Design And Implementation Of A Cross-domain Task Scheduling Algorithm Based On Federated Deep Reinforcement Learning

With the advancement of "The Belt and Road Initiative",the scientific and technological cooperation between China and countries along the Belt and Road has become increasingly close.In this context,in order to make full use of the cross-domain and heterogeneous network resources and computing power resources of various countries,and to be able to deal with different types of task loads in the international cross-domain heterogeneous network environment,this puts higher requirements on the task scheduling algorithm in this specific environment.However,due to the particularity and non-universality of international cross-domain task scheduling scenarios,research on cross-domain task scheduling is still in its infancy.Therefore,the research on efficient cross-domain task scheduling algorithm is very necessary and valuable.In cross-domain heterogeneous network environment,there are many issues such as the diversity of scheduling task types,the performance difference of computing nodes,the network delay caused by cross-domain scheduling,the execution efficiency of scheduling tasks under large-scale nodes,the load balancing of cluster,data security and privacy protection in different autonomous domains to be studied and solved.Therefore,the research on cross-domain task scheduling algorithms still faces a series of challenges.This thesis first investigates the research status of cross-domain task scheduling,analyzes and summarizes the limitations of existing task scheduling algorithms,and clarifies the optimization goals of cross-domain task scheduling,which guides the direction for the development of subsequent research content.Then,based on the basic characteristics of the cross-domain heterogeneous network environment,this thesis designs a cross-domain static task scheduling algorithm based on deep reinforcement learning,and a large number of comparative experiments show that the algorithm can significantly improve the execution efficiency of scheduled tasks and achieve load balancing of scheduling cluster.For the dynamic task scheduling scenario with multiple types of tasks,this thesis designs a type-aware dynamic task scheduling algorithm based on deep reinforcement learning,and optimizes the load balancing ability of the algorithm in the high concurrent task submission scenario.Experiments show that the algorithm exhibits good scheduling performance under different types of task loads.In order to solve the data island problem and data privacy problem of each autonomous domain node client in crossdomain environment,this thesis proposes a decentralized federated learning technology to realize inter-domain collaborative training and intra-domain privacy protection of deep reinforcement learning agents.Specifically,this thesis designs a federated model parameter merging algorithm that is more suitable for cross-domain environmental characteristics,uses differential privacy method to improve the security of federated learning,and uses decentralized design to improve the fault tolerance and availability of federated learning.Finally,this thesis summarizes the above work and looks forward to future work.