Optimization-Based Estimation for Real-World Systems: Blending System Identification and Machine Learning
In this thesis, a groundbreaking approach to solving estimation problems for real-world, physical systems is explored. Drawing inspiration from both classical system identification and modern machine learning techniques, the main focus is on optimization-based methods. By incorporating concepts such as regularization and basis-function expansions, this approach enables the encoding of prior knowledge and the addition of nonlinear modeling power while keeping practical data requirements.
The thesis covers a wide range of applications, with many inspired by robotics but extending beyond this field. The proposed methods and algorithms are illustrated through real-world applications that have motivated the research. Key topics covered include dynamics modeling and estimation, model-based reinforcement learning, spectral estimation, friction modeling, and state estimation and calibration in robotic machining.
One major aspect of the work is the development of regularization strategies that allow for the incorporation of prior domain knowledge into flexible, overparameterized models. By utilizing classical control theory and modern deep learning tools, the researchers gain valuable insights into training and regularization. A particular focus is placed on scenarios where gathering data comes with a high cost, making it crucial to optimize the methods for efficiency.
In the robotics-inspired parts of the thesis, the focus shifts towards practicality and implementation outside the research setting. The methods developed are motivated by real-world applications and are rigorously tested in realistic settings. To ensure accessibility, open-source implementations of all proposed methods and algorithms are made available, allowing others to benefit from the research.
While this thesis presents an optimization-based approach to estimation problems, it is important to note that there are various perspectives and approaches within the field. Nevertheless, this work contributes significantly to the understanding and practical implementation of estimation techniques for real-world systems. By blending classical system identification and modern machine learning, these methods offer innovative solutions to challenges faced in areas such as robotics and beyond.
In conclusion, the optimization-based estimation methods presented in this thesis provide a valuable contribution to the field of system identification and machine learning. By incorporating prior knowledge, leveraging modern deep learning tools, and addressing real-world challenges, these methods offer practical and efficient solutions for estimation problems. Through rigorous testing, open-source implementations, and a wide range of applications, the researchers showcase the potential of their approach in various domains. As the world becomes increasingly dependent on real-world systems, this research opens doors to more accurate and efficient estimation techniques that can drive advancements in diverse fields.
