Federated learning (FL) is a distributed machine learning (ML) paradigm which has lately enjoyed the attention of various different domains. Since it brings added privacy, security, and computability advantages, it is a natural fit for many different distributed ML environments. As a consequence of it being a paradigm with such broad applications, many different algorithmic, privacy, personalization, communication, and training efficiency challenges arise, just to name a few. This work focuses on optimizing the FL process in terms of strategy selection, communication, and computational efficiency, for both cross-silo and cross-device settings. This can be achieved during the setup phase by analysing the FL environment, and by selecting optimal network architectures.