[
Lists Home |
Date Index |
Thread Index
]
Because as has been said, the web reflects the society
that uses it (subcultures and all), it will have all
of those problems. Humans use negotiation to resolve
conflicts (modulo violence).
The Golem Problem: when assigning both power and authority to
an artificial system, how does one prevent the system
from becoming unpredictable. The agent negotiates. Does
it do better than humans, as well, or worse? It depends
on the negotiation constraints. Given what you say about the
semantic web, here are some starting rules for an
agent working to evolve reciprocity in a simulation that
assumes repeated encounters (repeated plays of prisoner's
dilemma: in a single play, defection is likely):
1. Know the source (keiretsu).
2. Negotiate with the source (if no ontology is
a priori in place, negotitate one).
2. Trust but verify (ontological commitment).
4. Be provocable (if cooperator defects, defect).
5. Be clear (use predictable strategy. communicate trust).
6. Be forgiving (with onset of cooperation, cooperate).
7. No massive retaliation (keep response proportional).
It's just an expanded set of rules from the Generous Tit for
Tat simulations. Then there is the win-stay, lose-shift
strategy (aka, Pavlov) which is said to be better at exploiting
cooperators and not incentivizing defectors. I'm not cynical
about the semantic web per se. The non-linear aspects of it
are not so much within itself, as in the coupling of the
semantic web to the humans who use it. The importance of
human training and agent training cannot be overstated. The
tools used to train the agent make all the difference.
And don't give the agent the power to exceed its authority,
so much of the solution is in how to define authority.
len
From: Didier PH Martin [mailto:martind@netfolder.com]
a) Who will be behind the chain of trust and what the costs are. Costs in
terms of efforts and money. I mean by that, how hard will it be for an
individual to get a certificate and how trustable it will be. In the past
and not necessarily in movies, some people were able to take a different
identity than their real one. If actually some people are smarter than the
system to break it, how can we prevent them to "worm" or "virus" ontologies
and therefore the whole reasoning of automated agents? I can imagine the fun
of some people to turn our agents into some paranoid machines a la Hal (ref:
2001 space Odyssey) or to create some unknown dummy behavior.
b) How will I know that a chain of trust is trustable and virus free?
c) How the web will struggle to remains democratic and prevent "reasoning
illnesses" caused by strange people with strange motivations and annoying
behavior.
|
