Advancing Edge Intelligence: Federated and Reinforcement Learning for Smarter Embedded Systems

Doctoral thesis, 2024

Context: The rapid growth of embedded devices and edge computing has brought new opportunities for creating intelligent systems. However, these systems face challenges such as limited computational power and the need to protect user privacy. As a result, there is a need for machine learning methods that can scale effectively, maintain privacy, and adapt to changing conditions in embedded applications.


Objective: This thesis focuses on improving the performance of machine learning models in embedded systems by using federated learning and reinforcement learning. The main goal is to develop methods that allow edge devices to work together without sharing raw data, which helps maintain privacy. Another goal is to make these systems more adaptable to dynamic environments, so they can perform better under changing conditions. Additionally, the research seeks to improve the efficiency of communication and computation across devices.

Method: The research uses a mix of case studies, simulations and real-world experiments. Federated learning is applied to allow edge devices to train models without centralizing the data, keeping sensitive information local. Reinforcement learning is used to help devices learn how to make better decisions by interacting with their environment. These two methods are tested in different scenarios to evaluate improvements in model accuracy, resource use, and adaptability.

Results: The results of this thesis highlight significant advancements in federated learning (FL) and reinforcement learning (RL) for embedded systems. A comprehensive literature review identified six key challenges and open research questions in FL, emphasizing the need for efficient communication, scalability, and privacy preservation. Case studies in telecommunications and automotive applications demonstrated that FL, particularly with asynchronous aggregation protocols, improves model performance, reduces communication overhead, and speeds up training in real-time, dynamic environments. Novel algorithms, such as AF-DNDF and deep RL approaches, further enhanced decision-making capabilities and adaptability in applications like autonomous driving and UAV base station deployment for disaster scenarios. The development of frameworks like EdgeFL provided practical solutions to overcome FL's implementation challenges, offering scalable, low-effort alternatives. Overall, the integration of FL and RL into embedded systems resulted in improved model accuracy, resource utilization, and adaptability, making these approaches highly suitable for real-world industrial use cases.

Conclusion: This research advances the field of edge intelligence by providing a practical approach to deploying machine learning models that are scalable, privacy-focused, and adaptive in embedded systems. The work demonstrates clear improvements in performance and offers a foundation for future research, which could explore more complex learning approaches and apply these techniques to a wider range of embedded systems.