A stochastic learning algorithm for multi-agent game in mobile network: A Cross-Silo federated learning perspective

Collaboration in mobile network involves multiple servers and smart devices, introducing the challenging of coordination. The inherent challenges arise from incomplete environmental information due to dynamic networks and privacy concerns, making collaboration for all participants a complex task. In this work, we introduce a Multi-agent step Refinement Stochastic Learning Algorithm (MARSL) empowered by neighbor search, achieving superior outcomes with low complexity compared to baseline algorithms. To demonstrate the performance of the proposed algorithm, we provide comprehensive theoretical analysis on the superior properties. We then formulate two Non-IID cross-silo Federated Learning scenarios as typical non-convex cases in mobile network collaboration. By conducting multiple experiments, we illustrate the algorithm’s superior performance in both final utility and computation complexity. This contribution addresses the cooperation challenge in Cross-silo FL, providing an effective solution for scenarios with incomplete environmental information.