Multi-objective reinforcement learning scheduling optimization algorithm assisted by multilayer graph attention network cascaded graph transformer

In the current hybrid edge and cloud computing network, efficient heterogeneous resource scheduling is crucial for data security, compliance and business growth. However, Kubernetes' default scheduling strategy has some problems, such as limited flexibility, coarse granularity and insufficient dynamic adjustment, especially in complex topology environments. To solve these problems, this paper proposes a multi-objective reinforcement learning scheduling optimization algorithm based on cascaded graph transformer of multi-layer graph attention network (GAT). Firstly, a fully connected weighted undirected graph is constructed to represent heterogeneous resources and their topological relationships, and the resource nodes and edge features are multi-dimensionally encoded. The local graph features are extracted by multi-layer GAT using multi-head attention mechanism, and the global graph structure is further modeled by graph transformer to generate a highdimensional embedded representation with deep fusion of local and global features. These embeddings are input into the reinforcement learning strategy network, and the optimization objectives (such as resource utilization, load balancing and task completion time) are adjusted through the reward mechanism to generate an adaptive scheduling strategy. Experimental results show that the algorithm is superior to the traditional Kubernetes scheduling algorithm, KCSS and GNN strategy in terms of resource utilization, pod deployment time, and the maximum deployment. It can effectively manage complex topology dependencies, achieve fine-grained scheduling of heterogeneous resources, and show excellent performance in load balancing and resource efficiency.