J2-0734

Dynamic systems, the state of which changes over time, are ubiquitous in both science and engineering. The task that we address in this proposal is the task of automated modeling of dynamic systems, i.e., the process of establishing models from observations and measurements of system behavior. Recent approaches to the task at hand, developed in the area of computational scientific discovery, suffer from a number of limitations: deterministic behavior of the modeled system is commonly assumed, where the state of the system in the future is completely determined by its present state; it is also assumed that a single model is valid over the entire lifetime of the modeled system; current methods employ standard approaches to parameter estimation, mainly based on gradient descent; finally, they suffer from computational complexity problems. The project will develop methods that overcome each of the four major limitations mentioned above: it will develop methods for learning probabilistic models of dynamic system behavior; methods for learning structurally dynamic models, whose structure and/or parameters change over time; improved methods for parameter estimation in the context of automated modeling of dynamic systems; and parallel algorithms for automated modeling. It will also evaluate the developed methods and demonstrate their utility by applying them to practical problems from the areas of ecological modeling (focusing mainly on aquatic ecosystems and waste-water treatment plants) and systems biology (learning metabolic and gene regulation networks).