This project focuses on developing efficient methods for training and evaluating modern machine learning models, with an emphasis on scalable optimization, robustness, and data‑efficient learning approaches. The research investigates algorithmic principles that enable reliable decision‑making across diverse environments, including reinforcement learning (RL) settings where agents must operate under uncertainty and limited feedback. We aim to understand how computational efficiency, communication constraints, and heterogeneous data sources influence the behavior and performance of contemporary ML systems. Part of the work is related to our ongoing Swedish Research Council (VR) project on Federated Reinforcement Learning (FedRL). This research area examines how multiple AI systems can learn collaboratively without sharing private data and explores challenges arising from agent heterogeneity, communication efficiency, and scalability in multi‑agent settings. These insights inform the broader goal of developing generalizable and computationally efficient ML methodologies suitable for a wide range of applications.