Kitty Corp Meow Mix Forums

Hi there,
My current goal right now is to make an ai for smash that learns. I've currently written a project draft and task list of things to do. and the theory is solid as long as i can figure out how to learn about brawl physics in general.

     I'm gonna send an email out tomorrow with a formal explanation on another board here. The gist of it is this : I just took a class on introductory artificial intelligence theory and practical techniques and now I believe that a learning ai can be achieved (since the non-learning ai is just a reactive agent but still finds a solution) and its only a matter of understanding the mechanics of smash accurately.
I apologize if any of this sounds confusing. I speak a lot of math plus abstractions and sometimes don't catch it.

But the three main problems of this project are this:
1. Get real time perfect information of the game during the game. (Create the sensors of the ai)
     This is to get the ai the information it needs to count features of the state he's in so he can wait which sequence of actions is optimal. This will be explained a whole lot more tomorrow, as this is a pretty vague overview of what this problem actually consists of. The project hinges on this step. The instant we have access to a body of the game and the context of the smash world, we can then build the brain.
2. With knowledge of the state, compute which features of the state contribute to getting a good score (Approximate the best response to the now)
    The method I'm going to use commonly takes 1000's of games possibly, so it'd be a while. But luckily the standard-ai is reactive so it can duplicate its behavior given the same exact scenario (or not if they used classical planning, mini-max, research those terms if you want to know what they are). This means i can automate the ai's training until it beats level 9's. By that time i have a plan to imlpement different learning scheme's for dynamic behavior against a certain player (average response to specific state-scenario's, learn how to predict a person's common reactions). This is on top of the layer that learns a general way to tell for every player what is the best way to move.
 Brawl is complex and so a lot of variables go into playing intelligently. That might make the convergence of this project small. So we would have to get the parameters right.
3.  Map the brain back to the body to complete the feedback loop. (The actuators of the ai, its arms and legs, or in this case fingers and thumbs)
    This is always the easiest step. When programming ai, this step comes after the brain and it literally feels like a walk in the park.

TL;DR I'm gonna ask for help making an adaptive ai because its 100% possible, keep a look out if you're interested in joining the effort. The ai will get the ass kicked out of it and then become super saiyan, i watched it with my own eyes after fighting my professor's ai for about 500 simulations. If you got this far thanks for listening.

Also to finish the introduction, I found this place while doing research and you guys seem like you're talented. I do not have experience doing nitty gritty hex hacking (i know that isn't an official term) but I have a lot of experience with logic and machine learning and programming, I just don't hack because the theory part and required classes for school get in my way to do fun stuff. So i'm gonna need help getting the information from a match while it's going on because i have no idea what is a possible way of going about that.

I'll be posting this on other boards as well depending on how all the recruitment goes.

Thanks again!

This is a thread to say hello and stuff, not to ask for help. No worries, a wonderful and probably manly moderator will move this to the correct area, like a super hero, on a friday.