(rec.games.mud.admin) Roleplaying (fwd)

[Nathan F Yospe]

I'm trolling again... decided this one might merit a repost to the list. It
may or may not interest any of you, but... oh well. *shrug* He's an LP guy,
from his earlier posts...

---------- Forwarded message ----------
Date: Fri, 27 Mar 1998 10:34:28 -1000 (HST)
From: Nathan F Yospe <yospe at hawaii.edu>
To: michael.willey at abnamro.com
Subject: Re: (rec.games.mud.admin) Roleplaying

You wrote:

:In article <6fc1ta$2ms at news.Hawaii.Edu>,
:  yospe at Hawaii.Edu (Nathan Fenenga Yospe) wrote:

:> Bad assumption... especially because everyone seems to learn the exact same
:> things in their travels at the exact same times... now, a skill system does
:> not just mean a "sensei" or something utterly generic like that... my word,
:> these stock mindsets are so ... so ... stock. Ugh. A skill system is one of
:> those either/or things if done well... combat skills, magic skills, any old
:> skill you might come up with... all are handleable in the same manner by an
:> arbitrary skill system. The concept is simple: skills are learned by a roll
:> or calculation or what have you when a character is exposed to a key... the
:> key can be a teacher, witnessing, or an accident when doing something close
:> to the new skill... skills are improved or weakened by related traits. That
:> means practice in related skills bleeds over... and using a skill only does
:> that particular skill and related skills any good.

:And having implemented a skill system very much like you're describing, I can
:tell you that it works, wonderfully.

I'd be interested in hearing more. I developed my skill web over two years,
and of course I'm rather proud of my creation. Physmud has always been more
of a platform for experimentation for me than anything else. The skill web,
as I wrote it, was one of my first efforts at migrating neural net theories
into simplified applications. It was designed as a triple layer neural net,
with attributes being the hidden layer and skills being double-linked. This
let me have skills modify output while being cross-modified. I train all of
the skills together; when a new skill or attribute is added, the entire web
must be retrained. Several test characters are used, responses are rated by
hand and fed back into the output. When it feels right, all neurons with no
more than a single percent effect (the number is flexible, the choice being
arbitrary) are deleted and the others are fixed; in short, the web has some
set of fixed relations between each skill and the attributes layer, and the
attributes layer has some fixed number of attributes. I have been making an
effort to name the attributes, but... the neural net sets the number, and I
end up half the time being completely stunned by skill relationships. There
is a problem with this: it ends up making relationships between skills pure
intuitive feedback. I thought about doing an analytic evaluation; the first
seed for the neural net had named attributes and I tried to make connection
weights on the basis of a sort of on the fly analysis; and the one thing it
had that I have not been able to retain is the linkage to physical keys; if
a skill is linked to the forearm, that physical key should have a directed,
modifiable weight. In short, I had to seperate damage to the forarm into an
applied modifier to the skill, instead of a direct component. I still think
I should be able to find a way to tie the suckers into the neural net; this
might require an extra half-layer somewhere in the attributes.
   If you found this at all interesting, get back to me, let me know that I
have it all wrong, or give me a description of how you did it... or just if
you want to continue discussion.
--

Nathan F. Yospe - Aimed High, Crashed Hard, In the Hanger, Back Flying Soon
Jr Software Engineer, Textron Systems Division (On loan to Rocketdyne Tech)
(Temporarily on Hold) Student, University of Hawaii at Manoa, Physics Dept.
yospe#hawaii.edu nyospe#premier.mhpcc.af.mil http://www2.hawaii.edu/~yospe/