Hello everyone. Somehow, I don't believe that my other account exists now. I tried logging in, and I was not successful. But anyhow, I used to go by SimDemon. I needed to change that anyhow. =) Now, to my question, what is the purpose of using a neural network? I read much of the article here on GameDev.net that provides an introduction to them. But I did not fully understand what they are used for. A friend of mine is in college, and he actually knows quite a bit about neural networks. But he does not know how to use them. He just has an understanding of their purpose. If anyone could enlighten me by answering my question, I would really appreciate it. Thanks a lot! =)

A Neural Net you give a few inputs - there's a web of weights and at the end a few outputs.

Inputs could be for a gaming domain problem, speed, direction, x location of enemy, y location of enemy - and the outputs could be move forward, back, left, right, fire.

You set thresholds for the outputs (so anything <.5 don't do it, anything >.5 do it).

You train the network by giving it situations and reenforcing good outputs (like firing at enemy when he's infront of the ai player) and in some structures backwardly properagate that through the network.

The weights inside the network hold the knowledge, and let the neural network try and generalise on things without you having to program anything.


So... at the end of the day neural nets you teach rather then program.

NN's give the academics something to jerk off to. Not much use in the real world though.

Alright. I understand neural networks more completely than I did before. The article somewhat gave me that kind of idea, but I was not sure. Thanks, paulecoyote, that helps me out quite a bit! =)

Four Cool Ways to Use Neural Networks in Games

Quote: Original post by Anonymous Poster
NN's give the academics something to jerk off to. Not much use in the real world though.

LOL

You could use neural nets for character or speech recognition.

The input would be some pattern, e.g. an area in a scanned image (a single letter). The net then has 26 nodes, one for each letter A to Z, and each gets the pattern as input. If you input something that looks like an C, the node for the C will respond on this the most. Others, like the one for letter O might respond as well.

This said, you can use neural nets where you dont have exact inputs available.

Yeah, and of course all the other technologies that we take for granted now are just things for the academics to jerk off to huh? No offense, but that was a fairly ignorant statement to make.