Closed samifruit514 closed 8 years ago
Yeah, the current model in trunk for some reason is generating puzzles only in a very narrow band of difficulties. This means that if you're trying to generate puzzles that are outside that narrow band it will keep on trying for a long time until it randomly hits on one, which can take basically forever.
This is something I've been digging into for the past week or so. See #253 for more information (and issues like #254 and others for work I've been doing to help me dig into the problem and understand it). Unfortunately whatever is causing the underlying problem is pretty persistent now even as I train new models using the freshest data, which implies it a long-term sustainable fix will take some more investigation.
I didn't actually realize anyone else was relying on trunk's difficulty model being in a good state so I was taking my time digging in and fixing it. I was just going to temporarily revert to an old hs_difficulty_weights.go model from January, but I just spot checked it and found it also had a weird band of results. I'll dig into this tonight to see what the best short-term approach is to get the model to a better place.
By the way, thanks for filing an issue--please feel free to provide any feedback or ask any questions you have!
I'm going to track the underlying problem and my investigation in #253 . For now I'm going to close this issue and encourage you to follow along there.
If you want a short-term solution while I track down the solution, checking out an old version of the repo and using that should work. 758540e833 should be pretty safe.
My investigation in #253 implies that a good short term solution for you is to use the 758540e snapshot for now to generate puzzles. I think I might know what the underlying problem is; tomorrow night hopefully I'll be able to implement a long-term fix (assuming my hypothesis is correct)
Is that normal? Thank you