Single notebook Reinforcement Learning for a local variant of the card game "Watten" in its south tyrolean form. It includes a simple evolutionary architecture search and is parallelized quite well.

This was written in a few afternoons after our professor of "Artificial Intelligence" claimed that current machine learning, while being able to learn chess or go cannot learn 'complex' games like watten. I wrote this just to convince myself that Watten is in fact quite easy to learn for a neural network to learn with its frequent and well interpretable rewards and relatively simple rule structure. Even without extensive optimzation or pretraining.
The resulting networks migth not be superhuman at the game but still work really well.

Link to a Colab-Notebook will be provided soon.

Watten (english but not all south tyrolen rules explained): https://en.wikipedia.org/wiki/Watten_(card_game)
Watten (german with pretty much all the rules): https://de.wikipedia.org/wiki/Watten