Fine-Grained Load Balancing with Multi-Agent Reinforcement Learning for Self-Organizing Networks

Abstract

Reinforcement learning (RL) has gained attention as a practical alternative to traditional approaches for mobility load balancing (MLB) in self-organizing networks (SON). However, most previous RL-based MLB schemes have focused on the centralized optimization, which may not be practical in real-world mobile networks. Moreover, the existing coarse-grained control has hampered the performance of optimization. In this paper, we propose a fine-grained load balancing scheme called FineBalancer, based on multi-agent reinforcement learning (MARL) that utilizes joint optimization with finer control of transmit power. We formulate a Markov decision process problem to maximize the average network throughput and employ the multi-agent deep deterministic policy gradient (MADDPG) algorithm to learn the optimal solution to the formulated problem. Extensive simulation results show that FineBalancer can improve the performance compared to state-of-the-art MLB schemes, achieving up to 37.41% better throughput with faster convergence time.