On 28 July 2010 23:09, Jan Klauck jkla...@uni-osnabrueck.de wrote:
Ian Parker wrote
If we program a machine for winning a war, we must think well what
we mean by winning.
I wasn't thinking about winning a war, I was much more thinking about
sexual morality and men kissing.
If we program a machine for doing X, we must think well what we mean
by X.
Now clearer?
Winning a war is achieving your political objectives in the war. Simple
definition.
Then define your political objectives. No holes, no ambiguity, no
forgotten cases. Or does the AGI ask for our feedback during mission?
If yes, down to what detail?
With Matt's ideas it does exactly that.
The axioms which we cannot prove
should be listed. You can't prove them. Let's list them and all the
assumptions.
And then what? Cripple the AGI by applying just those theorems we can
prove? That excludes of course all those we're uncertain about. And
it's not so much a single theorem that's problematic but a system of
axioms and inference rules that changes its properties when you
modify it or that is incomplete from the beginning.
No we simply add to the axiom pool. *All* I am saying is that we must always
have a lemma train taking us to the most fundamental Suppose I say
W=AσT4
Now I ask the system to prove this. At the bottom of the lemma trail will be
Clifford algebra. This relates Bose Einstein statistics to the spin, in this
case of the photon. It is Quantum Mechanics at a very fundamental level. A
Fermion has a half in its spin.
I can introduce as many axioms as I want. I can say that i = √-1. I can call
this statement an axiom, as a counter example of your natural numbers. In
constructing Clifford Algebra I make a number of statements.
This thinking in terms of axioms I repeat does not limit the power of AGI.
If we have a database you could almost say that a lemma trail was in essence
trivial.
What is does do is invalidate the biological model. *An absolute requirement
for AGI is openness.* In other words we must be able to examine the
arguments and their validity.
Example (very plain just to make it clearer what I'm talking about):
The natural numbers N are closed against addition. But N is not
closed against subtraction, since n - m 0 where m n.
You can prove the theorem that subtracting a positive number from
another number decreases it:
http://us2.metamath.org:88/mpegif/ltsubpos.html
but you can still have a formal system that runs into problems.
In the case of N it's missing closedness, i.e., undefined area.
Now transfer this simple example to formal systems in general.
You have to prove every formal system as it is, not just a single
theorem. The behavior of an AGI isn't a single theorem but a system.
The heuristics could be tested in an off line system.
Exactly. But by definition heuristics are incomplete, their solution
space is smaller than the set of all solutions. No guarantee for the
optimal solution, just probabilities 1, elaborated hints.
Unselfishness going wrong is in fact a frightening thought. It would
in
AGI be a symptom of incompatible axioms.
Which can happen in a complex system.
Only if the definitions are vague.
I bet against this.
Better to have a system based on *democracy* in some form or other.
The rules you mention are goals and constraints. But they are heuristics
you check during runtime.
That is true. Also see above. System cannot be inscruitable.
- Ian Parker
---
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
https://www.listbox.com/member/?;
Powered by Listbox: http://www.listbox.com
---
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c
Powered by Listbox: http://www.listbox.com