Efficient Large-Scale Personalizable Bidding for Multiagent Auction-Based Federated Learning

Auction-based Federated Learning (AFL) enables open collaboration among self-interested model users (MUs) and data owners (DOs). Muti-agent reinforcement learning-based bidding methods have gained traction in AFL due to their ability to deal with complex interactions among MUs and DOs. However, existing methods lack the efficiency required to manage a large number of AFL MUs. To bridge this gap, we propose the Cluster-based Personalizable Multi-Agent Reinforcement Learning bidding strategy for MUs in Auction-based Federated Learning (CPMARL-AFL) approach. It leverages the similarities among MUs to group them into distinct clusters, each being treated as a super MU and equipped with its own bidding agent. The core of our proposed multi-agent reinforcement learning formulation revolves around the interactions between cluster agents and DOs, as well as the interactions among the cluster agents themselves. To achieve personalized bidding for each MU within a cluster, the final bid price is fine-tuned based on the expected utility gain from a given DO and the mean utility of the entire cluster, which is influenced by the bid prices generated by the cluster's bidding agent. In addition, to prevent the bidding agent from converging to sub-optimal solutions, we design a novel reward method that facilitates efficient convergence towards the optimal solution. Through extensive experimentation conducted on six widely used benchmark datasets, CPMARL-AFL demonstrates superior performance compared to eight state-of-the-art approaches. It outperforms the best baseline by 1.28% in terms of the utility and 1.40% in terms of the test accuracy achieved by the resulting FL model.