Pedro Ferreira (Oxford)

An essential part of physics is finding algorithms or rules which approximate well a physical process (in the form of data or complex models). The current approach is to construct “black box” machine learning algorithms which are fast, predictive but opaque. I will discuss a different approach which uses symbolic regression to infer the simplest mathematical expressions that best approximate a data set. I will explain its underlying structure and the information theoretic aspects that underly it. I will then apply it to three test cases - the expansion of the Universe, the galaxy radial acceleration relation, inflation - and discuss its power and limitations.