The problem is that just like any other optimization, generalization to other settings is then limited and not necessarily predictable, potentially leading to much better or worse performance. That's the difference between similarly colored bars in this plot.

Monday I defended my thesis - with both words and swords! Luckily I did well enough in both to pass and survive the ordeal ⚔️💪

We designed a path choosing game to show people how one could go about solving an MDP. You can try however long you want to find the reward structure.

Hilariously, the older the people, the longer they took. Some small kids got it first try, older folks took up to 20 minutes 😅

- today we spent a few hours touring some of the AI labs in Aachen (of which there are many!). It was great to get demos and insights from diverse perspectives - plus they let me walk a spot! ♥️
(they only asked about my insurance afterwards...)

- 2/19 posters were on AutoRL and there were a few more on RL, so at least in this group of researchers, RL is alive and well

I spent an awesome two days in Aachen as one of 19 young researchers selected as the "Future of AI" in Germany.
It was a cool concept and very well organized event. Here are my personal highlights:

Brussels, arrived 30 minutes ago, already got our beers - #EWRL is kicking off nicely!