Hi ,
I'm doing some work with UT2k4 at present . Im using java bots in teams where they use a neural net ( amongst other techniques such as gaussian mixture models and dynamic path finding ) and have found that you should be fine to train neural nets on data on the fly within very tight time constraints as long as you dont train using back propogation but instead use either scaled conjugate gradients or quasi-newton methods . I use scaled conjugates because they are simpler . The toolkit that is have been using can be found at - http://homepage.mac.com/jhuwaldt/java/Packages/NeuralNets/NeuralNets.html . If your serious about using neural nets for this purpose then you shouldn't let anyone tell you its not viable , cause it is . As for lisp being " THE language of AI " , thats a load of bull . Firstly its use for numerical methods is limited ( This is a big downfall as many numerical and statistical based methods in AI require advanced linear algebraic and calculus applications ) , secondly its only big in certain countries , for instance Prolog is much bigger in britain and japan for similar applications .

Quote: Original post by Anowrexiya
Hi ,
I'm doing some work with UT2k4 at present . Im using java bots in teams where they use a neural net ( amongst other techniques such as gaussian mixture models and dynamic path finding ) and have found that you should be fine to train neural nets on data on the fly within very tight time constraints as long as you dont train using back propogation but instead use either scaled conjugate gradients or quasi-newton methods . I use scaled conjugates because they are simpler . The toolkit that is have been using can be found at - http://homepage.mac.com/jhuwaldt/java/Packages/NeuralNets/NeuralNets.html . If your serious about using neural nets for this purpose then you shouldn't let anyone tell you its not viable , cause it is . As for lisp being " THE language of AI " , thats a load of bull . Firstly its use for numerical methods is limited ( This is a big downfall as many numerical and statistical based methods in AI require advanced linear algebraic and calculus applications ) , secondly its only big in certain countries , for instance Prolog is much bigger in britain and japan for similar applications .

1) Mixture models are awesome.
2) viable doesnt mean its a good idea
3) Lisp is more "THE language of symbolic AI". Depend on your personnal definition of AI I guess. For a lot of people, including me, numerical methods arent AI, theyre numerical methods.

Viable doesnt mean its a bad idea either , i was just saying not to be told it wont ever work and that you shouldnt try it out . As for numerical methods not being AI , symbolic AI is prevalent in some institutes but in others there is none and everything AI is numerical and sub-symbolic , it just depends where you are . I don't think you can throw out a blanket over all of AI as either .