Dynamic Resource Configuration for Low-Power IoT Networks: A Multi-Objective Reinforcement Learning Method

Considering grant-free transmissions in low-power IoT networks with unknown time-frequency distribution of interference, we address the problem of Dynamic Resource Configuration (DRC), which amounts to a Markov decision process. Unfortunately, off-the-shelf methods based on single-objective reinforcement learning cannot guarantee energy-efficient transmission, especially when all frequency-domain channels in a time interval are interfered. Therefore, we propose a novel DRC scheme where configuration policies are optimized with a Multi-Objective Reinforcement Learning (MORL) framework. Numerical results show that the average decision error rate achieved by the MORL-based DRC can be even less than 12% of that yielded by the conventional R-learning-based approach.
Comment: Accepted to IEEE Communications Letters