Hierarchical Reinforcement Learning with Self-Distillation for Resource Scheduling in Complex Resource Networks

The resource scheduling problem still presents various limitations in traditional methods when dealing with a large number of resource packages. This paper proposes a novel hierarchical reinforcement learning with self-distillation (HRLSD) model to address the resource scheduling problem in complex service networks with a large-scale of resource packages. To reduce the complexity of the policy space, we introduce a student-teacher encoder as a hierarchical reinforcement learning model. The student Q-value and teacher Q-value are obtained from two separate Q-networks using the deep Q-learning method. By distilling the knowledge from the teacher Q-value to the student Q-value, the student encoder learns from the teacher encoder to enhance effectiveness without increasing computational complexity. We evaluate the effectiveness of our proposed method through numerical examples involving large-scale resource packages.