Fast-Convergence Federated Edge Learning via Bilevel Optimization
Sai Wang, Yi Gong · 2023
In this paper, we propose a fast-convergence federated edge learning by jointly optimizing the number of epochs and batch size. Specifically, we formulate a bilevel optimization problem. The upper level problem aims to trade off between aggregating time and loss error. The lower level problem is designed to eliminate synchronization waiting time. To solve this, we develop an approximate projection method. First, we obtain the optimal solution to the upper level problem using convex optimization. Based on the optimality conditions derived for the lower level problem, we formulate a projection optimization to minimize the distance between the projected points and upper level optimal solution. Our results demonstrate that the proposed method significantly outperforms other benchmark solvers on convergence speed for federated edge learning.