Transfer Learning in Coordinated Multi-Agent Deep Reinforcement Learning for Inter-Cell Slicing Resource Partitioning

Network slicing technology in next-generation RAN dedicates to supporting variant requirements of the network services. Correspondingly, slice-aware radio resource scheduler for RAN slicing management needs to partition the radio resource to slices dynamically and efficiently. However, it is difficult to derive an analytical solution for multi-cell scenario due to the complex inter-cell dependencies, inter-slice resource constraints and service specific individual objectives. State-of-the-arts use RL-based methods to derive solutions from dynamic interaction with networks, but suffer from low reproducibility and efficiency. To address these challenges, we proposed a TL aided MADRL approach for multi-cell RAN slicing management problem with inter-cell coordination, which alleviates the problems of inter-cell interference and improves model reproducibility. We analyzed and compared the alternatives of DRL schemes in centralized and distributed manners. We also investigated the performance of TL schemes with different transferred knowledge. The numerical evaluation in a system-level simulator reveals that our solution concludes competitive performance for multi-cell scenarios in comparison with other solutions, especially regarding interference mitigation and model reproducibility.