Navigating Boundaries in Quantifying Robustness: A DRL Expedition for Non-Linear Energy Harvesting IoT Networks

This letter investigates the uplink communication of an energy harvesting (EH)-enabled resource-constrained secondary device (RCSD) coexisting with primary devices in a cognitive radio-aided non-orthogonal multi-access (CR-NOMA) network. Assuming a non-linear EH model in practice, the data rate of the RCSD is maximized using deep reinforcement learning (DRL). We first derive the optimal solutions for the parameters of interest including the time-sharing coefficient and transmit power of the RCSD, using convex optimization and then implement the DRL to address a continuous action spaced optimization problem. To comprehensively assess the agent's performance and adaptability, we implement various DRL algorithms and compare them under non-linear EH, which reveals their suitability in various scenarios, aiding in selecting the most effective approach.