I'm working on a project that might interest some of /vr/
We're developing a neural network that can learn how to play SFII on the SNES (very similar concept to MarI/O). One of the first steps is to determine which memory addresses in the emulator correspond to things like X position, Y position, lifebar, etc.
We found an incomplete map here:
https://www.smwcentral.net/?p=viewthread&t=33039
But that's for the arcade version and it seems like the memory addresses used are different. Does anyone happen to know some obscure forum where we might find this information?
Why don't you find them yourself, nigga. Get Cheat Engine or something.
Wouldn't you rather chose a simpler game, like Super Mario Bros or something. The CPU is a cheating bastard in SF2.
>>3949180
>The idea is to train it by having 2 neural networks play against each other in VS mode.
Oh ok that's cool, makes sense.
>>3949180
>train it by having two neural networks fight each other
You might find that some of the learning it does will send it down the wrong path, putting you further away from beating the computer
I would ask on shoryuken's forums.
God speed, OP.
isnt this more or less what saltybet is
>>3949929
Doesn't look like that at all, anon.
>>3949672
True, we're expecting that we'll have to intervene if it finds that the best strategy is spamming low kicks or hadoukens in the corner.
If we find a good way to do it we might also post some sort of website where random people can help train the AI.
>>3949185
Good idea, I'll post a thread there too.
Have you asked in the ROM Hacking General? They might be interested in helping.
>>3949180
I wouldn't go for that route, since you can take a lot of time for the stabilization levels of the NN do something useful.
I would rather capture inputs of skilled players and fed those + the overcomes to the neural network. You should be able to bypass teaching them the basics.
Which model are you guys planning to use?
Not trying to rain on your parade, but wouldn't an AI knowing about the code itself be kinda omnipotent at the game? It'd get into an unbreakable winning streak pretty soon.
I know the alternative is not viable (movement recognition), but a neural network for this is kind of overkill.
>>3949160
If it's vanilla/HF you're probably just going to end up with Guile vs Guile, or in the latter case Guile vs M. Bison, matches.
>>3954060
Dic is irrelevant in HF.
Also, this project sounds like characters are picked by the player and then AIs do the rest.
>>3950779
It's not so much that they might cheese their opponent but that any iterative learning process becomes more and more restricted as iterations progress, like how evolution works, you end up subject to what has come before
Works fine for learning how to beat Mario, but I have doubts about beating the ai at a fighting game
>>3954035
>did I just swallow a meme
What's the disclaimer about what you read on 4chan anon?