[MUD-Dev] Player run reputation system

[J C Lawrence]

On Sat, 23 Jun 2001 09:24:56 -0700 (MST) 
Timothy Dang <tdang at U.Arizona.EDU> wrote:

> On Sat, 23 Jun 2001, J C Lawrence wrote:

>> Populations with common experiences, especially social
>> populations, like to talk.  Talking groups like to elect, if only
>> by self-election, leaders in the sense of "everybody knows" and
>> "everybody hears" (even if they "hear" only reputationally and
>> not in fact).  This creates an artificial and wide spread cluster
>> in the reputation graph.  I've done a couple models where,
>> literally, you get every node being no more than degrees removed
>> from the "famous" node (look at Advogato sometime for signs of
>> this (it used to be really bad)).

> What's the problem with everybody being closely connected to the
> famous? This seems very appropriate for a trust system.

Because Advogato's model is based similar to public key trust
metrics -- the expectation is that you'll personally know of and can
vouch for the individual you vote for, not that you've merely heard
good/bad things about them.  

  Real world equivalent:

    You've probably heard of and know about Larry Wall (original
    author and developer of perl, the scripting language).  For the
    sake of argument I'm going to assume that you've never met
    Larry, never spoken to or dealt with him directly, and have no
    more direct connection to him than reading his in perl-related
    discussions.

In Advogato's system a great many people were personally vouching
for Larry (or Alan Cox or Ted Tso or <isert favourite name>) not on
the basis that they knew them, but on the basis that they knew
__of__ them, and that by establishing the link from them to the
famous person that they could gain reflected glory (in Advogato's
system you can see who a person voted for) by the fact that they
knew and vouched for a famous person.

  Yes, I'm special.  I voted for Mother Theresa as a Good Person..

> I must admit I'm not entirely clear on Advogato's workings. 

They're well publicised and discussed on Advogato.

> From what I read, it delivers a binary trust/don't trust signal
> for each node, and the signal is a raandom variable (heavily
> skewed towards the proper signal). Is this correct?

Not even close.

> My preference is for decentralized reputations systems, along the
> model of collaborative filtering systems such as Firefly. In such
> a system, each person would have a unique trust level for every
> other person. I would set my trust level for some people, and the
> system would automatically calculate trust settings for others to
> match those of people who trust similarly to myself. If one
> prefers a more realistc dynamic, my trust could by updated by
> trusting those who people I trust trust (ugh). This latter is
> probably less exploitable than a pure collaborative-filtering
> model.

Please go read the Advogato design papers , and the (I think two
now, maybe three) rounds of discussion there on how well it has
worked out and the perceived problems.

Then come back.

> In such a system, there's no universal rating, so it can readily
> interact with political / faction systems. And, one would expect
> famous people or factions would be central to trust clusters.

The centrality of fame is not a problem.  Structures which are
literally flat except for infinite pillars on the famous are a
problem.

--
J C Lawrence                                             claw at kanga.nu
---------(*)                                http://www.kanga.nu/~claw/
I never claimed to be human.
_______________________________________________
MUD-Dev mailing list
MUD-Dev at kanga.nu
https://www.kanga.nu/lists/listinfo/mud-dev