With the rapid development of the Internet of Things and cloud computing, modern computing systems demand higher dynamic responsiveness and multi-objective optimization in CPU scheduling. Traditional rule-based scheduling algorithms struggle to adapt to heterogeneous tasks, dynamic workloads, and complex application scenarios, resulting in system performance bottlenecks. To address these challenges, this paper proposes an intelligent CPU scheduling method, ML_Prio, based on Proximal Policy Optimization. The method constructs a multi-dimensional state space encompassing the current process and scheduling queue context, and employs a dual-network architecture consisting of a policy network and a value network to achieve dynamic priority scheduling decisions. A curriculum learning strategy is introduced during training to progressively optimize for short task responsiveness, tail task handling, and mixed workloads, significantly enhancing the model’s generalization ability. Experimental results demonstrate that ML_Prio outperforms traditional algorithms such as RR, CFS, and MLQ in throughput, turnaround time, and response time metrics, especially under non-uniform loads and large-scale task scenarios. This work provides an effective solution to improve scheduling performance and resource utilization efficiency in modern computing environments.