Intelligent Backend Service Scheduling through the Integration of Reinforcement Learning and Meta-Learning

Abstract

This paper addresses the challenges of low scheduling efficiency, insufficient resource utilization, and weak policy transfer capability in backend service systems under dynamic environments. An adaptive service scheduling mechanism is proposed by integrating reinforcement learning and meta-learning. The mechanism models the service system as a Markov decision process. It employs a policy-gradient-based reinforcement learning framework to make resource scheduling decisions. A multi-task meta-learning optimizer is constructed to enhance policy generalization and transfer efficiency. For state modeling, a temporal embedding mechanism is introduced to encode service state sequences. This enhances the policy model's awareness of historical workload patterns. A multi-objective reward function is designed to jointly consider latency, resource utilization, and service stability. This improves the global responsiveness of the scheduling policy to performance requirements. The experimental section includes comparative evaluation and sensitivity analysis. The results demonstrate the superiority of the proposed method in key metrics such as average latency, CPU utilization, and service-level objective (SLO) assurance. The mechanism achieves stable operation under multiple load and task switching scenarios. It ensures efficient resource scheduling and service quality, showing strong decision-making and model transfer capabilities.