Multi-Agent Reinforcement Learning In AIoT For Dynamic Resource Allocation And Optimization

Due to the rapid rise in the use of artificial intelligence in things (AIoT), the dynamic resource allocation problem is now more complex than ever. In this research paper, a new dynamic resource allocation framework based on multi-agent reinforcement learning (MARL) for AIoT systems is described. A CTDE architecture based on improved MAPPO (IMAPPO) is utilized, which optimizes AoI for action spaces with both discrete and continuous variables. In simulation tests, the framework has achieved 99% successful task completion with 18 MEC servers available, resulting in a minimal probability of failure of 0.01% at 30 dBm. Furthermore, comparative studies show that MARL has better performance than conventional deep Q network (DQN) and proximal policy optimization (PPO) algorithms by 22.01% and 8.26%, respectively. Moreover, a maximum cumulative reward value of 62,306.58 and 99.98% accuracy were obtained, marking a 6.9% improvement over other MARL models.