CD-MAPPO: Centralized-Decentralized Multi-Agent Proximal Policy Optimization in Multi-Cell Networks

The rapid growth of mobile communication demands has intensified the need for efficient power allocation in multi-cell wireless networks to ensure optimal performance and scalability. This paper addresses power allocation in multi-cell wireless networks, where effective interference management is crucial for optimizing system performance. We formulate a detailed system model encompassing network topology, channel fading, SINR computation, and data rate analysis. To enable distributed yet coordinated power control, we propose Centralized-Decentralized Multi-Agent Proximal Policy Optimization (CD-MAPPO), a Multi-Agent Reinforcement Learning (MARL) approach based on the Centralized Training with Decentralized Execution (CTDE) paradigm. Each base station is modeled as an agent that learns to allocate power using only local observations, while a centralized critic leverages global state information to improve stability and coordination. The proposed CD-MAPPO, built upon the MAPPO algorithm, is well-suited for partially observable multi-agent environments. Simulation results demonstrate that CD-MAPPO achieves significant improvements in system throughput and inter-cell interference mitigation compared to traditional methods such as IPPO and WMMSE. These findings highlight the potential of CD-MAPPO in enabling scalable and adaptive resource management in dynamic wireless networks.