Slicing Resource Allocation via Transfer Learning aided Multi-Agent Deep Reinforcement Learning with Domain Similarity Analysis

Network slicing technology enables the system to support variant network services based on the common physical infrastructure. Facing the increasingly spread network scale, an efficient inter-slice resource allocation solution becomes significantly important for faster deployment. In this paper, we proposed a transfer learning (TL) aided multi agent deep reinforcement learning (MADRL) approach with domain similarity analysis for slicing resource allocation. To tackle the inter-cell interference, we introduce a coordination scheme with information sharing to MADRL method. We design a feature-based domain similarity measurement method to address the problem of "from whom to transfer" for inter-agent TL. Furthermore, we proposed three transfer schemes with knowledge from DRL model and instances. We evaluate the proposed solutions in a system-level simulator. The numerical shows that TL aided DRL approach outperforms in terms of convergence rate and performance. while the training process in target domains with selected source domain by domain similarity analysis provides better performance at the early stage and requires less time cost for policy finetuning.