A Multi-Agent Risk-Averse Reinforcement Learning Method for Reliability Enhancement in Sub-6GHz/mmWave Mobile Networks

The problem of enhancing the reliability of packet transmissions in a multi-Access Point (AP) Sub-6 GHz/mmWave integrated network is investigated. The goal is to improve system reliability in terms of average Packet Loss Rates (PLR), while enabling more IoT devices to satisfy their individual PLR, under dynamically varying mobile environments and without perfect Channel State Information. For this, we propose a multi-agent framework based on Risk-Averse Averaged Q-Learning (RAQL), where APs share partial information through a Centralized Unit with which they interact, enabling distributed APs to reinforce their learning process through a global reward function with minimal overhead. Numerical evaluations demonstrate that the proposed method outperforms benchmark risk-sensitive learning in terms of average PLR, number of satisfied users, and convergence speed.