Decentralized and Joint Resource Allocation, Beamforming, and Beamcombining for 5G Networks With Heterogeneous MARL

In this paper, we propose a novel Multi-Agent Reinforcement Learning (MARL) -based paradigm for distributed and joint resource allocation, beamforming (BF), and beam combining of uplink transmissions in 5G networks. The proposed paradigm employs two types of heterogenous agents that learn to perform and optimize different tasks in order to achieve the main objective of the system, as well as the objective of the individual agents. In the proposed paradigm, UEs can be multi-agents that optimize their own resource allocation and BF. In addition to these multi agents (i.e., UEs), the BS is a different type of agent that optimizes the combining of UEs' transmissions. We developed three different implementations of our proposal using three different MARL algorithms: Independent Q Learners (IQL), Multi-Agent Deep Deterministic Policy Gradient (MADDPG), and QTRAN. Various experiments were conducted to validate the usability of our proposal. Our results show that the proposed paradigm can successfully optimize the task of joint resource allocation, beamforming, and combining. Furthermore, we provide a comparative analysis of the three different implementations, highlighting noteworthy insights into the strengths and limitations of fully distributed algorithms, such as IQL, in comparison to algorithms employing the Centralized Training with Decentralized Execution (CTDE) framework, exemplified by QTRAN and MADDPG.