Appendix A Derivation of Equation (7) 561

Table 5 shows the positioning of FedL2P against existing literature. This personalized policy can either 1) be fixed, e.g. FedEx which randomly samples per-client hyperparameters from learned categorical distributions. Scenarios where it's expensive to train from scratch for a new group of clients, e.g. This is illustrated in Section 4.4 where we adapt a publicly available pretrained Scenarios where it's important to also maintain a global model with high initial accuracy - Note that our approach also does not critically depend on the global model's performance.