QIPPO/CA: A Quantized Communication-Efficient MARL Framework for Fully Distributed Channel Access in Next-Generation Wireless Networks

Abstract

Next-generation wireless networks (NGWNs) demand highly efficient, low-latency channel access schemes to support emerging applications. To capture the distributed nature of participating stations, we formulate the problem as a decentralized partially observable Markov decision process (Dec-POMDP). Building on this formulation, we propose QIPPO/CA, a communication-efficient distributed channel-access method for collision avoidance based on the listen-before-talk (LBT) protocol. QIPPO/CA leverages independent proximal policy optimization (IPPO) within a decentralized training and decentralized execution (DTDE) paradigm. It is further extended with a federated learning (FL)-inspired quantized gradient update, enabling efficient coordination among distributed stations without exchanging local information. Extensive simulations demonstrate that QIPPO/CA: 1) consistently outperforms legacy random access methods such as CSMA/CA; 2) achieves the performance of centralized training frameworks; 3) remains robust under dynamic channel sizes and traffic loads; and 4) substantially reduces communication overhead during agent training. These results highlight QIPPO/CA as a promising framework for scalable, efficient, and standard-compatible distributed channel access (DCA) in NGWNs.