A Game-Theoretic Incentive Mechanism for Multi-Distributor Multi-Agent Federated Learning

In a multi-distributor multi-agent federated learning architecture, it is desired to maximize the overall system benefits by optimizing the association relationship between the base station and users. To address this issue, a novel incentive mechanism based on Stackelberg game is proposed. Firstly, each mobile device determines its own strategy based on the utility function. Then, each base station adjusts its own strategy based on the optimal solution of the mobile devive to maximize its own utiltiy function. By relaxing the binary variables that represent the correlation relationship into continuous variables, the 0–1 optimization problem for base stations and mobile devices is transformed into a linear problem and solved. The simulation results show that the proposed algorithm is better than random algorithm in terms of training loss and testing accuracy.