Automated Telescience: Active Machine Learning Of Remote Dynamical Systems
Automated science is an emerging field of research and technology that aims to extend the role of computers in science from a tool that stores and analyzes data to one that generates hypotheses and designs experiments. Despite the tremendous discoveries and advancements brought forth by the scientific method, it is a process that is fundamentally driven by human insight and ingenuity. Automated science aims to develop algorithms, protocols and design philosophies that are capable of automating the scientific process. This work presents advances the field of automated science and the specific contributions of this work fall into three categories: coevolutionary search methods and applications, inferring the underlying structure of dynamical systems, and remote controlled automated science. First, a collection of coevolutionary search methods and applications are presented. These approaches include: a method to reduce the computational overhead of evolutionary algorithms via trainer selection strategies in a rank predictor framework, an approach for optimal experiment design for nonparametric models using Shannon information, and an application of coevolutionary algorithms to infer kinematic poses from RGBD images. Second, three algorithms are presented that infer the underlying structure of dynamical systems: a method to infer discrete-continuous hybrid dynamical systems from unlabeled data, an approach to discovering ordinary differential equations of arbitrary order, and a principle to uncover the existence and dynamics of hidden state variables that correspond to physical quantities from nonlinear differential equations. All of these algorithms are able to uncover structure in an unsupervised manner without any prior domain knowledge. Third, a remote controlled, distributed system is demonstrated to autonomously generate scientific models by perturbing and observing a system in an intelligent fashion. By automating the components of physical experimentation, scientific modeling and experimental design, models of luminescent chemical reactions and multi-compartmental pharmacokinetic systems were discovered without any human intervention, which illustrates how a set of distributed machines can contribute scientific knowledge while scaling beyond geographic constraints.
Automated science; Machine learning; Remote experimentation
Kress Gazit, Hadas; Hooker, Giles J.; Campbell, Mark
Ph.D. of Mechanical Engineering
Doctor of Philosophy
dissertation or thesis