Neural Networks and GalCiv

from Stardock Forums
as i'm hacking away at my own game design project, i'm learning various aspects of AI and ways to tackle AI related problems. One thing i've found very interesting is Neural Networks.

Does GalCiv use NNs and if so, to what capacity? [What would be a good resource/book on NNs?]

If it doesn't, it might be something usefull for building an *even smarter* adaptable AI. NNs basically take a slew of variables and, after training it properly, can predict and/or solve complex problems based on previous experience given a set of data. very cool.

if there is any interest, i've linked my project site as well.

http://www.project-axis.net/
303 views 4 replies
Reply #1 from Stardock Forums Top
AI networks can learn, but I'm not sure you can use them reliably for a game.
There are 2 kinds of NNs:
Those which you train, and those which learn by themselves.
The first will be useless once you mod the game, issue a patch or whatever. You'll have to retrain them entirely.
The latter learn more or less well, and you don't know what they do until you see them at work.
Additionnally, they provide predictable results, so you may have to add some randomness outside if you don't want players to know exactly what your ai will be doing next.
I wouldn't use neural nets for a complex thing like strategy planning. They are very effective at shape recognition, so you may try to put them in so that the ai would recognize a threat, or a pattern on a map... You must understand that a NN is effective only when given appropriate input, and that professionnals who use NNs start by removing everything irrelevant and shaping the data so that the NN will not drown in noise (f.e. in image recognition, they may make edge detection, shadow removal, whatever, before feeding the image to the NN).
So my opinion is don't use it. I don't for my game, and don't know of strategy games which use a NN ai.
Reply #2 from Stardock Forums Top
Hi,

NN have been used in games, I believe the first commercially released game to use one was named Fields of Battle.

Personally I wouldn't recommend them for the reasons mentioned above, plus my experience from FoB. Nice idea, but didn't work.
Ashes of the Singularity II
Reply #3 from Stardock Forums Top
They're just function fitters, and not infrequently require substantial care choosing inputs, outputs, and structure.
Reply #4 from Stardock Forums Top
Well, if they plan to analyse all the save games we've sent them, a neural net's just one of many machine learning techniques they could use. (These ones working on save games could learn by themselves, but are limited to predicting a player actions given a player's playing history.)

There's two different things for an AI to do: put together a full picture of the opponent's situation from limited information, and then figure out the best action to take. Two very different processes could be used for those two things.