Intelligent Decentralized Multiple Access via Multi- Agent Deep Reinforcement Learning

In this paper, we propose a multi-agent proximal policy optimization (MAPPO) based algorithm, PPO-DMA, with centralized training and decentralized execution (CTDE) framework for the sensing-free decentralized multiple access problem. To maximize the total network throughput while guaranteeing fair access among devices in a distributed manner, a single reward function for each node is introduced. Extensive simulation experiments are conducted to show that the proposed PPO-DMA scheme: 1) significantly outperforms the theoretical upper bound of slotted-Aloha with satisfactory fairness performance among nodes; 2) shows greater stability and robustness compared to independent deep Q network (IDQN) algorithm with decentralized training and decentralized execution (DTDE) framework; 3) can harmoniously coexist with Aloha-based nodes.