Existing techniques for system identification assume either partially known functional forms or moderate measurement noise.
A growing fraction of concrete applications, however, cannot rely on such simplifying assumptions, as often both the structure of the system, as well as the type of noise, are unknown to the modeler.
This talk discusses experimental design and model selection from the perspective of statistical learning and information theory, and introduces practical ways to perform active inference with noisy data.
Our work studies the interaction with uncertain systems to efficiently learn their internal causal structure from data, and introduces bounds with formal guarantees of near-optimality. We prove by reduction that our near-optimal design of experiments for nonlinear system identification yields a constant approximation factor which dominates all other efficient techniques, unless P=NP.
As our research is primarily motivated by concrete biomedical applications, I shall conclude by reporting our results in the context of cell signaling, automated microscopy, and personalized treatment of learning disabilities.
Alberto Giovanni Busetto is an Assistant Professor in the Department of Electrical and Computer Engineering at the University of California, Santa Barbara. He is also a member of the Center for Control, Dynamics and Computation and the Center for Bio-image Informatics. His research aims at automatically transforming “raw data” into “useful knowledge.” His work is strongly interdisciplinary, and includes work in the areas of system design, statistical learning, big data, fault-tolerant computing, Biocomputing, reinforcement learning, uncertainty quantification, design of experiments, and systems biology. He is also in collaboration with many other research labs in biomedicine, nanotechnology, finance and several branches of engineering.
