Multi-Session Multi-Objective Budget Optimization for Auction-based Federated Learning

Auction-based Federated Learning (AFL) has become a significant focus in recent years. Existing approaches for model users (MUs) in FL generally operate under the assumption that the entire set of essential data owners (DOs) must be gathered before the training process initiates. However, in practical scenarios, the MU can initiate the FL training process multiple times, gradually recruiting DOs across multiple FL training sessions. Current AFL methods are not equipped to handle such situations. The challenge of optimizing the AFL budget across multiple sessions and objectives remains unresolved. To address this gap, we propose the Efficient and Utility Return-Optimizing budget management strategy for MUs in Auction-based Federated Learning (EURO-AFL). By incorporating the hierarchical reinforcement learning framework, EURO-AFL concurrently optimizes inter-session budget pacing and intra-session budget allocation, with the dual objective of maximizing total utility while minimizing waiting time. Extensive experiments conducted on four real-world datasets illustrate the substantial advantages of EURO-AFL compared to five state-of-the-art baselines. It outperforms the best-performing baseline by 6.26%, 47.9% and 17.6% in terms of model accuracy, number of training sessions sustained, and training efficiency, respectively. To our best knowledge, EURO-AFL is the first multi-session multi-objective budget optimization approach designed for AFL MUs.