Master-Seminar: Machine Learning in Optimization Algorithms

Winter Term 2019/20

Content Requirements Dates Organization Contact

Content

Machine learning methods and optimization methods can be combined in manifold ways. In this seminar, we concentrate on the use of machine learning techniques to accelerate already existing optimization algorithms, in particular solvers for mixed-integer (linear) programs. Which subroutines of well-known optimization algorithms can profit from an application machine learning? Which machine learning algorithms are best suited to support decision making within optimization algorithms? Which problem classes benefit from combined approaches? Those question have been studied and partially answered in the literature of the past five years. Some highlights of recent research results are reviewed and we gain some insight into the current developments in this field.

Requirements

Students should have basic knowledge in mathematical optimization.

Dates

First meeting: Introduction and paper assignment

23.10.2019, 16:00-17:30 in room 3028 at ZIB.

Second meeting: Short talks

25.11.2019, 12:15-13:15 in room 4359 at ZIB.

Third meeting: Seminar talks

10.02.2020, 10:00-17:00 in room 2006 at ZIB.

Organization

At the first meeting we will shortly discuss the topic in general and distribute the assignments, i.e., papers. At the second meeting, you are supposed to give a short, introductory talk (at most 5 minutes) on your topic. To obtain credit points, you are required to hand in a short summary of your talk (LaTeX, 5 pages). The summary has to be sent to your advisor by E-mail two weeks before the third meeting. It will be graded and handed back to you before the final meeting in order to provide a first feedback. The seminar itself will take place on one or two days (depending on the number of attendees) in the last weeks of the semester. Talks should be prepared for 45 minutes, such that 15 minutes remain for questioning and discussions. Having submitted a summary is a requirement for giving the final presentation. Your final grade will be composed of 60% and 40% from the evaluation of your talk and summary, respectively.

Assignment

Candidate	Topic
M.B.	Learning MILP Resolution Outcomes Before Reaching Time-Limit
Y.K.	Learning a Classification of Mixed-Integer Quadratic Programming Problems
M.H.	Learning to Branch in Mixed Integer Programming
G.M.	Exact Combinatorial Optimization with Graph Convolutional Network
I.S.	Learning To Search via Retrospective Imitation
M.F.	Reinforcement Learning for Integer Programming: Learning to Cut
I.R.	Adaptive Large Neighborhood Search

Contact

	Sprechstunde	Raum	Telefon	email
Dr. Timo Berthold	n.V.	ZIB 3301	84185-425	bertholdzib.de

Zuletzt aktualisiert: 22. Oktober 2019