Research On Network Resource Management And Control Based On Deep Learning

With the increase in the number of users and the improvement of the requirement for QoS(Quality of Service),network resource management and control based on SDN(Software Defined Networking)becomes increasingly challenging.A large number of centralized control network control problems can be modeled as NP-hard combinatorial optimization problems,for which optimal solutions can hardly be obtained with the current solving ability of devices in a short period of time.In this work,a deep learning-based frame-work for network resource management and control problems is proposed based on the advantage that a trained neural network can perform rapid inference.This framework can employ neural networks to directly solve new network management and control problems with the use of historical solving experience.Based on the proposed framework,I try to solve the following three centralized control and delay-sensitive network resource man-agement and control scenarios:task placement and resource allocation problem in mobile edge computing,multiple video streams management and control problem in scenarios with shared bottleneck link bandwidth and multi-commodity stream load balancing prob-lem.In the problem of task placement and resource allocation problem in mobile edge computing,this work presents FAST-RAM,a deep learning solution based on multi-task learning.FAST-RAM can complete decision-making in milliseconds and obtain solutions close to optimal ones.A problem-splitting scheme is also proposed in this work to ad-dress the scale limitation of the multi-task learning framework,so that FAST-RAM can be generalized to scenarios with a larger task scale.To handle the problem of multiple video streams management and control problem in scenarios with shared bottleneck link band-width,FAIR-AREA,a solution which combines the feasibility assurance module with deep learning techniques,is presented.In the NS-3 simulation,FAIR-AREA is proved to give an almost optimal solution,and the loss function design which is more suitable for the combinatorial optimization problem is discussed.To cope with the problem of multi-commodity stream load balancing,I designed two deep learning solutions respec-tively based on the classification method and fitting method.The scope of application for these two methods is also discussed in several topological problem scenarios.The deep learning-based framework for network resource management and control problems has been successfully applied in the above three scenes in my work,which proves the performance advantage of this framework and provides a new solving idea for real-world centralized control and delay-sensitive network problems.