Machine Learning and System Identification for Estimation in Physical\n Systems

Abstract

In this thesis, we draw inspiration from both classical system identification\nand modern machine learning in order to solve estimation problems for\nreal-world, physical systems. The main approach to estimation and learning\nadopted is optimization based. Concepts such as regularization will be utilized\nfor encoding of prior knowledge and basis-function expansions will be used to\nadd nonlinear modeling power while keeping data requirements practical. The\nthesis covers a wide range of applications, many inspired by applications\nwithin robotics, but also extending outside this already wide field. Usage of\nthe proposed methods and algorithms are in many cases illustrated in the\nreal-world applications that motivated the research. Topics covered include\ndynamics modeling and estimation, model-based reinforcement learning, spectral\nestimation, friction modeling and state estimation and calibration in robotic\nmachining. In the work on modeling and identification of dynamics, we develop\nregularization strategies that allow us to incorporate prior domain knowledge\ninto flexible, overparameterized models. We make use of classical control\ntheory to gain insight into training and regularization while using flexible\ntools from modern deep learning. A particular focus of the work is to allow use\nof modern methods in scenarios where gathering data is associated with a high\ncost. In the robotics-inspired parts of the thesis, we develop methods that are\npractically motivated and ensure that they are implementable also outside the\nresearch setting. We demonstrate this by performing experiments in realistic\nsettings and providing open-source implementations of all proposed methods and\nalgorithms.\n