OPTIMAL (Optimization for and with Machine Learning) is a collaborative project involving researchers from Centrum Wiskunde & Informatica (or CWI, a national research institute for mathematics and computer science), Tilburg University, Delft University of Technology and UvA. Project manager Dick den Hertog is part of the research team and explains the challenges the project addresses. 'We want to analyse and develop better AI techniques like neural networks as well as improve various optimisation techniques through machine learning.'

Fighting famine and cancer

An example of optimisation with machine learning is a collaboration with Tilburg University which has led to the development of optimisation models for the United Nations world food programme. The UN has been using these models for a number of years now to combat humanitarian catastrophes. 'We can use optimisation techniques to calculate the ingredients a minimal meal should contain in order to meet food standards, with purchasing and transport costs kept as low as possible. Generally speaking, such problems are complex puzzles with hundreds of thousands of variables. Some aspects such as the palatability of a meal cannot easily be captured in a formula. But it can be done with machine learning and such aspects can then be added to a model.'

Another example comes from the world of medicine. Machine learning can help improve the radiation of tumours. Together with Harvard Medical School, the ABS is using the methodology to optimise the radiation treatment of cancer patients. In this application, the availability of large amounts of data plays a major role. 'It’s a model based on physics and augmented by model elements acquired via machine learning. We use historical treatment data to increase the probability that tumours will actually disappear through radiation.'

Towards a data-driven society

The research project consists of several parts, including research by various doctoral and post-doctoral students. 'We allocate tasks and meet regularly to keep everyone up to date with the progress we’ve made. At such moments, we also look at new ideas and how these could be applied by the companies and institutions involved.”

 “What are we hoping for? When it comes to optimisation for machine learning, I hope we’ll get a better understanding of how techniques can be used to improve machine learning. In addition, I hope we’ll be able to take a big step forward in the application of machine-learning models so that the work done in our society will be even more data-driven.'

Analytics for a Better World

OPTIMAL encompasses several cross-fertilisation efforts, including the Analytics for a Better Word initiative set up by Den Hertog together with Professor Dimitris Bertsimas (Massachusetts Institute of Technology). The project was launched to deploy analytics, in the broadest sense of the word, for the 17 UN SDGs. 'We organise regular webinars, where scientists can throw light on some great applications.'

Apart from Analytics for a Better World being offered as a course at both UvA and MIT, Den Hertog and his American colleagues want to create a scientific journal to show the research results to a broader audience. The programme is intended to stimulate the application of analytics as a way of achieving the SDGs and to help initiate new studies . 'Take the research we’re doing with Sanquin. Merel Wemelsfelder, a PhD student at the ABS, is looking at how the blood bank can improve its product distribution. Variables of interest are the various points of distribution, time pressure, supplies and the fact that Sanquin is dealing with a growing range of blood products.'

Hospitals in Asia

Yet another project set up under the Analytics for a Better World umbrella is a collaboration with the World Bank. 'In the past 18 months, we’ve been optimising the locations of new hospitals in East Timor. And there are many more examples. In Vietnam, for instance, we’re optimising locations for stroke centres. The goal is to reduce the number of deaths and patients with complications. These are complex puzzles but, with analytics and machine learning, we can make a real difference.'