Submitting note: I was surprised this hasn't been posted on HN before, since it's one of (maybe the?) earliest examples of proving how a computer can learn from failures over lots of training runs to gain "intelligence", as well as being a relatively easy to understand explanation of how that can work, so it seems interesting historically considering the current state of machine learning in the world. It's also the sort of thing you could make, and train, yourself if you're bored one weekend (or if you have kids I dare say there's an age that many could love doing it and learn at same time).

Something that has been submitted a couple of times before by users jordigh and JNRowe (but only getting 40 points and 3 comments[1] between them) is the website of Matthew Scroggs, who both writes about the physical MENACE he built and trained, and how to do that, and also provides a digital MENANCE simulator he built for anyone who wants to play around it without actually needing a few hundred matchboxes.

His MENACE main page is here: https://www.mscroggs.co.uk/menace/

[1] 3 previous comments on submission to the mscroggs page: https://news.ycombinator.com/item?id=30491888

also featured some years ago in a Matt Parker video who presented it on a math fest https://www.youtube.com/watch?v=R9c-_neaxeU

see also https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...

Ahh yes good point, thanks!

It's a shame it doesn't seem to be getting much voting traction, as with the previous couple of submissions to the MENACE simulation - I feel there must be interesting discussion to be had around the journey from this sort of ML to current state of the art and I was really hoping to see people smarter than me talking about that :(

Not sure if it applies rotation and reflection, but if not it is missing a trick. That reduces the possibilities a lot. There are only 3 opening moves in reality.