Neural Network(the evolving kind):
If you're not sure what I'm talking about, take a look at this: http://www.youtube.com/watch?v=lmPJeKRs8gE
well, you guys think its possible?
ID:151480
 Sep 18 2010, 7:08 pm
|
|
|
Sep 18 2010, 7:58 pm
|
|
Garthor
 |
Yes.
|
To me that just looks like normal path-finding with the robots always taking left turns when possible and avoiding the higher cost grass. That is up until the end when global collisions are avoided.
| |
The entire point is that it's all "learned" behavior from a very, very simple set of rules. The concept is that you can take a task that is difficult to encode (such as pathfinding for actual cars in real life) and simply generate it with a few simple rules and a lot of processing time.
It's too late for me to explain exactly what's going on. Maybe tomorrow. | |
I like this, if they would add data to a text file and it was made into a library then people could set up the rules for their AI to follow and have them do a few runs in the background while you do other stuff. In the meantime, there's a text file including locations, deviations, walls and shortcuts, compiling up in the background.
After a few hundred runs you'd have a text based map saved and accessible to the AI so they can make their way from one paint on a games map to another, completely avoiding obstacles. If you think about it, this is the way that Bethesda probably codes their AI to handle terrain. | |
Even at the highest generation, that AI seems pretty retarded. If its somehow teaching itself, with no human interaction, from generation 1-END, then I guess its somewhat impressive, depending on how its working exactly. But I could write more functional AI than that in like 10 minutes. And it doesn't even seem to have any path finding, it just seems to STAY ON THE ROAD as best it can. I sure as heck wouldn't ride in a car driving like those things.
| |
It's not suppose to have path finding. It's suppose to learn things depending on the mistakes the previous generation made. It learns based on the problems on previously encountered. This is not something people right in 10 minutes in my opinion, especially the more complicated ones like so http://www.youtube.com/watch?v=1iamM0SuPto (bad example really but i couldn't find the others), its not just a plain AI, it evolves based on a set of rules, not solid code. :P
EDIT: Much better example would be http://www.youtube.com/watch?v=lSG--GY2p2o&feature=related, but its not a neural network, its based on the genetic algorithm which i frankly have begun to find more interest in. EDIT2: I FREAKIN LOVE THIS ONE I WANT IT D: http://www.youtube.com/watch?v=lSG--GY2p2o&feature=related | |
Anything a Ph.D. student from MIT can do Falacy can do betterrr~~~
Falacy can do anything better than themmmm~~~ | |
So you could program this in ten minutes? (Hint: They are talking about the Euphoria engine)
| |
Popisfizzy wrote:
Anything a Ph.D. student from MIT can do Falacy can do betterrr~~~ I lol'd. Falacy wrote: In this case, sure. In most others, probably not. Saying "Hey, I could write pathfinding code that works better than a neural network which means I'm better than them" is a bit ignorant. You're talking about two entirely different things, and completely missing the purpose of a neural network. Go read about neural networks... They are intended to perform a task without being explicitly programmed to do so. Instead, they are given rules and goals and they figure out how to complete their tasks without being directly told... | |
A neural network is not "given" rules, or goals, or anything of the sort. A neural network is a complete, weighted, directed graph, with some vertices being labeled as "input" and given values, and some vertices being labeled as "output" and having their values computed.
In this case, 15 input nodes were based on the sensor data and 2 output nodes were fed into the "wheels" of the vehicle (though I'm guessing they behaved more like treads). | |
Garthor wrote:
The entire point is that it's all "learned" behavior from a very, very simple set of rules. Garthor wrote: A neural network is not "given" rules, or goals, or anything of the sort. I KNEW Garthor could fail. | |
<s> I think it's more a case of being ambiguous than being strictly self-contradicting. </s>
Edit; as per request, Garthor is clearly self-contradicting here. | |
Murrawhip wrote:
Garthor wrote: Two different meanings of the word. The neural network cannot process rules such as "do not drive on green grass". However, it DOES operate according to the rules of the network: input goes in, calculations get done, output is extracted. Those rules are ones of simple math and basic permutations, though, not of high-concept rules that you'd expect from an AI. | |
Garthor wrote:
Two different meanings of the word. The neural network cannot process rules such as "do not drive on green grass". However, it DOES operate according to the rules of the network: input goes in, calculations get done, output is extracted. Those rules are ones of simple math and basic permutations, though, not of high-concept rules that you'd expect from an AI. My use of that terminology was to imply a situation wherein the program would receive data, be given a calculation to determine if the data received was good or bad (a rule. I.E: You earn more points for travelling more distance), and react accordingly. Anyway, I have very little knowledge on this topic so my impression of how it works could very easily be wrong. What it appears to me is that the 'car' is given the instruction that the faster it moves the more points it gets. As the car begins each simulation with the intention of earning more points, it will eventually figure out that driving on terrain that does not slow it will yield the best results. Again, don't know anything, could be wrong, etc etc. | |