Federated, multi-agent, deep reinforcement learning

Abstract

The landscape of artificial intelligence (AI) is being reshaped by Federated Learning (FL), a decentralized approach to machine learning (ML) that enhances data privacy and collaborative model training. This thesis delves into the challenges and potential of FL, focusing on optimizing communication efficiency, enhancing model performance, and ensuring robustness in diverse settings. The research encompasses a detailed literature review and the identification of core challenges in FL. A series of studies were conducted to address specific aspects: optimizing data transmission and handling diverse model architectures, data partitioning and client selection, representation learning and federated distillation, incremental learning and knowledge retention, and training models with limited data. Each study contributed to the field by developing innovative algorithms, tested in simulated FL environments and compared with existing methods. Key findings of the research include improved communication efficiency with reduced overhead and bandwidth requirements, enhanced model performance in handling heterogeneous data and model architecture variability, effective strategies to combat catastrophic forgetting, and methodologies adept at working with limited and scattered data. The applicability of FL was demonstrated in practical scenarios, showcasing its potential in various domains. In conclusion, the dissertation significantly contributes to the advancement of FL. It addresses foundational challenges and demonstrates the adaptability and efficacy of FL in real-world applications. The findings emphasize FL's role as a method that ensures privacy, boosts efficiency, and showcases flexibility in the field of AI and ML.