I had an idea and I didn't know where to go with it or if it would go anywhere or how to explain it, so here is is.

I had an idea for a mapping bot for Osu!lazer in the future.

For Players:

It would take data from the maps people play and try to create a map within 7 days (not sure how song selection would work). there would be a survey asking how well it made the map, everyone would give feedback to the bot (nothing super complicated, just simple multiple choice question), and then the bot would try again. The questions will be simple, you go to a timeline on the map (select starting point and ending point to the place you want to give feedback), and you would select an option for the bot to review. The simple choices would be something like:

-It was/wasn't timed properly
-It was/wasn't representing the music well
-Did/Didn't follow the drums/beat

Stuff like that.

For mappers:

Standard:
before you start mapping, you'll have the option to have the bot watch you map. The bot will try to recognize
simple and complicated patterns to use in the map it's currently mapping. the first hard part in this whole idea is how the bot's going to recognize and utilize the patterns. One problem is the slider placement (when to use what slider where and when, how long or short it will be... etc.). While the bot's watching a player map, the bot will calculate best possible patterns and placements for the map.

CTB:
Not gonna lie, CTB will probably be easy, the only problem might be where to place the rain. I don't have much for CTB because I don't play it much, sorry.

Taiko:
One problem with taiko might be when to put a hitsound on a note, there could be many possibilities to apply a hitsound when and where. I don't play Taiko much too, sorry.
If there is any way for the bot not to make a map like a convert than it would be absolutely fine.

Mania:
For mania, I guess this all sums up to when where and how to place good notes and sliders (considering there is slider dedicated maps), although it'll find that out by watching the players map.

I'm sorry if I missed anything or got any info wrong, I'm not perfect and any help will be very appreciated.
I'm also very sorry for making this very long.

Have a good day!

-Fontdyla

Soooo... a map generator? possibly AI based?

abraker wrote:

Soooo... a map generator? possibly AI based?

kinda...

I think there are two independent steps you could break a "mapping bot" into:

1. given a song's audio file, identify at least with reasonable accuracy all of the timing, instruments, and sounds that those instruments make throughout the song. basically raw audio => sheet music. this is probably insanely difficult to get right and not in the scope of a fun project like this, but on the other hand there might be significant research done already since it's not specific to osu
2. knowing all of the info about the song (the "sheet music"), construct a beatmap

the second part is more relevant so I think that's the part worth focusing on. if you wanted to make it a task of machine learning, you could start by having experienced mappers make maps of songs that you already have the "sheet music" data for, and then train based on those examples + feedback... idk if that would really work though since your example set would be so small

however I wouldn't rule out the possibility of a deterministic solution, at least for some game modes. with the knowledge from many taiko mappers for example, I'd bet you could start to build some "rules" for a mapping bot to follow when creating a taiko map. you could put together a database of patterns and match them to rhythms/sounds/etc in the music, so your bot knows what to do, and add some element of RNG to decide among patterns when multiple could apply. you could add some bonuses that would encourage the bot's algorithm to favor mapping motifs across the whole track. just some ideas, idk how practical any of this is lol

this is the kind of thing I'd expect someone to pick up as a research project one day. iirc popner tried a simple version of this a long time ago

I guess this could be partly archived through a Tensor Flow Model, that learns how existing maps correlate to song, then the AI can try to recreate the map part for any song you throw at him.

I'm not deep enough into Machine Learning to know if there are big problems with this approach, all I'm saying is that it could work this way.

To the Idea of making the Bot review differnt map portions, this would be solved this way aswell, because you could just throw music snippets of the section at the bot and see what it would do differently from the current mapping.