DRL-Based Resource Allocation in NOMA-Aided Industrial IoT Towards Energy Productivity Maximization

This paper addresses the joint optimization problem of computing and communication resource allocation in a non-orthogonal multiple access (NOMA) supported industrial Internet of things (IoT) system. We introduce a decentralized deep reinforcement learning (DRL) based mechanism that collaboratively maximizes energy productivity (EP). Specifically, optimizing EP involves maximizing the amount of successfully processed bits while simultaneously minimizing task processing delay, energy consumption, and task drop ratio. With this goal, we design a comprehensive model that incorporates multiple factors, including NOMA-specific constraints, task queueing, partial processing, and multitasking, where each device can handle multiple concurrent tasks. Next, we propose a dual-decentralized multi-agent proximal policy optimization (MAPPO) based algorithm, where one MAPPO algorithm focuses explicitly on computing resource allocation for task processing at the edge servers, while the other MAPPO algorithm precisely manages communication resource allocation. Additionally, each MAPPO algorithm is equipped with a small residual network (ResNet) and a recurrent neural network (RNN) to effectively capture spatial features from complex channel conditions and the evolving task flow. Through extensive simulations, we validate the effectiveness of our resultant mechanism under various environmental conditions and demonstrate its superiority over the closely related mechanisms in the literature.