[MUD-Dev] Efficient AI?

[Miroslav Silovic]

lindahlb at hotmail.com wrote:

> Generally, neural networks CAN be hard-coded, but this removes any
> ability to train them in an online fashion (help generate better
> AI over time). Also, this would essentially be a genetic
> algorithm, using a large training set to generate a neural network
> that provides a reliable and believable output. Neural networks
> are best used to solve dynamic problems, but CAN be used
> statically.

I should note that static NN is actually a good idea. If your NPCs
keep getting better (and therefore stronger) in combat, your can say
goodbye to any long-term game balance.

>> Can you mix a state machine and neural networks efficently?
>> E.g. use the neural network to select states and then reconfigure
>> the network based on the selected state in order to get more
>> control and higher speed?

> For this, I would probably generate a neural network per
> state. You could dump the results into a database and use them to
> train the networks in an offline fashion.

Check

  http://groups-beta.google.com/group/comp.ai/browse_thread/thread/178b2078ebf3a290/08600b942ab64e35?fwc=1

Aparently the current real-world, high-end AI systems use rules (or
something more appropriate) for preprocessing, neural network for
decision making, and again rules for postprocessing (to ensure
hard-coded requirements).

I'm not sure how you'd go about feeding state machine output into a
NN. Note that using 1 input line per state into a single NN is
actually equivalent to having a different NN for each state. That's
because the 1st layer of the NN can act like a set of AND gates and
effectively split the network - but a single net would be able to
exploit cross-state analogies, if any.

    Miro
_______________________________________________
MUD-Dev mailing list
MUD-Dev at kanga.nu
https://kanga.nu/lists/listinfo/mud-dev