Multi-Agent Deep Reinforcement Learning for Resource Allocation in 5G and 6G Networks

Multi-Agent Deep Reinforcement Learning (MADRL) has been widely utilized in numerous next-generation wireless network functions, such as resource allocation in both centralized and distributed manners. This paper investigates the performance of centralized MADRL that leverages global knowledge, distributed MADRL that leverages local knowledge, and hybrid MADRL that leverages both local and global knowledge to self-learn a resource allocation strategy in wireless networks of varying sizes and mobility levels. The overall aim of the paper is to devise a resource allocation strategy by learning in a centralized, distributed, or hybrid manner. Experimental results show that, in small environments with low and high mobility, the centralized MADRL is four times better compared to the hybrid MADRL, while the hybrid MADRL is four times better compared to the distributed MADRL in the average quality of experience (QoE). However, in large environments with high mobility, distributed MADRL is three times better compared to the centralized MADRL and five times better compared to the hybrid MADRL in the average QoE. These findings demonstrate that the optimal MADRL approach in resource allocation depends on the resource allocation scheme, environment size, and mobility level. This paper offers insights into the strengths and weaknesses of various MADRL approaches in different scenarios, providing a foundation for further research and optimization in next-generation wireless networks.