On the geometry of optimization problems and their structure

Overview & Synopsis

In numerous fields such as machine learning, operational research or circuit design, a task is modeled by a set of parameters to be optimized in order to take the best possible decision. Formally, the problem amounts to minimize a function describing the desired objective with iterative algorithms. The development of these latter depends then on the characterization of the geometry of the function or the structure of the problem. In a first part, this thesis studies how sharpness of a function around its minimizers can be exploited by restarting classical algorithms. Optimal schemes are presented for general convex problems. They require however a complete description of the function that is rarely available. Adaptive strategies are therefore developed and shown to achieve nearly optimal rates. A specific analysis is then carried out for sparse problems that seek for compressed representation of the variables of the problem. Their underlying conic geometry, that describes sharpness of the objective, is shown to control both the statistical performance of the problem and the efficiency of dedicated optimization methods by a single quantity. A second part is dedicated to machine learning problems. These perform predictive analysis of data from large set of examples. A generic framework is presented to both solve the prediction problem and simplify it by grouping either features, samples or tasks. Systematic algorithmic approaches are developed by analyzing the geometry induced by partitions of the data. A theoretical analysis is then carried out for grouping features by analogy to sparse methods.