Federated Learning as a Mean-Field Game

In the following sections, we briefly review the necessary material from federated learning and mean-field games that will be used later when similarities start to appear. In federated learning (Section 1.1), we present the general idea and one of its standard algorithm called Federated Averaging (FedAvg). The general setting of stochastic games is introduced in Section 2. Two main pillars of mean-field games that are stochastic optimal control and stochastic differential games are introduced in Sections 2.1 and 2.2 respectively. The introduction of mean-field games is completed in Section 2.3. Finally, the connection between federated learning and mean-field games is established in Section 3.