RGMSC: Relational Graph Attention Assisted Multi-Agent Soft Actor-Critic for Task Allocation in Heterogeneous Computing Power Networks

As computing power networks evolve towards widearea heterogeneous architectures, and the dependency relationships of massive tasks become complex, on-demand and realtime access to ubiquitous distributed computing power resources is required. However, the distribution of heterogeneous node resources is uneven, and existing centralized allocation methods are difficult to meet the real-time requirements of wide area heterogeneous computing power networks, while existing distributed methods are difficult to effectively integrate dispersed local information to support global collaboration. In order to improve the collaboration and efficiency of task allocation in wide area heterogeneous computing networks, this paper proposes a multi-agent reinforcement learning algorithm RGMSC based on relational graph attention network (RGAT). Firstly, the task scheduling is modeled as a joint optimization problem, and a decentralized partially observable Markov decision process (DecPOMDP) model is constructed to coordinate multidimensional resources such as computing, memory, and storage. Secondly, a collaborative learning framework combining the RGAT and multi-agent soft actor critic (MASAC) is innovatively designed, which dynamically captures resource associations and topological features between agents through RGAT, enhances strategy exploration ability through maximum entropy reinforcement learning, and achieves efficient collaborative decision-making under local observation. The offline training, online distributed inference mode is adopted to solve the high latency of traditional centralized scheduling with the global information loss of distributed scheduling. Experiments have shown that compared to MASAC, deep deterministic policy gradient (DDPG), and multi-agent deep deterministic policy gradient (MADDPG), RGMSC reduces task completion time by $28.6 \%$, energy consumption by $24.7 \%$, and enhances energy efficiency by $32.5 \%$ in different scenarios.