The PhD Chronicles: Part 1
My journey through the archives of the Artificial Intelligence Lab at the VUB
As a summer job at the Artificial Intelligence Lab of the VUB, I was asked to tell the story of the lab. On my first day, they told me “have a look through the PhD theses and see what is interesting to highlight”. A little scared, but very excited, I started browsing through the archives. I had no clue what to expect… You know what I found out while scrolling them? AI researchers are problem solvers.
Each and every PhD thesis started off with a problem that the researchers were so passionate about that they embarked on a journey that would take years out of their lives to investigate. Years spent, solely focused on this problem with the single purpose of solving it to the best of their abilities and making humankind a little better than before.
“Multi-Criteria Reinforcement Learning for Sequential Decision Making Problems”
Impossible to read for some, fascinating breakthrough for others. This is the title of the first PhD thesis that I wish to place in the spotlight for a second. It was written by Kristof Van Moffaert, who received his doctorate in 2016. This thesis tells the story of the creation of super cool reinforcement learning algorithms that can help people in making better decisions. (For a crash course in RL, click here: link my blogpost of RL)
Quick context: Say, you are the boss of a factory, and you want some help in making executive decisions. An algorithm can help compute that. The problem is that you have many different criteria that are important to take into consideration. For example, you want to improve product quality, maximize profit and reduce environmental impact at the same time. However, when we talk about using algorithms that learn to make good decisions, usually these algorithms only know how to make good decisions with regards to one objective.
Now, the algorithms that Van Moffaert created are specifically designed for this type of complex decision where you need to take account of multiple criteria. His algorithms learn by interacting with the environment and discovering optimal balances of the different criteria efficiently. We can illustrate this again with our trusted factory boss. This time he needs to buy a new machine to install in his factory. In this case, he cares about the price of the machine, but also about the quality of the machine. But what is a good trade-off between the two? He decides this is too difficult of a decision to make, so he calls his friend Kristof Van Moffaert. Kristof presents you his new algorithm and lets it learn on some data. Finally, the algorithm offers you several optimal balances in price and quality, for you to choose from. Life for our factory boss becomes a lot easier, no?
In the end, he did not choose to help our fantasy factory boss. Instead, he tested his new algorithms in a different setting. He discovered that they could give him more than 300 unbeatable strategies between the objectives each time he ran it. I know that a PhD is very complicated and long, so this is just scratching the surface of his contributions, but, WOW!
As a scientist, your goal is always to solve a problem. In this case, the issue at hand was that current reinforcement learning techniques we’re not well suited to deal with complex decision making. To that extend, he spent FOUR YEARS trying to solve that problem. In the end, he developed new algorithms and tested these algorithms in both theoretical and practical settings, validating their success. I feel like this perfectly proves my original realization. AI researchers are problem solvers.
Interested in more PhD theses?
If the answer is YES take a look at this page on our website. You’ll find many exciting theses!
If the answer is YES but I don’t understand the lingo, then no worries! This is just the first article in a series. Actually, leave me a comment on this article with what thesis you would like to see me cover next and I’ll see what I can do!
In any case, I hope that I have clarified some of the vital work that the AI Lab at the VUB is doing and given some more insights about the people behind the research.
Who are we?
The Artificial Intelligence Lab (aka “AI Lab”) of the Vrije Universiteit Brussel is the first-established research lab for A.I. in Continental Europe.
Founded in 1983 by Prof. Dr. Luc Steels, it is internationally recognized for its research expertise in reinforcement learning, language and computational creativity. Its active research team has made major contributions to business, combining fundamental and applied research to enterprise. It is included in Reuters “Top 100: Europe’s Most Innovative Universities 2019” ranking and has alumni in MIT, Sony, Prowler and DeepMind.
Where can you find us?
Our website: https://ai.vub.ac.be
Twitter: https://twitter.com/aibrussels
Facebook: https://www.facebook.com/ailabbrussels
Linkedin: https://www.linkedin.com/company/vub-artificial-intelligence-lab
Blog: https://medium.com/artificial-intelligence-lab-brussels
