Abstract and Bio Speakers NGB/LNMB Seminar
Optimization for and with Machine Learning
Optimization for and with Machine Learning
Marleen Balvert (Tilburg University)
Short Bio: is assistant professor in Operations Research at Tilburg University and a member of the Zero Hunger Lab. In 2017 she obtained her PhD at Tilburg University. After a postdoc in genome data analysis at Centrum Wiskunde & Informatica and Utrecht University she moved back to Tilburg University. Her research focuses on the intersection between optimization and machine learning techniques and is application driven. Marleen develops optimization and machine learning algorithms for the analysis of large genome data, in particular for genotype-disease data. As a member of the Zero Hunger Lab she develops mathematical tools to help NGOs to achieve Sustainable Development Goal 2: Zero Hunger.Title: The optimization behind deep neural networks (tutorial)
Abstract:
Over the past years deep learning has seen a major increase in popularity. This is because recent developments in deep neural networks – both related to hardware and software – allow us to solve classification problems we were never able to solve before. At the core of deep neural networks lies a challenging optimization problem. In this presentation you will be introduced to the basic concepts in deep neural networks, the underlying optimization problem and current approaches to solve the optimization.
Taylan Cemgil (DeepMind UK, Bogazici University)
Short Bio: Taylan Cemgil is currently a research scientist at DeepMind UK working in the robust and verified AI team and a Professor on sabbatical leave from Bogazici University, Dept. of Computer Engineering. Istanbul, TR . He obtained his PhD from Radboud University Nijmegen and worked as a postdoc in Amsterdam Univ. and Cambridge Univ. He is interested into basic machine learning methodologies, in particular robust machine learning, Bayesian modeling/inference, Monte Carlo and Matrix and Tensor decomposition methods. In the past, he worked with his group in collaboration with the industry on several applications such as audio and music processing, anomaly detection, time series, bioinformatics, tracking, sensor fusion, recommendation systems, customer analyticsTitle: Robustness in Machine Learning
Abstract: Machine learning systems are not robust by default. Even systems that are reported to outperform humans in a particular domain can be shown to fail at solving problems with virtually small variations on the problem data. This talk will focus on robustness in supervised learning and representation learning. In particular, we will give an outline of the current work on robust training and verification, with an emphasis on the role played by optimization and model construction. Our goal will be to highlight the nature of the challenges that are faced in checking and ensuring that learning systems work according to desired specifications.
Martina Fischetti (Vattenfall BA Wind)
Short Bio: Martina Fischetti is currently lead engineer in Vattenfall BA Wind, specialized in operational research (OR). She has M.Sc. degrees from the University of Padova (March 2014) and the University of Aalborg (June 2014) in Automation Engineering. In March 2018, she finished her Industrial PhD in OR at Technical University of Denmark in collaboration with Vattenfall, entitled Mathematical Programming Models and Algorithms for Offshore Wind Park Design. Her PhD work on the optimization of wind farm design and cable routing has been awarded various international prizes, such as the Best Industrial PhD from Innovation Fund Denmark (2019), EURO Doctoral Dissertation Award (2019), Glover-Klingman Prize (2018), AIRO Best Application Paper award (2018), the Best Student Paper Award ICORES (2017), and finalist positions at the EURO Excellence in Practice award (2018) and the prestigious INFORMS Franz Edelman award (2019).Title: Operations Research + Machine Learning for the design of future offshore wind farms
Abstract: Sustainability is a key focus in our society that is today working to change towards a greener future. Wind energy, in particular, is attracting always more attention as source of renewable energy. In this picture, Vattenfall is working towards the ambitious goal of becoming fossil free within one generation. To achieve this goal, innovation (and optimization!) is of key importance.
This talk presents how Vattenfall is using advanced operations research and analytics for designing cheaper and more profitable offshore wind farms. The talk will focus on the design phase of offshore wind farms, explain in details the optimization challenges faced by companies as Vattenfall. In particular, we will focus on the Offshore Wind Farm Design problem, that is the task of deciding how to position turbines offshore in order to increase the overall farm production and reduce costs. This task is particularly challenging due to the interference effects among turbines, due to the stochasticity of wind and due to the high dimensionality of the problem in real applications. Mixed Integer Programming models and other state-of-the-art optimization techniques have been developed to solve this problem. These tools are nowadays fully deployed in Vattenfall and used for the design of all offshore wind farms. They have been used, for example, for the design of Hollandse Kust Zuid in the Netherlands, which will be the first offshore wind farm ever constructed without any subsidies. This is a huge milestone for the whole wind energy business.
These advanced optimization tools allowed Vattenfall to think out of the box, take more informed decision and perform different what-if-analyses. In particular, we can foresee the number of what-if analyses to quickly grow in the future. Therefore we have looked into Machine Learning techniques.
In the specific, we propose a combination of Mathematical Optimization and Machine Learning to estimate the value of optimized solutions. We investigate if a machine, trained on a large number of optimized solutions, could accurately estimate the value of the optimized solution for new (unseen) instances. This research question could be of general interest for the OR community, but we focus on the wind farm layout application in our research. Given the complexity of the wind farm layout problem and the big difference in production between optimized/non optimized solutions, it is not trivial to understand the potential value of a new site without running a complete optimization. This could be too time consuming if a lot of sites need to be evaluated, therefore we propose to use Machine Learning to quickly estimate the potential of new sites (i.e., to estimate the optimized production of a site without explicitly running the optimization). To do so, we trained and tested different Machine Learning models on a dataset of 3000+ optimized layouts found by the optimizer. Our results show that Machine Learning is able to efficiently estimate the value of optimized instances for the offshore wind farm layout problem.
Holger Hoos (Leiden University)
Short Bio: Holger H. Hoos is Professor of Machine Learning at Universiteit Leiden (the Netherlands) and Adjunct Professor of Computer Science at the University of British Columbia (Canada), where he also holds an appointment as Faculty Associate at the Peter Wall Institute for Advanced Studies. He is a Fellow of the Association for the Advancement of Artificial Intelligence (AAAI) and past president of the Canadian Association for Artificial Intelligence / Association pour l'intelligence artificielle au Canada (CAIAC).Holger's research interests span artificial intelligence, empirical algorithmics, bioinformatics and computer music. He is known for his work on machine learning and optimisation methods for the automated design of high-performance algorithms and for his work on stochastic local search. Based on a broad view of machine learning, he has developed - and vigorously pursues - the paradigm of programming by optimisation (PbO); he is also one of the originators of the concept of automated machine learning (AutoML). Holger has a penchant for work at the boundaries between computing science and other disciplines, and much of his work is inspired by real-world applications.
Title: Programming by Optimisation: Automated algorithm configuration, selection and beyond
Abstract:
In recent years, there has been a significant increase in the use of automated algorithm design methods, such as automated algorithm configuration and portfolio-based algorithm selection, across many areas within operations research, artificial intelligence and beyond. These methods are based on cutting-edge machine learning and optimisation techniques; they have also led to substantial advances in those areas.
In this tutorial, I will give an overview of these automated algorithm design methods and introduce Programming by Optimisation (PbO), a principled approach for developing high-performance software based on them. I will explain how PbO can fundamentally change the nature of developing solvers for challenging computational problems and give examples for its successful application to a range of prominent problems from OR and AI - notably, mixed integer programming, the travelling salesman problem, AI planning, automated reasoning and machine learning.
Andrea Lodi (Polytechnique Montréal)
Short Bio: Andrea Lodi received the PhD in System Engineering from the University of Bologna in 2000 and he has been Herman Goldstine Fellow at the IBM Mathematical Sciences Department, NY in 2005–2006. He has been full professor of Operations Research at DEI, University of Bologna between 2007 and 2015. Since 2015 is Canada Excellence Research Chair in “Data Science for Real-time Decision Making” at the École Polytechnique de Montréal. His main research interests are in Mixed-Integer Linear and Nonlinear Programming and Data Science and his work has received several recognitions including the IBM and Google faculty awards. He is author of more than 100 publications in the top journals of the field of Mathematical Optimization and Data Science. He serves as Editor for several prestigious journals in the area. He has been network coordinator and principal investigator of two large EU projects/networks, and, since 2006, consultant of the IBM CPLEX research and development team. Finally, Andrea Lodi is the co-principal investigator of the project "Data Serving Canadians: Deep Learning and Optimization for the Knowledge Revolution", recently generously funded by the Canadian Federal Government under the Apogée Programme and scientific co-director of IVADO, the Montréal Institute for Data Valorization.Title: On the interplay between Discrete Optimization and Machine Learning
Abstract: In this talk I review a couple of applications on Big Data that I personally like and I try to explain my point of view as a Mathematical Optimizer -- especially concerned with discrete (integer) decisions -- on the subject. I advocate a tight integration of Machine Learning and Mathematical Optimization (among others) to deal with the challenges of decision-making in Data Science. For such an integration I try to answer three questions:
1) what can optimization do for machine learning?
2) what can machine learning do for optimization?
3) which new applications can be solved by the combination of machine learning and optimization?
Tristan van Leeuwen (Utrecht University)
Short Bio: Tristan van Leeuwen was born in 1981 in the Netherlands. He received his BSc. and MSc. in Computational Science from Utrecht University in the Netherlands. He obtained his PhD. in geophysics from Delft University in 2010. After spending some time as a postdoctoral researcher at the University of British Columbia in Vancouver, Canada and the Centrum Wiskunde & Informatica in Amsterdam, the Netherlands, he returned to Utrecht University in 2014 as an assistant professor at the mathematical institute.His research interests include: inverse problems, computational imaging, tomography and numerical optimisation.
Title: Optimisation strategies for machine learning - harnessing inexactness
Abstract: Many (supervised) learning problems lead to optimisation problems. Often, the quantities required for finding an optimal solution (e.g., the gradient or Hessian) cannot be computed exactly. In this talk I will give an overview of recent approaches that allow one to use such approximated quantities in a rigorous way.