Communication

ABSTRACT In 5G networks, device‐to‐device (D2D) communication is considered for its potential to reduce latency. Conventional methods do not manage congestion issues efficiently, as they result in performance degradation issues. This research work proposes an Adaptive Deep Actor–Critic Reinforcement Learning (ADACRL) architecture to limit data traffic as well as optimally achieve spectrum management and reduce congestion issues in D2D communication. The developed ADACRL model exploits the best properties of actor–critic systems to acquire dynamic and environment‐sensitive spectrum access strategies. In order to enhance the performance of the ADACRL model, the tuning process is done over the parameters of the ADACRL model using Fused Artificial Lemming with Orangutan Optimization (FALOO), which is developed with the integration of Artificial Lemming Algorithm (ALA) and Orangutan Optimization Algorithm (OOA). The objective function considered for the tuning process is active cluster, latency, throughput, spectrum efficiency, and congestion factor to attain effective performance. Also, there is a congestion‐aware reward mechanism that tries to make the agent follow the least traffic routing path and yield better throughput performance. Simulations confirmed the higher performance of the proposed ADACRL over the current benchmarks, where a wide range of enhancements in spectrum efficiency, low communication overhead, and effective congestion control are observed.