Adaptive federated learning device selection strategy based on edge-end performance prediction

Aiming at the problems of performance fluctuation and unstable training of edge-end devices during training, this paper proposes an adaptive federated learning device selection strategy based on edge-end performance prediction. The method aims to solve the problem that the traditional federated device selection strategy does not consider the performance of edge-end devices, which leads to a large amount of wasted arithmetic and communication resources and slow convergence speed. Specifically, the method establishes a linear regression model based on the overhead of each resource by collecting the historical performance data of edge-end devices to realize the performance prediction based on the amount of local data; and realizes an efficient and stable federation training process in dynamic edge-end environments by realizing a device selection decision module based on the dobby slot machine algorithm, which adaptively learns the linkage between the performance characteristics of the devices and the device selection decision. Experimental results show that the method proposed in this paper can select stable devices for updating as much as possible, and has significant advantages in ensuring high global convergence speed and high training accuracy