To code neural net or not??

[Chiblue]

I have been working on my AI control mechanism and have gotten most of the functions I want to use working but now I hit a dilema and wonder if anyone else has been down this track?

 

I have a multi-layer, multi-resolution graph A* path finding routing that works very well for guiding a bot to a location in my scene, and thanks to help from you guys I have now got a inside building raycast mapping engine that builds a layer onto my exisiting graph with entry/exit links and works very well, thanks for the help.

 

Ok my problem or design consideration is this.... I want to addin the ability to control a group of bots that will move in a formation of sorts, but the question here is, I have 2 options, 1 I take on bot in the group as a leader and have the other bots maintain thier relative position to that bot in the formation, or I build a neural net control mechanism for all the bots to get the the desired location (formation relatively speaking). The reason I am asking this is that the option 1 would resuilt in a formation movement but will probably appear very mechanical, were as the neural net machanism will be more of a natural motion. I basically have options 1 and it is sort of a mechanical motion, but I would be interested in any experience anyone has had in doing this?

[Pixel Perfect]

I don't have any experience with Neural Net techniques for AI although I'm fully aware that various people have experimented with these. The thing is, I don't see any evidence of them being used widely in commercial games which leads me to believe there must be some fundamental issues with the Neural Net approach.

 

Would certainly be interested in seeing what progress you make if you attempt this!

[Chiblue]

Actually the coding is relatively simple compared with a A* graph or even moving entities together... the big difference with a neural net approach is like a neuron each bot on cares what it's nearest neighbors are doing... so rather than all following one, each ones behaviour is based on that of those bots within a sphere of influence... you have a lead bot and the others are basically connected by being within a sphere, each bot instance has code that defines it has the leader or not. If it is the leader then it follows the current behaviour, if it is not the leader then it follows those around it... so basically any action trickles down from the lead bot to those next and then the next furthest and so on, until it reached the furthest bots... by building in a small time delay for each bot to respond to the others around it (actually those nearest within the sphere of influence) you end up with much more organic flock motions. this is becuase the impact of the behavior by the lead affects all others over a space of time, not immediately, so in my scenario when I change formation it happens over time not immediately...

 

Hope this makes sense...

[Chris Paulson]

I would stay away from neural nets, they are for academics not real world games.

 

Check out: -

 

http://gamma.cs.unc.edu/RVO2/

 

http://digestingduck.blogspot.com/2010/10/rvo-sample-pattern.html

 

Also:-

 

http://aigamedev.com/insider/discussion/motion-planning-thoughts/

and

http://aigamedev.com/open/articles/simulating-crowd-flow-dynamics/

 

but you'll have to register to see these, but if you're doing a game with real game play you should already be registered...

[Pixel Perfect]

Hope this makes sense...

Yes it does, but it doesn't sound anything like a neural net but more like a close feedback system with some damping built in. Are you sure you were meaning a Neural Net in the first place (i.e. a simulation of the firing patterns of neurons in the brain which normally involves lots of teaching in order for it to learn to function in a desired manner) or more of a analogy to the physical arrangement of neurons?

[VeTaL]

I can help only with this link i saw some weeks ago:

http://www.codeproject.com/KB/graphics/GPUNN.aspx

[Laurens]

I recently came across this article: http://www.ai-blog.net/archives/000178.html

[Pixel Perfect]

That's an interesting article Laurens. I knew neural net type technology was not being used much in Games Houses, as I'd previously pointed out, and that goes some way towards explaining why. Very interesting!