This paper analyzed the characteristic of business flow and divided the business flow into two categories according to the flow distribution: elephant flow and mice flow, combined with the characteristics of the elephant flow’s high delay tolerance and bandwidth demand flexibility, focused on the multi-layer traffic coordination problem in the optical transport network, and the lack of fairness guarantee and multi-path transmission mechanism in the competition of bandwidth resources between elephant flows, realized the coordination of service traffic between the IP layer and the optical layer, proposed a periodic resource scheduling mechanism and a cross-layer global routing and bandwidth scheduling algorithm based on fairness and multi-path transmission under the unified control “IP over OTN” transport network architecture based on the SDN architecture and the OpenFlow protocol.

Based on the good performance of deep reinforcement learning (DRL) in policy optimization, a stereoscopic projection policy optimization method is proposed, which combines the simulation experiment method with the DRL method. On the basis of policy optimization research, a deep learning framework is selected according to the research problems, and a DRL stereoscopic project policy model based on the asynchronous advantage actor–critic (A3C) algorithm, which uses two groups of neural networks, is constructed. The optimized stereoscopic projection policy is obtained by the interactive learning between the DRL model and the simulation. The effectiveness of the cooperative optimization policy between the DRL and the simulation experiment is verified.