[pythian] Re: PROJECT - Combat (AI Update)
- From: Darryl Long <pythianproject@xxxxxxx>
- To: pythian@xxxxxxxxxxxxx
- Date: Mon, 24 Jun 2002 22:47:15 -0700
Robert Fakes wrote:
>It usually (if not always) takes alot of training to get a NN to act
>reasonably. Sending out an untrained NN is probably not an option.
>
>Having it learning once it is shipped may not always be a good use of
>resources, for the same reason. Depending on the algorithm, it may never
>change significantly during the games useful life on a particular machine.
>
Sadly, it's all true. Neural Networks suck for learning in-game.
The only reason I'm still toying with NNs is because a very simple NN is
equivalent to Mike's voter system plus my modifications (I'm fairly
pleased we invented the Perceptron, even though they originally did that
50 years ago). I'm hoping I can train and use one to direct the AIs
action in a smooth, organic manner. A NN can keep track of a lot more
inputs at once than I can with state-machine coding, so hopefully it's
actions will seem more natural (rather than abrupt). Once we've trained
the NN ourselves we'll ship it out with learning turned off. At that
point, it's just a complex decision maker.
Darryl
Other related posts:
