nu.kanga.list.mud-dev

11944: Re: [MUD-Dev] Neural Networks as the AI system for a MUD?

Andrew Norman wrote:
>
> Just my $0.02 here but what about instead of having the NN run
> individual mobs it is trained up as a generic player that controls
> 1000s of individual characters in the mud. I sort of envisage this
> as a sort of event driven state machine where the NN looks at the
> current state and decides on its next move. If the last couple
> of moves are stored as well as a "race" type memory the system could
> exhibit a very complex behaviour.

As a big, complex state machine it might work, as an NN? Nah. The
problem with NNs is that decisions are made based upon lots of
weighting values on the links between the inputs.  They are good
at making single decisions based upon a lot of (apparently unrelated)
input data where there is a fairly lear right/wrong answer.  While
they might help with strategic level decisions (should the mob help,
hinder or ignore the player?) they aren't much good for tactical
decisions (is moving north better than moving south?).  Having
different behaviour for different mobs would require a different
set of NN weights for each mob (or more input flags dictating the
behaviour) and that gets you back to the thraining problem.

Personally I favour a rule based system, where lots of rules (and
maybe some internal variables and state machines) are used to make
decisions.

Mik





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