I gave a talk a the AI group’s seminar series on a bit of work I recently did on modelling the magic hexagon problem.

There is a Colab notebook that shows the new model I developed [here](https://colab.research.google.com/drive/1Hjf3v6v3WOd4Q4uiGGRNJ56KZwnQNaYk?usp=sharing.

This work was prompted by the curiosity of one of our VIP students, Stella. Moreover the visualisations in the notebook above are done by Stella. Thanks!

There are a few interestign findings which are not in that notebook. Specifically, turns out linear relaxations in the OR Tools CP-SAT solver are extremely effective for this problem, which was not something I had expected. I did some investigation into why this may be, but I need to do a bit more yet. I might try to write this as a paper if the underlying reason turns out to be interesting…