Model-Based Reinforcement Learning With Kernels for Resource Allocation in RAN Slices

Network slicing is a key feature of 5G and beyond networks, allowing the deployment separate logical networks (network slices), sharing common underlying physical infrastructure, characterized by distinct descriptors behaviors. The dynamic allocation network resources among coexisting slices should address challenging trade-off: to use efficiently while assigning each slice sufficient meet its service level agreement (SLA). We consider time-frequency from new perspective: design control algorithm capable learning over operating network, keeping SLA violation rate under an acceptable during process. For this purpose, traditional model-free reinforcement (RL) methods present several drawbacks: low sample efficiency, extensive exploration policy space, inability discriminate between conflicting objectives, causing inefficient and/or frequent violations To overcome these limitations, we propose model-based RL approach built upon novel modeling strategy that comprises kernel-based classifier self-assessment mechanism. In numerical experiments, our proposal, referred as RL, clearly outperforms state-of-the-art algorithms in terms fulfillment, resource computational overhead.