Sign in to follow this  
Mizipzor

Attempt on my first neural net

Recommended Posts

Mizipzor    247
Ive trying to do this quite some time now. Have written a thread about it before but didnt really get anything out of it. Anyways, heres the scenario; a helicopter should learn to fly. Hehe not in a full dynamic 3d world of course. Think of it rather as a pixel that should stay in one place, the rest is just fancy graphics. My world where the helicopter will fly is even one dimensional, it can only move along the y axis. So, putting this into a neural net it should look something like this; inputs are the helicopters position (how many meters above the ground it is) and its vertical speed. The output is then a binary swith, should the engines be running or not?
height above ground 0--->
                          O ----> engines on/off
vertical speed      0--->
It cant really get any simpler than this. But still I cant manage to understand how to do it. Each input is multiplied by the weigth on that node. If all the calculated inputs combined are greater than the threashold, the output fires.Then, after each "run", or calculation, the weights are updated (evolved) with new, better values. Am I right so far? Can this be coded by someone on my level or have I missed something, like the hidden layers of neurons I skipped cause I wanted it simple, that makes it just doesnt work? I guess the big question mark is how the weights are updated. The neuron must include some kind of function to determine how successfull it were. Do I need to hardocde that? Because I know, of course, whats the AI is supposed to do. But is it possible to make it so that it isnt hardcoded? If so, then the AI would be able, to some extent, to relate to (understand) the worldwide phenomene of cause and effect. Im reading all the articles and websites I can find on this matters but still I cant get my head around how it actually should work. I found a very nice explanation here on gamedev.net. Some user wrote about it in his blog/diary. Some space shuttle that tried to land on the moon, learning the right speed dependning on how much it weight and such. Have forgotten where though, anyone knows which user Im talking about? Gonna stop writing now, otherwise noone will bother to read :P. But please tell me, what am I doing wrong? Which crucial part have I missed out? Why doesnt the "big picture" click in my head? Maybe 'cause Im to darn stupid? :P

Share this post


Link to post
Share on other sites
Cyprus2k1    122
hi mizipzor,

i guess u need to do a training set for the net, that is , a set of inputs and outputs so ur net can "learn" to compute the function necessary to do what you want.

there are several ways the weight can be updated, usually it is done by BackPropagation ex: http://www2.psy.uq.edu.au/~brainwav/Manual/BackProp.html

hope it helps ;)

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now

Sign in to follow this