Elastic Resource Scheduling Strategy Based on Supervised-Reinforcement Learning

In stream computing, existing deep reinforcement learning-based resource scheduling methods face challenges such as high computational complexity and difficulty in convergence. This paper proposes an elastic resource scheduling method based on a mixed model of supervised learning and deep reinforcement learning, aiming to maximize application throughput while minimizing resource consumption. The method models resource scheduling as Markov Decision Processes (MDP) and employs a deep neural network as the decision model. The network is initialized via supervised learning, and its adaptability is enhanced through reinforcement learning to derive an efficient scheduling strategy. Experiments demonstrate that compared with fully supervised, semi-supervised, and pure deep reinforcement learning models, the resource allocation results of this method are closer to the optimal values. It achieves maximized resource utilization while ensuring throughput.