sry, i replyed to Aurelien directly...

Aurelien Derouineau wrote:
> Just out of curiosity, how come you didn't use an already existing
> project for the network? I guess this leads to the question: do you
> think your ANN code is as well optimized as the other ANN projects out

there is only one big reason for doing this... it was a school project
;) on the other hand, i compared the code i made to the others nn's and
noticed the training process of my code is much faster and more
decisive. ie, i tried to train the other nn's to learn the xor function
or an identity function f(x)=x with minimum nodes in the hidden layer
(eg, 8x3x8) with little or no success... the training process is the
only real difference between all of them so once the training is done,
calculating nn's output is really easy.

> there? Or do you think it would be better to use such an existing
> project (if you'd like to "extend" your project to the public domain)
>

shure... use of my nn code is not mandatory. this plugin is mostly some
kind of proof of concept :)

> Also, I was wondering which would be better between:
> 1) Have the NN code written in all the plugins (all as in for all the
> games) and have the game server run the network as well as the game.

that is something i was trying to avoide... no
hacking/cracking/whatsoever on the client side.

> 2) Build some kind of public project (maybe a bit similar to
> steambans) where game servers would record game data periodically for
> x minutes, and then send that data to the project server. This data
> would then go into a queue, waiting to be analyzed by the neural
> network on that project server. This server would contain as best as
> possible the different patterns (cheating, non-cheating) for the
> different games. After primary analysis, if a player is clearly
> thought by the network as cheating, his id would go into the project
> database. If it's not clear, the data is re-sent to the queue to be
> reanalyzed. After 2-3 attempts, if it's still not clear, the data is
> put in a queue for a human to watch the demo (a demo would have to be
> recorded as well) or just throw out that data. Such project would help
> servers make sure the error rate is as low as possible, by feeding it
> more known data as more is gotten therefore "eliminating" the problem
> of out-of-date patterns.
>

this sounds very interesting but it grows out of scope for me as i don't
have the time for such a project (my graduation is very near). if anyone
whishes to pickup on the idea/code, please do... i would be very pleased
to see something like this in action :)

> I hope I haven't missed too many points and that it's not too hard to
> read and understand. I tend to express myself in complex ways :p
>

not to worry, me too :)

--
I doubt, therefore I might be.