Two-Layer Optimization With Utility Game and Resource Control for Federated Learning in Edge Networks

Federated learning (FL) is a distributed machine learning paradigm that can be organized in two layers. In the outer layer of users, there is a model interaction process between the task publisher and users, through which all parties obtain their respective utilities. However, these parties’ utilities are coupled, both depending on the training sample size and local iterations. In the inner layer of users, a user's multiple devices (e.g., computers and smart phones) can be used to jointly train local models efficiently. Yet, due to device heterogeneity, it is challenging for users to determine which devices to participate in local training and allocate how many computing and communication resources to minimize training costs. In this paper, we tackle this novel two-layer optimization problem by designing utility game and resource control strategies. In the outer layer, we model the relationship between the task publisher and users as a Stackelberg game and obtain the optimal solution for both parties by solving a unique Stackelberg equilibrium point; while in the inner layer, we formulate the optimization problem as a mixed integer nonlinear programming problem, which is decomposed into sub-problems and solved by devising resource control algorithm based on successive convex approximation. Finally, extensive experiments show that the proposed algorithms outperform baseline algorithms.