Evolutionary Game Theory Enhances Cooperation in Decentralized Federated Learning

A new research paper proposes an Evolutionary Game Theory (EGT) framework to analyze and improve cooperation in Decentralized Federated Learning (DFL) systems. DFL is a privacy-preserving machine learning approach that operates without a central coordinator, making it susceptible to opportunistic behaviors. Traditional EGT models assume perfect rationality and static strategies, which are unrealistic. The paper addresses this by modeling peer-to-peer interactions on a lattice network under bounded rationality. It formulates a comprehensive payoff matrix that includes training costs, communication overhead, and cooperative rewards, and tailors a strategy update rule to capture spatial propagation dynamics. The work is available on arXiv with identifier 2608.01197, submitted as a new announcement. The contributions are threefold: modeling P2P interactions on a lattice, formulating a payoff matrix, and designing a strategy update rule. This research is relevant to the fields of AI, machine learning, and decentralized systems, offering a theoretical foundation for enhancing cooperation in DFL.