Computation Rate Maximization for SCMA-Aided Edge Computing in IoT Networks: A Multi-Agent Reinforcement Learning Approach

Integrating sparse code multiple access (SCMA) and mobile edge computing (MEC) into the Internet of Things (IoT) networks can enable efficient connectivity and timely computation for resource-limited IoT users. This paper studies the computation rate maximization problem under task deadline constraints in dynamic SCMA-MEC networks. Specifically, we propose a predictive deep Q-network for SCMA resource allocation and computation offloading (PQ-RACO) algorithm for single-cell scenarios, where IoT devices use long short-term memory (LSTM) networks to predict the states and actions of other agents. However, the PQ-RACO algorithm is not scalable for increasing numbers of IoT devices. To address this issue, an improved multi-agent deep Q-network for SCMA resource allocation and computation offloading algorithm (MQ-RACO) is proposed for multi-cell scenarios. The algorithm is a centralized training and decentralized execution (CTDE) multi-agent reinforcement learning (MARL) algorithm with explicit rewards, which is tailored to the special structure of joint rewards. Simulation results demonstrate that the proposed algorithm outperforms several state-of-the-art MARL algorithms and other benchmark schemes in terms of convergence speed and computation rate.