Reinforcement Learning-Based Elastic Scaling Policy Optimization for Microservices

Abstract

This paper focuses on the intelligent optimization of elastic scaling strategies in microservice architectures. It proposes an adaptive resource control method based on reinforcement learning. The method constructs a multi-dimensional state space that incorporates key performance indicators such as CPU utilization, response time, and queue length. A policy network is used to make dynamic decisions on scaling actions. To improve learning efficiency and system responsiveness, the method introduces an advantage function for policy optimization. It also integrates a temporal dependency modeling mechanism to enhance the model's awareness of historical workload patterns. A reward function is designed based on multiple system metrics. It jointly considers resource cost, performance assurance, and system stability, enabling fine-grained guidance of the policy behavior. In the experimental section, a sensitivity analysis is conducted with different discount factors, dynamic workload patterns, and environmental disturbances. The evaluation systematically examines the stability and robustness of the proposed method. Comparative results with mainstream baselines show significant improvements in accuracy, precision, and recall. The results demonstrate strong policy convergence and efficient scheduling, confirming the effectiveness of the proposed approach in microservice elastic control tasks.