July 21, 2017: Invited Speaker Seminar: Polynomial Optimisation in Power Systems at IBM Research

Dr. Jakub Marecek
IBM Research

Friday, July 21, 2017 at 2:00 pm
Room EV003.309

Abstract

Problems in power systems modelling alternating-current transmission constraints are non-convex, but of a great and growing importance in practice. We have shown that: i) one can construct a hierarchy of convexifications, whose optima converge to the global optimum of the non-convex problem, ii) custom first- and second-order methods for solving the convexifications are competitive with leading heuristics, iii) the first-order methods have trivial per-iteration time and memory requirements, but their rates of convergence limit their direct application to large instances. Methods for switching from solving the convexification (e.g., using the first-order methods) to (any second-order methods on) the non-convex problem can guarantee convergence at the optimum of the convexification at quadratic rate of convergence. This allows one to tackle large-scale instances in practice and to guarantee global convergence in theory.

This summarises recent papers in IEEE T. Power Systems [31(1): 539–546], IEEE T. Smart Grid [DOI 10.1109/TSG.2017.2715282], and Optimization Methods and Software [32(4): 849-871], which are joint work with Bissan Ghaddar, Alan Liddell, Jie Liu, Martin Mevissen and Martin Takac.

Biography

Jakub Marecek is a research staff member at IBM Research. Together with some fabulous colleagues, Jakub develops solvers for optimisation problems in IBM's Smarter Cities Technology Centre. His recent work includes polynomial optimisation in power systems, policies for bi-level optimisation with uncertain dynamics, and a heterogeneous stream processing system for urban traffic management (called "Insight"), which has just won the 5th Annual Award for Excellence by ITS Ireland. Jakub is also the principal investigator for VaVeL, an H2020 project within the big data call, and a programme committee member for AISTATS 2017,ICAPS 2017, and Knowme 2017. Prior to joining IBM in August 2012, Jakub had worked on distributed solvers for non-smooth convex optimisation problems at the University of Edinburgh. Jakub grew up in Brno, the Czech Republic, where he had worked in two start-up companies before studying for his first degree.