Hi! Was just wondering: Usually an AI (neural net, one that doesnt die(like it would be for gentics alg.)) learns by having something which it gets a reward for... For example shooting the player. I thought about giving the AI a few interfaces, like to a Linux-shell, internet, infrared/radio, screen/keyboard, whatever I can think of.... And it learns (after a while) how to use the interfaces. Is it possible to design a learning AI without a target? What would this look like? How is the process of learning achieved? What does "general AI" have to do with that? Thx & sorry for my bad english Chickenman

An AI can learn anything, provided you include some "fitness functions". Without those, the AI has no way of interpreting data. The functions could do anything, example is any command done in the interface that causes alot of output is good, and low output is bad. All this will do is to teach your AI to do things such as ''more big_file.txt'' or ''dir c:\windows\system32''. It all depends on what you want it to learn and what it is designed to accomplish.

Just putting it at some interfaces and telling it to "learn" won''t do anything except gather data. Now, if thats what you want then great. Let it gather the data and then begin to give it goals. It can search through its "memory" and then solve the goals you want it to.

quote:Original post by Chickenman
Usually an AI (neural net, one that doesnt die(like it would be for gentics alg.)) learns by having something which it gets a reward for...
Is it possible to design a learning AI without a target?

Herr Chicken, there are (give or take) 3 types of learning in the AI community - supervised, reinforced and unsupervised. Unsupervised learns without a target or fitness function

So yes, you can learn

Know what i hate about "AI"? The terminology. A lot of it is vague and high level and means different things to different people. "Learning" is one of those words. You probably want to be more precise about what learning means to you

Most learning in AI is discovery of either an optimal solution or of a classification. A few techniques (the RL ones) supposedly do adaptation rather than discovery but they don''t do it very well

Unsupervised learning is where you give it a bunch of items and the algorithm basically groups the items. All the red shiny things go over here, the blue heavy things go over here, etc. Data mining and collaborative filtering are unsupervised. In a game, they might learn that certain patterns of movement are basically the same (say circle strafing or flanking). No one gives them feedback on it, it figures it out on its own

But what you can and can''t do really depends on your ability to define what precisely you''re looking for

-baylor