PAAI - Pattern Aware Artificial Intelligence

WIP games, tools & toy projects
#1

Hello,

I’ve started a new project with a few highschool friends and I for researching pattern recognition inspired by the human mind’s pattern recognition and ability to make use of arbitrary data from any given environment. The prototype will be programmed in Java, and open-source through git. I’m still working out the details, and PAAI isn’t really a justifying name for what I have planned.

When I say “arbitrary data” I mean the data comes from any given source, not that it has no specific form. What I mean is the system will find ways to link positive and negative correlations in a stream of LUA data to find patterns; the program will have ways to format this data, and other modules in the system can react to the patterns found.

If anyone has any ideas, or previous research that may involve this project, please PM me and I will post weekly updates on the project with compiled research notes.

This will be a huge project, and I am open to anyone helping me out.

Edit: Sorry if this doesn’t fit the sub-forum’s criteria.

#2

Sounds cool, if ambitious. I advise splitting “arbitrary data” into at least a few categories and tackling them one at a time, unless you have some specific idea of “generalized pattern.”

Are you familiar with neural networks? “Arbitrary pattern matching” is basically their job description.

#3

Yes, I’m no neurologist but I will be doing as much research as I can to make it as functional and similar to neural networks as possible, which is the point of the research notes that I will be posting weekly.

#4

I dont know how fare and advanced you are, but i recommend to at least skim the following: http://www.dkriesel.com/en/science/neural_networks (ger and eng)
-ClaasJG

#5

Really exciting area - sounds a bit like the Deep Mind thing that was in the news a few weeks ago learning how to play space invaders just from looking at the screen. I think I read somewhere they published some source code for that. Prob worth checking out if so…

The most profitable game AI of course would not be one which played but designed them :smiley:

Good luck!

#6

before you start with neural networks - read up on genetic-algorithm. its a major part in NN.

with a genetic approach you can already achieve interesting things.

#7

Aren’t they orthogonal? Obviously they can be combined, but I don’t think they’re a major part of NNs.

GAs are optimization, NNs are function approximation.

#8

Or one which improved itself :slight_smile:

#9

i was thinking about mutation of neurons for which GAs are prefect :slight_smile: … somehow a NN need to evolve, how would you go about it otherwise ?

besides optimisation, simulated annealing or ‘controlled random’ i think

GA = known result, unknown data
NN = known result, unknown function

#10

(I hope we’re not derailing your thread Ecumene)

NNs have dedicated algos for training that are specialized to exploit NNs’ nature, one of the better known is backpropagation, but GA can work as well, as seen in the link I provided. However I don’t think they are optimal, as they assume nothing about the problem they are optimizing (the NN).

A guy I watch on twitch has been experimenting with GA-on top of-NN for learning to play super mario: http://www.twitch.tv/sethbling/profile/past_broadcasts (the last few broadcasts)
The NN is what plays the game (a function of tile data around mario -> controller inputs)
The GA optimizes this function towards the goal (getting as far right in the level as possible)

Specifically, the ‘data’ is the argument to a function, the return value of which you are trying to optimize, whereas the NN’s ‘function’ is anything.
NNs approximate functions, but themselves do not preform optimization, that is the job of the training algo, which then of course can be any optimization algorithm, including GAs.
In this sense I say they are orthogonal.

#11

aye, was thinking the same :slight_smile: