Temel Kavramlar
The core message of this article is to propose a closed-loop model analytics framework that allows for effective evaluation of the trained global model in federated learning, and to study a goal-directed client selection problem based on this framework to improve the efficacy of the trained global model.
Özet
The article proposes two closed-loop model analytics frameworks, the FL&FA and FL&DA frameworks, to evaluate the trained global model in federated learning (FL). These frameworks connect the trained global model with the original dataset, enabling the evaluation results to guide the iterative adjustment of the FL training process.
Based on the proposed frameworks, the authors study a goal-directed client selection problem, which aims to find an optimal subset of clients for model training to maximize the clients' average opinions of the trained global model. This problem is formulated as a stochastic multi-armed bandit (SMAB) problem.
The authors propose two algorithms to solve the SMAB problem under the FL&FA and FL&DA frameworks, respectively:
Quick-Init UCB algorithm: This algorithm reduces the initialization phase of the standard UCB algorithm by dividing clients into groups and selecting each group once. It achieves an asymptotically optimal regret bound.
BP-UCB algorithm: This algorithm employs belief propagation (BP) to facilitate message exchange among clients in the decentralized FL&DA framework. It also achieves an asymptotically optimal regret bound.
The numerical results demonstrate that the proposed algorithms achieve nearly optimal performance, with a gap of less than 1.44% and 3.12% under the FA and DA frameworks, respectively.