"Adaptive Transfer DRL For Resource Allocation in 5G RAN Slicing"

"AN slicing is a key 5G radio access network (RAN) technology for enabling flexible resource allocation to support heterogeneous service requirements. The rapid growth of traffic across diverse vertical services necessitates intelligent resource management mechanisms capable of operating under stochastic and highly dynamic network conditions. Transfer learning has emerged as a promising approach for accelerating convergence and ensuring quality-of-service (QoS) guarantees; however, its effectiveness is often limited by negative transfer between the master and apprentice agents. To address this challenge, the Kullback--Leibler (KL) divergence, $D_{\\mathrm{KL}}$, is employed to regularize successive policy distributions in  the master agent network, thereby enabling stable convergence and reliable weight transfer from the master actor proximal policy optimization (MAPPO) agent to the apprentice. Furthermore, a long short-term memory (LSTM) architecture is integrated to enhance temporal feature extraction and to improve policy learning for both agents. Extensive simulation results demonstrate that the proposed framework achieves up to a significant improvement in data rate, latency, and energy compared to state-of-the-art approaches, including  meta-learning,  multi-agent deep reinforcement learning, and transfer deep reinforcement learning."