DeepDefrag: Spatio-Temporal Defragmentation of Time-Varying Virtual Networks in Computing Power Network based on Model-Assisted Reinforcement Learning

We propose DeepDefrag, a model-assisted reinforcement learning for spatio-temporal defragmentation of time-varying virtual networks in a cross-layer optical network testbed, which realizes the efficient utilization of computing nodes and lightpaths by co-optimizing scheduling and embedding with fragment matching, reduces >13.5% cost of computing power network.