Image Recognition

Recognition of images, letters, and geometric reconstruction is a huge area of science and technology. You can check Computer Vision online library to get an overview:
http://homepages.inf.ed.ac.uk/rbf/CVonline/
Some of methods are very fast and work even on the cellular phone with camera.
But the only way you would really need it if you AI control a real robot in the real world. If your AI is acting in the simulated enviroment you can get the same behavior a lot more easy if you just chek the line of site, distance to object, introduce artificial errors based in them and just send coordinates and identifiers to AI.

Search for those keywords:
pattern matching,
pattern recognition,
content-based indexing,
machine learning.

Some good journals/conference with papers on this subject:

journals: Journal of Pattern Matching, Artificial Intelligence (journal)

Conferences: ICPR, CVPR, ICCV

The 'popular' approach right now for image recognition is through "support vector machines". Tons of papers on this, it has rised in favor while people are abandoning connectivist methods.

Good luck!

edit: serg3d is right by the way, its a HUGE area, be prepared for a LONG litteracy review.

Thanks for the replies guys :-)

This will take all your RAM[random access memory] and all of allot of disk space.

Another thing to note is that your solution will be *very* specific to the game your AI is playing. Pong, for example, is relatively simple, you only need to recognize 3 simple objects (1 ball, 2 paddles), this can be made easier by using different colours for each object. Getting an AI to play Doom 3 this way however would be nigh impossible without a supercomputer.

lol dont worry my ai wont be playing doom anytime soon. I know i could do some simple solutions with pong, for example i could test only two columns to find the top left x and ys of the paddles and then just look for the rop left ball coordinate. Actually thats what ive started to do but it feels like my ai is cheating.

If you're only using one font, you can use a simple pattern matching routine (cross correlation) to recognize letters.

It doesn't have to be anything complicated, since you're not matching unknowns-to-knowns (like face detection), but are mapping knowns-to-knowns.

The more I think of it, the more I think a cross-correlation algorithm will work for you.

Like joanusdmentia said, this will be very specific to one application.


Almost forgot: What are you working on??? Can we see it? It sounds cool.
Will

If i ever finish ill post something, been busy lately though.

Quote:

Actually thats what ive started to do but it feels like my ai is cheating.

How is that? If you want a realistic AI for a game like Pong you should be modeling a delayed response like humans have, and including a margin of error on where it predicts the ball will land.

If you force the computer to 'view' the game the way the player does you'll never have a competetive AI on any 'real' game. That type of 'viewing' wouldn't even be desirable for a game AI.

Hes making a bot for a game!!

Quote:

Original post by Cubed3 Hes making a bot for a game!!

I had an idea to make a bot for Counter-Strike like this before. Just combining an aimbot and an image-recognition AI to run around the map and navigate for you.

Then you could just sit back and watch it do its thing.

But the project turned out to be harder than I thought it was going to be.

Someday!

One thing people keep forgetting about the human brain, is that we can't really process a full image of data 30 times a second. If you tried viewing a slideshow that way, it'd be a blur. Our brain with find components in an image, then assume it's either in the same area, or is moving to a predictable location, then look for that same object in one of the possible new locations.

Your Pong AI could look at the screen, find the paddles and ball, then predict possible future locations for each component. The AI's paddle is very predictable, the ball only changes direction when it strikes a paddle, so the only truely unpredictable component is the oponent's (human's) paddle.

If you want to be tricky, the human brain, at first, thinks that each component can move in any direction at any time. But after a while, it'll learn what the boundries for the movement are. Give your AI the ability to use "experience" and it'll be much more flexable in the future.

How would you give your AI the ability to use experience? What you mentioned now was only a specific case where I guess you have to program your AI to learn that a paddle often keeps going in the same direction as before. But how do you use that fact when it is discovered? And what if you want your AI to be able to learn more things than that?