Re: [MUD-Dev] Neural Networks as the AI system for a MUD?

[Hans-Henrik Staerfeldt <hhs#cbs,dtu.dk>]

On Thu, 2 Dec 1999 PartyG2816#aol,com wrote:

> The subject says it all.  I'm out researching various AI tools and i hit upon 
> NN's. From the bit I have read, they seem rather interesting, and very well 
> suited to a MUD's needs for AI.. But I still haven't read far enough to 
> really /know/ if you know what I mean... I thought maybe some of you more 
> experienced people on here might have used em or know ab it about them.. 

Artifical Neural Networks are a bit simplistic for anything i would use
for AI in a mud. Perhaps the neural net could be used as a tool for a 
rulebased AI, where the NN is used to try and identify 'situations' where
a given ruleset would apply, but i cannot see how NN could be accounting 
for AI by its own.

Specifically you need to 'display' the events in the mud intelligently to
the neural net. The better you port your event into the 'problem domain',
that the network is trying to solve, the better you probably will get the 
network to perform.

Another problem is CPU power, as someone else also commented on, training
or using (which is less demanding than training) the networks of thoustands
of NPC NN's would be very demanding, even on today's hardware.

What you perhaps should think very hard about, is where in your AI you want 
to use NN for in your AI. Remember you do not get 'problem solving ability'
just because you interconnect a few thousand neurons.

As before, you could try using the neural nets for guessing wants/needs or
threat/non threat, friend/enemy and other hard-definable feature, and use 
these ovservations later on.

Then of course you need to train the network, you need a set of situations
that you can display to the network, and then tell it if it performed
correctly or not. This may be, if possible, to train the network on-line, 
that the network at one time may predict some situation, and then at a later 
time be able to definitively tell if it predicted right or not, and that way 
learn from its mistake/success. However, you can also set up a set of cases 
offline and start by training on them first.

> And 
> if that fails, if you can also provide a few other ideas for me to look up 
> after I  get done digging into NN's :)
> 
> JD

Another idea (which better deal with time-sequences) is hidden markov models.
Although they are at-least as demanding to use as NN's. 


My background; I work with biological sequence analysis, where we use the
NN's/HMM's, to match patterns in biological sequences (i.e. DNA and proteins).

Hans Henrik Stærfeldt   |    bombman#diku,dk    | work:  hhs#cbs,dtu.dk      |
address:                |___  +45 40383492    __|__       +45 45252425     __|
 Dybendalsvej 74 2. th, | Scientific programmer at Center for Biological     |
 2720 Vanløse, Danmark. |  Sequence Analysis, Technical University of Denmark|




_______________________________________________
MUD-Dev maillist  -  MUD-Dev#kanga,nu
http://www.kanga.nu/lists/listinfo/mud-dev